[docs] dreamverse-integration: GPU4 smoke validated; flashinfer prereq

Live smoke test on GPU4 confirms public fastvideo serve from will/ltx2_sr_port HEAD d23e71c2 boots cleanly with bf16 fallback (~24s) and empirically validates drift item #4 (only /health exists; /healthz, /readyz, /status, /prompt-system-config, /curated-presets all 404 — Phase 4 promotion target).

Plan doc updates:

* CUDA_VISIBLE_DEVICES=4 pin convention documented (original Dreamverse-side server uses this; smoke commands now consistently include it).

* flashinfer-python + flash_attn marked as Phase 0 prerequisites for production-equivalent NVFP4 smoke; bf16 fallback path documented for hosts without these deps.

* Cleanup commands updated with the verified setsid/disown launch pattern that survived the smoke session.

* Smoke-test evidence (boot time, /health JSON, endpoint matrix) recorded inline as a 2026-05-05 verification anchor.

Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>

Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>

Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>

Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>
Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>
Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>
Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>
Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>

.agents/STATUS.md

@@ -1,94 +0,0 @@
-# Agent Infrastructure — Status Dashboard
-
-Developer-maintained overview of all agent components and their maturity.
-Use this to understand what exists, how complete it is, and how much to trust it.
-
-_Last synced: 2026-03-02_
-
-> To resync this dashboard, use the workflow: `.agents/workflows/sync-dashboard.md`
-
----
-
-## Summary
-
-| Category | Total | ✅ Ready | 🟡 Draft | 🔴 Stub | Trust |
-|----------|-------|---------|---------|---------|-------|
-| Skills | 8 | 0 | 8 | 0 | Low — newly created, untested |
-| Workflows (SOPs) | 4 | 0 | 4 | 0 | Low — newly created, untested |
-| Memory files | 4 | 1 | 3 | 0 | Medium — codebase_map is solid |
-| Lessons | 0 | — | — | — | N/A — empty |
-| Exploration logs | 0 | — | — | — | N/A — empty |
-
----
-
-## Skills (`.agents/skills/`)
-
-| Skill | File | Status | Trust | Tested | Notes |
-|-------|------|--------|-------|--------|-------|
-| Launch Experiment | `launch-experiment.md` | 🟡 Draft | Low | ❌ | Needs dry-run validation |
-| Monitor Experiment | `monitor-experiment.md` | 🟡 Draft | Low | ❌ | Requires W&B API access to test |
-| Summarize Run | `summarize-run.md` | 🟡 Draft | Low | ❌ | Pattern from existing test infra |
-| Log Experiment | `log-experiment.md` | 🟡 Draft | Low | ❌ | Journal formatting only |
-| Evaluate Video Quality | `evaluate-video-quality.md` | 🟡 Draft | Low | ❌ | SSIM section most mature |
-| Index Related Work | `index-related-work.md` | 🟡 Draft | Low | ❌ | Schema defined, no entries yet |
-| Search Related Work | `search-related-work.md` | 🟡 Draft | Low | ❌ | Depends on indexed entries |
-| Skill Template | `SKILL_TEMPLATE.md` | ✅ Ready | High | ✅ | Meta-template, stable |
-
-### Trust Level Definitions
-- **High**: Tested in production, validated against real experiments
-- **Medium**: Logic is sound, partially tested or based on existing patterns
-- **Low**: Newly written, not yet validated
-- **None**: Placeholder only
-
----
-
-## Workflows / SOPs (`.agents/workflows/`)
-
-| Workflow | File | Status | Trust | Tested | Notes |
-|----------|------|--------|-------|--------|-------|
-| Experiment Lifecycle | `experiment-lifecycle.md` | 🟡 Draft | Low | ❌ | End-to-end flow, untested |
-| Evaluation Development | `evaluation-development.md` | 🟡 Draft | Low | ❌ | Metric dev process |
-| Experiment Journaling | `experiment-journaling.md` | 🟡 Draft | Low | ❌ | Journaling cadence |
-| Lesson Capture | `lesson-capture.md` | 🟡 Draft | Low | ❌ | Post-experiment reflection |
-| Sync Dashboard | `sync-dashboard.md` | 🟡 Draft | Low | ❌ | This dashboard's updater |
-
----
-
-## Memory (`.agents/memory/`)
-
-| File | Status | Trust | Notes |
-|------|--------|-------|-------|
-| `codebase_map.md` | ✅ Ready | High | Synthesized from full repo research |
-| `experiment_journal.md` | 🟡 Draft | Medium | Schema defined, no entries yet |
-| `evaluation_registry.md` | 🟡 Draft | Medium | SSIM/loss metrics documented |
-| `related_work/README.md` | 🟡 Draft | Medium | Schema defined, no entries yet |
-
----
-
-## Lessons (`.agents/lessons/`)
-
-| File | Category | Severity | Notes |
-|------|----------|----------|-------|
-
-_No lessons captured yet._
-
----
-
-## Exploration Logs (`.agents/exploration/`)
-
-| File | Status | Topic | Notes |
-|------|--------|-------|-------|
-
-_No exploration logs yet._
-
----
-
-## What to Do Next
-
-1. **Validate skills**: Run a minimal training experiment using the
-   `experiment-lifecycle` SOP to test `launch-experiment` → `monitor-experiment`
-   → `summarize-run` end-to-end.
-2. **Index first related work**: Use `index-related-work` to add at least one
-   paper (e.g., the Self-Forcing paper used in the codebase).
-3. **Capture first lesson**: After the validation run, capture any findings.
-4. **Promote to Ready**: As each skill/SOP is tested, update its status here.

