update

Post-first-run corrections to the seed-ssim-references skill:
- modal volume get needs --force when the local parent directory
  already exists, otherwise it errors with [Errno 21] Is a directory.
- The downloaded tree has an extra generated_videos/ level (from the
  volume layout in _sync_generated_videos_to_volume), so copy-local's
  --generated-dir must include it.

.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.jsonl

@@ -5,3 +5,4 @@
 {"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"}

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

@@ -0,0 +1,250 @@
+---
+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:
+
+```bash
+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
+```
+
+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. |

.buildkite/pipeline.yml

@@ -9,11 +9,183 @@ notify:
   - github_commit_status:
       context: "full-suite-passed"
     if: build.env("TEST_SCOPE") == "full"
+  - github_commit_status:
+      context: "direct-test-completed"
+    if: build.env("TEST_SCOPE") == "direct"
 
 steps:
   # ============================================================
-    - label: ":dart: Direct Test (${TEST_TYPE})"
-      if: build.env("TEST_SCOPE") == "direct"
+  # Direct test: triggered by /test <name> slash command.
+  # Labels match fastcheck/full-suite counterparts so the GitHub
+  # check status overwrites the original failed check.
+  # Only ONE step executes per build (gated by TEST_TYPE).
+  # ============================================================
+
+  # --- Fastcheck-scope direct tests ---
+    - label: ":microscope: Encoder Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "encoder"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":microscope: VAE Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "vae"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":microscope: Transformer Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "transformer"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":microscope: Kernel Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "kernel_tests"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":microscope: Unit Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "unit_test"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+
+  # --- Full-suite-scope direct tests ---
+    - label: ":bar_chart: SSIM Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "ssim"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+          - exit_status: 1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: LoRA Inference Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "inference_lora"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Training Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "training"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Distillation DMD Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "distillation_dmd"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Self-Forcing Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "self_forcing"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: LoRA Training Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "training_lora"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+          - exit_status: 1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Training Tests VSA"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "training_vsa"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+          - exit_status: 1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Inference Tests VMoBA"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "inference_vmoba"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: Performance Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "performance"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
+    - label: ":test_tube: API Server Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "api_server"
       command: "timeout 90m .buildkite/scripts/pr_test.sh"
       retry:
         automatic:
@@ -135,6 +307,10 @@ steps:
                 label: ":bar_chart: SSIM Tests"
                 env:
                   - TEST_TYPE=ssim
+                retry:
+                  automatic:
+                    - exit_status: 1
+                      limit: 2
                 agents:
                   queue: "default"
             - path:
@@ -195,6 +371,10 @@ steps:
                 label: ":test_tube: LoRA Training Tests"
                 env:
                   - TEST_TYPE=training_lora
+                retry:
+                  automatic:
+                    - exit_status: 1
+                      limit: 2
                 agents:
                   queue: "default"
             - path:
@@ -207,6 +387,10 @@ steps:
                 label: ":test_tube: Training Tests VSA"
                 env:
                   - TEST_TYPE=training_vsa
+                retry:
+                  automatic:
+                    - exit_status: 1
+                      limit: 2
                 agents:
                   queue: "default"
             - path:

.github/mergify.yml

@@ -4,8 +4,10 @@ merge_protections:
       - base = main
     success_conditions:
       - "title~=(?i)^\\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model)\\]"
+      - "#approved-reviews-by>=1"
       - check-success~=pre-commit
       - check-success=fastcheck-passed
+      - check-success=full-suite-passed
 
 pull_request_rules:
 
@@ -103,7 +105,7 @@ pull_request_rules:
           - files~=^fastvideo/pipelines/samplers/
           - files~=^fastvideo/entrypoints/
           - files~=^fastvideo/worker/
-          - files~=^fastvideo/configs/sample/
+          - files~=^fastvideo/api/sampling_param
           - files~=^fastvideo/configs/pipelines/
           - files~=^examples/inference/
       - -closed
@@ -272,24 +274,15 @@ pull_request_rules:
       merge:
         method: squash
 
-  - name: auto-rebase when ready and Full Suite passed
+  - name: auto-update when ready
     conditions:
       - label=ready
       - "#approved-reviews-by>=1"
-      - check-success=full-suite-passed
       - -conflict
       - -closed
       - -draft
     actions:
-      rebase: {}
-
-  - name: remove ready label on Full Suite failure
-    conditions:
-      - label=ready
-      - check-failure=full-suite-passed
-    actions:
-      label:
-        remove: [ready]
+      update: {}
 
   # ============================================================
   # PR title format help
@@ -319,3 +312,5 @@ pull_request_rules:
 
           Please update your PR title and the merge protection check will pass automatically.
 
+merge_protections_settings:
+  reporting_method: check-runs

.github/workflows/ci-aggregate-status.yml

@@ -0,0 +1,80 @@
+name: Aggregate Test Status
+
+on:
+  status:
+
+permissions:
+  statuses: write
+
+jobs:
+  aggregate:
+    if: >-
+      github.event.context == 'direct-test-completed'
+      && github.event.state == 'success'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check and update aggregate status
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        with:
+          script: |
+            const sha = context.payload.sha;
+
+            const { data } = await github.rest.repos.getCombinedStatusForRef({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              ref: sha,
+              per_page: 100,
+            });
+
+            const bkStatuses = data.statuses.filter(
+              s => s.context.startsWith('buildkite/ci/')
+            );
+
+            const FASTCHECK_PREFIX = 'buildkite/ci/microscope-';
+            const FULL_SUITE_PREFIXES = [
+              'buildkite/ci/test-tube-',
+              'buildkite/ci/bar-chart-',
+            ];
+
+            const fastcheck = bkStatuses.filter(
+              s => s.context.startsWith(FASTCHECK_PREFIX)
+            );
+            const fullSuite = bkStatuses.filter(
+              s => FULL_SUITE_PREFIXES.some(p => s.context.startsWith(p))
+            );
+
+            if (
+              fastcheck.length > 0
+              && fastcheck.every(s => s.state === 'success')
+            ) {
+              core.info(
+                `All ${fastcheck.length} fastcheck tests passed — updating fastcheck-passed`
+              );
+              await github.rest.repos.createCommitStatus({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                sha,
+                state: 'success',
+                context: 'fastcheck-passed',
+                description:
+                  `All ${fastcheck.length} fastcheck tests passed`,
+              });
+            }
+
+            if (
+              fullSuite.length > 0
+              && fullSuite.every(s => s.state === 'success')
+            ) {
+              core.info(
+                `All ${fullSuite.length} full suite tests passed — updating full-suite-passed`
+              );
+              await github.rest.repos.createCommitStatus({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                sha,
+                state: 'success',
+                context: 'full-suite-passed',
+                description:
+                  `All ${fullSuite.length} full suite tests passed`,
+              });
+            }

.github/workflows/ci-precommit.yml

@@ -4,20 +4,23 @@ on:
   pull_request:
     branches: [main]
   workflow_call:
-
-concurrency:
-  group: pre-commit-${{ github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+    inputs:
+      ref:
+        description: 'Git ref to checkout (defaults to github.ref)'
+        required: false
+        type: string
 
 permissions:
   contents: read
 
 jobs:
   pre-commit:
-    if: github.event.pull_request.draft != true
+    if: github.event_name == 'workflow_call' || github.event.pull_request.draft != true
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
+      with:
+        ref: ${{ inputs.ref || '' }}
     - uses: actions/setup-python@v5
       with:
         python-version: "3.12"

.github/workflows/ci-slash-commands.yml

@@ -7,6 +7,7 @@ on:
 permissions:
   contents: read
   pull-requests: write
+  statuses: write
 
 jobs:
   handle-merge:
@@ -32,6 +33,7 @@ jobs:
             core.setOutput('has_write', String(hasWrite));
 
       - name: Add ready label and react
+        id: label
         if: steps.perm.outputs.has_write == 'true'
         uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
         with:
@@ -39,7 +41,6 @@ jobs:
             const owner = context.repo.owner;
             const repo = context.repo.repo;
             const prNumber = context.payload.issue.number;
-            // Remove ready first to allow re-trigger (labeled event fires on add, not if already present)
             try { await github.rest.issues.removeLabel({ owner, repo, issue_number: prNumber, name: 'ready' }); } catch {}
             await github.rest.issues.addLabels({ owner, repo, issue_number: prNumber, labels: ['ready'] });
             await github.rest.reactions.createForIssueComment({
@@ -47,6 +48,44 @@ jobs:
               comment_id: context.payload.comment.id,
               content: 'rocket',
             });
+            const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number: prNumber });
+            core.setOutput('pr_sha', pr.head.sha);
+            core.setOutput('pr_branch', pr.head.ref);
+            core.setOutput('pr_number', String(prNumber));
+
+      - name: Trigger Full Suite
+        if: steps.perm.outputs.has_write == 'true'
+        env:
+          BUILDKITE_API_TOKEN: ${{ secrets.BUILDKITE_API_TOKEN }}
+          PR_SHA: ${{ steps.label.outputs.pr_sha }}
+          PR_BRANCH: ${{ steps.label.outputs.pr_branch }}
+          PR_NUMBER: ${{ steps.label.outputs.pr_number }}
+          BK_ORG: ${{ vars.BUILDKITE_ORG_SLUG }}
+          BK_PIPELINE: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
+        run: |
+          curl -sS --fail-with-body -X POST \
+            "https://api.buildkite.com/v2/organizations/${BK_ORG}/pipelines/${BK_PIPELINE}/builds" \
+            -H "Authorization: Bearer $BUILDKITE_API_TOKEN" \
+            -H "Content-Type: application/json" \
+            --data-raw "$(jq -n \
+              --arg commit "$PR_SHA" \
+              --arg branch "$PR_BRANCH" \
+              --arg message "Full Suite for PR #${PR_NUMBER} (via /merge)" \
+              --argjson pr_id "$PR_NUMBER" \
+              '{
+                commit: $commit,
+                branch: $branch,
+                message: $message,
+                ignore_pipeline_branch_filters: true,
+                pull_request_id: $pr_id,
+                pull_request_base_branch: "main",
+                env: {
+                  TEST_SCOPE: "full",
+                  FULL_SUITE: "true",
+                  PR_NUMBER: ($pr_id | tostring)
+                }
+              }')"
+
   parse-command:
     if: >-
       github.event.issue.pull_request != null
@@ -86,7 +125,7 @@ jobs:
           set -euo pipefail
           TEST_NAME=$(echo "$COMMENT" | grep -oP '(?<=/test\s)\S+' | head -1 || true)
 
-          VALID="encoder vae transformer kernel unit ssim training lora-inference lora-training distillation self-forcing vsa vmoba performance api full fastcheck"
+          VALID="encoder vae transformer kernel unit ssim training lora-inference lora-training distillation self-forcing vsa vmoba performance api full fastcheck pre-commit"
           if [ -z "$TEST_NAME" ] || ! echo "$VALID" | grep -qw "$TEST_NAME"; then
             echo "Unknown test: '$TEST_NAME'. Valid: $VALID"
             exit 1
@@ -114,6 +153,12 @@ jobs:
               echo "test_scope=fastcheck"
               echo "full_suite=false"
             } >> "$GITHUB_OUTPUT"
+          elif [ "$TEST_NAME" = "pre-commit" ]; then
+            {
+              echo "test_type="
+              echo "test_scope=precommit"
+              echo "full_suite=false"
+            } >> "$GITHUB_OUTPUT"
           else
             {
               echo "test_type=${MAP[$TEST_NAME]}"
@@ -136,14 +181,8 @@ jobs:
             core.setOutput('sha', pr.head.sha);
             core.setOutput('branch', pr.head.ref);
 
-  trigger-buildkite:
-    needs: parse-command
-    if: >-
-      needs.parse-command.outputs.has_write == 'true'
-      && needs.parse-command.outputs.test_type != ''
-    runs-on: ubuntu-latest
-    steps:
       - name: React to comment
+        if: steps.perm.outputs.has_write == 'true'
         uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
         with:
           script: |
@@ -154,6 +193,43 @@ jobs:
               content: 'rocket',
             });
 
+  pre-commit:
+    needs: parse-command
+    if: >-
+      needs.parse-command.outputs.has_write == 'true'
+      && needs.parse-command.outputs.test_scope == 'precommit'
+    uses: ./.github/workflows/ci-precommit.yml
+    with:
+      ref: refs/pull/${{ github.event.issue.number }}/merge
+
+  post-precommit-status:
+    needs: [parse-command, pre-commit]
+    if: always() && needs.parse-command.outputs.test_scope == 'precommit'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        env:
+          PR_SHA: ${{ needs.parse-command.outputs.pr_sha }}
+          RESULT: ${{ needs.pre-commit.result }}
+        with:
+          script: |
+            const state = process.env.RESULT === 'success' ? 'success' : 'failure';
+            await github.rest.repos.createCommitStatus({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              sha: process.env.PR_SHA,
+              state,
+              context: 'pre-commit',
+              description: `Triggered via /test pre-commit (${state})`,
+            });
+
+  trigger-buildkite:
+    needs: parse-command
+    if: >-
+      needs.parse-command.outputs.has_write == 'true'
+      && needs.parse-command.outputs.test_type != ''
+    runs-on: ubuntu-latest
+    steps:
       - name: Trigger Buildkite
         env:
           BUILDKITE_API_TOKEN: ${{ secrets.BUILDKITE_API_TOKEN }}

.github/workflows/ci-trigger-full-suite.yml

@@ -1,7 +1,7 @@
 name: Trigger Full Suite
 
 on:
-  pull_request:
+  pull_request_target:
     types: [labeled, synchronize]
 
 permissions:
@@ -10,7 +10,7 @@ permissions:
 
 concurrency:
   group: full-suite-${{ github.event.pull_request.number }}
-  cancel-in-progress: true
+  cancel-in-progress: false
 
 jobs:
   trigger:
@@ -42,7 +42,7 @@ jobs:
           # Find running builds for this branch with TEST_SCOPE=full and cancel them
           builds=$(curl -sS -H "Authorization: Bearer $BUILDKITE_API_TOKEN" \
             "https://api.buildkite.com/v2/organizations/${{ vars.BUILDKITE_ORG_SLUG }}/pipelines/${{ vars.BUILDKITE_PIPELINE_SLUG }}/builds?branch=${PR_BRANCH}&state=running,scheduled" \
-            | jq -r '.[] | select(.env.TEST_SCOPE == "full") | .number')
+            | jq -r '.[] | select(try (.env.TEST_SCOPE == "full") catch false) | .number')
           for build_num in $builds; do
             echo "Cancelling Buildkite build #$build_num"
             curl -sS -X PUT -H "Authorization: Bearer $BUILDKITE_API_TOKEN" \

.gitignore

@@ -85,6 +85,7 @@ docs/distillation/examples/
 dmd_t2v_output/
 preprocess_output_text/
 
+# Next.js / Node artifacts under ui/: see ui/.gitignore
 
 .claude/
 .codex/

.nvimlog

@@ -0,0 +1 @@
+WRN 2026-03-26T13:46:33.469 ?.19646    server_start:193: Failed to start server: operation not permitted: /var/folders/z_/h_6myyk14d1b7z87z3vy4mjh0000gn/T/nvim.dsynkd/iSe0el/nvim.19646.0

.python-version

@@ -0,0 +1 @@
+3.12

docs/api/fastvideo.md

@@ -62,9 +62,9 @@ This page contains the complete API reference for the FastVideo library.
       show_root_toc_entry: true
       heading_level: 4
 
-#### fastvideo.configs.sample
+#### fastvideo.api.sampling_param
 
-::: fastvideo.configs.sample
+::: fastvideo.api.sampling_param
     options:
       show_source: true
       show_root_heading: true

docs/contributing/ci_architecture.md

@@ -24,7 +24,7 @@ PR push
           Runs on the PR branch directly
               │
           pass ──► Mergify auto-squash-merges to main, branch deleted
-          fail ──► Mergify removes 'ready' label; fix and /merge again
+          fail ──► fix the regression, push, and /merge again
 ```
 
 ---
@@ -102,8 +102,8 @@ failing test's output.
 | Performance Tests | `performance` | 30 min |
 | API Server Tests | `api_server` | 30 min |
 
-A Full Suite failure removes the `ready` label automatically. A Mergify comment links to
-the Buildkite build. Fix the regression, push, and comment `/merge` again.
+If a Full Suite test fails, check the Buildkite build log for the failing step's output.
+Fix the regression, push, and comment `/merge` again to re-trigger.
 
 ---
 
@@ -129,8 +129,8 @@ Suite passing directly on the PR branch.
    - No merge conflicts
 5. If all conditions pass, Mergify squash-merges to `main` automatically. The branch is
    deleted after merge.
-6. If the Full Suite fails, Mergify removes the `ready` label and posts a comment linking to
-   the Buildkite build. The developer fixes the issue, pushes, and comments `/merge` again.
+6. If the Full Suite fails, the developer fixes the issue, pushes, and comments `/merge`
+   again to re-trigger.
 
 **Merge conditions summary:**
 
@@ -173,7 +173,7 @@ Applied by Mergify based on which paths you modified. Multiple scope labels can
 | Label | File paths that trigger it |
 |-------|---------------------------|
 | `scope: training` | `fastvideo/train/`, `fastvideo/training/`, `fastvideo/distillation/`, `examples/train/`, `examples/training/`, `examples/distill/` |
-| `scope: inference` | `fastvideo/pipelines/basic/`, `fastvideo/pipelines/stages/`, `fastvideo/pipelines/samplers/`, `fastvideo/entrypoints/`, `fastvideo/worker/`, `fastvideo/configs/sample/`, `fastvideo/configs/pipelines/`, `examples/inference/` |
+| `scope: inference` | `fastvideo/pipelines/basic/`, `fastvideo/pipelines/stages/`, `fastvideo/pipelines/samplers/`, `fastvideo/entrypoints/`, `fastvideo/worker/`, `fastvideo/api/sampling_param.py`, `fastvideo/configs/pipelines/`, `examples/inference/` |
 | `scope: attention` | `fastvideo/attention/` |
 | `scope: kernel` | `fastvideo-kernel/`, `csrc/` |
 | `scope: data` | `fastvideo/dataset/`, `fastvideo/pipelines/preprocess/`, `examples/preprocessing/` |
@@ -279,6 +279,30 @@ Triggers a specific Buildkite test or suite on the current PR branch.
 | `/test api` | API server integration tests | `api_server` |
 | `/test full` | Entire Full Suite | all (with `TEST_SCOPE=full`) |
 | `/test fastcheck` | Entire Fastcheck suite | fastcheck (with `TEST_SCOPE=fastcheck`) |
+| `/test pre-commit` | Pre-commit checks on PR code | — (runs `ci-precommit.yml` via `workflow_call`) |
+
+**Re-running failed tests:** When you use `/test <name>` to re-run a specific failed test,
+the resulting Buildkite check uses the same name as the original (e.g., `/test encoder`
+creates `buildkite/ci/microscope-encoder-tests`). This overwrites the failed check status.
+Once all tests in a tier pass, the aggregate status (`fastcheck-passed` or
+`full-suite-passed`) is automatically updated to `success` by the `ci-aggregate-status.yml`
+workflow.
+
+**How aggregate status refresh works:**
+
+1. `/test <name>` triggers a Buildkite build with `TEST_SCOPE=direct`. The test step uses
+   the same label as its fastcheck/full-suite counterpart, so the resulting GitHub check
+   overwrites the original.
+2. When the build completes, Buildkite's `notify` posts a `direct-test-completed` commit
+   status. This is the only signal that triggers the aggregate workflow — intermediate step
+   status updates do not trigger it.
+3. `ci-aggregate-status.yml` fires, calls `getCombinedStatusForRef` to fetch the latest
+   status for every context on that commit (each context returns only its most recent
+   state), groups them by prefix (`microscope-*` → fastcheck, `test-tube-*`/`bar-chart-*`
+   → full suite), and posts `fastcheck-passed: success` or `full-suite-passed: success` if
+   all entries in the group are `success`.
+4. Tests that were never triggered (skipped by monorepo-diff) have no status entry and do
+   not block the aggregate.
 
 ---
 
@@ -296,6 +320,7 @@ Protected branches (`main`, `master`, `release/*`) are never deleted.
 | `ci-precommit.yml` | Every push / PR against `main` | Runs pre-commit hooks (yapf, ruff, mypy, codespell, pymarkdown, actionlint, check-filenames) |
 | `ci-trigger-full-suite.yml` | `ready` label added to a PR | Calls Buildkite API to run Full Suite on the PR branch |
 | `ci-slash-commands.yml` | PR comment starting with `/merge` or `/test` | Handles slash commands; adds `ready` label or triggers Buildkite |
+| `ci-aggregate-status.yml` | Any Buildkite commit status update | Checks if all tests in a tier passed; updates `fastcheck-passed` or `full-suite-passed` |
 | `community-issue-labeler.yml` | Issue opened or edited | Auto-labels issues by keyword matching against title and body |
 | `community-welcome.yml` | First contribution | Posts a welcome comment for first-time contributors |
 | `community-stale.yml` | Scheduled | Marks and closes stale issues and PRs |

docs/contributing/coding_agents.md

@@ -44,7 +44,7 @@ FastVideo maps a Diffusers-style repo into a pipeline like:
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
   weight name translation.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
-- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipeline logic built from stages.
 - `model_index.json`: the HF repo entrypoint that maps component names to
   classes and weight files.
@@ -55,7 +55,7 @@ Minimal usage example (based on `examples/inference/basic/basic.py`):
 