.agents/memory/codebase-map/README.md

@@ -119,7 +119,7 @@ FastVideo-WorldModel/
 ## Build & Test Commands
 
 ```bash
-uv pip install -e .[dev]                          # Editable install
+uv pip install -e ".[dev]"                        # Editable install
 pre-commit run --all-files                        # Lint/format/spell
 pytest tests/                                     # Top-level tests
 pytest fastvideo/tests/ -v                        # Package tests

.agents/memory/dreamverse-integration/README.md

@@ -0,0 +1,163 @@
+# Dreamverse Integration — Memory Index
+
+Living knowledge base for the FastVideo ↔ Dreamverse ↔ Dynamo integration.
+Tracks the public API refactor (PRs 0-17), the LTX-2 streaming server
+upstream, the Dreamverse switch from `FastVideo-internal` to public
+`FastVideo`, and the NVFP4 quantization landing.
+
+**Last reconciled:** 2026-05-05 (**D-18**: Option B+ chosen — Dreamverse
+becomes `apps/dreamverse/` subfolder under FastVideo; generic backend stays
+at `fastvideo.entrypoints.streaming.*`. See [integration-plan.md](integration-plan.md)
+for the executable 7-phase migration plan;
+[integration-review.md](integration-review.md) is **deprecated** but kept
+for the drift audit and OSS precedent citations.).
+FastVideo `will/ltx2_sr_port` @ HEAD (post-D-17 STACK.md removal +
+integration-review.md addition + integration-plan.md addition + D-18
+reconciliation). Dreamverse `will/integrate-public-fastvideo` @ `ec8ef92`.
+PRs #1257 / #1258 / #1284 / #1286 MERGED to main. **PR #1287 CLOSED
+(in favor of consolidation); PR #1288 OPEN as the single mega-PR
+landing the entire `will/ltx2_sr_port` chain at once** (LTX-2 SR
+runtime + NVFP4 + `generate_async`/Dynamo contract + agents memory dir).
+Split branches kept as historical bookmarks; STACK.md model **abandoned** —
+see [decisions-log.md D-17](decisions-log.md#d-17). Local backup
+`will/ltx2_sr_port-pre-1286-rebase` @ `1baa60bb` preserves the
+pre-rebase chain.
+
+## Fresh-context onboarding (read in order)
+
+If you're an agent picking up this work for the first time, do these
+**5 things in this order**. Once done, you have full context to continue
+any open thread, commit correctly, push, and propagate to the open PR.
+
+1. **Confirm worktree state** — run the "First 60 seconds" block in
+   [runbook.md](runbook.md). Tells you the branch is right, services
+   are up, and PR #1286's head matches what this dir claims.
+
+2. **Read [state.md](state.md)** — single-page snapshot of branch tips,
+   live services, test status, pre-existing failures, "do not pop"
+   stashes.
+
+3. **Read [pr-roadmap.md](pr-roadmap.md)** — what PRs landed, what's in
+   flight, what's planned. Identifies the active open PR (currently
+   #1286) and where it sits in the dependency chain.
+
+4. **Read [open-threads.md](open-threads.md)** — prioritized work items
+   with effort estimates and dependencies. The "Recommended pull order"
+   section is a ready-made TODO list if you need one.
+
+5. **Skim [runbook.md](runbook.md) end-to-end** — operational how-to:
+   verify, commit (with co-author trailers), push, propagate to PR
+   #1286, maintain the memory dir, and the "Common pitfalls" section
+   that catches the recurring traps.
+
+Skip the deep-context docs (design / streaming-server / cross-repo /
+quantization / decisions-log) until you need them — they're indexed in
+the "Deep-dive reading guide" below.
+
+Final check: run the "Self-test" block at the bottom of
+[runbook.md](runbook.md). If you can answer all 8 questions from this
+dir alone, you're ready. If you can't, the gap is a memory-dir bug —
+file it in [open-threads.md](open-threads.md) before continuing.
+
+## Deep-dive reading guide
+
+| Question / task | File |
+|---|---|
+| "What's running right now? What just landed?" | [state.md](state.md) |
+| "How do I commit / push / propagate to PR #1286?" | [runbook.md](runbook.md) |
+| "Why is the schema typed this way? What's the philosophy?" | [design.md](design.md) |
+| "What PRs landed? In flight? Planned?" | [pr-roadmap.md](pr-roadmap.md) |
+| "Streaming server, `generate_async`, `build_app` routes?" | [streaming-server.md](streaming-server.md) |
+| "How does Dreamverse use FastVideo? What about Dynamo?" | [cross-repo-surfaces.md](cross-repo-surfaces.md) |
+| "NVFP4? Layer profiles? `LinearBase` fallback? AbsMaxFP8?" | [quantization.md](quantization.md) |
+| "Why was decision X made? What's resolved vs. open?" | [decisions-log.md](decisions-log.md) |
+| "What should I work on next? Priority order?" | [open-threads.md](open-threads.md) |
+| "Who should be co-authored on commits in this scope?" | [authors.md](authors.md) |
+| "How do we execute the Dreamverse → FastVideo monorepo merge?" | [integration-plan.md](integration-plan.md) ← **CURRENT** |
+| "Historical drift audit + Option-D evaluation (deprecated by D-18)" | [integration-review.md](integration-review.md) (DEPRECATED) |
+
+## Repo + worktree paths
+
+| Repo | Path | Active branch |
+|---|---|---|
+| FastVideo (public) | `/home/william5lin/FastVideo` | `will/ltx2_sr_port` |
+| Dreamverse | `/home/william5lin/Dreamverse` | `will/integrate-public-fastvideo` |
+| FastVideo-internal (read-only ref) | `/home/william5lin/FastVideo-internal` | their `main` |
+| Dynamo (read-only ref) | `/home/william5lin/dynamo` | upstream |
+
+## Glossary
+
+- **NVFP4**: NVIDIA's specific block-scaled FP4 (e2m1 mantissa, fp32 alpha,
+  `layout_128x4` scale layout, group size 16). Distinct from MX-FP4 / OCP-FP4.
+- **`GeneratorConfig`**: typed init-time public config (model_path, engine,
+  pipeline). Replaces flat `from_pretrained(**kwargs)`.
+- **`GenerationRequest`**: typed per-call request (prompt, inputs, sampling,
+  runtime, output, stage_overrides, state, plan, extensions). Replaces flat
+  `generate_video(**kwargs)`.
+- **`ServeConfig`** / **`RunConfig`**: top-level YAML envelopes. ServeConfig
+  for `fastvideo serve`; RunConfig for offline `fastvideo generate`.
+- **`InferencePreset`**: model-owned named preset (e.g. `ltx2_two_stage`)
+  defining stage topology + per-stage defaults + valid override types.
+- **`ContinuationState`**: opaque round-trip state envelope `{kind, payload}`.
+  Hybrid: server-held for streaming WS, client-round-trip for stateless HTTP.
+- **`generate_async`**: future canonical async exec API (PR 7.10) yielding
+  `VideoProgressEvent` / `VideoPartialEvent` / `VideoFinalEvent`. Substrate
+  for streaming server, OpenAI server, AND Dynamo backend.
+- **`build_app`**: FastAPI app factory in
+  `fastvideo.entrypoints.streaming.server`. Currently exposes only
+  `/health` + `/v1/stream`. FE-required `/healthz`+`/readyz`+`/status`
+  migration is open follow-up #1.
+- **`LLMProvider`**: protocol abstraction for prompt enhancer providers
+  (cerebras, cerebras_ifm, groq). Public schema currently restricts to
+  `Literal["cerebras", "groq"]`; `cerebras_ifm` is internal-only.
+- **`compat.py`**: legacy kwargs translation layer (~370 lines). Scheduled
+  for death across PRs 14-17.
+- **`prepare_for_compile`**: duck-type protocol method called via
+  `getattr(module, "prepare_for_compile", None)` before `torch.compile`.
+  Currently only Gemma3 implements it.
+- **`SubprocessGpuPool`**: PR 7.6 public replacement for the internal
+  `realtime/local_runtime.GPUPool`. Per-GPU subprocess workers, typed
+  `GeneratorConfig` boundary.
+- **PR 5.5**: streaming server subpackage skeleton — adds
+  `fastvideo/entrypoints/streaming/` parallel to `openai/`.
+- **PR 7.10**: the unlock PR. Closes Q-5 (audio re-encode), Q-9 (Dynamo
+  progress), and PR 7.5's mid-segment cancellation TODO simultaneously.
+
+## Live process map (as of 2026-05-03)
+
+| Port | Service | Source |
+|---|---|---|
+| 8009 | `dreamverse-server` | running, `/readyz` 200, 1 warmed GPU worker |
+| 5274 | `next-server` (dev) | running |
+| 8000 | unknown FastAPI | not in handoff — verify before launching new BE |
+
+## How this directory is maintained
+
+- Source of truth for the integration story. Update when state changes.
+- Each file has a "Last updated" header; bump when you edit.
+- Cross-reference siblings via relative links; do NOT duplicate content.
+- New entries: register in `../index.jsonl`.
+- These files supersede the untracked source docs in the repo root and
+  `.agents/exploration/` — see [state.md](state.md) "Untracked but
+  present" section for disposition.
+
+## Source documents (archived 2026-05-03)
+
+The 7 source docs that this directory consolidates have been moved into
+[`source-archive/`](source-archive/). They remain available for agents
+who want the full unsynthesized rationale, but the synthesized memory
+files in this dir are the canonical source of truth.
+
+| Source doc | Lines | Synthesized into |
+|---|---|---|
+| [`source-archive/apirefactor.md`](source-archive/apirefactor.md) | 838 | [design.md](design.md) |
+| [`source-archive/PR-plan.md`](source-archive/PR-plan.md) | 1145 | [pr-roadmap.md](pr-roadmap.md) |
+| [`source-archive/dreamverse_review.md`](source-archive/dreamverse_review.md) | 390 | [state.md](state.md) + [decisions-log.md](decisions-log.md) |
+| [`source-archive/handoff-nvfp4-launch-demo.md`](source-archive/handoff-nvfp4-launch-demo.md) | 518 | [state.md](state.md) + [quantization.md](quantization.md) + [open-threads.md](open-threads.md) |
+| [`source-archive/streaming-server-upstream-plan.md`](source-archive/streaming-server-upstream-plan.md) | 539 | [streaming-server.md](streaming-server.md) + [decisions-log.md](decisions-log.md) |
+| [`source-archive/dreamverse_integration.md`](source-archive/dreamverse_integration.md) | 285 | [cross-repo-surfaces.md](cross-repo-surfaces.md) |
+| [`source-archive/video-generator-config-api-design.md`](source-archive/video-generator-config-api-design.md) | 93 | [design.md](design.md) (early-draft material) |
+| `.agents/exploration/pr-link-review.md` | 29 | already promoted to `.agents/skills/review-pr-link/` (kept in exploration dir) |
+
+See [`source-archive/README.md`](source-archive/README.md) for the
+archive policy.

.agents/memory/dreamverse-integration/authors.md

@@ -0,0 +1,150 @@
+# Authors — Dreamverse Integration
+
+**Status:** PERMANENT — keep around as the source of truth for who collaborated
+on the dreamverse-integration work, even after every PR in the integration
+scope has merged.
+**Last updated:** 2026-05-05 (strategy reversal — single mega-PR #1288 on `will/ltx2_sr_port` replaces planned 6-PR split; #1287 closed; per [decisions-log.md D-17](decisions-log.md#d-17))
+
+This file documents the human co-authors credited on every commit in the
+dreamverse-integration scope (FastVideo public-API refactor, streaming server
+upstream, GPU pool, prompt enhancer, NVFP4 wire-up, LTX-2 SR port). The
+4 collaborators below worked on the FastVideo-internal precursor of this code
+and are credited as co-authors on every public-side upstream commit via Git's
+standard
+[`Co-authored-by`](https://docs.github.com/en/pull-requests/committing-changes-to-your-project/creating-and-editing-commits/creating-a-commit-with-multiple-authors)
+trailer convention.
+
+Scope-wise this is the dreamverse-integration-flavored mirror of the
+top-level [`CO-AUTHORS.md`](../../../CO-AUTHORS.md), which is scoped to the
+broader `will/ltx2_sr_port` 10-PR stack. The roster is identical; this file
+exists so the dreamverse-integration memory dir is self-contained and
+discoverable without traversing to the repo root.
+
+## Co-author roster
+
+| GitHub user | Real name | GitHub ID | Trailer email |
+|---|---|---|---|
+| [`@Davids048`](https://github.com/Davids048) | Junda (David) Su | 90978028 | `90978028+Davids048@users.noreply.github.com` |
+| [`@RandNMR73`](https://github.com/RandNMR73) | Matthew Noto | 99706358 | `99706358+RandNMR73@users.noreply.github.com` |
+| [`@XOR-op`](https://github.com/XOR-op) | (unset) | 17672363 | `17672363+XOR-op@users.noreply.github.com` |
+| [`@jzhang38`](https://github.com/jzhang38) | Zhang Peiyuan | 42993249 | `42993249+jzhang38@users.noreply.github.com` |
+
+## Verification — where these trailers appear
+
+Verified via `gh pr view <PR> --json commits --jq '.commits[].messageBody'`
+across every PR in the integration scope:
+
+| PR | Branch | Status | Trailers present on every commit |
+|---|---|---|---|
+| #1257 | `will/api_7.6` (GPU pool upstream) | ✅ merged 2026-05-04 | yes (4/4) |
+| #1258 | `will/api_7.7` (prompt enhancer + LLMProvider) | ✅ merged 2026-05-04 | yes (3/3) |
+| #1284 | `will/api_7.8` (streaming auxiliaries) | ✅ merged 2026-05-04 | yes (2/2) |
+| #1286 | `will/api_7.9` (streaming router) | ✅ merged 2026-05-05 at `2aaeee2a` (squash) | yes on commits 1-3; commit `a152cb77` (`[fix] streaming: router polish`) was missing trailers but got squashed into the merge commit, so the merge commit on main inherits the trailers from the other 3. The trailerless cherry-pick partner (`40e265b8` on `will/ltx2_sr_port`) was dropped by the post-#1286 rebase — gap permanently resolved. |
+| #1287 | `will/api_7.10` (`generate_async` + `VideoEvent`) | ❌ CLOSED 2026-05-05 — superseded by #1288 per [D-17](decisions-log.md#d-17) | yes on all 3 commits (now part of #1288's chain) |
+| **#1288** | **`will/ltx2_sr_port`** (mega-PR — full stack: SR runtime + NVFP4 + generate_async + Dynamo contract + agents memory + integration-review) | 🟢 OPEN, MERGEABLE at `b36bdbc9`, 36 commits / 70 files / ~+13.0k LOC (post STACK.md removal) | yes on all 36 commits |
+
+Aggregate count across `will/ltx2_sr_port` (top of stack) at the time of
+writing: 32-33 commits per co-author, matching the 32 commits in the stack
+on top of base `cfccd292`. Numbers stay consistent because the rebase
+command (see "How the trailers were applied" below) walks every commit.
+
+## Trailer block (copy-paste ready)
+
+The trailers added to every commit on `will/ltx2_sr_port` and every
+dreamverse-integration PR:
+
+```
+Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>
+Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>
+Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>
+Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>
+```
+
+For one-off `git commit -m` invocations, use `--trailer` flags:
+
+```bash
+git commit -m "..." \
+  --trailer "Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>" \
+  --trailer "Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>" \
+  --trailer "Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>" \
+  --trailer "Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>"
+```
+
+`--trailer` is idempotent (dedupes by full `key: value`) so re-running is safe.
+
+## Why no-reply emails
+
+GitHub's `<id>+<username>@users.noreply.github.com` form is the most reliable
+way to link a `Co-authored-by` trailer to a GitHub account. It:
+
+- Always works regardless of whether the user has a public verified email
+- Survives the user changing their primary email
+- Doesn't expose anyone's personal email to git history
+- Is the format GitHub itself produces when you click "Add co-author" in the
+  web UI
+
+(All 4 collaborators have this email already used in `FastVideo-internal`
+git history, verified via `git log --all` on that repo.)
+
+## How the trailers were applied (bulk rebase)
+
+```bash
+git rebase --exec '
+  git commit --amend --no-edit \
+    --trailer "Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>" \
+    --trailer "Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>" \
+    --trailer "Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>" \
+    --trailer "Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>"
+' origin/main will/ltx2_sr_port
+```
+
+After running, re-slice all 10 split branches per [`STACK.md`](../../../STACK.md)
+and force-push the published branches (`will/api_7.9`, `will/ltx2_sr_port`).
+
+## How to add a new co-author later
+
+1. Add the user to the roster table above (and the top-level
+   [`CO-AUTHORS.md`](../../../CO-AUTHORS.md) — keep them in sync).
+2. Append their `Co-authored-by` line to the trailer block above.
+3. Re-run the bulk rebase command on `will/ltx2_sr_port` — git's trailer
+   dedupe handles the existing 4; the new one gets appended.
+4. Re-slice all split branches per [`STACK.md`](../../../STACK.md).
+5. Force-push the published branches.
+
+## What we do NOT add
+
+Per the repo's top-level [`AGENTS.md`](../../../AGENTS.md):
+
+> Never add any coding agent or models such as Claude (or Claude Code), GPT,
+> Codex or others as a co-author in commits or PRs. Do not include
+> `Co-Authored-By: Claude ...` trailers or "Generated with Claude Code" and
+> other such lines.
+
+So no `Co-authored-by: Claude <noreply@anthropic.com>`, no
+`Generated with Claude Code` footer, no `Cursor <cursoragent@cursor.com>`
+trailer (one such commit exists on `will/ltx2_sr_port` from a pre-policy
+external contribution and stays grandfathered; new commits MUST NOT introduce
+the pattern). Only human collaborators.
+
+## Known gaps
+
+**Resolved 2026-05-05 by the post-#1286 rebase.** The two trailerless
+commits (`a152cb77` on `will/api_7.9` and `40e265b8` on
+`will/ltx2_sr_port`) are no longer reachable from any active branch:
+
+- `a152cb77` was absorbed into squash merge `2aaeee2a` on main, which
+  inherits the trailers from the other 3 commits in the squash.
+- `40e265b8` was dropped by the post-#1286 rebase of
+  `will/ltx2_sr_port`.
+
+Both still exist on the local backup `will/ltx2_sr_port-pre-1286-rebase`
+for archeological reference. No further action needed.
+
+## See also
+
+- [`../../../CO-AUTHORS.md`](../../../CO-AUTHORS.md) — top-level stack-scoped
+  co-authors file (same roster, broader scope)
+- [`../../../STACK.md`](../../../STACK.md) — 10-PR split layout for
+  `will/ltx2_sr_port` (re-slice commands live here)
+- [`pr-roadmap.md`](pr-roadmap.md) — per-PR status within the
+  dreamverse-integration scope

.agents/memory/dreamverse-integration/cross-repo-surfaces.md

@@ -0,0 +1,277 @@
+# Cross-Repo Surfaces — Dreamverse + Dynamo
+
+How Dreamverse consumes FastVideo today, what's already shared, what's
+ad hoc, and what migrations land alongside each PR. Plus the Dynamo
+backend contract.
+
+For the streaming-server side see [streaming-server.md](streaming-server.md).
+For the API design see [design.md](design.md). For PR sequence see
+[pr-roadmap.md](pr-roadmap.md).
+
+**Last updated:** 2026-05-03.
+
+## The three surfaces
+
+Dreamverse depends on FastVideo across three surfaces (in order of
+stability):
+
+1. **Pipeline construction** (stable)
+2. **Realtime runtime** (in flight: PRs 7.5/7.6)
+3. **Continuation state** (PR 7 typed; PR 7.6 wires server-held)
+
+## Surface 1: Pipeline construction (stable)
+
+`Dreamverse/server/video_generation.py:VideoGenerationWorker` calls
+`VideoGenerator.from_pretrained(...)`.
+
+After PR 6 the typed `GeneratorConfig` path exists; **as of `d80c2a8`
+(May 2)** Dreamverse migrated to the typed path:
+
+| Dreamverse usage | FastVideo public surface (post-PR 6) |
+|---|---|
+| `VideoGenerator.from_pretrained(model_path, ltx2_refine_enabled=…, …)` | `VideoGenerator.from_pretrained(config=GeneratorConfig(...))` |
+| Flat `torch_compile_kwargs={…}` dict | `engine.compile.{backend,fullgraph,mode,dynamic,extras}` |
+| `ltx2_vae_tiling=True` | `pipeline.vae_tiling=True` |
+| `ltx2_refine_*` family | `pipeline.preset_overrides.refine.*` + `pipeline.components.upsampler_weights` |
+| `enable_torch_compile_text_encoder` | `engine.compile.text_encoder_enabled` |
+
+Refine knobs moved from `ltx2_refine_*` flat kwargs into
+`preset_overrides["refine"]`. **The in-memory `pipeline_config` pin**
+(`dit_config.quant_config = NVFP4Config()`) keeps using the legacy
+`experimental["pipeline_config"]` carrier because typed
+`transformer_quant: "NVFP4"` doesn't yet support setting
+`layer_profile` (see [open-threads.md](open-threads.md) follow-up #4 +
+[quantization.md](quantization.md)).
+
+Legacy flat-kwarg path stays supported via `compat.py`; migration is
+opt-in. PR 13's deprecation warnings are the eventual nudge.
+
+## Surface 2: Realtime runtime (in flight: PRs 7.5–7.6)
+
+`Dreamverse/server/runtime/factory.py` selects a runtime backend at
+process start:
+
+```python
+def create_runtime_pool() -> RuntimePool:
+    if os.getenv("FASTVIDEO_REALTIME_BASE_URL"):
+        return FastVideoRealtimePool(base_url=..., ws_url=..., default_model_id=...)
+    return GPUPool(get_available_gpus())  # in-process, wraps
+                                            # fastvideo.entrypoints.realtime.local_runtime
+```
+
+Both backends speak the same `RuntimePool` / `RuntimeSlot` Protocol
+(`server/runtime/interfaces.py`):
+
+- `acquire(client_id, websocket=None) -> (gpu_id, RuntimeSlot)`
+- `release(client_id)`
+- `RuntimeSlot.{join_user, user_step, leave_user, register_stream_queue, ...}`
+
+Today both impls reach into FastVideo-internal's
+`fastvideo.entrypoints.realtime.local_runtime` (which exposes
+`RealtimeRuntimeConfig`, `GPUPool`, `GPUSlot`). The remote backend talks
+HTTP+WS to a separately-deployed runtime of the same shape.
+
+**Contract that PR 7.5/7.6 must preserve:**
+
+- `RealtimeRuntimeConfig` accepts `model_registry`, `default_model_id`,
+  `default_height/width/num_frames/fps/num_inference_steps/guidance_scale/seed/negative_prompt`,
+  `default_ltx2_image_crf`, `startup_warmup_{enabled,prompt,timeout_seconds}`.
+- `GPUPool(gpu_ids: list[int], config: RealtimeRuntimeConfig)` constructor.
+- `pool.initialize() / shutdown() / acquire() / release() / get_status()`.
+- HTTP endpoints on the remote variant: `GET /healthz`, `GET /readyz`,
+  `GET /status`, `WS /ws`. (Already match what
+  `Dreamverse/server/routes/health.py` consumes.)
+
+**These three health routes still need to migrate into FastVideo's
+`build_app` to make `BE_FLAVOR=fastvideo` FE-compatible** — see
+[streaming-server.md](streaming-server.md) "build_app route contract" +
+[open-threads.md](open-threads.md) follow-up #1.
+
+When PR 7.6 lands the upstream of `fastvideo/entrypoints/realtime/`,
+Dreamverse should not need any code change unless the import path
+renames. Decided: keep `streaming/` (post-PR-5.5 public name); ship
+`realtime/__init__.py` as a re-export with `DeprecationWarning` for one
+release cycle.
+
+### Note on `default_ltx2_image_crf`
+
+Dreamverse's `RealtimeRuntimeConfig` includes `default_ltx2_image_crf`.
+The April 26 Dreamverse review (D-8) showed this getting passed to
+`SamplingParam(...)` and **silently dropped** by the public schema. Post
+`d80c2a8` (May 2 typed-config refactor), the migration target is
+`request.stage_overrides.refine.image_crf` (per
+[design.md](design.md) compatibility mapping table).
+
+**Whether `d80c2a8` actually wired this through, or it's still latent,
+is unverified.** See [open-threads.md](open-threads.md) item D-8.
+
+## Surface 3: Continuation state (PR 7)
+
+`Dreamverse/server/video_generation.py:89 ContinuationState` is
+Dreamverse's hand-rolled per-session state holder. PR 7 introduced the
+typed equivalent at
+[`fastvideo/pipelines/basic/ltx2/continuation.py`](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/continuation.py).
+
+### Field mapping
+
+| Dreamverse | PR 7 `LTX2ContinuationState` | Notes |
+|---|---|---|
+| `video_images: list[PIL.Image]` | `video_frames: list[np.ndarray]` (uint8 H×W×3) | numpy is leaner; Dreamverse already round-trips PIL→numpy→PIL just to add noise |
+| `audio_latents: torch.Tensor` `[B, C, T, mel]` | `audio_latents: torch.Tensor` (safetensors-serialized; bf16-safe) | unchanged shape; safetensors preserves bf16 |
+| `LTX2_VIDEO_CONDITIONING_FRAME_IDX` (env) | `video_conditioning_frame_idx: int` | env constant → per-state field |
+| `LTX2_VIDEO_CONDITIONING_STRENGTH` (env) | `video_conditioning_strength: float` | env constant → per-state field |
+| `AUDIO_CONDITIONING_NUM_FRAMES` (env) | `audio_conditioning_num_frames: int` | env constant → per-state field |
+| `AUDIO_CONDITIONING_STRENGTH` (env) | `audio_conditioning_strength: float` | env constant → per-state field |
+| `audio_lps` (passed into `apply_audio`) | `audio_sample_rate: int \| None` | analogous; rename worth confirming with audio team |
+| Computed `prefix_sec` per segment | `video_position_offset_sec: float` | **see open question below** |
+| `segment_idx` (param to `apply_*`) | `segment_index: int` | per-state field |
+| `VIDEO_CONTEXT_NOISE`, `AUDIO_CONTEXT_NOISE`, `ENABLE_AUDIO_COND` | not on state | runtime policy / regularization knobs, not portable session data |
+| `apply_video / apply_audio / save_video / save_audio_latents / clear` | not on PR-7 state class | state is a pure data carrier; runtime owns lifecycle policy |
+
+PR 7 is a strict superset of Dreamverse's data model **plus** lifts
+several env globals into per-session typed fields.
+
+### Lifecycle mapping
+
+| Dreamverse pattern | `SessionStore` API |
+|---|---|
+| `self.continuation = ContinuationState()` per session | `state = session_store.snapshot(sid) or LTX2ContinuationState()` |
+| `apply_video(req_kwargs, segment_idx)` + `apply_audio(req_kwargs, segment_idx, audio_lps)` | `state = session_store.snapshot(sid)`; runtime builds request from `state.video_frames` / `state.audio_latents` |
+| `save_video(frames)` + `save_audio_latents(latents)` | runtime constructs new `LTX2ContinuationState`, `session_store.store(sid, ...)` |
+| `clear()` at end of session | `session_store.drop(sid)` |
+
+`SessionStore` and `BlobStore` ABCs ship with thread-safe in-memory
+defaults (`InMemorySessionStore`, `InMemoryBlobStore`). Dreamverse can
+adopt them as-is for the local runtime; remote runtimes can plug in
+redis-backed implementations later.
+
+### Wire format (HTTP/WS round-trip)
+
+Dreamverse's `FastVideoRealtimePool` already speaks the realtime
+runtime's HTTP+WS protocol. When PR 7.5/7.6 land state emission on the
+server side, the on-the-wire payload is the public envelope:
+
+```json
+{
+  "kind": "ltx2.v1",
+  "payload": {
+    "schema_version": 1,
+    "segment_index": 3,
+    "video_conditioning_frame_idx": 9,
+    "video_conditioning_strength": 0.75,
+    "audio_sample_rate": 24000,
+    "audio_conditioning_num_frames": 5,
+    "audio_conditioning_strength": 0.5,
+    "video_position_offset_sec": 0.2,
+    "video": {"frames_b64": ["..."]},
+    "audio": {"safetensors_b64": "..."},
+    "metadata": {}
+  }
+}
+```
+
+JSON-serializable end-to-end; safetensors blob preserves audio dtype
+(incl. bf16). For payloads above the inline threshold a `BlobStore`
+indirection replaces the b64-encoded body with `{"blob_id": "..."}`;
+the blob itself stays inside the runtime that produced it.
+
+## Migration plan per PR
+
+| PR | Dreamverse action |
+|---|---|
+| PR 6 (landed) | Typed `GeneratorConfig` available; flat-kwarg path still works via compat. Optional migration. |
+| PR 7 (landed) | Typed `LTX2ContinuationState` available. ~50-line Dreamverse PR: replace `server/video_generation.py:89` import; move `apply_*`/`save_*`/`clear` off the state class onto `VideoGenerationWorker`; read knobs from typed state instead of env globals; swap `list[PIL.Image]` → `list[np.ndarray]`. |
+| PR 7.5 (open) | Streaming server skeleton — Dreamverse's `runtime/factory.py` either keeps building `GPUPool` from `RealtimeRuntimeConfig` (current path), or migrates to `ServeConfig.streaming` shape and invokes `fastvideo serve --config realtime.yaml`. Dreamverse's `RuntimePool`/`RuntimeSlot` Protocol can stay in place. |
+| PR 7.6 (branch ready) | GPU pool upstream — `local_runtime.py` import becomes a public import with same symbols (`RealtimeRuntimeConfig`, `GPUPool`, `get_available_gpus`). Per-GPU continuation state inside the worker becomes a `SessionStore` reference (Dreamverse doesn't see this). `request.state` / `result.state` round-trip starts working end-to-end on the local runtime. |
+| PR 7.10 (planned) | `generate_async` is canonical. Dreamverse's per-segment `user_step` flow can migrate from sync `generate_video(..., **kwargs)` to consuming the typed event stream. Optional; sync wrapper stays. |
+
+## Dynamo backend contract
+
+**FastVideo does not host any Dynamo code.** The backend package
+(`args.py`, `main.py`, `backend.py`, `register.py`, `health_check.py`,
+adapter, Dockerfile) lives entirely in the Dynamo repo at
+`components/src/dynamo/fastvideo/`, modeled on
+`components/src/dynamo/sglang/`.
+
+FastVideo's only obligation is to expose a stable, typed Python API
+that Dynamo's backend package imports.
+
+### Contract surface
+
+| Surface | Exposed as |
+|---|---|
+| Construction | `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` (typed_kwargs = a stable subset from `GeneratorConfig`; no flat LTX2 legacy) |
+| Sync execution | `generator.generate_video(request: GenerationRequest) -> VideoResult` |
+| Async execution | `generator.generate_async(request: GenerationRequest) -> AsyncGenerator[VideoEvent, None]` (PR 7.10) |
+| Typed request | `fastvideo.api.GenerationRequest`, `SamplingConfig`, `InputConfig` |
+| Typed result | `VideoResult` with `video_bytes` or tensor frames + optional `ContinuationState` |
+| Continuation | `ContinuationState(kind, payload)` — schema-versioned payloads |
+| Health-check input | `VideoGenerator.default_health_check_request() -> GenerationRequest` (256x256 / 8 frames / 1 step) |
+| Config dump | `GeneratorConfig.to_dict()` / `ServeConfig.to_dict()` |
+
+### Request/response mapping (Dynamo ↔ FastVideo)
+
+```
+NvCreateVideoRequest          ->  fastvideo.api.GenerationRequest
+  prompt                      ->    sampling.prompt
+  size="WxH"                  ->    sampling.width, sampling.height
+  seconds                     ->    (seconds * nvext.fps) -> sampling.num_frames
+  input_reference             ->    input.image_path / input.video_path
+  nvext.fps                   ->    sampling.fps
+  nvext.num_frames            ->    sampling.num_frames (overrides seconds*fps)
+  nvext.num_inference_steps   ->    sampling.num_inference_steps
+  nvext.guidance_scale        ->    sampling.guidance_scale
+  nvext.seed                  ->    sampling.seed
+  nvext.negative_prompt       ->    sampling.negative_prompt
+  response_format             ->    (handled by adapter at output)
+
+VideoFinalEvent               ->  NvVideosResponse
+  video_bytes                 ->    data[0].b64_json  (if response_format=b64_json)
+  video_url (after upload)    ->    data[0].url       (if response_format=url)
+  metadata.inference_time_s   ->    inference_time_s
+```
+
+All fields exist on FastVideo's typed schema after PR 6 expansion (typed
+LTX2 kwargs) + PR 7.10 (`generate_async` + health check).
+
+### Reference: PR ai-dynamo/dynamo#7544
+
+Closed draft establishing the Dynamo backend shape. Two frictions
+identified:
+
+1. Flat legacy LTX2 kwargs — solved by PR 6.
+2. Sync-only generation — solved by PR 7.10's `generate_async`.
+
+Next iteration of this PR (or its successor) will reopen against PR 8's
+docs reference and land cleanly.
+
+## Open questions across surfaces
+
+| # | Question | Source | Status |
+|---|---|---|---|
+| Q-1 | Multi-model GPU pool | dreamverse_review D-1 | Deferred (production single-model) |
+| Q-2 | LTX-2 prompt orchestration promotion to public | dreamverse_review D-2 | Open; consumer-side until 2nd consumer |
+| Q-3 | Race-based provider fallback | dreamverse_review D-3 | Open; sequential is current public |
+| Q-4 | Router upstream skip on Dreamverse | dreamverse_review D-4 | Resolved (PR 7.9 lands publicly, Dreamverse doesn't consume) |
+| Q-5 / D-5 | `generate_async` cutover (audio re-encode) | dreamverse_review | **Blocked on PR 7.10** |
+| D-6 | Don't upstream `realtime/local_runtime.py` | dreamverse_review | Resolved (Dreamverse switches to `streaming.gpu_pool.SubprocessGpuPool`) |
+| D-7 / Q-6 | FP4Config public colocation | dreamverse_review | **Resolved May 2** — public NVFP4 landed with lazy flashinfer |
+| D-8 | `ltx2_image_crf` silently dropped | dreamverse_review | **Unverified post-`d80c2a8`** — see [open-threads.md](open-threads.md) |
+| D-9 | `aarch64-conda-linux-gnu-cc` triton compile failure | dreamverse_review | Operational; `ENABLE_TORCH_COMPILE=0` workaround |
+| D-10 | Warmup OOM on shared GPU | dreamverse_review | Operational; idle-GPU pre-warm probe |
+| D-11 | ffmpeg `Broken pipe` on disconnect | dreamverse_review | Cosmetic logging cleanup |
+| — | `video_position_offset_sec` semantics (persistent vs per-segment) | dreamverse_integration | **Open — needs decision before PR 7.6 emits state** |
+| — | `SessionStore` / `BlobStore` lifecycle (TTL/eviction/blob-drop) | dreamverse_integration | Open — defer to PR 7.5 design pass |
+
+See [decisions-log.md](decisions-log.md) for full rationale per
+decision.
+
+## Don't / Cautions
+
+- **Don't pop the Dreamverse stash on this branch.** It's 3867 lines of
+  orphan modular refactor with broken absolute imports.
+- **Don't change `RealtimeRuntimeConfig` shape without coordinating
+  with Dreamverse `runtime/factory.py`.**
+- **Don't promise public compatibility for private Dreamverse-only
+  field aliases.** Those belong in the private adapter layer per design
+  spec.

.agents/memory/dreamverse-integration/decisions-log.md

@@ -0,0 +1,710 @@
+# Decisions Log — D + Q Resolutions
+
+Cross-doc consolidated decision log. Each entry: ID, source doc,
+question/decision, rationale, current status.
+
+For implementation status see [pr-roadmap.md](pr-roadmap.md). For
+follow-up actions see [open-threads.md](open-threads.md).
+
+**Last updated:** 2026-05-05 (added D-12 — GpuPool layer separation, Oracle review post-#1257-merge; added D-13 — prompt enhancer / LLMProvider abstraction shape, Oracle review pre-#1258-merge; added D-14 — streaming auxiliaries cohesion, Oracle review during #1284 review cycle; added D-15 — streaming router placement + sticky/active-active deferral, Oracle review during #1286 review cycle; added D-16 — streaming router polish round 2, second-pass review on top of D-15 covering bridge cancellation hygiene, registry state machine, httpx hard-fail, replica YAML parsing, and `websockets` dep; added D-17 — strategy reversal: abandon 6-PR split in favor of single mega-PR #1288 on `will/ltx2_sr_port`; added D-18 — Option B+ chosen: Dreamverse FE+product-server move into FastVideo as `apps/dreamverse/` subfolder while generic backend stays at `fastvideo.entrypoints.streaming.*`; integration-review.md deprecated, integration-plan.md is the executable migration plan).
+
+## Status legend
+
+- ✅ **Resolved** — decision made and implementation complete (or no implementation needed)
+- 🟡 **Deferred** — decision made, implementation deferred to a known PR
+- 🔴 **Open** — needs decision
+
+## Post-merge architecture decisions
+
+### D-18: Option B+ — Dreamverse becomes `apps/dreamverse/` subfolder under FastVideo
+
+**Status:** ✅ Resolved 2026-05-05. [integration-plan.md](integration-plan.md) is the executable migration plan; [integration-review.md](integration-review.md) is deprecated but kept for drift audit + OSS precedents.
+**Source:** User decision after reviewing [integration-review.md](integration-review.md)'s Option D recommendation.
+
+**Question:** [integration-review.md](integration-review.md) recommended **Option D** — Dreamverse stays a separate repo, generic backend (streaming runtime, GPU pool, prompt enhancer, router) merges into `fastvideo.entrypoints.streaming.*`. The user reviewed this and chose a different shape: keep the generic-backend principle from Option D but ALSO move the Dreamverse FE + product server into FastVideo as a subfolder (`apps/dreamverse/`). Combination is "Option B+" (Option B layout with Option D's backend principle).
+
+**Decision:** Option B+. Concrete shape:
+
+- **One repo**: `hao-ai-lab/FastVideo`. Dreamverse repo gets archived after migration completes.
+- **Python ML library** stays at root: `fastvideo/`, `fastvideo-kernel/`.
+- **Generic backend** stays at `fastvideo.entrypoints.streaming.*` (already there per #1257/#1258/#1284/#1286/#1288).
+- **Dreamverse product** moves into `apps/dreamverse/{server,web,prompts,serve_configs,scripts}/`.
+- **Tooling**: uv workspace for Python (`[tool.uv.workspace] members = ["apps/dreamverse/server"]`), standalone pnpm for the FE (no root `package.json`), split CI workflows with path-filter triggers.
+
+**Rationale:**
+
+- Drops the cross-repo coordination overhead identified in the post-#1286 rebase cycle (D-17 handled by consolidating into mega-PR; D-18 prevents the next round of cross-repo coordination from happening).
+- Keeps the architectural separation Option D recommended (FastVideo owns reusable runtime; product owns product). The boundary is now `apps/dreamverse/` directory rather than two repos.
+- Single repo means atomic cross-cutting refactors (e.g. GpuPool API change + Dreamverse adoption) ship as one PR.
+- OSS precedents support the shape (chainlit uv-workspace + pnpm; open-webui Python + Svelte with paths-ignore CI). The librarian explicitly noted no precedent for "Python ML library + Next.js product merged into library namespace" — but this isn't that pattern. Dreamverse goes into a sibling directory, NOT into `fastvideo.entrypoints.dreamverse.*`. Library namespace stays clean.
+
+**Why not Option D (separate repos):**
+
+- Each upstream merge into FastVideo invalidates Dreamverse's lockfile/imports; the post-#1286 rebase showed this requires coordination overhead that scales with feature velocity.
+- Cross-repo contract tests catch shape drift but not behavior drift.
+- Two repos means two `AGENTS.md`, two CI configs, two release stories, two Dependabot dashboards.
+
+**Why not Option C (full merge into `fastvideo.entrypoints.dreamverse.*`):**
+
+- Forces FastVideo to ship Tailwind config + curated preset JSON + Next.js build artifacts.
+- Locks Dreamverse product cadence to FastVideo PyPI releases.
+- Librarian: "no 1:1 precedent for Python ML library + Next.js product merged into library namespace" — argues against this.
+
+**Why not Option B (subfolder, but generic backend folded into `apps/dreamverse/server/`):**
+
+- Other consumers (Dynamo, future streaming clients) need the backend without the Dreamverse product. Folding the backend under `apps/dreamverse/server/` would force Dynamo to either depend on `apps/` paths (ugly) or carry a fork.
+
+**Implications:**
+
+- [integration-review.md](integration-review.md) is **deprecated** (banner header + reading-guide demotion). Kept in tree for drift audit + OSS precedent reference.
+- [integration-plan.md](integration-plan.md) is the **canonical executable plan** with 7 phases (Phase 0: land #1288; Phase 1: skeleton + tooling; Phase 2: backend move; Phase 3: FE move; Phase 4: promote generic-pending; Phase 5: prompt enhancer fork retirement; Phase 6: CI/release cutover; Phase 7: archive Dreamverse repo).
+- Dreamverse repo will be **archived** at end of Phase 7 — not before.
+- Dreamverse history does NOT migrate cross-repo via `git mv` (technical limitation); original history stays in archived Dreamverse repo, and Phase 2 PR body records the source SHA(s).
+- New top-level `apps/` directory created — must be excluded from FastVideo PyPI wheel via `[tool.setuptools.packages.find] exclude = ["apps*", ...]`.
+- Drift items from [integration-review.md](integration-review.md) get folded into specific phases of [integration-plan.md](integration-plan.md) (e.g. health routes → Phase 4, DR-1 → Phase 5).
+
+**Open questions deferred to phase planning:**
+
+- DR-2 (`cerebras_ifm`): public Literal vs Dreamverse-side custom provider — decide before Phase 5.
+- VPO (`video_position_offset_sec` semantics): persistent vs per-segment — decide in Phase 4.
+- Cross-repo history: fresh import vs `git subtree` import — decide before Phase 2.
+- CORS / write-endpoint security policy: dev-only vs auth vs firewall — decide before Phase 6.
+
+### D-17: Abandon 6-PR split — land everything as single mega-PR #1288
+
+**Status:** ✅ Resolved 2026-05-05. PR #1287 closed; PR #1288 opened on `will/ltx2_sr_port` covering the full chain.
+**Source:** User decision after observing the post-#1286 rebase + re-slice cycle.
+
+**Question:** The original plan ([STACK.md](../../../STACK.md), [pr-roadmap.md](pr-roadmap.md)) called for the remaining `will/ltx2_sr_port` content (after PRs 7.5/7.6/7.7/7.8/7.9 landed) to ship as 6 stacked PRs: 7.10 (#1287, generate_async), 8 (server contract docs), LTX-2 SR runtime, NVFP4, post-fixes, agents-cleanup. PR #1287 was opened on 2026-05-05 as the first slice. Should the remaining 5 slices be opened sequentially as planned, or should everything be consolidated into one PR?
+
+**Decision:** Consolidate. Close #1287; open one mega-PR (#1288) on `will/ltx2_sr_port` covering all 34 commits / 71 files / +13,074 LOC at once.
+
+**Rationale:**
+
+- The post-#1286 rebase + re-slice cycle exposed real overhead: backup branch, interactive rebase with manual `drop` directives, force-push, re-slice 6 bookmarks, push next slice as new remote, open new PR, update memory dir. Repeating that 6 more times for the remaining slices accumulates substantial review-coordination overhead with diminishing structural benefit.
+- The 6 layers are not independent in the way that landed PRs 7.5-7.9 were. PR 7.10 (`generate_async`) is the only API-shape change; PR 8 is docs+tests on top; LTX-2 SR / NVFP4 / post-fixes / agents-cleanup are feature/fix/docs work that doesn't shape the public API. Reviewing them as one ordered diff is at least as easy as reviewing 6 stacked PRs whose dependencies must be tracked manually.
+- Single PR keeps CI / merge queue simpler and avoids the 6-PR cascade where every upstream merge invalidates the chain below it.
+
+**Implications:**
+
+- [STACK.md](../../../STACK.md) (top-level, 10-PR split tracker) is **deprecated**. Kept in tree as a historical artifact with the merged half (PRs 1-4 of the 10) accurate. Safe to delete in a follow-up.
+- [authors.md](authors.md), [`CO-AUTHORS.md`](../../../CO-AUTHORS.md) — co-author roster is unchanged; trailers still apply per-commit on every commit in the consolidated PR.
+- [runbook.md](runbook.md) — "After a PR merges (re-slice protocol)" section replaced by a simpler "After PR #1288 merges" section.
+- Local split bookmarks (`will/api_7.10`, `will/api_8`, `will/ltx2_sr_runtime`, `will/ltx2_nvfp4`, `will/ltx2_post_fixes`, `will/agents_cleanup`) are no longer maintained; safe to delete locally.
+- `origin/will/api_7.10` — pushed during the #1287 cycle; can be deleted on origin once #1287 close-cleanup completes.
+
+**Watch outs:**
+
+- The PR is large (71 files, +13,074 LOC). Reviewers will need commit-by-commit review; the PR body structures the layers in commit order to make this tractable.
+- If #1288 becomes too large to merge cleanly later (e.g. main moves significantly underneath it), the fallback is to re-split — but the current expectation is to land it as-is.
+
+### D-12: `GpuPool` layer separation — keep distinct from `VideoGenerator`
+
+**Status:** ✅ Resolved (interim) + 🟡 Deferred long-term shape to PR 7.10.
+**Source:** Oracle review on 2026-05-04, post-PR-#1257 merge.
+
+**Question:** Should `fastvideo.entrypoints.streaming.GpuPool` (PR #1257) be
+folded into `fastvideo.entrypoints.video_generator.VideoGenerator`, or kept
+separate? Three alternatives were evaluated:
+
+| Alt | Approach | Verdict |
+|---|---|---|
+| A | Status quo — `VideoGenerator` (single inference call) and `GpuPool` (multi-session orchestration) stay separate | ✅ Correct as **interim** |
+| B | `VideoGenerator` absorbs the pool's role (`from_pretrained_pool`, `acquire/release/run`) | ❌ **Wrong layer.** Conflates execution with serving scheduler. |
+| C | `GpuPool` becomes a thin **session-aware async executor** over PR 7.10's `generate_async` | ✅ Correct **long-term destination** |
+
+**Decision:** Alt A as interim; evolve toward Alt C once PR 7.10 lands
+`generate_async`. Do NOT pursue Alt B.
+
+**Rationale:**
+
+- `VideoGenerator` is a library handle — "execute one request, possibly
+  across ranks via `MultiprocExecutor`/`RayDistributedExecutor`."
+- `GpuPool` is serving infrastructure — "schedule N concurrent sessions
+  across N independent replicas, with sticky session-to-GPU affinity for
+  cache locality."
+- These are different layers driven by different consumers (a Python
+  script doing `gen.generate(req)` vs. a WebSocket server with sticky
+  sessions). Folding them muddies both surfaces.
+
+**Key finding — `MultiprocExecutor` and `SubprocessGpuPool` are orthogonal,
+not redundant:**
+
+| Layer | Job | Granularity |
+|---|---|---|
+| `MultiprocExecutor` (`fastvideo/worker/`) | TP/SP shard ONE inference call across N GPU ranks | per-call |
+| `streaming_generator.py` (existing real-time path) | Per-frame streaming via `MultiprocExecutor.submit_step`/`get_result` | per-step within one generator |
+| `SubprocessGpuPool` (`entrypoints/streaming/`, PR #1257) | Serve N concurrent sessions on N replicas, sticky-bound | per-session |
+
+Both spawn subprocesses because **CUDA contexts demand process boundaries**,
+not because they solve the same problem. Sharing low-level lifecycle
+utilities (process spawn, queue plumbing, shutdown) is a future refactor;
+unifying the abstractions is wrong.
+
+**Sticky binding stays in the pool, NOT in `VideoGenerator`:** sticky
+session-to-GPU affinity is a serving policy driven by LTX-2's per-GPU
+continuation cache (last-9-decoded-frames + audio-latents). Different
+consumers want different policies — stateless OpenAI HTTP wants
+per-request leasing; LTX-2 streaming wants sticky affinity; per-frame
+real-time streaming wants a continuous queue. Keeping policy in the pool
+keeps `VideoGenerator` policy-free.
+
+**Specific risks flagged in PR #1257 (already merged):**
+
+| Risk | Mitigation (when relevant) |
+|---|---|
+| `GpuPool.run() -> Any` is sync — fine for whole-segment dispatch, blocks on cancellation | Replace with `run_async() -> AsyncIterator[VideoEvent]` in PR 7.10 cycle (`generate_async` makes this trivial) |
+| `PoolAssignment.gpu_id: int` assumes one-GPU-per-worker | Don't lock as public API. Future may need `device_ids: list[int]` for topology-aware pooling (one worker = group of GPUs running internal `MultiprocExecutor`) |
+| `GpuPool` could be documented as the canonical FastVideo serving API | Mark as **experimental / server-internal** in docstring until PR 7.10 lands. Don't include in user-facing API docs yet |
+| Memory: N processes = N model replicas (~10-50 GB each) | Expected for concurrent serving with crash isolation. CUDA IPC weight sharing loses isolation; CPU-shared-memory loading helps host RAM not device. Real scalable path is topology-aware pooling later. |
+
+**Action items (carried into post-7.10 cycle):**
+
+- [ ] Update `GpuPool` ABC docstring to note "API may change post-PR-7.10"
+- [ ] Plan to replace `run()` with `run_async() -> AsyncIterator[VideoEvent]` in PR 7.10 cycle
+- [ ] Don't promote `gpu_id: int` to public API; revisit shape post-7.10
+- [ ] Consider clarifying field naming (e.g. `worker_id` is the stable identifier; `gpu_id` is current-impl detail)
+- [ ] When opening 7.10's PR, have it consume `generate_async` from `GpuPool.run_async` end-to-end
+
+**Open thread it touches:** PR 7.10 (`open-threads.md` item D — generate_async)
+unblocks Alt C and is the natural place to land the API shape change.
+
+### D-15: Streaming router (PR #1286) — keep in-repo, defer sticky / active-active
+
+**Status:** ✅ Resolved (interim). Pre-merge polishes applied. Three follow-up
+items tracked.
+**Source:** Oracle review on 2026-05-05, during PR #1286 review cycle.
+
+**Question:** Where should the multi-replica WebSocket router live? Should it
+ship at all (vs. delegating to nginx/envoy)? Should sticky session routing
+or weighted/round-robin balancing be in the initial PR?
+
+| Alt | Approach | Verdict |
+|---|---|---|
+| A | Status quo — `fastvideo/entrypoints/streaming/router/`, FastAPI-based, single-primary failover, lazy `httpx`/`websockets` imports | ✅ **Keep** |
+| B | Move to separate package `fastvideo-router/` | ❌ **Premature** — adds packaging/release/compat overhead before evidence of independent adoption |
+| C | Fold router into the streaming server itself (one app, mode flag) | ❌ Conflates router/generator lifecycles, mode-dependent config, drags inference deps into routing deployments |
+| D | Replace with reverse proxy (nginx/envoy/HAProxy) recipes | ❌ Not as the SOLE answer — mature proxies don't naturally emit FastVideo typed `gpu_unavailable` frames or evolve with FastVideo session semantics. Recommend external proxies as a complement at high scale. |
+| E | Add sticky session routing now | ❌ **Defer** — implementing correctly depends on where `session_id` is available (URL/header is easy, first JSON frame is invasive). Reconnects are rare today. |
+| F | Add weighted / round-robin now | ❌ **Defer** — active-active without sticky routing is worse for LTX-2 continuation locality than active-passive failover |
+
+**Decision:** Alt A — keep current shape. Apply pre-merge polishes; preserve
+forward-compat for sticky routing.
+
+**Rationale:**
+
+- Python router is justified as a FastVideo-aware control-plane component,
+  not a replacement for Envoy/HAProxy. It can emit typed
+  `gpu_unavailable` frames, evolve with FastVideo session semantics,
+  and ship local/dev deployment without ceremony.
+- The current abstraction is small + testable: `RouterConfig`,
+  `ReplicaRegistry`, `ReplicaStatus`, `HttpProbe` (Protocol/structural alias).
+  Adding strategy registries / telemetry interfaces / active-active policies
+  now would be over-engineering.
+- Active-passive (single primary) is the right MVP for LTX-2 streaming —
+  preserves continuation cache locality (D-12 sticky binding rationale)
+  better than naive active-active.
+- The biggest architectural risk isn't placement; it's accidentally baking
+  in unstated semantics. Define single-primary behavior + config validation
+  now so future active-active or sticky routing becomes additive.
+
+**Pre-merge polishes applied (per gemini + Oracle review):**
+
+| # | What | Why |
+|---|---|---|
+| 1 | `ReplicaRegistry.select()` docstring rewrite | gemini flagged "round-robin via insertion order" claim was misleading — implementation always returns `[0]`. Replaced with explicit "first healthy primary, else first healthy non-primary; this MVP picks first match within tier; round-robin/weighted deferred". |
+| 2 | Refactored `run_health_check_loop` to share single `httpx.AsyncClient` across the loop's lifetime via `_build_default_probe()` async context manager | gemini flagged per-probe client instantiation as inefficient. With ~1 probe/second default polling, TCP/TLS handshake overhead is non-trivial; now reuses connection. Tests inject probes directly so the path stays bypassable. |
+| 3 | Probe all replicas concurrently per cycle via `asyncio.gather(..., return_exceptions=True)` | gemini flagged sequential probes risk falling behind `health_check_interval_seconds` if replicas time out. Now per-cycle wall time = max(probe latencies), not sum. |
+| 4 | `RouterConfig.__post_init__` validation | Oracle recommended: empty replicas, non-positive intervals/timeouts, thresholds < 1, non-`http(s)://` URLs, and >1 primary all `raise ValueError`. Surfaces misconfiguration at config-load instead of confusing runtime failures. |
+| 5 | Migrated `@app.on_event("startup"/"shutdown")` to `@contextlib.asynccontextmanager`-based `_lifespan()` | Pre-merge — FastAPI deprecated the old API. Was tracked as the 7.9 caveat in pr-roadmap.md. |
+
+**One review comment intentionally not implemented:**
+
+| Comment | Decision |
+|---|---|
+| gemini medium: `_load_router_config` duplicates `fastvideo.api.parser.parse_config` logic | Kept manual flat-from-nested mapping. The YAML schema has nested `health_check:` block but `RouterConfig` is flat; using `parse_config` directly would require either restructuring `RouterConfig` to have a nested `HealthCheckConfig` (schema change beyond this PR's scope) or accepting incomplete parsing. Manual mapping is intentional and well-typed. |
+
+All 4 review threads marked resolved on the GitHub PR.
+
+**Action items (deferred):**
+
+- [ ] Track sticky session routing extensibility — when needed, add
+  `ReplicaRegistry.select(routing_key: str | None = None)` so registry
+  evolution is additive; document upfront where `session_id` should
+  appear (URL/header preferred over first JSON frame to avoid
+  buffering/peeking)
+- [ ] Track `_bridge_session()` backpressure note — fine for MVP because
+  `websockets` library provides basic transport backpressure, but at
+  high scale add max_size/timeouts or recommend Envoy/HAProxy in front
+- [ ] If active-active multi-primary becomes a requirement, define
+  behavior (round-robin within healthy primaries, weighted, sticky-by-key)
+  rather than letting `select()` silently pick `[0]`
+
+**Watch outs:**
+
+- `session_id` in WebSocket URL/headers is the cleanest sticky-routing
+  hook. If it ends up only in the first JSON message, sticky routing
+  later will require buffering/peeking before backend selection.
+- Multi-primary configs are now explicitly rejected by validation;
+  documented + enforced.
+- `_bridge_session()` is fine for MVP (the libraries provide basic
+  backpressure), but not production-grade for edge load. Document the
+  limit.
+
+**Open thread it touches:** open-threads.md items #13 (sticky routing),
+#14 (bridge backpressure), #15 (multi-primary semantics).
+
+### D-16: Streaming router polish round 2 — second-pass fixes on top of D-15
+
+**Status:** ✅ Resolved. Applied as `[fix] streaming: router polish — bridge
+cancel + state machine + deps` (`a152cb77` on `will/api_7.9`, `40e265b8` on
+`will/ltx2_sr_port`).
+**Source:** Second-pass review on PR #1286, 2026-05-05, after D-15's pre-merge
+polishes landed.
+
+**Question:** D-15 closed the structural review (placement, sticky/active-active
+deferral, basic `__post_init__` validation). On a second pass through the same
+files, five latent issues surfaced that weren't covered by gemini's first pass
+or Oracle's structural review. Apply them on top of the merged D-15 polishes,
+or queue for a follow-up PR?
+
+**Decision:** Apply on top of `will/api_7.9` directly. All five are bug-class
+or DX-class — none are scope-expanding architecture changes — so folding them
+into PR #1286 keeps the router landing in one reviewable unit instead of
+shipping a router PR plus an immediate follow-up fix PR.
+
+**Fixes applied:**
+
+| # | File | What | Why |
+|---|---|---|---|
+| 1 | `router/main.py::_bridge_session` | Replaced `asyncio.gather()` with `wait(FIRST_COMPLETED)` + explicit `cancel()`/drain + `_is_normal_disconnect()` classifier | `gather` waited for both directions; on client disconnect, the backend-reader task leaked and stayed pending. Backend `ConnectionClosed` also surfaced as an unhandled exception in server logs. New shape: first task to finish triggers explicit cancel of the other, both are drained, and only non-routine exceptions re-raise. |
+| 2 | `router/registry.py::record_success` | Split state transitions: `UNKNOWN -> HEALTHY` is now immediate on first successful probe; only `UNHEALTHY -> HEALTHY` remains gated by `recovery_threshold` | Previously a fresh registry needed `recovery_threshold` consecutive successes before any replica was selectable. With default `recovery_threshold=2` and `health_check_interval=1s`, that meant 2-3s of `gpu_unavailable` rejections at startup. Now the first probe promotes immediately; recovery gating still protects against flapping replicas. |
+| 3 | `router/registry.py::_build_default_probe` | Missing `httpx` now raises `RuntimeError` with install hint instead of yielding a "disabled" probe stub | Previous behavior: silently returned `(0.0, "httpx not installed; ...")` for every probe, which `record_failure` then folded into `UNHEALTHY` after `failure_threshold` cycles. Operators saw replicas drop UNHEALTHY with a confusing reason and no clear remediation. Hard-fail at startup is the right surface. |
+| 4 | `router/config.py::__post_init__` | Extended D-15 polish #4 with: rejects `urlparse(url).path not in ("", "/")`, rejects `query`/`fragment`, rejects duplicate URLs across replicas | D-15's validation rejected non-`http(s)://` URLs and >1 primary; it didn't catch `http://host/api` (the router appends `/health` and `/v1/stream` itself, so a base-URL with path yields malformed routes) or `[{url: x}, {url: x}]` (replica registry keys by URL — duplicates would silently collapse to one entry, masking the misconfiguration). |
+| 5 | `cli/router_serve.py::_load_router_config` | Replaced silent list-comprehension filter (`for r in replicas_raw if isinstance(r, dict) and r.get("url")`) with per-index `raise ValueError` | Original parser silently dropped malformed YAML entries. A single typo in `replicas[2].url` would yield 2 replicas instead of 3 with no log line. New shape: explicit per-index error message ("missing required key 'url'", "must be a mapping"). |
+| 6 | `pyproject.toml::[streaming]` extra | Added `websockets` as explicit dep | `router/main.py::_bridge_session` does `import websockets` lazily and raises `RuntimeError` if missing. The `[streaming]` extra was an implicit transitive — anyone installing only `[streaming]` (and not the broader requirements) hit the runtime error. Now explicit. |
+
+**Tests added (7 cases in `fastvideo/tests/entrypoints/streaming/test_router.py`):**
+
+- `TestUnknownToHealthyImmediate.test_first_success_promotes_unknown` — first probe success transitions `UNKNOWN -> HEALTHY` regardless of `recovery_threshold`
+- `TestUnknownToHealthyImmediate.test_unhealthy_recovery_still_gated_by_threshold` — `UNHEALTHY -> HEALTHY` still requires `recovery_threshold` successes
+- `TestConfigValidation.test_rejects_path_in_url` / `test_rejects_query_in_url` / `test_rejects_fragment_in_url` / `test_rejects_duplicate_urls` / `test_accepts_trailing_slash` — `__post_init__` URL validation matrix
+
+**Verification:** 17/17 router tests pass on both branches. `pre-commit run`
+clean (yapf / ruff / codespell / mypy). `lsp_diagnostics` clean on changed
+regions; the one pre-existing `Task` generic-type warning at `main.py:37` is
+unrelated and predates this commit.
+
+**In-flight pre-commit corrections (not part of the 6 fixes themselves):**
+
+- yapf auto-reformatted 4 files (kept verbatim).
+- ruff `UP038`: rewrote `isinstance(exc, (CancelledError, WebSocketDisconnect))`
+  to `isinstance(exc, CancelledError | WebSocketDisconnect)`.
+- mypy `[misc]`: renamed loop var `exc` (inside `for task in done`) to
+  `task_exc` to avoid name collision with the outer
+  `except ImportError as exc` binding.
+
+**Open thread it touches:** None new. Item #14 (bridge backpressure) and
+item #13 (sticky routing) from D-15 remain deferred — this round addressed
+**cancellation/disconnect** semantics on the bridge, which is distinct from
+**throughput backpressure**. Item #14 still applies: at higher load, add
+`_bridge_session()` max-size + timeout limits or recommend Envoy/HAProxy
+in front.
+
+### D-14: Streaming auxiliaries (PR #1284) — cohesion + concrete-vs-Protocol scoping
+
+**Status:** ✅ Resolved (interim). Two polish items applied during review; one
+operational caveat tracked.
+**Source:** Oracle review on 2026-05-04, during PR #1284 review cycle.
+
+**Question:** Is PR #1284's bundle of 4 streaming-server auxiliary modules
+(`prompt/safety.py`, `prompt/rewrite.py`, `session_logger.py`,
+`mock_server.py`) correctly scoped? Should `mock_server` live in production
+module path? Should `PromptSafetyFilter` be a Protocol? Should the bundle
+have been split into 4 PRs?
+
+| Alt | Approach | Verdict |
+|---|---|---|
+| A | Status quo — single PR, 4 modules under `streaming/`, mock_server in production path, concrete safety filter | ✅ **Keep** |
+| B | Split into 4 separate PRs | ❌ Process overhead, not architectural improvement |
+| C | Move `mock_server.py` into `tests/` | ❌ Would reduce discoverability + install-time usability of `python -m fastvideo.entrypoints.streaming.mock_server` |
+| D | Move `session_logger.py` to `streaming/observability/` (or top-level `fastvideo/observability/`) | ❌ Premature — currently session-shaped + streaming-specific; promote when a non-streaming consumer appears |
+| E | Convert `PromptSafetyFilter` to Protocol (like `LLMProvider`) | ❌ Premature abstraction — only one classifier exists; small duck-typed surface preserves future Protocol introduction without breaking the concrete |
+| F | Convert `MockGenerator` to Protocol | ❌ Same — small duck-typed surface; no second mock generator exists |
+
+**Decision:** Alt A — keep current shape. Apply two polish items from
+Oracle's review before merge.
+
+**Rationale:**
+
+- "Streaming-server auxiliaries" is cohesive enough at 730 LOC with
+  isolated modules + tests. Each module has independent code path but
+  shared deployment context (the streaming server boots them all).
+- `mock_server.py` in production path is a strength: reuses
+  `build_app()` for protocol parity. Hiding it under `tests/` would lose
+  `python -m fastvideo.entrypoints.streaming.mock_server` CLI access for
+  FE devs.
+- Concrete `PromptSafetyFilter` matches "ship what we have, abstract
+  later" pattern. Internal had multi-classifier composition; public
+  ships single + leaves chaining as a Dreamverse-side concern (per D-2).
+- Same pattern for `MockGenerator`: small duck-typed `_GeneratorLike`
+  surface lets a second mock implementation drop in without inheritance.
+- `threading.Lock` (not `asyncio.Lock`) in `session_logger.py` is
+  correct — writes come from real encoder/control threads via
+  `run_in_executor`, not from coroutines directly. `asyncio.Lock` would
+  be the wrong primitive for cross-thread concurrency.
+
+**Pre-merge polishes applied (per Oracle):**
+
+| Polish | What | Why |
+|---|---|---|
+| 1 | Removed `RewriteOptions.user_system_prompt_override` | Inert public field — was declared but never threaded through to `enhancer.rewrite()`. Shipping unused public options is more likely to bite than any structural choice. Re-add when actually wired through. |
+| 2 | Sanitized `session_id` filename in `session_logger.SessionLogger._get_file()` | Defense-in-depth: today session_id is server-generated UUID, but a future code path that accepts client-supplied ids would otherwise allow path traversal via `../`. Added `_FILENAME_SANITIZE_RE = re.compile(r"[^A-Za-z0-9._-]")` + sub before `os.path.join`. |
+
+**Operational caveat tracked (not a code change):**
+
+- `SafetyDecision.UNAVAILABLE` is treated as `ALLOW` by callers — a
+  policy choice that's correct for an opt-in safety filter, but
+  callers should log loudly so operators know the filter is degraded.
+  Tracked as open-threads.md item #12.
+
+**Pre-merge review feedback (4 of 4 resolved on the GitHub PR):**
+
+| # | File:Line | Severity | Issue | Fix applied |
+|---|---|---|---|---|
+| 1 | `session_logger.py:57` | High | `log()` race vs `close()` — `KeyError` on `_locks[session_id]` | Atomic capture in `_get_file()`; master `_registry_lock`; `with lock, contextlib.suppress(ValueError):` |
+| 2 | `rewrite.py:71` | Medium | `re.compile()` in hot path | Module-level `_LEADING_MARKER_RE`, top-level `import re` |
+| 3 | `safety.py:105` | Medium | `_ensure_loaded()` race on concurrent fastText load | `_load_lock = threading.Lock()` + double-check pattern |
+| 4 | `pyproject.toml:145` | Medium | `streaming` extra missing `prompt-safety` | Added to aggregator |
+
+All 4 review threads marked resolved via GraphQL `resolveReviewThread`.
+
+**Action items (deferred):**
+
+- [ ] Track `SafetyDecision.UNAVAILABLE` log loudness in
+  open-threads.md item #12 — when streaming server starts using the
+  safety filter, ensure operator-visible logging on `UNAVAILABLE`
+  results
+- [ ] If a second safety classifier appears (Perspective API, Detoxify,
+  custom rules), promote `PromptSafetyFilter` to a Protocol — same
+  pattern as `LLMProvider` per D-13
+- [ ] If a second mock generator appears (different frame patterns,
+  different latency models), promote `MockGenerator` to a Protocol
+
+**Open thread it touches:** PR #1284 itself; future safety-classifier
+Protocol promotion; future observability module extraction.
+
+### D-13: Prompt enhancer / `LLMProvider` abstraction shape — keep streaming-scoped
+
+**Status:** ✅ Resolved (interim) + 🟡 Three deferred polishes after metrics or 2nd consumer.
+**Source:** Oracle review on 2026-05-04, pre-PR-#1258-merge.
+
+**Question:** Is PR #1258's `fastvideo.entrypoints.streaming.prompt.*` module
+correctly designed? Should it be (a) Protocol-based vs ABC, (b) under
+`streaming/` vs top-level `fastvideo.prompt.*`, (c) closed 3-op enum vs
+open `complete()` API?
+
+| Alt | Approach | Verdict |
+|---|---|---|
+| A | Status quo — `streaming/prompt/*`, Protocol provider, fixed 3 ops, lazy `httpx`, per-call `AsyncClient` | ✅ **Keep** |
+| B | Move to top-level `fastvideo.prompt.*` (decouple from streaming) | ❌ **Premature.** No second consumer exists yet. |
+| C | Convert `LLMProvider` Protocol → ABC with default impls + retry classification | ❌ **Wrong direction.** Biases extension toward OpenAI shape; `_openai_compat.py` already factors that as helper not inheritance. |
+
+**Decision:** Alt A as interim. Promote to Alt B only when a second
+non-streaming consumer (OpenAI server, batch generation, tooling) actually
+needs the prompt enhancer. Don't pursue Alt C.
+
+**Rationale:**
+
+- Public contract is tiny — `name: str` + `async complete(LLMRequest) -> LLMResponse`. ABC adds zero value.
+- `_openai_compat.py` is the right place for shared logic — helper, not base class. Anthropic / local / custom providers stay first-class.
+- The 3 ops (enhance / auto_extend / rewrite) are LTX-2 streaming concepts. `auto_extend` (continue prompt sequence) and `rewrite` (multi-line alternatives) come directly from session UX. Calling this "the FastVideo prompt API" misrepresents that.
+
+**Specific risks flagged in PR #1258 (already merged-pending review):**
+
+| Risk | Mitigation (when relevant) |
+|---|---|
+| API publicity — calling this "the FastVideo prompt API" before a second consumer exists | Document module as "streaming-server prompt enhancement" in user-facing docs; keep it nested under `entrypoints/streaming/` |
+| `httpx.AsyncClient` per-call (no connection pooling) | Acceptable for ~6-10 calls per LTX-2 session; LLM latency dominates. Add optional `client_factory` parameter LATER if metrics show connect overhead is meaningful. |
+| 3 fixed operations could constrain future generic use | Closed enum is right for application-level orchestration. Future generic consumers should either call `provider.complete()` directly, or get a thin separate enhancer that shares the provider/fallback machinery. |
+| `register_provider(priority=-1)` semantics rely on Python's negative-index `list.insert` | Cosmetic concern; docstring is clear. Could be tightened to explicit branch later. |
+| `runtime_checkable` Protocol with `name: str` instance attribute — static type checkers may miss missing `name` | Acceptable; runtime check via `isinstance(p, LLMProvider)` works for plugin discovery. |
+
+**Action items (deferred):**
+
+- [ ] Document `fastvideo.entrypoints.streaming.prompt.*` as streaming-scoped in user-facing docs (PR 12 docs migration); avoid promoting as framework-level
+- [ ] Add optional `client_factory` parameter to providers when metrics justify pooling
+- [ ] Plan future move to `fastvideo.prompt.*` (with import shim) when second non-streaming consumer materializes
+- [ ] Track Q-2 reactivation: promote LTX-2 prompt orchestration (locked segments, segment-prompts JSON parsing) to public `fastvideo.entrypoints.streaming.prompt.ltx2_orchestration` when a second LTX-2-style consumer appears
+
+**Open thread it touches:** Dreamverse migration (open-threads.md DR-1)
+will be the first real test of the public surface. Lessons learned there
+inform whether Alt B becomes feasible.
+
+## D-decisions (from `dreamverse_review.md`, Apr 26)
+
+### D-1: Realtime runtime → streaming GpuPool migration shape
+
+**Status:** ✅ Resolved.
+
+Internal `RealtimeRuntimeConfig` had a multi-model registry +
+flattened sampling defaults. Public `SubprocessGpuPool` is single-model
++ uses per-request `SamplingConfig`.
+
+**Decision:** Drop multi-model registry on integration branch (not used
+in production). Construct `GeneratorConfig` for chosen model and pass to
+`SubprocessGpuPool`. Move sampling defaults to a server-side
+`default_request: GenerationRequest` template.
+
+**Risk:** Migration branch surfaces missing-model errors if a flow
+silently relied on registry to swap models per-session. Integration
+tests exercise at least one segment per supported model id before
+merging.
+
+### D-2: PR 7.7 prompt enhancer API surface narrower than internal
+
+**Status:** ✅ Resolved.
+
+Public `PromptEnhancer.enhance/auto_extend/rewrite` returns
+`LLMResponse(content, provider, model, latency_ms, fallback_used)`.
+Internal returns `EnhanceResult(prompt, fallback_used, error, ...)` /
+`RewriteResult(prompts, ..., rollout_id, rollout_label, ...)`.
+
+**Decision:** Adapt at the call site via
+`Dreamverse/server/prompting/_internal_compat.py` shim. Locked-segment /
+next-segment-index plumbing stays Dreamverse-side. Public stays minimal
+and provider-agnostic.
+
+**Open question (Q-2):** Promote LTX-2-specific orchestration into
+`fastvideo.entrypoints.streaming.prompt.ltx2_orchestration` once a
+second consumer appears. Logged for future review.
+
+### D-3: Multi-stage provider race vs. sequential fallback
+
+**Status:** ✅ Resolved (public stays sequential).
+
+Internal enhancer runs all providers in a stage in parallel
+(`_run_provider_race`). Public enhancer runs sequentially with
+retryable-error fallback.
+
+**Decision:** Public stays sequential for PR 7.7. Race is a
+Dreamverse-specific tail-latency optimization that depends on parallel
+API budgets.
+
+**Risk / Q-3:** First-segment latency on Dreamverse may regress
+slightly when Cerebras has a bad minute (sequential waits 20s before
+trying Groq). If real production concern, add public
+`concurrency: int = 1` knob behind a race path — but only after measuring.
+
+### D-4: Skip PR 7.9 router for the integration branch
+
+**Status:** ✅ Resolved.
+
+Internal stack ships `router/main.py` for multi-replica load balancing.
+Dreamverse deployment uses single replica per region.
+
+**Decision:** Land PR 7.9 publicly (upstream the surface). Skip wiring
+into Dreamverse integration branch. Dreamverse's `server/main.py` does
+not import from `router/`.
+
+### D-5: Audio re-encode (PR 7.10) needed for streaming, deferred
+
+**Status:** 🟡 Deferred to PR 7.10.
+
+Internal streaming server's per-step path runs `_re_encode_audio` inside
+`_stream_av_fmp4_events` so each fMP4 segment ships with
+continuation-conditioning audio. Whole-segment `pool.run()` path doesn't
+need this.
+
+**Decision:** Land PR 7.10's `generate_async` publicly. Dreamverse
+integration branch initially keeps using `pool.run()` (whole segment, no
+re-encode). Follow-up branch swaps to `generate_async` + audio re-encode.
+
+**Open question (Q-5):** Acceptable for first switch, or does
+Dreamverse audio quality regress vs. internal until 7.10 wires in?
+
+### D-6: `realtime/local_runtime.py` is NOT upstreamed
+
+**Status:** ✅ Resolved.
+
+It was the FastVideo-internal precursor to `streaming.gpu_pool`.
+Upstreaming both would create two GPU pool implementations in public.
+
+**Decision:** Don't upstream `realtime/local_runtime.py`. Dreamverse
+switches to `streaming.gpu_pool.SubprocessGpuPool` on integration
+branch. Internal module can be deleted at follow-up.
+
+### D-7 / Q-6: `FP4Config` is private-only
+
+**Status:** ✅ **Resolved May 2.**
+
+April 26: `Dreamverse/server/video_generation.py:271` imported
+`fastvideo.layers.quantization.fp4_config.FP4Config` from
+FastVideo-internal only. The 411-line module hard-imported `flashinfer`.
+
+**Two options at the time:**
+
+1. Colocate publicly with `flashinfer` as optional extra
+   `pip install fastvideo[fp4]`; refactor `FP4QuantizeMethod` to take
+   layer-prefix list from a pipeline-config field instead of hardcoding
+   ltx2 paths.
+2. Keep private — Dreamverse imports from internal via thin shim.
+
+**Recommendation at the time:** option 1 once API refactor settles.
+
+**Resolution:** May 2 work chose option 1.
+- `365a66c7` upstreamed FP4Config with lazy `flashinfer` import in
+  loader helper (no public hard-dep)
+- `94c983a2` renamed FP4 → NVFP4 to disambiguate from MX-FP4 / OCP-FP4
+- `42b30bf9` wired through `fastvideo.layers.quantization`
+
+See [quantization.md](quantization.md) for full details.
+
+### D-8: `ltx2_image_crf` silently dropped by public schema
+
+**Status:** 🔴 **Unverified post-`d80c2a8`.**
+
+April 26: Dreamverse's `server/video_generation.py:406` passed
+`ltx2_image_crf=0.0` to `SamplingParam(...)`. Public
+`fastvideo.api.sampling_param.SamplingParam` did NOT have this field;
+the BE logged ERROR and silently dropped the kwarg.
+
+**Migration target** (per [design.md](design.md) compatibility map):
+`request.stage_overrides.refine.image_crf`.
+
+**Resolution status:** `d80c2a8` (May 2) refactored
+`server/video_generation.py` to use typed `GeneratorConfig` +
+`preset_overrides["refine"]`. Whether this PR routed `image_crf`
+through the typed `stage_overrides` path or left it silently dropped is
+unverified. See [open-threads.md](open-threads.md).
+
+### D-9: `aarch64-conda-linux-gnu-cc` triton compile failure
+
+**Status:** ✅ Resolved (operational).
+
+Conda env injected an ARM cross-compiler ahead of `gcc` on `$PATH`, so
+`torch._inductor`'s triton launcher failed compilation. Setting
+`ENABLE_TORCH_COMPILE=0` bypasses it.
+
+**Long-term fix:** clean conda env's compiler shadowing or add
+`CC=gcc` override in Dreamverse's worker bootstrap.
+
+### D-10: Warmup OOM on shared GPU
+
+**Status:** ✅ Resolved (operational).
+
+When `CUDA_VISIBLE_DEVICES` lands on a GPU another tenant uses, LTX-2
+warmup fails with OOM. Picking an idle GPU (4-7 in test setup) is a
+manual step.
+
+**Improvement:** pre-warm probe that checks free memory before booting
+the pool would prevent this.
+
+### D-11: ffmpeg fragment write `Broken pipe`
+
+**Status:** ✅ Resolved (cosmetic).
+
+When WS client closes before backend finishes streaming first segment,
+ffmpeg hits `[Errno 32] Broken pipe`. Currently propagates to
+"User step failed". Cosmetic — swallowing pipe-broken on intentional
+disconnect would clean up logs.
+
+## Q-questions (from `streaming-server-upstream-plan.md`, Apr 17)
+
+### Q-1: Router placement (in-repo or separate package)
+
+**Status:** ✅ Resolved (in-tree).
+
+**Recommendation at the time:** separate package `fastvideo-router/` or
+`fastvideo/contrib/router/`; defer final call to PR 7.9.
+
+**Resolution:** PR 7.9 implementation places router in-tree at
+`fastvideo/entrypoints/streaming/router/`.
+
+### Q-2: Session ID authority
+
+**Status:** ✅ Resolved (server-generated).
+
+**Recommendation:** server-generated UUID; accept externally provided
+session ID only for resume flows.
+
+### Q-3: Torch compile kwargs typing (opaque vs full vs hybrid)
+
+**Status:** ✅ Resolved (hybrid).
+
+**Recommendation:** hybrid — type the common four (`backend`,
+`fullgraph`, `mode`, `dynamic`) + allow `extras: dict[str, Any]`.
+
+**Resolution:** PR 6 + NVFP4 `221cb20a` shipped exactly this hybrid.
+
+### Q-4: Prompt safety / fasttext dependency
+
+**Status:** ✅ Resolved (optional extra).
+
+**Recommendation:** ship as optional extra `pip install fastvideo[prompt-safety]`.
+
+**Resolution:** PR 7.8 implements as optional extra.
+
+### Q-5: Audio-specific tensor payloads in continuation
+
+**Status:** ✅ Resolved (typed `LTX2ContinuationState`).
+
+`ltx2_audio_clean_latent`, `ltx2_audio_denoise_mask`,
+`ltx2_audio_latents` not in pre-refactor public schema.
+
+**Recommendation:** classify as opaque fields inside
+`LTX2ContinuationState.payload`, not top-level sampling fields.
+
+**Resolution:** PR 7's typed `LTX2ContinuationState` lifts these into
+typed fields (see [cross-repo-surfaces.md](cross-repo-surfaces.md)
+field mapping table).
+
+### Q-6: Dynamo subpackage home
+
+**Status:** ✅ Resolved (lives in Dynamo repo).
+
+**Resolution:** No Dynamo code in FastVideo. Full backend package
+(handler, adapter, registration, health check) owned by Dynamo repo at
+`components/src/dynamo/fastvideo/`, same pattern as vllm/sglang.
+FastVideo only guarantees the public API contract.
+
+### Q-7 (was Q-6 in dreamverse_review): How to land FP4Config publicly
+
+**Status:** ✅ Resolved May 2 — option 1 (colocate publicly).
+
+See D-7 above.
+
+### Q-8: Disaggregation readiness contract test
+
+**Status:** 🟡 Recommended; not yet shipped.
+
+PR ai-dynamo/dynamo#7544 is aggregated-only. `ContinuationState` hybrid
+already supports future prefill/decode split.
+
+**Recommendation:** PR 7.10 explicitly validate `ContinuationState`
+survives round-trip through Dynamo-style RPC (pickle or JSON), even
+though Dynamo isn't using it today. Cheap regression guard.
+
+### Q-9: Dynamo progress/status passthrough
+
+**Status:** 🟡 Deferred until Dynamo clarifies.
+
+`NvVideosResponse` has `status` and `progress` fields.
+
+**Recommendation:** PR 7.10 stays aggregated-final-only to match PR
+#7544 shape; revisit after Dynamo clarifies their streaming/progress
+semantics.
+
+## Cross-doc questions still 🔴 OPEN
+
+These need decisions; tracked also in [open-threads.md](open-threads.md):
+
+| ID | Question | Source | Why it matters |
+|---|---|---|---|
+| **D-8** | Did `d80c2a8` route `ltx2_image_crf` correctly, or is it still silently dropped? | dreamverse_review | Latent silent-drop bug; FP4-disabled paths may degrade |
+| **VPO** | `video_position_offset_sec` — persistent accumulation (a) vs per-segment hint (b) | dreamverse_integration | Needs decision before PR 7.6 emits state |
+| **SBS** | `SessionStore` / `BlobStore` lifecycle (TTL/eviction/blob-drop on state replacement) | dreamverse_integration | Needs decision in PR 7.5 design pass |
+| **#1** | Migrate `/healthz`+`/readyz`+`/status` into FastVideo `build_app` | streaming-upstream-plan + handoff | Closes BE_FLAVOR=fastvideo FE-compatibility |
+| **#3** | Add `cerebras_ifm` to public `PromptEnhancerConfig.provider` Literal | handoff | Internal supports it; public schema doesn't |
+| **#4** | Expose `layer_profile` on typed `engine.quantization` | handoff | Removes Dreamverse's `experimental["pipeline_config"]` dodge |
+| **#5** | Typed `dit_config.quant_config` carrier (design TBD) | handoff | Eliminates the `experimental["pipeline_config"]` escape hatch entirely |

.agents/memory/dreamverse-integration/design.md

@@ -0,0 +1,332 @@
+# Design — Typed Public Inference API
+
+Synthesis of the FastVideo public inference API refactor design philosophy.
+For PR-by-PR execution see [pr-roadmap.md](pr-roadmap.md). For the streaming
+extension see [streaming-server.md](streaming-server.md).
+
+**Last updated:** 2026-05-03.
+
+## Why the refactor
+
+The pre-refactor public boundary mixed three concerns through `**kwargs`:
+
+- `VideoGenerator.from_pretrained(..., **kwargs)` mixed engine/runtime,
+  pipeline init, and component overrides.
+- `VideoGenerator.generate_video(..., **kwargs)` mixed prompt+inputs,
+  sampling, output, and model-specific workflow knobs.
+- Unknown keys silently filtered or merely logged → API drift hard to detect.
+- Multi-stage models (LTX-2 two-stage, Hunyuan15 SR, LongCat distill+refine)
+  exposed via ad hoc top-level flags.
+
+This was already painful for LTX2/Dreamverse and would worsen as more
+multi-stage pipelines came in.
+
+## Core decision
+
+FastVideo has:
+
+1. **Typed nested public schema** — `RunConfig`, `ServeConfig`,
+   `GeneratorConfig`, `GenerationRequest`, `ContinuationState`.
+2. **Model-owned named pipeline presets** — `ltx2_two_stage`,
+   `longcat_distill_refine`, `hunyuan15_sr_1080p`, etc. All 13 model families
+   landed presets in PR 4.
+3. **Semantic stage overrides by stage name** —
+   `request.stage_overrides["refine"] = LTX2RefineStageOverride(...)`.
+4. **Optional advanced explicit plans** for power users — `GenerationPlan`
+   (escape hatch only; not the canonical surface).
+5. **YAML-first CLI** with dotted overrides —
+   `fastvideo generate --config run.yaml --request.sampling.seed 42`.
+
+The canonical user experience: choose a model → choose a preset → override
+a few typed fields → generate. Dicts/YAML/JSON are supported as
+serialization, but parse immediately into typed objects with strict
+unknown-key validation.
+
+## Schema surface
+
+Implemented in [`fastvideo/api/`](file:///home/william5lin/FastVideo/fastvideo/api/):
+
+| Type | Role |
+|---|---|
+| `RunConfig` | Offline envelope: `generator` + `request` |
+| `ServeConfig` | Serving envelope: `generator` + `server` + `default_request` + optional `streaming` |
+| `GeneratorConfig` | `model_path`, `revision`, `trust_remote_code`, `engine`, `pipeline` |
+| `EngineConfig` | parallelism / offload / compile / quantization / flags |
+| `PipelineSelection` | `workload_type`, `preset`, `preset_version`, `components`, `preset_overrides`, `experimental` |
+| `GenerationRequest` | `prompt`, `negative_prompt`, `inputs`, `sampling`, `runtime`, `output`, `stage_overrides`, `state`, `plan`, `extensions` |
+| `ContinuationState` | Opaque envelope `{kind: str, payload: dict[str, Any]}` |
+| `GenerationPlan` | Advanced/escape-hatch only; `{stages: list[PlannedStage], final_stage: str|None}` |
+
+Files:
+
+| File | Role |
+|---|---|
+| [`schema.py`](file:///home/william5lin/FastVideo/fastvideo/api/schema.py) | All public dataclasses |
+| [`parser.py`](file:///home/william5lin/FastVideo/fastvideo/api/parser.py) | `from_dict`, `to_dict`, `load_yaml`, `load_json`, validation |
+| [`overrides.py`](file:///home/william5lin/FastVideo/fastvideo/api/overrides.py) | Dotted override application |
+| [`compat.py`](file:///home/william5lin/FastVideo/fastvideo/api/compat.py) | Legacy kwargs translation (~370 lines, scheduled for death PRs 14-17) |
+| [`presets.py`](file:///home/william5lin/FastVideo/fastvideo/api/presets.py) | Preset registry |
+| [`sampling_param.py`](file:///home/william5lin/FastVideo/fastvideo/api/sampling_param.py) | Internal `SamplingParam` adapter (canonical home since PR 4) |
+| [`results.py`](file:///home/william5lin/FastVideo/fastvideo/api/results.py) | `GenerationResult` / `VideoResult` |
+| [`errors.py`](file:///home/william5lin/FastVideo/fastvideo/api/errors.py) | Path-aware validation errors |
+
+## Boundary normalization rule
+
+Every public inference entrypoint normalizes into typed config objects
+before touching legacy internals (`FastVideoArgs`, `SamplingParam`).
+Includes Python constructors, `generate*` calls, CLI `generate`, CLI
+`serve`, OpenAI server request translation, streaming server request
+translation.
+
+Legacy internals (`FastVideoArgs`, `SamplingParam`) may remain temporarily,
+but only behind a typed normalization boundary.
+
+## Strict-by-default validation
+
+All structured inputs are strict:
+
+- Unknown keys → error
+- Wrong types → error
+- Invalid stage names → error
+- Incompatible state/preset combinations → error
+
+The only intentional escape hatches:
+
+- `generator.pipeline.experimental` — for in-flight features without typed home
+- `request.extensions` — same, request-side
+
+These bypass validation by design. Intent: shrink as presets absorb
+model-specific fields. New fields should not land in `experimental` /
+`extensions` without a plan to either promote them to typed fields or
+remove them within two PR cycles.
+
+Error format includes nested path:
+
+```
+Invalid field: request.stage_overrides.refine.num_inference_steps
+Expected int, got "two"
+Preset: ltx2_two_stage
+Stage: refine
+```
+
+## Request mutation tracking
+
+When a `GenerationRequest` is parsed from raw dict (YAML/JSON/Python),
+FastVideo tracks which fields the user explicitly provided vs. which got
+schema defaults. Matters for `request_to_sampling_param()` — explicit
+values override model defaults; schema defaults do NOT.
+
+Mechanics:
+
+- At parse time, original raw dict + baseline snapshot stored on the request.
+- Dataclass field mutations (e.g. `request.sampling.seed = 7`) captured via
+  lightweight `__setattr__` dirty-path recording.
+- Dict-typed field mutations (e.g. `del request.stage_overrides["refine"]`)
+  detected at access time by diffing current dict vs. baseline.
+- Setting a field to its schema default value IS captured as explicit, so
+  it overrides model defaults.
+- Raw dict reconciled lazily when `normalize_generation_request()` is called.
+
+## Schema purity (model-specific fields still in shared schema)
+
+Remain for back-compat during initial migration; targeted for migration
+into preset-owned typed override classes:
+
+| Field | Owner | Migration target |
+|---|---|---|
+| `SamplingConfig.height_sr` / `width_sr` / `num_inference_steps_sr` | Hunyuan15 SR | `HunyuanSRStageOverride` (PR 10) |
+| `SamplingConfig.guidance_scale_2`, `boundary_ratio` | Wan2.2, LingBotWorld | preset-owned (per-family PR) |
+| `InputConfig.mouse_cond`, `keyboard_cond`, `grid_sizes` | MatrixGame | `request.extensions` or typed input config |
+| `InputConfig.c2ws_plucker_emb` | LingBotWorld | `request.extensions` or typed input config |
+| `InputConfig.refine_from`, `stage1_video` | LongCat | `LongCatRefineStageOverride` inputs (PR 9) |
+
+LTX-2 multi-modal CFG knobs (`ltx2_modality_scale_video/_audio`,
+`ltx2_rescale_scale`, `ltx2_stg_scale_video/_audio`,
+`ltx2_stg_blocks_video/_audio`) still leak into shared `SamplingParam` but
+only LTX-2 reads them today. Migration to typed `LTX2SamplingOverride` is
+deferred to per-model migration sweep.
+
+**LTX-2 CFG-force fix landed in PR 6**: defaults moved from `3.0/7.0` to
+`1.0/1.0` to stop force-enabling CFG for non-LTX-2 families.
+`ltx2_base` preset still sets `3.0/7.0` explicitly. Regression guard:
+`test_presets.py::TestPresetDefaultTypes::test_ltx2_cfg_defaults_are_off`.
+
+## Continuation state
+
+Public surface:
+
+```python
+@dataclass
+class ContinuationState:
+    kind: str            # e.g. "ltx2.v1"
+    payload: dict[str, Any]
+```
+
+Internally, model-specific typed subclasses (e.g. `LTX2ContinuationState`
+at [`fastvideo/pipelines/basic/ltx2/continuation.py`](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/continuation.py)).
+
+Payload must be JSON-serializable or use opaque blob-ID indirection for
+large tensors — supports both stateless OpenAI client round-trip AND
+future Dynamo prefill/decode disaggregation.
+
+Hybrid model: server-held for streaming WS, client-round-trip for
+stateless HTTP. See [streaming-server.md](streaming-server.md) D-1.
+
+## Pipeline package structure (target)
+
+Per-family colocation under `pipelines/basic/<family>/`:
+
+```
+fastvideo/pipelines/basic/<family>/
+├── <family>_pipeline.py      # pipeline implementation(s)
+├── presets.py                # user-facing presets (DONE in PR 4)
+├── pipeline_configs.py       # engine/arch config (from configs/pipelines/)
+└── stages/                   # model-specific stages (optional, if >2 files)
+```
+
+What stays shared:
+
+- `configs/pipelines/base.py` — `PipelineConfig` base class
+- `configs/models/` — architecture defs (dits/, vaes/, encoders/)
+- `pipelines/stages/` — shared stages only (denoising, encoding, decoding,
+  text_encoding, timestep_preparation, ...)
+
+What's gone (PR 4):
+
+- `fastvideo/configs/sample/` — directory removed entirely; defaults
+  absorbed into per-family `presets.py`.
+- All 12 `*_SamplingParam` subclass files — `SamplingParam` lives at
+  `fastvideo/api/sampling_param.py`; defaults flow through
+  `SamplingParam.from_pretrained()` → `_from_preset()`.
+
+What's pending: `configs/pipelines/<family>.py` colocation, optional
+`pipelines/stages/<family>_*.py` colocation. Per-model migration PRs
+(6/9/10) include the colocation step for that family.
+
+## YAML examples
+
+### Run config
+
+```yaml
+generator:
+  model_path: /models/ltx2
+  engine:
+    num_gpus: 1
+    parallelism: {tp_size: -1, sp_size: -1}
+    offload: {dit: false, text_encoder: false, vae: false, pin_cpu_memory: true}
+  pipeline:
+    workload_type: t2v
+    preset: ltx2_two_stage
+    components:
+      config_root: /models/ltx2-config
+      upsampler_weights: /models/ltx2-refine
+      lora_path: /models/ltx2-refine-lora
+    preset_overrides:
+      refine: {enabled: true, add_noise: true}
+
+request:
+  prompt: "a fox running through snow"
+  sampling: {num_frames: 121, height: 1024, width: 1536, num_inference_steps: 8, seed: 42}
+  output: {save_video: true, return_state: true}
+  stage_overrides:
+    refine: {num_inference_steps: 2, guidance_scale: 1.0}
+```
+
+### Serve config
+
+See [`Dreamverse/serve_configs/streaming_demo.yaml`](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml)
+for a canonical example matching internal/ui defaults (LTX-2 distilled,
+NVFP4, 121 frames @ 1088×1920 24fps, 5 inference steps, 2-step refine).
+
+## Compatibility mapping (legacy → typed)
+
+| Legacy field | New path |
+|---|---|
+| `model_path` | `generator.model_path` |
+| `num_gpus` | `generator.engine.num_gpus` |
+| `tp_size` / `sp_size` | `generator.engine.parallelism.{tp_size,sp_size}` |
+| `dit_cpu_offload` | `generator.engine.offload.dit` |
+| `enable_torch_compile` | `generator.engine.compile.enabled` |
+| `torch_compile_kwargs` | split: `generator.engine.compile.{backend,fullgraph,mode,dynamic}` + `.extras` |
+| `enable_torch_compile_text_encoder` | `generator.engine.compile.text_encoder_enabled` |
+| `prompt_txt` | `request.inputs.prompt_path` |
+| `image_path` / `video_path` | `request.inputs.{image_path,video_path}` |
+| `output_path` / `save_video` / `return_frames` | `request.output.*` |
+| `seed` / `num_frames` / `height` / `width` / `fps` / `num_inference_steps` / `guidance_scale` | `request.sampling.*` |
+| `enable_teacache` / `return_trajectory_*` | `request.runtime.*` |
+
+LTX-2 specific (private adapter, NOT public compat promise):
+
+| Legacy LTX-2 field | New path |
+|---|---|
+| `config_model_path` | `generator.pipeline.components.config_root` |
+| `ltx2_refine_enabled` | `generator.pipeline.preset_overrides.refine.enabled` |
+| `ltx2_refine_upsampler_path` | `generator.pipeline.components.upsampler_weights` |
+| `ltx2_refine_lora_path` | `generator.pipeline.components.lora_path` |
+| `ltx2_refine_num_inference_steps` | `request.stage_overrides.refine.num_inference_steps` |
+| `ltx2_refine_guidance_scale` | `request.stage_overrides.refine.guidance_scale` |
+| `ltx2_refine_add_noise` | `generator.pipeline.preset_overrides.refine.add_noise` |
+| `ltx2_image_crf` | `request.stage_overrides.refine.image_crf` |
+| `return_continuation_state` | `request.output.return_state` |
+
+LongCat:
+
+| Legacy | New |
+|---|---|
+| `refine_from` / `stage1_video` | `request.inputs.{refine_from,stage1_video}` |
+| `t_thresh` / `spatial_refine_only` / `num_cond_frames` | `request.stage_overrides.refine.*` |
+
+## External inspirations (and limits)
+
+| Source | Useful idea | Don't copy |
+|---|---|---|
+| Ray | YAML-first config interchange | Ray's package layout |
+| SGL `multimodal_gen` | Split instance/request config; dict input parsed into typed objects; merge user overrides on model defaults | `SamplingParams._adjust(ServerArgs)` (request depending on engine config); broad weakly-typed request bags |
+| vLLM-Omni | Model-owned pipeline presets; explicit stage topology; per-stage default sampling | Positional `sampling_params_list`; serving-engine stage-index semantics in primary Python API |
+
+## Naming guidance
+
+- Public schema names namespaced under `fastvideo.api`
+- Don't export from top-level `fastvideo/__init__.py` until migration further along
+- `RunConfig` / `ServeConfig` get sufficient disambiguation from training
+  config via the namespace
+- Future rename to `EngineQuantizationConfig` reserved if a collision
+  arises (deferred)
+
+## Public Python API (canonical form)
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    GeneratorConfig, GenerationRequest,
+    EngineConfig, OutputConfig,
+    PipelineSelection, SamplingConfig,
+)
+
+generator = VideoGenerator.from_pretrained(
+    config=GeneratorConfig(
+        model_path="/models/ltx2",
+        engine=EngineConfig(num_gpus=1),
+        pipeline=PipelineSelection(workload_type="t2v", preset="ltx2_two_stage"),
+    )
+)
+
+result = generator.generate(
+    GenerationRequest(
+        prompt="a fox running through snow",
+        sampling=SamplingConfig(num_frames=121, height=1024, width=1536,
+                                 num_inference_steps=8, seed=42),
+        output=OutputConfig(save_video=True, return_state=True),
+    )
+)
+```
+
+Accepted constructor forms:
+
+```python
+VideoGenerator.from_pretrained(config=GeneratorConfig(...))
+VideoGenerator.from_config(GeneratorConfig(...))
+VideoGenerator.from_file("run.yaml")
+VideoGenerator.from_pretrained("model-id", num_gpus=2, ...)  # stable convenience
+VideoGenerator.from_pretrained(model_path, **legacy_kwargs)   # compat (deprecated PR 13)
+```

.agents/memory/dreamverse-integration/integration-review.md

@@ -0,0 +1,888 @@
+# Integration Review — Drift Audit + Path Forward
+
+> # ⚠️ DEPRECATED — superseded by [integration-plan.md](integration-plan.md)
+>
+> This document recommended **Option D** (Dreamverse stays a separate repo;
+> generic backend merges into FastVideo). On 2026-05-05 the team chose
+> **Option B+** instead (Dreamverse FE + product server move into FastVideo
+> as `apps/dreamverse/`; generic backend stays at
+> `fastvideo.entrypoints.streaming.*` per Option D's principle).
+> See [decisions-log.md D-18](decisions-log.md#d-18) for the strategy
+> reversal rationale and [integration-plan.md](integration-plan.md) for the
+> executable migration plan.
+>
+> **What's still authoritative in this file:**
+> - **Part 1 — Drift audit** (the 17-row drift summary table). The drift
+>   findings remain valid; the migration plan in `integration-plan.md`
+>   folds them into specific phases.
+> - **OSS precedent citations** (vLLM, BentoML, Ray Serve, TGI+ChatUI,
+>   Transformers.js, ComfyUI, AUTOMATIC1111). Reused in `integration-plan.md`.
+>
+> **What's superseded:**
+> - **Part 2 — Recommendation (Option D)**. Replaced by Option B+ in the
+>   new plan. Read `integration-plan.md` for the current decision.
+> - **Part 3 — Action items**. Replaced by the phased migration plan.
+>
+> Kept in tree for historical reference and audit trail. Do not delete.
+
+**Last updated:** 2026-05-05 (deprecated header added).
+
+**Scope:** FastVideo public `will/ltx2_sr_port` at the requested audit
+anchor `b36bdbc9`; Dreamverse `will/integrate-public-fastvideo` at
+`ec8ef92`; FastVideo-internal `will/rebase-nbv` as read-only comparison.
+
+**Memory-dir context:** the current integration memory snapshot tracks the
+same public mega-PR lineage as `will/ltx2_sr_port`, with PRs #1257,
+#1258, #1284, and #1286 already merged, #1287 closed, and #1288 open as
+the consolidated landing vehicle for LTX-2 SR runtime, NVFP4,
+`generate_async`, Dynamo contract, and memory-dir cleanup. Source:
+[memory index](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/README.md#L8-L19)
+and [D-17](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L19-L45).
+
+**Bottom line:** Zero core typed API drift — typed construction, typed
+continuation state, NVFP4 wiring, and Dynamo-facing async events are either
+already public or in #1288. **Real drift remains on the realtime-runtime
+contract surface (`/healthz` / `/readyz` / `/status` routes per
+[cross-repo-surfaces.md](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L74-L88))
+and on operational/product edges**: stale Dreamverse docs/scripts, a
+1933-LOC Dreamverse prompt-enhancer fork, two unresolved per-session
+fields (`ltx2_image_crf` D-8, `video_position_offset_sec` VPO), one
+missing example config, and two internal-only utilities whose product
+relevance is not yet proven.
+
+---
+
+## Part 1 — Drift audit
+
+### Methodology
+
+1. **Compared three repositories and branches.**
+   - FastVideo public: `/home/william5lin/FastVideo`, branch
+     `will/ltx2_sr_port`.
+   - Dreamverse: `/home/william5lin/Dreamverse`, branch
+     `will/integrate-public-fastvideo`.
+   - FastVideo-internal: `/home/william5lin/FastVideo-internal`, branch
+     `will/rebase-nbv`.
+   - Canonical repo paths are listed in the integration memory index:
+     [repo paths](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/README.md#L72-L79).
+
+2. **Scoped the audit to the ultimate integration goal.**
+   - Dreamverse should depend on public `fastvideo`, not
+     `FastVideo-internal`.
+   - FastVideo should own the reusable backend subset that Dreamverse
+     currently needs from internal: streaming runtime, GPU pool, router,
+     prompt enhancer, NVFP4, continuation state, and typed generation.
+   - Dynamo should consume FastVideo through typed public Python APIs, not
+     through private modules.
+   - The three Dreamverse surfaces are documented as pipeline construction,
+     realtime runtime, and continuation state:
+     [cross-repo surfaces](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L13-L20).
+
+3. **Separated intentional refactor from drift.**
+   - A path rename is not drift if the public branch contains the same
+     responsibility under the typed design.
+   - A deleted file is not drift if the public design intentionally
+     consolidated it.
+   - A private alias is not drift if the public schema exposes a typed
+     replacement with contract tests.
+   - This matches the typed-public-boundary rule in
+     [design.md](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L24-L43).
+
+4. **Used memory docs for rationale and worktree files for concrete proof.**
+   - API schema and public exports:
+     [schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L68-L85),
+     [api exports](file:///home/william5lin/FastVideo/fastvideo/api/__init__.py#L49-L109).
+   - Streaming server current routes:
+     [build_app](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/server.py#L88-L160).
+   - Dreamverse dependency state:
+     [pyproject server extra](file:///home/william5lin/Dreamverse/pyproject.toml#L17-L22),
+     [uv lock editable source](file:///home/william5lin/Dreamverse/uv.lock#L716-L722).
+   - Contract tests:
+     [Dreamverse shape](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L1-L26),
+     [Dynamo shape](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dynamo_shape.py#L1-L19),
+     [generate_async](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_generate_async.py#L1-L7).
+
+5. **Did not treat product-only Dreamverse behavior as FastVideo drift.**
+   - Dreamverse keeps a local product server and Next.js UI today:
+     [README baseline](file:///home/william5lin/Dreamverse/README.md#L5-L16).
+   - Product-only routes, curated presets, devtools, and frontend-specific
+     behavior belong in Dreamverse unless a second non-Dreamverse consumer
+     needs them.
+
+6. **Risk scale used below.**
+   - **P0:** blocks Dreamverse from running without FastVideo-internal.
+   - **P1:** blocks clean `BE_FLAVOR=fastvideo` or Dynamo/public API use.
+   - **P2:** reproducibility or maintenance drag.
+   - **P3:** optional parity or future memory/perf improvement.
+
+### Findings: zero core typed API drift
+
+The public branch is aligned with the goal on the **core typed API
+surface** (construction, request, continuation state, async events).
+The table below lists items that look like drift only if compared by
+path name or legacy field name. They are intentional public refactors
+or already guarded by tests. **Note:** the realtime-runtime _contract_
+surface (FE-required health routes) is a separate matter — see "real
+drift items" §4 below.
+
+| Investigated item | Drift? | Evidence | Conclusion |
+|---|---:|---|---|
+| Dreamverse surface 1: pipeline construction | No | Dreamverse migrated from flat kwargs to typed `GeneratorConfig` at `d80c2a8`; mapping documented in [cross-repo surfaces](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L22-L47). | Stable public surface exists. |
+| Dreamverse surface 2: realtime runtime | No on architecture; some route work remains | Runtime migration target is public `streaming/`, not internal `realtime/`: [streaming upstream](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L14-L31). | Rename/refactor is intentional. |
+| Dreamverse surface 3: continuation state | No | Public typed `ContinuationState` plus LTX-2 state mapping are documented in [cross-repo surfaces](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L108-L146). | Public state is a superset of Dreamverse's data carrier. |
+| Internal `fastvideo/entrypoints/realtime/` | No | Public design chooses parallel `fastvideo/entrypoints/streaming/`: [layout decision](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L55-L60), [current build_app](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/server.py#L88-L160). | Intentional rename plus typed-config rewrite. |
+| Internal `configs/sample/` presets | No | Public PR 4 intentionally deleted `configs/sample/` and moved defaults to per-family presets plus `fastvideo/api/sampling_param.py`: [design](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L194-L205), [PR roadmap](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/pr-roadmap.md#L21-L29). | Intentional consolidation. |
+| LTX-2 pipeline presets | No | Public target is model-owned named presets and per-family colocation: [design](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L175-L205). | Public layout matches design. |
+| Internal `use_fp4_linear` flag | No | Public typed quant carrier is `engine.quantization.transformer_quant`; schema field exists in [schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L68-L85), compat resolves it in [compat.py](file:///home/william5lin/FastVideo/fastvideo/api/compat.py#L267-L279). | Replaced by typed NVFP4 surface. |
+| Public-only `transformer_quant` field | No | Public `FastVideoArgs` pins typed quant to `dit_config.quant_config`: [fastvideo_args](file:///home/william5lin/FastVideo/fastvideo/fastvideo_args.py#L220-L228), [apply logic](file:///home/william5lin/FastVideo/fastvideo/fastvideo_args.py#L260-L279). | Public superset, not drift. |
+| Internal `config_model_path` | No | Public typed home is `generator.pipeline.components.config_root`: [design mapping](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L261-L269), [compat mapping](file:///home/william5lin/FastVideo/fastvideo/api/compat.py#L295-L299). | Alias is covered. |
+| Internal flat video request fields | No | Public `GenerationRequest` nests `inputs`, `sampling`, `runtime`, `output`, `state`, `extensions`: [schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L193-L204). Internal legacy fields live in internal protocol at [protocol.py](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/openai/protocol.py#L64-L82). | Intentional request refactor. |
+| Dreamverse typed init kwargs | No | Contract test asserts current Dreamverse load kwargs all land on typed fields, not `experimental`: [test_dreamverse_shape](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L44-L135). | Guard in place. |
+| Dreamverse request path | No | Contract test asserts request fields round-trip through typed `GenerationRequest`: [test_dreamverse_shape](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L153-L197). | Guard in place. |
+| Dynamo native backend shape | No | FastVideo's only obligation is stable typed Python API: [cross-repo contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L188-L210). | Dynamo should stay out of FastVideo. |
+| `generate_async` event API | No | API exists in [video_generator](file:///home/william5lin/FastVideo/fastvideo/entrypoints/video_generator.py#L264-L332), event types exist in [results.py](file:///home/william5lin/FastVideo/fastvideo/api/results.py#L109-L164). | #1288 covers the async contract. |
+| Dynamo request mapping | No | Authoritative source is the contract test [test_dynamo_shape](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dynamo_shape.py#L90-L175) which asserts `req.prompt`, `req.sampling.{height,width,num_frames,fps,num_inference_steps,guidance_scale,seed,negative_prompt}`, and `req.inputs.{image_path,video_path}` against the actual nested [`GenerationRequest` schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L193-L204). The `streaming-server.md` Dynamo mapping table mis-cites a `prompt -> sampling.prompt` path that no longer exists; the test is correct, the doc is stale and tracked for refresh. | Guard in place; companion doc needs minor refresh. |
+| Public API exports | No | `VideoEvent`, `VideoResult`, and typed schema classes are exported from [fastvideo.api](file:///home/william5lin/FastVideo/fastvideo/api/__init__.py#L49-L109). | Integration imports resolve. |
+| FastVideo-internal FP4/NVFP4 paths | No | Public NVFP4 files and roles are documented in [quantization](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/quantization.md#L24-L35); actual `NVFP4Config` documents lazy FlashInfer and public naming in [nvfp4_config.py](file:///home/william5lin/FastVideo/fastvideo/layers/quantization/nvfp4_config.py#L1-L19). | Public is typed superset. |
+| AbsMaxFP8 refactor | No for Dreamverse | Public quant registry includes `AbsMaxFP8` and `NVFP4`: [quantization init](file:///home/william5lin/FastVideo/fastvideo/layers/quantization/__init__.py#L1-L8). AbsMaxFP8 failure is tracked as separate tech debt: [open threads](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L100-L117). | Not Dreamverse blocker. |
+| Internal realtime API regression test | No | Public contract tests replace it: [Dreamverse contract](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L1-L26), [Dynamo contract](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dynamo_shape.py#L1-L19), [generate_async tests](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_generate_async.py#L91-L230). | Better scoped guards exist. |
+| Dreamverse dependency declaration | No | `server` extra declares `fastvideo>=0.1.7`: [pyproject](file:///home/william5lin/Dreamverse/pyproject.toml#L17-L22). Dev lock resolves editable public `../FastVideo`: [uv.lock](file:///home/william5lin/Dreamverse/uv.lock#L716-L722), [package source](file:///home/william5lin/Dreamverse/uv.lock#L777-L780). | Dependency is already switched in metadata/lock. |
+
+#### Core conclusion for the zero-typed-drift section
+
+The public typed API no longer needs to mirror `FastVideo-internal` file
+paths. The correct test is whether Dreamverse and Dynamo can express their
+needs through public typed objects and public entrypoints. On that test,
+the **typed core** is covered (construction, request, continuation,
+async events). The **runtime contract** still has health-route gaps —
+see real drift §4. On the **typed core**:
+
+- `GeneratorConfig` and `GenerationRequest` cover construction and calls:
+  [schema surface](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L45-L72).
+- `ServeConfig.streaming` covers the server envelope:
+  [schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L244-L279).
+- `generate_async` covers streaming, OpenAI, and Dynamo on one substrate:
+  [video_generator](file:///home/william5lin/FastVideo/fastvideo/entrypoints/video_generator.py#L264-L332).
+- Contract tests now encode the cross-repo shapes:
+  [Dreamverse](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L70-L214),
+  [Dynamo](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dynamo_shape.py#L170-L331),
+  [async events](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_generate_async.py#L91-L273).
+
+### Findings: real drift items requiring action
+
+#### 1. Dreamverse README and bootstrap script still point at FastVideo-internal
+
+- **Priority:** P0 for a clean public-dependency story.
+- **Effort:** Small.
+- **Owner:** Dreamverse repo.
+- **Evidence:** Dreamverse metadata already points at public FastVideo:
+  [pyproject](file:///home/william5lin/Dreamverse/pyproject.toml#L17-L22),
+  [uv source](file:///home/william5lin/Dreamverse/pyproject.toml#L54-L61),
+  [uv.lock](file:///home/william5lin/Dreamverse/uv.lock#L716-L722).
+- **Drift:** README still tells users that `uv` resolves from
+  `../FastVideo-internal` and that bootstrap expects `../FastVideo-internal`:
+  [README](file:///home/william5lin/Dreamverse/README.md#L76-L109).
+- **Drift:** bootstrap script still defaults to cloning the private repo and
+  verifying imports from that clone:
+  [script defaults](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L7-L11),
+  [script clone flow](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L33-L63),
+  [script import assertion](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L66-L88).
+- **Action:** Replace private-fork bootstrap with public FastVideo bootstrap
+  or delete the bootstrap once PyPI publication is the default path.
+- **Do not overreach:** no FastVideo code change required.
+
+#### 2. Dreamverse carries a 1933-line prompt-enhancer fork
+
+- **Priority:** P1.
+- **Effort:** Medium.
+- **Owner:** Dreamverse repo, after public prompt enhancer is available.
+- **Evidence:** Dreamverse local fork starts at
+  [server/prompt_enhancer.py](file:///home/william5lin/Dreamverse/server/prompt_enhancer.py#L1-L80).
+- **Public replacement:** FastVideo now has provider-agnostic
+  `PromptEnhancer` with `enhance`, `auto_extend`, `rewrite`, and
+  `register_provider`:
+  [public enhancer](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/prompt/enhancer.py#L66-L142).
+- **Provider extension point:** custom providers implement `LLMProvider`:
+  [provider protocol](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/prompt/providers/base.py#L63-L75).
+- **Tracking:** DR-1 in open threads already defines the compat-shim shape:
+  [DR-1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L174-L206).
+- **Action:** Replace the fork with a small Dreamverse shim that adapts
+  public `LLMResponse` to Dreamverse's product response objects and keeps
+  only product-only extras.
+- **Do not overreach:** do not merge Dreamverse's full prompt product layer
+  into FastVideo unless a second consumer needs the same semantics.
+
+#### 3. `cerebras_ifm` provider is unresolved
+
+- **Priority:** P1 if Dreamverse needs IFM in production; P2 otherwise.
+- **Effort:** Small decision plus small/medium implementation.
+- **Owner:** Team decision; implementation either Dreamverse-side or public.
+- **Public state:** `PromptEnhancerConfig.provider` is currently
+  `Literal["cerebras", "groq"]`:
+  [schema](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L229-L235).
+- **Design note:** public Literal excludes `cerebras_ifm` today:
+  [streaming-server D-3](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L61-L100).
+- **Tracking:** DR-2 already frames the public-vs-Dreamverse decision:
+  [DR-2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L211-L229).
+- **Recommended default:** implement IFM as a Dreamverse-side custom provider