 ```python
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
 generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
@@ -319,7 +319,8 @@ Purpose:
 
 - `fastvideo/configs/pipelines/` describes pipeline wiring and model module
   names.
-- `fastvideo/configs/sample/` defines default runtime parameters.
+- `fastvideo/api/sampling_param.py` defines runtime sampling parameters.
+  Defaults come from profiles in `fastvideo/pipelines/basic/<family>/profiles.py`.
 
 Action:
 
@@ -474,7 +475,7 @@ FastVideo integration.
 3. Pipeline wiring.
    - Pipeline: `fastvideo/pipelines/basic/wan/wan_pipeline.py`
    - Pipeline config: `fastvideo/configs/pipelines/wan.py`
-   - Sampling defaults: `fastvideo/configs/sample/wan.py`
+   - Sampling defaults: `fastvideo/pipelines/basic/wan/profiles.py`
 
 4. Minimal example.
    - Script: `examples/inference/basic/basic.py`

docs/contributing/pull_requests.md

@@ -104,8 +104,9 @@ distillation, self-forcing, VSA, VMoBA, performance benchmarks, and API server t
 8. If all Full Suite tests pass and all merge conditions are met (approval, valid title,
    pre-commit green, fastcheck green, no draft, no conflicts), Mergify squash-merges to
    `main` automatically. Your branch is deleted.
-9. If a Full Suite test fails, Mergify removes the `ready` label and posts a comment with a
-   link to the Buildkite build. Fix the issue, push, and comment `/merge` again.
+9. If a Full Suite test fails, check the Buildkite build log for the failing step. Fix the
+   issue, push, and comment `/merge` again. You can also re-run individual failed tests
+   with `/test <name>` — see below.
 
 !!! note
     Only contributors with write permission to the repository can trigger slash commands.
@@ -149,10 +150,15 @@ Comment on your PR to trigger specific tests independently of the auto-merge flo
 /test vmoba            # VMoBA inference tests
 /test performance      # Performance benchmarks
 /test api              # API server integration tests
+/test pre-commit       # Pre-commit checks on PR code
 ```
 
 The workflow reacts with a 🚀 emoji to confirm the command was received.
 
+When you re-run an individual test with `/test <name>`, the new result overwrites the
+original failed check (same Buildkite check name). Once all tests in a tier pass, the
+`fastcheck-passed` or `full-suite-passed` status is automatically updated.
+
 ---
 
 ## Troubleshooting
@@ -199,9 +205,8 @@ Mergify removes the `needs-rebase` label automatically once conflicts are resolv
 
 ### Full Suite failed after `/merge`
 
-The Full Suite found a regression. Mergify removes the `ready` label and posts a comment
-linking to the Buildkite build. Check the failing step's output for assertion errors or
-tracebacks.
+The Full Suite found a regression. Check the failing Buildkite step's output for assertion
+errors or tracebacks.
 
 Common causes:
 

docs/design/inference_schema_parity_inventory.yaml

@@ -0,0 +1,447 @@
+status_definitions:
+  kept: "Public field remains on a public adapter surface with the same meaning."
+  moved: "Public field remains supported but normalizes into a different nested path."
+  preset_owned: "Public field remains supported only through a model/preset-specific surface."
+  compatibility_only: "Legacy public field remains adapter-only during migration and is not part of the canonical typed schema."
+  private_only: "Field should only be handled by private adapters and is not a public FastVideo compatibility promise."
+  internal_only: "Field is runtime/config plumbing and should not be part of the new public typed inference API."
+
+surfaces:
+  fastvideo_args:
+    moved:
+      model_path: generator.model_path
+      workload_type: generator.pipeline.workload_type
+      distributed_executor_backend: generator.engine.execution_backend
+      trust_remote_code: generator.trust_remote_code
+      revision: generator.revision
+      num_gpus: generator.engine.num_gpus
+      tp_size: generator.engine.parallelism.tp_size
+      sp_size: generator.engine.parallelism.sp_size
+      hsdp_replicate_dim: generator.engine.parallelism.hsdp_replicate_dim
+      hsdp_shard_dim: generator.engine.parallelism.hsdp_shard_dim
+      dist_timeout: generator.engine.parallelism.dist_timeout
+      lora_path: generator.pipeline.components.lora_path
+      dit_cpu_offload: generator.engine.offload.dit
+      use_fsdp_inference: generator.engine.use_fsdp_inference
+      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: generator.engine.compile.backend,fullgraph,mode,dynamic,extras
+      disable_autocast: generator.engine.disable_autocast
+      enable_stage_verification: generator.engine.enable_stage_verification
+      prompt_txt: request.inputs.prompt_path
+      override_text_encoder_safetensors: generator.pipeline.components.text_encoder_weights
+      override_text_encoder_quant: generator.engine.quantization.text_encoder_quant
+      override_transformer_cls_name: generator.pipeline.components.override_transformer_cls_name
+      init_weights_from_safetensors: generator.pipeline.components.transformer_weights
+      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
+    preset_owned:
+      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
+    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."
+      lora_nickname: "Legacy adapter-selection surface pending LoRA API cleanup."
+      lora_target_modules: "Legacy LoRA configuration surface pending dedicated component API."
+      output_type: "Legacy output formatting surface pending GenerationResult cleanup."
+      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."
+    private_only:
+      ray_placement_group: "Ray deployment-only field."
+      ray_runtime_env: "Ray deployment-only field."
+    internal_only:
+      pipeline_config: "Legacy internal carrier object."
+      preprocess_config: "Legacy preprocess carrier object."
+      moba_config: "Derived runtime config loaded from moba_config_path."
+      model_paths: "Runtime bookkeeping."
+      model_loaded: "Runtime bookkeeping."
+
+  pipeline_config_base:
+    moved:
+      pipeline_config_path: generator.pipeline.components.pipeline_config_path
+    preset_owned:
+      embedded_cfg_scale: generator.pipeline.preset_overrides.embedded_cfg_scale
+      flow_shift: generator.pipeline.preset_overrides.flow_shift
+      flow_shift_sr: generator.pipeline.preset_overrides.flow_shift_sr
+      is_causal: generator.pipeline.preset_overrides.is_causal
+      vae_tiling: generator.pipeline.preset_overrides.vae_tiling
+      vae_sp: generator.pipeline.preset_overrides.vae_sp
+      dmd_denoising_steps: generator.pipeline.preset_overrides.dmd_denoising_steps
+      ti2v_task: generator.pipeline.preset_overrides.ti2v_task
+      boundary_ratio: generator.pipeline.preset_overrides.boundary_ratio
+    compatibility_only:
+      model_path: "Redundant with generator.model_path."
+      disable_autocast: "Duplicated by generator.engine.disable_autocast during migration."
+      dit_precision: "Precision override pending dedicated typed component precision design."
+      upsampler_precision: "Precision override pending dedicated typed component precision design."
+      vae_precision: "Precision override pending dedicated typed component precision design."
+      image_encoder_precision: "Precision override pending dedicated typed component precision design."
+      text_encoder_precisions: "Precision override pending dedicated typed component precision design."
+    internal_only:
+      dit_config: "Legacy internal component config object."
+      upsampler_config: "Legacy internal component config object."
+      vae_config: "Legacy internal component config object."
+      image_encoder_config: "Legacy internal component config object."
+      text_encoder_configs: "Legacy internal component config object."
+      preprocess_text_funcs: "Internal text preprocessing hooks."
+      postprocess_text_funcs: "Internal text postprocessing hooks."
+
+  pipeline_config_extensions:
+    preset_owned:
+      conditioning_strategy:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      max_num_conditional_frames:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      min_num_conditional_frames:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      sigma_conditional:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      sigma_data:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      state_ch:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      state_t:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      text_encoder_class:
+        sources:
+          - fastvideo.configs.pipelines.cosmos.CosmosConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      autoregressive_chunk_frames:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      autoregressive_overlap_frames:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      cfg_behavior:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      default_camera_rotation:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      default_movement_distance:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      default_negative_prompt:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      default_trajectory_type:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      filter_points_threshold:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      fps:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      frame_buffer_max:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      moge_model_name:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      noise_aug_strength:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      num_frames:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      offload_moge_after_depth:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      use_moge_depth:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      video_resolution:
+        sources:
+          - fastvideo.configs.pipelines.gen3c.Gen3CConfig
+          - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      text_encoder_crop_start:
+        sources:
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15I2V480PStepDistilledConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15I2V720PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15SR1080PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15T2V480PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15T2V720PConfig
+          - fastvideo.configs.pipelines.hyworld.HYWorldConfig
+          - fastvideo.configs.pipelines.hyworld.Hunyuan15T2V480PConfig
+      text_encoder_max_lengths:
+        sources:
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15I2V480PStepDistilledConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15I2V720PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15SR1080PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15T2V480PConfig
+          - fastvideo.configs.pipelines.hunyuan15.Hunyuan15T2V720PConfig
+          - fastvideo.configs.pipelines.hyworld.HYWorldConfig
+          - fastvideo.configs.pipelines.hyworld.Hunyuan15T2V480PConfig
+      precision:
+        sources:
+          - fastvideo.configs.pipelines.lingbotworld.LingBotWorldI2V480PConfig
+          - fastvideo.configs.pipelines.lingbotworld.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_1_3B_Config
+          - fastvideo.configs.pipelines.wan.FastWan2_1_T2V_480P_Config
+          - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
+          - fastvideo.configs.pipelines.wan.MatrixGameBaseI2V480PConfig
+          - fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig
+          - fastvideo.configs.pipelines.wan.SelfForcingWan2_2_T2V480PConfig
+          - fastvideo.configs.pipelines.wan.SelfForcingWanT2V480PConfig
+          - fastvideo.configs.pipelines.wan.WANV2VConfig
+          - fastvideo.configs.pipelines.wan.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.wan.Wan2_2_T2V_A14B_Config
+          - fastvideo.configs.pipelines.wan.Wan2_2_TI2V_5B_Config
+          - fastvideo.configs.pipelines.wan.WanI2V480PConfig
+          - fastvideo.configs.pipelines.wan.WanI2V720PConfig
+          - fastvideo.configs.pipelines.wan.WanT2V480PConfig
+          - fastvideo.configs.pipelines.wan.WanT2V720PConfig
+      warp_denoising_step:
+        sources:
+          - fastvideo.configs.pipelines.lingbotworld.LingBotWorldI2V480PConfig
+          - fastvideo.configs.pipelines.lingbotworld.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_1_3B_Config
+          - fastvideo.configs.pipelines.wan.FastWan2_1_T2V_480P_Config
+          - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
+          - fastvideo.configs.pipelines.wan.MatrixGameBaseI2V480PConfig
+          - fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig
+          - fastvideo.configs.pipelines.wan.SelfForcingWan2_2_T2V480PConfig
+          - fastvideo.configs.pipelines.wan.SelfForcingWanT2V480PConfig
+          - fastvideo.configs.pipelines.wan.WANV2VConfig
+          - fastvideo.configs.pipelines.wan.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.wan.Wan2_2_T2V_A14B_Config
+          - fastvideo.configs.pipelines.wan.Wan2_2_TI2V_5B_Config
+          - fastvideo.configs.pipelines.wan.WanI2V480PConfig
+          - fastvideo.configs.pipelines.wan.WanI2V720PConfig
+          - fastvideo.configs.pipelines.wan.WanT2V480PConfig
+          - fastvideo.configs.pipelines.wan.WanT2V720PConfig
+      bsa_cdf_threshold:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_chunk_k:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_chunk_q:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_params:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_sparsity:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enable_bsa:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enable_kv_cache:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enhance_hf:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      offload_kv_cache:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      t_thresh:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      use_distill:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      scheduler_arch:
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
+      text_encoder_archs:
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
+      tokenizer_archs:
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
+      transformer_arch:
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
+      vae_arch:
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
+      expand_timesteps:
+        sources:
+          - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
+          - fastvideo.configs.pipelines.wan.Wan2_2_TI2V_5B_Config
+      context_noise:
+        sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
+      num_frames_per_block:
+        sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
+    compatibility_only:
+      batch_size: "Gen3C inference-only tuning field pending typed batching design."
+      gradient_checkpointing: "Gen3C inference-only compatibility field pending typed batching design."
+      guidance_scale: "Gen3C pipeline-level default pending preset/default-request cleanup."
+      num_inference_steps: "Gen3C pipeline-level default pending preset/default-request cleanup."
+    internal_only:
+      audio_decoder_config: "Legacy internal component config object."
+      audio_decoder_precision: "Precision override pending dedicated component precision design."
+      vocoder_config: "Legacy internal component config object."
+      vocoder_precision: "Precision override pending dedicated component precision design."
+
+  sampling_param_base:
+    moved:
+      image_path: request.inputs.image_path
+      pil_image: request.inputs.pil_image
+      video_path: request.inputs.video_path
+      mouse_cond: request.inputs.mouse_cond
+      keyboard_cond: request.inputs.keyboard_cond
+      grid_sizes: request.inputs.grid_sizes
+      pose: request.inputs.pose
+      c2ws_plucker_emb: request.inputs.c2ws_plucker_emb
+      refine_from: request.inputs.refine_from
+      stage1_video: request.inputs.stage1_video
+      prompt: request.prompt
+      negative_prompt: request.negative_prompt
+      prompt_path: request.inputs.prompt_path
+      output_path: request.output.output_path
+      output_video_name: request.output.output_video_name
+      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
+      height_sr: request.sampling.height_sr
+      width_sr: request.sampling.width_sr
+      fps: request.sampling.fps
+      num_inference_steps: request.sampling.num_inference_steps
+      num_inference_steps_sr: request.sampling.num_inference_steps_sr
+      guidance_scale: request.sampling.guidance_scale
+      guidance_scale_2: request.sampling.guidance_scale_2
+      guidance_rescale: request.sampling.guidance_rescale
+      boundary_ratio: request.sampling.boundary_ratio
+      sigmas: request.sampling.sigmas
+      enable_teacache: request.runtime.enable_teacache
+      save_video: request.output.save_video
+      return_frames: request.output.return_frames
+      return_trajectory_latents: request.runtime.return_trajectory_latents
+      return_trajectory_decoded: request.runtime.return_trajectory_decoded
+    preset_owned:
+      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
+      trajectory_type: request.extensions.gen3c.trajectory_type
+      movement_distance: request.extensions.gen3c.movement_distance
+      camera_rotation: request.extensions.gen3c.camera_rotation
+      prompt_attention_mask: request.extensions.hyworld.prompt_attention_mask
+      negative_attention_mask: request.extensions.hyworld.negative_attention_mask
+      camera_states: request.extensions.hunyuangamecraft.camera_states
+      camera_trajectory: request.extensions.hunyuangamecraft.camera_trajectory
+      action_list: request.extensions.hunyuangamecraft.action_list
+      action_speed_list: request.extensions.hunyuangamecraft.action_speed_list
+      gt_latents: request.extensions.hunyuangamecraft.gt_latents
+      conditioning_mask: request.extensions.hunyuangamecraft.conditioning_mask
+      ltx2_cfg_scale_video: request.extensions.ltx2.cfg_scale_video
+      ltx2_cfg_scale_audio: request.extensions.ltx2.cfg_scale_audio
+      ltx2_modality_scale_video: request.extensions.ltx2.modality_scale_video
+      ltx2_modality_scale_audio: request.extensions.ltx2.modality_scale_audio
+      ltx2_rescale_scale: request.extensions.ltx2.rescale_scale
+      ltx2_stg_scale_video: request.extensions.ltx2.stg_scale_video
+      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
+    internal_only:
+      data_type: "Derived from the request shape and not a public input."
+
+  sampling_param_extensions: {}
+
+  openai_image_request:
+    kept:
+      model: "HTTP adapter model-routing field."
+      response_format: "HTTP adapter response formatting field."
+      output_format: "HTTP adapter output-format field."
+      background: "HTTP adapter output-format field."
+      quality: "Compatibility field currently accepted by the adapter."
+      style: "Compatibility field currently accepted by the adapter."
+      user: "Compatibility field currently accepted by the adapter."
+    moved:
+      prompt: request.prompt
+      n: request.sampling.num_videos_per_prompt
+      size:
+        target: request.sampling.width,height
+        note: "Adapter parses OpenAI size strings as WIDTHxHEIGHT and forwards width then height."
+      num_inference_steps: request.sampling.num_inference_steps
+      guidance_scale: request.sampling.guidance_scale
+      true_cfg_scale: request.sampling.true_cfg_scale
+      seed: request.sampling.seed
+      negative_prompt: request.negative_prompt
+      enable_teacache: request.runtime.enable_teacache
+
+  openai_video_request:
+    kept:
+      model: "HTTP adapter model-routing field."
+    moved:
+      prompt: request.prompt
+      input_reference: request.inputs.image_path
+      reference_url: request.inputs.image_path
+      size:
+        target: request.sampling.width,height
+        note: "Adapter parses OpenAI size strings as WIDTHxHEIGHT and forwards width then height."
+      fps: request.sampling.fps
+      num_frames: request.sampling.num_frames
+      seed: request.sampling.seed
+      num_inference_steps: request.sampling.num_inference_steps
+      guidance_scale: request.sampling.guidance_scale
+      guidance_scale_2: request.sampling.guidance_scale_2
+      true_cfg_scale: request.sampling.true_cfg_scale
+      negative_prompt: request.negative_prompt
+      enable_teacache: request.runtime.enable_teacache
+      output_path: request.output.output_path
+    compatibility_only:
+      seconds:
+        target: request.sampling.num_frames
+        note: "HTTP adapter duration convenience field. If num_frames is omitted, the adapter computes num_frames = fps * seconds."
+
+cli:
+  notes:
+    - "CLI parity is checked against the actual generate/serve parser dest sets."
+    - "The inventory tracks parser dest names, excluding argparse's implicit help action."
+    - "The refactored inference CLI is config-only: subcommands expose only --config, and any additional CLI input must use dotted override paths."
+  generate:
+    explicit_local_fields:
+      - config
+    expected_dests:
+      - config
+  serve:
+    explicit_local_fields:
+      - config
+    expected_dests:
+      - config

docs/design/overview.md

@@ -12,7 +12,7 @@ FastVideo maps a Diffusers-style repo into a pipeline like this:
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
   weight name translation.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
-- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipelines.
 - `fastvideo/pipelines/stages/*`: reusable pipeline stages.
 - `fastvideo/models/loader/*`: component loaders for Diffusers-style repos.
@@ -26,7 +26,7 @@ Minimal usage (from `examples/inference/basic/basic.py`):
 