+  registered through `enhancer.register_provider(...)` unless there is a
+  non-Dreamverse public user.
+
+#### 4. `/healthz`, `/readyz`, and `/status` are not in public `build_app`
+
+- **Priority:** P1 for `BE_FLAVOR=fastvideo` frontend compatibility.
+- **Effort:** Medium/Large because route shapes need tests.
+- **Owner:** FastVideo public.
+- **Public current state:** `build_app` exposes `GET /health` and
+  `WS /v1/stream`:
+  [server.py](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/server.py#L126-L160).
+- **Dreamverse expected state:** Dreamverse exposes `GET /healthz`,
+  `GET /readyz`, and `GET /status`:
+  [routes/health.py](file:///home/william5lin/Dreamverse/server/routes/health.py#L34-L79).
+- **Tracking:** open item #1 documents route ownership and files likely to
+  touch:
+  [open threads](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L69-L99).
+- **Design note:** `/curated-presets`, `/prompt-system-config`, and devtools
+  stay Dreamverse-side, with feature detection:
+  [streaming route contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L232-L259).
+
+#### 5. `fastvideo/models/layerwise_offload.py` exists only internally
+
+- **Priority:** P3 unless memory-tight Dreamverse deployments require it.
+- **Effort:** Medium if adopted; low if documented as deferred.
+- **Owner:** FastVideo public only if a concrete deployment needs it.
+- **Internal evidence:** internal file defines async layerwise CPU offload
+  manager with pinned CPU memory and prefetch stream:
+  [layerwise_offload.py](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L1-L20),
+  [prefetch path](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L127-L180).
+- **Public state:** no equivalent public file was identified in this audit.
+- **Action:** defer unless Dreamverse or another public deployment hits a
+  memory ceiling that cannot be handled by existing offload knobs.
+- **Decision rule:** if adopted, port as a generic offload utility with
+  tests; do not make it Dreamverse-specific.
+
+#### 6. Standalone LTX-2 upsampler CLI exists only internally
+
+- **Priority:** P2 for reproducibility; P3 for product runtime.
+- **Effort:** Small/Medium after scope decision.
+- **Owner:** FastVideo public if standalone upsampling is a supported user
+  workflow.
+- **Internal utility:** `upscale_video_file(...)` reads an existing video,
+  prepares frame count/resolution, loads VAE + upsampler, and writes an mp4:
+  [upsample.py](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/upsample.py#L120-L180),
+  [write tail](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/upsample.py#L181-L202).
+- **Internal CLI:** `fastvideo upsample` wrapper exists internally:
+  [cli/upsample.py](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L15-L35),
+  [CLI args](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L48-L130).
+- **Public related functionality:** LTX-2 SR refine stage covers the
+  in-pipeline latent upsample/refine path:
+  [ltx2_refine.py](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L1-L22),
+  [upsample stage](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L116-L180).
+- **Action:** decide whether standalone file-to-file upsampling is a public
+  CLI promise or whether the SR refine stage is sufficient.
+
+#### 7. Reproducible streaming demo config lives only in Dreamverse
+
+- **Priority:** P2.
+- **Effort:** Small.
+- **Owner:** FastVideo public.
+- **Evidence:** canonical demo config currently lives at
+  [Dreamverse/serve_configs/streaming_demo.yaml](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L1-L12).
+- **Config content:** it documents LTX-2 distilled model, one GPU,
+  no offload, compile settings, NVFP4, refine overrides, default request,
+  and streaming settings:
+  [generator block](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L31-L87),
+  [streaming block](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L108-L149).
+- **Memory pointer:** design.md already treats this as the canonical
+  example:
+  [design YAML example](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L235-L239).
+- **Action:** copy/adapt it into
+  `examples/serving/streaming_demo.yaml` with public-safe comments.
+
+#### 8. LTX-2 stage equivalence is a verification gap, not proven drift
+
+- **Priority:** P2.
+- **Effort:** Medium if parity checks are added; small if only manual audit.
+- **Owner:** FastVideo public.
+- **Public state:** model-specific LTX-2 stages are colocated under
+  `fastvideo/pipelines/basic/ltx2/stages/`, consistent with the target
+  layout in [design.md](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/design.md#L175-L205).
+- **Example public stage:** `ltx2_refine.py` explicitly says it is a
+  public-side port of the internal stage and describes the three-stage SR
+  flow:
+  [ltx2_refine.py](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L1-L22).
+- **Action:** verify behavior for the six internal `ltx2_*` stage files
+  against public colocated stages. If a mismatch is found, file it as a
+  real drift item with a failing parity test.
+
+### Findings: deferred / accepted residual
+
+These items should not block the public-dependency transition.
+
+1. **StepVideo residual.**
+   - Dreamverse's model registry is LTX-2/LTX-2.3 only:
+     [Dreamverse config](file:///home/william5lin/Dreamverse/server/config.py#L28-L45).
+   - Internal local tests even stub StepVideo modules to keep LTX registry
+     tests focused:
+     [test_ltx2_registry.py](file:///home/william5lin/FastVideo-internal/tests/local_tests/test_ltx2_registry.py#L38-L61).
+   - Conclusion: accepted low-priority deferral unless Dreamverse adds a
+     StepVideo model.
+
+2. **Internal debug-only `FastVideoArgs` fields.**
+   - Internal debug fields exist around `FastVideoArgs` and stage/model sums:
+     [internal grep source](file:///home/william5lin/FastVideo-internal/fastvideo/fastvideo_args.py#L200-L203).
+   - They are debug-only and not a public user-facing integration surface.
+   - Conclusion: low-priority; do not add to public schema unless a debug
+     workflow requires them.
+
+3. **Private request aliases.**
+   - Public request schema is nested and strict:
+     [GenerationRequest](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L193-L204).
+   - Legacy OpenAI flat fields are compatibility input, not the canonical
+     public API:
+     [internal protocol](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/openai/protocol.py#L64-L82).
+   - Conclusion: no action beyond current compat tests.
+
+4. **`experimental["pipeline_config"]` escape hatch.**
+   - Dreamverse currently uses an explicit in-memory quant config because
+     typed `transformer_quant: "NVFP4"` does not expose `layer_profile`:
+     [quantization](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/quantization.md#L86-L97).
+   - Open thread #4 tracks `layer_profile`:
+     [open threads](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L244-L260).
+   - Conclusion: defer broader typed carrier design; add `layer_profile`
+     first if Dreamverse needs base/refine profile selection.
+
+5. **Router sticky routing and active-active semantics.**
+   - Public router intentionally ships active-passive first and defers
+     sticky/weighted routing:
+     [D-15](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L118-L155).
+   - Follow-ups are tracked:
+     [D-15 action items](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L175-L201).
+   - Conclusion: not drift; defer until load-balancing needs are real.
+
+6. **AbsMaxFP8 failure.**
+   - Pre-existing and not introduced by NVFP4:
+     [state](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/state.md#L154-L159),
+     [quantization](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/quantization.md#L202-L214).
+   - Conclusion: fix separately; not a Dreamverse public-dependency blocker.
+
+### Drift summary table
+
+| # | Item | Priority | Effort | Status | Tracked where | Next action |
+|---:|---|---|---|---|---|---|
+| 1 | Dreamverse README still names `../FastVideo-internal` | P0 | S | Real drift | [README lines](file:///home/william5lin/Dreamverse/README.md#L76-L109) | Update docs to public FastVideo / PyPI path. |
+| 2 | Dreamverse private bootstrap clones internal repo | P0 | S | Real drift | [bootstrap script](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L7-L11) | Replace or delete private bootstrap. |
+| 3 | Dreamverse `prompt_enhancer.py` fork | P1 | M | Real drift | [DR-1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L174-L206) | Build compat shim over public enhancer. |
+| 4 | `cerebras_ifm` provider path | P1/P2 | S-M | Real drift / decision | [DR-2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L211-L229) | Choose public provider vs Dreamverse custom provider. |
+| 5 | Health route mismatch | P1 | M-L | Real drift | [open item #1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L69-L99) | Add `/healthz`, `/readyz`, `/status` to public build_app. |
+| 6 | Missing public streaming demo config | P2 | S | Real drift | [Dreamverse config](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L1-L12) | Add `examples/serving/streaming_demo.yaml`. |
+| 7 | Standalone upsampler CLI | P2/P3 | S-M | Real drift if standalone CLI is desired | [internal CLI](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L15-L35) | Decide CLI promise; port or defer. |
+| 8 | Layerwise offload utility | P3 | M | Optional internal-only residual (no Dreamverse deployment requires it today) | [internal manager](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L15-L20) | Defer until memory-tight deployment needs it. |
+| 9 | LTX-2 stage equivalence | P2 | S-M | Verification gap | [public refine stage](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L1-L22) | Add targeted parity audit/test if needed. |
+| 10 | StepVideo | P3 | M | Accepted residual | [Dreamverse model registry](file:///home/william5lin/Dreamverse/server/config.py#L28-L45) | No action unless Dreamverse adds StepVideo. |
+| 11 | Debug-only fields | P3 | S | Accepted residual | [internal args](file:///home/william5lin/FastVideo-internal/fastvideo/fastvideo_args.py#L200-L203) | Do not publicize unless needed. |
+| 12 | `layer_profile` typed quant knob | P2 | M | Tracked gap | [open item #4](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L244-L260) | Add typed layer profile if Dreamverse drops escape hatch. |
+| 13 | `ltx2_image_crf` per-segment field flow (D-8) | P1 | S | Open verification gap — Dreamverse still passes `ltx2_image_crf=0.0` per [Dreamverse video_generation.py](file:///home/william5lin/Dreamverse/server/video_generation.py#L420-L435); needs trace-through to confirm it lands on `request.stage_overrides.refine.image_crf` rather than being silently dropped | [D-8 in open-threads](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L52-L68) | 10-min trace + add a Dreamverse-shape contract test pinning the field. |
+| 14 | `video_position_offset_sec` semantics (VPO) | P1 | S | Open decision — persistent-vs-per-segment ambiguity unresolved | [VPO in open-threads](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L120-L144) | Confirm semantics with audio team; document + add test. Decision deadline was "before PR 7.6 emits state" — that PR (7.6 / #1257) is now MERGED, so the decision is overdue. |
+| 15 | `GpuPool` ABC docstring missing experimental caveat (D-12-A) | P3 | trivial | Tracked gap — `GpuPool` ABC at [gpu_pool.py:74-83](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/gpu_pool.py#L74-L83) lacks the "API may change post-PR-7.10; experimental / server-internal" caveat | [D-12-A](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L301-L313) | Edit docstring; trivial. |
+| 16 | `GpuPool.run_async()` migration (D-12-B) | P2 | M | Tracked gap — `GpuPool.run() -> Any` should become `run_async() -> AsyncIterator[VideoEvent]` per D-12 / D-12-B | [D-12-B](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L317-L327) | Land alongside #1288 merge or in immediate follow-up. |
+| 17 | `SessionStore` / `BlobStore` lifecycle policy (SBS) | P2 | M | Tracked gap — in-memory defaults have no eviction/TTL/blob-cleanup policy | [SBS](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L278-L294) | Streaming server design pass needed before high-traffic deployment. |
+| 13 | Router sticky / active-active | P3 | M | Deferred | [D-15](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L175-L201) | Defer until reconnect/load evidence. |
+| 14 | AbsMaxFP8 test failure | P2 | S | Separate tech debt | [state](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/state.md#L154-L159) | Fix outside Dreamverse migration. |
+| 15 | Dynamo backend package | P1 | External | Not FastVideo drift | [Dynamo contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L188-L210) | Reopen Dynamo-side PR after public API lands. |
+
+---
+
+## Part 2 — Integration path tradeoffs
+
+### The four options
+
+#### Option A — Status quo: Dreamverse stays separate and depends on `fastvideo`
+
+**Shape**
+
+- FastVideo remains the Python library and reusable backend runtime.
+- Dreamverse remains the product repo with FastAPI product glue and Next.js
+  frontend.
+- Dreamverse `server` extra depends on `fastvideo>=0.1.7`:
+  [pyproject](file:///home/william5lin/Dreamverse/pyproject.toml#L17-L22).
+- Local development can keep using editable `../FastVideo` until PyPI
+  publication catches up:
+  [uv.lock](file:///home/william5lin/Dreamverse/uv.lock#L716-L722).
+
+**What it solves**
+
+- Directly satisfies "Dreamverse depends on public FastVideo".
+- Keeps frontend release cadence independent.
+- Keeps product-specific prompts, routes, and UI in the product repo.
+- Minimizes FastVideo packaging and CI growth.
+
+**What it does not solve by itself**
+
+- Does not remove Dreamverse prompt-enhancer fork unless DR-1 is executed.
+- Does not give Dreamverse FE compatibility with public `build_app` until
+  health routes migrate.
+- Does not make Dreamverse server itself reusable as a public entrypoint.
+
+**Best fit**
+
+- Default for the next release if the goal is to stop using
+  FastVideo-internal quickly and safely.
+
+#### Option B — Dreamverse as a subfolder under FastVideo
+
+**Shape**
+
+- One repository: FastVideo contains `dreamverse/server/` and
+  `dreamverse/apps/web/`.
+- Dreamverse can remain a separate package in the same repo, or FastVideo's
+  build can ignore Dreamverse by default.
+- CI must understand Python library tests plus Next.js install/build/test.
+
+**What it solves**
+
+- Eliminates sibling-checkout drift.
+- Makes cross-repo integration changes atomic.
+- Easier for a single reviewer to see library and product changes together.
+
+**Costs**
+
+- Adds frontend dependency management to a Python ML library repo.
+- Couples clone size, CI setup, issue tracking, and review load.
+- Forces maintainers to decide whether product assets are included in source
+  distributions, wheels, docs, and release notes.
+
+**Best fit**
+
+- Only if Dreamverse becomes the primary FastVideo product surface and the
+  team accepts a product monorepo.
+
+#### Option C — Full merge into `fastvideo.entrypoints.dreamverse.*`
+
+**Shape**
+
+- Dreamverse backend becomes FastVideo code.
+- Public import becomes something like
+  `from fastvideo.entrypoints.dreamverse import build_app`.
+- CLI becomes `fastvideo dreamverse-serve --config dreamverse.yaml`.
+- Frontend either ships as static assets in the package or as a frontend
+  extra.
+
+**What it solves**
+
+- One namespace and one release train for library plus product backend.
+- No dependency boundary between Dreamverse server and FastVideo internals.
+- Product route contract can be tested entirely inside FastVideo CI.
+
+**Costs**
+
+- Maximally expands FastVideo's public/security surface.
+- Locks product experiments to FastVideo release cadence.
+- Makes private prompt/provider/product assumptions look like framework API.
+- Has weak precedent for a Python ML library plus Next.js product being merged
+  into the library namespace.
+
+**Best fit**
+
+- Only if Dreamverse is no longer a separate product and becomes the
+  canonical FastVideo UI/serving mode.
+
+#### Option D — Hybrid: backend merges, frontend stays separate
+
+**Shape**
+
+- Reusable backend components merge into public FastVideo.
+- Frontend stays in a separate Dreamverse UI repo or Dreamverse product repo.
+- The backend should be generic where possible: `fastvideo.entrypoints.streaming`,
+  not product-only names, unless product-only routes are intentionally
+  accepted as public API.
+- This matches the current trajectory: streaming server, GPU pool, prompt
+  enhancer, safety/rewrite/session logging, router, NVFP4, and
+  `generate_async` are public-side work already tracked in the PR roadmap:
+  [pr-roadmap](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/pr-roadmap.md#L19-L42).
+
+**What it solves**
+
+- Removes FastVideo-internal dependency for reusable backend pieces.
+- Keeps product frontend cadence independent.
+- Gives non-Dreamverse users a streaming backend and typed API without
+  carrying the Dreamverse app.
+- Gives Dynamo a stable library API while leaving Dynamo package code in
+  Dynamo:
+  [Dynamo contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L188-L210).
+
+**Costs**
+
+- Requires careful boundary discipline: generic streaming/server code in
+  FastVideo; product routes/prompts/presets in Dreamverse.
+- Requires contract tests to prevent drift.
+- Some Dreamverse compatibility routes may become public and need support.
+
+**Best fit**
+
+- Best long-term target if the team wants FastVideo to own serving/runtime
+  infrastructure while keeping Dreamverse as a separately evolving product.
+
+### Comparison matrix
+
+| Criterion | A. Separate dep | B. Subfolder monorepo | C. Full namespace merge | D. Hybrid backend merge |
+|---|---|---|---|---|
+| Alignment with stated goal | High: Dreamverse depends on public package | Medium: no external dep, but product becomes repo-local | Medium: dependency disappears by absorption | High: reusable backend in public, product separate |
+| Time to remove `FastVideo-internal` | Fastest | Medium | Slowest | Medium-fast |
+| Build complexity | Low | High: Python + Next.js in one repo | High: Python package plus static/frontend extras | Medium: Python backend only in FastVideo |
+| Release cadence | Independent | Coupled clone; releases can still be separate but more friction | Fully coupled | Backend coupled to FastVideo, frontend independent |
+| Security surface in FastVideo | Low | Medium/High | Highest | Medium |
+| Contributor friction | Low for both repos | Higher for library contributors | Highest; product assumptions in library | Medium; clear backend boundary needed |
+| Dependency management | Normal package pin | Workspace/monorepo tooling needed | FastVideo extras/static asset decisions needed | FastVideo extras for backend; FE out-of-tree |
+| CI cost | Low/medium | High | High | Medium |
+| Contract-test value | High; cross-repo contract tests are essential | Medium; same repo but still useful | Medium; less boundary pressure | High; generic backend vs product boundary |
+| Precedent strength | Strong: library/server plus external UI patterns exist | Mixed | Weak for Python ML library + Next.js inside namespace | Strongest match: in-tree server/backend, external UI |
+| Packaging risk | Low | Medium/high | High | Medium |
+| Future Dynamo fit | Strong | Strong if API remains clean | Risky if product API bleeds in | Strong |
+| Frontend iteration speed | Highest | Lower | Lowest | Highest |
+| Risk of product-specific API leakage | Low | Medium | High | Medium; controllable with naming discipline |
+| Reversibility | High | Medium | Low | Medium/high |
+
+### OSS precedents (with citations)
+
+| Pattern | Project | What it supports | Citation |
+|---|---|---|---|
+| Library plus in-tree server | vLLM | A Python ML library can ship an in-tree OpenAI-compatible server while clients remain external. | https://github.com/vllm-project/vllm/blob/bcf5cac9fb956788f649d1f5297b74c886a9d6d3/README.md#L64-L74 |
+| Service packaging | BentoML | Packaging model + service + dependencies is supported, but CWD packaging creates discipline needs. | https://github.com/bentoml/BentoML/blob/32230a5276a8da8b23c4a06a9ec6272c1993451a/docs/source/build-with-bentoml/asgi.rst#L5-L18 |
+| YAML-driven production serving | Ray Serve | Production updates should avoid in-place mutation; use new deployment/traffic switch. | https://docs.ray.io/en/latest/serve/advanced-guides/inplace-updates.html |
+| Library/server plus external UI | TGI + ChatUI | Server can live with backend project while UI is separate. | https://github.com/huggingface/text-generation-inference/blob/b4adbf2f6e2e721280bd0ea5f91d70f7d033f5ed/docs/source/basic_tutorials/consuming_tgi.md#L182-L186 |
+| Lean library plus examples elsewhere | Transformers.js | Library stays lean; demos/examples can live outside core. | https://github.com/huggingface/transformers.js/blob/f7487c737aa8cafbc106c9adf69dc9578c8f3fe0/README.md#L26-L34 |
+| Product monorepo that later split frontend | ComfyUI | Product UI/server monorepo can hit release-cadence mismatch and split FE later. | https://github.com/comfyanonymous/ComfyUI/blob/fed8d5efa6b70d5b24c4c33cb643bfccc39d45b5/README.md#L131-L149 and https://github.com/Comfy-Org/ComfyUI_frontend/blob/60f789d58070a9d1d789b260f83c36d7293a39f0/README.md#L31-L60 |
+| Tightly coupled UI/server product | AUTOMATIC1111 SD WebUI | Product repos can couple UI/server tightly, but security surface becomes product-sized. | https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/82a973c04367123ae98bd9abdf80d9eda9b910e2/webui.py#L48-L104 |
+
+#### Precedent synthesis
+
+- Strong precedents exist for a Python ML library shipping a server entrypoint.
+- Strong precedents exist for keeping frontend/product UI out of the backend
+  library repo.
+- The cited set does not contain a clean precedent for merging a Next.js
+  product into a Python ML library namespace.
+- The most applicable pattern is **backend/server in the ML project,
+  product UI outside**.
+
+### Recommendation
+
+#### Recommend Option D, constrained: backend merges as generic FastVideo streaming; frontend stays separate
+
+Recommendation: follow **Option D** as the long-term architecture, but keep
+the backend merge generic. In practice, this means continuing the current
+public FastVideo path:
+
+- `fastvideo.entrypoints.streaming.*` owns reusable streaming runtime.
+- `fastvideo.entrypoints.streaming.gpu_pool` owns generic GPU worker pools.
+- `fastvideo.entrypoints.streaming.prompt.*` owns provider-agnostic prompt
+  operations.
+- `fastvideo.entrypoints.streaming.router.*` owns FastVideo-aware routing.
+- `fastvideo.api` owns typed construction, requests, results, events, and
+  continuation state.
+- Dreamverse keeps product-only FE, curated presets, prompt UX, product
+  routes, and launch scripts.
+
+This is effectively the path already underway in PRs #1257, #1258, #1284,
+#1286, and #1288:
+[PR roadmap](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/pr-roadmap.md#L21-L42).
+
+#### Why not Option A as the final answer?
+
+Option A is the fastest near-term release posture and should be used as the
+immediate migration posture. However, plain status quo is not enough for
+the ultimate goal because reusable backend pieces still need to live in
+public FastVideo so Dreamverse can stop reaching into internal code. That
+work is already partly complete:
+
+- GPU pool: [D-12](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L47-L116).
+- Prompt enhancer: [PR roadmap 7.7](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/pr-roadmap.md#L32-L35).
+- Streaming auxiliaries: [PR roadmap 7.8](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/pr-roadmap.md#L35-L36).
+- Router: [D-15](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L118-L155).
+- `generate_async`: [streaming-server unlock PR](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L314-L345).
+
+So the practical answer is:
+
+- **Near term:** Option A operationally, after docs/scripts are fixed.
+- **Architecture target:** Option D, with generic backend ownership in
+  FastVideo and product ownership in Dreamverse.
+
+#### Why not Option B?
+
+Option B makes cross-repo coordination easier but imports frontend build,
+package, and CI complexity into FastVideo. That is unnecessary while a
+normal package dependency plus contract tests can guard the integration.
+FastVideo's current repo structure is a Python package with examples and
+docs, not a product monorepo:
+[codebase map](file:///home/william5lin/FastVideo/.agents/memory/codebase-map/README.md#L5-L75).
+
+#### Why not Option C?
+
+Option C makes the product backend a public FastVideo namespace. That is
+only appropriate if the team wants to support Dreamverse as a first-class
+FastVideo product surface. Today the known public obligations are generic:
+typed requests, streaming server, GPU pool, prompt provider protocol,
+router, NVFP4, and Dynamo event APIs. Product-only Dreamverse behavior does
+not need to become framework API.
+
+#### Conditions that would change the recommendation
+
+Move from constrained D toward **C** only if all of these become true:
+
+1. Dreamverse is declared the canonical FastVideo serving product.
+2. Product routes such as curated presets and prompt-system config are
+   accepted as public FastVideo API.
+3. FastVideo maintainers accept the security and support surface.
+4. Release cadence for product UX and FastVideo core is intentionally
+   coupled.
+5. Frontend packaging/static asset strategy is explicitly owned by
+   FastVideo.
+
+Move from constrained D back toward **A** if any of these become true:
+
+1. Prompt enhancement, router, or GPU pool turn out to be Dreamverse-only.
+2. No second user appears for the streaming backend outside Dreamverse.
+3. FastVideo maintainers want to minimize serving surface and publish only
+   Python library APIs.
+4. Dreamverse needs product changes faster than FastVideo can release.
+5. Security review rejects in-tree serving/router responsibilities.
+
+### Migration sketch for the recommended path
+
+#### Phase 0 — Land the public backend stack
+
+- **Effort:** Large, already in flight.
+- **Owner:** FastVideo public.
+- **Files:** #1288 scope, especially `fastvideo/api/`,
+  `fastvideo/entrypoints/video_generator.py`,
+  `fastvideo/entrypoints/streaming/`, LTX-2 pipeline stages, NVFP4 files,
+  and contract tests.
+- **Exit criteria:** #1288 merges; public `fastvideo.api.VideoEvent` and
+  `VideoGenerator.generate_async` are available:
+  [results.py](file:///home/william5lin/FastVideo/fastvideo/api/results.py#L109-L164),
+  [video_generator.py](file:///home/william5lin/FastVideo/fastvideo/entrypoints/video_generator.py#L264-L332).
+
+#### Phase 1 — Fix Dreamverse dependency docs and bootstrap
+
+- **Effort:** Small.
+- **Owner:** Dreamverse.
+- **Files:**
+  - [README.md](file:///home/william5lin/Dreamverse/README.md#L76-L109)
+  - [bootstrap script](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L7-L11)
+  - [pyproject.toml](file:///home/william5lin/Dreamverse/pyproject.toml#L17-L22)
+  - [uv.lock](file:///home/william5lin/Dreamverse/uv.lock#L716-L722)
+- **Exit criteria:** no user-facing docs or scripts mention
+  `FastVideo-internal` as the expected dependency path.
+
+#### Phase 2 — Add public health/readiness/status route compatibility
+
+- **Effort:** Medium/Large.
+- **Owner:** FastVideo public.
+- **Files likely to touch:**
+  - `fastvideo/entrypoints/streaming/server.py::build_app`
+  - new `fastvideo/entrypoints/streaming/health.py`
+  - tests under `fastvideo/tests/entrypoints/streaming/`
+- **Source route shapes:**
+  [Dreamverse health routes](file:///home/william5lin/Dreamverse/server/routes/health.py#L34-L79).
+- **Tracking:** [open item #1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L69-L99).
+- **Exit criteria:** Dreamverse FE can target public `build_app` for
+  `/healthz`, `/readyz`, `/status`, and `/v1/stream`; product-only routes
+  remain feature-detected.
+
+#### Phase 3 — Replace Dreamverse prompt enhancer fork
+
+- **Effort:** Medium.
+- **Owner:** Dreamverse.
+- **Files likely to touch:**
+  - new `Dreamverse/server/prompting/_internal_compat.py`
+  - `Dreamverse/server/runtime.py`
+  - `Dreamverse/server/main.py`
+  - `Dreamverse/server/prompt_enhancer.py`
+- **Public API:**
+  [PromptEnhancer](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/prompt/enhancer.py#L66-L142),
+  [LLMProvider](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/prompt/providers/base.py#L63-L75).
+- **Tracking:** [DR-1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L174-L206).
+- **Exit criteria:** Dreamverse no longer carries a full local fork for the
+  generic prompt operations public FastVideo already owns.
+
+#### Phase 4 — Decide and implement `cerebras_ifm`
+
+- **Effort:** Small decision plus small/medium implementation.
+- **Owner:** Team decision, then Dreamverse or FastVideo.
+- **Default recommendation:** Dreamverse-side custom provider.
+- **Public schema source:**
+  [PromptEnhancerConfig](file:///home/william5lin/FastVideo/fastvideo/api/schema.py#L229-L235).
+- **Tracking:** [DR-2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L211-L229).
+- **Exit criteria:** Dreamverse IFM provider works after prompt fork removal.
+
+#### Phase 5 — Move streaming demo config into FastVideo examples
+
+- **Effort:** Small.
+- **Owner:** FastVideo public.
+- **Source:**
+  [Dreamverse streaming_demo.yaml](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L1-L149).
+- **Target:** `examples/serving/streaming_demo.yaml`.
+- **Exit criteria:** users can reproduce the typed streaming path from the
+  FastVideo repo without checking out Dreamverse.
+
+#### Phase 6 — Remove `experimental["pipeline_config"]` where practical
+
+- **Effort:** Medium for `layer_profile`; Large for a full typed
+  `dit_config.quant_config` carrier.
+- **Owner:** FastVideo public, then Dreamverse cleanup.
+- **Tracking:**
+  [open item #4](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L244-L260),
+  [quantization follow-up](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/quantization.md#L175-L200).
+- **Exit criteria:** Dreamverse can express its quant layer profile through
+  typed config instead of in-memory mutation.
+
+#### Phase 7 — Decide standalone upsampler CLI
+
+- **Effort:** Small/Medium.
+- **Owner:** FastVideo public.
+- **Input:** internal standalone utility
+  [upsample.py](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/upsample.py#L120-L180)
+  and internal CLI
+  [cli/upsample.py](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L48-L130).
+- **Public alternative:** SR refine stage already covers in-pipeline latent
+  upsampling:
+  [ltx2_refine.py](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L116-L180).
+- **Exit criteria:** explicit decision: port CLI, document refine-stage-only
+  support, or defer.
+
+#### Phase 8 — Validate LTX-2 stage parity and offload residuals
+
+- **Effort:** Small/Medium for stage parity; Medium for layerwise offload.
+- **Owner:** FastVideo public.
+- **Stage source:**
+  [public LTX-2 stages](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/).
+- **Offload source:**
+  [internal layerwise offload](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L15-L20).
+- **Exit criteria:** no known behavior gap between internal and public LTX-2
+  stages; offload is either deliberately deferred or ported with tests.
+
+### Open questions
+
+1. **Which provider path for `cerebras_ifm`?**
+   - Public provider or Dreamverse-side custom provider?
+   - Default recommendation: Dreamverse-side unless there is another user.
+   - Source: [DR-2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L211-L229).
+
+2. **Should public FastVideo support standalone LTX-2 file upsampling?**
+   - If yes, port internal CLI.
+   - If no, document that SR support is pipeline-refine only.
+   - Sources: [internal CLI](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L15-L35),
+     [public refine stage](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L1-L22).
+
+3. **Does Dreamverse need layerwise CPU offload?**
+   - If memory-tight deployments require it, port as generic FastVideo.
+   - Otherwise defer.
+   - Source: [internal offload manager](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L15-L20).
+
+4. **How much Dreamverse route surface should FastVideo own?**
+   - Health/readiness/status should migrate because they are part of
+     streaming-server compatibility.
+   - Curated presets and prompt-system config should stay Dreamverse-side
+     unless product policy changes.
+   - Source: [route contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/streaming-server.md#L232-L259).
+
+5. **Should `layer_profile` be the only near-term quant typed addition?**
+   - Adding `layer_profile` is bounded.
+   - A typed carrier for arbitrary mutated `PipelineConfig` is larger design
+     work.
+   - Source: [quantization follow-ups](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/quantization.md#L175-L200).
+
+6. **When does Option D become Option C?**
+   - Only if Dreamverse backend routes become public FastVideo product API.
+   - Until then, keep generic streaming code in FastVideo and product code in
+     Dreamverse.
+
+---
+
+## Part 3 — Action items
+
+1. **P0 / S — Update Dreamverse README dependency notes.**
+   - Replace `../FastVideo-internal` with public FastVideo instructions.
+   - Preserve local editable `../FastVideo` dev flow where useful.
+   - Source: [README stale lines](file:///home/william5lin/Dreamverse/README.md#L76-L109).
+
+2. **P0 / S — Replace or remove private FastVideo bootstrap script.**
+   - Current script clones `FastVideo-internal` and verifies imports from it.
+   - Source: [script](file:///home/william5lin/Dreamverse/.agents/skills/bootstrap-fastvideo-private-fork/scripts/bootstrap_fastvideo_private.sh#L7-L11).
+
+3. **P1 / M-L — Add `/healthz`, `/readyz`, and `/status` to public `build_app`.**
+   - Keep `/curated-presets` and prompt-system config in Dreamverse.
+   - Source: [open item #1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L69-L99).
+
+4. **P1 / M — Replace Dreamverse prompt-enhancer fork with compat shim.**
+   - Wrap public `PromptEnhancer`.
+   - Keep only Dreamverse-specific metadata and product fallback behavior.
+   - Source: [DR-1](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L174-L206).
+
+5. **P1 / S-M — Decide `cerebras_ifm` provider path.**
+   - Default: Dreamverse custom provider via `register_provider`.
+   - Source: [DR-2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L211-L229).
+
+6. **P2 / S — Add `examples/serving/streaming_demo.yaml` to FastVideo.**
+   - Start from Dreamverse config and remove Dreamverse-private comments.
+   - Source: [streaming_demo.yaml](file:///home/william5lin/Dreamverse/serve_configs/streaming_demo.yaml#L1-L149).
+
+7. **P2 / M — Verify each public LTX-2 colocated stage against internal behavior.**
+   - Start with refine, denoising, latent prep, image conditioning, text
+     encoding, and audio decoding.
+   - Source: [public refine stage](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py#L1-L22).
+
+8. **P2 / M — Add typed `transformer_quant_layer_profile` if Dreamverse needs it.**
+   - Thread schema → compat → `FastVideoArgs._apply_transformer_quant`.
+   - Source: [open item #4](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L244-L260).
+
+9. **P2 / S-M — Decide standalone LTX-2 upsampler CLI support.**
+   - Port internal CLI only if file-to-file upsampling is a public workflow.
+   - Source: [internal upsample CLI](file:///home/william5lin/FastVideo-internal/fastvideo/entrypoints/cli/upsample.py#L15-L35).
+
+10. **P2 / S — Fix pre-existing AbsMaxFP8 test failure separately.**
+    - Do not block Dreamverse migration on it.
+    - Source: [open item #2](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/open-threads.md#L100-L117).
+
+11. **P2 / S-M — Document public streaming install extras and dependencies.**
+    - Include router `websockets`, prompt enhancer provider SDKs, and optional
+      safety classifier extras.
+    - Source: [D-16 dependency note](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L224-L242).
+
+12. **P2 / S — Keep contract tests in the FastVideo CI path.**
+    - Guard Dreamverse shape, Dynamo shape, and async events.
+    - Sources: [Dreamverse test](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dreamverse_shape.py#L1-L26),
+      [Dynamo test](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_dynamo_shape.py#L1-L19),
+      [async test](file:///home/william5lin/FastVideo/fastvideo/tests/contract/test_generate_async.py#L1-L7).
+
+13. **P3 / M — Defer layerwise offload until a deployment needs it.**
+    - Port only as generic FastVideo utility with tests.
+    - Source: [internal offload](file:///home/william5lin/FastVideo-internal/fastvideo/models/layerwise_offload.py#L15-L20).
+
+14. **P3 / M — Defer StepVideo public parity for this integration.**
+    - Dreamverse model registry is LTX-2/LTX-2.3 only.
+    - Source: [Dreamverse config](file:///home/william5lin/Dreamverse/server/config.py#L28-L45).
+
+15. **P3 / S — Do not add debug-only fields to public schema by default.**
+    - Keep them private unless there is a user-facing debugging workflow.
+    - Source: [internal debug args](file:///home/william5lin/FastVideo-internal/fastvideo/fastvideo_args.py#L200-L203).
+
+16. **P3 / M — Keep router active-active and sticky routing deferred.**
+    - Add only when session-routing evidence justifies it.
+    - Source: [D-15 action items](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L175-L201).
+
+17. **P3 / S — Preserve Dynamo as an external backend package.**
+    - FastVideo should expose typed API; Dynamo code lives in Dynamo.
+    - Source: [Dynamo contract](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/cross-repo-surfaces.md#L188-L210).
+
+18. **P3 / S — After #1288 merges, update memory-dir state.**
+    - Mark item D resolved and update branch tips.
+    - Source: [runbook post-merge steps](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/runbook.md#L51-L70).
+
+19. **P3 / S — Remove stale split-PR mental model from follow-up docs.**
+    - #1288 is the current vehicle; split bookmarks are historical.
+    - Source: [D-17 implications](file:///home/william5lin/FastVideo/.agents/memory/dreamverse-integration/decisions-log.md#L34-L45).
+
+20. **P3 / S — Keep product-only Dreamverse frontend out of FastVideo unless explicitly re-scoped.**
+    - This preserves release cadence and avoids packaging bloat.
+    - Source: [Dreamverse baseline](file:///home/william5lin/Dreamverse/README.md#L5-L16).

.agents/memory/dreamverse-integration/open-threads.md

@@ -0,0 +1,498 @@
+# Open Threads — Active Follow-Ups
+
+Live work items with priority, effort estimate, dependencies, and
+recommended next action.
+
+For why each item is open see [decisions-log.md](decisions-log.md). For
+PR-level context see [pr-roadmap.md](pr-roadmap.md).
+
+**Last updated:** 2026-05-05 (strategy reversal — PR #1287 CLOSED, replaced
+by mega-PR #1288 on `will/ltx2_sr_port` @ `b36bdbc9` covering the full
+6-layer stack at once. See [decisions-log.md D-17](decisions-log.md#d-17).
+Item D resolution gate is now #1288 merge instead of #1287; same content,
+different vehicle.).
+
+## Priority overview
+
+| # | Pri | Item | Effort | Unblocks |
+|---|---|---|---|---|
+| **D-8** | High | Verify `ltx2_image_crf` post-`d80c2a8` | 10 min | Confirms typed stage-override path actually flows; closes a latent silent-drop bug |
+| **1** | High | Migrate `/healthz`+`/readyz`+`/status` into FastVideo `build_app` | M-L | Closes BE_FLAVOR=fastvideo FE-compatibility; closes streaming-upstream contract debt |
+| **2** | High | Fix pre-existing AbsMaxFP8 test failure | S | Self-contained quantization tech debt |
+| **VPO** | High | Decide `video_position_offset_sec` semantics (a vs b) | 30 min | Unblocks PR 7.6 state emission |
+| **D** | 🟢 in flight | Implement `generate_async` — content shipped in mega-PR **#1288** on `will/ltx2_sr_port` @ `b36bdbc9` (was #1287, CLOSED + re-routed per [D-17](decisions-log.md#d-17)) | L | Closes Q-5/Q-9/PR-7.5 TODOs simultaneously; enables Dynamo backend; unblocks audio re-encode; enables `GpuPool.run_async()` migration (D-12-B). Resolution gate: #1288 merge. |
+| **DR-1** | High | Dreamverse: create `prompting/_internal_compat.py` shim + replace local `prompt_enhancer.py` (1933 LOC) — **PR #1258 has merged (`f673423b`); now actionable** | M (~150-200 LOC shim, replace upstream wiring) | Lets Dreamverse stop carrying a 1933-LOC fork |
+| **DR-2** | Med | Decide `cerebras_ifm` provider path: (a) public Literal + `CerebrasIFMProvider` shipped, OR (b) Dreamverse-side custom provider via `enhancer.register_provider(...)` | S (decision) + S-M (impl) | Resolves the cerebras_ifm gap left by PR #1258. Same item as legacy #3 below; DR-2 is the Dreamverse-side framing. |
+| **3** | Med | Add `cerebras_ifm` to `PromptEnhancerConfig.provider` Literal + provider | S-M | Public-side resolution if DR-2 picks (a) |
+| **4** | Med | Expose `layer_profile` on typed `engine.quantization` | M | Removes Dreamverse's `experimental["pipeline_config"]` dodge for stage profiles |
+| **5** | Med | Design typed `dit_config.quant_config` carrier | L design + L impl | Removes broader `experimental["pipeline_config"]` escape hatch |
+| **SBS** | Med | `SessionStore` / `BlobStore` lifecycle policy | M design | Needed in PR 7.5 design pass |
+| **D-12-A** | Med | Update `GpuPool` ABC docstring: mark "API may change post-PR-7.10; experimental / server-internal" | trivial | Prevents accidental promotion of streaming-internal API to framework-level |
+| **D-12-B** | Med | Replace `GpuPool.run() -> Any` with `run_async() -> AsyncIterator[VideoEvent]` in PR 7.10 cycle | M | Closes the streaming-server cancellation TODO; converges with `generate_async` |
+| **D-13-A** | Med | Document `fastvideo.entrypoints.streaming.prompt.*` in user-facing docs as "streaming-server scoped"; avoid framework-level framing | trivial (docs only) | Keeps future move to `fastvideo.prompt.*` cheap |
+| **D-13-B** | Low | Add optional `client_factory` parameter to `LLMProvider` for `httpx.AsyncClient` pooling | S | Only if metrics show connect/TLS overhead is meaningful |
+| **D-12-C** | Low | Avoid locking `PoolAssignment.gpu_id: int` as public; rename to `worker_id` (already exists) or add `device_ids: list[int]` for topology-aware pooling | S | Future multi-GPU-per-worker refactor stays cheap |
+| **6** | Low | Audio attention quantization profile + test update | S | Future audio quant exploration |
+| **7** | Low | Schema parity inventory cleanup (env-driven prompt fields) | S-M | Long-term consistency |
+| **8** | Low | Stale `apps/web/test-results/` dir cleanup | trivial | Cosmetic |
+| **11** | Low | Promote LTX-2 prompt orchestration (locked segments, segment_prompts JSON shape, rollout id/label) to `fastvideo.entrypoints.streaming.prompt.ltx2_orchestration` | M | Resolves Q-2 from decisions-log when a second LTX-2-style consumer appears |
+| **12** | Low | When streaming server starts using `PromptSafetyFilter`, ensure operator-visible logging on `SafetyDecision.UNAVAILABLE` results | trivial | Surfaces degraded-safety state to operators (per D-14 Watch-Out item) |
+| **13** | Low | When sticky session routing is needed, add `ReplicaRegistry.select(routing_key: str | None = None)` and document where `session_id` lives (WS URL/header preferred over first JSON frame) | M | Forward-compat from D-15 — keeps the door open without buffering/peeking |
+| **14** | Low | At higher load, add `_bridge_session()` max-size + timeout limits OR recommend Envoy/HAProxy in front | S-M | The libraries' basic backpressure suffices for MVP; document the limit per D-15 |
+| **15** | Low | If active-active multi-primary becomes a requirement, define behavior (round-robin within healthy primaries, weighted, sticky-by-key) | M | Currently `RouterConfig.__post_init__` rejects multi-primary; D-15 deferred until evidence |
+| **~~Source-doc disposition~~** | ~~Med~~ | ~~Disposition of 7 untracked source docs~~ | ~~trivial~~ | ✅ **Resolved 2026-05-03** — moved into [source-archive/](source-archive/) |
+| **~~9~~** | ~~Low~~ | ~~Commit-message cleanup: PR 8's 3 commits still have `[8/n] Improve API:` prefix~~ | ~~S~~ | ✅ **Resolved 2026-05-04** — bundled into the will/api_7.8 prep rebase. PR 8's 3 commits now read `[type] streaming: ...` |
+| **~~10~~** | ~~Low~~ | ~~Commit-message cleanup: PR 7.8/7.9 commits have `streaming: streaming X` duplication~~ | ~~S~~ | ✅ **Resolved 2026-05-04** — bundled into the will/api_7.8 prep rebase. 3 commits dedup'd. |
+
+---
+
+## High priority
+
+### D-8: Verify `ltx2_image_crf` typed flow post-`d80c2a8`
+
+**Why:** Apr 26 dreamverse_review documented this field getting silently
+dropped by the public `SamplingParam`. May 2 `d80c2a8` (Dreamverse)
+refactored to typed `GeneratorConfig` + `preset_overrides`. Whether
+`image_crf` now flows through `request.stage_overrides.refine.image_crf`
+(per [design.md](design.md) mapping) or is still dropped is unverified.
+
+**Action:**
+1. Read [`Dreamverse/server/video_generation.py`](file:///home/william5lin/Dreamverse/server/video_generation.py)
+   post-`d80c2a8` for `image_crf` handling
+2. Trace through to FastVideo's `request.stage_overrides.refine.image_crf`
+3. Confirm runtime consumption in [`fastvideo/pipelines/basic/ltx2/`](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/)
+
+**Effort:** 10 min, no code changes.
+
+**Outcome:** Either confirms working OR identifies bug → opens fix item.
+
+### Item #1: Migrate `/healthz`+`/readyz`+`/status` into `build_app`
+
+**Why:** Today
+[`fastvideo.entrypoints.streaming.server.build_app`](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/server.py)
+exposes only `/health` + `/v1/stream`. Dreamverse FE expects all of
+`/healthz`, `/readyz`, `/status`, `/curated-presets`,
+`/prompt-system-config`, devtools.
+
+The streaming-server-upstream plan (line 84) explicitly lists
+`/healthz`+`/readyz`+`/status` as part of the contract that the upstream
+of `realtime/` → `streaming/` must preserve. They were deferred from
+PR 7.5's MVP. `/curated-presets` and `/prompt-system-config` are
+operator-side and stay in Dreamverse (FE feature-detects).
+
+**Action:**
+1. Read PR 7.5 (#1251) `build_app` to scope what's there
+2. Read [`Dreamverse/server/routes/health.py`](file:///home/william5lin/Dreamverse/server/routes/health.py)
+   for the route shapes Dreamverse already consumes
+3. Propose route migration as commit on top of `will/api_7.5` or as
+   part of PR 7.10 cycle
+4. Land
+
+**Effort:** Medium-Large (route shapes need preservation; tests).
+
+**Dependencies:** None blocking; can land anytime.
+
+**Files likely to touch:**
+- `fastvideo/entrypoints/streaming/server.py::build_app`
+- New `fastvideo/entrypoints/streaming/health.py`
+- Tests in `fastvideo/tests/entrypoints/streaming/`
+
+### Item #2: AbsMaxFP8 pre-existing test failure
+
+**Why:** [`fastvideo/tests/ops/quantization/test_absmax_fp8.py::test_create_weights_rejects_invalid_dtype`](file:///home/william5lin/FastVideo/fastvideo/tests/ops/quantization/test_absmax_fp8.py)
+fails with `AssertionError not raised`. Pre-existing on `main`; verified
+NOT introduced by NVFP4 work via `git stash`.
+
+**Action:**
+1. `git log --oneline fastvideo/tests/ops/quantization/test_absmax_fp8.py`
+   to find when it last passed
+2. Either:
+   - Restore the assert in `AbsMaxFP8LinearMethod.create_weights` if
+     intentional behavior was lost
+   - Drop the test if assert is no longer correct
+3. Verify
+
+**Effort:** Small.
+
+**Dependencies:** None.
+
+### Item VPO: `video_position_offset_sec` semantics
+
+**Why:** Per
+[`fastvideo/pipelines/basic/ltx2/continuation.py`](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/continuation.py),
+`LTX2ContinuationState.video_position_offset_sec` exists as a state
+field. Two valid interpretations:
+
+- **(a) Persistent across segments** — accumulating time offset for long
+  sessions; useful for time-coherent audio chaining.
+- **(b) Per-segment hint that rides on the carrier** — runtime
+  overwrites every time; field is harmless redundancy.
+
+Dreamverse computes `prefix_sec = float(audio_extra) / 24.0` per segment
+in `apply_audio` and currently does NOT persist it on
+`ContinuationState`. Field's docstring leans toward (b).
+
+**Decision deadline:** before PR 7.6 starts emitting/consuming the
+field (PR 7.6 branch is ready, not yet PR'd).
+
+**Action:**
+1. Confirm field's intended semantics with audio team
+2. If (a): document the accumulation rule explicitly + add tests
+3. If (b): leave docstring as-is + add test confirming overwrite
+
+**Effort:** 30 min discussion + small implementation.
+
+### Item D: Implement `generate_async` (PR 7.10)
+
+**Why:** Highest leverage. Closes:
+
+- D-5 / Q-5: audio re-encode for cross-segment continuity
+- Q-9: Dynamo progress passthrough (deferred)
+- PR 7.5's mid-segment cancellation TODO
+- Unblocks Dynamo native backend integration
+- **D-12-B**: enables `GpuPool.run() -> run_async() -> AsyncIterator[VideoEvent]` migration
+
+**Action:** See [streaming-server.md](streaming-server.md) "PR 7.10 — the
+unlock PR" section for scoping.
+
+**Effort:** Large.
+
+**Dependencies:** Best after PR 7.6 lands (gpu_pool upstream).
+
+**Files:**
+- `fastvideo/entrypoints/video_generator.py` — add `generate_async`,
+  refactor `generate_video` as wrapper
+- `fastvideo/api/results.py` — add `VideoEvent`/`VideoProgressEvent`/
+  `VideoPartialEvent`/`VideoFinalEvent`
+- `fastvideo/entrypoints/streaming/server.py` — consume `generate_async`,
+  remove TODO markers
+- `fastvideo/entrypoints/streaming/gpu_pool.py` — add `run_async()`
+  forwarding events from worker to caller
+- New `fastvideo/tests/entrypoints/test_generate_async.py`
+- New `fastvideo/tests/contract/test_dynamo_shape.py` (already in PR 8)
+
+### Item DR-1: Dreamverse — replace local `prompt_enhancer.py` with public + compat shim
+
+**Why:** Today Dreamverse carries `Dreamverse/server/prompt_enhancer.py`
+(1933 LOC) — a local copy/derivative of the FastVideo-internal version.
+After PR #1258 merges, Dreamverse should switch to the public
+`fastvideo.entrypoints.streaming.prompt.PromptEnhancer` and delete most
+of the local module.
+
+**Migration shape:**
+
+1. **Create** `Dreamverse/server/prompting/_internal_compat.py` (~150-200 LOC):
+   - Wraps public `PromptEnhancer.enhance()` → returns `EnhanceResult` shape Dreamverse expects
+   - Wraps public `PromptEnhancer.auto_extend()` — JSON-parses `LLMResponse.content` into `{"next_prompt": "..."}`
+   - Wraps public `PromptEnhancer.rewrite()` — JSON-parses into `{"segment_prompts": [...]}` with lenient fallback for malformed JSON
+   - Layers locked-segment + rollout_id + rollout_label metadata back on top
+2. **Update** `Dreamverse/server/runtime.py + main.py` — replace `from prompt_enhancer import PromptEnhancer` with `from prompting._internal_compat import PromptEnhancer`
+3. **Delete most of** `Dreamverse/server/prompt_enhancer.py` (1933 LOC). Keep only the bits that don't have a public equivalent:
+   - Race-based parallel fallback (`_run_provider_race`) — Dreamverse-specific tail-latency optimization
+   - `cerebras_ifm` provider — pending DR-2 decision
+   - Multi-classifier prompt safety (NSFW + hate-speech chained) — public ships single classifier
+4. **Tests** — verify Dreamverse session controllers still see the expected response shapes through the shim
+
+**Effort:** Medium (~150-200 LOC shim + replace upstream wiring + delete 1700+ LOC local module + test fixture updates).
+
+**Dependencies:**
+- PR #1258 must merge first (publishes `fastvideo.entrypoints.streaming.prompt.*`)
+- DR-2 informs the cerebras_ifm path
+
+**Files:**
+- New: `Dreamverse/server/prompting/_internal_compat.py`
+- Modified: `Dreamverse/server/runtime.py`, `Dreamverse/server/main.py`
+- Mostly deleted: `Dreamverse/server/prompt_enhancer.py`
+
+---
+
+## Medium priority
+
+### Item DR-2: Decide `cerebras_ifm` provider path
+
+**Why:** Public PR #1258's `PromptEnhancerConfig.provider` is
+`Literal["cerebras", "groq"]`. Internal supports `"cerebras_ifm"` (the
+Cerebras IFM API endpoint with different auth). Dreamverse needs
+`cerebras_ifm` working post-migration.
+
+Two options:
+
+| Option | Approach | Pros | Cons |
+|---|---|---|---|
+| **(a) Public** | Add `"cerebras_ifm"` to public Literal + ship `CerebrasIFMProvider` in `fastvideo/entrypoints/streaming/prompt/providers/cerebras_ifm.py` | Discoverable; users with IFM access can use typed config | Adds ~50 LOC + Literal extension to public surface |
+| **(b) Dreamverse-side** | Implement `CerebrasIFMProvider` Dreamverse-side as a custom `LLMProvider`, register via `enhancer.register_provider(CerebrasIFMProvider())` | Zero public surface change; private endpoint stays private | Slightly more boilerplate Dreamverse-side; not surfaced to non-Dreamverse users |
+
+**Recommendation:** Option (b) is more contained. Option (a) is more
+discoverable. Default to (b) unless there's a third-party user who needs
+IFM access. The Dreamverse-side PR carrying DR-1 is the natural place to
+make this decision.
+
+**Effort:** Small (decision) + Small-Medium (implementation).
+
+**Dependencies:** DR-1 (compat shim creation).
+
+### Item #3: `cerebras_ifm` provider in public Literal
+
+**Why:** Same item as DR-2 from the public-side framing. If DR-2 picks
+option (a), this is the implementation. If DR-2 picks option (b), this
+item is closed without implementation.
+
+**Action:** See DR-2.
+
+**Effort:** S-M.
+
+### Item #4: Expose `layer_profile` on typed `engine.quantization`
+
+**Why:** Today `transformer_quant: "NVFP4"` always constructs
+`NVFP4Config()` with default `layer_profile="refine"`. Dreamverse
+dodges via `experimental["pipeline_config"]`.
+
+**Action:**
+1. Add `transformer_quant_layer_profile: str | None = None` to
+   `QuantizationConfig` in [`schema.py`](file:///home/william5lin/FastVideo/fastvideo/api/schema.py)
+2. Thread through [`compat.py`](file:///home/william5lin/FastVideo/fastvideo/api/compat.py)
+3. Update `_apply_transformer_quant` in
+   [`fastvideo_args.py`](file:///home/william5lin/FastVideo/fastvideo/fastvideo_args.py)
+   to pass profile
+4. Update Dreamverse to drop the `experimental["pipeline_config"]`
+   dodge in favor of typed knob
+5. Tests in [`test_typed_quant_flow.py`](file:///home/william5lin/FastVideo/fastvideo/tests/api/test_typed_quant_flow.py)
+
+**Effort:** Medium.
+
+**Files:** schema.py, compat.py, fastvideo_args.py, test_typed_quant_flow.py,
++ Dreamverse/server/video_generation.py.
+
+### Item #5: Typed `dit_config.quant_config` carrier
+
+**Why:** The `experimental["pipeline_config"]` escape hatch in
+Dreamverse should eventually become a typed field. Design TBD.
+
+**Action:** Heaviest design work. Should consult Oracle.
+
+**Effort:** Large design + Large implementation.
+
+**Dependencies:** #4 should land first; this is the "final form" of #4.
+
+### Item SBS: `SessionStore` / `BlobStore` lifecycle policy
+
+**Why:** PR 7's in-memory implementations have no eviction, no TTL, no
+automatic blob cleanup on state replacement. Documented as per-deployment
+policy decision.
+
+When PR 7.5/7.6 land the live consumer, who owns:
+
+- bounded session capacity (LRU? TTL? hard max?)
+- blob `drop()` chained when state is replaced
+- session expiry on websocket disconnect
+
+**Recommendation:** streaming server's session manager. Worth stating
+explicitly in PR 7.5's design.
+
+**Effort:** Medium design + small implementation.
+
+### Item D-12-A: Update `GpuPool` ABC docstring — mark experimental
+
+**Why:** Per D-12 in [decisions-log.md](decisions-log.md), `GpuPool`
+should be documented as "API may change post-PR-7.10; experimental /
+server-internal" to prevent accidental promotion of streaming-internal
+API to framework-level. PR #1257 merged without this caveat.
+
+**Action:** Edit
+[`fastvideo/entrypoints/streaming/gpu_pool.py`](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/gpu_pool.py)
+class docstring on `GpuPool` ABC. Add a note: "API may change post-PR-7.10
+when run_async() lands; treat as server-internal for now."
+
+**Effort:** Trivial.
+
+**Dependencies:** None.
+
+### Item D-12-B: Replace `GpuPool.run() -> Any` with `run_async() -> AsyncIterator[VideoEvent]`
+
+**Why:** Per D-12, this is the canonical evolution post-PR-7.10. Closes
+the streaming server's cancellation TODO and converges the streaming +
+OpenAI + Dynamo consumers on a single async API.
+
+**Action:** As part of PR 7.10 cycle:
+1. Add `GpuPool.run_async(session_id, request) -> AsyncIterator[VideoEvent]`
+2. Worker forwards events through `result_queue` with type discriminator
+3. Streaming server replaces `await pool.run(...)` with `async for event in pool.run_async(...)`
+4. Sync `run()` becomes a thin compat wrapper that collects events and returns the final
+5. Cancellation propagates: client disconnect → `asyncio.CancelledError` → worker stops mid-step
+
+**Effort:** Medium. Adds ~50-100 LOC + tests.
+
+**Dependencies:** Item D (PR 7.10 — `generate_async` on `VideoGenerator`).
+
+### Item D-13-A: Document `streaming/prompt/*` as streaming-scoped
+
+**Why:** Per D-13 in [decisions-log.md](decisions-log.md), the prompt
+enhancer is currently scoped to streaming-server use even though the
+abstraction is general. Phrase user-facing docs as "streaming-server
+prompt enhancement" to keep future move to `fastvideo.prompt.*` cheap.
+
+**Action:** When PR 12 (docs migration) is written, the prompt enhancer
+section should:
+- Be titled "Streaming Server Prompt Enhancement", not "Prompt API"
+- Note the 3 fixed operations (`enhance` / `auto_extend` / `rewrite`) are
+  shaped by LTX-2 streaming session needs
+- Note that consumers wanting custom prompt operations can use
+  `provider.complete()` directly with their own LLMRequest
+- Avoid `from fastvideo import LLMProvider` exports until a second
+  consumer exists
+
+**Effort:** Trivial (docs only).
+
+**Dependencies:** PR 12 (docs migration).
+
+---
+
+## Low priority
+
+### Item D-13-B: Optional `client_factory` parameter for `httpx.AsyncClient` pooling
+
+**Why:** Today `_openai_compat.py` instantiates `httpx.AsyncClient` per
+call (no connection pooling). Reviewer flagged inefficient. Team chose
+simplicity for the expected scale (~6-10 enhancer calls per LTX-2
+session). If real-world metrics show connect/TLS overhead is meaningful,
+add an optional `client_factory: Callable[[], httpx.AsyncClient] | None`
+parameter to providers so they can share a pool.
+
+**Action:** Only when metrics justify. Add `client_factory=None` parameter
+to `CerebrasProvider` / `GroqProvider` constructors and pass through to
+`complete_openai_compatible()`. Default to current per-call behavior.
+
+**Effort:** Small.
+
+**Dependencies:** None blocking; only act on real perf data.
+
+### Item D-12-C: Avoid locking `PoolAssignment.gpu_id: int` as public
+
+**Why:** Today `PoolAssignment` exposes `gpu_id: int`, assuming
+one-GPU-per-worker. Future topology-aware pooling may need
+`device_ids: list[int]` (one worker = group of GPUs running internal
+`MultiprocExecutor`). Don't freeze the int field as public API.
+
+**Action:**
+- Treat `gpu_id` as a current-impl detail; prefer `worker_id` (already
+  exists, is stable identifier)
+- When a worker actually spans multiple GPUs, add
+  `PoolAssignment.device_ids: list[int]` and let `gpu_id` be `device_ids[0]`
+  for backward compat
+- Or rename to `gpu_id` → `device_id` with deprecation alias
+
+**Effort:** Small (1 field rename + alias).
+
+**Dependencies:** Driven by an actual future "one worker = many GPUs" use case. Don't act preemptively.
+
+### Item #6: Audio attention quantization profile
+
+**Why:** Today audio attn and FFN are bf16. If an audio-quant profile
+is added to `NVFP4Config.fp4_layers`, update
+[`test_basic_av_block_propagates_quant_config_to_all_children`](file:///home/william5lin/FastVideo/fastvideo/tests/ops/quantization/test_nvfp4_ltx2_wiring.py).
+
+**Effort:** Small (one test + one config field).
+
+### Item #7: Schema parity inventory cleanup
+
+**Why:** A few internal-only fields are not exposed publicly:
+
+- `PROMPT_HTTP_TIMEOUT_MS`
+- `PROMPT_INITIAL_STAGE_TIMEOUT_MS`
+- `PROMPT_TEMPERATURE`
+- `PROMPT_MAX_COMPLETION_TOKENS`
+- `PROMPT_AUTO_SLEEP_MS`
+- `PROMPT_AUTO_TIMEOUT_MS`
+- curated-presets file paths
+
+These flow via env vars on `dreamverse-server` today. If
+`fastvideo serve --config` becomes the canonical entrypoint, they need
+typed homes.
+
+**Effort:** Small-Medium.
+
+### Item #8: Stale `apps/web/test-results/` directory
+
+**Why:** Cosmetic. `.gitignore` entry hides it from `git status`, but
+the dir has stale `.last-run.json` (45 bytes) from a prior Playwright
+run.
+
+**Action:** `rm -rf apps/web/test-results` whenever convenient.
+
+**Effort:** Trivial.
+
+### ~~Item #9~~ + ~~#10~~: Commit-message cleanups — ✅ Resolved 2026-05-04
+
+Both items resolved during the `will/api_7.8` prep rebase. A targeted
+conditional script (`/tmp/opencode/cleanup_subjects_v2.sh` — only amends
+when text actually changes) ran across 33 commits, modified 6:
+
+- **#9 fix**: extended the regex from `\[\d+\.\d+/n\]` to
+  `\[\d+(\.\d+)?/n\]` so single-digit prefixes match. PR 8's 3 commits
+  now read `[type] streaming: ...` instead of `[type] [8/n] Improve API: ...`.
+- **#10 fix**: added second substitution `streaming: streaming X` →
+  `streaming: X`. PR 7.8 / 7.9 commits no longer have the duplication.
+
+The conditional check skipped pre-commit-hook flakiness on no-op amends
+(unlike the earlier first attempt). All affected commits verified clean
+post-rebase.
+
+### Item #11: Promote LTX-2 prompt orchestration to public (when 2nd consumer exists)
+
+**Why:** Per Q-2 in [decisions-log.md](decisions-log.md) and D-13's
+"missing alternative", the LTX-2-specific orchestration (locked
+segments, segment_prompts JSON shape, rollout id/label, lenient JSON
+parsing) currently stays Dreamverse-side per DR-1. If a second
+LTX-2-style consumer appears (e.g. another video model with multi-segment
+continuation needing the same prompt orchestration), promote this layer
+to `fastvideo.entrypoints.streaming.prompt.ltx2_orchestration`.
+
+**Action:** Wait for a second consumer to materialize. Until then, the
+orchestration stays in Dreamverse's `_internal_compat.py` shim (DR-1).
+
+**Effort:** Medium when triggered.
+
+**Dependencies:** A second consumer.
+
+---
+
+## Recommended pull order
+
+If you have unbounded time and want to maximize forward progress:
+
+1. **D-8 verify** (10 min) — eliminates uncertainty
+2. **D-12-A docstring** (trivial) — caveat the GpuPool API publicly
+3. **Item #2 AbsMaxFP8** (S) — clears tech debt
+4. **Item VPO video_position_offset_sec** (30 min) — unblocks PR 7.10 (since 7.6 has merged, this is now scoped to whatever consumer first reads the field)
+5. **DR-1 + DR-2 Dreamverse migration** (M) — **now unblocked since PR #1258 merged**; replaces 1700+ LOC of local fork
+6. **Item #4 layer_profile** (M) — closes Dreamverse quant escape hatch
+7. **Item #1 build_app routes** (M-L) — closes FE-compat
+8. **Item D generate_async** (L) — unlock PR; brings along D-12-B (run_async) + closes Q-5/Q-9/PR-7.5 TODOs
+9. **Item #5 typed quant_config carrier** (L+L) — final form
+10. **Items #6/#7/#8 + D-12-C/D-13-A/D-13-B + #11** — cleanup polish (#9, #10 resolved 2026-05-04)
+
+If you have a specific user goal (e.g. "ship `BE_FLAVOR=fastvideo`
+flavor end-to-end"), that goal dictates the order — read this list as a
+menu, not a prescription.
+
+---
+
+## Verification gates per item
+
+When implementing any item above, evidence required:
+
+| Phase | Check |
+|---|---|
+| Build | `lsp_diagnostics` clean on changed files |
+| Test | new + relevant existing tests pass; output captured |
+| Manual QA | actually run the affected feature end-to-end (per AGENTS.md MANUAL_QA_MANDATE) |
+| Regression | full `fastvideo/tests/api/` + `contract/` + relevant SSIM (if NVFP4 touch) |
+
+For NVFP4 touches: re-run `test_nvfp4_ltx2_wiring.py` +
+`test_typed_quant_flow.py` (CPU) + ideally a flashinfer-enabled path
+test (manual, not in CI).
+
+For Dreamverse-side items (DR-1, DR-2): re-run
+`Dreamverse/apps/web/npx playwright test e2e/preset-prompt-generation.spec.ts`
+end-to-end against the live BE+FE — this is the contract test that
+exercises the prompt enhancer through a real session.

.agents/memory/dreamverse-integration/pr-roadmap.md

@@ -0,0 +1,141 @@
+# PR Roadmap
+
+Status of all 17 PRs in the FastVideo public API refactor + streaming
+server upstream + Dynamo backend contract + post-deprecation cleanup.
+
+For design rationale see [design.md](design.md). For streaming-specific
+PRs (7.5-7.10) see [streaming-server.md](streaming-server.md). For NVFP4
+work that runs parallel to this sequence see [quantization.md](quantization.md).
+
+**Last updated:** 2026-05-05 (strategy reversal — single mega-PR #1288 replaces planned splits 7.10/8/LTX-2/NVFP4/post-fixes/agents_cleanup; see [decisions-log.md D-17](decisions-log.md#d-17)).
+
+## Status legend
+
+- ✅ **Landed on `origin/main`**
+- 🟢 **Open / in flight** — branch exists, may have open PR
+- 🟡 **Planned** — designed, not started
+- 🔵 **Future** — deferred to post-PR-13 cleanup
+
+## Landed PRs (0 → 7.7)
+
+| # | PR | Status | Merge commit | Scope |
+|---|---|---|---|---|
+| 0 | #1218 [1/n] | ✅ | merged | Parity inventory + typed inference schema |
+| 1 | #1218 [1/n] | ✅ | merged | Strict parser/validation/overrides + API tests |
+| 2 | #1220 [2/n] | ✅ | merged | Typed `VideoGenerator` constructors + request path + compat |
+| 3 | #1226 [3/n] | ✅ | merged | CLI/YAML-first typed config loading for `generate` and `serve` |
+| 4 | #1234 [4/n] | ✅ | merged | Preset registry + presets for all 13 model families; `SamplingParam` moved to `fastvideo/api/`; `configs/sample/` deleted entirely |
+| 5 | #1237 [5/n] | ✅ | merged | `ServeConfig.default_request` wired into stateless OpenAI server |
+| 5.5 | (`5d1d71fc`) | ✅ | merged | Streaming server package skeleton, typed `StreamingConfig`/`GpuPoolConfig`/`PromptEnhancerConfig`/`PromptSafetyConfig`/`WarmupConfig`, `streaming-serve` CLI stub |
+| 6 | #1239 [6/n] | ✅ | merged | LTX2 public preset + asset wiring + `gpu_pool.py` typed-kwarg translation |
+| 7 | #1250 [7/n] | ✅ | merged | Typed LTX2 continuation state + streaming session store + blob store |
+| **7.5** | **#1251** | ✅ | `95fd29e0` (merged 2026-04-26) | Streaming server skeleton (WebSocket + fMP4 + single generator). 8 commits. Deferred TODOs (per-step progress, mid-segment cancellation) carried forward to PR 7.10. |
+| **7.6** | **#1257** | ✅ | `eb0a4152` (merged 2026-05-04) | GPU pool upstream + worker subprocess + two-segment warmup. 7 commits squashed. APPROVED by Eigensystem. See [decisions-log.md D-12](decisions-log.md#d-12) for the architectural review. |
+| **7.7** | **#1258** | ✅ | `f673423b` (merged 2026-05-04) | Prompt enhancer with `LLMProvider` abstraction. Built-in providers: cerebras, groq. 3 commits squashed. **Public Literal does NOT include `cerebras_ifm`** — open-threads.md item DR-2 covers the gap. See [decisions-log.md D-13](decisions-log.md#d-13) for the architectural review. |
+| **7.8** | **#1284** | ✅ | `eb3a3942` (merged 2026-05-04) | Streaming auxiliaries — `prompt/safety.py` (optional fasttext, lazy import), `prompt/rewrite.py`, `session_logger.py` (thread-safe JSONL), `mock_server.py` (build_mock_app + MockGenerator for FE dev). 730 LOC, 2 commits. See [decisions-log.md D-14](decisions-log.md#d-14). |
+| **7.9** | **#1286** | ✅ | `2aaeee2a` (merged 2026-05-05) | Streaming router (multi-replica load balancer + WS proxy + `fastvideo router-serve` CLI). Squashed `cd76cf51 + 1ac1e732 + b0b7f59c + a152cb77` (router-polish second-pass; cherry-pick of `40e265b8` from `will/ltx2_sr_port`). See [decisions-log.md D-15](decisions-log.md#d-15) (structural review) + [D-16](decisions-log.md#d-16) (second-pass polish). |
+
+## In flight (mega-PR #1288)
+
+| # | PR | Status | Branch | Scope |
+|---|---|---|---|---|
+| **mega** | **#1288** | 🟢 OPEN, MERGEABLE | `will/ltx2_sr_port` (head `b36bdbc9`) | **Single consolidated landing of the full `will/ltx2_sr_port` chain.** Was originally planned as 6 stacked PRs (slices 1-3 / 4-6 / 7-15 / 16-21 / 22-23 / 24-34). Now landing as one PR — see [decisions-log.md D-17](decisions-log.md#d-17) for the strategy decision. **Contents** (commit-ordered): (1) streaming `generate_async` + `VideoEvent` + Dynamo backend contract (3 commits, was PR 7.10/#1287 closed); (2) server contract docs + Dreamverse/Dynamo shape tests (3 commits, was PR 8); (3) LTX-2 SR runtime port + i2v conditioning + alignment harness (9 commits); (4) NVFP4 wire-up + per-component compile + typed `transformer_quant` flow (6 commits); (5) LTX-2 post-handoff parity fixes — Gemma `to()`, list-of-generators (2 commits); (6) `.agents/memory/dreamverse-integration/` knowledge base + agents Phase 1 cleanup (11 commits). 34 commits total, 71 files, +13,074/-583 LOC. |
+
+## Closed PRs in this scope
+
+| # | PR | Status | Why closed |
+|---|---|---|---|
+| **7.10** | **#1287** | ❌ CLOSED 2026-05-05 | Superseded by mega-PR #1288 — strategy reversal to land everything in one go. Same 3 commits now form the head of #1288. |
+
+## Deprecated split bookmarks (D-17)
+
+`will/api_7.10` / `will/api_8` / `will/ltx2_sr_runtime` / `will/ltx2_nvfp4` / `will/ltx2_post_fixes` / `will/agents_cleanup` were the split-PR bookmarks under the abandoned 6-PR plan. They remain locally as historical references but are no longer maintained. STACK.md (top-level) is similarly deprecated.
+
+## Planned (post-#1288 merge)
+
+| # | Status | Branch | Scope |
+|---|---|---|---|
+| 9 | 🟡 | — | LongCat preset migration + colocation (9 model-specific stage files) |
+| 10 | 🟡 | — | Hunyuan15 SR preset migration + colocation + SR field migration POC |
+| 11 | 🟡 | — | SSIM/performance test migration off legacy `generate_video(..., **kwargs)` |
+| 12 | 🟡 | — | Docs + examples migration (includes streaming server + Dynamo) |
+| 13 | 🟡 | — | Deprecation cleanup (includes flat LTX2 kwargs the internal `gpu_pool.py` used to consume) |
+
+## Future (compat.py death sequence)
+
+After PR 13 lands deprecation warnings, `fastvideo/api/compat.py` (~370
+lines) is the last translation shim between typed public API and legacy
+internals (`FastVideoArgs`, `SamplingParam`).
+
+| # | Status | Scope | Lines removed |
+|---|---|---|---|
+| 14 | 🔵 reachable | Strip forward translation: `legacy_from_pretrained_to_config`, `legacy_generate_call_to_request`, `_sampling_param_to_request_raw`, `_LEGACY_REQUEST_ALIASES`, `_LTX2_REFINE_FLAT_KEYS`. Depends on PRs 11/12/7.6 callers being migrated. | ~100 |
+| 15 | 🔵 | `FastVideoArgs` becomes a `@dataclass` view over `GeneratorConfig` with `@property` accessors backing legacy field names. ~600-line god-object refactor. Depends on PR 14. | reverse-translation half (~150) trivial |
+| 16 | 🔵 | `ForwardBatch` reads `GenerationRequest` by reference; kills `request_to_sampling_param` and the `ForwardBatch(**shallow_asdict(sampling_param), …)` spread. `SamplingParam` demoted or deleted. Depends on PR 15. | rest |
+| 17 | 🔵 | Move `normalize_generator_config`, `normalize_generation_request`, `load_generator_config_from_file` to `parser.py`. Delete `compat.py`. | file gone |
+
+PRs 15-17 touch training, distributed, and worker code in addition to
+inference path; realistically 1-2 quarters beyond the current plan.
+
+## Dependency chain
+
+```
+PR 13 (deprecation)
+  ↓
+PRs 11, 12, 7.6 (migrate callers)
+  ↓
+PR 14 (forward translation gone)      ─── ~100 lines out of compat.py
+  ↓
+PR 15 (FastVideoArgs as view)         ─── reverse-translation trivial
+  ↓
+PR 16 (ForwardBatch reads request)    ─── SamplingParam demoted
+  ↓
+PR 17 (move normalizers, delete file)
+```
+
+## NVFP4 work (out-of-band, parallel to PR 7.5+)
+
+NOT in the canonical PR sequence. Lives on `will/ltx2_sr_port`
+(currently @ `156103b9`) — a separate stack alongside the public-API
+upstreaming. See [quantization.md](quantization.md) for what each commit
+locks in.
+
+| Commit range | Topic |
+|---|---|
+| `cfccd292..b6ac7630` | LTX-2 i2v + SR runtime port + alignment harness |
+| `a4760bae..c6c14c55` | NVFP4 LTX-2 wire-up + per-component compile + parity fixes (May 2 handoff) |
+| `a5fcd19c..156103b9` | Post-handoff parity/perf fixes |
+
+## Key landed artifacts (reference points)
+
+- Parity inventory: [`docs/design/inference_schema_parity_inventory.yaml`](file:///home/william5lin/FastVideo/docs/design/inference_schema_parity_inventory.yaml) + guard [`fastvideo/tests/api/test_schema_parity_inventory.py`](file:///home/william5lin/FastVideo/fastvideo/tests/api/test_schema_parity_inventory.py)
+- Typed schema: [`fastvideo/api/schema.py`](file:///home/william5lin/FastVideo/fastvideo/api/schema.py)
+- Compat layer: [`fastvideo/api/compat.py`](file:///home/william5lin/FastVideo/fastvideo/api/compat.py)
+- Preset system: [`fastvideo/api/presets.py`](file:///home/william5lin/FastVideo/fastvideo/api/presets.py) + per-family `pipelines/basic/<family>/presets.py`
+- Streaming package skeleton (PR 5.5): [`fastvideo/entrypoints/streaming/`](file:///home/william5lin/FastVideo/fastvideo/entrypoints/streaming/)
+- LTX2 typed continuation state (PR 7): [`fastvideo/pipelines/basic/ltx2/continuation.py`](file:///home/william5lin/FastVideo/fastvideo/pipelines/basic/ltx2/continuation.py)
+
+## Known notable decisions carried forward
+
+- **Public inference boundary stays plain dataclasses + plain dict/YAML/JSON**
+  — not OmegaConf, not runtime config wrappers.
+- **Every public entrypoint normalizes into typed config objects** before
+  touching legacy `FastVideoArgs` or `SamplingParam`.
+- **Legacy `generate_video(..., **kwargs)` stays on direct legacy execution
+  path until PR 11**'s SSIM/performance migration. Prevents golden
+  baselines from drifting during compat period.
+- **Typed requests use schema defaults**; legacy `generate_video(...)`
+  continues to inherit model-specific `SamplingParam` defaults during
+  compat period.
+- **Preset registry uses explicit `_register_presets()` pattern** matching
+  `_register_configs()`; lookup keyed by `model_family`.
+- **Stateless OpenAI server clones `ServeConfig.default_request`** and
+  merges user overrides; preset validation runs before legacy generation.
+- **Streaming server added as sibling `fastvideo/entrypoints/streaming/`**
+  rather than extending `fastvideo/entrypoints/openai/` (PR 5.5).
+
+## Per-PR commit-level detail
+
+For per-PR commit lists, test plans, and merge criteria, the archived
+source [`source-archive/PR-plan.md`](source-archive/PR-plan.md) (1145 lines)
+remains the deepest reference. This file is the navigable summary.

.agents/memory/dreamverse-integration/quantization.md

@@ -0,0 +1,229 @@
+# Quantization — NVFP4, LinearBase Fallback, Layer Profiles
+
+What landed in the May 2 NVFP4 stack, why it's load-bearing, and what's
+still owed (`layer_profile`, typed quant carrier, AbsMaxFP8 cleanup).
+
+For overall API design see [design.md](design.md). For the open
+follow-ups see [open-threads.md](open-threads.md).
+
+**Last updated:** 2026-05-03.
+
+## NVFP4 — what it is
+
+NVIDIA's specific block-scaled FP4 format:
+
+- e2m1 mantissa
+- fp32 alpha
+- `layout_128x4` scale layout
+- group size 16
+
+Distinct from MX-FP4 / OCP-FP4 / generic e3m0. The May 2 rename
+(`94c983a2`) disambiguated the naming throughout FastVideo's public
+surface.
+
+## Files (current)
+
+| File | Role |
+|---|---|
+| [`fastvideo/layers/quantization/nvfp4_config.py`](file:///home/william5lin/FastVideo/fastvideo/layers/quantization/nvfp4_config.py) | `NVFP4Config`, `NVFP4QuantizeMethod`, `convert_model_to_nvfp4` |
+| [`fastvideo/layers/quantization/__init__.py`](file:///home/william5lin/FastVideo/fastvideo/layers/quantization/__init__.py) | `QuantizationMethods` literal includes `"NVFP4"`; `get_quantization_config` resolves it |
+| [`fastvideo/layers/linear.py`](file:///home/william5lin/FastVideo/fastvideo/layers/linear.py) | `LinearBase.__init__` falls back to `UnquantizedLinearMethod` when `quant_config.get_quant_method` returns None — **load-bearing** |
+| [`fastvideo/models/loader/fsdp_load.py`](file:///home/william5lin/FastVideo/fastvideo/models/loader/fsdp_load.py) | `_maybe_convert_model_to_nvfp4` helper detects via `isinstance(quant_method, NVFP4QuantizeMethod)`; calls `convert_model_to_nvfp4` to materialize buffers |
+| [`fastvideo/models/dits/ltx2.py`](file:///home/william5lin/FastVideo/fastvideo/models/dits/ltx2.py) | `nn.Linear` → `ReplicatedLinear` for FP4-eligible subset; `_supports_prequantized_input` + `_linear_project_with_optional_prequant` helpers; quant_config + prefix= plumbing |
+| [`fastvideo/api/compat.py`](file:///home/william5lin/FastVideo/fastvideo/api/compat.py) | Typed `engine.quantization.transformer_quant: "NVFP4"` resolves to `NVFP4Config()` instance |
+| [`fastvideo/fastvideo_args.py`](file:///home/william5lin/FastVideo/fastvideo/fastvideo_args.py) | `__post_init__._apply_transformer_quant` pins `pipeline_config.dit_config.quant_config = NVFP4Config()` |
+
+## Buffer naming (post-rename)
+
+| Old | New |
+|---|---|
+| `_fp4_weight` / `_fp4_alpha` | `_nvfp4_weight` / `_nvfp4_alpha` |
+| `_weight_global_sf` | unchanged |
+| `convert_model_to_fp4` | `convert_model_to_nvfp4` |
+| `FP4QuantizeMethod` | `NVFP4QuantizeMethod` |
+| `QuantizationMethods` literal `"FP4"` | `"NVFP4"` |
+
+Internal-scope torch op namespace `fastvideo_fp4::*` and
+`_get_ltx2_fp4_stage_profile` deliberately left as-is — purely internal
+naming that mirrors FastVideo-internal.
+
+## Layer set asymmetry — by design
+
+`NVFP4Config.fp4_layers` (default `layer_profile="refine"`) covers:
+
+- `attn1.{to_q,to_k,to_v,to_out}` — full self-attention
+- `attn2.{to_q,to_out}` — cross-attn Q + out only (text context not quantized)
+- `audio_to_video_attn.{to_q,to_out}` — AV cross Q + out
+- `video_to_audio_attn.{to_k,to_v}` — VA cross K + V
+- `ffn.{fc_in,fc_out}` — video FFN
+- `adaln_single.linear` — but this is `nn.Linear` (not `LinearBase`),
+  so it never actually gets FP4'd. List entry has no effect; matches
+  internal.
+
+**NOT in the set:**
+
+- audio self-attention (`audio_attn1.*`)
+- audio cross-attention (`audio_attn2.*`)
+- audio FFN (`audio.ffn.*`)
+
+Audio path is cheap enough that quant overhead isn't worth it. Test
+[`test_basic_av_block_propagates_quant_config_to_all_children`](file:///home/william5lin/FastVideo/fastvideo/tests/ops/quantization/test_nvfp4_ltx2_wiring.py)
+locks this in — if you add audio quantization later, update the test.
+
+## `LinearBase` fallback — DO NOT REMOVE
+
+[`fastvideo/layers/linear.py:191-202`](file:///home/william5lin/FastVideo/fastvideo/layers/linear.py#L191-L202): when `quant_config.get_quant_method` returns
+`None` (layer not in the quant config's set), we fall back to
+`UnquantizedLinearMethod`.
+
+**Removing this fallback would break every non-tagged
+`ReplicatedLinear` constructed with an `NVFP4Config`** — the previous
+`assert quant_method is not None` would crash on unmatched layers (e.g.
+text-encoder K/V projections, audio attention, etc.).
+
+This is one of the load-bearing changes from `42b30bf9`.
+
+## `transformer_quant` precedence rules
+
+`FastVideoArgs._apply_transformer_quant` only writes
+`dit_config.quant_config` when it's currently `None`. **If a caller has
+explicitly set** `pipeline_config.dit_config.quant_config = NVFP4Config(...)`,
+the explicit setter wins.
+
+Dreamverse's `video_generation.py` relies on this precedence — it sets
+`NVFP4Config()` directly via `experimental["pipeline_config"]` because
+typed `transformer_quant: "NVFP4"` doesn't yet expose `layer_profile`.
+See "Open follow-ups" below.
+
+## Attention forward optimization
+
+[`models/dits/ltx2.py`](file:///home/william5lin/FastVideo/fastvideo/models/dits/ltx2.py)
+ports `_supports_prequantized_input` and
+`_linear_project_with_optional_prequant`. Attention forward
+pre-quantizes input once (`quantize_input`), reuses the
+`(x_fp4, x_scale, x_global_sf)` tuple for k/v projections when
+`context is x` — bit-matches internal's fused path.
+
+## `prepare_for_compile` protocol
+
+[`composed_pipeline_base._maybe_compile_pipeline_module`](file:///home/william5lin/FastVideo/fastvideo/pipelines/composed_pipeline_base.py)
+calls `getattr(module, "prepare_for_compile", None)` before invoking
+`torch.compile`. Defined as a duck-type protocol — no base class method.
+
+Currently only **Gemma3** implements it (to materialize HF weights
+outside Dynamo's tracer). Add to other models that have lazy external
+state if you observe compile-time graph breaks.
+
+## Per-component compile flags
+
+`CompileConfig` (in
+[`fastvideo/api/schema.py`](file:///home/william5lin/FastVideo/fastvideo/api/schema.py))
+gained per-component knobs in `221cb20a`:
+
+```python
+@dataclass
+class CompileConfig:
+    enabled: bool = False                 # master DiT switch
+    backend: str = "inductor"
+    fullgraph: bool = False
+    mode: str | None = None
+    dynamic: bool | None = None
+    extras: dict = field(default_factory=dict)
+
+    # Per-component overlays, None = inherit master `enabled`
+    text_encoder_enabled: bool | None = None
+    vae_enabled: bool | None = None
+    audio_vae_enabled: bool | None = None
+
+    # Per-component kwargs override master when non-empty
+    dit_kwargs: dict = field(default_factory=dict)
+    text_encoder_kwargs: dict = field(default_factory=dict)
+    vae_kwargs: dict = field(default_factory=dict)
+    audio_vae_kwargs: dict = field(default_factory=dict)
+```
+
+**`transformer_refine` is auto-compiled with the master DiT flag.** No
+separate `enable_torch_compile_refine` flag — by design, refine inherits
+DiT compile state to keep typed surface small. Decoupling would add a
+new flag, not repurpose existing ones.
+
+## Quantization commit chain (`will/ltx2_sr_port`)
+
+| Commit | Locks in |
+|---|---|
+| `365a66c7  feat(quantization): upstream LTX-2 FP4Config with lazy flashinfer` | Public colocation of FP4Config (resolves dreamverse_review Q-6 option 1); flashinfer lazy-imported in loader helper, no public hard-dep |
+| `a4760bae  fix(api): propagate generic refine_*` | `_resolve_refine_args()` copies generic `refine_*` knobs onto `ltx2_refine_*` runtime carriers; `_randn_ltx2_video_latents` reverts to `torch.randn` to bit-match internal under single-generator inference |
+| `221cb20a  feat(api): typed per-component CompileConfig` | `CompileConfig` per-component knobs; matching `FastVideoArgs` carriers; compat layer round-trip |
+| `6da342ba  feat(compile): per-component compile + transformer_refine + prepare hook` | `composed_pipeline_base.post_init` compiles `transformer_refine` alongside `transformer`/`transformer_2`; per-component compile loops; `prepare_for_compile` hook on Gemma3 |
+| `42b30bf9  feat(ltx2): wire FP4 inference` (largest) | `nn.Linear` → `ReplicatedLinear` for FP4-eligible LTX2 subset; `quant_config` + `prefix=` plumbing; `_maybe_convert_model_to_nvfp4` helper; `LinearBase` fallback to `UnquantizedLinearMethod`; typed `transformer_quant` resolution |
+| `94c983a2  refactor(quant): rename FP4 → NVFP4` | Mechanical rename across config, methods, buffers, tests |
+| `c6c14c55  test(nvfp4): lock LTX-2 wiring + typed transformer_quant flow` | 6+4 tests in `test_nvfp4_ltx2_wiring.py` + `test_typed_quant_flow.py` |
+| `a5fcd19c  [fix]: lazy-import flash_attn 2 fallback in attention backend` | post-handoff: lazy import to avoid hard flash_attn 2 dep |
+| `d4ee5be2  [fix]: avoid model.to() round-trip in Gemma encoder forward` | post-handoff: parity / perf fix |
+| `156103b9  [fix]: unwrap list-of-generator before torch.randn in LTX-2 latent prep` | post-handoff: parity fix for list-of-generators (was bit-matching only single-generator path) |
+
+## Tests
+
+| Test | Asserts |
+|---|---|
+| [`fastvideo/tests/ops/quantization/test_nvfp4_ltx2_wiring.py`](file:///home/william5lin/FastVideo/fastvideo/tests/ops/quantization/test_nvfp4_ltx2_wiring.py) (6 tests) | `LTXSelfAttention.to_q/to_k/to_v/to_out` are `ReplicatedLinear`; `NVFP4Config()` attaches `NVFP4QuantizeMethod` on the quantized subset with correct `layer_prefix`; non-tagged projections (cross-attn K/V, audio attn, audio FFN) fall back to `UnquantizedLinearMethod`; `BasicAVTransformerBlock` propagates `quant_config`+`prefix` correctly to all 4 attention modules + FFN |
+| [`fastvideo/tests/api/test_typed_quant_flow.py`](file:///home/william5lin/FastVideo/fastvideo/tests/api/test_typed_quant_flow.py) (4 tests) | typed `engine.quantization.transformer_quant: "NVFP4"` → `NVFP4Config()` instance flow; default leaves `transformer_quant` None; explicit `dit_config.quant_config = ...` wins over typed carrier |
+
+CPU-only by design; do NOT exercise actual FP4 kernels (no flashinfer in
+CI). Real kernel coverage requires a CI run with flashinfer installed.
+
+## Open follow-ups (quantization-specific)
+
+### #4: Expose `layer_profile` on typed `engine.quantization`
+
+Today `transformer_quant: "NVFP4"` always constructs `NVFP4Config()`
+with default `layer_profile="refine"`. To support stage-1 profiles (no
+`attn2.to_out`, no cross-modal AV) via typed config, add
+`transformer_quant_layer_profile: str | None = None` and thread it
+through:
+
+- `fastvideo/api/schema.py` — `QuantizationConfig` field
+- `fastvideo/api/compat.py` — typed → flat translation
+- `fastvideo/fastvideo_args.py` — `_apply_transformer_quant` consumes it
+
+Dreamverse currently dodges this by setting `NVFP4Config()` directly via
+`experimental["pipeline_config"]`. Exposing `layer_profile` removes the
+dodge. See [open-threads.md](open-threads.md) #4.
+
+### #5: Typed `dit_config.quant_config` carrier (replace `experimental["pipeline_config"]`)
+
+Long-term: design a typed home for an in-memory `PipelineConfig`
+instance with mutated `dit_config`. Today `compat.py` recognizes the
+`pipeline_config` key in `experimental` and threads it through to
+`FastVideoArgs.from_kwargs`. This is fine for short-term but not pretty.
+
+Heaviest design work in the open queue. May need Oracle consult.
+
+### #2: AbsMaxFP8 pre-existing test failure
+
+`fastvideo/tests/ops/quantization/test_absmax_fp8.py::test_create_weights_rejects_invalid_dtype`
+fails on `main` and on `will/ltx2_sr_port` with the same error
+(`AssertionError not raised`). Verified via `git stash` that the
+failure pre-dates NVFP4 work.
+
+Either:
+- Fix the test (`AbsMaxFP8LinearMethod.create_weights` no longer
+  asserts on invalid dtype — restore the assert if intentional, or drop
+  the test).
+
+Self-contained tech debt; small fix.
+
+## Don't / Cautions
+
+- **Don't change `NVFP4Config` buffer names back to `_fp4_*`.** Rename
+  is intentional to disambiguate from MX-FP4 / OCP-FP4.
+- **Don't remove the `LinearBase` `UnquantizedLinearMethod` fallback.**
+  Load-bearing for non-tagged layers when a `quant_config` is set.
+- **Don't repurpose `enable_torch_compile` to mean DiT-only.** It also
+  drives `transformer_refine` and `transformer_2` compile.
+- **Don't bypass the typed surface for new options.** New compile /
+  quant / refine knobs should land on the dataclass + compat.py +
+  parity inventory together. The existing test suite locks this in.
+- **Don't merge to main without a CI run that covers FP4.** Current CI
+  doesn't run flashinfer-dependent paths; the wiring tests are CPU-only
+  by design.

.agents/memory/dreamverse-integration/runbook.md

@@ -0,0 +1,385 @@
+# Runbook — How to Do Work in This Scope
+
+Operational how-to for the dreamverse-integration scope. Read after
+[state.md](state.md) and [open-threads.md](open-threads.md).
+
+For design rationale see [design.md](design.md). For who to credit see
+[authors.md](authors.md). For PR status see [pr-roadmap.md](pr-roadmap.md).
+
+**Last updated:** 2026-05-05 (strategy reversed to single mega-PR #1288 on `will/ltx2_sr_port`; #1287 closed; STACK.md split model deprecated per [decisions-log.md D-17](decisions-log.md#d-17)).
+
+## Worktree contract
+
+```
+Repo:   /home/william5lin/FastVideo
+Branch: will/ltx2_sr_port
+```
+
+Other agents and the user share this worktree concurrently. If `git status`
+shows changes you don't recognize, they belong to **someone else's work** —
+don't revert, don't `git stash drop`, don't `git checkout -- <file>`.
+Switch to `will/ltx2_sr_port` cleanly with `git checkout will/ltx2_sr_port`
+(safe if your own working tree is clean) and proceed.
+
+If your task requires a different branch (e.g. cherry-pick to
+`will/api_7.9` for PR #1286 propagation), return to `will/ltx2_sr_port`
+when done — that is the assumed default.
+
+## Branch topology (single mega-PR model)
+
+The dreamverse-integration work now ships as one PR (#1288) off
+`will/ltx2_sr_port`. The split-PR model documented in earlier revisions
+of this runbook (and in top-level `STACK.md`) is **abandoned** —
+see [decisions-log.md D-17](decisions-log.md#d-17).
+
+```
+origin/main
+  ↓ [public-API refactor: PRs 0..7.9 merged on main, latest #1286 = 2aaeee2a]
+will/ltx2_sr_port   (**PR #1288 head** — single mega-PR, 34 commits, 71 files, +13,074/-583)
+```
+
+| Branch | Role | Status |
+|---|---|---|
+| `will/ltx2_sr_port` | **PR #1288 head**, default working branch | OPEN, MERGEABLE |
+| `will/api_7.10` / `will/api_8` / `will/ltx2_sr_runtime` / `will/ltx2_nvfp4` / `will/ltx2_post_fixes` / `will/agents_cleanup` | deprecated split-PR bookmarks | local-only historical references; safe to delete |
+| `will/ltx2_sr_port-pre-1286-rebase` | safety backup | local-only; preserves the 4 commits dropped during the post-#1286 rebase |
+
+**Sanity check:** `git merge-base --is-ancestor origin/main will/ltx2_sr_port`
+should exit 0. If it doesn't, the branch is in an unexpected state — read
+[state.md](state.md) before continuing.
+
+## After PR #1288 merges
+
+When the mega-PR squash-merges into `main`:
+
+1. `git fetch origin main` to pull the merge commit.
+2. The entire `will/ltx2_sr_port` content is now on main; the branch can
+   be deleted (locally + on origin) once all consumers are notified.
+3. Delete deprecated split bookmarks: `git branch -D will/api_7.10
+   will/api_8 will/ltx2_sr_runtime will/ltx2_nvfp4 will/ltx2_post_fixes
+   will/agents_cleanup` (local-only, no remote).
+4. Optionally remove top-level `STACK.md` (now a historical artifact).
+   Keep `CO-AUTHORS.md` — still the canonical roster reference.
+5. Decide whether to keep `will/ltx2_sr_port-pre-1286-rebase` (safety
+   backup of the pre-rebase chain) — recommend deleting once #1288 is
+   merged and verified on main.
+6. Update memory dir to reflect the post-merge state — bump
+   `Last reconciled` headers, mark Item D resolved in
+   [open-threads.md](open-threads.md), record the merge commit in
+   [decisions-log.md](decisions-log.md).
+
+## Historical: split-PR re-slice protocol (deprecated)
+
+Prior revisions of this runbook documented a 10-step re-slice protocol
+for the abandoned 6-PR split model. That protocol is now obsolete.
+The post-#1286 rebase (2026-05-05) was the last execution of it; details
+are preserved in [state.md](state.md) "Post-#1286 rebase summary" and
+git history at commit `b34d9704`.
+
+## Verification
+
+### Lint (pre-commit)
+
+```bash
+pre-commit run --files <changed-paths...>
+```
+
+- Binary: `/home/william5lin/miniconda3/envs/fv-main/bin/pre-commit`.
+  NOT `.venv/bin/pre-commit` — that doesn't exist in this worktree.
+- Auto-applies yapf reformatting; re-stage modified files after.
+- Hook chain: yapf → ruff → codespell → mypy → spaces-check.
+- Memory dir (`.agents/memory/`) is yapf/ruff/mypy excluded — only
+  "spaces" runs. Memory edits don't need lint, but DO use UTF-8 and
+  consistent line endings.
+
+### Tests
+
+Router tests (PR #1286 scope):
+```bash
+.venv/bin/python -m pytest fastvideo/tests/entrypoints/streaming/test_router.py -v --no-header
+```
+
+Stack baseline (May 2 handoff suite — re-run when you change anything in
+api/, contract/, or LTX-2 paths):
+```bash
+.venv/bin/python -m pytest \
+    fastvideo/tests/api/ \
+    fastvideo/tests/contract/ \
+    fastvideo/tests/ops/quantization/test_nvfp4_*.py \
+    tests/local_tests/pipelines/test_ltx2_pipeline_smoke.py \
+    -q --no-header
+```
+
+Expected baselines:
+- May 2 handoff (`156103b9`): 222 passed, 1 skipped.
+- Post-D-16 (`a152cb77` / `09647a30`): +7 router tests pass on top.
+
+### LSP
+
+Use `lsp_diagnostics` on changed files BEFORE running build. Pre-existing
+warnings to ignore (predate this work):
+
+- `fastvideo/entrypoints/streaming/router/main.py:37` — `Task` generic.
+- `fastvideo/entrypoints/cli/router_serve.py:55` — `_SubParsersAction` generic.
+
+### gh CLI for PR status
+
+```bash
+# PR #1286 quick status
+gh pr view 1286 --json headRefOid,mergeable,statusCheckRollup \
+  --jq '{headRefOid, mergeable, checks: [.statusCheckRollup[] | {name, status, conclusion}]}'
+
+# All commits in a PR + co-author check
+gh pr view 1286 --json commits \
+  --jq '.commits[] | {oid: .oid[0:8], msg: .messageHeadline, author: .authors[0].login}'
+```
+
+## Commit workflow
+
+### Subject convention
+
+`[type] <scope>: <imperative summary>` — keep ≤ 72 chars.
+
+Types observed in this scope: `feat`, `fix`, `test`, `docs`, `chore`,
+`refactor`. Scopes observed: `streaming`, `dreamverse-integration`,
+`api`, `quant`, `ltx2`, `nvfp4`, etc.
+
+Examples:
+- `[fix] streaming: router polish — bridge cancel + state machine + deps`
+- `[docs] dreamverse-integration: add authors.md + track D-16 router polish`
+
+### Body convention
+
+Bullet list, one bullet per file or concern. Why-before-what. Wrap at
+~80 chars (yapf doesn't reformat commit messages; readability is on you).
+
+### Co-author trailers (REQUIRED on every commit)
+
+The 4 trailers in [authors.md](authors.md) MUST appear on every commit
+in this scope. Use `--trailer` flags or write the body to a file with
+`-F` — DO NOT use multiple `-m` blocks for the trailers (each `-m` is
+its own paragraph and git's trailer parser only reads the LAST paragraph,
+yielding 1 trailer parsed instead of 4).
+
+**Inline `--trailer` form (preferred for short commits):**
+
+```bash
+git commit -m "subject" -m "body..." \
+  --trailer "Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>" \
+  --trailer "Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>" \
+  --trailer "Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>" \
+  --trailer "Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>"
+```
+
+**File form (preferred for multi-paragraph bodies):**
+
+```bash
+cat > /tmp/opencode/msg.txt <<'EOF'
+[type] scope: subject
+
+* Bullet one with rationale.
+* Bullet two with rationale.
+
+Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>
+Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>
+Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>
+Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>
+EOF
+git commit -F /tmp/opencode/msg.txt
+```
+
+The trailers MUST be a single block at the end of the message with no
+blank lines between them.
+
+**Verify trailers parsed:**
+
+```bash
+git log -1 --format='%(trailers:key=Co-authored-by,valueonly)'
+```
+
+Should print 4 lines (one per author). If only 1 line, you have the
+multi-`-m` bug — amend with `-F` to fix (allowed if commit is unpushed
+and you authored it in this session per AGENTS.md amend rules).
+
+### NEVER add to commits
+
+Per [`AGENTS.md`](../../../AGENTS.md):
+
+- AI co-authors (Claude, GPT, Codex, Cursor, etc.) — explicitly forbidden
+- "Generated with Claude Code" footer — explicitly forbidden
+- `--no-verify` to skip pre-commit — explicitly forbidden
+
+## Push + PR propagation
+
+### Pushing `will/ltx2_sr_port` (top of stack)
+
+```bash
+git push origin will/ltx2_sr_port  # fast-forward, no force needed
+```
+
+If git wants to force-push, you've rewritten history. STOP and verify:
+
+```bash
+git log origin/will/ltx2_sr_port..will/ltx2_sr_port  # local-only commits
+git log will/ltx2_sr_port..origin/will/ltx2_sr_port  # remote-only commits
+```
+
+Force-push requires explicit user confirmation per `AGENTS.md`.
+
+### Propagating fixes to PR #1286 (`will/api_7.9`)
+
+When a fix is in router code (`fastvideo/entrypoints/streaming/router/`,
+`cli/router_serve.py`, `tests/entrypoints/streaming/test_router.py`,
+or `pyproject.toml` router-related), it must land on BOTH branches.
+Cherry-pick avoids any force-push:
+
+```bash
+# 1. Commit on will/ltx2_sr_port first (working branch)
+git add <files...>
+git commit -F /tmp/opencode/msg.txt   # with trailers per above
+
+# 2. Cherry-pick onto will/api_7.9 (creates a separate SHA, identical diff)
+git checkout will/api_7.9
+git cherry-pick <ltx2_sr_port-sha>
+git push origin will/api_7.9          # fast-forward, no force
+
+# 3. Return to working branch
+git checkout will/ltx2_sr_port
+
+# 4. Verify PR #1286 picked it up
+gh pr view 1286 --json headRefOid --jq '.headRefOid'
+```
+
+Two SHAs for the same diff — they'll dedupe naturally on the next
+bulk-rebase via the trailer-injection rebase command in
+[authors.md](authors.md).
+
+### When a fix is memory-dir-only
+
+`.agents/memory/dreamverse-integration/` lives in the `agents_cleanup`
+layer of the stack — it does NOT belong on `will/api_7.9`. Memory updates
+stay on `will/ltx2_sr_port` only.
+
+### When a fix is non-router code in the integration scope
+
+Land on `will/ltx2_sr_port`. If that fix needs to ship as a separate PR
+(e.g. extending PR 7.10 or starting PR 9), open a new branch off the
+right base per [pr-roadmap.md](pr-roadmap.md).
+
+## Memory dir maintenance
+
+When state changes, update the memory dir BEFORE moving on. Every file
+has a "Last updated" header — bump when you edit.
+
+| Change | File to update |
+|---|---|
+| Branch tip moves | [state.md](state.md) "Branch tips" + "Last reconciled" |
+| PR opens / merges | [pr-roadmap.md](pr-roadmap.md) status table |
+| New decision made | [decisions-log.md](decisions-log.md) — add D-N entry, bump header |
+| Open thread resolved | [open-threads.md](open-threads.md) — strikethrough + "Resolved" note |
+| New open thread | [open-threads.md](open-threads.md) — priority overview + section |
+| New collaborator credited | [authors.md](authors.md) roster + trailer block + bulk-rebase |
+| Source doc archived | [source-archive/README.md](source-archive/README.md) + [README.md](README.md) sources table |
+| Process / runbook detail changes | [runbook.md](runbook.md) (this file) |
+
+Cross-link siblings via relative paths. Never duplicate content — link.
+
+## Common pitfalls
+
+### `pre-commit` not in `.venv/bin`
+
+`pre-commit` lives at `/home/william5lin/miniconda3/envs/fv-main/bin/pre-commit`.
+The `.venv` here is for the FastVideo package itself, not pre-commit.
+
+### Trailers split across paragraphs
+
+`git commit -m A -m B -m C` makes A, B, C separate paragraphs. Git's
+trailer parser only reads the LAST paragraph — multiple `-m
+"Co-authored-by: ..."` produces 1 trailer parsed, not 4. Use `--trailer`
+flags or `-F` with the trailers in a single block at the end.
+
+### Stash 0 on FastVideo IS NOT yours
+
+`stash@{0}: WIP on main: 71bfc13d HunyuanVideo plugin` predates this work.
+**DO NOT POP.** See [state.md](state.md) "Stashes — DO NOT POP".
+
+### `AbsMaxFP8` test "failure" is pre-existing
+
+`fastvideo/tests/ops/quantization/test_absmax_fp8.py::test_create_weights_rejects_invalid_dtype`
+fails on `main` and on every branch in this scope. NOT introduced by
+integration work. See [open-threads.md](open-threads.md) item #2.
+
+### Untracked nested clones at repo root
+
+`dynamo/`, `ray/`, `vllm-omni/` are untracked nested git clones at the
+FastVideo repo root. Reference repos for cross-repo work. **Do not
+`rm -rf`** — they're someone else's working state.
+
+### Live services on 8009 / 5274
+
+`dreamverse-server` runs on 8009 (warmed GPU worker), Next.js dev server
+on 5274. Don't start new instances on those ports without checking
+[state.md](state.md) "Live services" first.
+
+### Branch may have been switched by another agent
+
+Other agents share this worktree. If `git branch --show-current` returns
+something other than `will/ltx2_sr_port`, switch back cleanly with
+`git checkout will/ltx2_sr_port` — don't disturb their work, don't
+discard their uncommitted changes.
+
+### Force-push policy
+
+Per `AGENTS.md`: never force-push without explicit user confirmation.
+For trailer fixes on already-pushed commits, prefer the bulk-rebase
+command in [authors.md](authors.md) — safe to re-run.
+
+### Two trailerless commits in PR #1286
+
+`a152cb77` (on `will/api_7.9`) and `40e265b8` (now-superseded ancestor
+on `will/ltx2_sr_port`) lack the 4 co-author trailers. **Accepted gap**
+per user decision — see [authors.md](authors.md) "Known gaps".
+
+## Self-test (verify your context is loaded)
+
+After reading the memory dir, you should be able to answer:
+
+1. What branch should I be on? → `will/ltx2_sr_port`
+2. What's the active open PR in this scope? → #1286 on `will/api_7.9`
+3. Where does PR #1286 land in the stack? → Bottom; ancestor of `will/ltx2_sr_port`
+4. Who do I credit on every commit? → 4 authors per [authors.md](authors.md)
+5. Where do memory updates land? → `will/ltx2_sr_port` only (NOT api_7.9)
+6. What's the next-priority open thread? → See [open-threads.md](open-threads.md) "Recommended pull order" — D-8 verify is current top
+7. What pre-existing failure can I ignore? → AbsMaxFP8 test (item #2)
+8. What's the bulk-rebase command for adding trailers across the stack? → See [authors.md](authors.md) "How the trailers were applied"
+
+If you can't answer one of these from the memory dir alone, the dir has
+a gap — file it as a new entry in [open-threads.md](open-threads.md)
+before continuing.
+
+## First 60 seconds — copy-paste orientation
+
+```bash
+# 1. Confirm branch
+cd /home/william5lin/FastVideo
+git branch --show-current   # should print: will/ltx2_sr_port
+# If not, recover: git checkout will/ltx2_sr_port
+
+# 2. Confirm worktree clean (untracked nested clones expected)
+git status --short
+
+# 3. Confirm PR #1286 head matches expected api_7.9 tip
+gh pr view 1286 --json headRefOid --jq '.headRefOid'
+git rev-parse will/api_7.9   # should match PR head
+
+# 4. Confirm your context vs the memory dir
+git log -1 --oneline
+cat .agents/memory/dreamverse-integration/state.md | head -30
+
+# 5. Confirm live services still running
+curl -s http://localhost:8009/readyz | head -c 200
+curl -s http://localhost:5274/ -o /dev/null -w "%{http_code}\n"
+```
+
+If any of those produce unexpected output, read [state.md](state.md)
+before changing anything.

.agents/memory/dreamverse-integration/source-archive/README.md

@@ -0,0 +1,44 @@
+# Source Archive
+
+These are the original unsynthesized design and integration docs that
+predate the consolidation in
+[`../`](../). They are **NOT** the source of truth — the synthesized
+sibling files in the parent directory are.
+
+Archived 2026-05-03. All previously untracked.
+
+## Contents
+
+| File | Original location | Date | Synthesized into |
+|---|---|---|---|
+| `apirefactor.md` | `FastVideo/` (repo root) | 2026-04-21 | [`../design.md`](../design.md) |
+| `PR-plan.md` (was `PR plan.md` at repo root) | `FastVideo/` (repo root) | 2026-04-25 | [`../pr-roadmap.md`](../pr-roadmap.md) |
+| `dreamverse_review.md` | `FastVideo/` (repo root) | 2026-04-26 | [`../decisions-log.md`](../decisions-log.md) + [`../state.md`](../state.md) |
+| `handoff-nvfp4-launch-demo.md` | `.agents/exploration/` | 2026-05-02 | [`../state.md`](../state.md) + [`../quantization.md`](../quantization.md) + [`../open-threads.md`](../open-threads.md) |
+| `streaming-server-upstream-plan.md` | `.agents/exploration/` | 2026-04-17 | [`../streaming-server.md`](../streaming-server.md) + [`../decisions-log.md`](../decisions-log.md) |
+| `dreamverse_integration.md` | `.agents/exploration/` | 2026-04-23 | [`../cross-repo-surfaces.md`](../cross-repo-surfaces.md) |
+| `video-generator-config-api-design.md` | `.agents/exploration/` | 2026-04-02 | [`../design.md`](../design.md) (early-draft material) |
+
+## Why archived (not deleted)
+
+- Future agents may want the **full unsynthesized rationale** for a
+  decision the synthesis abbreviated.
+- The originals remain useful as a **time machine** for understanding
+  how the design evolved.
+- These docs were never committed to git, so leaving them on disk costs
+  nothing.
+
+## When to read the archive vs. the synthesis
+
+- **Read the synthesis (`../*.md`)** for: current state, decision
+  status, action items, design rationale at the conceptual level.
+- **Read the archive (here)** for: deep historical context, exact wording
+  of design decisions, full PR plan with all sub-PR commit details,
+  the original Q-1..Q-9 / D-1..D-11 prose.
+
+## Maintenance rule
+
+Do NOT edit files in this archive. They are point-in-time snapshots.
+If new design material appears that supersedes an entry here, update the
+synthesis (the parent dir) and append a note to that synthesis file —
+do not mutate this archive.

.agents/memory/dreamverse-integration/source-archive/apirefactor.md

@@ -0,0 +1,838 @@
+# FastVideo API Refactor Design
+
+## Related Documents
+- [PR plan.md](PR%20plan.md) — PR-by-PR implementation plan for this design
+- [.agents/exploration/streaming-server-upstream-plan.md](.agents/exploration/streaming-server-upstream-plan.md) — streaming-server upstream + Dynamo backend contract (shapes PRs 5.5-7.10)
+- `../FastVideo-internal/.agents/exploration/rebase-upstream-fastvideo.md` — rebasing FastVideo-internal onto upstream (enables PRs 6-8)
+- `../FastVideo-internal/ui/ltx2-streaming/` — source for the streaming server being upstreamed (PRs 7.5-7.9)
+- `../dynamo/` — local clone of ai-dynamo/dynamo; `components/src/dynamo/sglang/` is the template for FastVideo's native backend landed in PR 7.10
+- https://github.com/ai-dynamo/dynamo/pull/7544 — closed draft PR that establishes the Dynamo backend shape this design must satisfy
+
+## Status
+
+Design spec for the public inference API refactor. PRs 0-5.5 are landed; see [PR plan.md](PR%20plan.md) for rollout status and the PR 6+ roadmap. The typed schema, strict parser, preset system, typed VideoGenerator, typed CLI, and stateless OpenAI server default-request merge are all implemented. Streaming package skeleton + typed streaming config types are in place; live streaming server + Dynamo contract are the next milestones.
+
+## Executive Summary
+
+FastVideo should move to a single typed nested inference schema that is shared across:
+
+- Python API
+- CLI
+- YAML/JSON config files
+- OpenAI/server request translation
+
+The core split is:
+
+- `GeneratorConfig`: generator-instance lifetime settings
+- `GenerationRequest`: per-call inputs, sampling, outputs, and continuation
+- `InferencePreset`: model-owned named multi-stage defaults
+
+The canonical user experience should be:
+
+1. Choose a model.
+2. Choose a pipeline preset.
+3. Override a few typed fields.
+4. Generate.
+
+FastVideo should not make a raw free-form string dict the primary API. Dicts and YAML/JSON should be supported as serialization/interchange layers, but they must be parsed immediately into typed config objects with strict unknown-key validation.
+
+The repo should also shift model-specific preset/default definitions closer to their pipeline implementations, while keeping the shared public schema and parsers centralized.
+
+## Why This Refactor Is Needed
+
+Today the public inference boundary is too flat and too forgiving.
+
+- `VideoGenerator.from_pretrained(..., **kwargs)` mixes:
+  - engine/runtime settings
+  - pipeline init settings
+  - component overrides
+- `VideoGenerator.generate_video(..., **kwargs)` mixes:
+  - prompt and inputs
+  - sampling parameters
+  - output settings
+  - model-specific workflow knobs
+- unknown or drifting keys can be silently filtered or merely logged instead of failing fast
+- model-specific multi-stage behavior is exposed through ad hoc top-level flags instead of a stable preset/stage abstraction
+
+This is already painful in LTX2/Dreamverse, and it will get worse as more multi-stage pipelines are upstreamed.
+
+## Design Goals
+
+- Keep the Python API typed and editor-friendly.
+- Make YAML/JSON a first-class serialization of the same schema.
+- Support CLI overrides cleanly without flattening the schema into hundreds of canonical flags.
+- Separate init-time config from request-time config.
+- Provide a stable public abstraction for multi-stage pipelines.
+- Support LTX2 two-stage and continuation behavior cleanly.
+- Keep the simple case simple.
+- Co-locate model-owned defaults and stage topology with the relevant pipeline.
+- Protect current public/server behavior with an explicit schema parity audit before freezing the new surface.
+- Preserve backward compatibility long enough to migrate examples, internal users, and servers safely.
+
+## Non-Goals
+
+- Do not make Ray a structural dependency or copy its package layout.
+- Do not make a raw free-form dict the primary Python API.
+- Do not force all models into one universal `RefineConfig`.
+- Do not expose stage indices as the primary user interface.
+- Do not move every shared config class into per-model directories.
+
+## External Inspiration
+
+### Ray
+
+Borrow only the ergonomic idea that user-facing config can be expressed as a string-keyed dict or YAML/JSON config. Do not copy Ray's structure into FastVideo.
+
+### SGL Multimodal Gen
+
+Useful ideas: split instance config from request config; allow dict input at the boundary; parse dicts immediately into typed request objects; merge request overrides onto model defaults; validate request params against pipeline/task type. Do not copy: request objects depending on server/engine config; broad weakly typed request bags as the canonical API.
+
+### vLLM-Omni
+
+Useful ideas: model-owned pipeline presets; explicit stage topology; per-stage default sampling params; clean separation between stage topology, engine defaults, and runtime overrides. Do not copy: positional `sampling_params_list` as the primary public API; serving-engine-oriented stage index semantics in the main Python interface.
+
+## Core Decision
+
+FastVideo should have:
+
+1. A shared typed public schema.
+2. Model-owned named pipeline presets.
+3. Semantic stage overrides by stage name.
+4. Optional advanced explicit plans for power users.
+5. YAML-first config loading with dotted CLI overrides.
+
+The public API should be stable at the schema level, while model-specific behavior should be contained in preset definitions and model-specific typed override classes.
+
+## Schema Parity Requirement
+
+Before the new schema is declared canonical, FastVideo should build a parity inventory across all current public inference surfaces (Python `VideoGenerator` kwargs, CLI flags, YAML/JSON config inputs, OpenAI/server request models, model-specific sampling/runtime fields). Each field must be marked: kept as-is, renamed, moved to a nested path, preset-owned, private-only adapter field, or intentionally dropped. No field should disappear implicitly.
+
+For any public field that remains supported, there should be either a normalized-config equivalence test, or an explicit parser/translation test. Fields that exist only in private Dreamverse integration code should be handled by a private adapter layer, not quietly converted into public FastVideo compatibility guarantees.
+
+Landed artifact: [inference_schema_parity_inventory.yaml](docs/design/inference_schema_parity_inventory.yaml) + guard [test_schema_parity_inventory.py](fastvideo/tests/api/test_schema_parity_inventory.py).
+
+## Canonical Public Schema
+
+The typed schema is implemented in [fastvideo/api/schema.py](fastvideo/api/schema.py). Envelope types:
+
+- `RunConfig` — offline: `generator` (GeneratorConfig) + `request` (GenerationRequest)
+- `ServeConfig` — serving: `generator` + `server` (ServerConfig) + `default_request` (GenerationRequest) + optional `streaming` (StreamingConfig)
+
+Key nested types (summary; full fields in `schema.py`):
+
+- `GeneratorConfig` → `model_path`, `revision`, `trust_remote_code`, `engine` (EngineConfig: parallelism/offload/compile/quantization/flags), `pipeline` (PipelineSelection: workload_type, preset, preset_version, components, preset_overrides, experimental)
+- `GenerationRequest` → `prompt`, `negative_prompt`, `inputs` (InputConfig), `sampling` (SamplingConfig), `runtime` (RequestRuntimeConfig), `output` (OutputConfig), `stage_overrides`, `state` (ContinuationState), `plan` (GenerationPlan), `extensions`
+- `ContinuationState` → opaque `{kind: str, payload: dict[str, Any]}`
+- `GenerationPlan` → `{stages: list[PlannedStage], final_stage: str | None}`; advanced/escape-hatch only
+
+### Important Semantics
+
+- Dataclasses are canonical for Python users.
+- Dict and YAML/JSON are parsed into these dataclasses immediately.
+- Unknown keys must raise validation errors.
+- Typed `GenerationRequest` defaults come from the public schema, not from model-specific `SamplingParam.from_pretrained(...)` defaults.
+- Legacy `generate_video(...)` continues to inherit model-specific sampling defaults until the SSIM/performance migration lands (PR 11).
+- The only open-ended escape hatches are:
+  - `generator.pipeline.experimental`
+  - `request.extensions`
+
+That keeps the public contract strict without blocking experimental work.
+
+### Request Mutation Tracking
+
+When a `GenerationRequest` is parsed from a raw dict (YAML, JSON, or Python mapping), FastVideo records which fields the user explicitly provided versus which received schema defaults. This matters because `request_to_sampling_param()` must distinguish user-provided values (which should override model defaults) from schema defaults (which should NOT override model defaults).
+
+The tracking contract:
+
+- At parse time, the original raw dict and a baseline snapshot of the parsed object are stored on the request.
+- Dataclass field mutations after parsing (e.g., `request.sampling.seed = 7`) are captured via lightweight `__setattr__` dirty-path recording.
+- Dict-typed field mutations (e.g., `del request.stage_overrides["refine"]`) are detected at access time by diffing the current dict against the baseline snapshot.
+- Setting a field to the schema default value IS captured as explicit, so it will override model defaults.
+- The raw dict is reconciled lazily when `normalize_generation_request()` is called, not on every individual mutation.
+
+### Schema Purity and Model-Specific Fields
+
+The shared schema currently contains fields that are specific to one or two model families. These remain for backward compatibility during the initial migration (PRs 0-3) but should migrate to preset-owned typed override classes as the preset system lands (PRs 4-10).
+
+**SamplingConfig fields to migrate:**
+
+- `height_sr`, `width_sr`, `num_inference_steps_sr`: Hunyuan15 SR only. Target: `HunyuanSRStageOverride` in PR 10.
+- `guidance_scale_2`, `boundary_ratio`: Wan2.2 and LingBotWorld only. Target: preset-owned overrides in the relevant model migration PR.
+
+**InputConfig fields to migrate:**
+
+- `mouse_cond`, `keyboard_cond`, `grid_sizes`: MatrixGame action control only. Target: `request.extensions` or a typed MatrixGame input config.
+- `c2ws_plucker_emb`: LingBotWorld camera control only. Target: `request.extensions` or a typed LingBotWorld input config.
+- `refine_from`, `stage1_video`: LongCat refinement only. Target: `LongCatRefineStageOverride` inputs or keep in `InputConfig` if they remain a public contract.
+
+**Universal fields that stay in the shared schema:**
+
+- `guidance_rescale`: used by multiple denoising stages across models, default 0.0. Universally applicable.
+- `true_cfg_scale`: OpenAI adapter surface. Keep for protocol compatibility.
+
+### Escape Hatch Sunset
+
+`generator.pipeline.experimental` and `request.extensions` are intentional escape hatches for experimental and private work. They bypass strict validation by design.
+
+Rules for escape hatch usage:
+
+- New fields should not be added to `experimental` or `extensions` without a plan to either promote them to typed fields or remove them within two PR cycles.
+- Each model migration PR (PRs 6-10) should review and shrink escape hatch usage for that model family.
+- The compatibility layer currently routes unrecognized legacy kwargs into `experimental`. This pass-through should shrink as presets absorb model-specific fields.
+
+## Public Python API
+
+### New Canonical API
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    GeneratorConfig, GenerationRequest,
+    EngineConfig, OutputConfig,
+    PipelineSelection, SamplingConfig,
+)
+
+generator = VideoGenerator.from_pretrained(
+    config=GeneratorConfig(
+        model_path="/models/ltx2",
+        engine=EngineConfig(num_gpus=1),
+        pipeline=PipelineSelection(
+            workload_type="t2v",
+            preset="ltx2_two_stage",
+        ),
+    )
+)
+
+result = generator.generate(
+    GenerationRequest(
+        prompt="a fox running through snow",
+        sampling=SamplingConfig(
+            num_frames=121, height=1024, width=1536,
+            num_inference_steps=8, seed=42,
+        ),
+        output=OutputConfig(save_video=True, return_state=True),
+    )
+)
+```
+
+### Accepted Construction Forms
+
+Canonical:
+
+```python
+VideoGenerator.from_pretrained(config=GeneratorConfig(...))
+VideoGenerator.from_config(GeneratorConfig(...))
+VideoGenerator.from_file("run.yaml")
+```
+
+Stable convenience constructor:
+
+```python
+VideoGenerator.from_pretrained("model-id")
+VideoGenerator.from_pretrained("model-id", num_gpus=2, use_fsdp_inference=False, ...)
+```
+
+Legacy compatibility:
+
+```python
+VideoGenerator.from_pretrained(model_path, **legacy_kwargs)
+```
+
+All constructor forms normalize through the same typed path. Stable convenience kwargs remain supported with no deprecation warning. Advanced model/pipeline-specific kwargs are accepted during migration but only as compatibility inputs that normalize into `GeneratorConfig`. The thing being deprecated over time is the unbounded legacy kwarg surface, not the `from_pretrained(...)` entrypoint itself.
+
+### Generation Entry Point
+
+Canonical: `generator.generate(request: GenerationRequest) -> GenerationResult`.
+
+Compatibility alias: `generator.generate_video(prompt=..., **legacy_kwargs)` — converts legacy calls into a `GenerationRequest` and emits a deprecation warning.
+
+During the compat period, `generate(request=...)` uses schema defaults while `generate_video(...)` preserves legacy model-default behavior. These paths intentionally differ until preset-owned defaults replace the remaining `SamplingParam` default logic (migrated in PR 11).
+
+### Boundary Normalization Rule
+
+Every public inference entrypoint normalizes into typed config objects before touching legacy internals. That includes Python constructors, generation calls, CLI `generate`, CLI `serve`, and OpenAI/server request translation. Legacy internals (`FastVideoArgs`, `SamplingParam`) may remain temporarily, but only behind a typed normalization boundary.
+
+## Pipeline Presets
+
+### Definition
+
+An `InferencePreset` is a named model-owned preset that defines:
+
+- workload selection
+- stage topology
+- per-stage defaults
+- stage names
+- allowed stage override types
+- init-time feature requirements
+
+The preset is not user-authored by default. It is supplied by the model integration.
+
+### Why Presets Are The Right Abstraction
+
+Users usually do not want to assemble a stage graph by hand. They want to say:
+
+- use LongCat distill + refine
+- use Hunyuan 1080p SR
+- use LTX2 two-stage continuation mode
+
+Presets provide a stable public noun for that behavior.
+
+### Preset Naming Rules
+
+- Use semantic names, not stage indices.
+- Keep names stable across releases.
+- If semantics change incompatibly, change `preset_version` or create a new preset name.
+
+Examples: `ltx2_base`, `ltx2_two_stage`, `longcat_distill_refine`, `hunyuan15_sr_720p`, `hunyuan15_sr_1080p`.
+
+### Preset-Owned Stage Names
+
+Stage names are public and stable within a preset.
+
+- LTX2: `base`, `refine`
+- LongCat: `distill`, `refine`
+- Hunyuan15: `base`, `sr_720p`, `sr_1080p`
+
+Public overrides should reference these stage names, never stage indices.
+
+## Stage Overrides
+
+The main user override surface for multi-stage pipelines is:
+
+```python
+request.stage_overrides["refine"] = ...
+```
+
+Each model family should expose typed override classes for its stage names. Examples for the model families that land in PRs 6/9/10:
+
+```python
+@dataclass
+class LTX2RefineStageOverride:
+    enabled: bool | None = None
+    num_inference_steps: int | None = None
+    guidance_scale: float | None = None
+    add_noise: bool | None = None
+    image_crf: int | None = None
+    video_position_offset_sec: float | None = None
+
+@dataclass
+class LongCatRefineStageOverride:
+    t_thresh: float | None = None
+    spatial_refine_only: bool | None = None
+    num_cond_frames: int | None = None
+
+@dataclass
+class HunyuanSRStageOverride:
+    num_inference_steps: int | None = None
+    guidance_scale: float | None = None
+```
+
+### Strictness Rules
+
+- Stage names must exist in the selected preset.
+- Override fields must be valid for that stage type.
+- Unknown stage names and unknown fields must error.
+
+## Advanced Explicit Plans
+
+Presets should be the default API. `GenerationPlan` exists only for advanced composition or experimentation:
+
+- building a custom workflow that is not yet standardized as a preset
+- debugging or benchmarking stage combinations
+- prototyping a future preset
+
+Do not require `GenerationPlan` for normal users.
+
+## Continuation State
+
+Continuation must be a first-class part of the API.
+
+### Public Contract
+
+- `GenerationResult.state` may return a `ContinuationState`.
+- `GenerationRequest.state` may accept a previously returned state.
+- Most users should treat `state` as opaque and round-trip it back into the next request.
+
+### Why This Matters
+
+Dreamverse/LTX2 currently leaks continuation internals into app-level request fields like video conditions, audio clean latent, audio denoise mask, and segment offsets. Those should not remain top-level app-owned public API.
+
+### State Design
+
+Public surface:
+
+```python
+@dataclass
+class ContinuationState:
+    kind: str
+    payload: dict[str, Any]
+```
+
+Internally, FastVideo should also define typed model-specific state subclasses, e.g. `LTX2ContinuationState` (PR 7) and `LongCatIntermediateState` if ever needed. Minimal stable surface: return state, pass state back in, validate that the state is compatible with the active preset.
+
+Payload serialization: fields must be JSON-serializable or use an opaque blob-ID indirection for large tensors. This supports both the stateless OpenAI client round-trip AND future Dynamo prefill/decode disaggregation where prefill yields a state that decode hydrates across workers.
+
+## YAML / JSON Design
+
+YAML and JSON should be exact serializations of the typed schema, not a second unrelated config system. YAML is the primary documented format. JSON is accepted with the same schema.
+
+### Run Config Example
+
+```yaml
+generator:
+  model_path: /models/ltx2
+  engine:
+    num_gpus: 1
+    parallelism: {tp_size: -1, sp_size: -1}
+    offload: {dit: false, text_encoder: false, vae: false, pin_cpu_memory: true}
+  pipeline:
+    workload_type: t2v
+    preset: ltx2_two_stage
+    components:
+      config_root: /models/ltx2-config
+      upsampler_weights: /models/ltx2-refine
+      lora_path: /models/ltx2-refine-lora
+    preset_overrides:
+      refine: {enabled: true, add_noise: true}
+
+request:
+  prompt: "a fox running through snow"
+  sampling:
+    num_frames: 121
+    height: 1024
+    width: 1536
+    num_inference_steps: 8
+    seed: 42
+  output: {save_video: true, return_state: true}
+  stage_overrides:
+    refine: {num_inference_steps: 2, guidance_scale: 1.0}
+```
+
+### Serve Config Example
+
+```yaml
+generator:
+  model_path: /models/ltx2
+  engine: {num_gpus: 1}
+  pipeline: {workload_type: t2v, preset: ltx2_two_stage}
+
+server: {host: 0.0.0.0, port: 8000, output_dir: outputs/}
+
+default_request:
+  sampling: {num_frames: 121, height: 1024, width: 1536, num_inference_steps: 8}
+  output: {save_video: false, return_frames: false}
+```
+
+### Validation Rules
+
+- top-level schema must match `RunConfig` or `ServeConfig`
+- unknown keys must fail
+- dotted CLI overrides are applied to the nested config before typed parsing
+- parse errors must include the exact nested path that failed
+
+## CLI Design
+
+Inference CLI reuses the best parts of the current training authoring flow (YAML-first authoring, dotted nested overrides, typed parsing after merge) but stays stricter than training at the public boundary because it is a user-facing API surface for Python, CLI, YAML/JSON, and serving.
+
+### Canonical CLI Forms
+
+```bash
+fastvideo generate --config run.yaml
+fastvideo generate --config run.yaml --request.sampling.seed 42
+fastvideo generate --config run.yaml --generator.engine.num_gpus 2
+
+fastvideo serve --config serve.yaml
+fastvideo serve --config serve.yaml --server.port 8090
+```
+
+The CLI is config-only. Beyond `--config`, CLI input uses dotted override paths into the nested schema rather than maintaining a second flat flag surface.
+
+Implementation: YAML/JSON is loaded into a nested dict, dotted CLI overrides are applied to the nested dict, then the result is parsed into typed config objects. Flat CLI flags are rejected so the nested schema stays canonical.
+
+## OpenAI / Server Mapping
+
+`fastvideo serve` loads `ServeConfig`. Incoming HTTP requests are translated into `GenerationRequest` by:
+
+1. cloning `default_request`
+2. applying API request fields onto that request
+3. validating against the selected preset
+
+This is similar in spirit to the SGL pattern of merging user overrides onto model defaults.
+
+Rules:
+
+- HTTP request translation must not bypass typed validation.
+- multi-stage defaults should come from the preset and `default_request`, not from ad hoc server-local logic.
+- stateful continuation requests should accept and return typed `ContinuationState` payloads.
+
+Landed in PR 5 for the stateless OpenAI server at `fastvideo/entrypoints/openai/`. The streaming/session server (PRs 7.5-7.9) uses the same preset/default_request merge through `ServeConfig.streaming`.
+
+## Streaming Server + Dynamo Backend
+
+The typed public API is consumed by three server-class integrations. They must share one execution substrate so we don't grow three near-duplicate progress loops.
+
+### The three consumers
+
+| Consumer | Transport | Request shape | State |
+|---|---|---|---|
+| Stateless OpenAI (`fastvideo/entrypoints/openai/`) | HTTP POST | `GenerationRequest` merged onto `ServeConfig.default_request` | Stateless; continuation via opaque payload if needed |
+| Streaming WebSocket (`fastvideo/entrypoints/streaming/`) | WebSocket JSON + binary fMP4 | `GenerationRequest` per segment, session-scoped | Server-held session (per-GPU continuation cache); snapshot on demand |
+| Dynamo native backend (`ai-dynamo/dynamo/components/src/dynamo/fastvideo/`) | Dynamo RPC endpoint | `NvCreateVideoRequest` ↔ adapter ↔ `GenerationRequest` | Aggregated today; disaggregated prefill/decode later via `ContinuationState` |
+
+### Shared execution substrate: `VideoGenerator.generate_async`
+
+The OpenAI server, streaming server, and Dynamo backend all want the same thing: a typed async API that yields progress events and a typed final result. FastVideo exposes exactly one canonical entry point:
+
+```python
+async def generate_async(
+    self,
+    request: GenerationRequest,
+) -> AsyncGenerator[VideoEvent, None]: ...
+```
+
+Events:
+
+```python
+@dataclass
+class VideoProgressEvent:
+    step: int
+    total_steps: int
+    stage: str  # "denoise" | "refine" | "decode" | ...
+
+@dataclass
+class VideoPartialEvent:
+    frames: np.ndarray           # shape: (num_frames, H, W, 3)
+    index: int                   # monotonic chunk index
+
+@dataclass
+class VideoFinalEvent:
+    video_bytes: bytes | None    # mp4-encoded if requested
+    tensor: torch.Tensor | None  # raw if requested
+    metadata: dict[str, Any]
+    continuation_state: ContinuationState | None
+
+VideoEvent = VideoProgressEvent | VideoPartialEvent | VideoFinalEvent
+```
+
+The sync `generate_video(request=...) -> VideoResult` becomes a thin `asyncio.run` wrapper over `generate_async` that collects events and returns the final.
+
+### Streaming server mapping
+
+`fastvideo/entrypoints/streaming/` owns per-session state:
+
+- `SessionStore.hydrate(state: ContinuationState) -> session_id`
+- `SessionStore.snapshot(session_id) -> ContinuationState`
+- Per-GPU implicit continuation cache (today's internal behavior) is wrapped as a `SessionStore` implementation.
+
+Per-segment, the session writes a `GenerationRequest`, pipes the event stream to the WebSocket (progress → JSON messages, partial → fMP4 frames), and persists the final's `ContinuationState` into the session.
+
+### Dynamo backend mapping
+
+Dynamo's backend pattern (from `components/src/dynamo/sglang/`) is a pure Python import. FastVideo does not host a `fastvideo/entrypoints/dynamo/` subpackage; the integration lives in the Dynamo repo. FastVideo exposes a stable contract:
+
+| Surface | Exposed as |
+|---|---|
+| Construction | `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` |
+| Execution (async) | `VideoGenerator.generate_async(request) -> AsyncGenerator[VideoEvent, None]` |
+| Execution (sync) | `VideoGenerator.generate_video(request=...) -> VideoResult` |
+| Typed request | `fastvideo.api.GenerationRequest`, `SamplingConfig`, `InputConfig` |
+| Typed result | `fastvideo.api.VideoResult`, `VideoEvent`, `ContinuationState` |
+| Health-check input | `VideoGenerator.default_health_check_request() -> GenerationRequest` |
+| Config dump | `config_to_dict(cfg)` (already exists) |
+
+Request/response mapping the Dynamo adapter must perform:
+
+```
+NvCreateVideoRequest          ->  fastvideo.api.GenerationRequest
+  prompt                      ->    sampling.prompt
+  size="WxH"                  ->    sampling.width, sampling.height
+  seconds                     ->    seconds * nvext.fps -> sampling.num_frames
+  input_reference             ->    input.image_path | input.video_path
+  nvext.fps                   ->    sampling.fps
+  nvext.num_frames            ->    sampling.num_frames (overrides seconds*fps)
+  nvext.num_inference_steps   ->    sampling.num_inference_steps
+  nvext.guidance_scale        ->    sampling.guidance_scale
+  nvext.seed                  ->    sampling.seed
+  nvext.negative_prompt       ->    sampling.negative_prompt
+  response_format             ->    (handled at the adapter's output stage)
+
+VideoFinalEvent               ->  NvVideosResponse
+  video_bytes                 ->    data[0].b64_json  (if response_format=b64_json)
+  uploaded URL                ->    data[0].url       (if response_format=url)
+  metadata.inference_time_s   ->    inference_time_s
+  continuation_state          ->    (reserved for future disaggregation)
+```
+
+All fields already exist (or will exist after PR 6's typed-kwarg expansion) on FastVideo's typed schema. **The adapter lives entirely in the Dynamo repo** at `components/src/dynamo/fastvideo/` — FastVideo does not host any Dynamo subpackage, dep, or CLI. The only FastVideo obligation is the stable public Python API listed above.
+
+### Constraints this places on other sections
+
+- **Continuation State** (see earlier section): `ContinuationState.payload` must be JSON-serializable or use an opaque blob-ID indirection for large tensors. This supports both the stateless OpenAI client round-trip *and* future Dynamo prefill/decode disaggregation, where prefill yields a state that decode hydrates across workers.
+- **Typed GeneratorConfig** (see Public Python API): every flat legacy LTX2 kwarg currently used by the internal `gpu_pool.py` must have a typed home reachable from `GeneratorConfig`. Dynamo's `FastVideoArgGroup` builds the config from its CLI and must not have to know any legacy LTX2 name.
+- **Public exports**: `from fastvideo import VideoGenerator`; `from fastvideo.api import GenerationRequest, SamplingConfig, ContinuationState, VideoResult, VideoEvent, VideoProgressEvent, VideoPartialEvent, VideoFinalEvent`.
+
+## Repo Layout
+
+### Shared Public API
+
+`fastvideo/api/` contains the shared public API package. Current files:
+
+- `schema.py` — `RunConfig`, `ServeConfig`, `ServerConfig`, `GeneratorConfig`, and all nested typed config dataclasses
+- `sampling_param.py` — `SamplingParam` + `CacheParams` (canonical home since PR 4; former `configs/sample/base.py` location removed)
+- `presets.py` — `InferencePreset`, `PresetStageSpec`, registry APIs
+- `results.py` — `GenerationResult` / `VideoResult`
+- `parser.py` — `from_dict`, `to_dict`, `load_yaml`, `load_json`, validation
+- `overrides.py` — dotted override application
+- `compat.py` — legacy Python kwargs translation
+- `errors.py` — path-aware validation errors
+
+May split further by concern in a future cleanup.
+
+### Pipeline-Local Model-Owned Config
+
+Model-owned presets and override types live next to the model pipeline:
+
+```text
+fastvideo/pipelines/basic/ltx2/
+  ltx2_pipeline.py, presets.py, stage_overrides.py, continuation.py
+
+fastvideo/pipelines/basic/longcat/
+  longcat_pipeline.py, presets.py, stage_overrides.py
+
+fastvideo/pipelines/basic/hunyuan15/
+  hunyuan15_pipeline.py, hunyuan15_sr_pipeline.py, hunyuan15_2sr_pipeline.py,
+  presets.py, stage_overrides.py
+```
+
+PR 4 landed `presets.py` for all 13 model families. Remaining colocation targets are `pipeline_configs.py` (moving `configs/pipelines/<family>.py`) and model-specific stages (moving `pipelines/stages/<family>_*.py`); see [PR plan.md](PR%20plan.md) "Pipeline Package Structure".
+
+### Registry
+
+Central registry (`fastvideo/registry.py`) registers preset providers rather than owning all model-specific defaults directly. It answers:
+
+- which pipeline class corresponds to a model path
+- which presets are available for that model family
+- which override/state classes are valid for a selected preset
+
+## Relationship To Current Internal Classes
+
+This refactor does not require deleting current internals immediately.
+
+- `FastVideoArgs` is an internal compatibility/input adapter, no longer the primary public inference type.
+- `SamplingParam` now lives in `fastvideo/api/sampling_param.py` and gets model-specific defaults from presets via `_from_preset()`. All 12 `SamplingParam` subclasses have been removed and the former `fastvideo/configs/sample/` directory has been deleted entirely (PR 4). It remains an internal adapter between the preset system and the runtime.
+- current `PipelineConfig` classes can remain temporarily as internal component config carriers
+- the new public schema is the stable boundary above them
+
+`VideoGenerator` accepts the new schema and translates down into current execution internals. Legacy `generate_video(..., **kwargs)` stays on the direct execution path during the compat period until SSIM/performance tests migrate in PR 11.
+
+## Model-Specific Design
+
+### LTX2 / Dreamverse
+
+LTX2 needs both:
+
+- init-time two-stage feature wiring
+- request-time continuation/refine behavior
+
+Expressed as:
+
+- preset: `ltx2_two_stage`
+- init-time fields: refine assets, optional config root, stage enablement
+- request-time fields: stage override for refine behavior, optional returned continuation state
+
+#### LTX2 Preset Example
+
+```yaml
+generator:
+  pipeline:
+    preset: ltx2_two_stage
+    components:
+      config_root: /models/ltx2-config
+      upsampler_weights: /models/ltx2-refine
+      lora_path: /models/ltx2-refine-lora
+    preset_overrides:
+      refine: {enabled: true, add_noise: true}
+```
+
+#### LTX2 Request Example
+
+```yaml
+request:
+  prompt: "continue the previous sequence"
+  state: ${previous_result.state}
+  stage_overrides:
+    refine:
+      num_inference_steps: 2
+      guidance_scale: 1.0
+      image_crf: 18
+  output:
+    return_state: true
+```
+
+#### LTX2 Explicit Decisions
+
+- `config_model_path` becomes `generator.pipeline.components.config_root`
+- `ltx2_refine_*` stops being a pile of top-level kwargs
+- continuation internals move into `ContinuationState`
+- app-level code should pass `state`, not raw latent/audio condition payloads
+
+### LongCat
+
+LongCat should expose a named preset like `longcat_distill_refine` with stage topology `distill` and `refine`.
+
+User-facing override knobs remain model-specific (`t_thresh`, `spatial_refine_only`, `num_cond_frames`) but live under:
+
+```yaml
+request:
+  stage_overrides:
+    refine:
+      t_thresh: 0.5
+      spatial_refine_only: false
+      num_cond_frames: 8
+```
+
+### Hunyuan 1.5 SR
+
+Hunyuan already behaves like an integrated multi-stage pipeline. Expose it via presets: `hunyuan15_sr_720p`, `hunyuan15_sr_1080p`. Users should not need to know the exact internal pipeline class split between base and SR stages. Per-stage override surface should stay small and mostly sampling-focused.
+
+Hunyuan15 presets (`hunyuan15_t2v_480p`, `hunyuan15_i2v_480p_distilled`, `hunyuan15_t2v_720p`, `hunyuan15_i2v_720p_distilled`, `hunyuan15_sr_1080p`) are implemented (PR 4). The `Hunyuan15_*_SamplingParam` subclasses have been removed; defaults (including precomputed sigmas) come from preset `defaults` dicts. Remaining work: adding typed `HunyuanSRStageOverride` classes and colocating PipelineConfig (PR 10).
+
+## Exact Compatibility Mapping
+
+Intended translation layer for common current fields.
+
+| Legacy Field | New Path |
+| --- | --- |
+| `model_path` | `generator.model_path` |
+| `revision` | `generator.revision` |
+| `trust_remote_code` | `generator.trust_remote_code` |
+| `workload_type` | `generator.pipeline.workload_type` |
+| `num_gpus` | `generator.engine.num_gpus` |
+| `tp_size` | `generator.engine.parallelism.tp_size` |
+| `sp_size` | `generator.engine.parallelism.sp_size` |
+| `dit_cpu_offload` | `generator.engine.offload.dit` |
+| `dit_layerwise_offload` | `generator.engine.offload.dit_layerwise` |
+| `text_encoder_cpu_offload` | `generator.engine.offload.text_encoder` |
+| `image_encoder_cpu_offload` | `generator.engine.offload.image_encoder` |
+| `vae_cpu_offload` | `generator.engine.offload.vae` |
+| `pin_cpu_memory` | `generator.engine.offload.pin_cpu_memory` |
+| `enable_torch_compile` | `generator.engine.compile.enabled` |
+| `torch_compile_kwargs` | split across `generator.engine.compile.backend`, `.fullgraph`, `.mode`, `.dynamic`; uncommon keys land in `.extras` |
+| `enable_torch_compile_text_encoder` | `generator.engine.compile.text_encoder_enabled` |
+| `enable_stage_verification` | `generator.engine.enable_stage_verification` |
+| `prompt_txt` | `request.inputs.prompt_path` |
+| `prompt` | `request.prompt` |
+| `negative_prompt` | `request.negative_prompt` |
+| `image_path` | `request.inputs.image_path` |
+| `video_path` | `request.inputs.video_path` |
+| `output_path` | `request.output.output_path` |
+| `output_video_name` | `request.output.output_video_name` |
+| `save_video` | `request.output.save_video` |
+| `return_frames` | `request.output.return_frames` |
+| `num_videos_per_prompt` | `request.sampling.num_videos_per_prompt` |
+| `seed` | `request.sampling.seed` |
+| `num_frames` | `request.sampling.num_frames` |
+| `height` | `request.sampling.height` |
+| `width` | `request.sampling.width` |
+| `fps` | `request.sampling.fps` |
+| `num_inference_steps` | `request.sampling.num_inference_steps` |
+| `guidance_scale` | `request.sampling.guidance_scale` |
+| `guidance_scale_2` | `request.sampling.guidance_scale_2` |
+| `guidance_rescale` | `request.sampling.guidance_rescale` |
+| `true_cfg_scale` | `request.sampling.true_cfg_scale` |
+| `boundary_ratio` | `request.sampling.boundary_ratio` |
+| `sigmas` | `request.sampling.sigmas` |
+| `enable_teacache` | `request.runtime.enable_teacache` |
+| `return_trajectory_latents` | `request.runtime.return_trajectory_latents` |
+| `return_trajectory_decoded` | `request.runtime.return_trajectory_decoded` |
+
+### Private Dreamverse Adapter Mapping
+
+The mappings below are useful for private Dreamverse migration, but they should not be treated as a public FastVideo backward-compatibility promise unless and until those fields actually exist in the public repo surfaces.
+
+| Private Adapter Field | New Path |
+| --- | --- |
+| `config_model_path` | `generator.pipeline.components.config_root` |
+| `ltx2_refine_enabled` | `generator.pipeline.preset_overrides.refine.enabled` |
+| `ltx2_refine_upsampler_path` | `generator.pipeline.components.upsampler_weights` |
+| `ltx2_refine_lora_path` | `generator.pipeline.components.lora_path` |
+| `ltx2_refine_num_inference_steps` | `request.stage_overrides.refine.num_inference_steps` |
+| `ltx2_refine_guidance_scale` | `request.stage_overrides.refine.guidance_scale` |
+| `ltx2_refine_add_noise` | `generator.pipeline.preset_overrides.refine.add_noise` |
+| `ltx2_image_crf` | `request.stage_overrides.refine.image_crf` |
+| `return_continuation_state` | `request.output.return_state` |
+
+### LongCat Legacy Mapping
+
+| Legacy Field | New Path |
+| --- | --- |
+| `refine_from` | `request.inputs.refine_from` |
+| `stage1_video` | `request.inputs.stage1_video` |
+| `t_thresh` | `request.stage_overrides.refine.t_thresh` |
+| `spatial_refine_only` | `request.stage_overrides.refine.spatial_refine_only` |
+| `num_cond_frames` | `request.stage_overrides.refine.num_cond_frames` |
+
+## Validation and Error Handling
+
+### Strict by Default
+
+All structured inputs should be strict by default: unknown keys error, wrong types error, invalid stage names error, incompatible state/preset combinations error.
+
+### Exceptions
+
+The only intentionally open-ended fields are `generator.pipeline.experimental` and `request.extensions`. These must be clearly documented as unstable and unsupported for long-term API compatibility.
+
+### Error Quality
+
+Validation errors should include the full nested path, expected type or valid choices, and preset/stage context when relevant:
+
+```text
+Invalid field: request.stage_overrides.refine.num_inference_steps
+Expected int, got "two"
+Preset: ltx2_two_stage
+Stage: refine
+```
+
+## Implementation Plan
+
+### Phases 0-5: Landed
+
+- Phase 0 — Schema Parity Inventory: inventory complete; field classifications live in `docs/design/inference_schema_parity_inventory.yaml`; parity test guard in `fastvideo/tests/api/test_schema_parity_inventory.py`.
+- Phase 1 — Shared Schema: `fastvideo/api/` with typed dataclasses, parser, validation, dotted overrides, `RunConfig`/`ServeConfig`.
+- Phase 2 — VideoGenerator Compat: `from_config`, `from_file`, `generate(request=...)`, legacy `from_pretrained(..., **kwargs)` and `generate_video(..., **kwargs)` as compat shims routed through typed normalization.
+- Phase 3 — CLI Refactor: `fastvideo generate` and `fastvideo serve` parse nested YAML/JSON with training-style dotted overrides; flat flag expansion removed as the canonical path.
+- Phase 4 — Preset System: shared registry + pipeline-local `presets.py` for all 13 families; all 12 `SamplingParam` subclasses removed; `SamplingParam` moved to `fastvideo/api/sampling_param.py`.
+- Phase 5 — Server Request Translation: `fastvideo serve` loads `ServeConfig`; stateless OpenAI endpoint clones `default_request` and merges validated user overrides.
+
+### Remaining Phases
+
+- **Phase 6 — LTX2 Public Upstream Path** (PR 6): upstream `ltx2_two_stage` preset; upstream continuation-state contract; upstream only repo-visible/public LTX2 surfaces into FastVideo.
+- **Phase 7 — Dreamverse Adapter Migration** (PR 7 + private repo work): translate private Dreamverse-only request/config fields in a private adapter; replace raw app-owned continuation kwargs with `state` in the private server; do not expand the public FastVideo compatibility promise just to match private adapter fields.
+- **Phase 7.5-7.10 — Streaming Server and Dynamo Contract** (PRs 7.5-7.10): upstream the streaming server (skeleton, GPU pool, prompt enhancer, auxiliaries, router) consuming `generate_async`; land the Dynamo backend contract (`VideoGenerator.generate_async`, health-check helper) with the Dynamo backend package itself living in the Dynamo repo.
+- **Phase 8 — Model Migration and Docs** (PRs 9-10, 12): colocate `configs/pipelines/<family>.py` with pipeline implementations; add typed stage override classes for multi-stage models; update basic examples to the new API; document YAML-first inference config and migration guidance.
+- **Phase 8.5 — Golden-Test Migration** (PR 11): keep SSIM/performance regression tests on legacy Python generation while preset defaults are still settling; one dedicated migration pass after the preset system and model-default behavior are stable; complete this migration before removing legacy Python inference entrypoints or kwargs.
+- **Phase 9 — Deprecation and Cleanup** (PR 13): deprecate direct public use of `FastVideoArgs`; deprecate direct public use of `SamplingParam`; gradually reduce public documentation for flat flags; eventually remove legacy kwargs after downstream migration is complete.
+
+## Final Recommendation
+
+The public FastVideo inference API is being rebuilt around:
+
+- typed nested configs
+- model-owned named presets
+- semantic stage overrides
+- first-class continuation state
+- YAML-first CLI with dotted overrides
+
+The primary abstraction is `InferencePreset`, not raw kwargs and not a fully manual stage graph.
+
+The repo is moving model-specific defaults closer to each pipeline, while keeping the public schema and parsing logic centralized.
+
+Regression and quality tests follow the rollout. Unit/entrypoint tests migrated to the typed API early, but SSIM/performance suites only move once the typed path can express all current knobs without compatibility exceptions and produces stable defaults through presets (PR 11).
+
+End state:
+
+- stable Python typing
+- clean YAML/JSON support
+- a much better CLI story
+- a sane path for Dreamverse/LTX2
+- a unified abstraction for LongCat, Hunyuan, and future multi-stage models

.agents/memory/dreamverse-integration/source-archive/dreamverse_integration.md

@@ -0,0 +1,285 @@
+# Dreamverse ↔ FastVideo Integration
+
+## Status
+
+Working integration record. Captures how Dreamverse consumes the
+FastVideo public API today, what's already shared, what's still ad
+hoc, and what migrations land alongside each PR in the API refactor
+sequence.
+
+Pinned versions (last reconciled this session):
+
+| Repo | Branch | Commit | Note |
+|---|---|---|---|
+| FastVideo (public) | `origin/main` | `70ee5d23` | PR 6 merged |
+| FastVideo (public) | `will/api_7` | `3de5f833` | PR 7 in flight (typed continuation state) |
+| FastVideo-internal | `will/rebase-nbv` | `1adc513e` | pre-PR-1 on the API refactor; has live realtime runtime |
+| Dreamverse | `master` | `dc500330` | uses local + remote FastVideo runtimes via `server/runtime/` |
+
+## Related Documents
+
+- [PR plan.md](../../PR%20plan.md) — PR-by-PR sequence for the API refactor
+- [apirefactor.md](../../apirefactor.md) — design spec
+- [streaming-server-upstream-plan.md](streaming-server-upstream-plan.md) — upstream plan for `ui/ltx2-streaming/server/`
+- `../../../Dreamverse/server/video_generation.py` — Dreamverse's worker + local `ContinuationState`
+- `../../../Dreamverse/server/runtime/{factory,backend,gpu_pool,interfaces}.py` — runtime abstraction
+- `../../../FastVideo-internal/fastvideo/entrypoints/realtime/{api_server,local_runtime}.py` — internal's realtime runtime (PR 7.5/7.6 upstream source)
+
+## Surface Area
+
+Dreamverse depends on FastVideo across three surfaces. Listed in order
+of how stable each is.
+
+### 1. Pipeline construction (stable)
+
+`Dreamverse/server/video_generation.py:VideoGenerationWorker` calls
+`VideoGenerator.from_pretrained(...)` with flat LTX-2 kwargs today.
+After PR 6 the typed `GeneratorConfig` path exists; Dreamverse can
+migrate at its own pace.
+
+| Dreamverse usage | FastVideo public surface (post-PR 6) |
+|---|---|
+| `VideoGenerator.from_pretrained(model_path, ltx2_refine_enabled=…, …)` | `VideoGenerator.from_pretrained(config=GeneratorConfig(...))` |
+| Flat `torch_compile_kwargs={…}` dict | `engine.compile.{backend,fullgraph,mode,dynamic,extras}` |
+| `ltx2_vae_tiling=True` | `pipeline.vae_tiling=True` |
+| `ltx2_refine_*` family | `pipeline.preset_overrides.refine.*` + `pipeline.components.upsampler_weights` |
+| `enable_torch_compile_text_encoder` | `engine.compile.text_encoder_enabled` |
+
+The legacy flat-kwarg path stays supported via `compat.py`; migration
+is opt-in. PR 13's deprecation warnings are the eventual nudge.
+
+### 2. Realtime runtime (in flight: PRs 7.5–7.6)
+
+`Dreamverse/server/runtime/factory.py` selects a runtime backend at
+process start:
+
+```python
+def create_runtime_pool() -> RuntimePool:
+    if os.getenv("FASTVIDEO_REALTIME_BASE_URL"):
+        return FastVideoRealtimePool(base_url=..., ws_url=..., default_model_id=...)
+    return GPUPool(get_available_gpus())  # in-process, wraps fastvideo.entrypoints.realtime.local_runtime
+```
+
+Both backends speak the same `RuntimePool` / `RuntimeSlot` Protocol
+(`server/runtime/interfaces.py`):
+
+- `acquire(client_id, websocket=None) -> (gpu_id, RuntimeSlot)`
+- `release(client_id)`
+- `RuntimeSlot.{join_user, user_step, leave_user, register_stream_queue, …}`
+
+Today both impls reach into FastVideo-internal's
+`fastvideo.entrypoints.realtime.local_runtime` (which exposes
+`RealtimeRuntimeConfig`, `GPUPool`, `GPUSlot`). The remote backend
+talks HTTP+WS to a separately-deployed runtime of the same shape.
+
+**Contract that PR 7.5/7.6 must preserve:**
+
+- `RealtimeRuntimeConfig` accepts `model_registry`, `default_model_id`,
+  `default_height/width/num_frames/fps/num_inference_steps/guidance_scale/seed/negative_prompt`,
+  `default_ltx2_image_crf`, `startup_warmup_{enabled,prompt,timeout_seconds}`.
+- `GPUPool(gpu_ids: list[int], config: RealtimeRuntimeConfig)` constructor.
+- `pool.initialize() / shutdown() / acquire() / release() / get_status()`.
+- HTTP endpoints on the remote variant: `GET /healthz`, `GET /readyz`,
+  `GET /status`, `WS /ws`. (These already match what
+  `Dreamverse/server/routes/health.py` consumes.)
+
+When PR 7.6 lands the upstream of `fastvideo/entrypoints/realtime/`,
+Dreamverse should not need any code change unless we rename the import
+path. **Open: do we rename `realtime/` → `streaming/` to match the
+public package introduced in PR 5.5?** A deprecation alias module
+keeps both working during transition.
+
+### 3. Continuation state (PR 7)
+
+`Dreamverse/server/video_generation.py:89 ContinuationState` is
+Dreamverse's hand-rolled per-session state holder. PR 7 introduces
+the typed equivalent at `fastvideo/pipelines/basic/ltx2/continuation.py`.
+
+#### Field mapping
+
+| Dreamverse | PR 7 `LTX2ContinuationState` | Notes |
+|---|---|---|
+| `video_images: list[PIL.Image]` | `video_frames: list[np.ndarray]` (uint8 H×W×3) | numpy is leaner; Dreamverse already round-trips PIL→numpy→PIL just to add noise |
+| `audio_latents: torch.Tensor` `[B, C, T, mel]` | `audio_latents: torch.Tensor` (safetensors-serialized; bf16-safe) | unchanged shape; safetensors preserves dtype incl. `bfloat16` |
+| `LTX2_VIDEO_CONDITIONING_FRAME_IDX` (env) | `video_conditioning_frame_idx: int` | env constant → per-state field |
+| `LTX2_VIDEO_CONDITIONING_STRENGTH` (env) | `video_conditioning_strength: float` | env constant → per-state field |
+| `AUDIO_CONDITIONING_NUM_FRAMES` (env) | `audio_conditioning_num_frames: int` | env constant → per-state field |
+| `AUDIO_CONDITIONING_STRENGTH` (env) | `audio_conditioning_strength: float` | env constant → per-state field |
+| `audio_lps` (passed into `apply_audio`) | `audio_sample_rate: int \| None` | analogous; rename worth confirming with audio team |
+| Computed `prefix_sec` per segment | `video_position_offset_sec: float` | **see open question below** |
+| `segment_idx` (param to apply_*) | `segment_index: int` | per-state field |
+| `VIDEO_CONTEXT_NOISE`, `AUDIO_CONTEXT_NOISE`, `ENABLE_AUDIO_COND` | not on state | runtime policy / regularization knobs, not portable session data |
+| `apply_video / apply_audio / save_video / save_audio_latents / clear` | not on PR-7 state class | state is a pure data carrier; runtime owns lifecycle policy |
+
+PR-7 is a strict superset of Dreamverse's data model **plus** lifts
+several env globals into per-session typed fields.
+
+#### Lifecycle mapping
+
+| Dreamverse pattern | `SessionStore` API |
+|---|---|
+| `self.continuation = ContinuationState()` per session | `state = session_store.snapshot(sid) or LTX2ContinuationState()` |
+| `apply_video(req_kwargs, segment_idx)` + `apply_audio(req_kwargs, segment_idx, audio_lps)` | `state = session_store.snapshot(sid)`; runtime builds request from `state.video_frames` / `state.audio_latents` etc. |
+| `save_video(frames)` + `save_audio_latents(latents)` | runtime constructs new `LTX2ContinuationState`, then `session_store.store(sid, new_state.to_continuation_state())` |
+| `clear()` at end of session | `session_store.drop(sid)` |
+
+`SessionStore` and `BlobStore` ABCs ship with thread-safe in-memory
+defaults (`InMemorySessionStore`, `InMemoryBlobStore`). Dreamverse can
+adopt them as-is for the local runtime; remote runtimes can plug in
+redis-backed implementations later.
+
+#### Wire format (HTTP/WS round-trip)
+
+Dreamverse's `FastVideoRealtimePool` already speaks the realtime
+runtime's HTTP+WS protocol. When PR 7.5/7.6 land state emission on
+the server side, the on-the-wire payload is the public envelope:
+
+```json
+{
+  "kind": "ltx2.v1",
+  "payload": {
+    "schema_version": 1,
+    "segment_index": 3,
+    "video_conditioning_frame_idx": 9,
+    "video_conditioning_strength": 0.75,
+    "audio_sample_rate": 24000,
+    "audio_conditioning_num_frames": 5,
+    "audio_conditioning_strength": 0.5,
+    "video_position_offset_sec": 0.2,
+    "video": {"frames_b64": ["..."]},
+    "audio": {"safetensors_b64": "..."},
+    "metadata": {}
+  }
+}
+```
+
+JSON-serializable end-to-end; safetensors blob preserves audio dtype
+(incl. bf16). For payloads above the inline threshold a `BlobStore`
+indirection replaces the b64-encoded body with `{"blob_id": "..."}`;
+the blob itself stays inside the runtime that produced it.
+
+## Migration Plan
+
+Per PR landed, Dreamverse adoption is opt-in.
+
+### After PR 7 merges
+
+Single-file change in Dreamverse, ~50-line PR:
+
+1. Replace `server/video_generation.py:89 ContinuationState` import
+   with `from fastvideo.pipelines.basic.ltx2.continuation import LTX2ContinuationState`.
+2. Move `apply_video`, `apply_audio`, `save_video`, `save_audio_latents`,
+   `clear` off the state class onto `VideoGenerationWorker` (these are
+   runtime policy that uses the state, not part of the state itself).
+3. Update `apply_audio` to read knobs from `state.audio_conditioning_num_frames`
+   and `state.audio_conditioning_strength` instead of the env globals
+   `AUDIO_CONDITIONING_NUM_FRAMES` / `AUDIO_CONDITIONING_STRENGTH`. The
+   env globals can stay as defaults that populate the state when a new
+   session starts.
+4. Same treatment for video knobs: `state.video_conditioning_frame_idx`,
+   `state.video_conditioning_strength`.
+5. Frame storage swaps `list[PIL.Image]` for `list[np.ndarray]` —
+   simpler `save_video` (no PIL conversion) and simpler `clear` (no
+   `.close()` loop).
+
+### After PR 7.5 lands streaming server skeleton
+
+Dreamverse's runtime/factory.py either:
+
+- Continues to construct `GPUPool` from `RealtimeRuntimeConfig` (the
+  current path), now backed by the upstreamed `fastvideo/entrypoints/realtime/`.
+- Or migrates to the upstream's `ServeConfig.streaming` shape and
+  invokes `fastvideo serve --config realtime.yaml` as the launch path.
+
+Either way, `Dreamverse/server/runtime/interfaces.py` `RuntimePool` /
+`RuntimeSlot` Protocol can stay in place — it was modeled after the
+realtime runtime's surface. No interface change needed.
+
+### After PR 7.6 lands the GPU pool upstream
+
+- The `local_runtime.py` import in
+  `Dreamverse/server/runtime/gpu_pool.py:24` becomes a public import
+  with the same symbols (`RealtimeRuntimeConfig`, `GPUPool`,
+  `get_available_gpus`).
+- Per-GPU continuation state inside the worker (`ltx2_continuation_images`,
+  `ltx2_continuation_audio_latents`) gets replaced by a `SessionStore`
+  reference. Dreamverse doesn't see this change — it's runtime-internal.
+- `request.state` / `result.state` round-trip starts working end-to-end
+  on the local runtime. Dreamverse's worker can begin reading
+  `result.state` and feeding `request.state` between segments.
+
+### After PR 7.10 lands the Dynamo backend contract
+
+- `VideoGenerator.generate_async(...) -> AsyncGenerator[VideoEvent, None]`
+  is the canonical API.
+- Dreamverse's per-segment `user_step` flow can migrate from the legacy
+  sync `generate_video(..., **kwargs)` path to consuming the typed
+  event stream. Optional; the sync wrapper stays.
+
+## Open Questions
+
+### `video_position_offset_sec` semantics
+
+Dreamverse computes `prefix_sec = float(audio_extra) / 24.0` per
+segment in `apply_audio`. Not persisted on `ContinuationState`.
+
+PR-7 has `video_position_offset_sec` as a **state field**. Two valid
+interpretations:
+
+(a) **Persistent across segments** — accumulating time offset for
+long sessions; useful for time-coherent audio chaining.
+(b) **Per-segment hint that rides on the carrier** — runtime
+overwrites every time; field is harmless redundancy.
+
+Field's docstring leans toward (b). Decide before PR 7.6 starts
+emitting/consuming it. If we land on (a), document the accumulation
+rule explicitly.
+
+### `BlobStore` / `SessionStore` lifecycle ownership
+
+PR 7's in-memory implementations have no eviction, no TTL, no
+automatic blob cleanup on state replacement. Documented as a
+per-deployment policy decision.
+
+When PR 7.5/7.6 land the live consumer, who owns:
+
+- bounded session capacity (LRU? TTL? hard max?)
+- blob `drop()` chained when a state is replaced
+- session expiry on websocket disconnect
+
+Probably the streaming server's session manager, but worth stating
+explicitly in PR 7.5's design.
+
+### `realtime/` vs `streaming/` package naming
+
+Currently:
+
+- Public PR 5.5 introduced `fastvideo/entrypoints/streaming/` (skeleton + typed config).
+- Internal has `fastvideo/entrypoints/realtime/` (live runtime).
+- Dreamverse imports from `fastvideo.entrypoints.realtime` (per the internal name).
+
+PR 7.5 either picks one or ships a deprecation alias module.
+Recommendation in `streaming-server-upstream-plan.md`: keep
+`streaming/` (it's the post-PR-5.5 public name), provide
+`realtime/__init__.py` as a re-export with a `DeprecationWarning` for
+one release cycle so internal/Dreamverse can land import updates.
+
+## Test Coverage on the FastVideo Side
+
+PR 7 ships:
+
+- `fastvideo/tests/api/test_ltx2_continuation.py` — typed
+  state round-trip (inline + blob), bf16 preservation, JSON
+  serializability, kind/version validation, schema_version guard.
+- `fastvideo/tests/entrypoints/streaming/test_session_store.py` —
+  store/snapshot/hydrate/drop behavior on `InMemorySessionStore`;
+  put/get/drop on `InMemoryBlobStore`; thread-safety of both.
+
+PR 7.5+ should add a contract test that exercises the round-trip via
+the same wire format Dreamverse's `FastVideoRealtimePool` consumes.
+
+## Changelog
+
+| Date | Change |
+|------|--------|
+| 2026-04-23 | Initial draft. Captures PR 6 / PR 7 mapping; open questions on `video_position_offset_sec`, lifecycle ownership, and `realtime/` vs `streaming/` naming. |

.agents/memory/dreamverse-integration/source-archive/dreamverse_review.md

@@ -0,0 +1,390 @@
+# Dreamverse Integration Review Log
+
+This document tracks design decisions, open questions, and integration-time
+choices made while landing the public-side stacked PRs (7.7 → 8) and switching
+Dreamverse from `FastVideo-internal` to public `FastVideo`. The user will
+review this carefully — entries are deliberately verbose about *why*.
+
+## Goal
+
+Replace Dreamverse's dependency on `FastVideo-internal` with the public
+`FastVideo` package, using the upstreamed streaming server stack
+(`fastvideo.entrypoints.streaming.*`) where Dreamverse currently has local
+copies or imports private modules.
+
+## Surfaces Dreamverse currently uses from FastVideo-internal
+
+(from `/home/william5lin/Dreamverse/server/`, scanned 2026-04-26):
+
+| Dreamverse import | Internal path | Public replacement |
+|---|---|---|
+| `fastvideo.entrypoints.realtime.local_runtime.RealtimeRuntimeConfig` | `FastVideo-internal/fastvideo/entrypoints/realtime/local_runtime.py` | (none) — Dreamverse rewires through `streaming.gpu_pool.SubprocessGpuPool` |
+| `fastvideo.entrypoints.realtime.local_runtime.GPUPool` | same as above | `fastvideo.entrypoints.streaming.gpu_pool.SubprocessGpuPool` (PR 7.6) |
+| `fastvideo.configs.pipelines.base.PipelineConfig` | already in public | unchanged |
+| `fastvideo.entrypoints.video_generator.VideoGenerator` | already in public | unchanged |
+| `fastvideo.layers.quantization.fp4_config.FP4Config` | already in public | unchanged |
+| `fastvideo.utils.maybe_download_model` | already in public | unchanged |
+| `fastvideo.models.audio.ltx2_audio_processing.AudioProcessor` | already in public | unchanged |
+| `fastvideo.models.loader.component_loader.ComponentLoader` | already in public | unchanged |
+| `fastvideo.models.dits.ltx2.*` | already in public | unchanged |
+| local copy: `Dreamverse/server/prompt_enhancer.py` (1933 lines) | mirrors `FastVideo-internal/.../prompt_enhancer.py` | `fastvideo.entrypoints.streaming.prompt.*` (PR 7.7) |
+| local copy: `Dreamverse/server/prompt_safety.py` | mirrors `FastVideo-internal/.../prompt_safety.py` | `fastvideo.entrypoints.streaming.prompt.safety` (PR 7.8) |
+| local copy: `Dreamverse/server/session_logger.py` | mirrors `FastVideo-internal/.../session_logger.py` | `fastvideo.entrypoints.streaming.session_logger` (PR 7.8) |
+| local copy: `Dreamverse/server/rewrite_prompt_payload.py` | mirrors `FastVideo-internal/.../rewrite_prompt_payload.py` | `fastvideo.entrypoints.streaming.prompt.rewrite` (PR 7.8) |
+| local copy: `Dreamverse/server/mock_server.py` (1200 lines) | mirrors `FastVideo-internal/.../mock_server.py` | `fastvideo.entrypoints.streaming.mock_server` (PR 7.8) |
+| local copy: `Dreamverse/server/session_init_image.py` | mirrors `FastVideo-internal/.../session_init_image.py` | `fastvideo.entrypoints.streaming.session_init_image` (PR 7.5 — already public) |
+
+## Design decisions made (auto-resolved)
+
+### D-1: Realtime runtime → streaming GpuPool migration shape
+
+**Context.** Dreamverse's `server/runtime/gpu_pool.py` thin-wraps
+`fastvideo.entrypoints.realtime.local_runtime.GPUPool`, which takes a
+`RealtimeRuntimeConfig(model_registry=…, default_model_id=…, default_height=…,
+default_width=…, default_num_frames=…, default_num_inference_steps=…,
+startup_warmup_*…)`. The public `streaming.gpu_pool.SubprocessGpuPool` takes a
+typed `GeneratorConfig` + `GpuPoolConfig` + `WarmupConfig`.
+
+The shapes differ in two important ways:
+
+1. The internal version had a multi-model registry (`model_id → model_config`
+   dict). The public version is single-model (one `GeneratorConfig`).
+2. The internal version flattened a few sampling defaults (height/width/frames/
+   steps) into the runtime config. The public version expects them as part of
+   the per-request `SamplingConfig`.
+
+**Decision.** Dreamverse will:
+1. Drop the multi-model registry on the integration branch (it is not used in
+   production today — Dreamverse boots one model per replica).
+2. Construct a `GeneratorConfig` for the chosen model from `MODEL_REGISTRY[id]`
+   and pass it to `SubprocessGpuPool`.
+3. Move the `default_height` / `default_width` / `default_num_frames` /
+   `default_num_inference_steps` defaults into a server-side
+   `default_request: GenerationRequest` template the session controller fills
+   from per-request input.
+
+**Why.** Multi-model is feasible to add back later (one pool per model id,
+acquire by `(session_id, model_id)`), but not on the migration branch — that
+would couple the upstream switch to a feature redesign. Punting keeps the
+upstream switch a pure mechanical refactor.
+
+**Risk.** If a Dreamverse code path silently relied on the registry to swap
+models per-session, the migration branch will surface that as a missing-model
+error. The integration tests must exercise at least one segment per supported
+model id before merging the Dreamverse branch.
+
+### D-2: PR 7.7 prompt enhancer API surface narrower than the internal one
+
+**Context.** The upstreamed `PromptEnhancer.enhance/auto_extend/rewrite` returns
+`LLMResponse(content, provider, model, latency_ms, fallback_used)`. The internal
+`enhance_prompt` / `generate_auto_prompt` / `rewrite_prompt_sequence` returns
+`EnhanceResult(prompt, fallback_used, error, provider, model, latency_ms)` /
+`RewriteResult(prompts, …, rollout_id, rollout_label, raw_response_text)`.
+
+**Decision.** The Dreamverse integration branch will adapt at the call site:
+- `enhancer.enhance_prompt(...)` → `enhancer.enhance(prompt)` + a thin shim
+  that maps the structured response into the existing `EnhanceResult` shape
+  for the session-controller code path. Move the shim to
+  `Dreamverse/server/prompting/_internal_compat.py`.
+- The locked-segment / next-segment-index plumbing the internal version
+  built into the user payload becomes Dreamverse-side template logic in
+  the shim.
+- The JSON-shaped responses the internal prompts assume (`{"next_prompt":
+  "..."}` / `{"segment_prompts": [...]}`) become Dreamverse-side
+  parsing in the shim, since the public `LLMResponse` is intentionally raw.
+
+**Why.** The public surface stays minimal and provider-agnostic; the
+LTX-2-specific orchestration (locked segments, rollout id/label, JSON
+schemas) is an internal-UI concern, not something every public consumer
+should wear. Dreamverse keeps its existing call shape; the public stays
+clean.
+
+**Open question for review:** Should we promote some of this into
+`fastvideo.entrypoints.streaming.prompt.ltx2_orchestration` (or similar)
+once a second consumer appears? Logging here so we have the option.
+
+### D-3: Multi-stage provider race (Dreamverse) vs sequential fallback (public)
+
+**Context.** The internal enhancer runs all providers in a stage in parallel
+and returns the first to succeed (`_run_provider_race`). The public
+enhancer runs providers strictly sequentially with retryable-error fallback.
+
+**Decision.** Public stays sequential for PR 7.7. The race-based fallback is
+a Dreamverse-specific tail-latency optimization that depends on parallel API
+budgets; promoting it would force every public consumer to have multiple
+provider keys configured. Dreamverse can keep `_run_provider_race` as an
+internal optimization on its side.
+
+**Risk.** First-segment latency on Dreamverse may regress slightly when
+Cerebras is having a bad minute (sequential fallback waits the full
+20s timeout before trying Groq). If this is a real production concern,
+add a public knob like `concurrency: int = 1` on `PromptEnhancer` that
+gates a race path — but only after measuring.
+
+### D-4: Skipping PR 7.9 router for the integration branch
+
+**Context.** The internal stack ships a `router/main.py` that load-balances
+across replicas with health checks. Dreamverse's deployment uses a single
+replica per region (per `gpu_pool.py:_parse_requested_gpu_limit`).
+
+**Decision.** Land PR 7.9 on the public side (so the surface is upstreamed)
+but skip wiring it into the Dreamverse integration branch. Dreamverse's
+`server/main.py` does not import from `router/`.
+
+### D-5: Audio re-encode (PR 7.10) needed for streaming, deferred
+
+**Context.** The internal streaming server's per-step path runs an audio
+re-encode (`_re_encode_audio` inside `_stream_av_fmp4_events` /
+`do_step_ltx2`) so each fMP4 segment ships with continuation-conditioning
+audio. The whole-segment `pool.run()` path the public streaming server
+currently uses doesn't need this. The PR plan defers re-encode integration
+to PR 7.10 (`generate_async` / per-step streaming).
+
+**Decision.** Land PR 7.10's `generate_async` on the public side. The
+Dreamverse integration branch initially keeps using `pool.run()` (whole
+segment, no re-encode); a follow-up branch swaps it to
+`generate_async` + audio re-encode once that path is exercised end-to-end.
+
+### D-6: `realtime/local_runtime.py` is *not* upstreamed
+
+**Context.** It is the FastVideo-internal precursor to `streaming.gpu_pool`.
+Upstreaming both would create two GPU pool implementations in the public
+repo.
+
+**Decision.** Don't upstream `realtime/local_runtime.py`. Dreamverse switches
+to `streaming.gpu_pool.SubprocessGpuPool` on the integration branch. The
+internal module can be deleted from FastVideo-internal at a follow-up.
+
+## Open questions for user review
+
+Each section below is a place the auto-decision could plausibly be wrong.
+Please flip / annotate these in review.
+
+### Q-1 Multi-model GPU pool (D-1)
+
+Does any current Dreamverse production flow load multiple model ids
+concurrently? If yes, we need to either (a) keep `realtime/local_runtime`
+alive on the internal side until the public side gains a multi-model pool,
+or (b) build the multi-model abstraction upstream as part of PR 7.6 follow-up
+work.
+
+### Q-2 Promoting LTX-2 prompt orchestration (D-2)
+
+The locked-segments / next-segment-index / JSON-response orchestration is
+LTX-2-specific. If Cosmos / Wan / Hunyuan ever grow a similar continuation
+flow, we'll regret keeping the orchestration on the consumer side. Worth
+promoting now?
+
+### Q-3 Race-based provider fallback (D-3)
+
+The sequential fallback in the public enhancer adds up to `timeout_ms` of
+extra latency per failing provider before the next is tried. For Dreamverse
+that's 20s. Should we land the race path now behind a `concurrency: int = 1`
+knob, or wait until we have data?
+
+### Q-4 Router upstream skip on Dreamverse branch (D-4)
+
+We're upstreaming PR 7.9 (router) but not consuming it in the Dreamverse
+integration branch. Is that right? Dreamverse currently has no router
+component, so the answer is probably yes — but flagging.
+
+### Q-5 generate_async cutover for the streaming path (D-5)
+
+The plan leaves Dreamverse using `pool.run` (whole segment) initially.
+Audio re-encode for cross-segment continuity is deferred to a follow-up.
+Is that acceptable for the first switch, or does Dreamverse audio quality
+regress relative to the internal path until 7.10 is wired in?
+
+## PR-by-PR execution log
+
+### PR 7.6 — already opened (#1257)
+
+`will/api_7.6` rebased onto `origin/main`, with subprocess-pool robustness
+review fixes pushed (boot_ok event, dead-worker detection, parallel shutdown,
+reader-exit pending-job cleanup). 17/17 gpu_pool tests + 89/89 streaming
+tests green at head.
+
+### PR 7.7 — already opened (#1258)
+
+`will/api_7.7` rebased onto the new 7.6 + LLM provider review fixes applied
+locally (per-instance `retryable`, 4xx-non-retryable, json-decode wrap,
+shared `_openai_compat.complete_openai_compatible`, `dataclasses.replace`
+for the fallback marker). 29/29 prompt tests + 120/120 streaming tests green.
+**Pending push** — the user opted to push this branch themselves.
+
+### PR 7.8 — rebased onto new 7.7
+
+`will/api_7.8` two commits replayed cleanly on the new 7.7. Adds
+`fastvideo/entrypoints/streaming/{prompt/safety,prompt/rewrite,session_logger,
+mock_server}.py` plus `test_auxiliaries.py`. 141/141 streaming tests green.
+
+Notable gap vs internal version: the public `PromptSafetyFilter` ships one
+classifier slot (`unsafe` label, single threshold) whereas the internal
+version chained an NSFW filter and a hate-speech filter with marker-based
+label matching. Multi-classifier composition is left to Dreamverse —
+operators chain two filters explicitly. See **D-7** below.
+
+### PR 7.9 — rebased onto new 7.8
+
+`will/api_7.9` three commits replayed cleanly. Adds streaming router
+(`router/{config,registry,main}.py`), `fastvideo router-serve` CLI
+subcommand, and `test_router.py`. 151/151 streaming tests green.
+
+Caveat: router/main.py uses the deprecated FastAPI `app.on_event("shutdown")`
+hook — emits a DeprecationWarning. Migration to lifespan handlers is a
+pre-merge cleanup item but not a blocker.
+
+### PR 7.10 — rebased onto new 7.9
+
+`will/api_7.10` three commits replayed with two trivial conflicts (line
+wrap in `server.py`, redundant test in `test_cli_translation.py`). Adds
+`VideoEvent` hierarchy, `VideoGenerator.generate_async`,
+`default_health_check_request`, plus `test_generate_async.py` (273-line
+contract test). 184/184 streaming + contract tests green.
+
+### PR 8 — rebased onto new 7.10
+
+`will/api_8` four commits → three (the 4th was a duplicate
+`streaming.md` doc that 7.5 already shipped, dropped during rebase).
+Adds `docs/design/server_contracts/{dynamo,index,openai}.md`,
+`mkdocs.yml` entries, and `fastvideo/tests/contract/test_{dreamverse,
+dynamo}_shape.py`. 206/206 streaming + contract tests green.
+
+### Dreamverse `will/integrate-public-fastvideo`
+
+Branch created from Dreamverse `master`. Single change: `pyproject.toml`
+swaps `fastvideo = { path = "../FastVideo-internal", editable = true }`
+to point at `../FastVideo`. Comment added linking back to this review
+doc.
+
+**Verified:** every TRACKED `from fastvideo.*` import in Dreamverse
+(`server/video_generation.py` only) resolves against the public
+package — except `fastvideo.layers.quantization.fp4_config.FP4Config`
+(see **D-7** / Q-6 below).
+
+**Untracked WIP** in `Dreamverse/server/{config,prompting,runtime,session}/`
+imports `fastvideo.entrypoints.realtime.local_runtime` (D-6); this
+branch does not migrate that WIP. The user's existing untracked work
+stays untouched and will need a separate follow-up to consume
+`streaming.gpu_pool.SubprocessGpuPool`.
+
+## Test ladder (built-up to e2e per user request)
+
+Each rung verifies the integration switch at one layer. Run from the
+narrowest to the broadest before running the full e2e against real
+GPU + model weights.
+
+| # | Layer | Command | Status against the switched stack |
+|---|---|---|---|
+| 1 | Public FastVideo unit + contract tests | `pytest fastvideo/tests/api/ fastvideo/tests/entrypoints/streaming/ fastvideo/tests/contract/` | 358/358 passing on `will/api_8` |
+| 2 | Public FastVideo FP4 lazy-import | `pytest fastvideo/tests/ops/quantization/test_fp4_config.py` | 3/3 passing |
+| 3 | Dreamverse Python tests | `cd Dreamverse && uv run pytest server/tests/ -k "not stress and not benchmark and not health_endpoint"` | 73/73 passing against public FastVideo |
+| 4 | Dreamverse FE unit/integration (vitest) | `cd Dreamverse/apps/web && npm test` | 54/86 passing — 32 failures are pre-existing copy-mismatches in `reducer.test.ts` etc., not caused by the switch |
+| 5 | Backend HTTP smoke (Playwright) | `cd Dreamverse/apps/web && PLAYWRIGHT_SKIP_WEBSERVER=1 PLAYWRIGHT_BASE_URL=http://127.0.0.1:8009 npx playwright test e2e/backend-health.spec.ts` | 4/4 passing (5th correctly skipped because devtools-only route is off) |
+| 6 | Frontend shell smoke (Playwright) | `npx playwright test e2e/frontend-shell.spec.ts` | Pending — requires Next.js dev server to be reachable; was stuck during this run, needs a clean restart |
+| 7 | Full e2e preset generation | `npx playwright test e2e/preset-prompt-generation.spec.ts` | **8/8 passing** end-to-end after restart with `CUDA_VISIBLE_DEVICES=4 ENABLE_TORCH_COMPILE=0 FASTVIDEO_GPU_COUNT=1 FASTVIDEO_ENABLE_DEVTOOLS=1`. BE warmup + GPU 4 idle slot let `/readyz` flip green; the spec verifies preset → WS → backend handshake → "Generating video…" state. |
+
+### How to reproduce e2e tier 7 from cold
+
+```
+# 1. BE — picks an idle GPU and skips torch.compile (avoids the
+#    aarch64 cross-compiler bug in the conda env's triton stack).
+cd ~/Dreamverse
+set -a; source ~/.env; set +a
+CUDA_VISIBLE_DEVICES=4 ENABLE_TORCH_COMPILE=0 \
+  FASTVIDEO_ENABLE_DEVTOOLS=1 FASTVIDEO_GPU_COUNT=1 \
+  uv run dreamverse-server &
+
+# 2. Wait for /readyz (~2 min for warmup x2 segments)
+until curl -fsS http://127.0.0.1:8009/readyz >/dev/null; do sleep 5; done
+
+# 3. FE
+cd ~/Dreamverse/apps/web
+BACKEND_URL=http://127.0.0.1:8009 NEXT_PUBLIC_INCLUDE_DEVTOOLS=1 \
+  npm run dev:devtools &
+
+# 4. Playwright
+cd ~/Dreamverse/apps/web
+PLAYWRIGHT_SKIP_WEBSERVER=1 \
+  PLAYWRIGHT_BASE_URL=http://127.0.0.1:5274 \
+  BACKEND_URL=http://127.0.0.1:8009 \
+  npx playwright test --project=chromium --reporter=list
+```
+
+### Surfaced during the e2e debug pass (logged here for follow-up)
+
+* **`SamplingParam has no field ltx2_image_crf`** — Dreamverse's
+  `server/video_generation.py:406` passes `ltx2_image_crf=0.0` to a
+  `SamplingParam(...)` constructor. The internal SamplingParam (in
+  `fastvideo/configs/sample/base.py`) declared this field; the public
+  `fastvideo.api.sampling_param.SamplingParam` does not. Currently
+  the BE logs an `ERROR` and silently drops the kwarg; warmup still
+  succeeds because the field is non-load-bearing for FP4-disabled
+  inference. Either re-add the field to the public schema or update
+  Dreamverse to stop passing it. **D-8.**
+
+* **`aarch64-conda-linux-gnu-cc` triton compile failure** — the conda
+  env we boot from injects an ARM cross-compiler ahead of `gcc` on
+  `$PATH`, so `torch._inductor`'s triton launcher fails compilation.
+  Setting `ENABLE_TORCH_COMPILE=0` bypasses it. Long-term fix: clean
+  the conda env's compiler shadowing or add a `CC=gcc` override in
+  Dreamverse's worker bootstrap. **D-9.**
+
+* **GPU pool starts but warmup OOMs on a shared GPU** — when
+  `CUDA_VISIBLE_DEVICES` lands on a GPU another tenant is using
+  (107 GiB-pegged training run on GPU 0 in this case), LTX-2 warmup
+  fails with OOM. Picking an idle GPU (4-7 here) is a manual step.
+  A pre-warm probe that checks free memory before booting the pool
+  would prevent this. **D-10.**
+
+* **ffmpeg fragment write `Broken pipe`** — when the WS client closes
+  before the backend finishes streaming the first segment, ffmpeg
+  hits `[Errno 32] Broken pipe`. Currently Dreamverse's
+  `gpu_pool.handle_command` re-raises this as a session error,
+  which then propagates to "User step failed". Cosmetic for now —
+  swallowing pipe-broken on intentional disconnect would clean up
+  the logs. **D-11.**
+
+## Additional integration gaps surfaced during the switch
+
+### D-7: `FP4Config` is private-only
+
+**Context.** `Dreamverse/server/video_generation.py:271` imports
+`fastvideo.layers.quantization.fp4_config.FP4Config` and assigns it to
+`pipeline_config.dit_config.quant_config`. The 411-line module lives only
+in `FastVideo-internal/fastvideo/layers/quantization/fp4_config.py` and
+hard-imports `flashinfer` at module top — it never made the public
+upstream pass. Public has `base_config.py` and `absmax_fp8.py` only.
+
+**Decision (provisional).** Don't upstream `fp4_config.py` in this
+session. Reasons:
+1. It introduces a new external dependency (`flashinfer`) the public
+   package has avoided so far.
+2. The class hard-codes LTX-2 layer paths
+   (`ltx2.blocks.{i}.attn1.to_q` etc.) — this is "LTX-2-specific FP4",
+   not generic FP4. Belongs colocated with `pipelines/basic/ltx2/` if
+   it goes anywhere.
+3. The FP4 pre-quantize/forward op surface is the kind of thing where
+   a careful review pass matters more than a bulk copy.
+
+**What this means for the integration branch.** Dreamverse will boot
+fine; only the FP4-quantized path inside `video_generation.py:283`
+will fail (lazy import). For workflows that don't enable FP4
+quantization, the integration is complete.
+
+### Q-6 (review): how to land FP4Config publicly?
+
+Two reasonable next steps:
+1. **Colocate.** Move FP4 code to `fastvideo/pipelines/basic/ltx2/quantization.py`
+   with `flashinfer` as an optional extra: `pip install fastvideo[fp4]`.
+   Refactor `FP4QuantizeMethod` to take its layer-prefix list from a
+   pipeline-config field instead of hardcoding ltx2 paths so the
+   approach generalizes.
+2. **Keep private.** Treat FP4 as a Dreamverse-side concern — Dreamverse
+   imports `fp4_config` from the internal repo via a thin shim. Public
+   FastVideo stays focused on generic surfaces. This means the
+   "FastVideo-internal removable" goal is partially undone.
+
+Recommendation: option 1 once the API refactor settles — wait until
+the LTX-2 colocation step (PR 9 / 10 territory) and land FP4 there.
+

.agents/memory/dreamverse-integration/source-archive/handoff-nvfp4-launch-demo.md

@@ -0,0 +1,518 @@
+# Handoff: LTX-2 NVFP4 wire-up + Dreamverse launch-demo skill
+
+This document hands off in-flight work to the next coding agent. It covers
+two related streams that landed across two repos:
+
+1. **FastVideo** (`will/ltx2_sr_port`): wire NVFP4 (NVIDIA's block-scaled
+   FP4) inference + per-component torch.compile + supporting parity fixes
+   so the public package matches `FastVideo-internal` for the LTX-2
+   distilled streaming path used by Dreamverse.
+2. **Dreamverse** (`will/integrate-public-fastvideo`): switch the GPU
+   worker to the typed `GeneratorConfig` API, rename `FP4Config` →
+   `NVFP4Config`, add a `launch-demo` skill + canonical
+   `serve_configs/streaming_demo.yaml` for `fastvideo serve --config`.
+
+Stack remains green: 222/222 FastVideo unit/contract/api tests pass; 8/8
+Playwright e2e tests pass against the live `dreamverse-server` + Next.js
+stack.
+
+---
+
+## Repo + branch state
+
+| Repo | Path | Branch | Tip |
+| --- | --- | --- | --- |
+| FastVideo | `/home/william5lin/FastVideo` | `will/ltx2_sr_port` | `c6c14c55` |
+| Dreamverse | `/home/william5lin/Dreamverse` | `will/integrate-public-fastvideo` | `3d7fd89` |
+| Reference (read-only) | `/home/william5lin/FastVideo-internal` | (their) `main` | source of truth for parity |
+
+> **Working branch on FastVideo is `will/ltx2_sr_port`, not the default checkout.**
+> The shell may report `will/uv-pip-install-everywhere` because that was
+> the earlier checkout. Run `git checkout will/ltx2_sr_port` before
+> picking up FastVideo work.
+
+### Live processes (do not duplicate)
+
+```
+:8009  dreamverse-server    pid 2453227   (warmed, /readyz returns 200)
+:5274  next-server (dev)    pid 2399103   (devtools build)
+```
+
+### Stashes
+
+* FastVideo: `stash@{0}: WIP on main: …HunyuanVideo plugin…` — pre-existing,
+  unrelated to this work, do not pop.
+* Dreamverse: `stash@{0}: wip: server modular refactor (split
+  config/prompting/runtime/session)` — 3867 lines of orphan modular split
+  off this branch. Do not pop on this branch; recover on a separate
+  feature branch if anyone wants to resurrect it.
+
+---
+
+## What landed (FastVideo: `cfccd292..c6c14c55`)
+
+Six commits on top of the i2v / continuation latent port:
+
+```
+c6c14c55 test(nvfp4): lock LTX-2 wiring + typed transformer_quant flow
+94c983a2 refactor(quant): rename FP4 → NVFP4 to disambiguate from other FP4 variants
+42b30bf9 feat(ltx2): wire FP4 inference through fastvideo.layers.quantization
+6da342ba feat(compile): per-component compile + transformer_refine + prepare hook
+221cb20a feat(api): typed per-component CompileConfig + FastVideoArgs carriers
+a4760bae fix(api): propagate generic refine_* args + match internal randn
+```
+
+Each commit message has the rationale. Highlights below.
+
+### `a4760bae` — three small parity fixes
+
+* `FastVideoArgs.__post_init__` now calls `_resolve_refine_args()` which
+  copies the public-facing generic `refine_*` knobs onto their
+  `ltx2_refine_*` runtime carriers. Was missing → callers that set
+  `refine_lora_path=...` saw "applied to 0 layers" warnings as the value
+  was silently dropped.
+* `_randn_ltx2_video_latents` patch path reverted from `randn_tensor` →
+  `torch.randn` to bit-match internal under single-generator inference.
+  Identical for a single `torch.Generator` but diverges for
+  `list[Generator]` (per-sample seeds).
+* Classified 19 `refine_*` / `ltx2_refine_*` / i2v / `ltx2_audio_*` /
+  `ltx2_conditioning_latent_*` / `ltx2_video_conditions` fields in the
+  schema-parity inventory yaml.
+
+### `221cb20a` — typed CompileConfig + FastVideoArgs carriers
+
+`CompileConfig` (in `fastvideo/api/schema.py`) gained per-component knobs:
+
+```python
+@dataclass
+class CompileConfig:
+    enabled: bool = False                 # master DiT switch
+    backend / fullgraph / mode / dynamic / extras  # master kwargs
+
+    # Per-component overlays, None = inherit master `enabled`
+    text_encoder_enabled: bool | None = None
+    vae_enabled: bool | None = None
+    audio_vae_enabled: bool | None = None
+
+    # Per-component kwargs override master when non-empty
+    dit_kwargs: dict = ...
+    text_encoder_kwargs: dict = ...
+    vae_kwargs: dict = ...
+    audio_vae_kwargs: dict = ...
+```
+
+Matching carrier fields on `FastVideoArgs`:
+`enable_torch_compile_text_encoder/vae/audio_vae` and
+`torch_compile_kwargs_dit/text_encoder/vae/audio_vae`. Compat layer
+round-trips them through `legacy_from_pretrained_to_config` and
+`generator_config_to_fastvideo_args`. **No behavior change yet** — these
+are surface ports only; consumed in the next commit.
+
+### `6da342ba` — refine + per-component compile + prepare_for_compile
+
+`composed_pipeline_base.post_init` now:
+
+* compiles `transformer_refine` alongside `transformer` and
+  `transformer_2` whenever the DiT compile flag is on (closes the LTX-2
+  stage-2 silent-eager bug);
+* dispatches per-component compile loops (text encoder, VAE, audio VAE)
+  with per-component kwargs falling back to master when empty;
+* calls `module.prepare_for_compile()` on each compiled submodule
+  before invoking `torch.compile` (hook protocol — model-specific).
+  Implemented on `Gemma3` to materialize HF weights outside Dynamo's
+  tracer.
+
+### `42b30bf9` — NVFP4 LTX-2 inference wire-up *(largest)*
+
+End-to-end:
+
+1. `models/dits/ltx2.py` — swap `nn.Linear` → `ReplicatedLinear` for the
+   FP4-eligible subset (`LTXSelfAttention`, `LTXDistributedSelfAttention`,
+   `FeedForward`/`GELUApprox`); plumb `quant_config` and `prefix=` from
+   `BasicAVTransformerBlock` → `_init_transformer_blocks` → `LTXModel`
+   → `LTX2Transformer3DModel`. Other linears
+   (`TimestepEmbedding`, `PixArtAlphaTextProjection`, `patchify_proj`,
+   `proj_out`, `AdaLayerNormSingle.linear`) stay `nn.Linear` —
+   matches internal exactly.
+2. Port `_supports_prequantized_input` and
+   `_linear_project_with_optional_prequant` helpers. Attention forward
+   pre-quantizes input once (`quantize_input`), reuses the
+   `(x_fp4, x_scale, x_global_sf)` tuple for k/v projections when
+   `context is x` — bit-matches internal's fused path.
+3. `models/loader/fsdp_load.py` — new `_maybe_convert_model_to_nvfp4`
+   helper detects via `isinstance(quant_method, NVFP4QuantizeMethod)`
+   (no flag); calls `convert_model_to_nvfp4` to materialize
+   `_nvfp4_weight*` / `_nvfp4_alpha` / `_weight_global_sf` buffers.
+   `flashinfer` import is lazy (inside the helper), so the loader is a
+   no-op on hosts without flashinfer.
+4. `layers/quantization/__init__.py` — registered `"NVFP4"` in
+   `QuantizationMethods` literal + `get_quantization_config`.
+5. `api/compat.py` + `fastvideo_args.py` — typed
+   `engine.quantization.transformer_quant: "NVFP4"` resolves to a
+   concrete `NVFP4Config()` instance, carried on `FastVideoArgs.transformer_quant`,
+   pinned onto `pipeline_config.dit_config.quant_config` in
+   `__post_init__._apply_transformer_quant`. **The explicit setter
+   (legacy mutation pattern) wins** if `dit_config.quant_config` is
+   already non-None.
+6. `layers/linear.py` — `LinearBase.__init__` now falls back to
+   `UnquantizedLinearMethod` when `quant_config.get_quant_method` returns
+   `None`. `NVFP4Config` only tags a curated subset of LTX-2 layers, and
+   the previous `assert quant_method is not None` would crash any
+   non-tagged layer that received a quant_config.
+
+### `94c983a2` — FP4 → NVFP4 rename
+
+NVIDIA's specific block-scaled fp4 format (e2m1 mantissa, fp32 alpha,
+`layout_128x4` scale layout, group size 16) — distinct from MX-FP4 /
+OCP-FP4 / generic e3m0. Mechanical rename, no behavior change:
+
+* `fp4_config.py` → `nvfp4_config.py`
+* `FP4Config` → `NVFP4Config`; `get_name()` returns `"nvfp4"`
+* `FP4QuantizeMethod` → `NVFP4QuantizeMethod`
+* `convert_model_to_fp4` → `convert_model_to_nvfp4`
+* `QuantizationMethods` literal: `"FP4"` → `"NVFP4"`
+* registered buffer names: `_fp4_weight`/`_fp4_alpha` →
+  `_nvfp4_weight`/`_nvfp4_alpha`
+* loader helper renamed
+* test file rename + symbol updates
+
+Internal-scope torch op namespace `fastvideo_fp4::*` and
+`_get_ltx2_fp4_stage_profile` deliberately left as-is — purely
+internal naming that mirrors FastVideo-internal.
+
+### `c6c14c55` — contract + numerical lock-in tests
+
+* `fastvideo/tests/ops/quantization/test_nvfp4_ltx2_wiring.py` (6 tests):
+  asserts that `LTXSelfAttention.to_q/to_k/to_v/to_out` are
+  `ReplicatedLinear`; `NVFP4Config()` attaches `NVFP4QuantizeMethod`
+  on the quantized subset with the correct `layer_prefix`; non-tagged
+  projections (cross-attn K/V, audio attn, audio FFN) fall back to
+  `UnquantizedLinearMethod`; `BasicAVTransformerBlock` propagates
+  `quant_config` and `prefix` correctly to all 4 attention modules +
+  FFN at once.
+* `fastvideo/tests/api/test_typed_quant_flow.py` (4 tests): asserts
+  typed `engine.quantization.transformer_quant: "NVFP4"` →
+  `NVFP4Config()` instance flow; default leaves `transformer_quant`
+  None; explicit `dit_config.quant_config = …` wins over typed carrier.
+
+---
+
+## What landed (Dreamverse: `248060b..3d7fd89`)
+
+Three commits on top of the e2e tier:
+
+```
+3d7fd89 feat(skill): launch-demo orchestrator + fastvideo serve YAML
+d80c2a8 refactor(server): drive FP4 + per-component compile via typed GeneratorConfig
+4cc6b30 chore: gitignore Playwright + Next.js build artifacts under apps/web
+```
+
+### `d80c2a8` — server/video_generation.py refactor
+
+Three coordinated changes in the GPU worker:
+
+* Replace legacy `load_kwargs` dict + `VideoGenerator.from_pretrained(model_root, **kwargs)`
+  call with the typed `GeneratorConfig` (`EngineConfig` /
+  `OffloadConfig` / `CompileConfig` / `PipelineSelection` /
+  `ComponentConfig`). Refine knobs move from `ltx2_refine_*` flat
+  kwargs into `preset_overrides["refine"]`. **The in-memory
+  `pipeline_config` pin** (`dit_config.quant_config = NVFP4Config()`)
+  keeps using the legacy `experimental["pipeline_config"]` carrier
+  because typed `transformer_quant: "NVFP4"` doesn't yet support
+  setting `layer_profile`.
+* Rename FP4 → NVFP4.
+* Re-enable `"mode": "max-autotune-no-cudagraphs"` (was commented out).
+  Closes the last known divergence vs FastVideo-internal in the
+  worker-level path trace.
+
+### `4cc6b30` — gitignore Playwright/Next.js artifacts
+
+Added `apps/web/{node_modules,.next,test-results,playwright-report}` to
+`.gitignore`. Mirror of the existing `prod-ui/` ignore set.
+
+### `3d7fd89` — launch-demo skill
+
+```
+.agents/skills/launch-demo/
+├── SKILL.md
+└── scripts/
+    ├── launch_demo.sh                 # orchestrator: BE + FE + health probes + Ctrl-C trap
+    ├── launch_backend_dreamverse.sh   # uv run dreamverse-server (default)
+    ├── launch_backend_fastvideo.sh    # uv run fastvideo serve --config (typed path)
+    └── launch_frontend.sh             # next dev (devtools/dev/single5s)
+serve_configs/
+└── streaming_demo.yaml                # canonical ServeConfig matching internal/ui
+```
+
+YAML has every field annotated with the internal source line it mirrors:
+LTX-2 distilled, NVFP4, 121 frames @ 1088×1920 24fps, 5 inference steps,
+2-step refine gs=1.0 add_noise=true, max-autotune-no-cudagraphs compile,
+121-frame default request, 300s session timeout, 6 segment cap, av_fmp4
+streaming, cinematic-drone warmup prompt, 2400s warmup timeout, 9
+conditioning frames + 0 end-offset, prompt enhancer on with cerebras /
+gpt-oss-120b / 20s timeout.
+
+**Two BE flavors documented in SKILL.md:**
+
+| `BE_FLAVOR=` | Boots | Routes served | FE compatible |
+| --- | --- | --- | --- |
+| `dreamverse` (default) | `dreamverse-server` | `/healthz`, `/readyz`, `/curated-presets`, `/v1/stream`, devtools, session monitor | ✓ full |
+| `fastvideo` | `fastvideo serve --config <yaml>` | `/health`, `/v1/stream` | ⚠ FE will surface fetch errors for `/curated-presets`, `/readyz` until those routes migrate into FastVideo's `build_app` |
+
+The fastvideo flavor exists today as the verifiable typed-config path
+(YAML parses, streaming worker boots, dotted overrides work). It is not
+yet a drop-in for the FE — see "Open follow-ups" below.
+
+---
+
+## Verified
+
+* `222 passed, 1 skipped` across `fastvideo/tests/api/`,
+  `fastvideo/tests/contract/`,
+  `fastvideo/tests/ops/quantization/test_nvfp4_*`,
+  `tests/local_tests/pipelines/test_ltx2_pipeline_smoke.py`.
+* `8 passed` Playwright e2e (backend-health 5, frontend-shell 2,
+  preset-prompt-generation 1) against the live `dreamverse-server`
+  + Next.js stack.
+* `streaming_demo.yaml` parses cleanly against `ServeConfig`; the
+  validation path of `fastvideo serve --config <yaml>` runs without
+  error and accepts dotted overrides like `--server.port 8010`.
+* FastVideo `bash -n` clean across all four launch scripts.
+
+---
+
+## Critical context (gotchas a successor should know)
+
+### NVFP4 layer set is asymmetric — by design
+
+`NVFP4Config.fp4_layers` covers:
+
+* `attn1.{to_q,to_k,to_v,to_out}` — full self-attention
+* `attn2.{to_q,to_out}` — cross-attn Q + out only (text context not quantized)
+* `audio_to_video_attn.{to_q,to_out}` — AV cross Q + out
+* `video_to_audio_attn.{to_k,to_v}` — VA cross K + V
+* `ffn.{fc_in,fc_out}` — video FFN
+* `adaln_single.linear` — but this is `nn.Linear` (not `LinearBase`),
+  so it never actually gets FP4'd. List entry has no effect; matches
+  internal.
+
+**NOT in the set:** audio self-attention (`audio_attn1.*`), audio
+cross-attention (`audio_attn2.*`), audio FFN (`audio.ffn.*`). Audio
+path is cheap enough that quant overhead isn't worth it. Test
+`test_basic_av_block_propagates_quant_config_to_all_children` locks
+this in — if you add audio quantization later, update the test.
+
+### `LinearBase` fallback is load-bearing
+
+`fastvideo/layers/linear.py:191-202`: when `quant_config.get_quant_method`
+returns `None` (layer not in the quant config's set), we fall back to
+`UnquantizedLinearMethod`. **Do not remove this fallback** — it would
+break every non-tagged `ReplicatedLinear` constructed with a
+`NVFP4Config`, and `assert quant_method is not None` in
+`ReplicatedLinear.__init__` would fire on unmatched layers.
+
+### Typed `transformer_quant` precedence
+
+`FastVideoArgs._apply_transformer_quant` only writes
+`dit_config.quant_config` when it's currently `None`. If a caller has
+explicitly set `pipeline_config.dit_config.quant_config = NVFP4Config(...)`,
+the explicit setter wins. Dreamverse's `video_generation.py` relies on
+this — it sets `NVFP4Config()` directly because the typed
+`transformer_quant: "NVFP4"` doesn't expose `layer_profile`.
+
+### Pre-existing AbsMaxFP8 test failure is NOT mine
+
+`fastvideo/tests/ops/quantization/test_absmax_fp8.py::test_create_weights_rejects_invalid_dtype`
+fails on `main` and on this branch with the same error
+("AssertionError not raised"). I confirmed via `git stash` that the
+failure pre-dates my changes. Not blocking; tracked as separate tech
+debt.
+
+### `transformer_refine` is auto-compiled with the master DiT flag
+
+Set `enable_torch_compile=True` and `transformer_refine` compiles
+along with `transformer` and `transformer_2`. There is **no separate
+`enable_torch_compile_refine` flag** — by design, refine inherits the
+DiT compile state to keep the typed surface small. If you need them
+decoupled, add a new field; don't repurpose existing ones.
+
+### `prepare_for_compile` is a duck-type protocol, not a base class method
+
+Defined nowhere; called via `getattr(module, "prepare_for_compile", None)`
+in `composed_pipeline_base._maybe_compile_pipeline_module`. Currently
+only Gemma implements it (to materialize HF weights outside Dynamo).
+Add to other models that have lazy external state if you observe
+compile-time graph breaks.
+
+### Public typed `PromptEnhancerConfig.provider` is `Literal["cerebras", "groq"]`
+
+Internal supports `"cerebras_ifm"` (config.py:143). The public typed
+schema does not. The `streaming_demo.yaml` defaults to `"cerebras"`.
+For agents that need `cerebras_ifm`, the `dreamverse-server` flavor
+respects the `FASTVIDEO_PROMPT_PROVIDER` env var (legacy path);
+`fastvideo serve --config` does not currently expose it.
+
+### Dreamverse `pipeline_config` is still a Python object passed via `experimental`
+
+The typed `GeneratorConfig` doesn't have a clean home for an
+in-memory `PipelineConfig` instance with mutated `dit_config`. We
+pass it via `pipeline.experimental["pipeline_config"]` — the
+`compat.py` legacy adapter recognizes that key and threads it through
+to `FastVideoArgs.from_kwargs`. This is fine but not pretty; if
+someone designs a typed `dit_config` carrier later, this becomes
+obsolete.
+
+### `fastvideo serve --config` is not yet a drop-in for the FE
+
+`fastvideo.entrypoints.streaming.server.build_app` exposes only
+`/health` and `/v1/stream`. The Dreamverse Next.js shell expects
+`/healthz`, `/readyz`, `/status`, `/curated-presets`,
+`/curated-presets/append`, `/prompt-system-config`, and the devtools
+routes. These all live in `Dreamverse/server/main.py` +
+`Dreamverse/server/routes/`. Until they migrate into FastVideo's
+`build_app` (or are exposed via a Dreamverse-side proxy), the
+`BE_FLAVOR=fastvideo` flavor is for verifying the typed serve config
+path only — not for full FE compatibility.
+
+---
+
+## Open follow-ups (prioritized)
+
+### High
+
+1. **Migrate FE-required routes into FastVideo's `build_app`.**
+   `/healthz`, `/readyz`, `/status` look obviously fastvideo-side
+   (they're streaming-server health). `/curated-presets` and
+   `/prompt-system-config` are operator-side surfaces and should
+   probably stay in Dreamverse (or migrate as opt-in routes that the
+   FE feature-detects). Without this, `BE_FLAVOR=fastvideo` is
+   permanently a "diagnostic" flavor. Closes the
+   `launch-demo` skill TODO.
+
+2. **AbsMaxFP8 test failure cleanup.** Pre-existing. Either fix the
+   test (`AbsMaxFP8LinearMethod.create_weights` no longer asserts on
+   invalid dtype — restore the assert if intentional, otherwise drop
+   the test).
+
+### Medium
+
+3. **Add `cerebras_ifm` to public `PromptEnhancerConfig.provider`
+   Literal.** Trivial schema change; needs paired enhancer-side
+   provider implementation in
+   `fastvideo/entrypoints/streaming/prompt/providers/`.
+
+4. **Expose `layer_profile` on typed `engine.quantization`.** Today
+   `transformer_quant: "NVFP4"` always constructs `NVFP4Config()`
+   with the default `layer_profile="refine"`. To support stage-1
+   profiles (no `attn2.to_out`, no cross-modal AV) via typed config,
+   add `transformer_quant_layer_profile: str | None = None` and
+   thread it through `compat.py`. Dreamverse currently dodges this
+   by setting `NVFP4Config()` directly via `experimental`.
+
+5. **Typed `dit_config.quant_config` carrier.** The
+   `experimental["pipeline_config"]` escape hatch in Dreamverse
+   should eventually become a typed field. Design TBD.
+
+### Low
+
+6. **Audio attention quantization profile.** If an audio-quant
+   profile is added to `NVFP4Config.fp4_layers` (currently audio attn
+   and FFN are bf16), update
+   `test_basic_av_block_propagates_quant_config_to_all_children`.
+
+7. **Schema parity inventory.** A few internal-only fields are not
+   exposed publicly (`PROMPT_HTTP_TIMEOUT_MS`,
+   `PROMPT_INITIAL_STAGE_TIMEOUT_MS`, `PROMPT_TEMPERATURE`,
+   `PROMPT_MAX_COMPLETION_TOKENS`, `PROMPT_AUTO_SLEEP_MS`,
+   `PROMPT_AUTO_TIMEOUT_MS`, the curated-presets file paths).
+   These all flow via env vars on `dreamverse-server` today; if
+   `fastvideo serve --config` becomes the canonical entrypoint,
+   they'll need typed homes.
+
+8. **Empty `apps/web/test-results/` directory locally.** The
+   `.gitignore` entry I added makes it invisible to `git status`,
+   but the dir itself still has a stale `.last-run.json` (45 bytes)
+   from a prior Playwright run. Harness blocked auto-cleanup
+   ("pre-existing files"); the user can `rm -rf
+   apps/web/test-results` whenever convenient.
+
+---
+
+## How to pick up work
+
+### Quick orientation (run these first)
+
+```bash
+# FastVideo state
+cd /home/william5lin/FastVideo
+git checkout will/ltx2_sr_port
+git log --oneline cfccd292..HEAD     # six commits added this round
+.venv/bin/python -m pytest fastvideo/tests/api/ \
+    fastvideo/tests/contract/ \
+    fastvideo/tests/ops/quantization/test_nvfp4_*.py \
+    tests/local_tests/pipelines/test_ltx2_pipeline_smoke.py \
+    -q --no-header   # expect 222 passed, 1 skipped
+
+# Dreamverse state
+cd /home/william5lin/Dreamverse
+git log --oneline 248060b..HEAD      # three commits added this round
+cat serve_configs/streaming_demo.yaml | head -40
+ls .agents/skills/launch-demo/
+
+# Live stack health (already running on this host)
+curl -s http://localhost:8009/readyz | head -c 200
+curl -s http://localhost:5274/ | head -c 100
+( cd apps/web && npx playwright test --reporter=line )   # expect 8 passed
+```
+
+### Reference docs
+
+* **FastVideo internal/ui parity source:** `../FastVideo-internal/ui/ltx2-streaming/server/config.py`
+* **NVFP4 source on internal:** `../FastVideo-internal/fastvideo/layers/quantization/fp4_config.py`
+* **Worker-trace audit:** `../FastVideo/dreamverse_review.md` (D-1
+  multi-model, D-5 audio re-encode, prior gap inventory)
+* **Schema parity inventory:** `docs/design/inference_schema_parity_inventory.yaml`
+* **PR-plan for the broader migration:** `../FastVideo/PR plan.md`
+
+### Files most likely to need touches in follow-ups
+
+* `fastvideo/api/schema.py` — `CompileConfig`, `QuantizationConfig`,
+  `PromptEnhancerConfig` Literal extension.
+* `fastvideo/api/compat.py` — typed → flat translation.
+* `fastvideo/fastvideo_args.py` — carrier fields and
+  `_apply_transformer_quant`.
+* `fastvideo/entrypoints/streaming/server.py::build_app` — add
+  `/healthz`, `/readyz`, `/status` routes for FE compatibility (high
+  priority follow-up #1).
+* `Dreamverse/server/video_generation.py` — typed `GeneratorConfig`
+  builder (current).
+* `Dreamverse/serve_configs/streaming_demo.yaml` — every parity
+  knob; edit here, not in shell scripts.
+
+---
+
+## Don't / Cautions
+
+* **Don't pop the Dreamverse stash on this branch.** It's 3867 lines
+  of orphan modular refactor (server/{config,prompting,runtime,session}/)
+  with broken absolute imports. If anyone wants to resurrect it, do so
+  on a separate feature branch.
+* **Don't remove the `LinearBase` `UnquantizedLinearMethod` fallback.**
+  See "Critical context" above.
+* **Don't repurpose `enable_torch_compile` to mean DiT-only.** It also
+  drives `transformer_refine` and `transformer_2` compile. Add a new
+  flag if decoupling is needed.
+* **Don't change `NVFP4Config` buffer names back to `_fp4_*`.** The
+  rename is intentional to disambiguate from MX-FP4 / OCP-FP4.
+* **Don't bypass the typed surface for new options.** New compile /
+  quant / refine knobs should land on the dataclass + compat.py +
+  parity inventory together. The existing test suite locks this in.
+* **Don't merge to main without a CI run that covers FP4.** Current
+  CI doesn't run flashinfer-dependent paths; the wiring tests in
+  `test_nvfp4_ltx2_wiring.py` are CPU-only by design and don't
+  exercise the actual FP4 kernels.
+
+---
+
+*Last updated: end of session that landed `c6c14c55` on FastVideo and
+`3d7fd89` on Dreamverse. Stack remains green; no dirty state.*

.agents/memory/dreamverse-integration/source-archive/streaming-server-upstream-plan.md

@@ -0,0 +1,539 @@
+# FastVideo Streaming Server Upstream — Design & Plan
+
+## Status
+Exploration / design draft. Captures the re-evaluation triggered by the
+decision to upstream `FastVideo-internal/ui/ltx2-streaming/server/` into
+the public repo. Not yet approved for execution.
+
+## Related Documents
+- [PR plan.md](../../PR%20plan.md) — PR-by-PR implementation plan for the API refactor
+- [apirefactor.md](../../apirefactor.md) — design spec this plan implements
+- `../../../FastVideo-internal/ui/ltx2-streaming/` — upstream source (server side)
+- `../../../dynamo/` — local clone of ai-dynamo/dynamo; backend patterns at
+  `components/src/dynamo/{vllm,sglang,trtllm}/` and `CLAUDE.md` files
+- https://github.com/ai-dynamo/dynamo/pull/7544 — draft PR that promotes
+  FastVideo to a native Dynamo backend (CLOSED, superseded — but establishes
+  the integration shape)
+
+## Context
+
+The internal `FastVideo-internal/ui/ltx2-streaming/` directory contains a
+complete LTX2 streaming service. The user has decided:
+
+- **Frontend clients** (`client/`, `prod-ui/`) stay in the internal repo
+- **Everything server-side** — FastAPI/WebSocket server, GPU pool, prompt
+  enhancer, router, auxiliaries — will be upstreamed to FastVideo
+
+In parallel, FastVideo is becoming a **first-class Dynamo backend** (same
+tier as vllm, sglang, trtllm). The refactor must produce an API that
+Dynamo's `components/src/dynamo/fastvideo/` package can consume as a
+pure Python import, without re-introducing the legacy flat-kwarg
+surface. Draft PR ai-dynamo/dynamo#7544 defines the concrete integration
+shape we need to support.
+
+This materially changes the tail of the API refactor plan. The current
+PR 5 ("wire `ServeConfig.default_request` into the OpenAI-compatible
+HTTP server") addresses only the stateless endpoint; the real upstream
+target is a much larger, session-based stack **plus** a clean Dynamo
+backend contract.
+
+This document captures:
+- what's being upstreamed and where it lands
+- four design decisions that shape the upstream (continuation model,
+  streaming server layout, LLM provider abstraction, Dynamo backend
+  integration)
+- a revised PR sequence for the tail of the refactor
+
+## What's being upstreamed
+
+| Internal path | Size | Role | Upstream target |
+|---|---|---|---|
+| `server/main.py` | 94KB | FastAPI + WebSocket, session lifecycle, segment orchestration | `fastvideo/entrypoints/streaming/server.py` + handlers |
+| `server/gpu_pool.py` | 66KB | GPU orchestration, subprocess workers | `fastvideo/entrypoints/streaming/gpu_pool.py` |
+| `server/prompt_enhancer.py` | 69KB | LLM orchestration (cerebras_ifm, cerebras, groq) | `fastvideo/entrypoints/streaming/prompt/` package |
+| `server/mock_server.py` | 45KB | Mock backend for dev/tests | `fastvideo/entrypoints/streaming/mock_server.py` |
+| `server/prompt_safety.py` | 7KB | Optional fasttext-gated prompt safety | `fastvideo/entrypoints/streaming/prompt/safety.py` |
+| `server/session_init_image.py` | 3KB | i2v init image handling | `fastvideo/entrypoints/streaming/session_init_image.py` |
+| `server/rewrite_prompt_payload.py` | 3KB | Rewrite flow payload builder | `fastvideo/entrypoints/streaming/prompt/rewrite.py` |
+| `server/session_logger.py` | 1KB | Session JSONL logs | `fastvideo/entrypoints/streaming/session_logger.py` |
+| `server/config.py` | 9KB | Env-driven server config | Typed `ServeConfig` extensions |
+| `router/main.py` | 27KB | Multi-replica load balancer + WS proxy | `fastvideo/entrypoints/streaming/router/` (or separate package) |
+| `slurm/` | — | Deployment scripts | Likely stays internal |
+
+## FastVideo contact surface today
+
+Direct calls from the internal stack into FastVideo, all in `gpu_pool.py`:
+
+| Location | Call | Notes |
+|---|---|---|
+| `gpu_pool.py:164` | `from fastvideo.entrypoints.video_generator import VideoGenerator` | Subprocess-level import, post-`CUDA_VISIBLE_DEVICES` setup |
+| `gpu_pool.py:230` | `PipelineConfig.from_pretrained(config_model_path)` | Direct access to legacy `PipelineConfig` |
+| `gpu_pool.py:231` | `pipeline_config.dit_config.quant_config = FP4Config()` | Direct internals mutation |
+| `gpu_pool.py:264-267` | `VideoGenerator.from_pretrained(model_root, **load_kwargs)` | Flat legacy kwargs |
+| `gpu_pool.py:837` | `generator.generate_video(**request_kwargs)` | Per-segment flat kwargs |
+| `gpu_pool.py:282-288` | `LTX2AudioEncoder`, `AudioProcessor`, `get_diffusers_config` | Audio re-encode path |
+
+`load_kwargs` at `gpu_pool.py:233-260` contains:
+`ltx2_refine_enabled`, `ltx2_refine_upsampler_path`, `ltx2_refine_lora_path`,
+`ltx2_refine_num_inference_steps`, `ltx2_refine_guidance_scale`,
+`ltx2_refine_add_noise`, `pipeline_config`, `torch_compile_kwargs`,
+`dit_cpu_offload`, `dit_layerwise_offload`, `vae_cpu_offload`,
+`text_encoder_cpu_offload`, `pin_cpu_memory`, `ltx2_vae_tiling`,
+`use_fsdp_inference`, `enable_torch_compile`.
+
+`request_kwargs` at `gpu_pool.py:837` includes:
+`ltx2_audio_clean_latent`, `ltx2_audio_denoise_mask`,
+`ltx2_video_conditions`, `video_position_offset_sec`, standard sampling
+fields.
+
+**Implication**: upstreaming `gpu_pool.py` as-is perpetuates the flat
+kwarg surface inside the public server. We need a typed translation
+(PR 6 expansion) at the worker boundary before, or as part of, the
+gpu_pool upstream.
+
+## Session / continuation semantics today
+
+Per-session state (in `server/main.py`):
+- `locked_segment_prompts`, `curated_prompts`, `segment_idx`,
+  `generated_segment_count`, `loop_iteration`
+
+Per-**GPU** (not per-session) continuation cache (in `gpu_pool.py`):
+- `ltx2_continuation_images` — last 9 decoded frames for clip conditioning
+- `ltx2_continuation_audio_latents` — denoised audio latents for audio conditioning
+
+Segment N+1 automatically conditions on segment N's trailing frames and
+audio. On session reset or handoff (`USER_JOIN`), the per-GPU cache is
+cleared. There is currently **no way for a client to serialize and
+resume continuation state elsewhere** — it lives on the GPU only.
+
+## Design Decision 1: Continuation model
+
+### Options
+
+**A. Opaque client-round-trip payload** (current plan PR 7 design)
+- Server returns `ContinuationState(kind, payload)`; client sends it back.
+- Pro: stateless server, trivially load-balanceable, survives disconnects.
+- Con: large payloads (frames + audio latents) over every request hop;
+  bandwidth heavy on multi-segment WebSocket sessions.
+
+**B. Server-held session state** (internal reality)
+- Continuation lives per-GPU; implicit between adjacent segments.
+- Pro: zero client bandwidth; fast; matches today.
+- Con: needs GPU affinity, no resume after disconnect, harder to scale horizontally.
+
+**C. Hybrid** (recommended)
+- Server-held is the default for streaming WebSocket sessions.
+- Server exposes a `snapshot_state` message that returns the opaque
+  payload form for migration/retry.
+- Stateless HTTP endpoints always use round-trip opaque payloads.
+- One serialization format underlies both surfaces.
+
+### Decision: **C (Hybrid)**
+
+Rationale: matches both internal streaming use (server-held, fast) and
+stateless API use (client-round-trip, resumable). Cost is one serialization
+layer that serves both.
+
+### Implications
+- `ContinuationState.kind` identifies the payload schema
+  (e.g. `"ltx2.v1"`).
+- `ContinuationState.payload` must cover:
+  - trailing conditioning frames (or a tensor reference)
+  - audio latents (or a tensor reference)
+  - segment index / rollout position
+  - any model-specific conditioning metadata (e.g. audio sample rate,
+    `video_position_offset_sec`)
+- For large tensors, payload may reference a server-side blob by ID
+  rather than inline everything.
+- Streaming server has a `SessionStore` keyed by session ID that holds
+  a typed `LTX2ContinuationState` object.
+- `SessionStore.snapshot(session_id) -> ContinuationState` serializes
+  the current state for export.
+- `SessionStore.hydrate(state: ContinuationState) -> session_id` loads
+  a state into a new session.
+- Plan PR 7 expands to cover both surfaces and define the payload schema.
+
+## Design Decision 2: Streaming server layout
+
+### Options
+
+- **A. `fastvideo/entrypoints/streaming/`** — parallel to
+  `fastvideo/entrypoints/openai/`
+- **B. `fastvideo/entrypoints/server/{stateless,streaming}/`** — reorg both
+- **C. `fastvideo/streaming/`** — top-level package, not under entrypoints
+
+### Decision: **A (parallel subpackage)**
+
+Rationale: lowest-friction, no existing code moves, both servers share
+the same `fastvideo/entrypoints/*` namespace and import style. Shared
+utilities can be factored into `fastvideo/entrypoints/server_common/`
+later if needed. Option B creates churn across every openai/ import for
+marginal organizational win.
+
+### Target layout
+
+```text
+fastvideo/entrypoints/
+├── openai/                        # existing: stateless HTTP POST
+│   ├── api_server.py
+│   ├── video_api.py
+│   ├── image_api.py
+│   ├── common_api.py
+│   ├── protocol.py
+│   ├── state.py
+│   ├── stores.py
+│   └── utils.py
+├── streaming/                     # NEW: session WebSocket
+│   ├── server.py                  # FastAPI + WebSocket entry
+│   ├── session.py                 # session lifecycle, state machine
+│   ├── session_store.py           # typed session state + snapshot/hydrate
+│   ├── protocol.py                # JSON WebSocket message schemas
+│   ├── stream.py                  # fMP4 encoding (av_fmp4 mode)
+│   ├── gpu_pool.py                # subprocess workers
+│   ├── worker.py                  # per-GPU worker loop
+│   ├── continuation.py            # typed LTX2 state payload
+│   ├── session_init_image.py
+│   ├── session_logger.py
+│   ├── mock_server.py
+│   ├── prompt/
+│   │   ├── enhancer.py            # provider-agnostic prompt ops
+│   │   ├── rewrite.py
+│   │   ├── safety.py              # optional fasttext
+│   │   ├── payload.py             # rewrite payload builder
+│   │   └── providers/
+│   │       ├── base.py            # LLMProvider protocol
+│   │       ├── cerebras.py
+│   │       ├── cerebras_ifm.py
+│   │       └── groq.py
+│   └── router/                    # or separate top-level package
+│       ├── main.py
+│       └── registry.py
+├── cli/                           # existing
+└── video_generator.py             # existing
+```
+
+### Config integration
+
+`ServeConfig` gets an optional `streaming: StreamingConfig | None` field:
+
+```python
+@dataclass
+class StreamingConfig:
+    session_timeout_seconds: int = 300
+    generation_segment_cap: int = 6
+    stream_mode: Literal["av_fmp4", "legacy_jpeg"] = "av_fmp4"
+    warmup: WarmupConfig = field(default_factory=WarmupConfig)
+    pool: GpuPoolConfig = field(default_factory=GpuPoolConfig)
+    prompt: PromptEnhancerConfig | None = None
+    safety: PromptSafetyConfig | None = None
+
+@dataclass
+class GpuPoolConfig:
+    num_workers: int | None = None  # default: CUDA_VISIBLE_DEVICES count
+    enable_audio_reencode: bool = True
+    conditioning_num_frames: int = 9
+    conditioning_end_offset: int = 0
+
+@dataclass
+class PromptEnhancerConfig:
+    provider: Literal["cerebras_ifm", "cerebras", "groq"] = "cerebras_ifm"
+    model: str = "gpt-oss-120b"
+    timeout_ms: int = 20000
+    system_prompt_dir: str | None = None  # hot-reloadable system prompts
+
+@dataclass
+class PromptSafetyConfig:
+    enabled: bool = False
+    classifier_path: str | None = None
+```
+
+## Design Decision 3: LLM provider abstraction
+
+### Problem
+
+`prompt_enhancer.py` (69KB) hard-codes three providers (cerebras_ifm,
+cerebras, groq) with provider-specific request/response handling
+scattered throughout. Upstreaming as-is locks FastVideo to those three
+providers and couples the prompt operations to their response shapes.
+
+### Shape
+
+Introduce an `LLMProvider` protocol:
+
+```python
+from typing import Protocol, AsyncIterator, Literal
+from dataclasses import dataclass
+
+@dataclass
+class LLMMessage:
+    role: Literal["system", "user", "assistant"]
+    content: str
+
+@dataclass
+class LLMRequest:
+    messages: list[LLMMessage]
+    model: str
+    max_tokens: int | None = None
+    temperature: float | None = None
+    timeout_ms: int | None = None
+
+@dataclass
+class LLMResponse:
+    content: str
+    provider: str
+    model: str
+    latency_ms: float
+    fallback_used: bool = False
+
+class LLMProvider(Protocol):
+    name: str
+    async def complete(self, request: LLMRequest) -> LLMResponse: ...
+```
+
+### Decision: **Protocol + built-in implementations for cerebras, cerebras_ifm, groq**
+
+Rationale: keeps the prompt enhancer free of provider-specific branching;
+users (and future OpenAI/Anthropic/local additions) can register their
+own provider without modifying FastVideo. Each built-in provider is
+100-200 LOC; the enhancer becomes provider-agnostic prompt orchestration.
+
+### Implications
+- `prompt_enhancer.py` splits into `enhancer.py` (prompt operations) +
+  `providers/` (IO).
+- Config moves from scattered env vars to typed `PromptEnhancerConfig`
+  under `ServeConfig.streaming.prompt`.
+- Hot-reloadable system prompts stay — exposed as a management endpoint
+  on the streaming server.
+- Fallback behavior (retry across providers in priority order) moves
+  into the enhancer layer, orthogonal to provider implementations.
+
+## Design Decision 4 preamble: what Dynamo expects from FastVideo
+
+Dynamo's backend pattern (observed in
+`dynamo/components/src/dynamo/sglang/` and confirmed by PR #7544) is a
+**pure Python import** pattern. Dynamo owns the backend subpackage in its
+own repo; FastVideo only needs to expose a stable, typed, aggregated
+and (later) streaming generation surface.
+
+### Contract surface Dynamo consumes
+
+| Surface | Shape | Notes |
+|---|---|---|
+| Constructor | `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` | Already exists; `typed_kwargs` must be a stable subset from `GeneratorConfig` — no flat LTX2 legacy kwargs. |
+| Sync execution | `generator.generate_video(request: GenerationRequest) -> VideoResult` | Aggregated mode; Dynamo wraps in `asyncio.to_thread` under an `asyncio.Lock`. |
+| Async execution | `generator.generate_async(request: GenerationRequest) -> AsyncGenerator[VideoEvent, None]` | Needed for: (a) streaming server fMP4 chunks; (b) future Dynamo disaggregation. Events: `Progress`, `Partial?`, `Final`. |
+| Typed request | `fastvideo.api.GenerationRequest`, `SamplingConfig`, `InputConfig` | Stable import path; Dynamo's adapter builds this from `NvCreateVideoRequest` + `VideoNvExt`. |
+| Typed result | `VideoResult` with `video_bytes` or tensor frames, plus `ContinuationState?` | Must be picklable / JSON-serializable enough for Dynamo RPC. |
+| Continuation | `ContinuationState(kind, payload)` with schema-versioned payloads | Used by FastVideo's session store today; tomorrow by Dynamo disaggregated workers. |
+| Health check input | `VideoGenerator.default_health_check_request() -> GenerationRequest` | Minimal 256x256 / 8 frames / 1 step; lets Dynamo's `FastVideoHealthCheckPayload.to_dict()` produce the Dynamo `health_check_payload` kwarg without knowledge of FastVideo internals. |
+| Config dump | `GeneratorConfig.to_dict()` / `ServeConfig.to_dict()` | Dynamo calls `dynamo.common.config_dump.dump_config(path, config)` at worker start; we already have `config_to_dict()`. |
+
+### Request/response mapping (Dynamo ↔ FastVideo)
+
+Dynamo's video protocol (`NvCreateVideoRequest` / `NvVideosResponse`):
+
+```
+NvCreateVideoRequest          ->  fastvideo.api.GenerationRequest
+  prompt                      ->    sampling.prompt
+  size="WxH"                  ->    sampling.width, sampling.height
+  seconds                     ->    (seconds * nvext.fps) -> sampling.num_frames
+  input_reference             ->    input.image_path / input.video_path
+  nvext.fps                   ->    sampling.fps
+  nvext.num_frames            ->    sampling.num_frames (overrides seconds*fps)
+  nvext.num_inference_steps   ->    sampling.num_inference_steps
+  nvext.guidance_scale        ->    sampling.guidance_scale
+  nvext.seed                  ->    sampling.seed
+  nvext.negative_prompt       ->    sampling.negative_prompt
+  response_format             ->    (handled by adapter at output)
+
+VideoFinalEvent               ->  NvVideosResponse
+  video_bytes                 ->    data[0].b64_json  (if response_format=b64_json)
+  video_url (after upload)    ->    data[0].url       (if response_format=url)
+  metadata.inference_time_s   ->    inference_time_s
+```
+
+All fields already exist (or will exist after PR 6 expansion) on
+FastVideo's typed schema. No FastVideo changes required beyond what the
+rest of this plan already covers **except**:
+
+1. `generate_async` must exist (new in PR 7.10).
+2. `default_health_check_request()` helper (new in PR 7.10).
+3. The sync `generate_video(request=...)` path must be reachable without
+   extra wrapping (exists since PR 2; confirm stability).
+
+### Where the Dynamo subpackage lives
+
+The Dynamo-side integration (`FastVideoHandler`, `register_fastvideo_model`,
+`FastVideoHealthCheckPayload`, args parsing, main.py, Dockerfile,
+request/response mapping) lives **entirely in the Dynamo repo** at
+`components/src/dynamo/fastvideo/`, matching the pattern used by vllm
+and sglang. FastVideo does **not** host any Dynamo-related subpackage,