 ```python
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
 generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
@@ -49,8 +49,9 @@ runtime parameters consistent:
 - `fastvideo/configs/models/`: architecture definitions, layer shapes, and
   `param_names_mapping` rules for key renaming.
 - `fastvideo/configs/pipelines/`: pipeline wiring and required components.
-- `fastvideo/configs/sample/`: default sampling parameters (steps, frames,
-  guidance scale, resolution, fps).
+- `fastvideo/api/sampling_param.py`: sampling parameters (steps, frames,
+  guidance scale, resolution, fps). Defaults come from profiles in
+  `fastvideo/pipelines/basic/<family>/profiles.py`.
 - `fastvideo/registry.py`: unified registry for pipeline config + sampling
   defaults and model metadata resolution, defined via explicit
   `register_configs(...)` blocks (no separate dict registries).
@@ -142,7 +143,7 @@ How this maps to FastVideo:
 - `T5TokenizerFast` -> loaded via HF in `fastvideo/models/loader/`
 - `UniPCMultistepScheduler` -> loaded via Diffusers scheduler utilities
 - Pipeline defaults -> `fastvideo/configs/pipelines/wan.py`
-- Sampling defaults -> `fastvideo/configs/sample/wan.py`
+- Sampling defaults -> `fastvideo/pipelines/basic/wan/profiles.py`
 
 ## Pipeline system
 

docs/distillation/dmd.md

@@ -16,7 +16,8 @@ Both models are trained on **61×448×832** resolution but support generating vi
 First install [VSA](../attention/vsa/index.md). Set `MODEL_BASE` to your own model path and run:
 
 ```bash
-bash scripts/inference/v1_inference_wan_dmd.sh
+FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN \
+  fastvideo generate --config scripts/inference/inference_wan_VSA_DMD_1_3B.yaml
 ```
 
 ## 🗂️ Dataset

docs/getting_started/v1_api.md

@@ -33,7 +33,7 @@ The following two classes `PipelineConfig` and `SamplingParam` are used to confi
 
 ### SamplingParam
 
-::: fastvideo.configs.sample.base.SamplingParam
+::: fastvideo.api.sampling_param.SamplingParam
     options:
       show_root_heading: true
       show_source: false

docs/inference/architecture.md

@@ -128,19 +128,14 @@ Concrete hierarchy: `DiTConfig` → `DiTArchConfig`, `VAEConfig` →
 - `dump_to_json()` / `load_from_json()` — JSON persistence. Callable
   fields and `arch_config` are excluded from dumps.
 
-### SamplingParam (`fastvideo/configs/sample/`)
+### SamplingParam (`fastvideo/api/sampling_param.py`)
 
 Generation parameters separate from pipeline config. Each model family
-provides defaults:
+provides defaults via a profile (see `fastvideo/pipelines/basic/<family>/profiles.py`):
 
 ```python
-@dataclass
-class WanT2V_1_3B_SamplingParam(SamplingParam):
-    height: int = 480
-    width: int = 832
-    num_frames: int = 81
-    guidance_scale: float = 3.0
-    num_inference_steps: int = 50
+sp = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+# sp.height == 480, sp.width == 832, sp.num_frames == 81, etc.
 ```
 
 ## Component Loading
@@ -430,9 +425,9 @@ User: generator.generate_video(prompt, ...)
    `fastvideo/configs/pipelines/<model>.py`. Set DiT/VAE/encoder configs,
    flow_shift, precision defaults.
 
-2. **Sampling param** — Create a `SamplingParam` subclass in
-   `fastvideo/configs/sample/<model>.py`. Set default height, width,
-   num_frames, guidance_scale, num_inference_steps.
+2. **Sampling param profile** — Create a profile in
+   `fastvideo/pipelines/basic/<model>/profiles.py` with default height,
+   width, num_frames, guidance_scale, num_inference_steps.
 
 3. **Register configs** — In `fastvideo/registry.py`, add a
    `register_configs()` call inside `_register_configs()` with
@@ -455,6 +450,6 @@ User: generator.generate_video(prompt, ...)
    `fastvideo/pipelines/stages/`, implement `forward()`, optionally
    implement `verify_input()`/`verify_output()`.
 
-7. **Verify** — Run `fastvideo generate --model-path <path> --prompt
-   "test" --num-inference-steps 2` to confirm the pipeline loads and
-   generates output.
+7. **Verify** — Run `fastvideo generate --config <config.yaml>` with a
+   minimal nested config to confirm the pipeline loads and generates
+   output.

docs/inference/cli.md

@@ -1,71 +1,29 @@
 # FastVideo CLI Inference
 
-The FastVideo CLI exposes the same core inference controls as the Python API.
+The FastVideo CLI is config-first. Inference runs are driven by a nested JSON or
+YAML config, with optional dotted-path overrides on the command line. The
+contract matches training: use an explicit subcommand plus `--config`, then add
+any dotted overrides you need.
 
 ## Basic Usage
 
-Use either:
-
-1. `--model-path` + `--prompt`
-2. `--model-path` + `--prompt-txt` (batch prompts, one line per prompt)
-3. `--config` (JSON/YAML)
-
 ```bash
-fastvideo generate --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
-  --prompt "A cat playing with a ball of yarn"
+fastvideo generate --config config.yaml
+fastvideo serve --config serve.yaml
 ```
 
-```bash
-fastvideo generate --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
-  --prompt-txt prompts.txt
-```
-
-You cannot provide both `--prompt` and `--prompt-txt` in the same run.
-
 ## View All Arguments
 
 ```bash
 fastvideo generate --help
 ```
 
-Arguments come from:
+The subcommands intentionally expose only `--config`. Any per-run CLI changes
+must use dotted override paths such as:
 
-- FastVideo runtime args (`FastVideoArgs`)
-- Sampling args (`SamplingParam`)
-- Pipeline config args (`PipelineConfig`)
-
-## Common Arguments
-
-### Parallelism
-
-- `--num-gpus`
-- `--sp-size`
-- `--tp-size`
-
-### Sampling
-
-- `--num-frames`
-- `--height` / `--width`
-- `--num-inference-steps`
-- `--guidance-scale`
-- `--seed`
-- `--negative-prompt`
-
-### Output
-
-- `--output-path`
-- `--save-video` / `--no-save-video`
-- `--return-frames`
-
-### Offloading and Performance
-
-- `--dit-layerwise-offload`
-- `--use-fsdp-inference`
-- `--text-encoder-cpu-offload`
-- `--image-encoder-cpu-offload`
-- `--vae-cpu-offload`
-- `--enable-torch-compile`
-- `--torch-compile-kwargs`
+- `--generator.engine.num_gpus 2`
+- `--request.sampling.seed 42`
+- `--server.port 9000`
 
 ## Using Config Files
 
@@ -73,50 +31,53 @@ Arguments come from:
 fastvideo generate --config config.yaml
 ```
 
-Config files can be JSON or YAML. CLI flags override config-file values.
+Config files can be JSON or YAML. Dotted CLI overrides take precedence over
+config-file values.
 
 Example `config.yaml`:
 
 ```yaml
-model_path: "FastVideo/FastHunyuan-diffusers"
-prompt: "A capybara lounging in a hammock"
-output_path: "outputs/"
-num_gpus: 2
-sp_size: 2
-tp_size: 1
-num_frames: 45
-height: 720
-width: 1280
-num_inference_steps: 6
-seed: 1024
-dit_precision: "bf16"
-vae_precision: "fp16"
-vae_tiling: true
-vae_sp: true
-enable_torch_compile: false
+generator:
+  model_path: FastVideo/FastHunyuan-diffusers
+  engine:
+    num_gpus: 2
+    parallelism:
+      sp_size: 2
+      tp_size: 1
+request:
+  prompt: A capybara lounging in a hammock
+  sampling:
+    num_frames: 45
+    height: 720
+    width: 1280
+    num_inference_steps: 6
+    seed: 1024
+  output:
+    output_path: outputs/
 ```
 
 Notes:
 
-- Use `dit_precision` / `vae_precision` (not `precision`).
-- Nested config objects are supported, for example `vae_config` and
-  `dit_config`.
+- `generator` and `request` are the top-level keys for generation configs.
+- `serve` configs use `generator`, `server`, and optional `default_request`.
+- Prompt text files belong under `request.inputs.prompt_path`.
 
 ## Examples
 
 Simple generation:
 
 ```bash
-fastvideo generate \
-  --model-path FastVideo/FastHunyuan-diffusers \
-  --prompt "A cat playing with a ball of yarn" \
-  --num-frames 45 --height 720 --width 1280 \
-  --num-inference-steps 6 --seed 1024 \
-  --output-path outputs/
+fastvideo generate --config config.yaml
 ```
 
-Config + CLI override:
+Config + dotted override:
 
 ```bash
-fastvideo generate --config config.yaml --prompt "A panda skiing at sunset"
+fastvideo generate --config config.yaml --request.prompt "A panda skiing at sunset"
+```
+
+Helper wrapper with positional config path:
+
+```bash
+bash scripts/inference/run.sh scripts/inference/inference_wan.yaml
 ```

docs/inference/configuration.md

@@ -73,32 +73,40 @@ if __name__ == '__main__':
 
 ## JSON/YAML Config Files (CLI)
 
-The CLI supports `--config` with JSON or YAML. Command-line arguments override
-config file values.
-By default, `fastvideo generate` uses `return_frames=false` unless you set
-`--return-frames` (or `return_frames: true` in config).
+The inference CLI is config-first. Use an explicit subcommand with `--config`,
+then apply optional dotted overrides on top, matching the training CLI style.
+By default, CLI generation uses `return_frames=false` unless you set
+`request.output.return_frames: true` in config or via a dotted override.
 
 ```bash
 fastvideo generate --config config.yaml
 ```
 
-Use CLI argument names as keys (underscore or hyphen is accepted). Example:
+Example nested config:
 
 ```yaml
-model_path: "FastVideo/FastHunyuan-diffusers"
-prompt: "A capybara relaxing in a hammock"
-num_gpus: 2
-sp_size: 2
-num_frames: 45
-height: 720
-width: 1280
-num_inference_steps: 6
-seed: 1024
-dit_precision: "bf16"
-vae_precision: "fp16"
-vae_tiling: true
-vae_sp: true
-enable_torch_compile: false
+generator:
+  model_path: FastVideo/FastHunyuan-diffusers
+  engine:
+    num_gpus: 2
+    parallelism:
+      sp_size: 2
+request:
+  prompt: A capybara relaxing in a hammock
+  sampling:
+    num_frames: 45
+    height: 720
+    width: 1280
+    num_inference_steps: 6
+    seed: 1024
+  output:
+    output_path: outputs/
+```
+
+Override individual values from the CLI with dotted paths:
+
+```bash
+fastvideo generate --config config.yaml --request.sampling.seed 42
 ```
 
 ## Performance Optimization

docs/inference/gen3c.md

@@ -0,0 +1,129 @@
+# GEN3C: 3D-Informed Camera-Controlled Video Generation
+
+[GEN3C](https://arxiv.org/abs/2503.03751) is NVIDIA's Cosmos-7B-based video model for camera-controlled generation from a single image. The FastVideo integration supports the GEN3C I2V workflow, including 3D cache conditioning and tokenizer-based conditioning latents.
+
+## Key Features
+
+- **Camera trajectory control**: `left/right/up/down/zoom_in/zoom_out/clockwise/counterclockwise`
+- **3D cache conditioning**: depth prediction -> point cloud cache -> forward warping -> latent conditioning
+- **Single-image to video generation**: 121-frame generation with camera motion
+- **Official raw checkpoint conversion**: `model.pt` -> Diffusers/FastVideo layout
+
+## Model Sources
+
+- Official raw checkpoint (not Diffusers): `nvidia/GEN3C-Cosmos-7B`
+- Diffusers-format checkpoint: `FastVideo/GEN3C-Cosmos-7B-Diffusers`
+
+## Prerequisites
+
+- Install MoGe:
+
+```bash
+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:
+
+```bash
+sudo apt-get update
+sudo apt-get install -y libgl1 libglib2.0-0 libsm6 libxext6 libxrender1
+```
+
+## Quick Start
+
+### Option A: Use Diffusers-format weights directly
+
+```bash
+python examples/inference/basic/basic_gen3c.py \
+  --model_path FastVideo/GEN3C-Cosmos-7B-Diffusers \
+  --image_path /path/to/input.png \
+  --prompt "" \
+  --trajectory left \
+  --movement_distance 0.3 \
+  --camera_rotation center_facing \
+  --num_inference_steps 35 \
+  --guidance_scale 1.0 \
+  --output_path outputs_video/gen3c_output.mp4
+```
+
+### Option B: Convert official raw checkpoint locally
+
+1. Download:
+
+```bash
+huggingface-cli download nvidia/GEN3C-Cosmos-7B --local-dir official_weights/GEN3C-Cosmos-7B
+```
+
+1. Convert:
+
+```bash
+python scripts/checkpoint_conversion/convert_gen3c_to_fastvideo.py \
+  --source official_weights/GEN3C-Cosmos-7B/model.pt \
+  --output converted_weights/GEN3C-Cosmos-7B
+```
+
+1. Run:
+
+```bash
+python examples/inference/basic/basic_gen3c.py \
+  --model_path converted_weights/GEN3C-Cosmos-7B \
+  --image_path /path/to/input.png \
+  --prompt "" \
+  --trajectory left \
+  --movement_distance 0.3 \
+  --camera_rotation center_facing \
+  --num_inference_steps 35 \
+  --guidance_scale 1.0 \
+  --output_path outputs_video/gen3c_output.mp4
+```
+
+## FastVideo Defaults
+
+GEN3C defaults in FastVideo:
+
+- `height=704`, `width=1280`
+- `num_frames=121`
+- `num_inference_steps=35`
+- `guidance_scale=1.0`
+- `fps=24`
+
+These values are defined in:
+
+- `fastvideo/pipelines/basic/gen3c/profiles.py`
+- `fastvideo/configs/pipelines/gen3c.py`
+
+and align with the official GEN3C inference defaults in:
+
+- `tmp/GEN3C/cosmos_predict1/diffusion/inference/inference_utils.py`
+
+## Scheduler Note
+
+The converted GEN3C Diffusers layout may include a FlowMatch scheduler config, but GEN3C denoising uses EDM preconditioning behavior. FastVideo's GEN3C pipeline enforces an EDM scheduler at runtime for parity with official inference behavior.
+
+Implementation path:
+
+- `fastvideo/pipelines/basic/gen3c/gen3c_pipeline.py`
+
+## 3D Cache Conditioning Path
+
+FastVideo GEN3C conditioning stage performs:
+
+1. MoGe depth estimation from input image
+2. 3D cache initialization
+3. Camera trajectory generation
+4. Forward rendering of warped frames + masks
+5. VAE/tokenizer encoding of conditioning buffers
+6. Denoising with condition mask + condition pose channels
+
+Main implementation:
+
+- `fastvideo/pipelines/basic/gen3c/gen3c_pipeline.py`
+- `fastvideo/pipelines/basic/gen3c/cache_3d.py`
+- `fastvideo/pipelines/basic/gen3c/depth_estimation.py`
+- `fastvideo/models/vaes/gen3c_tokenizer_vae.py`
+
+## References
+
+- [GEN3C Paper](https://arxiv.org/abs/2503.03751)
+- [Official Repository](https://github.com/nv-tlabs/GEN3C)
+- [Official Checkpoint (raw)](https://huggingface.co/nvidia/GEN3C-Cosmos-7B)

docs/inference/support_matrix.md

@@ -73,6 +73,7 @@ pipeline initialization and sampling.
 | Matrix Game 2.0 Base | `FastVideo/Matrix-Game-2.0-Base-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Matrix Game 2.0 GTA | `FastVideo/Matrix-Game-2.0-GTA-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Matrix Game 2.0 TempleRun | `FastVideo/Matrix-Game-2.0-TempleRun-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| GEN3C Cosmos 7B | `FastVideo/GEN3C-Cosmos-7B-Diffusers` | 704px1280p | ❌ | ❌ | ❌ | ⭕ | ⭕ |
 
 **Note**: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.
 
@@ -85,6 +86,11 @@ The authoritative source for model-ID recognition is
 `fastvideo/registry.py`. If a model ID is registered there, FastVideo can
 resolve default pipeline and sampling configuration for it.
 
+**Note (GEN3C)**: The official `nvidia/GEN3C-Cosmos-7B` repo provides a raw
+`model.pt` checkpoint. Use a Diffusers-format repo (for example,
+`FastVideo/GEN3C-Cosmos-7B-Diffusers`) or convert locally with
+`scripts/checkpoint_conversion/convert_gen3c_to_fastvideo.py`.
+
 ## Special requirements
 
 ### Sliding Tile Attention

examples/inference/basic/README.md

@@ -28,6 +28,11 @@ For an example running DMD+VSA inference:
 python examples/inference/basic/basic_dmd.py
 ```
 