+Dynamo dependency, or Dynamo-specific CLI. FastVideo's only obligation
+is to expose a clean, stable, typed Python API that Dynamo's backend
+package can import.
+
+## Design Decision 4: Dynamo as first-class backend target
+
+### Problem
+
+PR #7544 (closed) shows two frictions with the pre-refactor API:
+
+1. **Flat legacy kwargs** — the Dynamo handler had to know about
+   LTX2-specific flat names.
+2. **Sync-only generation** — Dynamo's async handler wrapped
+   `generator.generate(...)` in `asyncio.to_thread` under a lock; no
+   progress streaming, no disaggregation path.
+
+The refactor's stateless OpenAI server, WebSocket streaming server, and
+Dynamo backend all want the same thing: **a typed async API that yields
+progress events and a typed final result**. If we build it once in
+`VideoGenerator`, all three adapters become thin.
+
+### Options
+
+**A. Keep sync-only, each adapter wraps**
+- Simple; matches PR #7544.
+- Con: streaming server needs its own async runner; Dynamo loses progress
+  streaming; no path to disaggregation.
+
+**B. Add async event stream to `VideoGenerator`**
+- `generate_async(request) -> AsyncGenerator[VideoEvent, None]`.
+- Sync `generate_video` becomes a thin `asyncio.run` wrapper internally.
+- Pro: one canonical execution API; streaming server, OpenAI server,
+  and Dynamo all consume events directly.
+- Con: larger delta in `VideoGenerator` — must thread async through the
+  pipeline step loop.
+
+**C. Queue-based `generate(request, event_cb)` callback**
+- Middle ground; callback receives events.
+- Pro: no async rewrite needed.
+- Con: callers have to invert control; awkward for Dynamo's async
+  handler.
+
+### Decision: **B (async event stream)**
+
+Rationale: one substrate serves all three consumers. The cost is a
+`generate_async` implementation that runs the pipeline step loop in a
+thread and bridges events back via an asyncio queue — standard pattern,
+limited surface area.
+
+### Implications
+
+- New PR 7.10 adds `generate_async` on `VideoGenerator` with three event
+  types: `VideoProgressEvent(step, total_steps, stage)`,
+  `VideoPartialEvent(frames_ndarray, index)` (optional; emitted only in
+  the streaming path), `VideoFinalEvent(video_bytes_or_tensor, metadata,
+  continuation_state?)`.
+- Sync `generate_video(request=...)` becomes `asyncio.run(...)` over
+  `generate_async`, collecting events and returning the final.
+- Streaming server's fMP4 encoder consumes `VideoPartialEvent` frames
+  directly, never re-decoding through disk.
+- Dynamo adapter consumes `generate_async` and yields one
+  `NvVideosResponse` per `VideoFinalEvent` (aggregated mode; ignores
+  intermediate events today; can surface progress via Dynamo's
+  status/progress fields in the future).
+- `ContinuationState` can be attached to `VideoFinalEvent.metadata`,
+  giving Dynamo a first-class way to surface state for disaggregation
+  later.
+- Stable public exports: `from fastvideo import VideoGenerator`;
+  `from fastvideo.api import GenerationRequest, SamplingConfig,
+  ContinuationState, VideoResult, VideoEvent`.
+- No Dynamo subpackage, dep, or CLI lives in FastVideo. The adapter
+  (`NvCreateVideoRequest ↔ GenerationRequest` mapping, handler,
+  registration) lives entirely in the Dynamo repo at
+  `components/src/dynamo/fastvideo/`.
+
+### Constraints this adds to earlier PRs
+
+- **PR 6** (typed LTX2 kwargs): every flat kwarg must have a typed home
+  **reachable from `GeneratorConfig`**, so Dynamo can construct the
+  generator without importing internal compat paths.
+- **PR 7** (continuation state): `ContinuationState.payload` must be
+  JSON/YAML serializable (no raw torch tensors inline; use blob
+  indirection) so it survives Dynamo RPC transport.
+- **PR 7.5** (streaming skeleton): consume `generate_async` rather than
+  re-implementing a progress loop around `generate_video`.
+- **PR 2/3/4 already landed**: the typed request shape is fixed and
+  matches Dynamo's mapping needs — no backtracking required.
+
+## Revised PR sequence (PR 5 onwards)
+
+PRs 0-4 are unchanged and already landed. PR 5 is narrowed; PRs 5.5-7.9
+are new inserts; PRs 8-13 are reshaped or kept.
+
+| # | Title | Change | Key deliverables |
+|---|---|---|---|
+| **5** | Stateless `ServeConfig.default_request` merge | **Narrowed.** Wire typed default-request into `fastvideo/entrypoints/openai/`. | `_merge_default_request` helper, validated-against-preset, tests for default+user-override precedence |
+| **5.5** | Server architecture split | **NEW.** Introduce `fastvideo/entrypoints/streaming/` subpackage skeleton. No behavior change. | Empty subpackage + stub server.py; CLI subcommand `fastvideo streaming-serve` (raises NotImplementedError); doc on layout |
+| **6** | LTX2 public preset + stage overrides + config colocation | **Expanded.** Also add typed replacements for every flat kwarg used by internal `gpu_pool.py`. | `ltx2_two_stage` preset, `LTX2RefineStageOverride`, `CompileConfig` field types, typed `FP4Config` integration, colocation |
+| **7** | Continuation state (public + session) | **Expanded.** Define both opaque payload AND server-held session store. | `ContinuationState.payload` schema, `LTX2ContinuationState` typed subclass, `SessionStore` interface, snapshot/hydrate APIs |
+| **7.5** | Streaming server skeleton | **NEW.** Minimum viable WebSocket server: session lifecycle, JSON messages, fMP4 output, single-generator. | `server.py`, `session.py`, `protocol.py`, `stream.py` (fMP4), typed `StreamingConfig` |
+| **7.6** | GPU pool upstream | **NEW.** Upstream `gpu_pool.py` with typed config boundary. | `gpu_pool.py`, `worker.py`, job queue, session-to-GPU binding, session timeout handling |
+| **7.7** | Prompt enhancer upstream | **NEW.** Upstream `prompt_enhancer.py` with `LLMProvider` abstraction. | `prompt/enhancer.py`, `prompt/providers/{base,cerebras,cerebras_ifm,groq}.py`, hot-reloadable system prompts |
+| **7.8** | Streaming auxiliaries | **NEW.** Small, isolated. | `prompt/safety.py`, `session_init_image.py`, `prompt/rewrite.py`, `session_logger.py`, `mock_server.py` |
+| **7.9** | Router upstream | **NEW.** Multi-replica load balancer + WS proxy. | `streaming/router/` (or separate top-level package), health checks, WS proxy |
+| **7.10** | Dynamo backend contract | **NEW.** Add `VideoGenerator.generate_async` event stream + `default_health_check_request()` helper. FastVideo exposes the async API only; the Dynamo backend package (handler, adapter, registration) lives entirely in the Dynamo repo at `components/src/dynamo/fastvideo/`. Streaming server (PR 7.5) and Dynamo backend both consume the same async API. | `generate_async` with `VideoProgressEvent`/`VideoPartialEvent`/`VideoFinalEvent`; sync `generate_video` becomes a thin wrapper; contract tests against a mock Dynamo-style handler that imports only public FastVideo APIs |
+| **8** | Internal-UI ↔ public-server contract docs & tests | **Reframed.** Was "Dreamverse Server Adaptation Layer." Also covers Dynamo integration reference. | WebSocket protocol reference, contract tests, migration examples, Dynamo adapter example that upstream PR can copy verbatim |
+| **9** | LongCat preset migration + colocation | **Keep.** | Stage overrides, colocation |
+| **10** | Hunyuan15 SR preset migration + colocation | **Keep.** | Stage overrides, SR field migration POC, colocation |
+| **11** | SSIM / perf test migration | **Keep.** Now blocked on PR 6 expansion. | Typed API migration of golden tests |
+| **12** | Docs + examples | **Keep, expand.** | Streaming server docs now part of scope |
+| **13** | Deprecation + cleanup | **Keep, expand.** | Also deprecate flat kwargs that internal gpu_pool uses today |
+
+Total PR count: 13 → ~20 (13 original + 5 streaming-upstream inserts +
+1 architecture split + 1 Dynamo contract). Each new PR is small and
+self-contained because the streaming components are already cleanly
+separated in the internal repo, and the Dynamo contract rides on top of
+the async API that the streaming server already needs.
+
+## Open questions
+
+1. **Router: in-repo or separate package?** — It's orthogonal to inference;
+   in-repo couples deploy cycles, separate leaves FastVideo cleaner.
+   Recommendation: separate package `fastvideo-router/` or
+   `fastvideo/contrib/router/`; defer final call to PR 7.9.
+2. **Session ID authority** — internal uses ad-hoc client IDs.
+   Recommendation: server-generated UUID, accept externally provided
+   session ID only for resume flows.
+3. **Torch compile kwargs typing** — `CompileConfig.kwargs: dict[str, Any]`
+   today accepts `mode`, `backend`, `fullgraph`, `dynamic`. Options: keep
+   as opaque dict; fully type; hybrid (type the common four + allow
+   extras). Recommendation: hybrid, type common fields.
+4. **Prompt safety / fasttext dependency** — heavy for users who don't
+   need it. Recommendation: ship as optional extra
+   `pip install fastvideo[prompt-safety]`.
+5. **Audio-specific tensor payloads** — `ltx2_audio_clean_latent`,
+   `ltx2_audio_denoise_mask`, `ltx2_audio_latents` are not in the current
+   public schema. PR 7 should classify them (probably as opaque fields
+   inside `LTX2ContinuationState.payload`, not top-level sampling fields).
+6. **Batching behavior** — internal `test_batching.py` suggests batching
+   is exercised. Scope this into PR 7.5 or defer to a post-cleanup perf PR?
+7. ~~**Dynamo subpackage home**~~ — **Resolved.** No Dynamo code lives
+   in FastVideo. The full backend package (handler, adapter,
+   registration, health check) is owned by the Dynamo repo at
+   `components/src/dynamo/fastvideo/`, same pattern as vllm/sglang.
+   FastVideo only guarantees the public API contract listed above.
+8. **Disaggregation readiness** — PR #7544 is aggregated-only. Our
+   `ContinuationState` hybrid already supports a future prefill/decode
+   split (prefill yields state; decode hydrates it). Should PR 7.10
+   explicitly validate that `ContinuationState` survives round-trip
+   through a Dynamo-style RPC (pickle or JSON), even though Dynamo
+   isn't using it today? Recommendation: yes; cheap contract test that
+   prevents drift.
+9. **Dynamo progress/status passthrough** — `NvVideosResponse` has
+   `status` and `progress` fields. Should PR 7.10's handler contract
+   emit intermediate `NvVideosResponse` chunks keyed off
+   `VideoProgressEvent`, or stay aggregated-final-only to match PR
+   #7544? Recommendation: stay aggregated-final for PR 7.10; revisit
+   after Dynamo clarifies their streaming/progress semantics.
+
+## Immediate path forward
+
+1. Land `will/api_5` cleanup commits — **done** (`e03ca7d9`, `41f93179`
+   force-pushed without Claude co-author).
+2. Review this plan with a human — commit the doc to capture the state.
+3. Execute PR 5 (narrow stateless merge) and PR 5.5 (subpackage split)
+   in parallel. Both small; both unblock the streaming upstream that
+   follows.
+4. Start PR 6 expansion (typed replacements for flat LTX2 kwargs) as the
+   critical path for PR 7.6 (gpu_pool upstream).

.agents/memory/dreamverse-integration/source-archive/video-generator-config-api-design.md

@@ -0,0 +1,93 @@
+# Exploration Log: Video Generator Config API Design
+
+## Status: draft
+
+## Context
+FastVideo's Python inference API currently mixes generator-instance settings,
+pipeline initialization settings, and per-request sampling/runtime settings
+through broad `**kwargs` surfaces on `VideoGenerator.from_pretrained(...)` and
+`VideoGenerator.generate_video(...)`.
+
+This exploration compares the current FastVideo design with
+`sglang/multimodal_gen` and examines how to upstream multi-stage LTX2 /
+Dreamverse behavior without growing more ad hoc top-level flags.
+
+## Progress
+- [x] Read FastVideo onboarding, codebase map, and relevant design docs.
+- [x] Inspect current FastVideo generator, args, sampling, registry, and
+  workflow abstractions.
+- [x] Inspect internal LTX2 streaming server usage and current two-stage /
+  continuation requirements.
+- [x] Inspect SGL diffusion generator, server args, sampling params, and
+  request preparation boundary.
+- [x] Inspect vLLM-Omni stage config, stage metadata, request, and orchestration
+  surfaces for multi-stage pipeline ideas.
+- [x] Inspect current FastVideo CLI/config-file loading and compare with the
+  training YAML-only entrypoint.
+- [ ] Convert findings into a concrete implementation plan for FastVideo.
+
+## Findings
+- FastVideo already has the right internal separation points:
+  `FastVideoArgs`, `PipelineConfig`, `SamplingParam`, and `ForwardBatch`.
+- The public boundary is the unstable part:
+  init-time and request-time knobs are mixed through `**kwargs`.
+- Unknown init keys can be silently filtered, while unknown request keys can be
+  only logged rather than rejected. This makes API drift hard to detect.
+- SGL's split is cleaner:
+  `ServerArgs` for engine/runtime, `PipelineConfig` for model-family wiring,
+  and `SamplingParams` for per-request settings.
+- SGL also has better merge semantics for user request overrides:
+  it preserves model defaults, tracks explicitly provided fields, and validates
+  request params against pipeline task type.
+- SGL still has a design smell worth avoiding in FastVideo:
+  `SamplingParams._adjust(...)` depends on `ServerArgs`, which leaks
+  engine/pipeline concerns back into the request object.
+- vLLM-Omni contributes a useful extra abstraction beyond SGL:
+  model-owned multi-stage topology via `ModelPipeline` and `StageConfig`,
+  with per-stage defaults (`default_sampling_params`) and runtime override
+  layering.
+- vLLM-Omni's best reusable idea for FastVideo is not the serving stack, but
+  the separation between:
+  1. model-defined stage topology and per-stage defaults,
+  2. runtime engine overrides,
+  3. request-time sampling/state handoff.
+- vLLM-Omni also shows the downside of exposing stage-indexed request lists too
+  directly: `sampling_params_list` works for a serving engine, but is too
+  positional and low-level for FastVideo's higher-level Python API.
+- FastVideo already supports YAML/JSON config files for inference CLI, but the
+  current mechanism flattens nested documents back into argparse flags. This
+  preserves backward compatibility but keeps the CLI surface as the canonical
+  schema instead of a typed document model.
+- The training stack has a cleaner precedent: a YAML-first config loaded into a
+  typed schema, with dotted CLI overrides applied onto the nested document
+  before parsing. Inference can likely adopt a lighter variant of that pattern.
+- Multi-stage generation should be unified at the orchestration layer, not by
+  forcing LongCat refine, Hunyuan SR, and LTX2 continuation into one leaf config.
+
+## Mistakes / Dead Ends
+- A fully free-form string-dict API would lose too much type safety and would
+  likely recreate the current drift problem under a different shape.
+- A single universal `RefineConfig` for all models would become a sparse bag of
+  nullable fields and would not map cleanly to existing model families.
+
+## Proposed Standardization
+- Introduce a typed public split:
+  `GeneratorConfig` for instance-lifetime engine/init settings and
+  `GenerationRequest` for per-call inputs/sampling/output.
+- Allow dict input only as an interchange layer that is parsed immediately into
+  typed configs with strict unknown-key validation.
+- Add a typed `GenerationPlan` / multi-stage orchestration layer with
+  discriminated stage configs:
+  `SampleStageConfig`, `LongCatRefineStageConfig`,
+  `HunyuanSRStageConfig`, `LTX2ContinuationStageConfig`.
+- Let model families own stage defaults and stage topology through named
+  profiles or model-defined stage plans, similar in spirit to vLLM-Omni's
+  pipeline YAMLs, but expose them through typed Python config objects rather
+  than raw stage-indexed lists in the primary API.
+- Make YAML/JSON a first-class serialization of the same typed inference
+  schema, not just a file format that expands into CLI flags.
+- Prefer a YAML-first CLI pattern for nested configs:
+  `fastvideo generate --config run.yaml --request.sampling.seed 42`,
+  while keeping a compatibility layer for existing flat flags during migration.
+- Upstream LTX2 two-stage / continuation behavior as a first-class stage or
+  pipeline profile rather than more `ltx2_*` top-level kwargs.

.agents/memory/dreamverse-integration/state.md

@@ -0,0 +1,194 @@
+# Current State — 2026-05-05 (strategy reversal — single mega-PR #1288)
+
+Point-in-time snapshot of branches, commits, and live infrastructure.
+Update whenever commits land or services restart.
+
+For HOW to commit / push / verify see [runbook.md](runbook.md). For
+roster of co-authors to credit on every commit see
+[authors.md](authors.md).
+
+## Branch tips
+
+| Repo | Branch | Tip | Distance |
+|---|---|---|---|
+| FastVideo | `will/ltx2_sr_port` (**PR #1288 head**) | `b36bdbc9` | 36 commits ahead of `origin/main`; OPEN, MERGEABLE; +integration-review.md (Part 1 drift + Part 2 tradeoffs + recommended Option D) |
+| FastVideo | `will/api_7.10` | `6ae7a99f` | **deprecated** — PR #1287 closed in favor of #1288. Branch can be deleted on origin and locally; kept for now as historical reference. |
+| FastVideo | `will/api_8`, `will/ltx2_sr_runtime`, `will/ltx2_nvfp4`, `will/ltx2_post_fixes`, `will/agents_cleanup` | (various) | **deprecated** split bookmarks. Strategy reversed to single mega-PR (D-17). Safe to delete locally; not pushed to origin. |
+| FastVideo | `will/ltx2_sr_port-pre-1286-rebase` | `1baa60bb` | **local-only safety backup** of pre-rebase chain (37 commits); keep until next slice merges |
+| Dreamverse | `will/integrate-public-fastvideo` | `ec8ef92` | 10 commits ahead of `737f3c1` (the dep switch) |
+| FastVideo-internal | their `main` | (read-only ref) | — |
+
+FastVideo worktree default branch is `will/ltx2_sr_port`. Other agents
+share this worktree — if `git branch --show-current` shows something
+else, switch back cleanly with `git checkout will/ltx2_sr_port` (don't
+disturb their uncommitted work). I observed this happen repeatedly in
+the 2026-05-05 session — confirmed harmless; switching back was always
+safe with a clean working tree.
+
+## Post-#1286 rebase summary
+
+PR #1286 merged at `2aaeee2a` (squash). `will/ltx2_sr_port` was rebased
+onto new `origin/main`, dropping 4 commits whose content is now in main:
+
+- `cd76cf51` `[feat] streaming: router (multi-replica load balancer)`
+- `1ac1e732` `[feat] streaming: fastvideo router-serve CLI`
+- `b0b7f59c` `[test] streaming: router registry + health loop ...`
+- `40e265b8` `[fix] streaming: router polish — bridge cancel + state
+  machine + deps` (squashed into `2aaeee2a` via cherry-pick `a152cb77`)
+
+Rebase was clean — no conflicts. All 33 surviving commits got new SHAs
+(rebase rewrites). The pre-rebase tip `1baa60bb` is preserved on the
+local backup branch `will/ltx2_sr_port-pre-1286-rebase`.
+
+## New linearized chain (33 commits, slice indices for STACK.md)
+
+| Slice | PR | Commits | Tip SHA | Subject |
+|---|---|---|---|---|
+| 1-3 | 7.10 (PR #1287) | 3 | `6ae7a99f` | `[test] streaming: generate_async coverage + refreshed streaming test` |
+| 4-6 | 8 | 3 | `f32e31ec` | `[test] streaming: contract tests for Dreamverse + Dynamo shapes` |
+| 7-15 | LTX-2 SR | 9 | `e7297519` | `feat(ltx2): full i2v conditioning + continuation latent port` |
+| 16-21 | NVFP4 | 6 | `6793166b` | `test(nvfp4): lock LTX-2 wiring + typed transformer_quant flow` |
+| 22-23 | LTX-2 post-fixes | 2 | `25897b67` | `[fix]: unwrap list-of-generator before torch.randn in LTX-2 latent prep` |
+| 24-33 | agents_cleanup | 10 | `b34d9704` | `[docs] dreamverse-integration: add runbook + fresh-context onboarding` |
+
+5 PRs landed (7.5, 7.6, 7.7, 7.8, 7.9), 1 in flight (7.10), 5 remaining
+(8 / LTX-2 SR / NVFP4 / post-fixes / agents_cleanup).
+
+## Historical commit chain analysis (pre-#1286 rebase)
+
+The layered chain analysis below documented the pre-rebase SHAs (LTX-2
+SR layer, NVFP4 layer, post-handoff fixes layer). Those SHAs no longer
+exist on `will/ltx2_sr_port` — they live only on
+`will/ltx2_sr_port-pre-1286-rebase`. Content semantics are unchanged;
+SHAs were rewritten by the rebase. Kept here for narrative continuity.
+
+## FastVideo: commit chain `cfccd292..156103b9`
+
+Three layers since LTX-2 i2v port:
+
+### Layer 1 — LTX-2 SR port + alignment harness (5 commits)
+
+```
+365a66c7  feat(quantization): upstream LTX-2 FP4Config with lazy flashinfer
+433d26b2  feat(ltx2): port LTX-2 SR runtime — upsampler, refine stages, refine args
+751d05de  feat(ltx2): wire SR pipeline graph + port denoising/latent-prep stages
+af6bbfea  test(ltx2-sr): add numerical alignment harness — public vs internal
+974cd430  fix(ltx2-sr): close port gaps surfaced by alignment harness retries
+b6ac7630  test(ltx2-sr): pin ltx2 sampling knobs in harness for parity diff
+b043d550  fix(api): align public SamplingParam ltx2 defaults with distilled
+663dda80  fix(registry): order LTX-2 detectors so distilled wins for distilled paths
+cfccd292  feat(ltx2): full i2v conditioning + continuation latent port  (BASE)
+```
+
+(Predates the May 2 handoff.)
+
+### Layer 2 — NVFP4 wire-up + per-component compile (6 commits, May 2 handoff)
+
+```
+a4760bae  fix(api): propagate generic refine_* args + match internal randn
+221cb20a  feat(api): typed per-component CompileConfig + FastVideoArgs carriers
+6da342ba  feat(compile): per-component compile + transformer_refine + prepare hook
+42b30bf9  feat(ltx2): wire FP4 inference through fastvideo.layers.quantization
+94c983a2  refactor(quant): rename FP4 → NVFP4 to disambiguate from other FP4 variants
+c6c14c55  test(nvfp4): lock LTX-2 wiring + typed transformer_quant flow
+```
+
+See [quantization.md](quantization.md) for what each commit locks in.
+
+### Layer 3 — Post-handoff parity/perf fixes (3 commits, since May 2)
+
+```
+a5fcd19c  [fix]: lazy-import flash_attn 2 fallback in attention backend
+d4ee5be2  [fix]: avoid model.to() round-trip in Gemma encoder forward
+156103b9  [fix]: unwrap list-of-generator before torch.randn in LTX-2 latent prep  (HEAD)
+```
+
+Three small fixes — no new features. Continued parity tightening with internal.
+
+## Dreamverse: commit chain `737f3c1..ec8ef92`
+
+```
+737f3c1   chore: switch fastvideo dep from FastVideo-internal to public FastVideo
+4cc6b30   chore: gitignore Playwright + Next.js build artifacts under apps/web
+33caa92   test(e2e): align Playwright specs with the actual production composer
+6fd137c   test(e2e): tighten frontend-shell + preset specs to match actual UI
+248060b   test(e2e): add Playwright tier with backend-health smoke + preset run
+d80c2a8   refactor(server): drive FP4 + per-component compile via typed GeneratorConfig
+3d7fd89   feat(skill): launch-demo orchestrator + fastvideo serve YAML
+72f69b9   Update ffmpeg installation instructions.
+1ba5635   fix(server): block startup on GPU warmup readiness, propagate failures
+ec8ef92   fix(server): detect worker death in _send_command via proc.sentinel  (HEAD)
+```
+
+The post-handoff trio (`72f69b9`, `1ba5635`, `ec8ef92`) hardens server
+startup robustness — ffmpeg install docs, GPU warmup readiness gate, and
+worker-death detection.
+
+## Live services (do not duplicate)
+
+| Port | Service | PID | Status |
+|---|---|---|---|
+| 8009 | `dreamverse-server` | 2453227 | `/readyz` returns 200, 1 warmed GPU worker, queue 0 |
+| 5274 | `next-server` (dev) | 2399103 | 200, ~13.6 KB shell |
+| 8000 | unknown FastAPI | — | **Not in handoff.** Probably stray `fastvideo serve`. Verify with `lsof -i :8000` before launching a new BE on the default port. |
+
+## Stashes — DO NOT POP
+
+| Repo | Stash | Reason |
+|---|---|---|
+| FastVideo | `stash@{0}: WIP on main: 71bfc13d HunyuanVideo plugin` | Pre-existing, unrelated to integration work |
+| Dreamverse | `stash@{0}: wip: server modular refactor (split config/prompting/runtime/session)` | 3867-line orphan modular split, **not part of `will/integrate-public-fastvideo`**. Recover on a separate branch if needed. |
+
+## Test status (from May 2 handoff, not re-verified post-Layer-3)
+
+| Suite | Status |
+|---|---|
+| FastVideo `fastvideo/tests/api/` + `contract/` + `nvfp4_*` + `ltx2_pipeline_smoke` | 222 passed, 1 skipped |
+| Playwright e2e against live BE+FE | 8 passed (5 backend-health + 2 frontend-shell + 1 preset-prompt-generation) |
+| `fastvideo serve --config streaming_demo.yaml` validation | parses cleanly; dotted overrides work |
+| `bash -n` on launch-demo skill scripts | clean across all 4 |
+
+The 3 post-handoff commits are small parity fixes; full re-verification is
+recommended but not required to read this state.
+
+## Pre-existing failures (NOT caused by this work)
+
+| Test | Failure | Notes |
+|---|---|---|
+| `fastvideo/tests/ops/quantization/test_absmax_fp8.py::test_create_weights_rejects_invalid_dtype` | `AssertionError not raised` | Pre-existing on `main`. Verified via `git stash` that NVFP4 work doesn't introduce it. See [open-threads.md](open-threads.md) item #2. |
+
+## Source docs (archived 2026-05-03)
+
+The 7 source docs that this memory dir consolidates have been moved into
+[`source-archive/`](source-archive/) — see the
+[archive README](source-archive/README.md) for the archive policy and
+synthesis mapping.
+
+Other untracked items at the FastVideo repo root:
+- Nested clones: `dynamo/`, `ray/`, `vllm-omni/`
+- Lock files: `uv.lock`, `fastvideo/tests/ssim/.reference_videos_download.lock`
+- Skill dirs: `.agents/skills/diagnose-ssim-failure/`, `.agents/skills/review-pr-link/`
+- `.agents/exploration/pr-link-review.md` (kept; already promoted to a skill)
+
+## Quick orientation commands
+
+```bash
+# FastVideo state
+cd /home/william5lin/FastVideo
+git log --oneline cfccd292..HEAD     # 14 commits this round
+
+# Dreamverse state
+cd /home/william5lin/Dreamverse
+git log --oneline 737f3c1..HEAD      # 10 commits this round
+
+# Live stack health (already running)
+curl -s http://localhost:8009/readyz | head -c 300
+curl -s http://localhost:5274/ -o /dev/null -w "%{http_code}\n"
+
+# Re-verify test suite
+.venv/bin/python -m pytest fastvideo/tests/api/ \
+    fastvideo/tests/contract/ \
+    fastvideo/tests/ops/quantization/test_nvfp4_*.py \
+    tests/local_tests/pipelines/test_ltx2_pipeline_smoke.py \
+    -q --no-header
+```

.agents/memory/dreamverse-integration/streaming-server.md

@@ -0,0 +1,389 @@
+# Streaming Server Upstream — PRs 5.5 → 7.10
+
+The `FastVideo-internal/ui/ltx2-streaming/server/` stack is being
+upstreamed into public FastVideo at `fastvideo/entrypoints/streaming/`.
+In parallel, FastVideo is becoming a first-class Dynamo backend (same
+tier as vllm, sglang, trtllm). This file covers both threads since they
+share `generate_async` as the substrate.
+
+For PR sequence/status see [pr-roadmap.md](pr-roadmap.md). For the
+Dreamverse-side adoption see [cross-repo-surfaces.md](cross-repo-surfaces.md).
+
+**Last updated:** 2026-05-03.
+
+## What's being upstreamed
+
+| Internal path | Size | Role | Public target |
+|---|---|---|---|
+| `server/main.py` | 94 KB | FastAPI + WebSocket, session lifecycle, segment orchestration | `fastvideo/entrypoints/streaming/server.py` + handlers |
+| `server/gpu_pool.py` | 66 KB | GPU orchestration, subprocess workers | `fastvideo/entrypoints/streaming/gpu_pool.py` |
+| `server/prompt_enhancer.py` | 69 KB | LLM orchestration (cerebras_ifm, cerebras, groq) | `fastvideo/entrypoints/streaming/prompt/` package |
+| `server/mock_server.py` | 45 KB | Mock backend for dev/tests | `fastvideo/entrypoints/streaming/mock_server.py` |
+| `server/prompt_safety.py` | 7 KB | Optional fasttext-gated prompt safety | `prompt/safety.py` |
+| `server/session_init_image.py` | 3 KB | i2v init image handling | `streaming/session_init_image.py` (PR 7.5, already public) |
+| `server/rewrite_prompt_payload.py` | 3 KB | Rewrite flow payload builder | `prompt/rewrite.py` |
+| `server/session_logger.py` | 1 KB | Session JSONL logs | `streaming/session_logger.py` |
+| `server/config.py` | 9 KB | Env-driven server config | typed `ServeConfig.streaming` extensions |
+| `router/main.py` | 27 KB | Multi-replica load balancer + WS proxy | `fastvideo/entrypoints/streaming/router/` |
+| `slurm/` | — | Deployment scripts | Stays internal |
+
+Frontend clients (`client/`, `prod-ui/`) stay in the internal repo.
+
+## Four design decisions that shape the upstream
+
+### D-1: Continuation model — Hybrid (server-held + opaque client-round-trip)
+
+Streaming WebSocket sessions hold continuation per-GPU (matches today's
+internal behavior, fast, zero client bandwidth). Stateless HTTP endpoints
+use opaque round-trip payloads. Server exposes a `snapshot_state` message
+that returns the opaque form for migration/retry.
+
+One serialization layer underlies both surfaces.
+
+Implementation: `SessionStore` (in-memory default, pluggable for
+redis/etc.) keyed by session ID, holds typed `LTX2ContinuationState`.
+- `snapshot(session_id) -> ContinuationState` exports for migration
+- `hydrate(state: ContinuationState) -> session_id` loads state into new session
+
+Payload schema covers: trailing conditioning frames (or tensor-blob ID),
+audio latents (or blob ID), segment index, audio sample rate,
+`video_position_offset_sec`, model-specific metadata.
+
+Landed in PR 7. See [cross-repo-surfaces.md](cross-repo-surfaces.md) for
+the full wire format.
+
+### D-2: Streaming server layout — Parallel subpackage `fastvideo/entrypoints/streaming/`
+
+Sits next to `fastvideo/entrypoints/openai/`. No existing code moves.
+Both servers share the `entrypoints/*` namespace. Shared utilities can be
+factored into `fastvideo/entrypoints/server_common/` later if needed.
+
+### D-3: LLM provider abstraction — `LLMProvider` protocol + built-in providers
+
+`prompt_enhancer.py` (69 KB) hard-coded three providers (cerebras_ifm,
+cerebras, groq) with provider-specific request/response handling
+scattered throughout. Upstreaming as-is would lock FastVideo to those
+providers.
+
+Protocol shape:
+
+```python
+@dataclass
+class LLMRequest:
+    messages: list[LLMMessage]
+    model: str
+    max_tokens: int | None = None
+    temperature: float | None = None
+    timeout_ms: int | None = None
+
+@dataclass
+class LLMResponse:
+    content: str
+    provider: str
+    model: str
+    latency_ms: float
+    fallback_used: bool = False
+
+class LLMProvider(Protocol):
+    name: str
+    async def complete(self, request: LLMRequest) -> LLMResponse: ...
+```
+
+PR 7.7 ships built-in providers for cerebras, groq. **Public Literal
+currently restricts to `Literal["cerebras", "groq"]`** — `cerebras_ifm`
+is internal-only and remains environment-driven on `dreamverse-server`.
+See [open-threads.md](open-threads.md) follow-up #3.
+
+Hot-reloadable system prompts via management endpoint. Sequential
+fallback across providers in priority order — race-based fallback (the
+internal optimization) deferred per [decisions-log.md](decisions-log.md)
+D-3.
+
+### D-4: Dynamo as first-class backend target — async event stream
+
+PR ai-dynamo/dynamo#7544 (closed draft) showed two frictions:
+
+1. Flat legacy kwargs — Dynamo handler had to know LTX-2-specific names.
+2. Sync-only generation — Dynamo wrapped `generator.generate(...)` in
+   `asyncio.to_thread` under a lock; no progress streaming, no
+   disaggregation path.
+
+Decision: **add `generate_async`** as the canonical execution API.
+
+```python
+async def generate_async(
+    self,
+    request: GenerationRequest,
+) -> AsyncGenerator[VideoEvent, None]: ...
+```
+
+Events:
+
+```python
+@dataclass
+class VideoProgressEvent:
+    step: int
+    total_steps: int
+    stage: str  # "denoise" | "refine" | "decode" | ...
+
+@dataclass
+class VideoPartialEvent:
+    frames: np.ndarray           # (num_frames, H, W, 3)
+    index: int                   # monotonic chunk index
+
+@dataclass
+class VideoFinalEvent:
+    video_bytes: bytes | None
+    tensor: torch.Tensor | None
+    metadata: dict[str, Any]
+    continuation_state: ContinuationState | None
+```
+
+The sync `generate_video(request=...) -> VideoResult` becomes a thin
+`asyncio.run` wrapper over `generate_async` that collects events and
+returns the final.
+
+**Three consumers, one substrate:**
+
+| Consumer | Transport | Request shape | State |
+|---|---|---|---|
+| Stateless OpenAI (`fastvideo/entrypoints/openai/`) | HTTP POST | `GenerationRequest` merged onto `ServeConfig.default_request` | Stateless; opaque payload |
+| Streaming WebSocket (`fastvideo/entrypoints/streaming/`) | WebSocket JSON + binary fMP4 | `GenerationRequest` per segment, session-scoped | Server-held; per-GPU continuation cache |
+| Dynamo native backend (`ai-dynamo/dynamo/components/src/dynamo/fastvideo/`) | Dynamo RPC endpoint | `NvCreateVideoRequest` ↔ adapter ↔ `GenerationRequest` | Aggregated today; future disaggregated via `ContinuationState` |
+
+**FastVideo does NOT host any Dynamo code.** The full backend package
+(`args.py`, `main.py`, `backend.py`, `register.py`, `health_check.py`)
+lives entirely in the Dynamo repo at `components/src/dynamo/fastvideo/`,
+matching the vllm/sglang pattern. FastVideo's only obligation is the
+stable public Python API.
+
+PR 7.10 lands the FastVideo-side contract. Dynamo backend code lives in
+ai-dynamo/dynamo (next iteration of #7544 reopens against PR 8 reference
+docs).
+
+## Target package layout
+
+```
+fastvideo/entrypoints/
+├── openai/                        # existing: stateless HTTP POST
+├── streaming/                     # NEW: session WebSocket
+│   ├── server.py                  # FastAPI + WebSocket entry
+│   ├── session.py                 # session lifecycle, state machine
+│   ├── session_store.py           # typed session state + snapshot/hydrate
+│   ├── protocol.py                # JSON WebSocket message schemas
+│   ├── stream.py                  # fMP4 encoding (av_fmp4 mode)
+│   ├── gpu_pool.py                # subprocess workers (PR 7.6)
+│   ├── worker.py                  # per-GPU worker loop
+│   ├── continuation.py            # typed LTX2 state payload
+│   ├── session_init_image.py
+│   ├── session_logger.py
+│   ├── mock_server.py
+│   ├── prompt/
+│   │   ├── enhancer.py            # provider-agnostic prompt ops
+│   │   ├── rewrite.py
+│   │   ├── safety.py              # optional fasttext
+│   │   └── providers/
+│   │       ├── base.py            # LLMProvider protocol
+│   │       ├── cerebras.py
+│   │       ├── cerebras_ifm.py
+│   │       └── groq.py
+│   └── router/
+│       ├── main.py
+│       └── registry.py
+├── cli/
+└── video_generator.py
+```
+
+## Typed config integration
+
+`ServeConfig` gets an optional `streaming: StreamingConfig | None`:
+
+```python
+@dataclass
+class StreamingConfig:
+    session_timeout_seconds: int = 300
+    generation_segment_cap: int = 6
+    stream_mode: Literal["av_fmp4", "legacy_jpeg"] = "av_fmp4"
+    warmup: WarmupConfig = field(default_factory=WarmupConfig)
+    pool: GpuPoolConfig = field(default_factory=GpuPoolConfig)
+    prompt: PromptEnhancerConfig | None = None
+    safety: PromptSafetyConfig | None = None
+
+@dataclass
+class GpuPoolConfig:
+    num_workers: int | None = None  # default: CUDA_VISIBLE_DEVICES count
+    enable_audio_reencode: bool = True
+    conditioning_num_frames: int = 9
+    conditioning_end_offset: int = 0
+
+@dataclass
+class PromptEnhancerConfig:
+    provider: Literal["cerebras", "groq"] = "cerebras"  # cerebras_ifm pending
+    model: str = "gpt-oss-120b"
+    timeout_ms: int = 20000
+    system_prompt_dir: str | None = None  # hot-reloadable
+
+@dataclass
+class PromptSafetyConfig:
+    enabled: bool = False
+    classifier_path: str | None = None
+```
+
+## `build_app` route contract — open follow-up
+
+Today `fastvideo.entrypoints.streaming.server.build_app` exposes only:
+
+- `GET /health`
+- `WS /v1/stream`
+
+The Dreamverse Next.js shell expects these additional routes that the
+upstream plan (and Dreamverse FE today) require:
+
+| Route | Owner per upstream plan | Status |
+|---|---|---|
+| `GET /healthz` | Streaming-server-side health (FastVideo) | 🔴 NOT YET MIGRATED |
+| `GET /readyz` | Streaming-server-side health (FastVideo) | 🔴 NOT YET MIGRATED |
+| `GET /status` | Streaming-server-side health (FastVideo) | 🔴 NOT YET MIGRATED |
+| `GET /curated-presets` | Operator-side surface (Dreamverse) | 🟡 stays in Dreamverse, FE feature-detects |
+| `POST /curated-presets/append` | Operator-side surface (Dreamverse) | 🟡 stays in Dreamverse |
+| `GET /prompt-system-config` | Operator-side surface (Dreamverse) | 🟡 stays in Dreamverse |
+| Devtools routes | Dreamverse-only | 🟡 stays in Dreamverse |
+
+Until the three health routes migrate into FastVideo's `build_app`, the
+`BE_FLAVOR=fastvideo` flavor of `launch_demo.sh` is a "diagnostic" flavor
+only (verifies typed serve-config path) — not FE-compatible. See
+[open-threads.md](open-threads.md) follow-up #1.
+
+The streaming-upstream plan listed `/healthz`, `/readyz`, `/status`,
+`/ws` as the contract that the upstream of `realtime/` → `streaming/`
+must preserve. They were deferred from PR 7.5's MVP.
+
+## PR 7.5 status — open as #1251
+
+Single-generator WebSocket end-to-end shipped (8 commits):
+
+1. `feat(streaming): protocol schemas + session state machine`
+2. `feat(streaming): fMP4 encoder + session init-image persistence`
+3. `feat(streaming): single-generator WebSocket server entry`
+4. `test(streaming): server lifecycle + protocol + fMP4 coverage`
+5. `docs(streaming): server contract spec`
+6. `fix(streaming): restore missing-streaming-block guard + retire stub-era test`
+7. `simplify(streaming): review follow-ups (idle timeout via asyncio.wait_for, _send_error helper, _cleanup_session, Protocol-typed generator, cleanup-on-disconnect)`
+8. `fix(streaming): enforce idle timeout on receive_json + flag generator-cancellation gap (TODO → PR 7.10)`
+
+Deferred TODOs (intentionally) blocking on PR 7.10:
+
+- **Per-step progress events** — only terminal `step_complete` today;
+  needs `generate_async` for per-step `VideoProgressEvent` emission.
+- **Mid-segment cancellation on client disconnect** — TODO marker in
+  `server.py` near `pool.run`. Needs `generate_async`'s cancellation
+  propagation.
+
+## PR 7.6 status — branch ready, not yet PR'd
+
+`will/api_7.6` (5 commits, rebased on 7.5):
+
+1. `feat [7.6/n]: GPU pool manager with typed worker boundary`
+2. `refactor [7.6/n]: route streaming server through GpuPool`
+3. `test [7.6/n]: GPU pool coverage (in-process + subprocess)`
+4. `fix(streaming): restore missing asyncio import in server` (rebase fixup)
+5. `feat(streaming): extract worker.py and add two-segment warmup`
+
+Tests: 17/17 gpu_pool tests + 89/89 streaming tests green.
+
+Ships:
+- `GpuPool` ABC + `InProcessGpuPool` + `SubprocessGpuPool` +
+  `PoolAssignment` / `PoolHealth` / `PoolAcquireTimeout`
+- `worker.py` — per-GPU `worker_main` and two-segment warmup helper
+- Subprocess startup uses typed `GeneratorConfig`, NOT flat kwargs
+- Session-to-GPU binding with timeout + queue for contention
+- Two-segment startup warmup per worker (segment 1 fresh + segment 2
+  with returned `ContinuationState` so both compile branches are primed)
+- `SessionStore` (from PR 7) wired for per-GPU continuation cache
+
+Deferred to PR 7.10:
+
+- **Audio re-encode (`LTX2AudioEncoder`, `AudioProcessor`)**: internal
+  `_re_encode_audio` runs *inside* the per-step streaming loop
+  (`_stream_av_fmp4_events` / `do_step_ltx2`). The whole-segment
+  `pool.run()` path PR 7.6 ships doesn't need it. Re-encode is a
+  per-step streaming concern that belongs with `generate_async`.
+- **Deprecate `VideoGenerator.from_pretrained(**flat_kwargs)`**: belongs
+  with PR 13 cleanup.
+
+## PR 7.10 — the unlock PR
+
+PR 7.10 adds `generate_async` and closes three open threads
+simultaneously:
+
+- Q-5 / D-5: audio re-encode for cross-segment continuity
+- Q-9: Dynamo progress passthrough
+- PR 7.5's mid-segment cancellation TODO (client disconnect →
+  `asyncio.CancelledError` → GPU work stops)
+
+Plus health-check helper:
+
+```python
+def default_health_check_request(self) -> GenerationRequest: ...
+# Returns 256x256, 8 frames, 1 step. Lets Dynamo's
+# FastVideoHealthCheckPayload.to_dict() produce a Dynamo
+# health_check_payload kwarg without knowledge of FastVideo internals.
+```
+
+Stable public exports:
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    GenerationRequest, SamplingConfig, ContinuationState,
+    VideoResult, VideoEvent,
+    VideoProgressEvent, VideoPartialEvent, VideoFinalEvent,
+)
+```
+
+Streaming server (PR 7.5) gets rewired to consume `generate_async`
+directly — no wrapper duplication.
+
+## Dynamo request/response mapping
+
+```
+NvCreateVideoRequest          ->  fastvideo.api.GenerationRequest
+  prompt                      ->    sampling.prompt
+  size="WxH"                  ->    sampling.width, sampling.height
+  seconds                     ->    seconds * nvext.fps -> sampling.num_frames
+  input_reference             ->    input.image_path | input.video_path
+  nvext.fps                   ->    sampling.fps
+  nvext.num_frames            ->    sampling.num_frames (overrides seconds*fps)
+  nvext.num_inference_steps   ->    sampling.num_inference_steps
+  nvext.guidance_scale        ->    sampling.guidance_scale
+  nvext.seed                  ->    sampling.seed
+  nvext.negative_prompt       ->    sampling.negative_prompt
+  response_format             ->    (handled at adapter's output stage)
+
+VideoFinalEvent               ->  NvVideosResponse
+  video_bytes                 ->    data[0].b64_json  (response_format=b64_json)
+  uploaded URL                ->    data[0].url       (response_format=url)
+  metadata.inference_time_s   ->    inference_time_s
+  continuation_state          ->    (reserved for future disaggregation)
+```
+
+## Open questions
+
+1. **Router placement** — in-tree at `fastvideo/entrypoints/streaming/router/`
+   (current implementation per PR 7.9) or separate package
+   `fastvideo-router/` / `fastvideo/contrib/router/`. Effectively
+   resolved in-tree by the PR 7.9 implementation.
+2. **Session ID authority** — server-generated UUID; accept externally
+   provided session ID only for resume flows.
+3. **Disaggregation readiness contract test** — should PR 7.10 validate
+   `ContinuationState` survives round-trip through Dynamo-style RPC
+   (pickle or JSON), even though Dynamo isn't using it today?
+   Recommended: yes; cheap regression guard.
+4. **Dynamo progress/status passthrough** — should PR 7.10's handler
+   contract emit intermediate `NvVideosResponse` chunks keyed off
+   `VideoProgressEvent`, or stay aggregated-final-only? Recommended:
+   stay aggregated-final for PR 7.10; revisit after Dynamo clarifies.
+5. **`video_position_offset_sec` semantics** — see [decisions-log.md](decisions-log.md)
+   open question; needs decision before PR 7.6 emits state.
+6. **`SessionStore` / `BlobStore` lifecycle** — eviction, TTL, blob-drop
+   on state replacement; defer to PR 7.5 design pass.

.agents/memory/index.jsonl

@@ -2,3 +2,4 @@
 {"name": "evaluation-registry", "description": "Catalog of all evaluation metrics with detailed explanations, implementation status, and usage guides", "path": "evaluation-registry/README.md", "status": "draft", "trust": "medium"}
 {"name": "experiment-journal", "description": "Living log of all experiments with hypotheses, configs, metrics, and insights", "path": "experiment-journal/README.md", "status": "draft", "trust": "medium"}
 {"name": "related-work", "description": "Index of related papers, repos, and blog posts with structured comparisons to FastVideo", "path": "related-work/README.md", "status": "draft", "trust": "low"}
+{"name": "dreamverse-integration", "description": "Consolidated knowledge base for the FastVideo public API refactor (PRs 0-17), LTX-2 streaming server upstream, Dreamverse migration from FastVideo-internal, and NVFP4 quantization landing", "path": "dreamverse-integration/README.md", "status": "ready", "trust": "high"}

.agents/scripts/sync-skills.sh

@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+# Sync .agents/skills/ into .claude/skills/ via per-skill symlinks.
+#
+# Why: Claude Code only scans .claude/skills/ and ~/.claude/skills/ for
+# user-invocable skills (no skillsPath config exists — see
+# https://code.claude.com/docs/en/skills.md). This repo's skills live
+# in .agents/skills/ so they travel with the repo and stay under git.
+# Run this once after cloning (or after adding/removing a skill) to
+# expose them to Claude Code without maintaining a parallel tree.
+#
+# Usage:
+#   .agents/scripts/sync-skills.sh
+#
+# Idempotent and safe to re-run. Prunes stale symlinks whose source
+# has been removed from .agents/skills/. Leaves hand-written
+# .claude/skills/<name>/ directories untouched (only symlinks are
+# managed).
+
+set -euo pipefail
+
+REPO_ROOT="$(git -C "$(dirname "$0")" rev-parse --show-toplevel)"
+SRC_DIR="$REPO_ROOT/.agents/skills"
+DST_DIR="$REPO_ROOT/.claude/skills"
+
+if [[ ! -d "$SRC_DIR" ]]; then
+    echo "Error: $SRC_DIR does not exist." >&2
+    exit 1
+fi
+
+mkdir -p "$DST_DIR"
+
+linked=0
+unchanged=0
+skipped=0
+pruned=0
+
+link_skill() {
+    local name="$1"
+    local src="$SRC_DIR/$name"
+    local dst="$DST_DIR/$name"
+    # Relative target keeps symlinks portable across clones.
+    local rel="../../.agents/skills/$name"
+
+    if [[ -L "$dst" ]]; then
+        if [[ "$(readlink "$dst")" == "$rel" ]]; then
+            unchanged=$((unchanged + 1))
+            return
+        fi
+        rm "$dst"
+    elif [[ -e "$dst" ]]; then
+        echo "Skipped (not a symlink): .claude/skills/$name" >&2
+        skipped=$((skipped + 1))
+        return
+    fi
+
+    ln -s "$rel" "$dst"
+    echo "Linked: .claude/skills/$name -> $rel"
+    linked=$((linked + 1))
+}
+
+prune_stale() {
+    local link="$1"
+    local target
+    target="$(readlink "$link")"
+    case "$target" in
+        ../../.agents/skills/*) ;;
+        *) return ;;
+    esac
+    local name="${target##*/}"
+    if [[ ! -d "$SRC_DIR/$name" ]]; then
+        rm "$link"
+        echo "Pruned stale: .claude/skills/$(basename "$link")"
+        pruned=$((pruned + 1))
+    fi
+}
+
+for src in "$SRC_DIR"/*/; do
+    [[ -d "$src" ]] || continue
+    name="$(basename "$src")"
+    # Only treat directories that actually contain a SKILL.md as skills.
+    [[ -f "$src/SKILL.md" ]] || continue
+    link_skill "$name"
+done
+
+shopt -s nullglob
+for link in "$DST_DIR"/*; do
+    [[ -L "$link" ]] || continue
+    prune_stale "$link"
+done
+shopt -u nullglob
+
+printf "\nSummary: %d linked, %d unchanged, %d pruned" "$linked" "$unchanged" "$pruned"
+if [[ "$skipped" -gt 0 ]]; then
+    printf ", %d skipped (non-symlink collision)" "$skipped"
+fi
+printf "\n"

.agents/skills/index-related-work/SKILL.md

@@ -1,94 +0,0 @@
----
-name: index-related-work
-description: Ingest a paper or repository into the related work index
----
-
-# Index Related Work
-
-## Purpose
-Create a structured summary of a related paper, repository, or blog post and
-add it to `.agents/memory/related-work/` for future reference. This builds the
-agent's knowledge base for making informed decisions about training, evaluation,
-and architecture choices.
-
-## Prerequisites
-- Access to the paper/repo (URL, PDF, or local clone).
-
-## Inputs
-
-| Parameter | Required | Description |
-|-----------|----------|-------------|
-| `source` | Yes | URL, citation, or local path |
-| `type` | Yes | `paper`, `repo`, or `blog` |
-| `tags` | No | List of tags (default: inferred from content) |
-
-## Steps
-
-### 1. Extract key information
-
-For **papers**: Read abstract, method section, experimental setup, and results.
-For **repos**: Read README, key source files, and training scripts.
-For **blogs**: Read the full post.
-
-Focus on:
-- What problem does it solve?
-- What architecture/technique is used?
-- How does it relate to FastVideo's approach?
-
-### 2. Create the index entry
-
-Write to `.agents/memory/related-work/<slug>.md`:
-
-```markdown
----
-title: <title>
-source: <URL or citation>
-type: paper | repo | blog
-date_indexed: <ISO-8601>
-tags: [world-model, distillation, evaluation, ...]
----
-
-## Summary
-<1-2 paragraph summary.>
-
-## Key Differences from FastVideo
-- <comparison points>
-
-## Actionable Insights
-- <what we could adopt or adapt>
-```
-
-### 3. Update the catalog
-
-If `.agents/memory/related-work/_catalog.md` exists, append the new entry.
-If not, create it:
-
-```markdown
-# Related Work Catalog
-
-| Slug | Title | Type | Tags | Date |
-|------|-------|------|------|------|
-| <slug> | <title> | <type> | <tags> | <date> |
-```
-
-## Outputs
-- New file in `.agents/memory/related-work/<slug>.md`.
-- Updated catalog.
-
-## Example Usage
-
-```
-Index the Self-Forcing paper:
-
-  source: https://arxiv.org/abs/2406.xxxxx
-  type: paper
-  tags: [world-model, self-forcing, distillation]
-```
-
-## References
-- `.agents/memory/related-work/README.md` — schema documentation
-
-## Changelog
-| Date | Change |
-|------|--------|
-| 2026-03-02 | Initial version |

.agents/skills/index.jsonl

@@ -3,5 +3,5 @@
 {"name": "summarize-run", "description": "Extract a W&B run summary into a structured experiment report", "path": "summarize-run/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "log-experiment", "description": "Append or update an experiment entry in the experiment journal", "path": "log-experiment/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "evaluate-video-quality", "description": "Evaluate generated video quality using available metrics (SSIM, loss trajectory, caption consistency)", "path": "evaluate-video-quality/SKILL.md", "status": "draft", "trust": "low"}
-{"name": "index-related-work", "description": "Ingest a paper or repository into the related work index", "path": "index-related-work/SKILL.md", "status": "draft", "trust": "low"}
-{"name": "search-related-work", "description": "Query the related work index for relevant papers, repos, or comparisons", "path": "search-related-work/SKILL.md", "status": "draft", "trust": "low"}
+{"name": "seed-ssim-references", "description": "Run a new or updated fastvideo/tests/ssim/ test on Modal, pull generated videos, and upload them to FastVideo/ssim-reference-videos so the test has a regression baseline", "path": "seed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}
+{"name": "reseed-ssim-references", "description": "Re-seed (overwrite) HF reference videos for an existing fastvideo/tests/ssim/ test and a single model id on Modal L40S. Always backs up current refs first, regenerates on Modal, pauses for the user to eyeball before-vs-after, then uploads with --force scoped to --model-id. Sister skill to seed-ssim-references; use when intentional code change has invalidated existing refs", "path": "reseed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}

.agents/skills/launch-experiment/SKILL.md

@@ -12,7 +12,7 @@ automates the boilerplate of setting environment variables, picking the right
 entrypoint, and applying defaults from the closest example script.
 
 ## Prerequisites
-- The repo is cloned and `fastvideo` is installed (`uv pip install -e .[dev]`).
+- The repo is cloned and `fastvideo` is installed (`uv pip install -e ".[dev]"`).
 - Dataset is preprocessed (see `docs/training/data_preprocess.md`).
 - `WANDB_API_KEY` is set in the environment (or `WANDB_MODE=offline` for local).
 - GPU resources are available (multi-GPU requires NCCL).

.agents/skills/reseed-ssim-references/SKILL.md

@@ -0,0 +1,343 @@
+---
+name: reseed-ssim-references
+description: Re-seed HF reference videos for a single existing SSIM test on Modal L40S. Always backs up current refs locally first, regenerates on Modal, pauses for the user to eyeball before-vs-after quality, then overwrites the targeted `<model_id>` subtree on `FastVideo/ssim-reference-videos` with `--force`. Use when an intentional code change (model port fix, attention backend swap, kernel upgrade, hyperparameter change) has invalidated existing refs and they need to be regenerated. Pairs with `seed-ssim-references`, which is for first-time seeding only.
+---
+
+# Re-seed SSIM Reference Videos
+
+## Purpose
+
+Replace the existing SSIM reference videos for a single `(test_file, model_id)`
+pair on the HF dataset (`FastVideo/ssim-reference-videos`). This is **destructive**
+on HF — the old refs are overwritten — so the skill always:
+
+1. Confirms intent with a one-liner the user has to type.
+2. Downloads the existing refs as a local, timestamped backup.
+3. Regenerates on Modal L40S (same code path that CI uses).
+4. Pauses for a side-by-side eyeball of backup vs new mp4s.
+5. Uploads with `--force`, scoped to the single `--model-id`.
+6. Reminds the user to keep the backup until the PR lands.
+
+Pairs with `seed-ssim-references`, which is the inverse (first-time seeding
+only, refuses to overwrite). Re-seeding is intentionally a separate, more
+ceremonial operation because mistakenly clobbering production refs is much
+harder to recover from than failing closed.
+
+## When to use
+
+- An intentional code change (model port fix, kernel upgrade, attention
+  backend swap, hyperparameter change in the test itself) has shifted the
+  expected SSIM output and the existing refs no longer represent the new
+  ground truth.
+- A test is failing in CI **for the right reason** (the new code is correct,
+  the old refs are stale).
+
+## When not to use
+
+- A test is failing for the **wrong** reason (the port is buggy, not the
+  refs). Fix the port; re-seeding hides the bug.
+- A brand-new test that has no refs on HF yet. Use `seed-ssim-references`.
+- "Just to clean up drift" without a concrete code change to point at. The
+  PR description has to justify *why* refs changed; without a concrete
+  change, there's nothing to write.
+
+## Inputs
+
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `test_file` | Yes | Path to the SSIM test, e.g. `fastvideo/tests/ssim/test_matrixgame_similarity.py`. Validated against `fastvideo/tests/ssim/test_*_similarity.py`. |
+| `model_id` | Yes | Single model id from the test's `*_MODEL_TO_PARAMS`, e.g. `Matrix-Game-2.0-Diffusers-Base`. Re-seed runs are **per model**. For multi-model tests, invoke the skill once per model. |
+| `intent_rationale` | Yes | One-line explanation of *why* refs are being regenerated (e.g. "Relax FA-2 head_size whitelist to include 80 — matrix_game now uses FLASH_ATTN instead of TORCH_SDPA"). Recorded in the backup directory and reused in the PR description. |
+
+Hardcoded:
+
+- Modal GPU: **L40S** (matches CI; re-seeding from another SKU produces refs
+  that L40S CI cannot match).
+- Quality tier: **`default`**. `full_quality` is a separate, deliberate
+  operation.
+- HF repo: `FastVideo/ssim-reference-videos` (override via
+  `FASTVIDEO_SSIM_REFERENCE_HF_REPO`).
+- Device folder: `L40S_reference_videos`.
+
+## Prerequisites
+
+The user has confirmed:
+
+- `modal` CLI authenticated.
+- `hf` CLI authenticated, **and** `HF_API_KEY` (or `HUGGINGFACE_HUB_TOKEN` /
+  `HF_TOKEN`) exported with **write** access to
+  `FastVideo/ssim-reference-videos`.
+- The current branch's code is the change that motivated the re-seed (i.e.
+  `git rev-parse HEAD` is the commit that intentionally invalidated refs).
+
+Fail fast if any of these are missing.
+
+## Steps
+
+### 1. Validate inputs and confirm intent
+
+- Verify `test_file` exists and matches `fastvideo/tests/ssim/test_*_similarity.py`.
+- Grep the file for `*_MODEL_TO_PARAMS` and assert `model_id` is one of its
+  keys. If the file has only a single hardcoded model, accept that model id
+  as the only valid value.
+- Print the rationale and ask the user to type **`confirm reseed`** (not just
+  `y` — make it deliberate):
+
+  > About to RE-SEED references for model `<model_id>` from test `<test_file>`.
+  > This will OVERWRITE existing refs on
+  > `FastVideo/ssim-reference-videos/reference_videos/default/L40S_reference_videos/<model_id>/`
+  > after backup + Modal regen + eyeball.
+  >
+  > Reason: `<intent_rationale>`
+  > HEAD: `<git rev-parse --short=12 HEAD>`
+  >
+  > Reply `confirm reseed` to proceed, anything else to abort.
+
+  Stop until the user types exactly `confirm reseed`. Anything else aborts
+  with no side effects.
+
+### 2. Back up existing refs
+
+Always required. The backup is the only graceful path back if anything goes
+wrong later.
+
+```bash
+SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
+TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
+MODEL_SAFE=$(echo "<model_id>" | tr '/' '_')
+BACKUP_DIR="ssim_reseed_backup/${TIMESTAMP}_${SHORT_COMMIT}_${MODEL_SAFE}"
+mkdir -p "$BACKUP_DIR"
+
+hf download \
+    --repo-type dataset FastVideo/ssim-reference-videos \
+    --include "reference_videos/default/L40S_reference_videos/<model_id>/**" \
+    --local-dir "$BACKUP_DIR"
+
+mp4_count=$(find "$BACKUP_DIR" -name "*.mp4" | wc -l)
+echo "Backup mp4 count: $mp4_count"
+[ "$mp4_count" -gt 0 ] || {
+    echo "ERROR: backup is empty for <model_id>. Either the model id is wrong"
+    echo "or there are no existing refs (use seed-ssim-references instead)."
+    exit 1
+}
+
+# Provenance — used in the PR description
+cat > "$BACKUP_DIR/PROVENANCE.txt" <<EOF
+test_file: <test_file>
+model_id: <model_id>
+head_commit: $(git rev-parse HEAD)
+timestamp_utc: $(date -u +%FT%TZ)
+reason: <intent_rationale>
+EOF
+```
+
+If the `hf download` produces zero mp4s, abort — the user has either picked a
+non-existent `model_id` or there are no refs yet (in which case
+`seed-ssim-references` is the right tool).
+
+### 3. Regenerate on Modal L40S
+
+Mirror CI's exact env recipe so the regenerated refs are byte-comparable to
+what CI will produce on the same commit. Two differences from CI:
+
+1. **Pass the same env prefix CI uses** (`IMAGE_VERSION`, `BUILDKITE_*`) — see
+   `.buildkite/pipeline.yml:1-3` and `.buildkite/scripts/pr_test.sh:62-83`.
+   Without this, `ssim_test.py:17-18` resolves a different GHCR image tag
+   (default is `latest`, CI is `py3.12-latest`), and `ssim_test.py:38-46`
+   bakes different values into the image's frozen env block. **Mismatched
+   image or env is the most common source of SSIM drift between reseed and
+   CI runs.**
+2. **Do not pass `--skip-reference-download`**. Letting the test fetch the
+   existing refs and run the full SSIM compare gives "before" SSIM numbers
+   for the PR description, and the test still produces the new mp4s
+   regardless of whether the comparison passes or fails.
+
+```bash
+SUBDIR="${TIMESTAMP}_${SHORT_COMMIT}"
+
+IMAGE_VERSION="py3.12-latest" \
+BUILDKITE_REPO="$(git config --get remote.origin.url)" \
+BUILDKITE_COMMIT="$(git rev-parse HEAD)" \
+BUILDKITE_PULL_REQUEST="${BUILDKITE_PULL_REQUEST:-false}" \
+modal run fastvideo/tests/modal/ssim_test.py \
+    --git-repo="$(git config --get remote.origin.url)" \
+    --git-commit="$(git rev-parse HEAD)" \
+    --hf-api-key="$HF_API_KEY" \
+    --test-files="<test_file>" \
+    --sync-generated-to-volume \
+    --generated-volume-subdir="$SUBDIR" \
+    --no-fail-fast
+```
+
+Capture the printed `modal volume get ...` hint — its `<SUBDIR>` matches
+`$SUBDIR` and is needed for step 4. Capture the SSIM numbers from the test
+output (or from the JSON next to the generated mp4) for the PR description.
+
+### 4. Download generated videos
+
+```bash
+modal volume get --force hf-model-weights \
+    ssim_generated_videos/default/"$SUBDIR"/generated_videos \
+    ./generated_videos_modal/default
+```
+
+After this, the new mp4s live at:
+
+```
+./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4
+```
+
+`--force` is required when `./generated_videos_modal/default` already exists
+from a prior run; safe on the first run too.
+
+### 5. PAUSE — user reviews quality side-by-side
+
+Print the diff and the comparison:
+
+```bash
+echo "=== File list diff (backup vs new) ==="
+diff -u \
+    <(find "$BACKUP_DIR/reference_videos/default/L40S_reference_videos/<model_id>" -name "*.mp4" \
+        | sed "s|$BACKUP_DIR/reference_videos/default/L40S_reference_videos/||" | sort) \
+    <(find ./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id> -name "*.mp4" \
+        | sed "s|./generated_videos_modal/default/generated_videos/L40S_reference_videos/||" | sort) \
+    || true
+
+echo
+echo "=== SSIM numbers from this run (paste into PR) ==="
+find ./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id> -name "*_ssim.json" -exec cat {} \;
+```
+
+Then stop and tell the user:
+
+> Old refs backed up to `$BACKUP_DIR`.
+> New videos in `./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id>/`.
+>
+> Open both in a video player. Confirm the new videos:
+>   1. Look correct (no obvious artifacts, no black/static frames).
+>   2. Are *intentionally* different from the backup in the way described
+>      in `<intent_rationale>` (e.g. slight numerical drift only, not a
+>      different scene / different motion / corrupted output).
+>
+> Reply **`upload`** to overwrite HF, anything else to abort.
+> Aborting leaves the backup and new videos on disk for inspection — nothing
+> on HF changes.
+
+Do not proceed until the user types exactly `upload`. If they abort, leave
+everything on disk and stop here.
+
+### 6. Copy into the local reference layout
+
+Same as `seed-ssim-references` step 5:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --generated-dir ./generated_videos_modal/default/generated_videos/L40S_reference_videos
+```
+
+Result: `fastvideo/tests/ssim/reference_videos/default/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+
+### 7. Upload with `--force`, scoped to `--model-id`
+
+The `--force` flag is what makes this skill different from `seed-ssim-references`.
+Always pair it with `--model-id` so a typo cannot accidentally overwrite a
+neighboring model's refs.
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py upload \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --model-id "<model_id>" \
+    --force
+```
+
+The CLI's overwrite guard refuses without `--force`; with `--force` it
+overwrites only files under
+`reference_videos/default/L40S_reference_videos/<model_id>/`.
+
+### 8. Report success and retention guidance
+
+Print:
+
+- The HF path that was overwritten (`<repo>/reference_videos/default/L40S_reference_videos/<model_id>/`).
+- The local backup directory path.
+- The new SSIM numbers from step 5.
+- This restore command, in case the PR review surfaces a problem after
+  upload:
+
+  ```bash
+  python fastvideo/tests/ssim/reference_videos_cli.py upload \
+      --quality-tier default \
+      --device-folder L40S_reference_videos \
+      --model-id "<model_id>" \
+      --reference-dir "$BACKUP_DIR/reference_videos/default/L40S_reference_videos" \
+      --force
+  ```
+
+- This PR-description checklist (see `fastvideo/tests/ssim/AGENTS.md` →
+  *Updating Reference Videos*):
+  1. Source commit that produced the new refs (HEAD at re-seed time).
+  2. Test command and GPU SKU (`L40S`).
+  3. Before/after SSIM numbers.
+  4. The `<intent_rationale>` from step 1.
+  5. A note that the backup lives at `$BACKUP_DIR` and should be retained
+     until CI on the PR is green.
+
+Do **not** auto-rerun the SSIM test — the user does that as part of the PR.
+
+## Failure modes and how to handle them
+
+- **`HF_API_KEY` unset.** Stop before step 2.
+- **Backup is empty (zero mp4s).** Stop before step 3 — the model id is
+  wrong or the refs don't exist yet (use `seed-ssim-references`).
+- **Modal run fails before generation.** No mp4s on the volume. Don't
+  upload. Investigate the failure (test crash, OOM, partition exhaustion),
+  fix, then retry from step 3. Backup is still intact.
+- **Quality regressed (visual or metric).** User aborts at step 5. Backup
+  retained. New videos retained on disk for inspection. Nothing on HF
+  changed. Either fix the underlying code change or abandon the re-seed.
+- **User confirmed `upload` but later realized the new refs are wrong.**
+  Run the restore command from step 8 with the backup `--reference-dir`.
+  This is exactly why the backup exists.
+- **Multi-model test, only one model is being re-seeded.** Run the skill
+  once per model id. The `--model-id` scope on upload guarantees the others
+  are untouched.
+
+## Design notes (for future skill maintainers)
+
+- Per-`model_id` scope is mandatory. The dataset houses many model subtrees;
+  re-seeding the wrong one is hard to undo without backup.
+- `default` tier only; `full_quality` is a separate, deliberate operation
+  with different params and ~doubled runtime, and isn't what CI gates on.
+- The skill deliberately does **not** pass `--skip-reference-download` to
+  Modal so we get pre-reseed SSIM numbers for the PR. The `seed`-skill
+  passes it because no refs exist yet; for re-seed, refs do exist and
+  exposing the comparison is informative.
+- The two-token confirm (`confirm reseed`, then `upload`) is intentional.
+  Re-seeding is high-blast-radius and should not be one-keystroke.
+- The backup directory is plain mp4s + `PROVENANCE.txt`. No HF metadata is
+  preserved; the restore path uses `reference_videos_cli.py upload
+  --reference-dir` which doesn't need it.
+
+## References
+
+- `.agents/skills/seed-ssim-references/SKILL.md` — the first-time seed
+  skill this one parallels. Read it for the Modal flag rationale shared
+  between the two flows.
+- `fastvideo/tests/ssim/AGENTS.md` — directory rules, including the PR
+  expectations for any reference-video change (rationale, before/after
+  SSIM, source commit/model/backend).
+- `fastvideo/tests/ssim/reference_videos_cli.py` — `copy-local`, `upload`
+  (with `--model-id`, `--force`), `download`. The overwrite guard at
+  `upload_reference_videos` is the safety net this skill leans on.
+- `fastvideo/tests/modal/ssim_test.py` — Modal orchestrator;
+  `--sync-generated-to-volume`, `--generated-volume-subdir`,
+  `--skip-reference-download`, `--no-fail-fast`.
+
+## Changelog
+
+| Date | Change |
+|------|--------|
+| 2026-05-02 | Initial version. Sister skill to `seed-ssim-references`, scoped to single `(test_file, model_id)` re-seeds, with mandatory backup and two-token confirm. |

.agents/skills/search-related-work/SKILL.md

@@ -1,82 +0,0 @@
----
-name: search-related-work
-description: Query the related work index for relevant papers, repos, or comparisons
----
-
-# Search Related Work
-
-## Purpose
-Search through `.agents/memory/related-work/` to find indexed papers, repos,
-or blog posts relevant to a query. Use this when you need to understand how
-other work compares to FastVideo's approach, or when looking for techniques
-to adopt.
-
-## Prerequisites
-- The related work index has entries (`.agents/memory/related-work/*.md`).
-
-## Inputs
-
-| Parameter | Required | Description |
-|-----------|----------|-------------|
-| `query` | Yes | Natural language query |
-| `tags` | No | Filter by tags (e.g., `[distillation, evaluation]`) |
-| `type` | No | Filter by type (`paper`, `repo`, `blog`) |
-
-## Steps
-
-### 1. Search the index
-
-Use grep-based search through `.agents/memory/related-work/`:
-
-```bash
-# Search by content
-grep -rl "<query>" .agents/memory/related-work/
-
-# Search by tags (in frontmatter)
-grep -l "tags:.*<tag>" .agents/memory/related-work/*.md
-```
-
-### 2. Rank results
-
-For each matching file:
-1. Read the file.
-2. Score relevance to the query based on:
-   - Title match
-   - Tag match
-   - Content match (summary, differences, insights)
-3. Return top results.
-
-### 3. Format output
-
-```markdown
-## Related Work Search: "<query>"
-
-### 1. <Title> (relevance: high)
-- **Source**: <URL>
-- **Tags**: <tags>
-- **Key insight**: <most relevant excerpt>
-- **File**: `.agents/memory/related-work/<slug>.md`
-
-### 2. <Title> (relevance: medium)
-...
-```
-
-## Outputs
-- Ranked list of relevant related work entries with excerpts.
-
-## Example Usage
-
-```
-Search for work related to video quality evaluation metrics:
-
-  query: "video generation quality evaluation metrics"
-  tags: [evaluation]
-```
-
-## References
-- `.agents/memory/related-work/README.md` — index schema
-
-## Changelog
-| Date | Change |
-|------|--------|
-| 2026-03-02 | Initial version |

.agents/skills/seed-ssim-references/SKILL.md

@@ -0,0 +1,273 @@
+---
+name: seed-ssim-references
+description: Seed HF reference videos for a single newly-added SSIM test. Runs the test on Modal L40S, downloads the generated mp4s via `modal volume get`, pauses for the user to eyeball quality, then uploads only that test's files to `FastVideo/ssim-reference-videos`. Use when a new `fastvideo/tests/ssim/test_*_similarity.py` has just been added and has no references on HF yet.
+---
+
+# Seed SSIM Reference Videos
+
+## Purpose
+
+A brand-new SSIM test in `fastvideo/tests/ssim/` fails forever until its
+reference videos exist on the HF dataset (`FastVideo/ssim-reference-videos`).
+This skill:
+
+1. Runs the test on Modal's L40S pool to generate the videos.
+2. Downloads them to the local repo via `modal volume get`.
+3. Pauses so the user can eyeball the mp4s and confirm quality.
+4. Uploads only the new test's files to HF, with a guard that refuses to
+   overwrite anything already present.
+
+The skill is run **manually**, once per new test. Before invoking it, the user
+has already sanity-tested the new test locally — it launches `VideoGenerator`
+and writes an mp4 without crashing. The skill does not re-test locally; it
+goes straight to Modal L40S (which is what CI uses).
+
+## When to use
+
+- A new `test_*_similarity.py` file has been added in `fastvideo/tests/ssim/`
+  and the HF dataset has no `reference_videos/default/L40S_reference_videos/<model_id>/`
+  subtree for it yet.
+
+## When not to use
+
+- Regular CI runs — once refs exist, `pytest fastvideo/tests/ssim/` downloads
+  them automatically.
+- Re-seeding an existing test. That requires `--force` on the upload step, and
+  is out of scope here; treat as a separate, deliberate operation.
+
+## Inputs
+
+The skill has **one required input**: the path to the new SSIM test file.
+Prompt the user for it if they didn't supply it.
+
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `test_file` | Yes | e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`. The skill's first action is to ask for this if missing. |
+
+Everything else is fixed:
+
+- Modal runner GPU: **L40S** (hardcoded in `fastvideo/tests/modal/ssim_test.py`).
+- Device folder: `L40S_reference_videos`.
+- Quality tier: `default` (the tier CI runs). The `full_quality` tier is not
+  seeded by this skill.
+- HF repo: `FastVideo/ssim-reference-videos` (dataset).
+- Multi-model test files: all model ids in `*_MODEL_TO_PARAMS` are seeded
+  together; the Modal run produces one mp4 per (model, prompt, backend) and
+  the upload scopes by `--model-id`, looping if there is more than one.
+
+## Prerequisites
+
+The user has confirmed:
+
+- `modal` CLI authenticated.
+- `HF_API_KEY` (or `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`) exported with write
+  access to `FastVideo/ssim-reference-videos`.
+- The test file runs locally end-to-end (generates an mp4; SSIM assertion
+  failure due to missing reference is expected and fine).
+
+Fail fast if the token env var is missing.
+
+## Steps
+
+### 1. Ask for the test file
+
+If the user didn't name one, ask: *"Which SSIM test file do you want to seed
+references for? (e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`)"*.
+
+Validate:
+
+- Path exists and matches `fastvideo/tests/ssim/test_*_similarity.py`.
+- File defines a `*_MODEL_TO_PARAMS` dict — grep it to extract the set of
+  model ids. Those ids drive step 5.
+
+If either check fails, stop and tell the user what's wrong.
+
+### 2. Run the test on Modal L40S
+
+Pick a subdir name so repeated runs don't collide:
+
+```bash
+SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
+TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
+SUBDIR="${TIMESTAMP}_${SHORT_COMMIT}"
+```
+
+Then launch the Modal run. The `IMAGE_VERSION` and `BUILDKITE_*` env-prefix
+**must** match what CI exports in `.buildkite/scripts/pr_test.sh`, otherwise
+`fastvideo/tests/modal/ssim_test.py` resolves a different GHCR image tag
+(default is `latest`, CI is `py3.12-latest`) and bakes different values into
+the image's frozen env block (`ssim_test.py:17-18, 38-46`). Mismatched image
+or env produces SSIM drift that doesn't show up until the same commit runs
+in CI.
+
+```bash
+IMAGE_VERSION="py3.12-latest" \
+BUILDKITE_REPO="$(git config --get remote.origin.url)" \
+BUILDKITE_COMMIT="$(git rev-parse HEAD)" \
+BUILDKITE_PULL_REQUEST="${BUILDKITE_PULL_REQUEST:-false}" \
+modal run fastvideo/tests/modal/ssim_test.py \
+    --git-repo="$(git config --get remote.origin.url)" \
+    --git-commit="$(git rev-parse HEAD)" \
+    --hf-api-key="$HF_API_KEY" \
+    --test-files="<test_file>" \
+    --sync-generated-to-volume \
+    --generated-volume-subdir="$SUBDIR" \
+    --skip-reference-download \
+    --no-fail-fast
+```
+
+Env prefix rationale (parity with CI; see `.buildkite/pipeline.yml:1-3` and
+`.buildkite/scripts/pr_test.sh:62-83`):
+- `IMAGE_VERSION=py3.12-latest`: pins the Modal image tag to the same one CI
+  uses. Without this, `ssim_test.py:17` falls back to `latest`, which on
+  GHCR is built from `Dockerfile.python3.10` — different Python, torch, and
+  flash-attn wheel than CI's `py3.12-latest` (`infra-build-image.yml:51-67`,
+  `_template-build-image.yml:65-101`).
+- `BUILDKITE_REPO`/`BUILDKITE_COMMIT`/`BUILDKITE_PULL_REQUEST`: mirror what
+  Buildkite exports. `ssim_test.py:38-46` bakes these into the image's
+  `.env(...)` block; mismatched values can perturb in-container code paths
+  that branch on PR-vs-non-PR. `false` for `BUILDKITE_PULL_REQUEST` matches
+  Buildkite's "non-PR build" sentinel.
+
+Flag rationale:
+- `--skip-reference-download`: no refs exist yet, so conftest must not try to
+  pull them.
+- `--no-fail-fast`: lets the test finish generation before `_assert_similarity`
+  raises `FileNotFoundError: Reference video folder does not exist`. The
+  expected failure is what we want — the mp4 has already been written.
+- `--sync-generated-to-volume` + `--generated-volume-subdir`: copies the
+  generated mp4s to the `hf-model-weights` Modal volume under
+  `ssim_generated_videos/default/<SUBDIR>/generated_videos/` so we can pull
+  them locally.
+
+The Modal run will end with a nonzero exit (expected) and print a
+`modal volume get hf-model-weights ssim_generated_videos/default/<SUBDIR>/generated_videos ./generated_videos_modal/default`
+command. Capture that `<SUBDIR>` — you need it for step 3.
+
+### 3. Download generated videos locally
+
+```bash
+modal volume get --force hf-model-weights \
+    ssim_generated_videos/default/"$SUBDIR"/generated_videos \
+    ./generated_videos_modal/default
+```
+
+`--force` is required when the parent `./generated_videos_modal/default`
+already exists; without it, `modal volume get` errors with `[Errno 21] Is a
+directory`. Safe to pass on the first run too.
+
+After this, the mp4s live at
+`./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+The extra `generated_videos/` level comes from the volume layout in
+`_sync_generated_videos_to_volume` (`ssim_test.py`) — the command copies
+`<repo>/fastvideo/tests/ssim/generated_videos/<tier>` to
+`ssim_generated_videos/<tier>/<SUBDIR>/generated_videos/`, and `modal volume
+get` preserves that trailing `generated_videos/` segment.
+
+### 4. PAUSE — user reviews quality
+
+Print the list of downloaded mp4s and their paths, then stop. Tell the user:
+
+> "Generated videos downloaded to `./generated_videos_modal/default/generated_videos/L40S_reference_videos/`. Please open them and confirm the quality looks correct. Reply **`upload`** to continue, or anything else to abort."
+
+Do not proceed until the user explicitly says `upload`. If they abort, leave
+everything on disk so they can inspect further — no cleanup.
+
+### 5. Copy into the local reference layout
+
+Scoped copy — only the new test's mp4s. Loop over each `<model_id>` extracted
+in step 1:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --generated-dir ./generated_videos_modal/default/generated_videos/L40S_reference_videos
+```
+
+(The `--generated-dir` points at the device-folder root inside the
+downloaded tree; `copy-local` walks all `<model>/<backend>/*.mp4`
+underneath it. Since the Modal run was scoped to a single test file via
+`--test-files`, only that test's model(s) are present — so the copy is
+implicitly per-test.)
+
+Result: `fastvideo/tests/ssim/reference_videos/default/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+
+### 6. Upload to HF — scoped per model_id, with overwrite guard
+
+For each `<model_id>`:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py upload \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --model-id "<model_id>"
+```
+
+The upload command:
+
+- Uploads **only** `reference_videos/default/L40S_reference_videos/<model_id>/`.
+- **Refuses** if any file already exists at that path on HF (this is the
+  guard — seeding a new test should never clobber existing refs). To override,
+  the user must re-run with `--force`. If the guard fires, stop and report
+  exactly which files exist; do not silently `--force`.
+
+Reads the HF token from `HF_API_KEY` / `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`.
+
+### 7. Report success
+
+List what was uploaded (paths in repo) and remind the user to push any
+related code changes. Do **not** auto-verify by re-running Modal — the user
+can run `pytest fastvideo/tests/ssim/<test_file>` later to confirm end-to-end;
+it will auto-download the refs they just uploaded.
+
+## Failure modes and how to handle them
+
+- **`HF_API_KEY` unset.** Stop before step 2. The Modal run needs it (passed
+  via `--hf-api-key`), and step 6 needs it for upload.
+- **Modal run fails before generation.** No mp4s on the volume — nothing to
+  download. Fix the test locally (`pytest fastvideo/tests/ssim/<test_file>`)
+  and retry from step 2.
+- **`./generated_videos_modal/default/L40S_reference_videos/` missing after
+  `modal volume get`.** The run didn't produce videos (most likely the test
+  crashed before writing, or `REQUIRED_GPUS` exceeded the partition capacity
+  — see Modal logs).
+- **Upload guard fires (files already exist).** The test name / model id
+  collides with something already on HF. Verify the user actually wants to
+  replace existing refs; if so, re-run the upload with `--force`. If not,
+  rename the model id in `*_MODEL_TO_PARAMS` and re-seed.
+- **Quality looks wrong in step 4.** Abort. The mp4s stay on disk for
+  inspection. The fix is usually in the test's params (resolution, steps,
+  seed) — edit the test, then re-run the skill.
+
+## Design notes (for future skill maintainers)
+
+- The skill deliberately runs on Modal, **not** locally, because the CI
+  runner is L40S. Seeding from a different GPU SKU produces refs that CI's
+  L40S runs can't match (SSIM drifts across SKUs).
+- The skill is default-tier only. `full_quality` refs are seeded by a
+  separate, deliberate operation — they double runtime and aren't what CI
+  gates on.
+- The overwrite guard in `reference_videos_cli.py upload` is default-on
+  specifically because this skill exists. Re-seeding is a distinct operation
+  that requires explicit `--force`.
+
+## References
+
+- `fastvideo/tests/modal/ssim_test.py` — Modal orchestrator; see
+  `--sync-generated-to-volume`, `--generated-volume-subdir`,
+  `--skip-reference-download`, `--no-fail-fast`.
+- `fastvideo/tests/ssim/reference_videos_cli.py` — `copy-local`, `upload`
+  (with `--model-id`, `--force`), `download`, `ensure` subcommands.
+- `fastvideo/tests/ssim/README.md` — reference layout, HF repo conventions.
+- `fastvideo/tests/ssim/inference_similarity_utils.py` —
+  `run_text_to_video_similarity_test` + `_build_init_kwargs`: what each test
+  config passes to `VideoGenerator.from_pretrained`.
+
+## Changelog
+
+| Date | Change |
+|------|--------|
+| 2026-04-17 | Initial version (Modal sync-to-volume flow). |
+| 2026-04-21 | Rewrite: single-test scope, explicit user-review pause, per-`model_id` upload, HF overwrite guard. Dropped `scripts/seed_ssim.sh`. |
+| 2026-04-21 | Post-first-run fixes: `modal volume get` needs `--force` when parent exists; download tree has an extra `generated_videos/` level so `--generated-dir` must reflect it. |

.agents/workflows/sync-dashboard.md

@@ -1,67 +0,0 @@
----
-description: Synchronize the STATUS.md dashboard by scanning .agents/ directories
----
-
-# Sync Dashboard
-
-Updates `.agents/STATUS.md` by scanning the skills, workflows, memory, lessons,
-and exploration directories to reflect what actually exists on disk.
-
-## When to Use
-
-- After adding, removing, or renaming any file in `.agents/`.
-- Periodically (e.g., at end of each conversation session).
-- When the dashboard feels out of date.
-
-## Steps
-
-### 1. Scan directories
-
-List all files in each directory:
-
-```bash
-echo "=== Skills ==="
-ls -1 .agents/skills/*.md 2>/dev/null | grep -v SKILL_TEMPLATE
-
-echo "=== Workflows ==="
-ls -1 .agents/workflows/*.md 2>/dev/null
-
-echo "=== Memory ==="
-ls -1 .agents/memory/*.md 2>/dev/null
-ls -1 .agents/memory/related-work/*.md 2>/dev/null | grep -v README
-
-echo "=== Lessons ==="
-ls -1 .agents/lessons/*.md 2>/dev/null | grep -v README
-
-echo "=== Exploration ==="
-ls -1 .agents/exploration/*.md 2>/dev/null | grep -v README
-```
-
-### 2. Compare with STATUS.md
-
-For each file found:
-- If it's in STATUS.md → leave it (preserve status/trust/tested fields).
-- If it's NOT in STATUS.md → add it with status `🔴 Stub`, trust `None`, tested `❌`.
-
-For each entry in STATUS.md:
-- If the file no longer exists → mark it as `❌ Removed` or delete the row.
-
-### 3. Update counts
-
-Recalculate the summary table at the top:
-- Count files per category.
-- Count by status (Ready, Draft, Stub).
-
-### 4. Update timestamp
-
-Set `_Last synced: <current date>_` at the top of STATUS.md.
-
-### 5. Review
-
-Read through the updated STATUS.md for accuracy. Flag anything that looks wrong.
-
-## Notes
-
-- Do NOT change trust levels during sync — those are set manually after testing.
-- Do NOT change status during sync — status changes require actual validation.
-- This workflow only handles structural sync (file existence), not content review.

.buildkite/performance-benchmarks/tests/wan-t2v-1.3b.json

@@ -29,8 +29,8 @@
     "Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of a bustling street food market. The scene captures a mix of humor and authenticity. Mid-shot framing, vibrant lighting."
   ],
   "run_config": {
-    "num_warmup_runs": 1,
-    "num_measurement_runs": 3,
+    "num_warmup_runs": 2,
+    "num_measurement_runs": 5,
     "required_gpus": 2
   },
   "thresholds": {

.buildkite/scripts/pr_test.sh

@@ -15,8 +15,21 @@ log "Project root: $PROJECT_ROOT"
 # Install Modal if not available
 if ! python3 -m modal --version &> /dev/null; then
     log "Modal not found, installing..."
-    python3 -m pip install modal
-    
+    if ! command -v uv &> /dev/null; then
+        log "uv not found, bootstrapping..."
+        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
+            log "Error: Failed to bootstrap uv via astral.sh installer."
+            exit 1
+        fi
+        export PATH="$HOME/.local/bin:$PATH"
+        if ! command -v uv &> /dev/null; then
+            log "Error: uv still not on PATH after bootstrap."
+            exit 1
+        fi
+    fi
+    # --break-system-packages preserves prior `pip install --user` semantics on PEP 668 agents.
+    uv pip install --system --break-system-packages modal
+
     # Verify installation
     if ! python3 -m modal --version &> /dev/null; then
         log "Error: Failed to install modal. Please install it manually."
@@ -63,7 +76,72 @@ EFFECTIVE_PR=${BUILDKITE_PULL_REQUEST:-false}
 if [ "$EFFECTIVE_PR" = "false" ] && [ -n "${PR_NUMBER:-}" ]; then
     EFFECTIVE_PR=$PR_NUMBER
 fi
-MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR IMAGE_VERSION=$IMAGE_VERSION"
+MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR BUILDKITE_BRANCH=${BUILDKITE_BRANCH:-} TEST_SCOPE=${TEST_SCOPE:-} IMAGE_VERSION=$IMAGE_VERSION"
+
+POST_RUN_HOOK=""
+
+upload_performance_artifacts() {
+    SHORT_SHA=${BUILDKITE_COMMIT:0:7}
+    LOCAL_DIR="downloaded_reports"
+
+    _download_reports() {
+        log "Downloading perf_reports/ from Modal Volume..."
+        mkdir -p "$LOCAL_DIR"
+        if ! modal volume get hf-model-weights "perf_reports/" "$LOCAL_DIR"; then
+            log "Error: Failed to download perf_reports/ from Modal Volume."
+            return 1
+        fi
+    }
+
+    _upload_dashboard() {
+        local target
+        target=$(find "$LOCAL_DIR" -name "dashboard_${SHORT_SHA}_*" | head -n 1)
+        log "TARGET dashboard: '$target'"
+
+        if [ -n "$target" ]; then
+            log "Found dashboard: $target. Uploading to Buildkite..."
+            buildkite-agent artifact upload "$target"
+            buildkite-agent annotate --style info --context "perf-dashboard" < "$target"
+        else
+            log "Warning: Could not find a dashboard file matching $SHORT_SHA"
+        fi
+    }
+
+    _upload_perf_summary() {
+        local target
+        target=$(find "$LOCAL_DIR" -name "perf_${SHORT_SHA}_*" | head -n 1)
+        log "TARGET perf summary: '$target'"
+
+        if [ -n "$target" ]; then
+            log "Found perf summary: $target. Uploading to Buildkite..."
+            buildkite-agent artifact upload "$target"
+            buildkite-agent annotate --style info --context "perf-summary" < "$target"
+        else
+            log "Warning: Could not find a perf summary file matching $SHORT_SHA"
+        fi
+    }
+
+    _cleanup_modal_volume() {
+        log "Cleaning up perf_reports/ from Modal Volume..."
+        if modal volume rm hf-model-weights "perf_reports/" --recursive; then
+            log "Successfully deleted perf_reports/ from Modal Volume."
+        else
+            log "Warning: Failed to delete perf_reports/ from Modal Volume. Manual cleanup may be required."
+        fi
+    }
+
+    _cleanup_local() {
+        log "Cleaning up local download directory..."
+        rm -rf "$LOCAL_DIR"
+    }
+
+    # --- Main flow ---
+    _download_reports || { _cleanup_local; return 1; }
+    _upload_dashboard
+    _upload_perf_summary
+    _cleanup_modal_volume
+    _cleanup_local
+}
 
 case "$TEST_TYPE" in
     "encoder")
@@ -124,8 +202,9 @@ case "$TEST_TYPE" in
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_lora_extraction_tests"
         ;;
     "performance")
-        log "Running performance tests..."
+        log "Running performance tests on Modal..."
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_performance_tests"
+        POST_RUN_HOOK="upload_performance_artifacts"
         ;;
     "api_server")
         log "Running API server integration tests..."
@@ -147,5 +226,10 @@ else
     log "Error: Modal test failed with exit code: $TEST_EXIT_CODE"
 fi
 
+if [ -n "$POST_RUN_HOOK" ]; then
+    log "Executing post-run hook: $POST_RUN_HOOK"
+    "$POST_RUN_HOOK"
+fi
+
 log "=== Test execution completed with exit code: $TEST_EXIT_CODE ==="
 exit $TEST_EXIT_CODE

.buildkite/scripts/pre_commit.sh

@@ -13,8 +13,21 @@ log "Project root: $PROJECT_ROOT"
 
 if ! python3 -m pre_commit --version &> /dev/null; then
     log "pre-commit not found, installing..."
-    python3 -m pip install --user pre-commit==4.0.1
-    
+    if ! command -v uv &> /dev/null; then
+        log "uv not found, bootstrapping..."
+        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
+            log "Error: Failed to bootstrap uv via astral.sh installer."
+            exit 1
+        fi
+        export PATH="$HOME/.local/bin:$PATH"
+        if ! command -v uv &> /dev/null; then
+            log "Error: uv still not on PATH after bootstrap."
+            exit 1
+        fi
+    fi
+    # --break-system-packages preserves prior `pip install --user` semantics on PEP 668 agents.
+    uv pip install --system --break-system-packages pre-commit==4.0.1
+
     if ! python3 -m pre_commit --version &> /dev/null; then
         log "Error: Failed to install pre-commit."
         exit 1

.github/workflows/infra-docs.yml

@@ -37,10 +37,11 @@ jobs:
         with:
           python-version: '3.12'
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
       - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-mkdocs.txt
+        run: uv pip install --system -r requirements-mkdocs.txt
 
       - name: Setup Pages
         uses: actions/configure-pages@v4

.github/workflows/publish-fastvideo.yml

@@ -56,10 +56,11 @@ jobs:
         with:
           python-version: '3.10'
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
       - name: Install build dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install build twine wheel
+        run: uv pip install --system build twine wheel
 
       - name: Build package
         run: |

.github/workflows/publish-kernel.yml

@@ -131,11 +131,13 @@ jobs:
           clang-11 --version
           nvcc --version
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
       - name: Install PyTorch ${{ matrix.torch-cuda.torch-version }}+cu${{ matrix.torch-cuda.cuda-version }}
         run: |
-          pip install --upgrade pip
-          pip install typing-extensions==4.12.2
-          pip install --no-cache-dir torch==${{ matrix.torch-cuda.torch-version }} --index-url https://download.pytorch.org/whl/${{matrix.torch-cuda.torch-cuda-short}}
+          uv pip install --system typing-extensions==4.12.2
+          uv pip install --system --no-cache-dir torch==${{ matrix.torch-cuda.torch-version }} --index-url https://download.pytorch.org/whl/${{matrix.torch-cuda.torch-cuda-short}}
           nvcc --version
           python --version
           python -c "import torch; print('PyTorch:', torch.__version__)"
@@ -145,20 +147,20 @@ jobs:
       - name: Build wheel
         run: |
           export PYTHONPATH=$GITHUB_WORKSPACE:$PYTHONPATH
-          
-          pip install setuptools ninja packaging wheel triton scikit-build-core cmake build
-          
+
+          uv pip install --system setuptools ninja packaging wheel triton scikit-build-core cmake build
+
           cd fastvideo-kernel
           git submodule update --init --recursive  # Ensure ThunderKittens submodule is initialized
           # Release builds are produced on GPU-less runners, so force-enable TK and target Hopper.
           export TORCH_CUDA_ARCH_LIST="9.0a"
           export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DCMAKE_CUDA_ARCHITECTURES=90a"
-          
+
           # Build standard wheel (no local version suffix) for PyPI
           python -m build --wheel --outdir dist
-          
+
           # Fix the wheel to be manylinux compliant
-          pip install auditwheel
+          uv pip install --system auditwheel
           # Point auditwheel at torch libs, but do not vendor them into the wheel.
           TORCH_LIB_DIR=$(python - <<'PY'
           import os
@@ -211,10 +213,13 @@ jobs:
           pattern: 'fastvideo_kernel-py*'
           merge-multiple: true
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
       - name: Build source distribution
         run: |
-          pip install build scikit-build-core cmake ninja
-          
+          uv pip install --system build scikit-build-core cmake ninja
+
           cd fastvideo-kernel
           # We don't need full CUDA/Torch to just package the source (sdist)
           python -m build --sdist --outdir dist

.pre-commit-config.yaml

@@ -7,20 +7,13 @@ exclude: |
     fastvideo-kernel/.*|
     assets/.*|
     tests/.*|
-    demo/.*|
-    predict\.py|
     scripts/.*|
-    assets/prompts/.*|
-    fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
     fastvideo/models/.*|
-    fastvideo/sample/.*|
-    fastvideo/train\.py|
-    fastvideo/utils/.*|
     examples/.*|
+    \.agents/.*|
     .github/workflows/publish-fastvideo.yml|
-    .github/workflows/_template-build-image.yml|
-    docs/source/inference/support_matrix.md
+    .github/workflows/_template-build-image.yml
   )
 repos:
 - repo: https://github.com/google/yapf

AGENTS.md

@@ -11,7 +11,7 @@
 - Static assets: `assets/` (including `assets/images/`, `assets/videos/`, and `assets/prompts/`) and `comfyui/assets/`.
 
 ## Build, Test, and Development Commands
-- `uv pip install -e .[dev]`: editable install with lint/test extras.
+- `uv pip install -e ".[dev]"`: editable install with lint/test extras.
 - `pre-commit install --hook-type pre-commit --hook-type commit-msg`: enable local hooks.
 - `pre-commit run --all-files`: run formatter/lint/type/spelling checks.
 - `pytest tests/`: run top-level test suite.
@@ -23,7 +23,8 @@
 - Python 3.10+; 4-space indentation; keep code and imports readable and explicit.
 - Style tools are configured in `pyproject.toml` and `.pre-commit-config.yaml`:
   - `yapf` (format), `ruff` (lint, auto-fix), `mypy` (typing), `codespell`.
-- Target line length is 80.
+- Lint via `pre-commit run --files <changed paths>` (or `pre-commit run --all-files` for a full sweep) before committing. Do not shell out to `yapf`/`ruff`/`codespell`/`mypy` directly — pre-commit chains them with the project's config and respects the `.pre-commit-config.yaml` excludes (e.g. `fastvideo/tests/` is intentionally skipped). If pre-commit reports `(no files to check)` for your paths, that exclude is deliberate — don't bypass it.
+- Target line length is 120 (configured in `pyproject.toml` for ruff, yapf, and isort).
 - Naming: `snake_case` for functions/files, `PascalCase` for classes, `UPPER_SNAKE_CASE` for constants.
 
 ## Testing Guidelines
@@ -54,3 +55,31 @@ This repository is agent-friendly. Before doing any work, read:
 If you are exploring a new procedure that has no existing SOP, document your
 progress in `.agents/exploration/` and flag it for review at the end of your
 session.
+
+## Per-Directory AGENTS.md
+
+Local guidance lives next to the code. Read the in-scope file before editing:
+
+| Directory | What it covers |
+|-----------|----------------|
+| `fastvideo/AGENTS.md` | Core package map, public API, registry-driven model dispatch |
+| `fastvideo/configs/AGENTS.md` | Arch + pipeline config dataclasses, `param_names_mapping` |
+| `fastvideo/models/AGENTS.md` | DiT / VAE / encoder / scheduler / loader layout (pre-commit excluded) |
+| `fastvideo/layers/AGENTS.md` | Tensor-parallel linear/attention layer rules for ports |
+| `fastvideo/attention/AGENTS.md` | Backend registry + env-var override |
+| `fastvideo/pipelines/AGENTS.md` | Stage ABC, `basic/<model>/`, `preprocess/`, presets |
+| `fastvideo/training/AGENTS.md` | Legacy monolithic pipelines (frozen for existing models) |
+| `fastvideo/train/AGENTS.md` | New modular trainer (methods × models × callbacks, YAML) |
+| `fastvideo/tests/AGENTS.md` | Test taxonomy, conftest, pre-commit-excluded path |
+| `fastvideo/tests/ssim/AGENTS.md` | GPU SSIM regression authoring + reference video sync |
+| `scripts/checkpoint_conversion/AGENTS.md` | Adding a converter for a new HF/official checkpoint |
+
+## Critical: Two Training Stacks Coexist
+
+- `fastvideo/training/` — legacy, monolithic per-model `*_training_pipeline.py` and
+  `*_distillation_pipeline.py`. Still authoritative for shipped models.
+- `fastvideo/train/` — new modular framework (composable methods × models × callbacks
+  driven by YAML). Preferred for new training work.
+
+Pick the matching stack before editing. Do not migrate a pipeline between them
+without an explicit ask — the conventions and config surfaces differ.

CLEANUP.md

@@ -0,0 +1,182 @@
+# `.agents/` Cleanup Log — Phase 1 (Deletes Only)
+
+**Status:** TEMPORARY — delete this file after the cleanup is reviewed/committed.
+**Date:** 2026-05-04
+**Branch:** `will/ltx2_sr_port`
+**Scope:** Phase 1 of the `.agents/` cleanup plan (deletes only; no rewrites or additions).
+
+For the full multi-phase plan, see the prior session analysis. This file tracks
+exactly what got deleted, why, and what cross-references still point at deleted
+content (to fix in a future phase).
+
+---
+
+## Deletions executed
+
+### Files deleted
+
+| Path | Size | Reason |
+|---|---|---|
+| `.agents/STATUS.md` | 3.85 KB | Stale dashboard, last synced 2026-03-02. Counts wrong (claimed 8 skills/4 workflows/4 memory; actual 9/5/5). References old snake_case filenames (`codebase_map.md`/`experiment_journal.md`) that don't exist. Hand-maintained derivative of `.agents/{memory,skills}/index.jsonl` — strictly redundant. |
+| `.agents/exploration/pr-link-review.md` | 1.11 KB | Status: "promoted" to `.agents/skills/review-pr-link/`. Per `.agents/exploration/README.md` lifecycle, promoted exploration logs should not linger after the skill exists. |
+| `.agents/workflows/sync-dashboard.md` | 1.87 KB | SOP for maintaining `STATUS.md` (which is also deleted). Contained obsolete file paths (`.agents/skills/launch-experiment.md` flat layout vs. actual `<skill>/SKILL.md` per-dir layout). Has never been run successfully (judging by stale dates everywhere). |
+
+### Skill directories deleted
+
+| Path | Size | Reason |
+|---|---|---|
+| `.agents/skills/index-related-work/` | 2.18 KB | Vapor-skill operating on the empty `.agents/memory/related-work/` registry. Never used (the registry has zero entries despite ~6 weeks since skill creation). Re-add when the related-work catalog gains entries. |
+| `.agents/skills/search-related-work/` | 1.91 KB | Same: vapor-skill against empty registry. The skill description literally requires "The related work index has entries" as a prerequisite, and there are none. |
+
+**Total deleted: 5 items, ~10.9 KB.**
+
+### Registry updates
+
+| File | Change |
+|---|---|
+| `.agents/skills/index.jsonl` | Removed entries for `index-related-work` and `search-related-work`. Was 9 entries; now 7. |
+
+### Symlink hygiene
+
+`.agents/scripts/sync-skills.sh` was run to prune now-stale symlinks under
+`.claude/skills/` that pointed at the deleted skill directories. Output captured
+in the run log.
+
+---
+
+## What was KEPT (despite being candidates)
+
+| Path | Why kept |
+|---|---|
+| `.agents/scripts/sync-skills.sh` | User explicitly requested keep. **Verified**: this script is INDEPENDENT of STATUS.md / sync-dashboard.md. It mirrors `.agents/skills/` → `.claude/skills/` via symlinks for Claude Code skill discovery. Self-contained, useful, prunes its own stale symlinks. |
+| `.agents/memory/related-work/README.md` | Empty placeholder, but the schema/template is reusable. Kept for when first related-work entry is added. |
+| `.agents/memory/experiment-journal/README.md` | Same: empty placeholder with template; kept for when journaling begins. |
+| `.agents/lessons/README.md` | Same: empty placeholder, reusable schema. |
+| `.agents/exploration/README.md` | Active template for new exploration logs. Kept. |
+
+---
+
+## Remaining broken cross-references (FOLLOW-UP NEEDED)
+
+These files still reference deleted content. **NOT fixed in Phase 1** — track for
+the next pass (Phase 2: rewrites/dedupe).
+
+### References to deleted `STATUS.md`
+
+| Referencing file | Action needed |
+|---|---|
+| `.agents/onboarding/README.md` | Quick-reference tree (line ~65) lists `STATUS.md ← dashboard: completeness & trust of all components`. Remove that line + the `ONBOARDING.md` typo (file is `README.md`). |
+
+### References to deleted `pr-link-review.md`
+
+| Referencing file | Action needed |
+|---|---|
+| `.agents/memory/dreamverse-integration/state.md` | "Untracked but present" / "Source docs (archived)" sections still mention `pr-link-review.md` as kept. Update to reflect deletion. |
+| `.agents/memory/dreamverse-integration/README.md` | Same — table row for `pr-link-review.md` says "kept in exploration dir". Update or remove the row. |
+
+### References to deleted skills (`index-related-work`, `search-related-work`)
+
+| Referencing file | Action needed |
+|---|---|
+| `.agents/memory/related-work/README.md` | Says "Use the `index-related-work` skill". Either remove that hint or note "skill removed; re-add when registry has entries". |
+| `.agents/workflows/evaluation-development.md` | Step 1 says "Search `.agents/memory/related-work/` for existing evaluation approaches" — that's still valid (manual search). No change needed. |
+
+### References to deleted `sync-dashboard.md`
+
+| Referencing file | Action needed |
+|---|---|
+| `.agents/memory/evaluation-registry/README.md` | Doesn't reference sync-dashboard directly. No change. |
+| `.agents/STATUS.md` | Already being deleted. |
+
+---
+
+## Other registry inconsistencies discovered (NOT FIXED in Phase 1)
+
+While editing `.agents/skills/index.jsonl`, two skill directories were found
+that exist on disk but **are not registered** in `index.jsonl`:
+
+| Skill dir | Status | Why missing from index |
+|---|---|---|
+| `.agents/skills/diagnose-ssim-failure/` | Untracked locally; NOT on `origin/main`. 12.3 KB SKILL.md + `scripts/compare_latent_pt.py`. Recent mtime (2026-05-01). | Created in a prior session but the registration step was skipped. |
+| `.agents/skills/review-pr-link/` | Untracked locally; NOT on `origin/main`. 2.9 KB SKILL.md + `scripts/prepare_pr_review.py` + `agents/openai.yaml`. The promotion target of the deleted `pr-link-review.md` exploration log. | Skipped registration when promoted from exploration log. |
+
+Both skills are functional and exposed via `sync-skills.sh` symlinks (just verified in
+`.claude/skills/`), but agents reading `index.jsonl` to discover skills will miss them.
+
+**Action for Phase 2**: Add entries to `.agents/skills/index.jsonl` for both,
+likely with `trust: medium` since they have working scripts and recent use.
+
+---
+
+## Skill registry parity check
+
+After Phase 1, `.agents/skills/` contains 9 directories but `index.jsonl` lists 7:
+
+| In `index.jsonl` | On disk |
+|---|---|
+| ✓ launch-experiment | ✓ launch-experiment/ |
+| ✓ monitor-experiment | ✓ monitor-experiment/ |
+| ✓ summarize-run | ✓ summarize-run/ |
+| ✓ log-experiment | ✓ log-experiment/ |
+| ✓ evaluate-video-quality | ✓ evaluate-video-quality/ |
+| ✓ seed-ssim-references | ✓ seed-ssim-references/ |
+| ✓ reseed-ssim-references | ✓ reseed-ssim-references/ |
+| ❌ (missing) | ⚠ diagnose-ssim-failure/ |
+| ❌ (missing) | ⚠ review-pr-link/ |
+
+`.claude/skills/` symlinks (the runtime-discoverable surface) include all 9 ✓.
+
+---
+
+## Phase 2+ items (NOT executed in this session)
+
+For future cleanup sessions, the prior plan identified:
+
+**Phase 2 (rewrites)**:
+- Rewrite `.agents/onboarding/worldmodel-training/README.md` to drop ~50% structural duplication with `codebase-map/README.md`
+- Refresh `.agents/memory/codebase-map/README.md` (last updated 2026-03-08; missing `fastvideo/api/`, `fastvideo/entrypoints/streaming/`, etc.)
+- Refresh `.agents/memory/evaluation-registry/README.md` (last updated 2026-03-02; references old `evaluation_registry.md` filename)
+- Merge `.agents/workflows/experiment-journaling.md` into `experiment-lifecycle.md` (one SOP per workflow)
+- Fix the broken cross-references listed above
+
+**Phase 3 (additions)**:
+- `fastvideo/api/AGENTS.md`
+- `fastvideo/entrypoints/AGENTS.md`
+- `tests/AGENTS.md` (top-level, distinct from `fastvideo/tests/AGENTS.md`)
+- `fastvideo/distributed/AGENTS.md`
+- `examples/AGENTS.md`
+- `docs/AGENTS.md`
+- `benchmarks/AGENTS.md`
+
+**Phase 4 (registry)**:
+- Add `.agents/workflows/index.jsonl`
+- Standardize all three index.jsonl schemas
+
+**Phase 5 (skills quality)**:
+- Promote tested skills (`seed-ssim-references`, `reseed-ssim-references`, `diagnose-ssim-failure`, `review-pr-link`) from `trust: low` to `trust: medium`
+- Mark untested skills (`launch-experiment`, `monitor-experiment`, `summarize-run`, `log-experiment`, `evaluate-video-quality`) explicitly with their gating prerequisite (e.g. "operates on empty registry")
+
+---
+
+## Recovery
+
+All deletions are local (`will/ltx2_sr_port`, not committed). To restore any
+deleted file:
+
+```bash
+git restore --source=HEAD .agents/STATUS.md
+git restore --source=HEAD .agents/exploration/pr-link-review.md
+git restore --source=HEAD .agents/workflows/sync-dashboard.md
+git restore --source=HEAD .agents/skills/index-related-work/SKILL.md
+git restore --source=HEAD .agents/skills/search-related-work/SKILL.md
+```
+
+---
+
+## When to delete THIS file
+
+Once:
+1. The Phase 1 deletions are committed (or merged), AND
+2. Phase 2 (broken cross-reference cleanup) is also committed,
+
+remove this file. Its purpose is transient bookkeeping for a multi-phase cleanup.

CO-AUTHORS.md

@@ -0,0 +1,88 @@
+# Co-Authors — `will/ltx2_sr_port` Stack
+
+**Status:** PERMANENT — keep around as the source of truth for who collaborated on this work, even after the stack merges.
+**Last updated:** 2026-05-04
+
+This file documents the human co-authors credited on every commit in the
+`will/ltx2_sr_port` stack and its 10 split PRs. The 4 collaborators below
+worked on the FastVideo-internal precursor of this code (LTX-2 streaming
+server, NVFP4 wire-up, GPU pool, prompt enhancer, etc.) and are credited as
+co-authors on the public-side upstream commits via Git's standard
+[`Co-authored-by`](https://docs.github.com/en/pull-requests/committing-changes-to-your-project/creating-and-editing-commits/creating-a-commit-with-multiple-authors)
+trailer convention.
+
+The trailers are added to every commit on `will/ltx2_sr_port` (see
+[`STACK.md`](STACK.md)), which means GitHub will:
+
+- Show the 4 co-authors on every commit detail page
+- Show them on the merge commit / squash commit summary
+- Display their avatars in the PR's "Contributors" sidebar
+- Surface them in [`/contributors`](https://github.com/hao-ai-lab/FastVideo/contributors) once the stack lands
+
+## Co-author roster
+
+| GitHub user | Real name | GitHub ID | Trailer email |
+|---|---|---|---|
+| [`@Davids048`](https://github.com/Davids048) | Junda (David) Su | 90978028 | `90978028+Davids048@users.noreply.github.com` |
+| [`@RandNMR73`](https://github.com/RandNMR73) | Matthew Noto | 99706358 | `99706358+RandNMR73@users.noreply.github.com` |
+| [`@XOR-op`](https://github.com/XOR-op) | (unset) | 17672363 | `17672363+XOR-op@users.noreply.github.com` |
+| [`@jzhang38`](https://github.com/jzhang38) | Zhang Peiyuan | 42993249 | `42993249+jzhang38@users.noreply.github.com` |
+
+## Why no-reply emails
+
+GitHub's `<id>+<username>@users.noreply.github.com` form is the most reliable
+way to link a `Co-authored-by` trailer to a GitHub account. It:
+
+- Always works regardless of whether the user has a public verified email
+- Survives the user changing their primary email
+- Doesn't expose anyone's personal email to git history
+- Is the format GitHub itself produces when you click "Add co-author" in the
+  web UI
+
+(All 4 collaborators have this email already used in `FastVideo-internal`
+git history, verified via `git log --all` on that repo.)
+
+## Trailer block (copy-paste ready)
+
+The trailers added to every commit on `will/ltx2_sr_port`:
+
+```
+Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>
+Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>
+Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>
+Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>
+```
+
+## How the trailers were applied
+
+```bash
+git rebase --exec '
+  git commit --amend --no-edit \
+    --trailer "Co-authored-by: Junda (David) Su <90978028+Davids048@users.noreply.github.com>" \
+    --trailer "Co-authored-by: Matthew Noto <99706358+RandNMR73@users.noreply.github.com>" \
+    --trailer "Co-authored-by: XOR-op <17672363+XOR-op@users.noreply.github.com>" \
+    --trailer "Co-authored-by: Zhang Peiyuan <42993249+jzhang38@users.noreply.github.com>"
+' origin/main will/ltx2_sr_port
+```
+
+Git's `--trailer` flag is idempotent (it dedupes by the full `key: value`
+string), so re-running the rebase is safe and won't add duplicates.
+
+## How to add a new co-author later
+
+1. Add the user to the roster table above.
+2. Append their `Co-authored-by` line to the trailer block.
+3. Re-run the rebase command above on `will/ltx2_sr_port` — git's
+   trailer dedupe handles the existing 4; the new one gets appended.
+4. Re-slice all 10 split branches per [`STACK.md`](STACK.md).
+5. Force-push `will/api_7.6`, `will/api_7.7`, and `will/ltx2_sr_port`.
+
+## What we do NOT add
+
+Per [`AGENTS.md`](AGENTS.md):
+
+> Never add any coding agent or models such as Claude (or Claude Code), GPT,
+> Codex or others as a co-author in commits or PRs.
+
+So no `Co-authored-by: Claude <noreply@anthropic.com>` or similar. Only
+human collaborators.

benchmarks/fvd/feature_extractors.py

@@ -128,7 +128,7 @@ class CLIPFeatureExtractor(BaseFeatureExtractor):
 
     def __init__(self, device: str = 'cuda', model_name: str = "openai/clip-vit-base-patch32"):
         if not TRANSFORMERS_AVAILABLE:
-            raise ImportError("Please install transformers: pip install transformers")
+            raise ImportError("Please install transformers: uv pip install transformers")
         super().__init__(device)
         self.processor = CLIPProcessor.from_pretrained(model_name)
         self.model = CLIPModel.from_pretrained(model_name).to(self.device)
@@ -171,7 +171,7 @@ class VideoMAEFeatureExtractor(BaseFeatureExtractor):
 
     def __init__(self, device: str = 'cuda', model_name: str = "MCG-NJU/videomae-base"):
         if not TRANSFORMERS_AVAILABLE:
-            raise ImportError("Please install transformers: pip install transformers")
+            raise ImportError("Please install transformers: uv pip install transformers")
         super().__init__(device)
         self.model = VideoMAEModel.from_pretrained(model_name).to(self.device)
         self.model.eval()

benchmarks/fvd/i3d_model.py

@@ -57,7 +57,7 @@ class I3DFeatureExtractor(nn.Module):
         except Exception as e:
             raise RuntimeError(f"Failed to load I3D model from Hugging Face Hub. Error: {e}\n"
                                f"Ensure you have internet connection and huggingface_hub installed:\n"
-                               f"pip install huggingface_hub") from e
+                               f"uv pip install huggingface_hub") from e
 
     def preprocess(self, videos: torch.Tensor) -> torch.Tensor:
         """

benchmarks/scripts/run.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 # 1. Install missing dependency
-pip install -q opencv-python-headless transformers huggingface_hub
+uv pip install -q opencv-python-headless transformers huggingface_hub
 
 # 2. Run FVD script
 python benchmarks/fvd/run_fvd.py

benchmarks/scripts/setup_fvd.sh

@@ -1,4 +1,4 @@
 #!/bin/bash
 
 # 1. Install missing dependency
-pip install -q opencv-python-headless
+uv pip install -q opencv-python-headless

comfyui/README.md

@@ -38,10 +38,10 @@ cp -r /path/to/FastVideo/comfyui /path/to/ComfyUI/custom_nodes/FastVideo
 
 #### Install dependencies:
 
-Currently, the only dependency is `fastvideo`, which can be installed using pip.
+Currently, the only dependency is `fastvideo`, which can be installed with `uv`.
 
 ```bash
-pip install fastvideo
+uv pip install fastvideo
 ```
 
 #### Install missing custom nodes:

docker/Dockerfile.python3.10

@@ -42,15 +42,15 @@ RUN source $HOME/.local/bin/env && \
     uv venv --python 3.10 --seed /opt/venv && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
-    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp310-cp310-linux_x86_64.whl
+    uv pip install --no-cache-dir ".[dev]" && \
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.4/flash_attn-2.8.3+cu128torch2.11-cp310-cp310-linux_x86_64.whl
 
 COPY . .
 
 # Install dependencies using uv and set up shell configuration
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
+    uv pip install --no-cache-dir -e ".[dev]" && \
     git config --unset-all http.https://github.com/.extraheader || true && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

docker/Dockerfile.python3.11

@@ -42,15 +42,15 @@ RUN source $HOME/.local/bin/env && \
     uv venv --python 3.11 --seed /opt/venv && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
-    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp311-cp311-linux_x86_64.whl
+    uv pip install --no-cache-dir ".[dev]" && \
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.4/flash_attn-2.8.3+cu128torch2.11-cp311-cp311-linux_x86_64.whl
 
 COPY . .
 
 # Install dependencies using uv and set up shell configuration
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
+    uv pip install --no-cache-dir -e ".[dev]" && \
     git config --unset-all http.https://github.com/.extraheader || true && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

docker/Dockerfile.python3.12

@@ -42,15 +42,15 @@ RUN source $HOME/.local/bin/env && \
     uv venv --python 3.12 --seed /opt/venv && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
-    uv pip install --no-cache-dir .[dev] && \
-    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3+cu128torch2.10-cp312-cp312-linux_x86_64.whl
+    uv pip install --no-cache-dir ".[dev]" && \
+    uv pip install --no-cache-dir https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.4/flash_attn-2.8.3+cu128torch2.11-cp312-cp312-linux_x86_64.whl
 
 COPY . .
 
 # Install dependencies using uv and set up shell configuration
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
+    uv pip install --no-cache-dir -e ".[dev]" && \
     git config --unset-all http.https://github.com/.extraheader || true && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

docker/Dockerfile.python3.12.cuda12.9.1

@@ -42,7 +42,7 @@ RUN source $HOME/.local/bin/env && \
     uv venv --python 3.12 --seed /opt/venv && \
     source /opt/venv/bin/activate && \
     uv pip install --no-cache-dir --upgrade pip && \
-    uv pip install --no-cache-dir .[dev] && \
+    uv pip install --no-cache-dir ".[dev]" && \
     uv pip install --no-cache-dir flash-attn==2.8.3 --no-build-isolation
 
 COPY . .
@@ -50,7 +50,7 @@ COPY . .
 # Install dependencies using uv and set up shell configuration
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[dev] && \
+    uv pip install --no-cache-dir -e ".[dev]" && \
     git config --unset-all http.https://github.com/.extraheader || true && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

docker/rocm7_1.Dockerfile

@@ -43,7 +43,7 @@ COPY . .
 # Install dependencies using uv and set up shell configuration
 RUN source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
-    uv pip install --no-cache-dir -e .[rocm] && \
+    uv pip install --no-cache-dir -e ".[rocm]" && \
     git config --unset-all http.https://github.com/.extraheader || true && \
     echo 'source /opt/venv/bin/activate' >> /root/.bashrc && \
     echo 'if [ -n "$ZSH_VERSION" ] && [ -f ~/.zshrc ]; then . ~/.zshrc; elif [ -f ~/.bashrc ]; then . ~/.bashrc; fi' > /root/.profile

docs/README.md

@@ -6,7 +6,7 @@ This directory contains the FastVideo documentation built with MkDocs.
 