+For the typed config/request path added during the inference API refactor:
+```
+python examples/inference/basic/basic_dmd_new_api.py
+```
+
 ## Basic Walkthrough
 
 All you need to generate videos using multi-gpus from state-of-the-art diffusion pipelines is the following few lines!

examples/inference/basic/basic.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples"
 def main():

examples/inference/basic/basic_cosmos2_5_i2w.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 
 def main():

examples/inference/basic/basic_cosmos2_5_t2w.py

@@ -1,5 +1,5 @@
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 
 def main():

examples/inference/basic/basic_cosmos2_5_v2w.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 
 def main():

examples/inference/basic/basic_dmd.py

@@ -2,7 +2,7 @@ import os
 import time
 from fastvideo import VideoGenerator
 
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_dmd2"
 def main():

examples/inference/basic/basic_dmd_new_api.py

@@ -0,0 +1,98 @@
+import os
+import time
+
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    OffloadConfig,
+    OutputConfig,
+    PipelineSelection,
+)
+
+OUTPUT_PATH = "video_samples_dmd2_typed"
+
+
+def main():
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
+
+    model_name = "FastVideo/FastWan2.1-T2V-1.3B-Diffusers"
+    generator_config = GeneratorConfig(
+        model_path=model_name,
+        engine=EngineConfig(
+            num_gpus=1,
+            use_fsdp_inference=False,
+            offload=OffloadConfig(
+                text_encoder=True,
+                pin_cpu_memory=True,
+                dit=False,
+                vae=False,
+            ),
+        ),
+        # PR 2 still routes a few advanced inference knobs through the
+        # compatibility bridge until they get first-class typed fields.
+        pipeline=PipelineSelection(
+            experimental={
+                "VSA_sparsity": 0.8,
+            },
+        ),
+    )
+
+    load_start_time = time.perf_counter()
+    generator = VideoGenerator.from_config(generator_config)
+    load_end_time = time.perf_counter()
+    load_time = load_end_time - load_start_time
+
+    prompt = (
+        "A neon-lit alley in futuristic Tokyo during a heavy rainstorm at night. "
+        "The puddles reflect glowing signs in kanji, advertising ramen, karaoke, "
+        "and VR arcades. A woman in a translucent raincoat walks briskly with an "
+        "LED umbrella. Steam rises from a street food cart, and a cat darts "
+        "across the screen. Raindrops are visible on the camera lens, creating "
+        "a cinematic bokeh effect."
+    )
+    request = GenerationRequest(
+        prompt=prompt,
+        output=OutputConfig(
+            output_path=OUTPUT_PATH,
+            save_video=True,
+            return_frames=False,
+        ),
+    )
+
+    start_time = time.perf_counter()
+    result = generator.generate(request)
+    end_time = time.perf_counter()
+    gen_time = end_time - start_time
+
+    prompt2 = (
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently "
+        "in the breeze, enhancing the lion's commanding presence. The tone is "
+        "vibrant, embodying the raw energy of the wild. Low angle, steady "
+        "tracking shot, cinematic."
+    )
+    request2 = GenerationRequest(
+        prompt=prompt2,
+        output=OutputConfig(
+            output_path=OUTPUT_PATH,
+            save_video=True,
+            return_frames=False,
+        ),
+    )
+
+    start_time = time.perf_counter()
+    result2 = generator.generate(request2)
+    end_time = time.perf_counter()
+    gen_time2 = end_time - start_time
+
+    print(f"Time taken to load model: {load_time} seconds")
+    print(f"Time taken to generate video: {gen_time} seconds")
+    print(f"First output written to: {result.video_path}")
+    print(f"Time taken to generate video2: {gen_time2} seconds")
+    print(f"Second output written to: {result2.video_path}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_gen3c.py

@@ -0,0 +1,109 @@
+"""
+GEN3C: 3D-aware camera-controlled video generation.
+
+This example generates a video from a single input image with camera control.
+The pipeline uses MoGe depth estimation, 3D point cloud forward warping,
+and the GEN3C diffusion model.
+
+Requirements:
+  1. Install MoGe:
+     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:
+     huggingface-cli download nvidia/GEN3C-Cosmos-7B --local-dir official_weights/GEN3C-Cosmos-7B
+     python scripts/checkpoint_conversion/convert_gen3c_to_fastvideo.py \
+       --source ./official_weights/GEN3C-Cosmos-7B/model.pt \
+       --output ./converted_weights/GEN3C-Cosmos-7B \
+       --components-source nvidia/Cosmos-Predict2-2B-Video2World
+  3. Provide an input image for 3D-conditioned generation.
+"""
+
+import argparse
+
+from fastvideo import VideoGenerator
+
+
+def main():
+    parser = argparse.ArgumentParser(description="GEN3C video generation")
+    parser.add_argument("--model_path",
+                        type=str,
+                        default="converted_weights/GEN3C-Cosmos-7B")
+    parser.add_argument("--image_path",
+                        type=str,
+                        default=None,
+                        help="Input image for 3D cache conditioning")
+    parser.add_argument("--prompt",
+                        type=str,
+                        default="A slow camera pan over a sunlit landscape.")
+    parser.add_argument(
+        "--negative_prompt",
+        type=str,
+        default=(
+            "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
+            "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
+            "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, "
+            "jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special "
+            "effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and "
+            "flickering. Overall, the video is of poor quality."
+        ),
+    )
+    parser.add_argument("--trajectory",
+                        type=str,
+                        default="left",
+                        choices=[
+                            "left", "right", "up", "down", "zoom_in",
+                            "zoom_out", "clockwise", "counterclockwise", "none"
+                        ])
+    parser.add_argument("--movement_distance", type=float, default=0.3)
+    parser.add_argument("--camera_rotation",
+                        type=str,
+                        default="center_facing",
+                        choices=[
+                            "center_facing", "no_rotation",
+                            "trajectory_aligned"
+                        ])
+    parser.add_argument("--height", type=int, default=704)
+    parser.add_argument("--width", type=int, default=1280)
+    parser.add_argument("--num_frames", type=int, default=121)
+    parser.add_argument("--num_inference_steps", type=int, default=35)
+    parser.add_argument("--guidance_scale", type=float, default=1.0)
+    parser.add_argument("--output_path",
+                        type=str,
+                        default="outputs_video/gen3c.mp4")
+    parser.add_argument("--seed", type=int, default=42)
+    args = parser.parse_args()
+
+    generator = VideoGenerator.from_pretrained(
+        args.model_path,
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+    )
+
+    video = generator.generate_video(
+        args.prompt,
+        negative_prompt=args.negative_prompt,
+        image_path=args.image_path,
+        trajectory_type=args.trajectory,
+        movement_distance=args.movement_distance,
+        camera_rotation=args.camera_rotation,
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        fps=24,
+        seed=args.seed,
+        output_path=args.output_path,
+        save_video=True,
+    )
+
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_hy15.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 import json
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_hy15"
 def main():

examples/inference/basic/basic_hy15_1080p.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 import json
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_hy15_1080p"
 def main():

examples/inference/basic/basic_lingbotworld_base_cam.py

@@ -1,7 +1,7 @@
 from fastvideo import VideoGenerator
 from fastvideo.models.dits.lingbotworld.cam_utils import prepare_camera_embedding
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 OUTPUT_PATH = "video_samples_lingbotworld"
 def main():
     # FastVideo will automatically use the optimal default arguments for the

examples/inference/basic/basic_mps.py

@@ -1,5 +1,5 @@
 from fastvideo import VideoGenerator, PipelineConfig
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 def main():
     config = PipelineConfig.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")

examples/inference/basic/basic_ray.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples"
 def main():

examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py

@@ -2,7 +2,7 @@
 
 from fastvideo import VideoGenerator, SamplingParam
 import json
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_self_forcing_causal_wan2_2_14B_i2v"
 def main():

examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py

@@ -2,7 +2,7 @@
 
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_self_forcing_causal_wan2_2_14B_t2v"
 def main():

examples/inference/basic/basic_wan2_2.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_wan2_2_14B_t2v"
 def main():

examples/inference/basic/basic_wan2_2_Fun.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_wan2_1_Fun"
 OUTPUT_NAME = "wan2.1_test"

examples/inference/basic/basic_wan2_2_i2v.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-# from fastvideo.configs.sample import SamplingParam
+# from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "video_samples_wan2_2_14B_i2v"
 def main():

examples/inference/gradio/local/gradio_local_demo.py

@@ -5,7 +5,7 @@ import time
 
 import gradio as gr
 from fastvideo.entrypoints.video_generator import VideoGenerator
-from fastvideo.configs.sample.base import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 from copy import deepcopy
 
 

examples/inference/gradio/serving/gradio_frontend.py

@@ -9,7 +9,7 @@ import tempfile
 
 import gradio as gr
 
-from fastvideo.configs.sample.base import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 
 MODEL_PATH_MAPPING = {

examples/inference/gradio/serving/ray_serve_backend.py

@@ -185,7 +185,7 @@ class BaseModelDeployment:
 
     def _initialize_generator(self, config: Dict[str, Any]) -> None:
         from fastvideo.entrypoints.video_generator import VideoGenerator
-        from fastvideo.configs.sample.base import SamplingParam
+        from fastvideo.api.sampling_param import SamplingParam
 
         print(f"Initializing model: {self.model_path}")
         self.generator = VideoGenerator.from_pretrained(

examples/inference/lora/wan_lora_inference.py

@@ -1,5 +1,5 @@
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "./lora_out"
 def main():

examples/inference/lora/wan_lora_inference_from_ckpt.py

@@ -2,7 +2,7 @@
 Inference using a LoRA checkpoint from FastVideo trainer.
 """
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 OUTPUT_PATH = "./lora_out"
 def main():

fastvideo-kernel/build.sh

@@ -10,6 +10,35 @@ set -ex
 
 echo "Building fastvideo-kernel..."
 
+# ---------------------------------------------------------------------------
+# Neutralise conda-injected compiler toolchains.
+#
+# Conda compiler packages (gcc_linux-aarch64, gxx_linux-64, etc.) set
+# CMAKE_ARGS, CFLAGS, CXXFLAGS, and LDFLAGS on activation.  When multiple
+# toolchains are installed the variables can reference a *cross*-compiler
+# that doesn't match the host (e.g. aarch64-conda-linux-gnu-c++ on x86_64).
+# Even when the correct toolchain is active, the flags it injects
+# (-march=nocona, -mtune=haswell, …) can conflict with nvcc's host-compiler
+# expectations.  Clear them so CMake discovers the system compiler instead.
+# ---------------------------------------------------------------------------
+if [[ -n "${CONDA_PREFIX:-}" ]]; then
+    _need_clean=0
+    # Detect conda cross-compiler that doesn't match the host.
+    _host_arch="$(uname -m)"
+    if [[ "${CXX:-}" == *"conda"* ]] || [[ "${CC:-}" == *"conda"* ]]; then
+        _need_clean=1
+    fi
+    if [[ "${CMAKE_ARGS:-}" == *"conda"* ]]; then
+        _need_clean=1
+    fi
+    if (( _need_clean )); then
+        echo "NOTE: Clearing conda-injected compiler settings (CC/CXX/CMAKE_ARGS/CFLAGS/...)"
+        echo "      to use the system compiler for CUDA extension builds."
+        unset CC CXX CMAKE_ARGS CFLAGS CXXFLAGS LDFLAGS
+    fi
+    unset _need_clean _host_arch
+fi
+
 # Ensure submodules are initialized if needed (tk)
 git submodule update --init --recursive
 
@@ -32,7 +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/pyproject.toml

@@ -23,7 +23,7 @@ classifiers = [
 ]
 dependencies = [
     "torch>=2.5.0",
-    "triton>=2.0.0",
+    "triton>=2.0.0; sys_platform == 'linux'",
 ]
 
 [project.urls]

fastvideo/__init__.py

@@ -1,5 +1,5 @@
 from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.entrypoints.video_generator import VideoGenerator
 from fastvideo.version import __version__
 

fastvideo/api/__init__.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+from fastvideo.api.schema import (
+    CompileConfig,
+    ComponentConfig,
+    ContinuationState,
+    EngineConfig,
+    GenerationPlan,
+    GenerationRequest,
+    GeneratorConfig,
+    GpuPoolConfig,
+    InputConfig,
+    OffloadConfig,
+    OutputConfig,
+    ParallelismConfig,
+    PipelineSelection,
+    PlannedStage,
+    PromptEnhancerConfig,
+    PromptSafetyConfig,
+    QuantizationConfig,
+    RequestRuntimeConfig,
+    RunConfig,
+    SamplingConfig,
+    ServeConfig,
+    ServerConfig,
+    StreamingConfig,
+    WarmupConfig,
+)
+from fastvideo.api.errors import ConfigValidationError
+from fastvideo.api.overrides import apply_overrides, parse_cli_overrides
+from fastvideo.api.presets import (
+    InferencePreset,
+    PresetStageSpec,
+    get_all_preset_names,
+    get_preset,
+    get_presets_for_family,
+    register_preset,
+    validate_preset_selection,
+    validate_stage_names,
+    validate_stage_overrides,
+)
+from fastvideo.api.parser import (
+    config_to_dict,
+    load_config,
+    load_raw_config,
+    load_run_config,
+    load_serve_config,
+    parse_config,
+)
+from fastvideo.api.results import GenerationResult
+from fastvideo.api.sampling_param import SamplingParam
+
+__all__ = [
+    "CompileConfig",
+    "ComponentConfig",
+    "ContinuationState",
+    "ConfigValidationError",
+    "EngineConfig",
+    "GenerationResult",
+    "GenerationPlan",
+    "GenerationRequest",
+    "GeneratorConfig",
+    "GpuPoolConfig",
+    "InputConfig",
+    "OffloadConfig",
+    "OutputConfig",
+    "ParallelismConfig",
+    "PipelineSelection",
+    "PlannedStage",
+    "PromptEnhancerConfig",
+    "PromptSafetyConfig",
+    "QuantizationConfig",
+    "RequestRuntimeConfig",
+    "RunConfig",
+    "SamplingConfig",
+    "SamplingParam",
+    "ServeConfig",
+    "ServerConfig",
+    "StreamingConfig",
+    "WarmupConfig",
+    "InferencePreset",
+    "PresetStageSpec",
+    "apply_overrides",
+    "config_to_dict",
+    "load_config",
+    "load_raw_config",
+    "load_run_config",
+    "load_serve_config",
+    "parse_cli_overrides",
+    "get_all_preset_names",
+    "get_preset",
+    "get_presets_for_family",
+    "parse_config",
+    "register_preset",
+    "validate_preset_selection",
+    "validate_stage_names",
+    "validate_stage_overrides",
+]