 ```bash
 # Install dependencies
-pip install -r requirements-mkdocs.txt
+uv pip install -r requirements-mkdocs.txt
 
 # Serve docs with live reload (recommended for development)
 mkdocs serve

docs/contributing/coding_agents.md

@@ -296,8 +296,10 @@ Action:
 
 - Add or reuse a numerical parity test that loads the official model and the
   FastVideo model and compares outputs.
-- See examples in `tests/local_tests/` (e.g., `tests/local_tests/upsamplers/`)
-  and the commands in `tests/local_tests/README.md`.
+- See examples in `tests/local_tests/` organized by model family
+  (e.g., `tests/local_tests/sd35/`, `tests/local_tests/ltx2/`,
+  `tests/local_tests/stable_audio/`) and the navigation index in
+  `tests/local_tests/README.md`.
 - If there are discrepancies, add opt‑in logging to both models and compare
   activation summaries (layer output sums, per‑stage logs).
 - First align the loaded weights (validate `param_names_mapping`).
@@ -348,7 +350,8 @@ Purpose:
 
 Action:
 
-- Add a pipeline parity test under `tests/local_tests/pipelines/`.
+- Add a pipeline parity test under `tests/local_tests/<family>/`
+  (e.g., `tests/local_tests/<family>/test_<family>_pipeline_parity.py`).
 - See the [Testing Guide](testing.md) for test conventions.
 
 ### 7) Add user‑facing examples

docs/contributing/developer_env/runpod.md

@@ -99,7 +99,7 @@ cd /FastVideo
 **Install the package**
 
 ```bash
-uv pip install -e .[dev]
+uv pip install -e ".[dev]"
 ```
 
 The Docker image already includes Flash Attention and most heavy dependencies, so this is fast.

docs/contributing/overview.md

@@ -49,7 +49,7 @@ git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
 Install FastVideo in editable mode and set up hooks:
 
 ```bash
-uv pip install -e .[dev]
+uv pip install -e ".[dev]"
 
 # Optional: FlashAttention (builds native kernels)
 uv pip install flash-attn --no-build-isolation -v

docs/design/inference_schema_parity_inventory.yaml

@@ -29,7 +29,15 @@ surfaces:
       vae_cpu_offload: generator.engine.offload.vae
       pin_cpu_memory: generator.engine.offload.pin_cpu_memory
       enable_torch_compile: generator.engine.compile.enabled
-      torch_compile_kwargs: generator.engine.compile.kwargs
+      enable_torch_compile_text_encoder: generator.engine.compile.text_encoder_enabled
+      enable_torch_compile_vae: generator.engine.compile.vae_enabled
+      enable_torch_compile_audio_vae: generator.engine.compile.audio_vae_enabled
+      torch_compile_kwargs: generator.engine.compile.backend,fullgraph,mode,dynamic,extras
+      torch_compile_kwargs_dit: generator.engine.compile.dit_kwargs
+      torch_compile_kwargs_text_encoder: generator.engine.compile.text_encoder_kwargs
+      torch_compile_kwargs_vae: generator.engine.compile.vae_kwargs
+      torch_compile_kwargs_audio_vae: generator.engine.compile.audio_vae_kwargs
+      transformer_quant: generator.engine.quantization.transformer_quant
       disable_autocast: generator.engine.disable_autocast
       enable_stage_verification: generator.engine.enable_stage_verification
       prompt_txt: request.inputs.prompt_path
@@ -40,13 +48,26 @@ surfaces:
       init_weights_from_safetensors_2: generator.pipeline.components.transformer_2_weights
       override_pipeline_cls_name: generator.pipeline.components.override_pipeline_cls_name
       boundary_ratio: request.sampling.boundary_ratio
+      ltx2_vae_tiling: generator.pipeline.vae_tiling
+      refine_enabled: generator.pipeline.preset_overrides.refine.enabled
+      refine_upsampler_path: generator.pipeline.components.upsampler_weights
+      refine_lora_path: generator.pipeline.components.lora_path
+      refine_num_inference_steps: request.stage_overrides.refine.num_inference_steps
+      refine_guidance_scale: request.stage_overrides.refine.guidance_scale
+      refine_add_noise: generator.pipeline.preset_overrides.refine.add_noise
+      ltx2_refine_enabled: generator.pipeline.preset_overrides.refine.enabled
+      ltx2_refine_upsampler_path: generator.pipeline.components.upsampler_weights
+      ltx2_refine_lora_path: generator.pipeline.components.lora_path
+      ltx2_refine_num_inference_steps: request.stage_overrides.refine.num_inference_steps
+      ltx2_refine_guidance_scale: request.stage_overrides.refine.guidance_scale
+      ltx2_refine_add_noise: generator.pipeline.preset_overrides.refine.add_noise
     preset_owned:
-      ltx2_vae_tiling: generator.pipeline.preset_overrides.ltx2.vae_tiling
       ltx2_vae_spatial_tile_size_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_size_in_pixels
       ltx2_vae_spatial_tile_overlap_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_overlap_in_pixels
       ltx2_vae_temporal_tile_size_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_size_in_frames
       ltx2_vae_temporal_tile_overlap_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_overlap_in_frames
       ltx2_initial_latent_path: request.extensions.ltx2.initial_latent_path
+      ltx2_audio_latent_path: request.extensions.ltx2.audio_latent_path
     compatibility_only:
       mode: "Legacy multi-mode FastVideoArgs switch; typed inference config should not expose execution mode."
       inference_mode: "Legacy boolean mirror of mode; kept only through adapters while FastVideoArgs remains."
@@ -56,6 +77,12 @@ surfaces:
       VSA_sparsity: "Model-specific inference optimization not yet represented in the typed public schema."
       moba_config_path: "Model-specific MoBA optimization surface not yet represented in the typed public schema."
       master_port: "Executor/bootstrap compatibility field; not part of the canonical inference schema."
+      refine_transformer_path: "Generic stage-2 refine transformer override; no typed equivalent yet."
+      refine_noise_path: "Generic stage-2 refine noise override; no typed equivalent yet."
+      refine_audio_noise_path: "Generic stage-2 refine audio noise override; no typed equivalent yet."
+      ltx2_refine_transformer_path: "LTX-2 refine transformer carrier; no typed equivalent yet."
+      ltx2_refine_noise_path: "LTX-2 refine noise carrier; no typed equivalent yet."
+      ltx2_refine_audio_noise_path: "LTX-2 refine audio noise carrier; no typed equivalent yet."
     private_only:
       ray_placement_group: "Ray deployment-only field."
       ray_runtime_env: "Ray deployment-only field."
@@ -306,6 +333,30 @@ surfaces:
         sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
       num_frames_per_block:
         sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
+      audio_channels:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
+      audio_end_in_s:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
+      audio_start_in_s:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
+      max_audio_duration_s:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
+      sample_size:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
+      sampling_rate:
+        sources:
+          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
+          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
     compatibility_only:
       batch_size: "Gen3C inference-only tuning field pending typed batching design."
       gradient_checkpointing: "Gen3C inference-only compatibility field pending typed batching design."
@@ -354,6 +405,8 @@ surfaces:
       return_frames: request.output.return_frames
       return_trajectory_latents: request.runtime.return_trajectory_latents
       return_trajectory_decoded: request.runtime.return_trajectory_decoded
+      continuation_state: request.state
+      return_continuation_state: request.output.return_state
     preset_owned:
       t_thresh: request.stage_overrides.refine.t_thresh
       spatial_refine_only: request.stage_overrides.refine.spatial_refine_only
@@ -378,6 +431,18 @@ surfaces:
       ltx2_stg_scale_audio: request.extensions.ltx2.stg_scale_audio
       ltx2_stg_blocks_video: request.extensions.ltx2.stg_blocks_video
       ltx2_stg_blocks_audio: request.extensions.ltx2.stg_blocks_audio
+      ltx2_images: request.extensions.ltx2.images
+      ltx2_image_crf: request.stage_overrides.refine.image_crf
+      ltx2_conditioning_latent_stage1: request.extensions.ltx2.conditioning_latent_stage1
+      ltx2_conditioning_latent_stage2: request.extensions.ltx2.conditioning_latent_stage2
+      ltx2_video_conditions: request.extensions.ltx2.video_conditions
+      audio_start_in_s: request.extensions.stable_audio.audio_start_in_s
+      audio_end_in_s: request.extensions.stable_audio.audio_end_in_s
+      init_audio: request.extensions.stable_audio.init_audio
+      init_audio_strength: request.extensions.stable_audio.init_audio_strength
+      init_noise_level: request.extensions.stable_audio.init_noise_level
+      inpaint_audio: request.extensions.stable_audio.inpaint_audio
+      inpaint_mask: request.extensions.stable_audio.inpaint_mask
     internal_only:
       data_type: "Derived from the request shape and not a public input."
 

docs/design/server_contracts/dynamo.md

@@ -0,0 +1,354 @@
+# Dynamo Native Backend Integration
+
+FastVideo exposes a stable Python API that the
+[ai-dynamo/dynamo](https://github.com/ai-dynamo/dynamo) project consumes
+as a pure-Python import, same tier as `vllm`, `sglang`, `trtllm`.
+
+**FastVideo hosts no Dynamo code.** The backend subpackage
+(`components/src/dynamo/fastvideo/`) lives in the Dynamo repo. This doc
+is the reference integrators copy when standing up that package — it
+mirrors the structure used by `dynamo/components/src/dynamo/sglang/`
+and is known to satisfy the (closed) draft
+[ai-dynamo/dynamo#7544](https://github.com/ai-dynamo/dynamo/pull/7544)
+pattern.
+
+## What FastVideo provides
+
+The public surface Dynamo imports is intentionally small:
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    ContinuationState,
+    GenerationRequest,
+    InputConfig,
+    OutputConfig,
+    SamplingConfig,
+    # Post-PR 7.10:
+    VideoEvent, VideoProgressEvent, VideoPartialEvent, VideoFinalEvent,
+    VideoResult,
+)
+```
+
+| Surface | Availability | Notes |
+| --- | --- | --- |
+| `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` | Today | `typed_kwargs` is a stable subset from `GeneratorConfig` — no flat legacy LTX-2 kwargs (guaranteed after PR 6) |
+| `VideoGenerator.generate(request: GenerationRequest) -> GenerationResult` | Today | Aggregated; Dynamo wraps in `asyncio.to_thread` under `asyncio.Lock` |
+| `VideoGenerator.generate_async(request) -> AsyncGenerator[VideoEvent, None]` | **PR 7.10** | Canonical execution substrate; sync wrapper reroutes through this |
+| `VideoGenerator.default_health_check_request() -> GenerationRequest` | **PR 7.10** | 256x256 / 8 frames / 1 step; lets Dynamo build its health payload without knowing any FastVideo internals |
+| `fastvideo.api.GenerationRequest` / `SamplingConfig` / `InputConfig` | Today | Stable public dataclasses |
+| `fastvideo.api.ContinuationState` | Today (PR 7) | JSON-safe envelope; kind-versioned payloads |
+| `fastvideo.api.VideoResult` | Today | `frames`, `video_path`, `state`, `metadata` |
+| `config_to_dict(cfg)` | Today | Used by Dynamo's `dump_config(path, config)` |
+
+## Backend package layout
+
+Modeled on `components/src/dynamo/sglang/`:
+
+```
+components/src/dynamo/fastvideo/
+├── __init__.py
+├── __main__.py               # Entry: python -m dynamo.fastvideo
+├── main.py                   # worker() dispatch — mirrors sglang/main.py
+├── args.py                   # FastVideoArgGroup — CLI → GeneratorConfig
+├── backend_args.py           # Dynamo runtime flags (namespace, fs_url, ...)
+├── init_video_generation.py  # init_video_generation(runtime, config)
+├── register.py               # register_video_generation_model() for Dynamo
+├── backend.py                # VideoGenerationWorkerHandler
+├── health_check.py           # FastVideoHealthCheckPayload
+├── protocol.py               # NvCreateVideoRequest ↔ GenerationRequest adapter
+├── request_handlers/
+│   └── video_generation/
+│       └── video_generation_handler.py  # async generate(req, ctx)
+├── README.md
+└── CLAUDE.md                 # per-backend guidance
+```
+
+None of these files live in FastVideo.
+
+## Request/response mapping
+
+Dynamo's `NvCreateVideoRequest` / `VideoNvExt` / `NvVideosResponse` map
+one-to-one onto FastVideo's typed schema:
+
+```
+NvCreateVideoRequest          ->  fastvideo.api.GenerationRequest
+  prompt                      ->    request.prompt
+  size="WxH"                  ->    request.sampling.width, height
+  seconds                     ->    seconds * nvext.fps -> request.sampling.num_frames
+  input_reference             ->    request.inputs.image_path / video_path
+  nvext.fps                   ->    request.sampling.fps
+  nvext.num_frames            ->    request.sampling.num_frames  (overrides seconds*fps)
+  nvext.num_inference_steps   ->    request.sampling.num_inference_steps
+  nvext.guidance_scale        ->    request.sampling.guidance_scale
+  nvext.seed                  ->    request.sampling.seed
+  nvext.negative_prompt       ->    request.sampling.negative_prompt
+  nvext.continuation_state    ->    request.state  (opaque ContinuationState)
+  response_format             ->    (handled by adapter at output)
+
+VideoFinalEvent               ->  NvVideosResponse
+  video_bytes                 ->    data[0].b64_json   (if response_format=b64_json)
+  uploaded URL                ->    data[0].url        (if response_format=url)
+  metadata.inference_time_s   ->    inference_time_s
+  continuation_state          ->    nvext.continuation_state  (reserved for disagg)
+```
+
+## Example: aggregated handler (sync wrap)
+
+Satisfies the PR #7544 shape; works today against
+`VideoGenerator.generate`, upgrades cleanly to `generate_async` after
+PR 7.10.
+
+```python
+# components/src/dynamo/fastvideo/request_handlers/video_generation/
+# video_generation_handler.py
+from __future__ import annotations
+
+import asyncio
+import base64
+import time
+from typing import Any, AsyncGenerator
+
+from fastvideo import VideoGenerator
+from fastvideo.api import GenerationRequest, InputConfig, OutputConfig, SamplingConfig
+
+
+class VideoGenerationWorkerHandler:
+    def __init__(self, generator: VideoGenerator, config, fs=None):
+        self.generator = generator
+        self.config = config
+        self.fs = fs
+        self._lock = asyncio.Lock()  # aggregated = one-in-flight
+
+    async def generate(
+        self,
+        request: dict[str, Any],
+        context,
+    ) -> AsyncGenerator[dict[str, Any], None]:
+        req = _to_fastvideo_request(request)
+        t0 = time.perf_counter()
+        async with self._lock:
+            result = await asyncio.to_thread(self.generator.generate, req)
+        elapsed = time.perf_counter() - t0
+
+        video_bytes = _materialize(result, self.fs, request.get("response_format"))
+        yield {
+            "data": [video_bytes],
+            "inference_time_s": elapsed,
+            "model": request.get("model"),
+        }
+
+
+def _to_fastvideo_request(request: dict[str, Any]) -> GenerationRequest:
+    nvext = request.get("nvext") or {}
+    fps = nvext.get("fps", 24)
+    num_frames = nvext.get("num_frames") or (request.get("seconds") or 4) * fps
+    width, height = _parse_size(request.get("size"))
+
+    return GenerationRequest(
+        prompt=request["prompt"],
+        negative_prompt=nvext.get("negative_prompt"),
+        inputs=InputConfig(
+            image_path=request.get("input_reference"),
+        ),
+        sampling=SamplingConfig(
+            width=width, height=height,
+            num_frames=num_frames, fps=fps,
+            num_inference_steps=nvext.get("num_inference_steps", 50),
+            guidance_scale=nvext.get("guidance_scale", 1.0),
+            seed=nvext.get("seed", 1024),
+        ),
+        output=OutputConfig(save_video=False, return_frames=False),
+        state=nvext.get("continuation_state"),  # public ContinuationState
+    )
+```
+
+`_parse_size` and `_materialize` are small adapter helpers owned by the
+Dynamo backend package; they never appear in FastVideo.
+
+## Example: streaming handler (post-PR 7.10)
+
+```python
+async def generate(self, request, context):
+    req = _to_fastvideo_request(request)
+    async for event in self.generator.generate_async(req):
+        if event.__class__.__name__ == "VideoProgressEvent":
+            yield {"status": "generating", "progress": event.step / event.total_steps}
+        elif event.__class__.__name__ == "VideoFinalEvent":
+            yield {
+                "data": [{"b64_json": base64.b64encode(event.video_bytes).decode()}],
+                "inference_time_s": event.metadata.get("inference_time_s"),
+                "nvext": {"continuation_state": _serialize_state(event.continuation_state)},
+            }
+```
+
+Aggregated and streaming differ only in which events the handler
+forwards; both share one `generate_async` substrate.
+
+## Example: health check
+
+```python
+# components/src/dynamo/fastvideo/health_check.py
+from dynamo.health_check import HealthCheckPayload
+from fastvideo import VideoGenerator
+
+
+class FastVideoHealthCheckPayload(HealthCheckPayload):
+    def __init__(self, generator: VideoGenerator) -> None:
+        # Post-PR 7.10: generator.default_health_check_request() returns a
+        # typed GenerationRequest; dump it into the same dict shape that
+        # FastVideo's adapter accepts.
+        req = generator.default_health_check_request()
+        self.default_payload = {
+            "prompt": req.prompt or "test",
+            "size": f"{req.sampling.width}x{req.sampling.height}",
+            "response_format": "b64_json",
+            "nvext": {
+                "fps": req.sampling.fps,
+                "num_frames": req.sampling.num_frames,
+                "num_inference_steps": req.sampling.num_inference_steps,
+                "guidance_scale": req.sampling.guidance_scale,
+            },
+        }
+        super().__init__()
+```
+
+Fallback (pre-PR 7.10) — hardcoded 256×256 / 8 frames / 1 step, matching
+[`VideoGenerationHealthCheckPayload`](https://github.com/ai-dynamo/dynamo/blob/main/components/src/dynamo/sglang/health_check.py#L198-L226).
+
+## Init function sketch
+
+```python
+# components/src/dynamo/fastvideo/init_video_generation.py
+async def init_video_generation(runtime, config, shutdown_endpoints):
+    from fastvideo import VideoGenerator
+    from fastvideo.api import config_to_dict
+
+    server_args, dynamo_args = config.server_args, config.dynamo_args
+    generator = VideoGenerator.from_pretrained(**config.fastvideo_kwargs())
+
+    dump_config(dynamo_args.dump_config_to, config)
+
+    endpoint = runtime.endpoint(
+        f"{dynamo_args.namespace}.{dynamo_args.component}.{dynamo_args.endpoint}"
+    )
+    shutdown_endpoints[:] = [endpoint]
+
+    handler = VideoGenerationWorkerHandler(
+        generator, config, fs=get_fs(dynamo_args.media_output_fs_url)
+    )
+    payload = FastVideoHealthCheckPayload(generator).to_dict()
+
+    await asyncio.gather(
+        endpoint.serve_endpoint(
+            handler.generate,
+            graceful_shutdown=True,
+            health_check_payload=payload,
+        ),
+        register_video_generation_model(
+            generator, endpoint, server_args,
+        ),
+    )
+```
+
+## Args adapter
+
+`FastVideoArgGroup` (Dynamo-side) converts CLI flags into a typed
+`GeneratorConfig` — **never** into legacy flat kwargs. Because PR 6
+added typed homes for every kwarg the internal `gpu_pool.py` used,
+this adapter can build the config purely from the public typed schema:
+
+```python
+def build_generator_config(args) -> "GeneratorConfig":
+    from fastvideo.api import (
+        CompileConfig, ComponentConfig, EngineConfig, GeneratorConfig,
+        OffloadConfig, ParallelismConfig, PipelineSelection,
+    )
+    return GeneratorConfig(
+        model_path=args.model_path,
+        engine=EngineConfig(
+            num_gpus=args.num_gpus,
+            parallelism=ParallelismConfig(tp_size=args.tp_size, sp_size=args.sp_size),
+            offload=OffloadConfig(dit=args.dit_offload, text_encoder=args.te_offload),
+            compile=CompileConfig(enabled=args.compile, mode=args.compile_mode),
+        ),
+        pipeline=PipelineSelection(
+            workload_type=args.workload or "t2v",
+            preset=args.preset,  # e.g. "ltx2_two_stage"
+            components=ComponentConfig(
+                upsampler_weights=args.refine_upsampler,
+                lora_path=args.refine_lora,
+            ),
+        ),
+    )
+```
+
+## Registration
+
+Dynamo's Rust side skips HuggingFace `config.json` downloads for
+`ModelType::Videos`, same fast path used by image diffusion. The
+Python-side registration:
+
+```python
+# components/src/dynamo/fastvideo/register.py
+from dynamo.llm import ModelDeploymentCard, ModelType, register_model
+
+
+async def register_video_generation_model(generator, endpoint, server_args):
+    mdc = ModelDeploymentCard.with_name_only(server_args.model_name or server_args.model_path)
+    await register_model(endpoint, mdc, ModelType.Videos, readiness_gate=asyncio.Event())
+```
+
+## Contract guarantees
+
+These guardrails let the Dynamo backend be written once and not
+re-chase FastVideo drift:
+
+1. `GenerationRequest` field paths are stable across PR 6 onward. Any
+   breaking rename triggers a major bump and appears in
+   [`inference_schema_parity_inventory.yaml`](../inference_schema_parity_inventory.yaml).
+2. `ContinuationState.payload` is JSON-serializable or references
+   opaque blob ids. Dynamo can round-trip it through RPC without
+   special-casing torch tensors.
+3. `VideoGenerator.from_pretrained` accepts a typed `GeneratorConfig`;
+   legacy flat kwargs are compatibility-only and deprecate in PR 13.
+4. `generate_async` (PR 7.10+) emits events in order
+   `Progress* → Partial* → Final`; the final event always has exactly
+   one occurrence per request.
+5. `default_health_check_request()` (PR 7.10+) returns a request that
+   passes `parse_config` and produces a non-zero-latency but bounded
+   workload (256×256 / 8 frames / 1 step).
+
+FastVideo's contract tests (`fastvideo/tests/contract/`) assert these
+with mocked Dynamo-style handlers that import only the public surface.
+If a change to FastVideo breaks the adapter pattern, those tests fail
+at FastVideo's CI — before the Dynamo-side integration even knows.
+
+## What the Dynamo adapter MUST NOT import
+
+* Anything under `fastvideo.pipelines.*` directly (pipelines are
+  internal; presets identify them by name on
+  `PipelineSelection.preset`).
+* `fastvideo.fastvideo_args.FastVideoArgs` (legacy compat type).
+* `fastvideo.api.compat.*` private helpers
+  (`_validate_continuation_state` etc.) — the public boundary is
+  `VideoGenerator` + `fastvideo.api`.
+* Any flat legacy LTX-2 kwarg (`ltx2_refine_upsampler_path`,
+  `torch_compile_kwargs`, etc.) — all have typed homes in
+  `GeneratorConfig`.
+
+## Future: disaggregated prefill/decode
+
+PR 7's continuation state was designed to survive RPC transport, so a
+future Dynamo split where prefill yields state and decode hydrates it
+is expressible without changing the contract. The streaming server
+(PR 7.6) already uses the `SessionStore` pattern; Dynamo's disagg
+could wire a distributed `SessionStore` backend by the same interface.
+
+## See also
+
+* [OpenAI HTTP contract](openai.md)
+* [Streaming WebSocket protocol](streaming.md)
+* Draft PR reference: [ai-dynamo/dynamo#7544](https://github.com/ai-dynamo/dynamo/pull/7544)
+* Dynamo SGLang backend (template this doc is modeled on):
+  [ai-dynamo/dynamo](https://github.com/ai-dynamo/dynamo/tree/main/components/src/dynamo/sglang)

docs/design/server_contracts/index.md

@@ -0,0 +1,41 @@
+# Server Contracts
+
+FastVideo's typed public API (`fastvideo.api`) is consumed by three
+server-class integrations that must share one execution substrate so we
+don't grow three near-duplicate progress loops:
+
+| Consumer | Transport | Request shape | State model |
+| --- | --- | --- | --- |
+| [Stateless OpenAI](openai.md) | HTTP POST `/v1/videos` | `VideoGenerationsRequest` → `GenerationRequest` merged onto `ServeConfig.default_request` | Stateless; optional `ContinuationState` round-trip |
+| [Streaming WebSocket](streaming.md) | WebSocket JSON + binary fMP4 | `GenerationRequest` per segment | Server-held `SessionStore`, snapshot-on-demand |
+| [Dynamo native backend](dynamo.md) | Dynamo RPC | `NvCreateVideoRequest` → adapter → `GenerationRequest` | Aggregated today; disaggregated via `ContinuationState` later |
+
+All three consume the same underlying surface:
+
+```python
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    ContinuationState,
+    GenerationRequest,
+    InputConfig,
+    OutputConfig,
+    SamplingConfig,
+    ServeConfig,
+)
+
+# Sync today, async after PR 7.10 lands VideoGenerator.generate_async.
+result = generator.generate(request)
+```
+
+These docs lock down the request/response shapes so drift between
+FastVideo, the internal UI, and Dynamo can be caught at review time.
+PR 8 does not ship runtime code; it ships the contract reference and
+the contract tests that guard it.
+
+## Related
+
+- [API refactor design](../overview.md)
+- Parity inventory: [`inference_schema_parity_inventory.yaml`](../inference_schema_parity_inventory.yaml)
+- [Streaming server upstream plan](../../../.agents/exploration/streaming-server-upstream-plan.md)
+- Draft PR (closed) that establishes the Dynamo shape:
+  https://github.com/ai-dynamo/dynamo/pull/7544

docs/design/server_contracts/openai.md

@@ -0,0 +1,116 @@
+# OpenAI-compatible HTTP Contract
+
+The stateless FastVideo HTTP server lives at
+[`fastvideo/entrypoints/openai/`](https://github.com/hao-ai-lab/FastVideo/tree/main/fastvideo/entrypoints/openai).
+Launch: `fastvideo serve --config serve.yaml`.
+
+## Endpoints
+
+| Method | Path | Description |
+| --- | --- | --- |
+| `POST` | `/v1/videos/generations` | Synchronous video generation |
+| `GET` | `/v1/videos` | List prior jobs held in the in-memory store |
+| `GET` | `/v1/videos/{id}` | Job status / result |
+| `GET` | `/v1/videos/{id}/content` | Download the MP4 once ready |
+| `POST` | `/v1/images/generations` | Synchronous image generation |
+| `GET` | `/v1/models` | Enumerate registered models |
+| `GET` | `/health` | Liveness probe |
+
+## `VideoGenerationsRequest` shape
+
+Mirrors the OpenAI `POST /v1/videos/generations` shape:
+
+```json
+{
+  "prompt": "a fox running through snow",
+  "size": "1024x1536",
+  "seconds": 5,
+  "fps": 24,
+  "num_frames": 121,
+  "seed": 42,
+  "num_inference_steps": 8,
+  "guidance_scale": 1.0,
+  "negative_prompt": "blurry, low quality",
+  "input_reference": "/path/to/init.png"
+}
+```
+
+SGLang-compatible extensions carried today:
+`num_inference_steps`, `guidance_scale`, `guidance_scale_2`,
+`true_cfg_scale`, `negative_prompt`, `enable_teacache`, `output_path`.
+
+## Merge precedence
+
+The server builds a `GenerationRequest` each call using three layers,
+highest first:
+
+1. **Request body (client-explicit)** — only fields carried in
+   `request.model_fields_set` (Pydantic v2). Unset fields do not count,
+   even if the Pydantic model has a schema default for them.
+2. **`ServeConfig.default_request` (operator-explicit)** — projected via
+   [`explicit_request_updates()`](../../../fastvideo/api/compat.py);
+   only fields the operator actually wrote into the YAML count as
+   defaults. Every other field inherits the schema default rather than
+   being pinned.
+3. **Hardcoded fallback** — e.g. `fps = 24`.
+
+The gate matters: both surfaces carry schema defaults. Without
+`model_fields_set` / explicit-path tracking, schema defaults would
+masquerade as intent and silently shadow the other side.
+
+See [`video_api.py::_build_generation_kwargs`](../../../fastvideo/entrypoints/openai/video_api.py)
+for the canonical implementation; the per-request assembly lives there,
+not in pipeline code.
+
+## Continuation state
+
+The stateless surface accepts an opaque `ContinuationState` round-trip.
+Clients that want continuation pass the prior `state` blob back on the
+next request, and receive a new one on the response when
+`request.output.return_state = true`.
+
+Shape:
+
+```json
+{
+  "state": {
+    "kind": "ltx2.v1",
+    "payload": { "schema_version": 1, "segment_index": 3, ... }
+  }
+}
+```
+
+Payload is always JSON-serializable. Large tensors may live in an
+opaque blob-store reference the client simply round-trips; see
+[`LTX2ContinuationState`](../../../fastvideo/pipelines/basic/ltx2/continuation.py).
+
+Continuation is not yet wired all the way through to
+`generator.generate_video(...)` — PR 7.6 (GPU pool upstream) is the
+pipeline-level consumer. PR 7 locked the envelope so this surface is
+stable ahead of that plumbing.
+
+## Error codes
+
+| HTTP | Condition |
+| --- | --- |
+| `400 Bad Request` | Parse/validation failure (unknown field, type mismatch, incompatible preset/state) |
+| `404 Not Found` | `GET /v1/videos/{id}` for an unknown job |
+| `409 Conflict` | Job id already exists |
+| `500 Internal Server Error` | Pipeline raised; body mirrors upstream OpenAI error envelope |
+| `503 Service Unavailable` | No generator loaded, or shutdown in progress |
+
+Errors include a JSON body with
+`{"error": {"type": "...", "message": "..."}}` matching the OpenAI
+Python SDK's expectation.
+
+## What does not cross this boundary
+
+* Flat legacy kwargs (`ltx2_refine_enabled`, `torch_compile_kwargs`,
+  etc.) — these are init-time, configured via `ServeConfig.generator`,
+  never per-request.
+* Private Dreamverse-only fields — those live in a private adapter on
+  the Dreamverse side; the public FastVideo surface never promises
+  backward compatibility for them.
+* Raw tensor payloads (`ltx2_audio_clean_latent` et al.) — these are
+  derived by the pipeline from `ContinuationState`, never shipped as
+  request fields.

docs/design/server_contracts/streaming.md

@@ -0,0 +1,177 @@
+# Streaming WebSocket Server Contract
+
+The streaming server (`fastvideo/entrypoints/streaming/server.py`) speaks
+a JSON-over-WebSocket protocol with binary fMP4 chunks for media. This
+document is the authoritative spec for the message catalogue and the
+session state machine. Any change to either must update this document
+in the same PR that touches `protocol.py` or `session.py`.
+
+## Endpoint
+
+| Path | Protocol | Purpose |
+|---|---|---|
+| `WS /v1/stream` | WebSocket (JSON + binary) | Per-session realtime streaming |
+| `GET /health` | HTTP | Liveness probe (`status`, `stream_mode`, active `sessions`) |
+
+The server is launched by `fastvideo serve --config <serve.yaml>` when
+the config carries a `streaming:` block. Without that block the same CLI
+launches the OpenAI stateless HTTP server instead.
+
+## Connection lifecycle
+
+Every WebSocket connection holds exactly one `Session`. Sessions move
+through the states in `SessionState` (`fastvideo/entrypoints/streaming/session.py`).
+
+```
+                    ┌──────────────┐
+                    │ INITIALIZING │  ← WebSocket accepted, before init frame
+                    └──────┬───────┘
+                           │ session_init_v2 received
+            ┌──────────────┼──────────────┐
+            ▼              ▼              ▼
+         QUEUED      GPU_BINDING       REJECTED
+            │              │              ↑
+            │ slot ready   │              │ max-sessions hit
+            ▼              ▼              │ or invalid init
+                       ┌────────┐         │
+                       │ ACTIVE │ ────────┘
+                       └────┬───┘
+              segment loop  │
+                            │
+                ┌───────────┼───────────┐
+                ▼           ▼           ▼
+            COMPLETE      ERROR      TIMEOUT
+        (clean leave)  (any failure)  (idle / segment_cap reached)
+```
+
+Terminal states (`COMPLETE`, `ERROR`, `TIMEOUT`, `REJECTED`) are sinks —
+no transitions out. The transition matrix is enforced in
+`session.py::_VALID_TRANSITIONS`; bad transitions raise.
+
+`SessionManager` enforces the per-process budgets pulled from
+`StreamingConfig`:
+
+- `session_timeout_seconds` — idle reaper drops sessions that haven't
+  advanced; non-terminal sessions transition to `TIMEOUT`.
+- `generation_segment_cap` — a session that hits the cap transitions to
+  `COMPLETE` after the last segment ships.
+
+## Message catalogue
+
+Every JSON frame carries `{"type": <str>, ...}`. Pydantic models in
+`protocol.py` are the source of truth; this table is the human-readable
+view.
+
+### Client → server
+
+| `type` | Required fields | Purpose |
+|---|---|---|
+| `session_init_v2` | — | Opening frame. Carries preset, curated prompts, optional initial image, feature toggles, optional `continuation_state` to resume from a snapshot. |
+| `segment_prompt_source` | `prompt` | Request the next segment using the supplied prompt; optional sampling overrides (`seed`, `num_inference_steps`, `guidance_scale`, `negative_prompt`). |
+| `seed_prompts_updated` | `seed_prompts` | Replace the session's seed-prompt list; takes effect on the next segment. |
+| `enhancement_updated` | `enabled` | Toggle prompt enhancement for subsequent segments. |
+| `auto_extension_updated` | `enabled` | Toggle automatic per-segment prompt extension. |
+| `loop_generation_updated` | `enabled` | Toggle loop-generation mode. |
+| `generation_paused_updated` | `paused` | Pause/resume segment generation; queued requests defer. |
+| `snapshot_state` | — | Request the current `ContinuationState` for export; server replies with `continuation_state_snapshot`. |
+
+The opening frame must be `session_init_v2`. Any other first frame is
+rejected with an `error` (code `invalid_message`) and the WebSocket is
+closed.
+
+### Server → client
+
+| `type` | Carries | When emitted |
+|---|---|---|
+| `queue_status` | `position`, `queue_depth` | After `session_init_v2` accepted, before GPU binding. |
+| `gpu_assigned` | GPU id, model id | Once a generator slot is bound. |
+| `ltx2_stream_start` | session-level metadata | Once the session enters `ACTIVE`. |
+| `ltx2_segment_start` | `segment_idx`, `prompt`, prompt source | When a `segment_prompt_source` request begins generation. |
+| `step_complete` | `segment_idx`, denoise timings | After the segment's denoising loop finishes (before media emission). |
+| `media_init` | `segment_idx`, mime, stream id | First frame of fMP4 output for the segment. |
+| binary frame | fMP4 fragment bytes | Subsequent media chunks; the protocol enforces that `media_init` precedes any binary frames. |
+| `media_segment_complete` | `segment_idx`, chunk count, byte count | Last media chunk for the segment. |
+| `ltx2_segment_complete` | `segment_idx`, segment summary | Segment fully shipped; ready for the next `segment_prompt_source`. |
+| `ltx2_stream_complete` | session summary | Session reached `generation_segment_cap` or client requested clean shutdown. |
+| `session_timeout` | reason | Session hit `session_timeout_seconds`; immediately followed by close. |
+| `continuation_state_snapshot` | `kind`, `payload` | Reply to `snapshot_state`. The payload is the same shape produced by `LTX2ContinuationState.to_continuation_state(...)`. |
+| `error` | `code`, `message` | Any validation/runtime error. Non-fatal errors keep the connection open; fatal errors precede a `close`. |
+
+## Continuation state
+
+The session optionally accepts a `continuation_state` dict inside the
+opening `session_init_v2` frame. When present, the server hydrates it
+into a `ContinuationState(kind, payload)` envelope and feeds it as the
+`request.state` on the first segment's `GenerationRequest` — letting a
+client resume after a disconnect, migrate sessions across processes,
+or replay a prior session.
+
+After every segment, if the runtime returns a fresh state, the server
+persists it to the `SessionStore` so a `snapshot_state` request can
+export it. The store and serialization contracts live with the model
+family (e.g. `fastvideo/pipelines/basic/ltx2/continuation.py` for LTX-2).
+
+## Example flow
+
+```
+client                                                    server
+──────                                                    ──────
+WS /v1/stream  ─────── connect ─────────────────────────►
+               ◄────── (accept)
+
+{"type": "session_init_v2",
+ "preset": "ltx2_two_stage",
+ "curated_prompts": ["a fox in snow", "the fox jumps"],
+ "initial_image": {...},
+ "stream_mode": "av_fmp4"} ─────────────────────────────►
+
+                                                          (validate, queue, bind)
+               ◄──── {"type": "queue_status",
+                       "position": 0, "queue_depth": 0}
+               ◄──── {"type": "gpu_assigned",
+                       "gpu_id": 0, "model_id": "..."}
+               ◄──── {"type": "ltx2_stream_start", ...}
+
+{"type": "segment_prompt_source",
+ "prompt": "a fox in snow",
+ "source": "curated"} ───────────────────────────────────►
+                                                          (run pipeline)
+               ◄──── {"type": "ltx2_segment_start",
+                       "segment_idx": 1, ...}
+               ◄──── {"type": "step_complete",
+                       "segment_idx": 1, "timings": {...}}
+               ◄──── {"type": "media_init",
+                       "segment_idx": 1,
+                       "mime": "video/mp4", ...}
+               ◄──── <binary fMP4 init segment>
+               ◄──── <binary fMP4 fragment>
+               ◄──── <binary fMP4 fragment>
+               ◄──── {"type": "media_segment_complete",
+                       "segment_idx": 1, "chunks": 12}
+               ◄──── {"type": "ltx2_segment_complete",
+                       "segment_idx": 1, ...}
+
+{"type": "segment_prompt_source",
+ "prompt": "the fox jumps"} ─────────────────────────────►
+                                                          (segment 2 …)
+
+{"type": "snapshot_state"} ──────────────────────────────►
+               ◄──── {"type": "continuation_state_snapshot",
+                       "kind": "ltx2.v1",
+                       "payload": {"schema_version": 1, ...}}
+
+(close) ──────────────────────────────────────────────────►
+                                                          (session → COMPLETE)
+```
+
+## Backward / forward compatibility
+
+- Adding a new client message: append a Pydantic model to `protocol.py`
+  with a unique `type`; add the discriminator entry to `ClientMessage`;
+  add a row to the table above. Old clients that don't send the new
+  message remain compatible.
+- Adding a new server message: emit only when a new feature flag is
+  enabled (or always emit, since clients ignore unknown types).
+- Changing an existing message: bump the `type` (e.g. `session_init_v2`
+  → `session_init_v3`) and accept both for one release cycle. Never
+  silently change field semantics under the same `type`.

docs/design/training_architecture.md

@@ -243,7 +243,7 @@ for step in range(start_step, max_steps):
 
 ```bash
 # Install
-uv pip install -e .[dev]
+uv pip install -e ".[dev]"
 
 # Run DMD2 distillation on Wan 2.1
 torchrun --nproc_per_node=8 -m fastvideo.train.entrypoint.train \

docs/distillation/dmd.md

@@ -86,3 +86,25 @@ sbatch examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B.sh
 - Learning rate: 2e-5
 - Training steps: 3000 (~12 hours)
 - HSDP shard dim: 1
+
+## 🧭 Note on `real_score_guidance_scale`
+
+The teacher CFG used inside the DMD loss follows the DMD2 reference
+implementation and uses the parameterization
+
+```
+x = x_cond + w * (x_cond - x_uncond)
+```
+
+rather than the Ho & Salimans form `x_uncond + w * (x_cond - x_uncond)`. The
+two are mathematically equivalent up to a constant offset:
+
+| `real_score_guidance_scale` (`w`) | Equivalent standard CFG (`w + 1`) | Output                |
+|-----------------------------------|-----------------------------------|-----------------------|
+| `-1`                              | `0`                               | unconditional         |
+| `0`                               | `1`                               | conditional           |
+| `3.5` (default)                   | `4.5`                             | strong guidance       |
+
+So `real_score_guidance_scale` should be read as the **extra** guidance
+strength added on top of the conditional prediction. When porting values
+from a paper that uses the Ho & Salimans form, subtract 1.

docs/getting_started/installation.md

@@ -27,7 +27,7 @@ uv pip install fastvideo
 conda create -n fastvideo python=3.12 -y
 conda activate fastvideo
 
-pip install fastvideo
+uv pip install fastvideo
 ```
 
 ### From source
@@ -41,11 +41,11 @@ uv pip install -e .
 uv pip install flash-attn --no-build-isolation -v
 ```
 
-Alternative with Conda environment:
+Alternative with Conda environment (still drives installs through `uv`):
 
 ```bash
-pip install -e .
-pip install flash-attn --no-build-isolation -v
+uv pip install -e .
+uv pip install flash-attn --no-build-isolation -v
 ```
 
 ## Hardware Requirements

docs/getting_started/installation/gpu.md

@@ -58,14 +58,16 @@ uv pip install flash-attn --no-build-isolation -v
 
 #### With Conda environment (alternative)
 
+`uv` works inside an active conda env too, so prefer `uv pip` for the actual install:
+
 ```bash
-pip install fastvideo
+uv pip install fastvideo
 ```
 
 Also optionally install FlashAttention:
 
 ```bash
-pip install flash-attn --no-build-isolation -v
+uv pip install flash-attn --no-build-isolation -v
 ```
 
 ### Installation from Source
@@ -87,7 +89,7 @@ uv pip install -e .
 Alternative with Conda environment:
 
 ```bash
-pip install -e .
+uv pip install -e .
 ```
 
 ### Optional Dependencies
@@ -101,7 +103,7 @@ uv pip install flash-attn --no-build-isolation -v
 Alternative with Conda environment:
 
 ```bash
-pip install flash-attn --no-build-isolation -v
+uv pip install flash-attn --no-build-isolation -v
 ```
 
 ## Set up using Docker

docs/getting_started/installation/mps.md

@@ -57,8 +57,10 @@ uv pip install fastvideo
 
 #### With Conda environment (alternative)
 
+`uv` works inside an active conda env too, so prefer `uv pip` for the actual install:
+
 ```bash
-pip install fastvideo
+uv pip install fastvideo
 ```
 
 ### Installation from Source
@@ -80,7 +82,7 @@ uv pip install -e .
 Alternative with Conda environment:
 
 ```bash
-pip install -e .
+uv pip install -e .
 ```
 
 ## Development Environment Setup

docs/inference/gen3c.md

@@ -19,7 +19,7 @@
 - Install MoGe:
 
 ```bash
-pip install git+https://github.com/microsoft/MoGe.git
+uv pip install git+https://github.com/microsoft/MoGe.git
 ```
 
 - If you hit `ImportError: libGL.so.1` (common on Ubuntu/headless nodes), you can try installing OpenCV runtime libs:

docs/inference/optimizations.md

@@ -54,7 +54,7 @@ FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 We recommend always installing [Flash Attention 2](https://github.com/Dao-AILab/flash-attention):
 
 ```bash
-pip install flash-attn==2.7.4.post1 --no-build-isolation
+uv pip install flash-attn==2.7.4.post1 --no-build-isolation
 ```
 
 And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#flashattention-3-beta-release) by compiling it from source (takes about 10 minutes for me):
@@ -63,7 +63,7 @@ And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://git
 git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention
 
 cd hopper
-pip install ninja
+uv pip install ninja
 python setup.py install
 ```
 
@@ -98,7 +98,7 @@ To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please co
 ```bash
 git clone https://github.com/thu-ml/SageAttention.git
 cd sageattention
-python setup.py install  # or pip install -e .
+python setup.py install  # or uv pip install -e .
 ```
 
 ### Sage Attention 3

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/README.md

@@ -4,7 +4,7 @@ These are end-to-end example scripts for distilling Wan2.1 T2V 1.3B model using
 ### 0. Make sure you have installed VSA
 
 ```bash
-pip install vsa
+uv pip install vsa
 ```
 
 ### 1. Download dataset:

examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/README.md

@@ -4,7 +4,7 @@ These are end-to-end example scripts for distilling Wan2.2 TI2V 5B model DMD+VSA
 ### 0. Make sure you have installed VSA
 
 ```bash
-pip install vsa
+uv pip install vsa
 ```
 
 ### Data-free Distillation

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/README.md

@@ -4,7 +4,7 @@ These are end-to-end example scripts for distilling Wan2.2 TI2V 5B model DMD+VSA
 ### 0. Make sure you have installed VSA
 