fastvideo/api/compat.py

@@ -0,0 +1,568 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from collections.abc import Mapping
+from copy import deepcopy
+from dataclasses import fields, is_dataclass
+from pathlib import Path
+from typing import Any
+
+from fastvideo.api.overrides import apply_overrides, normalize_overrides
+from fastvideo.api.parser import config_to_dict, load_raw_config, parse_config
+from fastvideo.api.request_metadata import (
+    EXPLICIT_PATHS_ATTR,
+    bind_generation_request_raw,
+    get_explicit_paths,
+    reset_tracking_roots,
+)
+from fastvideo.api.schema import (
+    CompileConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    InputConfig,
+    OutputConfig,
+    RequestRuntimeConfig,
+    SamplingConfig,
+)
+from fastvideo.api.sampling_param import SamplingParam
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.basic.ltx2.stage_overrides import REFINE_FLAT_KEYS
+from fastvideo.utils import shallow_asdict
+
+_INPUT_FIELD_NAMES = {field.name for field in fields(InputConfig)}
+_SAMPLING_FIELD_NAMES = {field.name for field in fields(SamplingConfig)}
+_RUNTIME_FIELD_NAMES = {field.name for field in fields(RequestRuntimeConfig)}
+_OUTPUT_FIELD_NAMES = {field.name for field in fields(OutputConfig)}
+_MISSING = object()
+_LEGACY_REQUEST_ALIASES = {
+    "neg_prompt": "negative_prompt",
+}
+_REQUEST_PIPELINE_OVERRIDE_FIELDS = frozenset({
+    "embedded_cfg_scale",
+})
+_COMPILE_TYPED_KEYS = ("backend", "fullgraph", "mode", "dynamic")
+
+
+def normalize_generator_config(config: GeneratorConfig | Mapping[str, Any], ) -> GeneratorConfig:
+    if isinstance(config, GeneratorConfig):
+        return config
+    return parse_config(GeneratorConfig, config)
+
+
+def load_generator_config_from_file(
+    path: str | Path,
+    overrides: list[str] | Mapping[str, Any] | None = None,
+) -> GeneratorConfig:
+    raw = load_raw_config(path)
+    normalized_overrides = normalize_overrides(overrides)
+
+    if _looks_like_run_or_serve_config(raw):
+        if normalized_overrides:
+            raw = apply_overrides(raw, normalized_overrides)
+        return parse_config(GeneratorConfig, raw["generator"])
+
+    if normalized_overrides:
+        adjusted = normalized_overrides
+        if all(key.startswith("generator.") for key in adjusted):
+            adjusted = {key[len("generator."):]: value for key, value in adjusted.items()}
+        raw = apply_overrides(raw, adjusted)
+
+    return parse_config(GeneratorConfig, raw)
+
+
+def legacy_from_pretrained_to_config(
+    model_path: str,
+    kwargs: Mapping[str, Any],
+) -> GeneratorConfig:
+    raw: dict[str, Any] = {"model_path": model_path}
+    engine: dict[str, Any] = {}
+    parallelism: dict[str, Any] = {}
+    offload: dict[str, Any] = {}
+    compile_config: dict[str, Any] = {}
+    pipeline: dict[str, Any] = {}
+    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":
+            raw["revision"] = value
+        elif key == "trust_remote_code":
+            raw["trust_remote_code"] = value
+        elif key == "num_gpus":
+            engine["num_gpus"] = value
+        elif key == "distributed_executor_backend":
+            engine["execution_backend"] = value
+        elif key in {"tp_size", "sp_size", "hsdp_replicate_dim", "hsdp_shard_dim", "dist_timeout"}:
+            parallelism[key] = value
+        elif key == "dit_cpu_offload":
+            offload["dit"] = value
+        elif key == "dit_layerwise_offload":
+            offload["dit_layerwise"] = value
+        elif key == "text_encoder_cpu_offload":
+            offload["text_encoder"] = value
+        elif key == "image_encoder_cpu_offload":
+            offload["image_encoder"] = value
+        elif key == "vae_cpu_offload":
+            offload["vae"] = value
+        elif key == "pin_cpu_memory":
+            offload["pin_cpu_memory"] = value
+        elif key == "enable_torch_compile":
+            compile_config["enabled"] = value
+        elif key == "torch_compile_kwargs":
+            remaining: dict[str, Any] = (dict(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 == "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":
+            quantization["text_encoder_quant"] = value
+        elif key == "workload_type":
+            pipeline["workload_type"] = value
+        elif key == "lora_path":
+            components["lora_path"] = value
+        elif key == "override_pipeline_cls_name":
+            components["override_pipeline_cls_name"] = value
+        elif key == "override_transformer_cls_name":
+            components["override_transformer_cls_name"] = value
+        elif key == "pipeline_config":
+            if isinstance(value, str):
+                components["pipeline_config_path"] = value
+            else:
+                experimental[key] = deepcopy(value)
+        elif key == "override_text_encoder_safetensors":
+            components["text_encoder_weights"] = value
+        elif key == "init_weights_from_safetensors":
+            components["transformer_weights"] = value
+        elif key == "init_weights_from_safetensors_2":
+            components["transformer_2_weights"] = value
+        else:
+            experimental[key] = deepcopy(value)
+
+    if parallelism:
+        engine["parallelism"] = parallelism
+    if offload:
+        engine["offload"] = offload
+    if compile_config:
+        engine["compile"] = compile_config
+    if quantization:
+        engine["quantization"] = quantization
+    if engine:
+        raw["engine"] = engine
+
+    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:
+        raw["pipeline"] = pipeline
+
+    return parse_config(GeneratorConfig, raw)
+
+
+def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, Any], ) -> FastVideoArgs:
+    normalized = normalize_generator_config(config)
+    unsupported = []
+    if normalized.pipeline.preset is not None:
+        unsupported.append("pipeline.preset")
+    if normalized.pipeline.preset_version is not None:
+        unsupported.append("pipeline.preset_version")
+    if normalized.pipeline.components.vae_weights is not None:
+        unsupported.append("pipeline.components.vae_weights")
+    if unsupported:
+        joined = ", ".join(unsupported)
+        raise NotImplementedError(f"VideoGenerator compatibility adapter does not support {joined} yet")
+
+    engine = normalized.engine
+    kwargs: dict[str, Any] = {
+        "model_path": normalized.model_path,
+        "revision": normalized.revision,
+        "trust_remote_code": normalized.trust_remote_code,
+        "num_gpus": engine.num_gpus,
+        "distributed_executor_backend": engine.execution_backend,
+        "tp_size": engine.parallelism.tp_size,
+        "sp_size": engine.parallelism.sp_size,
+        "hsdp_replicate_dim": engine.parallelism.hsdp_replicate_dim,
+        "hsdp_shard_dim": engine.parallelism.hsdp_shard_dim,
+        "dist_timeout": engine.parallelism.dist_timeout,
+        "dit_cpu_offload": engine.offload.dit,
+        "dit_layerwise_offload": engine.offload.dit_layerwise,
+        "text_encoder_cpu_offload": engine.offload.text_encoder,
+        "image_encoder_cpu_offload": engine.offload.image_encoder,
+        "vae_cpu_offload": engine.offload.vae,
+        "pin_cpu_memory": engine.offload.pin_cpu_memory,
+        "enable_torch_compile": engine.compile.enabled,
+        "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
+
+    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
+
+    components = normalized.pipeline.components
+    if components.pipeline_config_path is not None:
+        kwargs["pipeline_config"] = components.pipeline_config_path
+    if components.lora_path is not None:
+        kwargs["lora_path"] = components.lora_path
+    if components.override_pipeline_cls_name is not None:
+        kwargs["override_pipeline_cls_name"] = components.override_pipeline_cls_name
+    if components.override_transformer_cls_name is not None:
+        kwargs["override_transformer_cls_name"] = components.override_transformer_cls_name
+    if components.text_encoder_weights is not None:
+        kwargs["override_text_encoder_safetensors"] = components.text_encoder_weights
+    if components.transformer_weights is not None:
+        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
+
+    preset_overrides = deepcopy(normalized.pipeline.preset_overrides)
+    refine = preset_overrides.pop("refine", None)
+    if isinstance(refine, Mapping):
+        for key in 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)
+
+
+def normalize_generation_request(request: GenerationRequest | Mapping[str, Any], ) -> GenerationRequest:
+    normalized = (request if isinstance(request, GenerationRequest) else parse_config(GenerationRequest, request))
+
+    if not hasattr(normalized, EXPLICIT_PATHS_ATTR):
+        # Request wasn't bound through the parser (e.g. constructed
+        # directly). Treat every currently-set field as explicit.
+        bind_generation_request_raw(normalized, _serialize_generation_request(normalized))
+    return normalized
+
+
+def legacy_generate_call_to_request(
+    prompt: str | None,
+    sampling_param: SamplingParam | None,
+    *,
+    mouse_cond: Any | None = None,
+    keyboard_cond: Any | None = None,
+    grid_sizes: Any | None = None,
+    legacy_kwargs: Mapping[str, Any] | None = None,
+) -> GenerationRequest:
+    raw = _sampling_param_to_request_raw(sampling_param)
+    if prompt is not None:
+        raw["prompt"] = prompt
+
+    for key, value in (legacy_kwargs or {}).items():
+        _apply_request_field(raw, key, value)
+
+    if mouse_cond is not None:
+        raw.setdefault("inputs", {})["mouse_cond"] = mouse_cond
+    if keyboard_cond is not None:
+        raw.setdefault("inputs", {})["keyboard_cond"] = keyboard_cond
+    if grid_sizes is not None:
+        raw.setdefault("inputs", {})["grid_sizes"] = grid_sizes
+
+    normalized = parse_config(GenerationRequest, raw)
+    bind_generation_request_raw(normalized, raw)
+    return normalized
+
+
+def request_to_sampling_param(
+    request: GenerationRequest,
+    *,
+    model_path: str,
+) -> 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)
+    updates = explicit_request_updates(request)
+
+    for key, value in updates.items():
+        if hasattr(sampling_param, key):
+            setattr(sampling_param, key, deepcopy(value))
+        elif key in _REQUEST_PIPELINE_OVERRIDE_FIELDS:
+            continue
+        elif value == _SCHEMA_DEFAULT_UPDATES.get(key, _MISSING):
+            # Schema-default field that isn't on SamplingParam; tolerated
+            # because direct GenerationRequest(...) construction has no
+            # way to distinguish "user set" from "schema default".
+            continue
+        else:
+            raise ValueError(f"Request field {key!r} is not supported by sampling params for {model_path}")
+
+    sampling_param.__post_init__()
+    sampling_param.check_sampling_param()
+    return sampling_param
+
+
+def expand_request_prompt_batch(request: GenerationRequest, ) -> list[GenerationRequest]:
+    if not isinstance(request.prompt, list):
+        return [request]
+
+    requests: list[GenerationRequest] = []
+    for index, prompt in enumerate(request.prompt):
+        single_request = deepcopy(request)
+        # deepcopy preserves the tracking-root cycle, but re-pin roots
+        # defensively so that subsequent setattrs record on the copy.
+        reset_tracking_roots(single_request)
+        single_request.prompt = prompt
+        _fan_out_batched_input_value(request, single_request, "image_path", index)
+        _fan_out_batched_input_value(request, single_request, "video_path", index)
+        requests.append(single_request)
+    return requests
+
+
+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 the legacy
+    ``torch_compile_kwargs`` dict, emitting only explicitly-set typed
+    fields and merging ``extras`` on top."""
+    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 {}
+
+    raw: dict[str, Any] = {}
+    for key, value in shallow_asdict(sampling_param).items():
+        if key == "prompt":
+            continue
+        _apply_request_field(raw, key, deepcopy(value))
+    return raw
+
+
+def _apply_request_field(
+    raw: dict[str, Any],
+    key: str,
+    value: Any,
+) -> None:
+    key = _LEGACY_REQUEST_ALIASES.get(key, key)
+    if key == "negative_prompt":
+        raw["negative_prompt"] = value
+        return
+    if key in _INPUT_FIELD_NAMES:
+        raw.setdefault("inputs", {})[key] = value
+        return
+    if key in _SAMPLING_FIELD_NAMES:
+        raw.setdefault("sampling", {})[key] = value
+        return
+    if key in _RUNTIME_FIELD_NAMES:
+        raw.setdefault("runtime", {})[key] = value
+        return
+    if key in _OUTPUT_FIELD_NAMES:
+        raw.setdefault("output", {})[key] = value
+        return
+    raw.setdefault("extensions", {})[key] = value
+
+
+def request_to_pipeline_overrides(request: GenerationRequest) -> dict[str, Any]:
+    overrides: dict[str, Any] = {}
+    for key, value in explicit_request_updates(request).items():
+        if key in _REQUEST_PIPELINE_OVERRIDE_FIELDS:
+            overrides[key] = deepcopy(value)
+    return overrides
+
+
+def explicit_request_updates(request: GenerationRequest) -> dict[str, Any]:
+    """Project a ``GenerationRequest`` down to *explicitly set* fields only.
+
+    Returns a flat kwargs dict suitable for merging into a generator call.
+    The projection uses ``_fastvideo_explicit_paths`` (populated during
+    ``parse_config`` / raw binding) so schema defaults on the dataclass
+    are **not** emitted — only paths the caller/operator actually wrote.
+
+    This is what makes ``ServeConfig.default_request`` work as an
+    operator-pinned baseline rather than a full override: a YAML with just
+    ``sampling.seed: 42`` yields ``{"seed": 42}``, not the full sampling
+    config with its 15 schema defaults.
+
+    Precondition: the request must carry ``_fastvideo_explicit_paths`` —
+    populated by :func:`fastvideo.api.parser.parse_config` or
+    :func:`fastvideo.api.compat.normalize_generation_request`. Calling on
+    a raw ``GenerationRequest()`` asserts.
+    """
+    assert hasattr(request,
+                   EXPLICIT_PATHS_ATTR), ("GenerationRequest reached explicit_request_updates without tracking; "
+                                          "every entry point must route through normalize_generation_request "
+                                          "or parse_config first")
+    paths = get_explicit_paths(request)
+    raw = _build_sparse_raw_from_paths(request, paths)
+    return _extract_request_updates(raw)
+
+
+def _build_sparse_raw_from_paths(
+    request: GenerationRequest,
+    paths: frozenset[str],
+) -> dict[str, Any]:
+    result: dict[str, Any] = {}
+    for path in paths:
+        parts = path.split(".")
+        value = _read_dotted_path(request, parts)
+        if value is _MISSING:
+            continue
+        _set_dotted_path(result, parts, deepcopy(value))
+    return result
+
+
+def _read_dotted_path(obj: Any, parts: list[str]) -> Any:
+    for part in parts:
+        if is_dataclass(obj) and not isinstance(obj, type):
+            if not hasattr(obj, part):
+                return _MISSING
+            obj = getattr(obj, part)
+        elif isinstance(obj, Mapping):
+            if part not in obj:
+                return _MISSING
+            obj = obj[part]
+        else:
+            return _MISSING
+    return obj
+
+
+def _set_dotted_path(
+    target: dict[str, Any],
+    parts: list[str],
+    value: Any,
+) -> None:
+    cursor = target
+    for part in parts[:-1]:
+        nxt = cursor.get(part)
+        if not isinstance(nxt, dict):
+            nxt = {}
+            cursor[part] = nxt
+        cursor = nxt
+    cursor[parts[-1]] = value
+
+
+def _extract_request_updates(raw: Mapping[str, Any]) -> dict[str, Any]:
+    updates: dict[str, Any] = {}
+    if "negative_prompt" in raw:
+        updates["negative_prompt"] = deepcopy(raw["negative_prompt"])
+
+    for section_name in ("inputs", "sampling", "runtime", "output"):
+        section = raw.get(section_name)
+        if not isinstance(section, Mapping):
+            continue
+        for key, value in section.items():
+            updates[key] = deepcopy(value)
+
+    stage_overrides = raw.get("stage_overrides")
+    if stage_overrides:
+        updates.update(_flatten_stage_overrides(stage_overrides))
+
+    extensions = raw.get("extensions")
+    if isinstance(extensions, Mapping):
+        for key, value in extensions.items():
+            updates[key] = deepcopy(value)
+
+    return updates
+
+
+def _flatten_stage_overrides(stage_overrides: Any) -> dict[str, Any]:
+    if not isinstance(stage_overrides, Mapping):
+        raise ValueError("GenerationRequest.stage_overrides must be a mapping")
+
+    flattened: dict[str, Any] = {}
+    for stage_name, overrides in stage_overrides.items():
+        if not isinstance(overrides, Mapping):
+            raise ValueError(f"GenerationRequest.stage_overrides.{stage_name} must be a mapping")
+        for key, value in overrides.items():
+            if key in flattened and flattened[key] != value:
+                raise ValueError(f"Conflicting stage override for {key!r} across stages")
+            flattened[key] = deepcopy(value)
+    return flattened
+
+
+def _serialize_generation_request(request: GenerationRequest) -> dict[str, Any]:
+    return deepcopy(config_to_dict(request))
+
+
+_SCHEMA_DEFAULT_UPDATES = _extract_request_updates(config_to_dict(GenerationRequest()))
+
+
+def _fan_out_batched_input_value(
+    source_request: GenerationRequest,
+    target_request: GenerationRequest,
+    field_name: str,
+    index: int,
+) -> None:
+    value = getattr(source_request.inputs, field_name)
+    if not isinstance(value, list):
+        return
+    _validate_batched_input_length(source_request.prompt, value, field_name)
+    setattr(target_request.inputs, field_name, deepcopy(value[index]))
+
+
+def _validate_batched_input_length(
+    prompts: str | list[str] | None,
+    values: list[Any],
+    field_name: str,
+) -> None:
+    if not isinstance(prompts, list):
+        return
+    if len(values) != len(prompts):
+        raise ValueError(f"GenerationRequest.inputs.{field_name} must have the same length as request.prompt")
+
+
+__all__ = [
+    "explicit_request_updates",
+    "generator_config_to_fastvideo_args",
+    "legacy_from_pretrained_to_config",
+    "legacy_generate_call_to_request",
+    "load_generator_config_from_file",
+    "normalize_generation_request",
+    "normalize_generator_config",
+    "request_to_pipeline_overrides",
+    "request_to_sampling_param",
+]

fastvideo/api/errors.py

@@ -0,0 +1,16 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+
+class ConfigValidationError(ValueError):
+    """Validation error that keeps track of the nested config path."""
+
+    def __init__(self, path: str, message: str):
+        self.path = path
+        self.message = message
+        super().__init__(str(self))
+
+    def __str__(self) -> str:
+        if self.path:
+            return f"{self.path}: {self.message}"
+        return self.message

fastvideo/api/overrides.py

@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from copy import deepcopy
+from typing import Any
+from collections.abc import Mapping
+
+import yaml
+
+from fastvideo.api.errors import ConfigValidationError
+
+
+def parse_cli_overrides(overrides: list[str]) -> dict[str, Any]:
+    """Parse ``--dotted.key value`` style overrides into a flat mapping."""
+    parsed: dict[str, Any] = {}
+    index = 0
+    while index < len(overrides):
+        token = overrides[index]
+        if not token.startswith("--"):
+            raise ValueError(f"Expected --dotted.key, got {token!r}")
+
+        key = token[2:]
+        if not key:
+            raise ValueError("Override key cannot be empty")
+
+        if "=" in key:
+            key, raw_value = key.split("=", 1)
+        else:
+            index += 1
+            if index >= len(overrides):
+                raise ValueError(f"Missing value for override {token!r}")
+            raw_value = overrides[index]
+
+        parsed[_normalize_override_key(key)] = _cast_override_value(raw_value)
+        index += 1
+
+    return parsed
+
+
+def apply_overrides(config: Mapping[str, Any], overrides: Mapping[str, Any]) -> dict[str, Any]:
+    """Return a copy of ``config`` with dotted-key overrides applied."""
+    merged = deepcopy(dict(config))
+    for dotted_key, value in overrides.items():
+        _apply_single_override(merged, dotted_key, value)
+    return merged
+
+
+def normalize_overrides(overrides: list[str] | Mapping[str, Any] | None, ) -> dict[str, Any] | None:
+    """Normalize a CLI list or mapping of overrides into a flat dict."""
+    if not overrides:
+        return None
+    if isinstance(overrides, list):
+        return parse_cli_overrides(overrides)
+    return dict(overrides)
+
+
+def _apply_single_override(config: dict[str, Any], dotted_key: str, value: Any) -> None:
+    parts = dotted_key.split(".")
+    if not all(parts):
+        raise ValueError(f"Invalid override path {dotted_key!r}")
+
+    cursor = config
+    for depth, part in enumerate(parts[:-1]):
+        existing = cursor.get(part)
+        if existing is None:
+            existing = {}
+            cursor[part] = existing
+        elif not isinstance(existing, dict):
+            raise ConfigValidationError(
+                ".".join(parts[:depth + 1]),
+                "cannot apply nested override through a non-mapping value",
+            )
+        cursor = existing
+
+    cursor[parts[-1]] = value
+
+
+def _cast_override_value(raw: str) -> Any:
+    lowered = raw.lower()
+    if lowered == "true":
+        return True
+    if lowered == "false":
+        return False
+    if lowered in {"none", "null"}:
+        return None
+
+    try:
+        return int(raw)
+    except ValueError:
+        pass
+
+    try:
+        return float(raw)
+    except ValueError:
+        pass
+
+    if raw.startswith("[") or raw.startswith("{"):
+        try:
+            return yaml.safe_load(raw)
+        except yaml.YAMLError:
+            pass
+
+    return raw
+
+
+def _normalize_override_key(key: str) -> str:
+    return key.replace("-", "_")
+
+
+__all__ = ["apply_overrides", "normalize_overrides", "parse_cli_overrides"]

fastvideo/api/parser.py

@@ -0,0 +1,324 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import dataclasses
+import json
+import types
+from pathlib import Path
+from collections.abc import Mapping
+from typing import Any, Literal, TypeVar, Union, get_args, get_origin, get_type_hints
+
+import yaml
+
+from fastvideo.api.errors import ConfigValidationError
+from fastvideo.api.overrides import apply_overrides, normalize_overrides
+from fastvideo.api.request_metadata import (
+    bind_generation_request_raw,
+    bind_run_config_raw,
+    bind_serve_config_raw,
+)
+from fastvideo.api.schema import GenerationRequest, RunConfig, ServeConfig
+
+T = TypeVar("T")
+_UNION_ORIGINS = {types.UnionType, Union}
+
+
+@dataclasses.dataclass(frozen=True)
+class _DataclassSpec:
+    cls: type[Any]
+    type_hints: dict[str, Any]
+    fields_by_name: dict[str, dataclasses.Field[Any]]
+
+
+def parse_config(config_type: type[T], raw: Mapping[str, Any] | T) -> T:
+    """Parse a nested mapping into a typed inference config object."""
+    if isinstance(raw, config_type):
+        return raw
+    if not isinstance(raw, Mapping):
+        raise ConfigValidationError("", f"expected mapping for {config_type.__name__}")
+    parsed = _SchemaParser().parse_dataclass(config_type, raw, "")
+    if config_type is GenerationRequest:
+        return bind_generation_request_raw(parsed, raw)
+    if config_type is RunConfig:
+        return bind_run_config_raw(parsed, raw)
+    if config_type is ServeConfig:
+        return bind_serve_config_raw(parsed, raw)
+    return parsed
+
+
+def config_to_dict(config: Any) -> Any:
+    """Serialize a typed config object into plain Python containers."""
+    if dataclasses.is_dataclass(config) and not isinstance(config, type):
+        return {field.name: config_to_dict(getattr(config, field.name)) for field in dataclasses.fields(config)}
+    if isinstance(config, list):
+        return [config_to_dict(item) for item in config]
+    if isinstance(config, dict):
+        return {key: config_to_dict(value) for key, value in config.items()}
+    return config
+
+
+def load_config(
+    config_type: type[T],
+    path: str | Path,
+    overrides: list[str] | Mapping[str, Any] | None = None,
+) -> T:
+    """Load a typed config object from YAML or JSON."""
+    raw = load_raw_config(path)
+    normalized_overrides = normalize_overrides(overrides)
+    if normalized_overrides:
+        raw = apply_overrides(raw, normalized_overrides)
+    return parse_config(config_type, raw)
+
+
+def load_run_config(
+    path: str | Path,
+    overrides: list[str] | Mapping[str, Any] | None = None,
+) -> RunConfig:
+    return load_config(RunConfig, path, overrides)
+
+
+def load_serve_config(
+    path: str | Path,
+    overrides: list[str] | Mapping[str, Any] | None = None,
+) -> ServeConfig:
+    return load_config(ServeConfig, path, overrides)
+
+
+def load_raw_config(path: str | Path) -> dict[str, Any]:
+    config_path = Path(path)
+    if not config_path.exists():
+        raise FileNotFoundError(f"Config file not found: {config_path}")
+
+    with config_path.open(encoding="utf-8") as handle:
+        raw = _load_raw_mapping(handle, config_path)
+
+    if raw is None:
+        return {}
+    if not isinstance(raw, Mapping):
+        raise ConfigValidationError("", f"{config_path} must contain a top-level mapping")
+    return dict(raw)
+
+
+def _load_raw_mapping(handle: Any, config_path: Path) -> Any:
+    suffix = config_path.suffix.lower()
+    if suffix in {".yaml", ".yml"}:
+        return yaml.safe_load(handle)
+    if suffix == ".json":
+        return json.load(handle)
+    raise ValueError(f"Unsupported config file format: {config_path}")
+
+
+class _SchemaParser:
+
+    def parse_dataclass(
+        self,
+        config_type: type[T],
+        raw: Mapping[str, Any],
+        path: str,
+    ) -> T:
+        if not isinstance(raw, Mapping):
+            raise ConfigValidationError(path, f"expected mapping for {config_type.__name__}")
+
+        spec = _get_dataclass_spec(config_type)
+        self._validate_keys(raw, spec, path)
+
+        values: dict[str, Any] = {}
+        for name, field in spec.fields_by_name.items():
+            field_path = _join_path(path, name)
+            if name in raw:
+                values[name] = self.parse_value(spec.type_hints[name], raw[name], field_path)
+                continue
+            if _field_is_required(field):
+                raise ConfigValidationError(field_path, "missing required field")
+
+        return config_type(**values)
+
+    def parse_value(self, annotation: Any, value: Any, path: str) -> Any:
+        if annotation is Any:
+            return value
+
+        origin = get_origin(annotation)
+        if origin in _UNION_ORIGINS:
+            return self._parse_union(annotation, value, path)
+        if origin is Literal:
+            return self._parse_literal(annotation, value, path)
+        if origin is list:
+            return self._parse_list(annotation, value, path)
+        if origin is dict:
+            return self._parse_dict(annotation, value, path)
+        if origin is tuple:
+            return self._parse_tuple(annotation, value, path)
+        if isinstance(annotation, type) and dataclasses.is_dataclass(annotation):
+            return self.parse_dataclass(annotation, value, path)
+
+        scalar_parser = _SCALAR_PARSERS.get(annotation)
+        if scalar_parser is not None:
+            return scalar_parser(value, path)
+
+        return self._parse_instance(annotation, value, path)
+
+    def _validate_keys(
+        self,
+        raw: Mapping[str, Any],
+        spec: _DataclassSpec,
+        path: str,
+    ) -> None:
+        for key in raw:
+            if not isinstance(key, str):
+                raise ConfigValidationError(path, "expected mapping keys to be strings")
+            if key not in spec.fields_by_name:
+                raise ConfigValidationError(_join_path(path, key), "unknown field")
+
+    def _parse_union(self, annotation: Any, value: Any, path: str) -> Any:
+        candidates = [candidate for candidate in get_args(annotation) if candidate is not type(None)]
+        if value is None and len(candidates) != len(get_args(annotation)):
+            return None
+        if len(candidates) == 1:
+            return self.parse_value(candidates[0], value, path)
+
+        errors: list[str] = []
+        for candidate in candidates:
+            try:
+                return self.parse_value(candidate, value, path)
+            except ConfigValidationError as exc:
+                errors.append(exc.message)
+
+        expected = ", ".join(_type_name(candidate) for candidate in candidates)
+        detail = errors[0] if errors else f"expected one of ({expected})"
+        raise ConfigValidationError(path, detail)
+
+    def _parse_literal(self, annotation: Any, value: Any, path: str) -> Any:
+        allowed = get_args(annotation)
+        if value not in allowed:
+            raise ConfigValidationError(path, f"expected one of {sorted(allowed)!r}")
+        return value
+
+    def _parse_list(self, annotation: Any, value: Any, path: str) -> list[Any]:
+        if not isinstance(value, list):
+            raise ConfigValidationError(path, "expected list")
+        item_type = get_args(annotation)[0] if get_args(annotation) else Any
+        return [self.parse_value(item_type, item, f"{path}[{index}]") for index, item in enumerate(value)]
+
+    def _parse_dict(self, annotation: Any, value: Any, path: str) -> dict[Any, Any]:
+        if not isinstance(value, Mapping):
+            raise ConfigValidationError(path, "expected mapping")
+
+        key_type, value_type = (get_args(annotation) + (Any, Any))[:2]
+        parsed: dict[Any, Any] = {}
+        for key, item in value.items():
+            parsed_key = self._parse_dict_key(key_type, key, path)
+            item_path = _join_path(path, str(key))
+            parsed[parsed_key] = self.parse_value(value_type, item, item_path)
+        return parsed
+
+    def _parse_tuple(self, annotation: Any, value: Any, path: str) -> tuple[Any, ...]:
+        if not isinstance(value, list | tuple):
+            raise ConfigValidationError(path, "expected tuple")
+
+        item_types = get_args(annotation)
+        if len(item_types) == 2 and item_types[1] is Ellipsis:
+            return tuple(self.parse_value(item_types[0], item, f"{path}[{index}]") for index, item in enumerate(value))
+
+        if len(value) != len(item_types):
+            raise ConfigValidationError(path, f"expected tuple of length {len(item_types)}")
+
+        return tuple(
+            self.parse_value(item_type, item, f"{path}[{index}]")
+            for index, (item_type, item) in enumerate(zip(item_types, value, strict=True)))
+
+    def _parse_dict_key(self, annotation: Any, value: Any, path: str) -> Any:
+        if annotation is Any:
+            return value
+        if annotation is str:
+            if not isinstance(value, str):
+                raise ConfigValidationError(path, "expected string dictionary keys")
+            return value
+        if annotation is int:
+            if not isinstance(value, int) or isinstance(value, bool):
+                raise ConfigValidationError(path, "expected integer dictionary keys")
+            return value
+        return value
+
+    def _parse_instance(self, annotation: Any, value: Any, path: str) -> Any:
+        if isinstance(annotation, type) and not isinstance(value, annotation):
+            raise ConfigValidationError(path, f"expected {annotation.__name__}")
+        return value
+
+
+def _parse_bool(value: Any, path: str) -> bool:
+    if type(value) is not bool:
+        raise ConfigValidationError(path, "expected bool")
+    return value
+
+
+def _parse_int(value: Any, path: str) -> int:
+    if not isinstance(value, int) or isinstance(value, bool):
+        raise ConfigValidationError(path, "expected int")
+    return value
+
+
+def _parse_float(value: Any, path: str) -> float:
+    if not isinstance(value, int | float) or isinstance(value, bool):
+        raise ConfigValidationError(path, "expected float")
+    return float(value)
+
+
+def _parse_str(value: Any, path: str) -> str:
+    if not isinstance(value, str):
+        raise ConfigValidationError(path, "expected str")
+    return value
+
+
+_SCALAR_PARSERS: dict[Any, Any] = {
+    bool: _parse_bool,
+    int: _parse_int,
+    float: _parse_float,
+    str: _parse_str,
+}
+
+
+def _field_is_required(field: dataclasses.Field[Any]) -> bool:
+    return (field.default is dataclasses.MISSING and field.default_factory is dataclasses.MISSING)
+
+
+def _get_dataclass_spec(config_type: type[Any]) -> _DataclassSpec:
+    spec = _DATACLASS_SPEC_CACHE.get(config_type)
+    if spec is not None:
+        return spec
+
+    spec = _DataclassSpec(
+        cls=config_type,
+        type_hints=get_type_hints(config_type),
+        fields_by_name={field.name: field
+                        for field in dataclasses.fields(config_type)},
+    )
+    _DATACLASS_SPEC_CACHE[config_type] = spec
+    return spec
+
+
+_DATACLASS_SPEC_CACHE: dict[type[Any], _DataclassSpec] = {}
+
+
+def _join_path(prefix: str, suffix: str) -> str:
+    if not prefix:
+        return suffix
+    return f"{prefix}.{suffix}"
+
+
+def _type_name(annotation: Any) -> str:
+    origin = get_origin(annotation)
+    if origin is not None:
+        return str(annotation)
+    if hasattr(annotation, "__name__"):
+        return annotation.__name__
+    return str(annotation)
+
+
+__all__ = [
+    "config_to_dict",
+    "load_config",
+    "load_raw_config",
+    "load_run_config",
+    "load_serve_config",
+    "parse_config",
+]

fastvideo/api/presets.py

@@ -0,0 +1,261 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Pipeline preset registry.
+
+A *preset* is a named inference preset for a model family. It bundles:
+
+  * ``defaults`` — sampling values applied when the user does not
+    override them (consumed at runtime via ``SamplingParam.from_pretrained``);
+  * ``stage_schemas`` — **validation-only** metadata describing which
+    user-facing stage names (``"denoise"``, ``"sr"``) the preset recognises
+    and which ``stage_overrides`` keys each stage accepts.
+
+The ``stage_schemas`` tuple does **not** drive pipeline execution. The
+concrete execution DAG (text encoding, denoising, VAE decoding, …) is
+hard-coded per-pipeline in ``create_pipeline_stages()``. Schemas exist
+purely so that ``PipelineSelection.preset`` and
+``GenerationRequest.stage_overrides`` can be type-checked up front
+without touching the pipeline.
+
+Preset base types and the registry API live here (public API surface).
+Preset *instances* are defined in pipeline-local ``presets.py`` files
+(e.g. ``fastvideo/pipelines/basic/wan/presets.py``) and registered
+explicitly from :func:`_register_presets` in ``fastvideo/registry.py``.
+"""
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from typing import Any
+
+from fastvideo.api.errors import ConfigValidationError
+
+# -------------------------------------------------------------------
+# Types
+# -------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class PresetStageSpec:
+    """A user-facing stage name within a preset, used only to validate
+    ``stage_overrides`` keys. Not read by pipeline execution — the real
+    execution DAG lives in each pipeline's ``create_pipeline_stages()``.
+    """
+
+    name: str
+    """Short user-facing name, e.g. ``"denoise"``, ``"sr"``."""
+
+    kind: str
+    """Semantic kind, e.g. ``"denoising"``, ``"super_resolution"``."""
+
+    description: str = ""
+
+    allowed_overrides: frozenset[str] = field(default_factory=frozenset)
+    """Keys that may appear in ``stage_overrides[name]``."""
+
+
+@dataclass(frozen=True)
+class InferencePreset:
+    """A named inference preset for a model family."""
+
+    name: str
+    """Preset name, e.g. ``"wan_t2v_1_3b"``."""
+
+    version: int
+    """Preset schema version; bump on breaking schema changes."""
+
+    model_family: str
+    """Model family key, e.g. ``"wan"``, ``"ltx2"``."""
+
+    description: str = ""
+
+    workload_type: str | None = None
+    """Optional workload hint: ``"t2v"``, ``"i2v"``, etc."""
+
+    stage_schemas: tuple[PresetStageSpec, ...] = ()
+    """User-facing stage names for ``stage_overrides`` validation.
+
+    Validation-only: this tuple is consumed by
+    :func:`validate_stage_overrides` and is **not** used to drive
+    pipeline execution. Omit or leave empty if the preset exposes no
+    per-stage override surface.
+    """
+
+    defaults: dict[str, Any] = field(default_factory=dict)
+    """Preset-level default sampling/runtime values."""
+
+    stage_defaults: dict[str, dict[str, Any]] = field(default_factory=dict)
+    """Per-stage default overrides, keyed by stage name."""
+
+
+# -------------------------------------------------------------------
+# Registry
+# -------------------------------------------------------------------
+
+# Keyed by (model_family, name, version).
+_PRESET_REGISTRY: dict[tuple[str, str, int], InferencePreset] = {}
+
+
+def register_preset(preset: InferencePreset) -> None:
+    """Register a preset definition.
+
+    Raises :class:`ValueError` on duplicate
+    ``(model_family, name, version)`` keys.
+    """
+    key = (preset.model_family, preset.name, preset.version)
+    if key in _PRESET_REGISTRY:
+        raise ValueError(f"Duplicate preset registration: "
+                         f"model_family={key[0]!r}, name={key[1]!r}, "
+                         f"version={key[2]!r}")
+    _PRESET_REGISTRY[key] = preset
+
+
+def get_preset(
+    name: str,
+    model_family: str,
+    version: int | None = None,
+) -> InferencePreset:
+    """Look up a registered preset.
+
+    When *version* is ``None`` the highest registered version for the
+    given *(model_family, name)* pair is returned.
+
+    Raises :class:`~fastvideo.api.errors.ConfigValidationError` when the
+    preset cannot be found.
+    """
+    if version is not None:
+        key = (model_family, name, version)
+        preset = _PRESET_REGISTRY.get(key)
+        if preset is not None:
+            return preset
+        raise ConfigValidationError(
+            "pipeline.preset",
+            f"unknown preset {name!r} version {version!r} "
+            f"for model family {model_family!r}; "
+            f"registered: {_format_registered(model_family)}",
+        )
+
+    # Find the highest version for (model_family, name).
+    candidates = [prof for (fam, n, _v), prof in _PRESET_REGISTRY.items() if fam == model_family and n == name]
+    if not candidates:
+        raise ConfigValidationError(
+            "pipeline.preset",
+            f"unknown preset {name!r} for model family "
+            f"{model_family!r}; "
+            f"registered: {_format_registered(model_family)}",
+        )
+    return max(candidates, key=lambda p: p.version)
+
+
+def get_presets_for_family(model_family: str, ) -> list[InferencePreset]:
+    """Return all presets registered for *model_family*."""
+    return [prof for (fam, _n, _v), prof in _PRESET_REGISTRY.items() if fam == model_family]
+
+
+def get_all_preset_names() -> list[str]:
+    """Return the sorted list of all registered preset names."""
+    return sorted({prof.name for prof in _PRESET_REGISTRY.values()})
+
+
+# -------------------------------------------------------------------
+# Validation helpers
+# -------------------------------------------------------------------
+
+
+def validate_stage_names(
+    preset: InferencePreset,
+    stage_overrides: Mapping[str, Any],
+) -> None:
+    """Check that *stage_overrides* keys are valid stage names.
+
+    Raises :class:`~fastvideo.api.errors.ConfigValidationError` with a
+    path-qualified message for unknown stage names.
+    """
+    valid_names = {stage.name for stage in preset.stage_schemas}
+    for stage_name in stage_overrides:
+        if stage_name not in valid_names:
+            raise ConfigValidationError(
+                f"stage_overrides.{stage_name}",
+                f"unknown stage for preset {preset.name!r}; "
+                f"valid stages: {sorted(valid_names)}",
+            )
+
+
+def validate_stage_overrides(
+    preset: InferencePreset,
+    stage_overrides: Mapping[str, Any],
+) -> None:
+    """Validate stage override keys against the preset.
+
+    Calls :func:`validate_stage_names` first, then checks that each
+    override key is in the stage's ``allowed_overrides``.
+    """
+    validate_stage_names(preset, stage_overrides)
+    stages_by_name = {stage.name: stage for stage in preset.stage_schemas}
+    for stage_name, overrides in stage_overrides.items():
+        if not isinstance(overrides, Mapping):
+            raise ConfigValidationError(
+                f"stage_overrides.{stage_name}",
+                "must be a mapping",
+            )
+        stage_spec = stages_by_name[stage_name]
+        if not stage_spec.allowed_overrides:
+            if overrides:
+                raise ConfigValidationError(
+                    f"stage_overrides.{stage_name}",
+                    f"stage {stage_name!r} does not accept "
+                    f"overrides",
+                )
+            continue
+        for key in overrides:
+            if key not in stage_spec.allowed_overrides:
+                raise ConfigValidationError(
+                    f"stage_overrides.{stage_name}.{key}",
+                    f"not an allowed override for stage "
+                    f"{stage_name!r}; allowed: "
+                    f"{sorted(stage_spec.allowed_overrides)}",
+                )
+
+
+def validate_preset_selection(
+    preset_name: str | None,
+    model_family: str,
+    *,
+    preset_version: int | None = None,
+    stage_overrides: Mapping[str, Any] | None = None,
+) -> InferencePreset | None:
+    """Resolve and validate a preset selection end-to-end.
+
+    Returns the resolved :class:`InferencePreset`, or ``None`` if
+    *preset_name* is ``None`` (no preset requested).
+    """
+    if preset_name is None:
+        return None
+    preset = get_preset(preset_name, model_family, version=preset_version)
+    if stage_overrides:
+        validate_stage_overrides(preset, stage_overrides)
+    return preset
+
+
+# -------------------------------------------------------------------
+# Internal helpers
+# -------------------------------------------------------------------
+
+
+def _format_registered(model_family: str) -> str:
+    names = sorted({prof.name for (fam, _n, _v), prof in _PRESET_REGISTRY.items() if fam == model_family})
+    if not names:
+        return "(none)"
+    return ", ".join(repr(n) for n in names)
+
+
+__all__ = [
+    "InferencePreset",
+    "PresetStageSpec",
+    "get_all_preset_names",
+    "get_preset",
+    "get_presets_for_family",
+    "register_preset",
+    "validate_preset_selection",
+    "validate_stage_names",
+    "validate_stage_overrides",
+]