 ```bash
-pip install vsa
+uv pip install vsa
 ```
 
 ### 1. Download dataset:

examples/inference/basic/basic_gen3c.py

@@ -7,7 +7,7 @@ and the GEN3C diffusion model.
 
 Requirements:
   1. Install MoGe:
-     pip install git+https://github.com/microsoft/MoGe.git
+     uv pip install git+https://github.com/microsoft/MoGe.git
      If you hit `ImportError: libGL.so.1`, install:
      sudo apt-get update && sudo apt-get install -y libgl1 libglib2.0-0 libsm6 libxext6 libxrender1
   2. Download and convert weights:

examples/inference/basic/basic_stable_audio.py

@@ -0,0 +1,77 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Stable Audio Open 1.0 — text-to-audio (baseline) example.
+
+User story (game-audio designer, prototyping):
+    "I'm prototyping a level and I need 6 seconds of background
+    ambience — gentle wind, distant thunder, a hint of birdsong. I
+    don't want to dig through a sound library; I want to type what I
+    hear in my head and get a wav back. If it's wrong I'll iterate
+    on the prompt. This is the first stop."
+
+User story (musician sketching ideas):
+    "I want to bounce a 30s lo-fi drum loop to use as a placeholder
+    bed while I build the rest of the track. Type prompt, get audio,
+    drop into the DAW. The actual production beat I'll record
+    myself, but I need *something* to write the chords against."
+
+User story (researcher exploring the model):
+    "First time touching Stable Audio Open — what does it sound
+    like at default settings? This is the smallest amount of code
+    that goes from prompt to mp4."
+
+How it works:
+    Pure text-to-audio (T2A). The pipeline runs:
+        T5 + NumberConditioner -> StableAudioDiT -> Oobleck VAE
+    via the `dpmpp-3m-sde` k-diffusion sampler. All components are
+    FastVideo-native — no diffusers / transformers model imports at
+    runtime (see REVIEW item 30). Mirrors upstream
+    `stable_audio_tools.inference.generation.generate_diffusion_cond`
+    bit-for-bit (~0.2% abs_mean drift on 25 steps).
+
+Tunable knobs (the "creative dials"):
+    audio_end_in_s
+        1–6   — quick ideation (sub-10s wall clock at 100 steps)
+        10–30 — full musical phrase / loop length (the README example
+                uses 30s)
+        47.5  — model maximum (full sample_size = 2097152 / 44100 Hz)
+    num_inference_steps
+        25  — fast preview, occasional artifacts
+        100 — preset default (matches the HF model card)
+        250 — diminishing returns past here
+    guidance_scale
+        3   — looser, more variation per seed
+        7   — preset default; matches README
+        12+ — sharper but can sound "fried"
+
+Prerequisites:
+  1. Accept the terms on https://huggingface.co/stabilityai/stable-audio-open-1.0
+     and export your HF token in the shell:
+         export HF_TOKEN=hf_...
+  2. Install optional inference deps (one-time):
+         uv pip install k_diffusion einops_exts alias_free_torch torchsde
+"""
+from fastvideo import VideoGenerator
+
+PROMPT = "Lo-fi hip hop instrumental with vinyl crackle and gentle piano."
+
+
+def main() -> None:
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/stable-audio-open-1.0-Diffusers",
+        num_gpus=1,
+    )
+    output_path = "outputs_audio/stable_audio_basic/output_stable_audio.wav"
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path=output_path,
+        save_video=True,
+        # 6-second clip; the model max is ~47.5s.
+        audio_end_in_s=6.0,
+        # The registered preset gives 100 steps + CFG=7.0 by default;
+        # override num_inference_steps / guidance_scale here for QA.
+    )
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_stable_audio_a2a.py

@@ -0,0 +1,77 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Stable Audio Open 1.0 — audio-to-audio variation example.
+
+User story (musician, late at night):
+    "I generated this 12-second lo-fi loop earlier and I love the chord
+    progression and overall vibe, but the snare hit at 0:08 sounds wrong
+    and the rhythm feels stiff. I don't want to start over from scratch
+    and lose what's working — I want the model to keep the harmony and
+    mood but reroll the percussion + groove."
+
+User story (sound designer, on a deadline):
+    "I have one good 'sword clang' SFX. The art director wants 8 sibling
+    variations that all feel like the same sword from different angles —
+    same metal, same weight, slightly different impact. I'd rather
+    refine my one good take than text-prompt my way through 50 misses."
+
+Pass `init_audio=path/to/clip` (any wav/mp3/mp4/m4a/flac the standard
+deps decode) and the model will use it as a starting point for the
+text prompt instead of pure noise.
+
+Picking `init_audio_strength` (0.0 to 1.0):
+
+    Higher = closer to the source clip. Lower = more transformation.
+    (Same convention as the "Input Audio Strength" slider in
+    Stability's commercial Stable Audio web UI, so values transfer
+    directly.)
+
+      | strength | what you get                                       |
+      |----------|----------------------------------------------------|
+      |   1.00   | Output ≈ reference. No transformation.             |
+      |   0.85   | Texture micro-variation only.                      |
+      |   0.70   | Light reroll, same instruments.                    |
+      |   0.60   | Default. Instrument identity is replaceable        |
+      |          | (cello can take over from piano on the same notes).|
+      |   0.50   | Heavy — only melody / chord progression survives.  |
+      |   0.30   | Reference acts as a loose mood prompt.             |
+      |   0.00   | Plain T2A — reference ignored.                     |
+
+    Rule of thumb by intent:
+      * "Fix one part of this clip"            -> 0.75 .. 0.85
+      * "Same notes, different instrument"     -> 0.55 .. 0.65
+      * "Same chord progression, new content"  -> 0.40 .. 0.55
+      * "Use this as a loose mood prompt"      -> 0.20 .. 0.35
+
+    If the reference timbre is bleeding through more than you want,
+    lower it; if the structure is gone, raise it.
+
+Prerequisites: same as `basic_stable_audio.py`.
+"""
+from fastvideo import VideoGenerator
+
+PROMPT = "Change the piano to a cello playing the same notes"
+# Path to any audio-bearing file (wav, mp3, mp4, m4a, flac, ...).
+# Set to `None` to skip A2A and run plain T2A.
+INIT_AUDIO_PATH: str | None = None
+# Reference fidelity in [0, 1] -- higher = closer to source.
+INIT_AUDIO_STRENGTH = 0.6
+
+
+def main() -> None:
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/stable-audio-open-1.0-Diffusers",
+        num_gpus=1,
+    )
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path="outputs_audio/stable_audio_a2a/output_a2a.wav",
+        save_video=True,
+        audio_end_in_s=6.0,
+        init_audio=INIT_AUDIO_PATH,
+        init_audio_strength=INIT_AUDIO_STRENGTH,
+    )
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_stable_audio_inpaint.py

@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Stable Audio Open 1.0 — inpainting / outpainting (loop extension) example.
+
+User story (loop extension — the killer app):
+    "I have a 6-second drum loop my client likes. They want it as
+    background bed for a 30-second ad. I need it to loop seamlessly,
+    but a hard cut every 6s sounds bad. Let me extend it to 30s,
+    keeping the first 6s exactly as-is and letting the model continue
+    the groove for the remaining 24s."
+
+User story (audio repair):
+    "There's a microphone bump at 0:14 in this 30-second field
+    recording — really obvious in headphones. Mask out 0:13 to 0:15
+    and let the model regenerate plausible ambience that blends in.
+    Everything else stays exactly as I recorded it."
+
+User story (transition smoothing):
+    "I have two 10-second clips I want to crossfade. Mask out a 1s
+    overlap region in the middle and let the model invent a coherent
+    transition between the two."
+
+How it works (RePaint-style blending):
+    Stable Audio Open 1.0 wasn't trained as an inpainting model
+    (`model_type=diffusion_cond`, not `diffusion_cond_inpaint`), so we
+    can't use the upstream's mask-conditioned approach directly. We
+    use the RePaint trick instead, which works on any v-prediction
+    diffusion model:
+
+      1. Encode the reference clip into latent space.
+      2. At every denoising step `i`, replace the kept region of the
+         in-flight latent (where mask == 1) with the reference
+         re-noised to the next timestep's sigma. Only the unkept
+         region (mask == 0) is freely denoised.
+      3. After the loop, the kept region is exactly the reference;
+         the unkept region is freshly generated content.
+
+    This is approximate compared to a properly trained inpainting
+    checkpoint — the seam between kept/unkept can have slight EQ
+    discontinuity — but it works on the existing public model.
+
+Tunable: the mask is a 1-D tensor in {0, 1} at the model's sample
+rate. Conventions:
+    1.0 = keep this sample from the reference
+    0.0 = regenerate this sample
+
+Prerequisites: same as `basic_stable_audio.py`.
+"""
+import os
+
+from fastvideo import VideoGenerator
+
+PROMPT = "Steady lo-fi hip hop drum loop with vinyl crackle."
+# Required: path to the reference audio file (wav, mp3, mp4, m4a, flac,
+# ...) you want to extend or repair. The pipeline raises if a mask is
+# passed without a reference, so this must be a real path.
+REFERENCE_AUDIO_PATH = "path/to/your/loop.wav"
+KEEP_SECONDS = 6.0       # first KEEP_SECONDS preserved exactly
+TOTAL_SECONDS = 12.0     # extend the loop to this duration
+
+
+def main() -> None:
+    if not os.path.isfile(REFERENCE_AUDIO_PATH):
+        raise FileNotFoundError(
+            f"REFERENCE_AUDIO_PATH={REFERENCE_AUDIO_PATH!r} does not exist. "
+            "Edit this script to point at a real audio file (wav/mp3/mp4/"
+            "m4a/flac) before running.")
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/stable-audio-open-1.0-Diffusers",
+        num_gpus=1,
+    )
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path="outputs_audio/stable_audio_inpaint/output_inpaint.wav",
+        save_video=True,
+        audio_end_in_s=TOTAL_SECONDS,
+        inpaint_audio=REFERENCE_AUDIO_PATH,
+        # Tuple form: keep first KEEP_SECONDS, regenerate the rest.
+        inpaint_mask=(KEEP_SECONDS, TOTAL_SECONDS),
+    )
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_stable_audio_small.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Stable Audio Open Small — fast / lightweight T2A example.
+
+User story (interactive UI builder):
+    "I'm building a sound-design UI where the user types a prompt and
+    we want sub-2-second feedback so the experience feels like
+    autocomplete, not a render queue. The full Stable Audio Open 1.0
+    takes ~8s on a single GPU; the small variant takes a fraction of
+    that — quality is lower but completely usable for real-time
+    iteration."
+
+User story (overnight batch jobs):
+    "I'm generating 10,000 short SFX variants for a procedural game.
+    Wall-clock matters more than per-clip polish — give me the small
+    model so I can fit the run in one night instead of a week."
+
+How it works:
+    The small variant is a separate Stability AI checkpoint
+    (`stabilityai/stable-audio-open-small`) that ships the same Oobleck
+    VAE as the 1.0 base model but a smaller / faster DiT (`embed_dim=1024`,
+    `depth=16`, `qk_norm="ln"`) and only one duration conditioner
+    (`seconds_total`, no `seconds_start`). FastVideo loads from the
+    converted Diffusers-format repo `FastVideo/stable-audio-open-small-Diffusers`
+    via the standard component loader; per-variant arch fields come
+    from `transformer/config.json` and `conditioner/config.json`.
+
+Prerequisites: same as `basic_stable_audio.py`. The converted repo is
+public so no gated-access flow is required.
+"""
+from fastvideo import VideoGenerator
+
+PROMPT = "Lo-fi hip hop instrumental with vinyl crackle and gentle piano."
+
+
+def main() -> None:
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/stable-audio-open-small-Diffusers",
+        num_gpus=1,
+    )
+    output_path = "outputs_audio/stable_audio_small/output_stable_audio_small.wav"
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path=output_path,
+        save_video=True,
+        # Small variant trains on a ~11.9s window — keep `audio_end_in_s`
+        # at or below that.
+        audio_end_in_s=6.0,
+    )
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/train/configs/fine_tuning/cosmos/t2v.yaml

@@ -0,0 +1,73 @@
+# Cosmos Predict2 2B T2V finetune config.
+#
+# Data must be preprocessed with Cosmos VAE + T5 text encoder
+# into parquet format before training.
+
+models:
+  student:
+    _target_: fastvideo.train.models.cosmos.CosmosModel
+    init_from: nvidia/Cosmos-Predict2-2B-Video2World
+    trainable: true
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 8
+    hsdp_shard_dim: 1
+
+  data:
+    data_path: data/cosmos_preprocessed
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    # Cosmos VAE: 4x temporal, 8x spatial compression.
+    # 93 frames -> 24 latent frames, 480x832 -> 60x104
+    num_latent_t: 24
+    num_height: 480
+    num_width: 832
+    num_frames: 93
+
+  optimizer:
+    learning_rate: 1.0e-5
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 5000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/cosmos_finetune
+    training_state_checkpointing_steps: 500
+    checkpoints_total_limit: 3
+    resume_from_checkpoint: latest
+
+  tracker:
+    project_name: fastvideo_cosmos
+    run_name: cosmos_finetune
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+  validation:
+    _target_: fastvideo.train.callbacks.validation.ValidationCallback
+    pipeline_target: fastvideo.pipelines.basic.cosmos.cosmos_pipeline.Cosmos2VideoToWorldPipeline
+    dataset_file: data/cosmos_preprocessed/validation_prompts.json
+    every_steps: 100
+    sampling_steps: [50]
+    guidance_scale: 6.0
+
+pipeline:
+  flow_shift: 1.0

examples/train/configs/overfit_cosmos25_t2w.yaml

@@ -0,0 +1,79 @@
+# Cosmos-Predict2.5-2B Text-to-World overfitting test config.
+#
+# Overfits on a few short videos (480x832, 93 frames) to verify the
+# Cosmos 2.5 training plugin works end-to-end.
+#
+# Preprocess data first:
+#   CUDA_VISIBLE_DEVICES=0 python fastvideo/pipelines/preprocess/preprocess_cosmos25_overfit.py
+#
+# Run:
+#   bash examples/train/run.sh examples/train/configs/overfit_cosmos25_t2w.yaml
+
+models:
+  student:
+    _target_: fastvideo.train.models.cosmos.CosmosModel
+    init_from: KyleShao/Cosmos-Predict2.5-2B-Diffusers
+    trainable: true
+    enable_gradient_checkpointing_type: full
+    flow_shift: 1.0
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
+
+training:
+  distributed:
+    num_gpus: 1
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 1
+
+  data:
+    data_path: data/cosmos25_overfit_preprocessed
+    dataloader_num_workers: 0
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 42
+    num_latent_t: 24
+    num_height: 480
+    num_width: 832
+    num_frames: 93
+
+  optimizer:
+    learning_rate: 5.0e-5
+    betas: [0.9, 0.999]
+    weight_decay: 0.0
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 300
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/cosmos25_overfit
+    training_state_checkpointing_steps: 50
+    checkpoints_total_limit: 2
+
+  tracker:
+    project_name: fastvideo_cosmos25
+    run_name: cosmos25_overfit
+
+  model:
+    precondition_outputs: false
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+  validation:
+    _target_: fastvideo.train.callbacks.validation.ValidationCallback
+    pipeline_target: fastvideo.pipelines.basic.cosmos.cosmos2_5_pipeline.Cosmos2_5Pipeline
+    dataset_file: data/cosmos25_overfit_preprocessed/validation_prompts.json
+    every_steps: 150
+    sampling_steps: [35]
+    guidance_scale: 7.0
+
+pipeline:
+  flow_shift: 1.0

fastvideo-kernel/CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.26 FATAL_ERROR)
 project(fastvideo-kernel LANGUAGES CXX)
 
-# Prefer environment variable (used by CI or pip install git+repo_addr) if CMake var is not explicitly set.
+# Prefer environment variable (used by CI or uv pip install git+repo_addr) if CMake var is not explicitly set.
 if(NOT DEFINED GPU_BACKEND AND DEFINED ENV{GPU_BACKEND})
     set(GPU_BACKEND "$ENV{GPU_BACKEND}")
 endif()

fastvideo-kernel/build.sh

@@ -61,7 +61,16 @@ has_cmake_arg() {
 }
 
 detect_with_torch() {
-    uv run --active --no-project python -c "import torch
+    # Prefer the active venv's python directly over `uv run --active --no-project`,
+    # which on some uv versions provisions its own interpreter and misses packages
+    # installed into VIRTUAL_ENV.
+    local py
+    if [[ -n "${VIRTUAL_ENV:-}" && -x "${VIRTUAL_ENV}/bin/python" ]]; then
+        py="${VIRTUAL_ENV}/bin/python"
+    else
+        py="$(command -v python3 || command -v python)"
+    fi
+    "${py}" -c "import torch
 if not torch.cuda.is_available():
     raise RuntimeError('torch.cuda.is_available() is false')
 mj, mn = torch.cuda.get_device_capability(0)

fastvideo-kernel/python/fastvideo_kernel/__init__.py

@@ -5,6 +5,11 @@ from fastvideo_kernel.ops import (
     video_sparse_attn,
 )
 
+from fastvideo_kernel.block_sparse_attn import (
+    block_sparse_attn,
+    block_sparse_attn_from_indices,
+)
+
 from fastvideo_kernel.vmoba import (
     moba_attn_varlen,
     process_moba_input,
@@ -22,6 +27,8 @@ from fastvideo_kernel.turbodiffusion_ops import (
 __all__ = [
     "sliding_tile_attention",
     "video_sparse_attn",
+    "block_sparse_attn",
+    "block_sparse_attn_from_indices",
     "moba_attn_varlen",
     "process_moba_input",
     "process_moba_output",

fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn.py

@@ -1,3 +1,5 @@
+"""Autograd-enabled block-sparse attention. Index-native ops with a bool-mask compat shim."""
+
 from __future__ import annotations
 
 import os
@@ -6,6 +8,11 @@ from typing import Tuple
 import torch
 
 
+# ---------------------------------------------------------------------------
+# Backend selection helpers
+# ---------------------------------------------------------------------------
+
+
 def _get_sm90_ops():
     try:
         from fastvideo_kernel._C import fastvideo_kernel_ops  # type: ignore
@@ -25,38 +32,66 @@ def _is_sm90() -> bool:
 
 
 def _force_triton() -> bool:
-    # Force Triton even on SM90 and even if the compiled extension is available.
-    # Useful for CI / debugging / parity testing.
     return os.environ.get("FASTVIDEO_KERNEL_VSA_FORCE_TRITON", "0") == "1"
 
 
-def _map_to_index(block_map: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Preferred map->index conversion used by the wrapper.
+# ---------------------------------------------------------------------------
+# Index helpers
+# ---------------------------------------------------------------------------
 
-    This wrapper **requires** the Triton implementation.
-    If Triton (or the Triton map_to_index module) is not available, it raises.
-    """
+
+def _map_to_index(block_map: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Compact a bool block_map to (q2k_idx, q2k_num). Legacy path only."""
     if block_map.dim() == 3:
         block_map = block_map.unsqueeze(0)
     if block_map.dim() != 4:
-        raise ValueError(f"block_map must be [B,H,Q,KV] (or [H,Q,KV]), got shape={tuple(block_map.shape)}")
+        raise ValueError(
+            f"block_map must be [B,H,Q,KV] (or [H,Q,KV]), "
+            f"got shape={tuple(block_map.shape)}"
+        )
     if block_map.dtype != torch.bool:
         block_map = block_map.to(torch.bool)
-
     if not block_map.is_cuda:
-        raise RuntimeError("block_map must be a CUDA tensor (Triton map_to_index required).")
+        raise RuntimeError(
+            "block_map must be a CUDA tensor (Triton map_to_index required)."
+        )
 
     try:
-        from fastvideo_kernel.triton_kernels.index import map_to_index as triton_map_to_index  # local import
-    except Exception as e:
+        from fastvideo_kernel.triton_kernels.index import map_to_index as triton_map_to_index
+    except Exception as e:  # pragma: no cover - environment issue
         raise ImportError(
             "Triton map_to_index is required but not available. "
-            "Ensure Triton is installed and fastvideo_kernel.triton_kernels.index is importable."
+            "Ensure Triton is installed and "
+            "fastvideo_kernel.triton_kernels.index is importable."
         ) from e
     return triton_map_to_index(block_map)
 
 
+def _invert_indices_for_backward(
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    num_kv_blocks: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    from fastvideo_kernel.triton_kernels.index import invert_indices
+    return invert_indices(q2k_idx, q2k_num, num_kv_blocks=num_kv_blocks)
+
+
+def _as_int32_contig(t: torch.Tensor, name: str) -> torch.Tensor:
+    """Return `t` as a contiguous int32 tensor, raising a clear error on CPU input."""
+    if not t.is_cuda:
+        raise RuntimeError(f"{name} must be a CUDA tensor, got device={t.device}")
+    if t.dtype != torch.int32:
+        t = t.to(torch.int32)
+    if not t.is_contiguous():
+        t = t.contiguous()
+    return t
+
+
+# ---------------------------------------------------------------------------
+# Triton backend custom ops (index-native)
+# ---------------------------------------------------------------------------
+
+
 @torch.library.custom_op(
     "fastvideo_kernel::block_sparse_attn_triton",
     mutates_args=(),
@@ -66,34 +101,40 @@ def block_sparse_attn_triton(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-
-    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (
         triton_block_sparse_attn_forward,
     )
 
-    o, M = triton_block_sparse_attn_forward(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
+    o, M = triton_block_sparse_attn_forward(
+        q.contiguous(),
+        k.contiguous(),
+        v.contiguous(),
+        q2k_idx,
+        q2k_num,
+        variable_block_sizes,
+    )
     return o, M
 
 
-
 @torch.library.register_fake("fastvideo_kernel::block_sparse_attn_triton")
 def _block_sparse_attn_triton_fake(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     o = torch.empty_like(q)
-    M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+    M = torch.empty(
+        (q.shape[0], q.shape[1], q.shape[2]),
+        device=q.device,
+        dtype=torch.float32,
+    )
     return o, M
 
 
@@ -109,20 +150,32 @@ def block_sparse_attn_backward_triton(
     v: torch.Tensor,
     o: torch.Tensor,
     M: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    grad_output = grad_output.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-    k2q_idx, k2q_num = _map_to_index(block_map.transpose(-1, -2).contiguous())
-
-    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (
         triton_block_sparse_attn_backward,
     )
 
+    num_kv_blocks = int(variable_block_sizes.numel())
+    k2q_idx, k2q_num = _invert_indices_for_backward(
+        q2k_idx, q2k_num, num_kv_blocks
+    )
+    # q/k/v are saved from the user-facing inputs and may be non-contiguous;
+    # o/M are kernel outputs so are already contiguous.
     dq, dk, dv = triton_block_sparse_attn_backward(
-        grad_output, q, k, v, o, M, q2k_idx, q2k_num, k2q_idx, k2q_num, variable_block_sizes
+        grad_output.contiguous(),
+        q.contiguous(),
+        k.contiguous(),
+        v.contiguous(),
+        o,
+        M,
+        q2k_idx,
+        q2k_num,
+        k2q_idx,
+        k2q_num,
+        variable_block_sizes,
     )
     return dq, dk, dv
 
@@ -135,7 +188,8 @@ def _block_sparse_attn_backward_triton_fake(
     v: torch.Tensor,
     o: torch.Tensor,
     M: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     dq = torch.empty_like(q)
@@ -144,19 +198,28 @@ def _block_sparse_attn_backward_triton_fake(
     return dq, dk, dv
 
 
-def _backward_triton(ctx, grad_o, grad_M):
-    q, k, v, o, M, block_map, variable_block_sizes = ctx.saved_tensors
-    dq, dk, dv = block_sparse_attn_backward_triton(grad_o, q, k, v, o, M, block_map, variable_block_sizes)
-    return dq, dk, dv, None, None
-
-
 def _setup_context_triton(ctx, inputs, output):
-    q, k, v, block_map, variable_block_sizes = inputs
+    q, k, v, q2k_idx, q2k_num, variable_block_sizes = inputs
     o, M = output
-    ctx.save_for_backward(q, k, v, o, M, block_map, variable_block_sizes)
+    ctx.save_for_backward(q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes)
 
 
-block_sparse_attn_triton.register_autograd(_backward_triton, setup_context=_setup_context_triton)
+def _backward_triton(ctx, grad_o, grad_M):
+    q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_triton(
+        grad_o, q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes
+    )
+    return dq, dk, dv, None, None, None
+
+
+block_sparse_attn_triton.register_autograd(
+    _backward_triton, setup_context=_setup_context_triton
+)
+
+
+# ---------------------------------------------------------------------------
+# SM90 backend custom ops (index-native)
+# ---------------------------------------------------------------------------
 
 
 @torch.library.custom_op(
@@ -168,21 +231,21 @@ def block_sparse_attn_sm90(
     q_padded: torch.Tensor,
     k_padded: torch.Tensor,
     v_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     block_sparse_fwd, _ = _get_sm90_ops()
     if block_sparse_fwd is None:
         raise ImportError("fastvideo_kernel_ops.block_sparse_fwd is not available")
 
-    q_padded = q_padded.contiguous()
-    k_padded = k_padded.contiguous()
-    v_padded = v_padded.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-
     o_padded, lse_padded = block_sparse_fwd(
-        q_padded, k_padded, v_padded, q2k_idx, q2k_num, variable_block_sizes.int()
+        q_padded.contiguous(),
+        k_padded.contiguous(),
+        v_padded.contiguous(),
+        q2k_idx,
+        q2k_num,
+        variable_block_sizes,
     )
     return o_padded, lse_padded
 
@@ -192,11 +255,16 @@ def _block_sparse_attn_sm90_fake(
     q_padded: torch.Tensor,
     k_padded: torch.Tensor,
     v_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     o = torch.empty_like(q_padded)
-    lse = torch.empty((q_padded.shape[0], q_padded.shape[1], q_padded.shape[2], 1), device=q_padded.device, dtype=torch.float32)
+    lse = torch.empty(
+        (q_padded.shape[0], q_padded.shape[1], q_padded.shape[2], 1),
+        device=q_padded.device,
+        dtype=torch.float32,
+    )
     return o, lse
 
 
@@ -212,30 +280,34 @@ def block_sparse_attn_backward_sm90(
     v_padded: torch.Tensor,
     o_padded: torch.Tensor,
     lse_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     _, block_sparse_bwd = _get_sm90_ops()
     if block_sparse_bwd is None:
         raise ImportError("fastvideo_kernel_ops.block_sparse_bwd is not available")
 
-    grad_output_padded = grad_output_padded.contiguous()
-    block_map = block_map.to(torch.bool)
-    k2q_idx, k2q_num = _map_to_index(block_map.transpose(-1, -2).contiguous())
+    num_kv_blocks = int(variable_block_sizes.numel())
+    k2q_idx, k2q_num = _invert_indices_for_backward(
+        q2k_idx, q2k_num, num_kv_blocks
+    )
 
+    # q/k/v are saved from user-facing inputs; o/lse are kernel outputs.
     dq, dk, dv = block_sparse_bwd(
-        q_padded,
-        k_padded,
-        v_padded,
+        q_padded.contiguous(),
+        k_padded.contiguous(),
+        v_padded.contiguous(),
         o_padded,
         lse_padded,
-        grad_output_padded,
+        grad_output_padded.contiguous(),
         k2q_idx,
         k2q_num,
-        variable_block_sizes.int(),
+        variable_block_sizes,
     )
-    # C++ kernel returns fp32 grads; cast back to match PyTorch convention if needed
-    return dq.to(grad_output_padded.dtype), dk.to(grad_output_padded.dtype), dv.to(grad_output_padded.dtype)
+    # C++ kernel returns fp32 grads; cast back to the input dtype.
+    out_dtype = grad_output_padded.dtype
+    return dq.to(out_dtype), dk.to(out_dtype), dv.to(out_dtype)
 
 
 @torch.library.register_fake("fastvideo_kernel::block_sparse_attn_backward_sm90")
@@ -246,7 +318,8 @@ def _block_sparse_attn_backward_sm90_fake(
     v_padded: torch.Tensor,
     o_padded: torch.Tensor,
     lse_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     dq = torch.empty_like(q_padded)
@@ -255,21 +328,57 @@ def _block_sparse_attn_backward_sm90_fake(
     return dq, dk, dv
 
 
-def _backward_sm90(ctx, grad_o, grad_lse):
-    q, k, v, o, lse, block_map, variable_block_sizes = ctx.saved_tensors
-    dq, dk, dv = block_sparse_attn_backward_sm90(
-        grad_o, q, k, v, o, lse, block_map, variable_block_sizes
-    )
-    return dq, dk, dv, None, None
-
-
 def _setup_context_sm90(ctx, inputs, output):
-    q, k, v, block_map, variable_block_sizes = inputs
+    q, k, v, q2k_idx, q2k_num, variable_block_sizes = inputs
     o, lse = output
-    ctx.save_for_backward(q, k, v, o, lse, block_map, variable_block_sizes)
+    ctx.save_for_backward(q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes)
 
 
-block_sparse_attn_sm90.register_autograd(_backward_sm90, setup_context=_setup_context_sm90)
+def _backward_sm90(ctx, grad_o, grad_lse):
+    q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_sm90(
+        grad_o, q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes
+    )
+    return dq, dk, dv, None, None, None
+
+
+block_sparse_attn_sm90.register_autograd(
+    _backward_sm90, setup_context=_setup_context_sm90
+)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+def block_sparse_attn_from_indices(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Block-sparse attention with autograd, taking compact per-row KV indices."""
+    # Normalize index tensors once at the public boundary so the custom ops
+    # and their fakes can assume int32/contiguous. No-op on well-formed input.
+    q2k_idx = _as_int32_contig(q2k_idx, "q2k_idx")
+    q2k_num = _as_int32_contig(q2k_num, "q2k_num")
+    variable_block_sizes = _as_int32_contig(variable_block_sizes, "variable_block_sizes")
+
+    block_sparse_fwd, block_sparse_bwd = _get_sm90_ops()
+    use_sm90 = (
+        (not _force_triton())
+        and _is_sm90()
+        and block_sparse_fwd is not None
+        and block_sparse_bwd is not None
+    )
+    if use_sm90:
+        return block_sparse_attn_sm90(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
+    # Triton path: supports q_seq_len != kv_seq_len as long as both are padded
+    # to a multiple of the block size (64 tokens).
+    return block_sparse_attn_triton(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
 
 
 def block_sparse_attn(
@@ -279,16 +388,8 @@ def block_sparse_attn(
     block_map: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Unified block-sparse attention op with autograd support.
-    - On SM90 with compiled extension present: uses fastvideo_kernel_ops.block_sparse_fwd/bwd.
-    - Otherwise: uses Triton implementation (requires q/k/v to have same padded length today).
-    """
-    block_sparse_fwd, block_sparse_bwd = _get_sm90_ops()
-    if (not _force_triton()) and _is_sm90() and (block_sparse_fwd is not None) and (block_sparse_bwd is not None):
-        return block_sparse_attn_sm90(q, k, v, block_map, variable_block_sizes)
-    # Triton path: supports q_seq_len != kv_seq_len as long as both are padded
-    # to a multiple of the block size (64 tokens).
-    return block_sparse_attn_triton(q, k, v, block_map, variable_block_sizes)
-
-
+    """Bool-mask compat wrapper; prefer block_sparse_attn_from_indices."""
+    q2k_idx, q2k_num = _map_to_index(block_map)
+    return block_sparse_attn_from_indices(
+        q, k, v, q2k_idx, q2k_num, variable_block_sizes
+    )

fastvideo-kernel/python/fastvideo_kernel/ops.py

@@ -1,6 +1,6 @@
 import math
 import torch
-from .block_sparse_attn import block_sparse_attn
+from .block_sparse_attn import block_sparse_attn, block_sparse_attn_from_indices
 from .triton_kernels.st_attn_triton import sliding_tile_attention_triton
 
 # Try to load the C++ extension
@@ -125,13 +125,18 @@ def video_sparse_attn(
     out_c = out_c.repeat(1, 1, 1, block_elements,
                          1).view(batch, heads, q_seq_len, dim)
 
-    # Sparse branch
+    # Sparse branch: feed top-k indices directly, skipping the bool-mask round-trip.
     topk_idx = torch.topk(scores, topk, dim=-1).indices
-    mask = torch.zeros_like(scores,
-                            dtype=torch.bool).scatter_(-1, topk_idx, True)
-
-    # out_s = block_sparse_attn(q, k, v, mask, variable_block_sizes)[0]
-    out_s = block_sparse_attn(q, k, v, mask, variable_block_sizes)[0]
+    q2k_idx = topk_idx.to(torch.int32).contiguous()
+    q2k_num = torch.full(
+        (batch, heads, q_num_blocks),
+        topk,
+        dtype=torch.int32,
+        device=q.device,
+    )
+    out_s = block_sparse_attn_from_indices(
+        q, k, v, q2k_idx, q2k_num, variable_block_sizes
+    )[0]
 
     if compress_attn_weight is not None:
         return out_c * compress_attn_weight + out_s

fastvideo-kernel/python/fastvideo_kernel/triton_kernels/index.py

@@ -1,9 +1,10 @@
 ## pytorch sdpa version of block sparse ##
+from typing import Tuple
+
 import triton
 import triton.language as tl
 import torch
 
-
 @triton.jit
 def topk_index_to_map_kernel(
     map_ptr,
@@ -153,3 +154,114 @@ def map_to_index(block_map: torch.Tensor):
     )
 
     return index, index_num
+
+
+@triton.jit
+def _invert_indices_kernel(
+    q2k_idx_ptr,
+    q2k_num_ptr,
+    k2q_idx_ptr,
+    k2q_num_ptr,
+    q2k_idx_b, q2k_idx_h, q2k_idx_q, q2k_idx_k,
+    q2k_num_b, q2k_num_h, q2k_num_q,
+    k2q_idx_b, k2q_idx_h, k2q_idx_k, k2q_idx_q,
+    k2q_num_b, k2q_num_h, k2q_num_k,
+    MAX_KV_PER_Q: tl.constexpr,
+):
+    # One program per (b, h, q): reserve a slot in k2q via atomicAdd, write q.
+    pid_b = tl.program_id(0)
+    pid_h = tl.program_id(1)
+    pid_q = tl.program_id(2)
+
+    n = tl.load(
+        q2k_num_ptr
+        + pid_b * q2k_num_b
+        + pid_h * q2k_num_h
+        + pid_q * q2k_num_q
+    )
+
+    q2k_row = (
+        q2k_idx_ptr
+        + pid_b * q2k_idx_b
+        + pid_h * q2k_idx_h
+        + pid_q * q2k_idx_q
+    )
+
+    for i in tl.range(0, MAX_KV_PER_Q):
+        if i < n:
+            kv = tl.load(q2k_row + i * q2k_idx_k)
+            count_ptr = (
+                k2q_num_ptr
+                + pid_b * k2q_num_b
+                + pid_h * k2q_num_h
+                + kv * k2q_num_k
+            )
+            pos = tl.atomic_add(count_ptr, 1)
+            tl.store(
+                k2q_idx_ptr
+                + pid_b * k2q_idx_b
+                + pid_h * k2q_idx_h
+                + kv * k2q_idx_k
+                + pos * k2q_idx_q,
+                pid_q,
+            )
+
+
+def invert_indices(
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    num_kv_blocks: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Transpose a Q->KV index list into a K->Q one via atomic compaction (GPU)."""
+    if q2k_idx.dim() != 4:
+        raise ValueError(
+            f"q2k_idx must be [B, H, Nq, Mk], got shape={tuple(q2k_idx.shape)}"
+        )
+    if q2k_num.dim() != 3:
+        raise ValueError(
+            f"q2k_num must be [B, H, Nq], got shape={tuple(q2k_num.shape)}"
+        )
+    if not q2k_idx.is_cuda or not q2k_num.is_cuda:
+        raise RuntimeError("invert_indices requires CUDA tensors.")
+
+    B, H, Nq, Mk = q2k_idx.shape
+    if q2k_num.shape != (B, H, Nq):
+        raise ValueError(
+            f"q2k_num shape {tuple(q2k_num.shape)} does not match q2k_idx "
+            f"[B, H, Nq] = {(B, H, Nq)}"
+        )
+
+    q2k_idx = q2k_idx.contiguous()
+    q2k_num = q2k_num.contiguous()
+    if q2k_idx.dtype != torch.int32:
+        q2k_idx = q2k_idx.to(torch.int32)
+    if q2k_num.dtype != torch.int32:
+        q2k_num = q2k_num.to(torch.int32)
+
+    # Any KV block is attended by at most Nq Q blocks (one per Q row), so
+    # `Nq` is a tight upper bound on the compacted K->Q slots.
+    k2q_idx = torch.empty(
+        (B, H, num_kv_blocks, Nq),
+        dtype=torch.int32,
+        device=q2k_idx.device,
+    )
+    k2q_num = torch.zeros(
+        (B, H, num_kv_blocks),
+        dtype=torch.int32,
+        device=q2k_idx.device,
+    )
+
+    grid = (B, H, Nq)
+    _invert_indices_kernel[grid](
+        q2k_idx,
+        q2k_num,
+        k2q_idx,
+        k2q_num,
+        q2k_idx.stride(0), q2k_idx.stride(1), q2k_idx.stride(2), q2k_idx.stride(3),
+        q2k_num.stride(0), q2k_num.stride(1), q2k_num.stride(2),
+        k2q_idx.stride(0), k2q_idx.stride(1), k2q_idx.stride(2), k2q_idx.stride(3),
+        k2q_num.stride(0), k2q_num.stride(1), k2q_num.stride(2),
+        MAX_KV_PER_Q=Mk,
+    )
+
+    return k2q_idx, k2q_num

fastvideo-kernel/python/fastvideo_kernel/vmoba.py

@@ -11,7 +11,7 @@ except ImportError:
 
     def _unsupported(*args, **kwargs):
         raise ImportError(
-            "flash-attn is not installed. Please install it, e.g., `pip install flash-attn`."
+            "flash-attn is not installed. Please install it, e.g., `uv pip install flash-attn`."
         )
 
     _flash_attn_varlen_forward = _unsupported

fastvideo/AGENTS.md

@@ -0,0 +1,67 @@
+# `fastvideo/` — Core Package
+
+**Generated:** 2026-05-02
+
+Inference + training framework for video DiTs. Public API entry: `from fastvideo import VideoGenerator, PipelineConfig, SamplingParam`.
+
+## Public Surface (`__init__.py`)
+
+```python
+VideoGenerator   # entrypoints/video_generator.py — high-level inference handle
+PipelineConfig   # configs/pipelines/base.py — pipeline wiring dataclass
+SamplingParam    # api/sampling_param.py — runtime sampling knobs
+```
+
+CLI entry: `fastvideo` script → `entrypoints/cli/main.py` (subcommands: `generate`, `serve`, `bench`).
+
+## Layout
+
+```
+fastvideo/
+├── api/             # Schema + presets for the OpenAI-compatible serving layer
+├── attention/       # Backends + selector (FlashAttn / SageAttn / SDPA / VSA / VMoBA / SLA)
+├── configs/         # Per-model arch configs + per-pipeline configs (registry-driven)
+├── dataset/         # Dataloaders (pre-commit excluded — minimal lint surface)
+├── distributed/     # SP/TP groups, device communicators, init helpers
+├── entrypoints/     # cli/, openai/, streaming/, video_generator.py
+├── hooks/           # Runtime hook system for pipelines
+├── layers/          # Tensor-parallel linears + attention wrappers (port targets)
+├── models/          # DiT / VAE / encoder / scheduler / loader (pre-commit excluded)
+├── pipelines/       # basic/<model>/, preprocess/, stages/, training/
+├── platforms/       # CUDA/ROCm capability + AttentionBackendEnum
+├── third_party/     # Vendored externals (lint excluded; do not reformat)
+├── train/           # NEW modular trainer — methods × models × callbacks
+├── training/        # LEGACY monolithic *_training/distillation_pipeline.py
+├── worker/          # Multi-process / Ray executors
+├── workflow/        # Preprocessing workflow base class
+├── registry.py      # Pipeline-config + model-class lookup (canonical)
+├── envs.py          # Env-var declarations
+├── fastvideo_args.py# Runtime arg dataclass passed through pipelines
+└── utils.py         # FlexibleArgumentParser, qualname resolver, etc.
+```
+
+## Where to Look
+
+| Task | Location |
+|------|----------|
+| Add a new pipeline class | `pipelines/basic/<model>/` + `configs/pipelines/<model>.py` + register in `registry.py` |
+| Add a new model component | `models/<role>/<model>.py` + `configs/models/<role>/<model>.py` |
+| Wire an existing model into a new pipeline | `pipelines/basic/<model>/presets.py` + reuse stages from `pipelines/stages/` |
+| Add a converter | `scripts/checkpoint_conversion/<model>_to_*.py` (separate dir, separate AGENTS.md) |
+| Add an attention backend | `attention/backends/<name>.py` + register in selector |
+| Add a runtime CLI flag | `fastvideo_args.py` (avoid `argparse` ad-hoc inside stages) |
+
+## Conventions Specific Here
+
+- `PipelineStage` subclasses (`pipelines/stages/`) own one verb each (encode, schedule, denoise, decode). Compose, don't fork.
+- Every pipeline reads from a `PipelineConfig` subclass and a `SamplingParam`. Never read raw env vars inside a stage — go through `fastvideo.envs`.
+- Logger setup: `from fastvideo.logger import init_logger; logger = init_logger(__name__)`. Do not call `logging.getLogger` directly.
+- Imports between `train/` and `training/` are **forbidden** — they are independent stacks.
+
+## Pre-Commit Exclusions (do not assume linted)
+
+These dirs are listed in `.pre-commit-config.yaml` `exclude`:
+
+- `fastvideo/third_party/`, `fastvideo/dataset/`, `fastvideo/models/`
+
+Editing files there will NOT trigger yapf/ruff/mypy/codespell. Format manually if a sibling file shows clear style; do not introduce new violations.

fastvideo/api/__init__.py

@@ -46,7 +46,14 @@ from fastvideo.api.parser import (
     load_serve_config,
     parse_config,
 )
-from fastvideo.api.results import GenerationResult
+from fastvideo.api.results import (
+    GenerationResult,
+    VideoEvent,
+    VideoFinalEvent,
+    VideoPartialEvent,
+    VideoProgressEvent,
+    VideoResult,
+)
 from fastvideo.api.sampling_param import SamplingParam
 
 __all__ = [
@@ -76,6 +83,11 @@ __all__ = [
     "ServeConfig",
     "ServerConfig",
     "StreamingConfig",
+    "VideoEvent",
+    "VideoFinalEvent",
+    "VideoPartialEvent",
+    "VideoProgressEvent",
+    "VideoResult",
     "WarmupConfig",
     "InferencePreset",
     "PresetStageSpec",

fastvideo/api/compat.py

@@ -16,6 +16,8 @@ from fastvideo.api.request_metadata import (
     reset_tracking_roots,
 )
 from fastvideo.api.schema import (
+    CompileConfig,
+    ContinuationState,
     GenerationRequest,
     GeneratorConfig,
     InputConfig,
@@ -25,6 +27,10 @@ from fastvideo.api.schema import (
 )
 from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.basic.ltx2.stage_overrides import (
+    refine_preset_override_fields,
+    refine_stage_override_fields,
+)
 from fastvideo.utils import shallow_asdict
 
 _INPUT_FIELD_NAMES = {field.name for field in fields(InputConfig)}
@@ -38,6 +44,10 @@ _LEGACY_REQUEST_ALIASES = {
 _REQUEST_PIPELINE_OVERRIDE_FIELDS = frozenset({
     "embedded_cfg_scale",
 })
+# torch.compile kwargs that map to first-class CompileConfig fields.
+_COMPILE_TYPED_KEYS = ("backend", "fullgraph", "mode", "dynamic")
+# LTX-2 refine flat kwargs (init + per-request) known to FastVideoArgs.
+_LTX2_REFINE_FLAT_KEYS = (refine_preset_override_fields() | refine_stage_override_fields())
 
 
 def normalize_generator_config(config: GeneratorConfig | Mapping[str, Any], ) -> GeneratorConfig:
@@ -80,6 +90,8 @@ def legacy_from_pretrained_to_config(
     components: dict[str, Any] = {}
     quantization: dict[str, Any] = {}
     experimental: dict[str, Any] = {}
+    preset_overrides: dict[str, Any] = {}
+    preset_refine: dict[str, Any] = {}
 
     for key, value in kwargs.items():
         if key == "revision":
@@ -106,8 +118,45 @@ def legacy_from_pretrained_to_config(
             offload["pin_cpu_memory"] = value
         elif key == "enable_torch_compile":
             compile_config["enabled"] = value
+        elif key == "enable_torch_compile_text_encoder":
+            compile_config["text_encoder_enabled"] = value
+        elif key == "enable_torch_compile_vae":
+            compile_config["vae_enabled"] = value
+        elif key == "enable_torch_compile_audio_vae":
+            compile_config["audio_vae_enabled"] = value
         elif key == "torch_compile_kwargs":
-            compile_config["kwargs"] = deepcopy(value)
+            remaining: dict[str, Any] = (dict(deepcopy(value)) if isinstance(value, Mapping) else {})
+            for first_class in _COMPILE_TYPED_KEYS:
+                if first_class in remaining:
+                    compile_config[first_class] = remaining.pop(first_class)
+            if remaining:
+                compile_config["extras"] = remaining
+        elif key in {
+                "torch_compile_kwargs_dit",
+                "torch_compile_kwargs_text_encoder",
+                "torch_compile_kwargs_vae",
+                "torch_compile_kwargs_audio_vae",
+        }:
+            compile_config[key[len("torch_compile_kwargs_"):] +
+                           "_kwargs"] = (dict(deepcopy(value)) if isinstance(value, Mapping) else {})
+        elif key == "ltx2_vae_tiling":
+            pipeline["vae_tiling"] = value
+        elif key == "config_model_path":
+            components["config_root"] = value
+        elif key == "ltx2_refine_enabled":
+            preset_refine["enabled"] = value
+        elif key == "ltx2_refine_upsampler_path":
+            # Empty string means "no upsampler"; keep typed None.
+            components["upsampler_weights"] = value or None
+        elif key == "ltx2_refine_lora_path":
+            # Empty string means "no refine LoRA"; keep typed None.
+            components["lora_path"] = value or None
+        elif key == "ltx2_refine_add_noise":
+            preset_refine["add_noise"] = value
+        elif key == "ltx2_refine_num_inference_steps":
+            preset_refine["num_inference_steps"] = value
+        elif key == "ltx2_refine_guidance_scale":
+            preset_refine["guidance_scale"] = value
         elif key in {"enable_stage_verification", "use_fsdp_inference", "disable_autocast"}:
             engine[key] = value
         elif key == "override_text_encoder_quant":
@@ -147,6 +196,10 @@ def legacy_from_pretrained_to_config(
 
     if components:
         pipeline["components"] = components
+    if preset_refine:
+        preset_overrides["refine"] = preset_refine
+    if preset_overrides:
+        pipeline["preset_overrides"] = preset_overrides
     if experimental:
         pipeline["experimental"] = experimental
     if pipeline:
@@ -162,12 +215,8 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         unsupported.append("pipeline.preset")
     if normalized.pipeline.preset_version is not None:
         unsupported.append("pipeline.preset_version")
-    if normalized.pipeline.components.config_root is not None:
-        unsupported.append("pipeline.components.config_root")
     if normalized.pipeline.components.vae_weights is not None:
         unsupported.append("pipeline.components.vae_weights")
-    if normalized.pipeline.components.upsampler_weights is not None:
-        unsupported.append("pipeline.components.upsampler_weights")
     if unsupported:
         joined = ", ".join(unsupported)
         raise NotImplementedError(f"VideoGenerator compatibility adapter does not support {joined} yet")
@@ -191,19 +240,43 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         "vae_cpu_offload": engine.offload.vae,
         "pin_cpu_memory": engine.offload.pin_cpu_memory,
         "enable_torch_compile": engine.compile.enabled,
-        "torch_compile_kwargs": deepcopy(engine.compile.kwargs),
+        "torch_compile_kwargs": _compile_config_to_torch_kwargs(engine.compile),
         "enable_stage_verification": engine.enable_stage_verification,
         "use_fsdp_inference": engine.use_fsdp_inference,
         "disable_autocast": engine.disable_autocast,
     }
     if normalized.pipeline.workload_type is not None:
         kwargs["workload_type"] = normalized.pipeline.workload_type
+    if normalized.pipeline.vae_tiling is not None:
+        kwargs["ltx2_vae_tiling"] = normalized.pipeline.vae_tiling
+    if engine.compile.text_encoder_enabled is not None:
+        kwargs["enable_torch_compile_text_encoder"] = (engine.compile.text_encoder_enabled)
+    if engine.compile.vae_enabled is not None:
+        kwargs["enable_torch_compile_vae"] = engine.compile.vae_enabled
+    if engine.compile.audio_vae_enabled is not None:
+        kwargs["enable_torch_compile_audio_vae"] = (engine.compile.audio_vae_enabled)
+    if engine.compile.dit_kwargs:
+        kwargs["torch_compile_kwargs_dit"] = deepcopy(engine.compile.dit_kwargs)
+    if engine.compile.text_encoder_kwargs:
+        kwargs["torch_compile_kwargs_text_encoder"] = deepcopy(engine.compile.text_encoder_kwargs)
+    if engine.compile.vae_kwargs:
+        kwargs["torch_compile_kwargs_vae"] = deepcopy(engine.compile.vae_kwargs)
+    if engine.compile.audio_vae_kwargs:
+        kwargs["torch_compile_kwargs_audio_vae"] = deepcopy(engine.compile.audio_vae_kwargs)
 
     quantization = engine.quantization
     if quantization is not None and quantization.text_encoder_quant is not None:
         kwargs["override_text_encoder_quant"] = quantization.text_encoder_quant
     if quantization is not None and quantization.transformer_quant is not None:
-        kwargs["transformer_quant"] = quantization.transformer_quant
+        # Resolve the typed quant name to a concrete ``QuantizationConfig``
+        # instance and pin it on ``dit_config.quant_config``. The legacy
+        # path expected callers to do this themselves via
+        # ``pipeline_config.dit_config.quant_config = NVFP4Config()``; the
+        # typed surface accepts a string and does the wiring here so
+        # downstream code can rely on a single source of truth.
+        from fastvideo.layers.quantization import get_quantization_config
+        _resolved_quant_cls = get_quantization_config(quantization.transformer_quant)
+        kwargs["transformer_quant"] = _resolved_quant_cls()
 
     components = normalized.pipeline.components
     if components.pipeline_config_path is not None:
@@ -220,8 +293,18 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         kwargs["init_weights_from_safetensors"] = components.transformer_weights
     if components.transformer_2_weights is not None:
         kwargs["init_weights_from_safetensors_2"] = components.transformer_2_weights
+    if components.config_root is not None:
+        kwargs["config_model_path"] = components.config_root
+    if components.upsampler_weights is not None:
+        kwargs["ltx2_refine_upsampler_path"] = components.upsampler_weights
 
-    kwargs.update(deepcopy(normalized.pipeline.preset_overrides))
+    preset_overrides = deepcopy(normalized.pipeline.preset_overrides)
+    refine = preset_overrides.pop("refine", None)
+    if isinstance(refine, Mapping):
+        for key in _LTX2_REFINE_FLAT_KEYS:
+            if key in refine:
+                kwargs[f"ltx2_refine_{key}"] = refine[key]
+    kwargs.update(preset_overrides)
     kwargs.update(deepcopy(normalized.pipeline.experimental))
     return FastVideoArgs.from_kwargs(**kwargs)
 
@@ -271,10 +354,13 @@ def request_to_sampling_param(
 ) -> SamplingParam:
     if request.plan is not None:
         raise NotImplementedError("GenerationRequest.plan is not wired into VideoGenerator yet")
-    if request.state is not None:
-        raise NotImplementedError("GenerationRequest.state is not wired into VideoGenerator yet")
 
     sampling_param = SamplingParam.from_pretrained(model_path)
+    if request.state is not None:
+        _validate_continuation_state(request.state)
+        sampling_param.continuation_state = request.state
+    if request.output.return_state:
+        sampling_param.return_continuation_state = True
     updates = explicit_request_updates(request)
 
     for key, value in updates.items():
@@ -316,6 +402,25 @@ def _looks_like_run_or_serve_config(raw: Mapping[str, Any]) -> bool:
     return isinstance(raw.get("generator"), Mapping)
 
 
+def _compile_config_to_torch_kwargs(compile_config: CompileConfig, ) -> dict[str, Any]:
+    """Flatten typed ``CompileConfig`` back to a ``torch_compile_kwargs``
+    dict that the legacy ``FastVideoArgs`` path still expects.
+
+    Typed first-class fields (:attr:`backend`, :attr:`fullgraph`,
+    :attr:`mode`, :attr:`dynamic`) are only emitted when the user set
+    them explicitly (non-``None``). ``extras`` is merged on top for any
+    uncommon kwargs.
+    """
+    out: dict[str, Any] = {}
+    for key in _COMPILE_TYPED_KEYS:
+        value = getattr(compile_config, key)
+        if value is not None:
+            out[key] = value
+    if compile_config.extras:
+        out.update(deepcopy(compile_config.extras))
+    return out
+
+
 def _sampling_param_to_request_raw(sampling_param: SamplingParam | None, ) -> dict[str, Any]:
     if sampling_param is None:
         return {}
@@ -476,6 +581,37 @@ def _serialize_generation_request(request: GenerationRequest) -> dict[str, Any]:
 
 _SCHEMA_DEFAULT_UPDATES = _extract_request_updates(config_to_dict(GenerationRequest()))
 
+_KNOWN_CONTINUATION_KINDS: set[str] = set()
+
+
+def register_continuation_kind(kind: str) -> None:
+    """Register a :class:`ContinuationState.kind` as recognized.
+
+    PR 7 wires the envelope through; per-kind payload deserializers live
+    with each model family (e.g. ``fastvideo.pipelines.basic.ltx2.
+    continuation.LTX2ContinuationState``). The registry lets the
+    public-API compat layer validate the kind early, before the state
+    reaches the pipeline.
+    """
+    if not isinstance(kind, str) or not kind:
+        raise ValueError("ContinuationState kind must be a non-empty string")
+    _KNOWN_CONTINUATION_KINDS.add(kind)
+
+
+def _validate_continuation_state(state: ContinuationState) -> None:
+    if not isinstance(state.kind, str) or not state.kind:
+        raise ValueError("GenerationRequest.state.kind must be a non-empty string; got "
+                         f"{state.kind!r}")
+    if not isinstance(state.payload, Mapping):
+        raise ValueError(f"GenerationRequest.state.payload must be a mapping; got "
+                         f"{type(state.payload).__name__}")
+    if state.kind not in _KNOWN_CONTINUATION_KINDS:
+        known = sorted(_KNOWN_CONTINUATION_KINDS)
+        raise ValueError(f"Unknown ContinuationState kind {state.kind!r}; registered "
+                         f"kinds: {known}. Import the model family that owns this kind "
+                         "(e.g. `import fastvideo.pipelines.basic.ltx2.continuation`) "
+                         "to register it, or drop the state field.")
+
 
 def _fan_out_batched_input_value(
     source_request: GenerationRequest,
@@ -509,6 +645,7 @@ __all__ = [
     "load_generator_config_from_file",
     "normalize_generation_request",
     "normalize_generator_config",
+    "register_continuation_kind",
     "request_to_pipeline_overrides",
     "request_to_sampling_param",
 ]

fastvideo/api/results.py

@@ -15,6 +15,7 @@ class GenerationResult:
     samples: Any | None = None
     frames: Any | None = None
     audio: Any | None = None
+    audio_sample_rate: int | None = None
     size: tuple[int, int, int] | None = None
     generation_time: float | None = None
     logging_info: Any | None = None
@@ -44,6 +45,7 @@ class GenerationResult:
                 "samples",
                 "frames",
                 "audio",
+                "audio_sample_rate",
                 "size",
                 "generation_time",
                 "logging_info",
@@ -62,6 +64,7 @@ class GenerationResult:
             samples=result.get("samples"),
             frames=result.get("frames"),
             audio=result.get("audio"),
+            audio_sample_rate=result.get("audio_sample_rate"),
             size=result.get("size"),
             generation_time=result.get("generation_time"),
             logging_info=result.get("logging_info"),
@@ -80,6 +83,7 @@ class GenerationResult:
             "samples": self.samples,
             "frames": self.frames,
             "audio": self.audio,
+            "audio_sample_rate": self.audio_sample_rate,
             "size": self.size,
             "generation_time": self.generation_time,
             "logging_info": self.logging_info,
@@ -98,4 +102,72 @@ class GenerationResult:
         return result
 
 
-__all__ = ["GenerationResult"]
+# Alias the canonical result type; matches the public docs.
+VideoResult = GenerationResult
+
+
+@dataclass
+class VideoProgressEvent:
+    """Per-step progress event emitted by :meth:`VideoGenerator.generate_async`.
+
+    Consumers treat these as best-effort telemetry; ``total_steps`` is
+    the count the pipeline reported at the start of the run, not a
+    rolling estimate.
+    """
+
+    step: int
+    total_steps: int
+    stage: str = "denoise"
+    """Logical stage name (``denoise`` | ``refine`` | ``decode`` | …)."""
+
+
+@dataclass
+class VideoPartialEvent:
+    """Chunk of decoded frames ready for streaming.
+
+    Emitted only on the streaming path; the aggregated code path never
+    yields partials. ``frames`` is a numpy ``(N, H, W, 3)`` uint8
+    ndarray; ``index`` is a monotonic chunk index starting at 0.
+    """
+
+    frames: Any
+    index: int
+
+
+@dataclass
+class VideoFinalEvent:
+    """Terminal event carrying the generated video and metadata.
+
+    Exactly one ``VideoFinalEvent`` is emitted per request. When
+    ``request.output.return_state`` is True the event also carries the
+    :class:`ContinuationState` the caller needs to resume.
+    """
+
+    video_bytes: bytes | None = None
+    tensor: Any | None = None
+    frames: Any | None = None
+    metadata: dict[str, Any] = field(default_factory=dict)
+    continuation_state: ContinuationState | None = None
+    result: VideoResult | None = None
+    """The full :class:`VideoResult` for callers that want everything.
+
+    Streaming consumers typically only care about ``frames`` /
+    ``continuation_state``; keeping the full result here avoids a
+    second code path."""
+
+
+VideoEvent = VideoProgressEvent | VideoPartialEvent | VideoFinalEvent
+"""Union of every event :meth:`VideoGenerator.generate_async` yields.
+
+Consumers match by ``isinstance`` rather than ``type`` so subclasses
+(e.g. a future ``VideoAudioSegmentEvent``) slot in without breaking
+existing code."""
+
+__all__ = [
+    "GenerationResult",
+    "VideoEvent",
+    "VideoFinalEvent",
+    "VideoPartialEvent",
+    "VideoProgressEvent",
+    "VideoResult",
+]

fastvideo/api/sampling_param.py

@@ -1,11 +1,16 @@
 # SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
 import copy
 from dataclasses import dataclass, field, fields
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from fastvideo.logger import init_logger
 from fastvideo.utils import StoreBoolean
 
+if TYPE_CHECKING:
+    from fastvideo.api.schema import ContinuationState
+
 logger = init_logger(__name__)
 
 
@@ -92,17 +97,71 @@ class SamplingParam:
     movement_distance: float | None = None
     camera_rotation: str | None = None
 
-    # LTX2 multi-modal CFG and STG
-    ltx2_cfg_scale_video: float = 3.0
-    ltx2_cfg_scale_audio: float = 7.0
-    ltx2_modality_scale_video: float = 3.0
-    ltx2_modality_scale_audio: float = 3.0
-    ltx2_rescale_scale: float = 0.7
-    ltx2_stg_scale_video: float = 1.0
-    ltx2_stg_scale_audio: float = 1.0
+    # LTX-2 multi-modal CFG and STG.
+    # Class-level defaults match the *distilled* LTX-2 schedule
+    # (mirrors ``FastVideo-internal/.../LTX2DistilledSamplingParam``):
+    # the distilled model expects neutral guidance scales — modality 1,
+    # rescale 0, STG 0 — and explicit-CFG callers (full LTX-2) opt back
+    # in by selecting the ``LTX2_BASE`` preset, which overrides these
+    # to mod=3.0 / rescale=0.7 / stg=1.0 in its ``defaults`` dict.
+    # cfg_scale defaults stay at 1.0 (CFG off) so
+    # ``ForwardBatch.__post_init__`` doesn't force CFG on non-LTX-2
+    # models that never override these fields.
+    ltx2_cfg_scale_video: float = 1.0
+    ltx2_cfg_scale_audio: float = 1.0
+    ltx2_modality_scale_video: float = 1.0
+    ltx2_modality_scale_audio: float = 1.0
+    ltx2_rescale_scale: float = 0.0
+    ltx2_stg_scale_video: float = 0.0
+    ltx2_stg_scale_audio: float = 0.0
     ltx2_stg_blocks_video: list[int] = field(default_factory=lambda: [29])
     ltx2_stg_blocks_audio: list[int] = field(default_factory=lambda: [29])
 
+    # LTX-2 image / video / continuation conditioning. These flow from
+    # generate_video(...) kwargs through ``sampling_param.update(kwargs)``
+    # onto the ForwardBatch fields of the same name. ``ltx2_image_crf``
+    # gates the conditioning-image H.264 re-encode; the streaming
+    # session controller passes ``ltx2_image_crf=0.0`` because it
+    # conditions on already-decoded VAE-quality frames.
+    ltx2_images: list[tuple[str, int, float]] | None = None
+    ltx2_image_crf: float = 33.0
+    ltx2_conditioning_latent_stage1: Any | None = None
+    ltx2_conditioning_latent_stage2: Any | None = None
+    ltx2_video_conditions: list[tuple[list[str], int, float]] | None = None
+
+    # Stable Audio (T2A): clip start/end in seconds. Honored by
+    # `StableAudioConditioningStage` + `StableAudioDecodingStage`. Other
+    # families ignore them.
+    audio_start_in_s: float | None = None
+    audio_end_in_s: float | None = None
+
+    # Stable Audio audio-to-audio (variation):
+    #   `init_audio` -- a path or `[B, C, samples]` waveform at the model
+    #                   sample rate; the pipeline encodes it via the VAE
+    #                   and uses it as the starting latent.
+    #   `init_audio_strength` -- 0..1, higher = closer to the reference
+    #                            (matches the convention of Stability's
+    #                            commercial Stable Audio 2.0 UI). 1.0 ~=
+    #                            VAE round-trip, 0.0 ~= plain T2A.
+    #   `init_noise_level` -- legacy raw `sigma_max` override (0.3..500,
+    #                         higher = more freedom). Kept for callers
+    #                         that already use it; prefer `init_audio_strength`.
+    init_audio: Any = None
+    init_audio_strength: float | None = None
+    init_noise_level: float | None = None
+
+    # Stable Audio inpainting (RePaint-style): `inpaint_audio` is the
+    # reference clip, `inpaint_mask` is a [samples] tensor in {0, 1} where
+    # 1 means *keep the reference* and 0 means *regenerate*.
+    inpaint_audio: Any = None
+    inpaint_mask: Any = None
+
+    # Continuation state carried across streaming/multi-segment calls.
+    continuation_state: ContinuationState | None = None
+    # When True, the pipeline returns a ContinuationState on the result so
+    # the caller can resume from the generated segment.
+    return_continuation_state: bool = False
+
     # Misc
     save_video: bool = True
     return_frames: bool = True
@@ -127,7 +186,7 @@ class SamplingParam:
         self.__post_init__()
 
     @classmethod
-    def from_pretrained(cls, model_path: str) -> "SamplingParam":
+    def from_pretrained(cls, model_path: str) -> SamplingParam:
         sampling_param = cls._from_preset(model_path)
         if sampling_param is not None:
             return sampling_param
@@ -143,7 +202,7 @@ class SamplingParam:
     def _from_preset(
         cls,
         model_path: str,
-    ) -> "SamplingParam | None":
+    ) -> SamplingParam | None:
         """Build a SamplingParam from preset defaults.
 
         Returns ``None`` when no preset is configured for

fastvideo/api/schema.py

@@ -33,8 +33,36 @@ class OffloadConfig:
 
 @dataclass
 class CompileConfig:
+    """Typed ``torch.compile`` configuration.
+
+    ``backend``/``fullgraph``/``mode``/``dynamic`` are the four most
+    common ``torch.compile`` knobs. ``extras`` holds any remaining
+    ``torch.compile`` kwargs (e.g. ``options``, ``disable``).
+
+    The ``enabled`` switch covers the DiT transformer path (including
+    ``transformer_2`` and the LTX-2 stage-2 ``transformer_refine``).
+    Per-component flags below are independent overlays — set to ``True``
+    to compile that component, ``None`` to leave it eager. Each
+    ``*_kwargs`` dict overrides the master ``backend``/``fullgraph``/
+    ``mode``/``dynamic``/``extras`` for that component when non-empty;
+    leaving it empty inherits the master kwargs.
+    """
+
     enabled: bool = False
-    kwargs: dict[str, Any] = field(default_factory=dict)
+    backend: str | None = None
+    fullgraph: bool | None = None
+    mode: str | None = None
+    dynamic: bool | None = None
+    extras: dict[str, Any] = field(default_factory=dict)
+
+    text_encoder_enabled: bool | None = None
+    vae_enabled: bool | None = None
+    audio_vae_enabled: bool | None = None
+
+    dit_kwargs: dict[str, Any] = field(default_factory=dict)
+    text_encoder_kwargs: dict[str, Any] = field(default_factory=dict)
+    vae_kwargs: dict[str, Any] = field(default_factory=dict)
+    audio_vae_kwargs: dict[str, Any] = field(default_factory=dict)
 
 
 @dataclass
@@ -76,6 +104,8 @@ class PipelineSelection:
     preset: str | None = None
     preset_version: int | None = None
     components: ComponentConfig = field(default_factory=ComponentConfig)
+    vae_tiling: bool | None = None
+    """Tile-based VAE decode. ``None`` keeps the model's default."""
     preset_overrides: dict[str, Any] = field(default_factory=dict)
     experimental: dict[str, Any] = field(default_factory=dict)
 

fastvideo/attention/AGENTS.md

@@ -0,0 +1,58 @@
+# `fastvideo/attention/` — Attention Backends
+
+**Generated:** 2026-05-02
+
+Backend registry + selector wrapping FlashAttn / SageAttn / SageAttn3 / SDPA / VSA / VMoBA / SLA / BSA.
+
+## Layout
+
+```
+attention/
+├── __init__.py            # Exports DistributedAttention, LocalAttention, get_attn_backend
+├── layer.py               # DistributedAttention, DistributedAttention_VSA, LocalAttention
+├── selector.py            # get_attn_backend (cached) + env-var override
+├── backends/
+│   ├── abstract.py        #   AttentionBackend / AttentionMetadata / AttentionMetadataBuilder
+│   ├── flash_attn.py      #   FA2/FA3
+│   ├── sage_attn.py       #   SageAttention v1
+│   ├── sage_attn3.py      #   SageAttention v3
+│   ├── sdpa.py            #   torch SDPA fallback
+│   ├── video_sparse_attn.py  # VSA (paper: Video Sparse Attention)
+│   ├── vmoba.py           #   Video-MoBA
+│   ├── sla.py             #   Sliding-window (STA)
+│   └── bsa_attn.py        #   Block-sparse
+└── utils/
+    ├── flash_attn_cute.py
+    └── flash_attn_no_pad.py
+```
+
+## Selection Order
+
+`get_attn_backend()` resolves via:
+
+1. Env-var override `FASTVIDEO_ATTENTION_BACKEND` (see `STR_BACKEND_ENV_VAR` in `fastvideo/utils.py`).
+2. Per-platform default from `fastvideo/platforms/`.
+3. Heuristic fallback to SDPA.
+
+The result is `@lru_cache`d. Tests that need a specific backend must use the
+`global_force_attn_backend(...)` context manager from `selector.py`, never set
+the env var mid-process.
+
+## Adding a Backend
+
+1. Subclass `AttentionBackend` in `backends/<name>.py`.
+2. Implement `AttentionMetadata` + `AttentionMetadataBuilder` for the new path.
+3. Register the enum value in `fastvideo/platforms/interface.py` (`AttentionBackendEnum`).
+4. Wire string → class resolution in `selector.py`.
+5. Verify the new backend works with `DistributedAttention` (sequence parallel)
+   and `LocalAttention` (single-rank). If it cannot support SP, document the
+   gap in the backend file's module docstring.
+
+## Anti-Patterns
+
+- Calling `torch.nn.functional.scaled_dot_product_attention` directly inside a
+  model's forward — go through `DistributedAttention` / `LocalAttention`.
+- Reading `os.environ[STR_BACKEND_ENV_VAR]` from arbitrary call sites. Use
+  `get_env_variable_attn_backend()`.
+- Caching backend instances per-module. The selector cache is process-wide; do
+  not duplicate it.

fastvideo/attention/backends/flash_attn.py

@@ -2,7 +2,6 @@
 
 import torch
 import torch.nn.functional as F
-from flash_attn import flash_attn_func as flash_attn_2_func
 from dataclasses import dataclass
 
 try:
@@ -18,6 +17,7 @@ except ImportError:
         flash_attn_func = flash_attn_3_func
         fa_version = "3"
     except ImportError:
+        from flash_attn import flash_attn_func as flash_attn_2_func
         flash_attn_func = flash_attn_2_func
         fa_version = "2"
 

fastvideo/attention/backends/sla.py

@@ -405,7 +405,7 @@ class SageSLAAttentionImpl(AttentionImpl, nn.Module):
 
         if not SAGESLA_ENABLED:
             raise ImportError("SageSLA requires spas_sage_attn. "
-                              "Install with: pip install git+https://github.com/thu-ml/SpargeAttn.git")
+                              "Install with: uv pip install git+https://github.com/thu-ml/SpargeAttn.git")
 
         assert head_size in [64, 128], f"SageSLA requires head_size in [64, 128], got {head_size}"
 

fastvideo/configs/AGENTS.md

@@ -0,0 +1,53 @@
+# `fastvideo/configs/` — Config-Driven Model Registry
+
+**Generated:** 2026-05-02
+
+Two layers of dataclass configs feed every pipeline: **arch configs** (what the model is) and **pipeline configs** (how to run it).
+
+## Layout
+
+```
+configs/
+├── configs.py                  # Dataset / loader enums (DatasetType, VideoLoaderType)
+├── utils.py                    # update_config_from_args, shallow_asdict helpers
+├── backend/                    # Attention backend defaults
+├── models/
+│   ├── base.py                 # ModelConfig ABC
+│   ├── dits/                   # DiTConfig per model (wanvideo, ltx2, hunyuan, ...)
+│   ├── vaes/                   # VAEConfig per model
+│   ├── encoders/               # EncoderConfig (t5, clip, llama, qwen2_5, gemma, siglip, ...)
+│   ├── upsamplers/             # UpsamplerConfig (hunyuan15)
+│   └── audio/                  # Audio-model configs (ltx2_audio_vae, ...)
+├── pipelines/
+│   ├── base.py                 # PipelineConfig ABC + (de)serialization
+│   └── <model>.py              # Concrete configs (HunyuanConfig, WanT2V480PConfig, ...)
+└── *.json                      # Frozen reference configs for shipped models
+```
+
+## How Configs Hook Into the Registry
+
+`fastvideo/registry.py` imports every concrete `PipelineConfig` and exposes
+`get_pipeline_config_cls_from_name(...)`. Adding a new pipeline config requires:
+
+1. Subclass `PipelineConfig` in `pipelines/<model>.py`.
+2. Reference its component arch configs (DiT / VAE / encoder / upsampler).
+3. Add the import + name mapping in `fastvideo/registry.py`.
+
+Configs that do not appear in `registry.py` are unreachable from `VideoGenerator`.
+
+## Arch vs Pipeline — Where Does This Field Go?
+
+| Field type | Lives on |
+|-----------|----------|
+| Architecture constants (hidden dim, num heads, layer count) | `configs/models/<role>/<model>.py` |
+| Default sampling params (steps, cfg, shift, fps) | `configs/pipelines/<model>.py` |
+| Runtime overrides (precision, sp_size, tp_size, attention backend) | `configs/pipelines/base.py` defaults + CLI flags via `fastvideo_args.py` |
+| `param_names_mapping` for HF → FastVideo state-dict | Arch config (lives with the model definition) |
+
+If a knob is tunable per inference call → `SamplingParam`, not `PipelineConfig`.
+
+## Anti-Patterns
+
+- Hard-coding architecture constants inside model classes — always read from the arch config.
+- Using `argparse` directly here. Configs deserialize from dicts via `update_config_from_args`.
+- Importing from `fastvideo.pipelines` here. Configs are the lower layer; the dependency is one-way.

fastvideo/configs/models/base.py

@@ -48,8 +48,6 @@ class ModelConfig:
         for key, value in source_model_dict.items():
             if key in valid_fields:
                 setattr(arch_config, key, value)
-            else:
-                raise AttributeError(f"{type(arch_config).__name__} has no field '{key}'")
 
         if hasattr(arch_config, "__post_init__"):
             arch_config.__post_init__()

fastvideo/configs/models/dits/__init__.py

@@ -5,11 +5,13 @@ from fastvideo.configs.models.dits.hunyuanvideo import HunyuanVideoConfig
 from fastvideo.configs.models.dits.hunyuanvideo15 import HunyuanVideo15Config
 from fastvideo.configs.models.dits.longcat import LongCatVideoConfig
 from fastvideo.configs.models.dits.ltx2 import LTX2VideoConfig
+from fastvideo.configs.models.dits.stable_audio import StableAudioConfig
 from fastvideo.configs.models.dits.wanvideo import WanVideoConfig
 from fastvideo.configs.models.dits.hyworld import HYWorldConfig
 from fastvideo.configs.models.dits.kandinsky5 import Kandinsky5VideoConfig
 
 __all__ = [
     "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "CosmosVideoConfig",
-    "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig"
+    "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig",
+    "StableAudioConfig"
 ]

fastvideo/configs/models/dits/cosmos.py

@@ -50,7 +50,8 @@ class CosmosArchConfig(DiTArchConfig):
         })
 
     # Cosmos-specific config parameters based on transformer_cosmos.py
-    in_channels: int = 16
+    # in_channels includes the condition_mask channel (16 latent + 1 cond = 17)
+    in_channels: int = 17
     out_channels: int = 16
     num_attention_heads: int = 16
     attention_head_dim: int = 128

fastvideo/configs/models/dits/stable_audio.py

@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Config for the Stable Audio Open 1.0 DiT.
+
+Note: the SA pipeline bypasses the standard `ComposedPipelineBase`
+component loader because the published HF repo ships a single monolithic
+`model.safetensors` (no Diffusers-style `model_index.json` or
+per-subfolder layout). The arch fields and `param_names_mapping` here
+document the architecture and key remap so the same conventions used by
+the rest of the DiT family apply (FSDP shard conditions, supported
+attention backends, future loader integrations) — they are not currently
+consumed by `fastvideo/models/loader/fsdp_load.py` for SA.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+from fastvideo.platforms import AttentionBackendEnum
+
+
+def _is_transformer_layer(n: str, m) -> bool:
+    # Matches `transformer.layers.{i}` in the SA DiT module tree.
+    parts = n.split(".")
+    return (len(parts) >= 3 and parts[-3] == "transformer" and parts[-2] == "layers" and parts[-1].isdigit())
+
+
+@dataclass
+class StableAudioArchConfig(DiTArchConfig):
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [_is_transformer_layer])
+
+    # SA's checkpoint is `stable_audio_tools` raw format (not Diffusers),
+    # so the only remaps are: strip the `model.model.` host-pipeline
+    # prefix, and rename `nn.LayerNorm`'s `gamma`/`beta` to torch's
+    # canonical `weight`/`bias`. Linear / cross-attention naming already
+    # matches FastVideo's conventions, so no further remap is needed.
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^model\.model\.(.*?)\.gamma$": r"\1.weight",
+            r"^model\.model\.(.*?)\.beta$": r"\1.bias",
+            r"^model\.model\.(.*)$": r"\1",
+        })
+
+    # SA only supports backends compatible with single-GPU LocalAttention.
+    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+    )
+
+    # Architecture constants (from the published `model_config.json` for
+    # `stabilityai/stable-audio-open-1.0`).
+    io_channels: int = 64
+    embed_dim: int = 1536
+    depth: int = 24
+    num_attention_heads: int = 24
+    cond_token_dim: int = 768
+    global_cond_dim: int = 1536
+    project_cond_tokens: bool = False
+    project_global_cond: bool = True
+    # Set to "ln" to wrap attention Q/K in LayerNorm (used by
+    # `stable-audio-open-small`; absent in the 1.0 base).
+    qk_norm: str | None = None
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.hidden_size = self.embed_dim
+        self.in_channels = self.io_channels
+        self.out_channels = self.io_channels
+        self.num_channels_latents = self.io_channels
+        self.attention_head_dim = self.embed_dim // self.num_attention_heads
+
+
+@dataclass
+class StableAudioConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=StableAudioArchConfig)
+
+    prefix: str = "StableAudio"

fastvideo/configs/models/encoders/__init__.py

@@ -7,9 +7,12 @@ from fastvideo.configs.models.encoders.qwen2_5 import Qwen2_5_VLConfig
 from fastvideo.configs.models.encoders.siglip import SiglipVisionConfig
 from fastvideo.configs.models.encoders.reason1 import Reason1ArchConfig, Reason1Config
 from fastvideo.configs.models.encoders.gemma import LTX2GemmaConfig
+from fastvideo.configs.models.encoders.stable_audio_conditioner import (StableAudioConditionerArchConfig,
+                                                                        StableAudioConditionerConfig)
 
 __all__ = [
     "EncoderConfig", "TextEncoderConfig", "ImageEncoderConfig", "BaseEncoderOutput", "CLIPTextConfig",
     "CLIPVisionConfig", "WAN2_1ControlCLIPVisionConfig", "LlamaConfig", "T5Config", "T5LargeConfig", "Qwen2_5_VLConfig",
-    "Reason1ArchConfig", "Reason1Config", "LTX2GemmaConfig", "SiglipVisionConfig"
+    "Reason1ArchConfig", "Reason1Config", "LTX2GemmaConfig", "SiglipVisionConfig", "StableAudioConditionerArchConfig",
+    "StableAudioConditionerConfig"
 ]