fastvideo/api/request_metadata.py

@@ -0,0 +1,233 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Track which GenerationRequest fields the user explicitly provided.
+
+When translating a GenerationRequest into a legacy SamplingParam we must
+distinguish user-provided values (which should override model defaults)
+from schema defaults (which should NOT override model defaults).
+
+The mechanism: a single ``_fastvideo_explicit_paths`` set stored on the
+root ``GenerationRequest``. It holds dotted leaf paths (e.g.
+``"sampling.guidance_scale"``) the user has touched, either via raw
+config at bind time or via attribute assignment at runtime. A patched
+``__setattr__`` on the request dataclass types records assignments into
+this set.
+
+The set holds leaf paths only. Nested dataclass or mapping assignments
+are flattened to their leaves at record time.
+"""
+from __future__ import annotations
+
+from collections.abc import Callable, Mapping
+import dataclasses
+from typing import Any, cast
+
+from fastvideo.api.schema import (
+    ContinuationState,
+    GenerationPlan,
+    GenerationRequest,
+    InputConfig,
+    OutputConfig,
+    PlannedStage,
+    RequestRuntimeConfig,
+    RunConfig,
+    SamplingConfig,
+    ServeConfig,
+)
+
+EXPLICIT_PATHS_ATTR = "_fastvideo_explicit_paths"
+
+_TRACKING_ROOT_ATTR = "_fastvideo_request_tracking_root"
+_TRACKING_PATH_ATTR = "_fastvideo_request_tracking_path"
+_TRACKING_PATCHED_ATTR = "_fastvideo_request_tracking_patched"
+
+_TRACKED_REQUEST_TYPES = (
+    GenerationRequest,
+    InputConfig,
+    SamplingConfig,
+    RequestRuntimeConfig,
+    OutputConfig,
+    ContinuationState,
+    PlannedStage,
+    GenerationPlan,
+)
+
+
+def bind_generation_request_raw(
+    request: GenerationRequest,
+    raw: Mapping[str, Any] | None,
+) -> GenerationRequest:
+    """Install explicit-path tracking on *request*.
+
+    *raw* is the parsed config dict (YAML/JSON/kwargs); every leaf key
+    in it becomes an explicit path. Subsequent attribute assignments on
+    *request* or its nested dataclasses are recorded automatically via a
+    patched ``__setattr__``.
+    """
+    _ensure_request_tracking()
+    # Disable recording while we walk the tree to install roots.
+    object.__setattr__(request, EXPLICIT_PATHS_ATTR, None)
+    _set_tracking_roots(request, request, "")
+    paths: set[str] = set()
+    _record_value_paths(raw or {}, "", paths)
+    object.__setattr__(request, EXPLICIT_PATHS_ATTR, paths)
+    return request
+
+
+def bind_run_config_raw(
+    config: RunConfig,
+    raw: Mapping[str, Any],
+) -> RunConfig:
+    request_raw = raw.get("request")
+    if isinstance(request_raw, Mapping):
+        bind_generation_request_raw(config.request, request_raw)
+    else:
+        bind_generation_request_raw(config.request, {})
+    return config
+
+
+def bind_serve_config_raw(
+    config: ServeConfig,
+    raw: Mapping[str, Any],
+) -> ServeConfig:
+    default_request_raw = raw.get("default_request")
+    if isinstance(default_request_raw, Mapping):
+        bind_generation_request_raw(config.default_request, default_request_raw)
+    else:
+        bind_generation_request_raw(config.default_request, {})
+    return config
+
+
+def get_explicit_paths(request: GenerationRequest) -> frozenset[str]:
+    """Return a snapshot of the explicit paths set on *request*."""
+    paths = getattr(request, EXPLICIT_PATHS_ATTR, None)
+    if isinstance(paths, set | frozenset):
+        return frozenset(paths)
+    return frozenset()
+
+
+def reset_tracking_roots(request: GenerationRequest) -> None:
+    """Re-install tracking roots after a deepcopy or manual clone.
+
+    The paths set itself deepcopies correctly; we only need to repoint
+    the tracking root on nested dataclasses at the new root.
+    """
+    _ensure_request_tracking()
+    _set_tracking_roots(request, request, "")
+
+
+# ---------------------------------------------------------------------------
+# Path recording
+# ---------------------------------------------------------------------------
+
+
+def _record_value_paths(
+    value: Any,
+    prefix: str,
+    out: set[str],
+) -> None:
+    """Add every leaf path under *value* to *out*.
+
+    A leaf is any terminal value (non-dataclass, non-mapping, or empty
+    mapping/dataclass). ``prefix`` is the dotted path at which *value*
+    sits. When called with an empty ``prefix`` (the root), leaves are
+    recorded at their own key.
+    """
+    if dataclasses.is_dataclass(value) and not isinstance(value, type):
+        dc_fields = dataclasses.fields(value)
+        if not dc_fields:
+            if prefix:
+                out.add(prefix)
+            return
+        for field in dc_fields:
+            child = getattr(value, field.name)
+            path = f"{prefix}.{field.name}" if prefix else field.name
+            _record_value_paths(child, path, out)
+        return
+    if isinstance(value, Mapping):
+        if not value:
+            if prefix:
+                out.add(prefix)
+            return
+        for key, child in value.items():
+            path = f"{prefix}.{key}" if prefix else key
+            _record_value_paths(child, path, out)
+        return
+    if prefix:
+        out.add(prefix)
+
+
+# ---------------------------------------------------------------------------
+# __setattr__ patching
+# ---------------------------------------------------------------------------
+
+
+def _ensure_request_tracking() -> None:
+    for config_type in _TRACKED_REQUEST_TYPES:
+        _patch_tracking_setattr(config_type)
+
+
+def _patch_tracking_setattr(config_type: type[Any]) -> None:
+    if getattr(config_type, _TRACKING_PATCHED_ATTR, False):
+        return
+
+    original_setattr = cast(
+        Callable[[Any, str, Any], None],
+        config_type.__setattr__,
+    )
+    field_names = {field.name for field in dataclasses.fields(config_type)}
+
+    def _tracking_setattr(self: Any, name: str, value: Any) -> None:
+        if name.startswith("_fastvideo_") or name not in field_names:
+            original_setattr(self, name, value)
+            return
+
+        original_setattr(self, name, value)
+
+        root = getattr(self, _TRACKING_ROOT_ATTR, None)
+        if root is None:
+            return
+        paths = getattr(root, EXPLICIT_PATHS_ATTR, None)
+        if not isinstance(paths, set):
+            return
+
+        prefix = getattr(self, _TRACKING_PATH_ATTR, "")
+        path = f"{prefix}.{name}" if prefix else name
+        # Wholesale dataclass replacement: install roots on the new
+        # instance so its future mutations are tracked too.
+        if dataclasses.is_dataclass(value) and not isinstance(value, type):
+            _set_tracking_roots(root, value, path)
+        _record_value_paths(value, path, paths)
+
+    type.__setattr__(config_type, "__setattr__", _tracking_setattr)
+    setattr(config_type, _TRACKING_PATCHED_ATTR, True)
+
+
+# ---------------------------------------------------------------------------
+# Tree walk to set tracking root/path on nested dataclasses
+# ---------------------------------------------------------------------------
+
+
+def _set_tracking_roots(
+    root: GenerationRequest,
+    obj: Any,
+    prefix: str,
+) -> None:
+    if not dataclasses.is_dataclass(obj) or isinstance(obj, type):
+        return
+    object.__setattr__(obj, _TRACKING_ROOT_ATTR, root)
+    object.__setattr__(obj, _TRACKING_PATH_ATTR, prefix)
+    for field in dataclasses.fields(obj):
+        child = getattr(obj, field.name)
+        child_path = f"{prefix}.{field.name}" if prefix else field.name
+        if dataclasses.is_dataclass(child) and not isinstance(child, type):
+            _set_tracking_roots(root, child, child_path)
+
+
+__all__ = [
+    "EXPLICIT_PATHS_ATTR",
+    "bind_generation_request_raw",
+    "bind_run_config_raw",
+    "bind_serve_config_raw",
+    "get_explicit_paths",
+    "reset_tracking_roots",
+]

fastvideo/api/results.py

@@ -0,0 +1,101 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+from collections.abc import Mapping
+
+from fastvideo.api.schema import ContinuationState
+
+
+@dataclass
+class GenerationResult:
+    prompt: str | None = None
+    prompt_index: int | None = None
+    samples: Any | None = None
+    frames: Any | None = None
+    audio: Any | None = None
+    size: tuple[int, int, int] | None = None
+    generation_time: float | None = None
+    logging_info: Any | None = None
+    trajectory: Any | None = None
+    trajectory_timesteps: Any | None = None
+    trajectory_decoded: Any | None = None
+    video_path: str | None = None
+    peak_memory_mb: float | None = None
+    state: ContinuationState | None = None
+    extra: dict[str, Any] = field(default_factory=dict)
+
+    @classmethod
+    def from_legacy_result(
+        cls,
+        result: Mapping[str, Any],
+    ) -> GenerationResult:
+        prompt = result.get("prompt")
+        if prompt is None:
+            prompt = result.get("prompts")
+
+        extra = {
+            key: value
+            for key, value in result.items() if key not in {
+                "prompt",
+                "prompt_index",
+                "prompts",
+                "samples",
+                "frames",
+                "audio",
+                "size",
+                "generation_time",
+                "logging_info",
+                "trajectory",
+                "trajectory_timesteps",
+                "trajectory_decoded",
+                "video_path",
+                "peak_memory_mb",
+                "state",
+            }
+        }
+
+        return cls(
+            prompt=prompt,
+            prompt_index=result.get("prompt_index"),
+            samples=result.get("samples"),
+            frames=result.get("frames"),
+            audio=result.get("audio"),
+            size=result.get("size"),
+            generation_time=result.get("generation_time"),
+            logging_info=result.get("logging_info"),
+            trajectory=result.get("trajectory"),
+            trajectory_timesteps=result.get("trajectory_timesteps"),
+            trajectory_decoded=result.get("trajectory_decoded"),
+            video_path=result.get("video_path"),
+            peak_memory_mb=result.get("peak_memory_mb"),
+            state=result.get("state"),
+            extra=extra,
+        )
+
+    def to_legacy_dict(self) -> dict[str, Any]:
+        result = {
+            "prompts": self.prompt,
+            "samples": self.samples,
+            "frames": self.frames,
+            "audio": self.audio,
+            "size": self.size,
+            "generation_time": self.generation_time,
+            "logging_info": self.logging_info,
+            "trajectory": self.trajectory,
+            "trajectory_timesteps": self.trajectory_timesteps,
+            "trajectory_decoded": self.trajectory_decoded,
+            "video_path": self.video_path,
+            "peak_memory_mb": self.peak_memory_mb,
+        }
+        if self.prompt_index is not None:
+            result["prompt_index"] = self.prompt_index
+            result["prompt"] = self.prompt
+        if self.state is not None:
+            result["state"] = self.state
+        result.update(self.extra)
+        return result
+
+
+__all__ = ["GenerationResult"]

fastvideo/api/sampling_param.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
+import copy
+from dataclasses import dataclass, field, fields
 from typing import Any
 
 from fastvideo.logger import init_logger
@@ -30,6 +31,16 @@ class SamplingParam:
 
     # Camera control inputs (HYWorld)
     pose: str | None = None  # Camera trajectory: pose string (e.g., 'w-31') or JSON file path
+    prompt_attention_mask: list = field(default_factory=list)
+    negative_attention_mask: list = field(default_factory=list)
+
+    # Camera/action control inputs (GameCraft)
+    camera_states: Any | None = None  # Plücker coordinates [B, T_video, 6, H, W]
+    camera_trajectory: str | None = None
+    action_list: list[str] | None = None
+    action_speed_list: list[float] | None = None
+    gt_latents: Any | None = None  # Ground truth latents [B, 16, T, H, W]
+    conditioning_mask: Any | None = None  # Mask [B, 1, T, H, W]
 
     # Camera control inputs (LingBotWorld)
     c2ws_plucker_emb: Any | None = None  # Plucker embedding: [B, C, F_lat, H_lat, W_lat]
@@ -68,10 +79,30 @@ class SamplingParam:
     num_inference_steps: int = 50
     num_inference_steps_sr: int = 50
     guidance_scale: float = 1.0
+    guidance_scale_2: float | None = None
     guidance_rescale: float = 0.0
     boundary_ratio: float | None = None
     sigmas: list[float] | None = None
 
+    # TeaCache parameters
+    enable_teacache: bool = False
+
+    # GEN3C camera control
+    trajectory_type: str | None = None
+    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
+    ltx2_stg_blocks_video: list[int] = field(default_factory=lambda: [29])
+    ltx2_stg_blocks_audio: list[int] = field(default_factory=lambda: [29])
+
     # Misc
     save_video: bool = True
     return_frames: bool = True
@@ -86,26 +117,58 @@ class SamplingParam:
             raise ValueError("prompt_path must be a txt file")
 
     def update(self, source_dict: dict[str, Any]) -> None:
+        valid_fields = {f.name for f in fields(self)}
         for key, value in source_dict.items():
-            if hasattr(self, key):
+            if key in valid_fields:
                 setattr(self, key, value)
             else:
-                logger.exception("%s has no attribute %s", type(self).__name__, key)
+                logger.error("%s has no field %s", type(self).__name__, key)
 
         self.__post_init__()
 
     @classmethod
     def from_pretrained(cls, model_path: str) -> "SamplingParam":
-        from fastvideo.registry import get_sampling_param_cls_for_name
-        sampling_cls = get_sampling_param_cls_for_name(model_path)
-        if sampling_cls is not None:
-            sampling_param: SamplingParam = sampling_cls()
-        else:
-            logger.warning("Couldn't find an optimal sampling param for %s. Using the default sampling param.",
-                           model_path)
-            sampling_param = cls()
+        sampling_param = cls._from_preset(model_path)
+        if sampling_param is not None:
+            return sampling_param
 
-        return sampling_param
+        logger.warning(
+            "Couldn't find a preset for %s."
+            " Using the default sampling param.",
+            model_path,
+        )
+        return cls()
+
+    @classmethod
+    def _from_preset(
+        cls,
+        model_path: str,
+    ) -> "SamplingParam | None":
+        """Build a SamplingParam from preset defaults.
+
+        Returns ``None`` when no preset is configured for
+        *model_path*, letting the caller fall back to the legacy
+        subclass lookup.
+        """
+        from fastvideo.registry import get_preset_selection
+
+        try:
+            preset_name, model_family = get_preset_selection(model_path)
+        except (ValueError, RuntimeError):
+            return None
+        if preset_name is None or model_family is None:
+            return None
+
+        from fastvideo.api.presets import get_preset
+
+        preset = get_preset(preset_name, model_family)
+        sp = cls()
+        valid_fields = {f.name for f in fields(cls)}
+        for key, value in preset.defaults.items():
+            if key in valid_fields:
+                setattr(sp, key, copy.deepcopy(value))
+        sp.__post_init__()
+        return sp
 
     @staticmethod
     def add_cli_args(parser: Any) -> Any:

fastvideo/api/schema.py

@@ -0,0 +1,290 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Literal
+
+
+@dataclass
+class ServerConfig:
+    host: str = "0.0.0.0"
+    port: int = 8000
+    output_dir: str = "outputs/"
+
+
+@dataclass
+class ParallelismConfig:
+    tp_size: int = -1
+    sp_size: int = -1
+    hsdp_replicate_dim: int = 1
+    hsdp_shard_dim: int = -1
+    dist_timeout: int | None = None
+
+
+@dataclass
+class OffloadConfig:
+    dit: bool = True
+    dit_layerwise: bool = True
+    text_encoder: bool = True
+    image_encoder: bool = True
+    vae: bool = True
+    pin_cpu_memory: bool = True
+
+
+@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``).
+    """
+
+    enabled: bool = False
+    backend: str | None = None
+    fullgraph: bool | None = None
+    mode: str | None = None
+    dynamic: bool | None = None
+    extras: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class QuantizationConfig:
+    text_encoder_quant: str | None = None
+    transformer_quant: str | None = None
+
+
+@dataclass
+class EngineConfig:
+    num_gpus: int = 1
+    execution_backend: Literal["mp", "ray"] = "mp"
+    parallelism: ParallelismConfig = field(default_factory=ParallelismConfig)
+    offload: OffloadConfig = field(default_factory=OffloadConfig)
+    compile: CompileConfig = field(default_factory=CompileConfig)
+    enable_stage_verification: bool = True
+    use_fsdp_inference: bool = False
+    disable_autocast: bool = False
+    quantization: QuantizationConfig | None = None
+
+
+@dataclass
+class ComponentConfig:
+    config_root: str | None = None
+    pipeline_config_path: str | None = None
+    text_encoder_weights: str | None = None
+    transformer_weights: str | None = None
+    transformer_2_weights: str | None = None
+    vae_weights: str | None = None
+    upsampler_weights: str | None = None
+    lora_path: str | None = None
+    override_pipeline_cls_name: str | None = None
+    override_transformer_cls_name: str | None = None
+
+
+@dataclass
+class PipelineSelection:
+    workload_type: Literal["t2v", "i2v", "t2i", "i2i"] | None = None
+    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)
+
+
+@dataclass
+class GeneratorConfig:
+    model_path: str
+    revision: str | None = None
+    trust_remote_code: bool = False
+    engine: EngineConfig = field(default_factory=EngineConfig)
+    pipeline: PipelineSelection = field(default_factory=PipelineSelection)
+
+
+@dataclass
+class InputConfig:
+    prompt_path: str | None = None
+    image_path: str | list[str] | None = None
+    video_path: str | list[str] | None = None
+    pil_image: Any | None = None
+    pose: str | None = None
+    mouse_cond: Any | None = None
+    keyboard_cond: Any | None = None
+    grid_sizes: Any | None = None
+    c2ws_plucker_emb: Any | None = None
+    refine_from: str | None = None
+    stage1_video: Any | None = None
+
+
+@dataclass
+class SamplingConfig:
+    num_videos_per_prompt: int = 1
+    seed: int = 1024
+    num_frames: int = 125
+    height: int = 720
+    width: int = 1280
+    height_sr: int = 1072
+    width_sr: int = 1920
+    fps: int = 24
+    num_inference_steps: int = 50
+    num_inference_steps_sr: int = 50
+    guidance_scale: float = 1.0
+    guidance_scale_2: float | None = None
+    guidance_rescale: float = 0.0
+    true_cfg_scale: float | None = None
+    boundary_ratio: float | None = None
+    sigmas: list[float] | None = None
+
+
+@dataclass
+class RequestRuntimeConfig:
+    enable_teacache: bool = False
+    return_trajectory_latents: bool = False
+    return_trajectory_decoded: bool = False
+
+
+@dataclass
+class OutputConfig:
+    output_path: str = "outputs/"
+    output_video_name: str | None = None
+    save_video: bool = True
+    return_frames: bool = True
+    return_state: bool = False
+
+
+@dataclass
+class ContinuationState:
+    kind: str
+    payload: dict[str, Any]
+
+
+@dataclass
+class PlannedStage:
+    name: str
+    kind: str
+    source: str | None = None
+    overrides: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class GenerationPlan:
+    stages: list[PlannedStage]
+    final_stage: str | None = None
+
+
+@dataclass
+class GenerationRequest:
+    prompt: str | list[str] | None = None
+    negative_prompt: str | None = None
+    inputs: InputConfig = field(default_factory=InputConfig)
+    sampling: SamplingConfig = field(default_factory=SamplingConfig)
+    runtime: RequestRuntimeConfig = field(default_factory=RequestRuntimeConfig)
+    output: OutputConfig = field(default_factory=OutputConfig)
+    stage_overrides: dict[str, Any] = field(default_factory=dict)
+    state: ContinuationState | None = None
+    plan: GenerationPlan | None = None
+    extensions: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class RunConfig:
+    generator: GeneratorConfig
+    request: GenerationRequest
+
+
+@dataclass
+class WarmupConfig:
+    enabled: bool = True
+    prompt: str = ("A cinematic drone shot over coastal cliffs at sunrise, "
+                   "golden light, gentle ocean waves, ultra detailed")
+    timeout_seconds: int = 2400
+
+
+@dataclass
+class GpuPoolConfig:
+    num_workers: int | None = None
+    enable_audio_reencode: bool = True
+    conditioning_num_frames: int = 9
+    conditioning_end_offset: int = 0
+
+
+@dataclass
+class PromptEnhancerConfig:
+    enabled: bool = False
+    provider: Literal["cerebras", "groq"] = "cerebras"
+    model: str = "gpt-oss-120b"
+    timeout_ms: int = 20000
+    system_prompt_dir: str | None = None
+
+
+@dataclass
+class PromptSafetyConfig:
+    enabled: bool = False
+    classifier_path: str | None = None
+
+
+@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 = field(default_factory=PromptEnhancerConfig)
+    safety: PromptSafetyConfig = field(default_factory=PromptSafetyConfig)
+
+
+@dataclass
+class ServeConfig:
+    """Typed serve config loaded from ``fastvideo serve --config``.
+
+    ``default_request`` is a full :class:`GenerationRequest` — the same type
+    clients POST to ``/v1/videos``. At request time the server merges it into
+    the incoming body as the operator-pinned baseline.
+
+    Important nuance: only fields the operator **explicitly wrote** in the
+    serve YAML/JSON count as defaults. Although the in-memory object is
+    fully populated (schema defaults fill every unset field), the merge
+    walks ``_fastvideo_explicit_paths`` — populated during parse — so
+    unset fields are *not* forced onto requests. Per-request precedence:
+
+        body (client-explicit) > default_request (operator-explicit)
+                               > hardcoded fallback (e.g. ``fps=24``)
+
+    See :func:`fastvideo.api.compat.explicit_request_updates` for the
+    projection and ``entrypoints/openai/video_api.py::_build_generation_kwargs``
+    for the merge.
+    """
+    generator: GeneratorConfig
+    server: ServerConfig = field(default_factory=ServerConfig)
+    default_request: GenerationRequest = field(default_factory=GenerationRequest)
+    streaming: StreamingConfig | None = None
+
+
+__all__ = [
+    "CompileConfig",
+    "ComponentConfig",
+    "ContinuationState",
+    "EngineConfig",
+    "GenerationPlan",
+    "GenerationRequest",
+    "GeneratorConfig",
+    "GpuPoolConfig",
+    "InputConfig",
+    "OffloadConfig",
+    "OutputConfig",
+    "ParallelismConfig",
+    "PipelineSelection",
+    "PlannedStage",
+    "PromptEnhancerConfig",
+    "PromptSafetyConfig",
+    "QuantizationConfig",
+    "RequestRuntimeConfig",
+    "RunConfig",
+    "SamplingConfig",
+    "ServeConfig",
+    "ServerConfig",
+    "StreamingConfig",
+    "WarmupConfig",
+]

fastvideo/attention/backends/bsa_attn.py

@@ -0,0 +1,738 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Bidirectional Sparse Attention (BSA) backend for FastVideo.
+
+Pure-PyTorch reference implementation from:
+"Bidirectional Sparse Attention for Faster Video Diffusion Training"
+(arXiv:2509.01085)
+
+BSA sparsifies both queries (pruning redundant tokens per block) and
+key-value pairs (keeping only relevant KV blocks per query block).
+
+This is a training-free inference backend: it works with any model
+trained with full attention by applying BSA sparsity at inference time.
+"""
+
+import functools
+import math
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+
+from fastvideo.attention.backends.abstract import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+from fastvideo.distributed import get_sp_group
+from fastvideo.logger import init_logger
+
+try:
+    from fastvideo.attention.utils.flash_attn_no_pad import (
+        flash_attn_varlen_func_impl, )
+    FLASH_ATTN_AVAILABLE = True
+except ImportError:
+    try:
+        from flash_attn import flash_attn_varlen_func as flash_attn_varlen_func_impl
+        FLASH_ATTN_AVAILABLE = True
+    except ImportError:
+        FLASH_ATTN_AVAILABLE = False
+
+logger = init_logger(__name__)
+
+BSA_TILE_SIZE = (4, 4, 4)
+
+# ---------------------------------------------------------------------------
+# Cached index helpers (same pattern as VSA)
+# ---------------------------------------------------------------------------
+
+
+@functools.lru_cache(maxsize=10)
+def get_tile_partition_indices(
+    dit_seq_shape: tuple[int, int, int],
+    tile_size: tuple[int, int, int],
+    device: torch.device,
+) -> torch.LongTensor:
+    """Map raster-order tokens to tile-contiguous order."""
+    T, H, W = dit_seq_shape
+    ts, hs, ws = tile_size
+    indices = torch.arange(T * H * W, device=device, dtype=torch.long).reshape(T, H, W)
+    ls = []
+    for t in range(math.ceil(T / ts)):
+        for h in range(math.ceil(H / hs)):
+            for w in range(math.ceil(W / ws)):
+                ls.append(indices[
+                    t * ts:min(t * ts + ts, T),
+                    h * hs:min(h * hs + hs, H),
+                    w * ws:min(w * ws + ws, W),
+                ].flatten())
+    return torch.cat(ls, dim=0)
+
+
+@functools.lru_cache(maxsize=10)
+def get_reverse_tile_partition_indices(
+    dit_seq_shape: tuple[int, int, int],
+    tile_size: tuple[int, int, int],
+    device: torch.device,
+) -> torch.LongTensor:
+    """Inverse mapping: tile-contiguous order back to raster order."""
+    return torch.argsort(get_tile_partition_indices(dit_seq_shape, tile_size, device))
+
+
+# ---------------------------------------------------------------------------
+# BSA core operations
+# ---------------------------------------------------------------------------
+
+
+def _prune_queries(
+    q_blocks: torch.Tensor,
+    keep_ratio: float,
+) -> tuple[torch.Tensor, torch.Tensor, int]:
+    """
+    Prune redundant query tokens within each block.
+
+    Scores tokens by cosine similarity to the block center.
+    Keeps the LEAST similar (most informative) tokens.
+
+    Args:
+        q_blocks: [B, N_heads, N_blocks, block_size, D]
+        keep_ratio: fraction of tokens to keep
+
+    Returns:
+        sparse_q: [B, N_heads, N_blocks, keep_size, D]
+        keep_indices: [B, N_heads, N_blocks, keep_size]
+        keep_size: int
+    """
+    B, H, N, S, D = q_blocks.shape
+    keep_size = max(1, int(S * keep_ratio))
+
+    if keep_size >= S:
+        idx = torch.arange(S, device=q_blocks.device)
+        idx = idx.view(1, 1, 1, S).expand(B, H, N, S)
+        return q_blocks, idx, S
+
+    center_idx = S // 2
+    center = q_blocks[:, :, :, center_idx:center_idx + 1, :]
+
+    q_norm = F.normalize(q_blocks, dim=-1)
+    c_norm = F.normalize(center, dim=-1)
+    similarity = (q_norm * c_norm).sum(dim=-1)  # [B, H, N, S]
+
+    # lowest similarity = most distinctive = keep
+    _, indices = similarity.topk(keep_size, dim=-1, largest=False)
+    indices, _ = indices.sort(dim=-1)
+
+    idx_expand = indices.unsqueeze(-1).expand(-1, -1, -1, -1, D)
+    sparse_q = torch.gather(q_blocks, 3, idx_expand)
+
+    return sparse_q, indices, keep_size
+
+
+def _select_kv_blocks(
+    sparse_q: torch.Tensor,
+    k_blocks: torch.Tensor,
+    cumulative_threshold: float,
+    min_kv_blocks: int,
+) -> torch.Tensor:
+    """
+    Dynamically select KV blocks for each query block.
+
+    Mean-pools to block level, computes block attention scores,
+    admits blocks in descending order until cumulative mass
+    exceeds threshold.
+
+    Args:
+        sparse_q: [B, H, N, Sq, D]
+        k_blocks:  [B, H, N, Sk, D]
+        cumulative_threshold: e.g. 0.9
+        min_kv_blocks: minimum blocks to keep
+
+    Returns:
+        kv_mask: [B, H, N, N] boolean
+    """
+    B, H, N, _, D = sparse_q.shape
+
+    q_repr = sparse_q.mean(dim=3)
+    k_repr = k_blocks.mean(dim=3)
+
+    scores = torch.matmul(q_repr, k_repr.transpose(-1, -2)) / (D**0.5)
+    block_attn = F.softmax(scores, dim=-1)
+
+    sorted_attn, sorted_idx = block_attn.sort(dim=-1, descending=True)
+    cumsum = sorted_attn.cumsum(dim=-1)
+
+    keep_sorted = torch.ones_like(cumsum, dtype=torch.bool)
+    keep_sorted[..., 1:] = cumsum[..., :-1] < cumulative_threshold
+
+    min_mask = torch.zeros_like(keep_sorted)
+    min_mask[..., :min(min_kv_blocks, N)] = True
+    keep_sorted = keep_sorted | min_mask
+
+    kv_mask = torch.zeros_like(block_attn, dtype=torch.bool)
+    kv_mask.scatter_(-1, sorted_idx, keep_sorted)
+
+    return kv_mask
+
+
+def _compute_sparse_attention(
+    sparse_q: torch.Tensor,
+    k_blocks: torch.Tensor,
+    v_blocks: torch.Tensor,
+    kv_mask: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Compute attention for each query block against selected KV blocks.
+
+    Handles per-batch and per-head KV masks correctly.
+    Uses flash_attn_varlen_func when available on GPU.
+    Falls back to pure-PyTorch reference on CPU.
+
+    Args:
+        sparse_q: [B, H, N, Sq, D]
+        k_blocks:  [B, H, N, Sk, D]
+        v_blocks:  [B, H, N, Sk, D]
+        kv_mask:   [B, H, N, N] boolean (per-batch, per-head)
+
+    Returns:
+        output: [B, H, N, Sq, D]
+    """
+    if FLASH_ATTN_AVAILABLE and sparse_q.is_cuda:
+        return _compute_sparse_attention_flash(sparse_q, k_blocks, v_blocks, kv_mask)
+    else:
+        return _compute_sparse_attention_reference(sparse_q, k_blocks, v_blocks, kv_mask)
+
+
+def _compute_sparse_attention_reference(
+    sparse_q: torch.Tensor,
+    k_blocks: torch.Tensor,
+    v_blocks: torch.Tensor,
+    kv_mask: torch.Tensor,
+) -> torch.Tensor:
+    """Pure-PyTorch fallback with per-batch, per-head mask support."""
+    B, H, N, Sq, D = sparse_q.shape
+    output = torch.zeros_like(sparse_q)
+
+    for b in range(B):
+        for h in range(H):
+            for qb in range(N):
+                selected = kv_mask[b, h, qb]  # [N] boolean
+                sel_idx = selected.nonzero(as_tuple=True)[0]
+
+                if sel_idx.shape[0] == 0:
+                    continue
+
+                # [num_sel * Sk, D]
+                sel_k = k_blocks[b, h, sel_idx].reshape(-1, D)
+                sel_v = v_blocks[b, h, sel_idx].reshape(-1, D)
+
+                q = sparse_q[b, h, qb]  # [Sq, D]
+                scores = torch.matmul(q, sel_k.transpose(-1, -2)) / (D**0.5)
+                weights = F.softmax(scores, dim=-1)
+                output[b, h, qb] = torch.matmul(weights, sel_v)
+
+    return output
+
+
+def _compute_sparse_attention_flash(
+    sparse_q: torch.Tensor,
+    k_blocks: torch.Tensor,
+    v_blocks: torch.Tensor,
+    kv_mask: torch.Tensor,
+) -> torch.Tensor:
+    """
+    FlashAttention implementation with per-batch, per-head mask support.
+
+    Strategy: check if all heads share the same mask. If so, use a single
+    FlashAttention call per batch (fast path). If not, process each head
+    separately (correct path).
+
+    Args:
+        sparse_q: [B, H, N, Sq, D]
+        k_blocks:  [B, H, N, Sk, D]
+        v_blocks:  [B, H, N, Sk, D]
+        kv_mask:   [B, H, N, N] boolean
+
+    Returns:
+        output: [B, H, N, Sq, D]
+    """
+    B, H, N, Sq, D = sparse_q.shape
+    Sk = k_blocks.shape[3]
+    device = sparse_q.device
+    output = torch.zeros_like(sparse_q)
+
+    for b in range(B):
+        # Check if all heads share the same mask for this batch element
+        # Compare each head's mask to head 0's mask
+        head0_mask = kv_mask[b, 0]  # [N, N]
+        all_heads_same = all(torch.equal(kv_mask[b, h], head0_mask) for h in range(1, H))
+
+        if all_heads_same:
+            # Fast path: all heads share the same mask, single FA call
+            _flash_attn_single_mask(
+                sparse_q[b],
+                k_blocks[b],
+                v_blocks[b],
+                head0_mask,
+                output[b],
+                H,
+                N,
+                Sq,
+                Sk,
+                D,
+                device,
+            )
+        else:
+            # Per-head path: process each head individually
+            for h in range(H):
+                head_mask = kv_mask[b, h]  # [N, N]
+                # Process single head: squeeze head dim, run FA, put back
+                _flash_attn_single_head(
+                    sparse_q[b, h],
+                    k_blocks[b, h],
+                    v_blocks[b, h],
+                    head_mask,
+                    output,
+                    b,
+                    h,
+                    N,
+                    Sq,
+                    Sk,
+                    D,
+                    device,
+                )
+
+    return output
+
+
+def _flash_attn_single_mask(
+    sparse_q_b: torch.Tensor,  # [H, N, Sq, D]
+    k_blocks_b: torch.Tensor,  # [H, N, Sk, D]
+    v_blocks_b: torch.Tensor,  # [H, N, Sk, D]
+    mask: torch.Tensor,  # [N, N] boolean
+    output_b: torch.Tensor,  # [H, N, Sq, D] (modified in-place)
+    H: int,
+    N: int,
+    Sq: int,
+    Sk: int,
+    D: int,
+    device: torch.device,
+) -> None:
+    """Run FlashAttention for all heads sharing the same KV mask."""
+    q_list = []
+    k_list = []
+    v_list = []
+    cu_seqlens_q = [0]
+    cu_seqlens_k = [0]
+    active_blocks = []
+
+    for qb in range(N):
+        selected = mask[qb]  # [N] boolean
+        sel_idx = selected.nonzero(as_tuple=True)[0]
+
+        if sel_idx.shape[0] == 0:
+            continue
+
+        active_blocks.append(qb)
+        num_kv_tokens = sel_idx.shape[0] * Sk
+
+        # [H, Sq, D] -> [Sq, H, D]
+        q_block = sparse_q_b[:, qb].permute(1, 0, 2)
+        q_list.append(q_block)
+
+        # [H, num_sel, Sk, D] -> [num_kv_tokens, H, D]
+        sel_k = k_blocks_b[:, sel_idx].permute(1, 2, 0, 3).reshape(num_kv_tokens, H, D)
+        sel_v = v_blocks_b[:, sel_idx].permute(1, 2, 0, 3).reshape(num_kv_tokens, H, D)
+        k_list.append(sel_k)
+        v_list.append(sel_v)
+
+        cu_seqlens_q.append(cu_seqlens_q[-1] + Sq)
+        cu_seqlens_k.append(cu_seqlens_k[-1] + num_kv_tokens)
+
+    if not q_list:
+        return
+
+    flat_q = torch.cat(q_list, dim=0)
+    flat_k = torch.cat(k_list, dim=0)
+    flat_v = torch.cat(v_list, dim=0)
+
+    cu_seqlens_q_t = torch.tensor(cu_seqlens_q, dtype=torch.int32, device=device)
+    cu_seqlens_k_t = torch.tensor(cu_seqlens_k, dtype=torch.int32, device=device)
+
+    max_seqlen_q = Sq
+    max_seqlen_k = int((cu_seqlens_k_t[1:] - cu_seqlens_k_t[:-1]).max().item())
+
+    orig_dtype = flat_q.dtype
+    compute_dtype = orig_dtype
+    if compute_dtype not in (torch.float16, torch.bfloat16):
+        compute_dtype = torch.bfloat16
+        flat_q = flat_q.to(compute_dtype)
+        flat_k = flat_k.to(compute_dtype)
+        flat_v = flat_v.to(compute_dtype)
+
+    flat_out = flash_attn_varlen_func_impl(
+        flat_q,
+        flat_k,
+        flat_v,
+        cu_seqlens_q_t,
+        cu_seqlens_k_t,
+        max_seqlen_q,
+        max_seqlen_k,
+        causal=False,
+    )
+
+    if compute_dtype != orig_dtype:
+        flat_out = flat_out.to(orig_dtype)
+
+    idx = 0
+    for qb in active_blocks:
+        block_out = flat_out[idx:idx + Sq]  # [Sq, H, D]
+        output_b[:, qb] = block_out.permute(1, 0, 2)  # [H, Sq, D]
+        idx += Sq
+
+
+def _flash_attn_single_head(
+    sparse_q_bh: torch.Tensor,  # [N, Sq, D]
+    k_blocks_bh: torch.Tensor,  # [N, Sk, D]
+    v_blocks_bh: torch.Tensor,  # [N, Sk, D]
+    mask: torch.Tensor,  # [N, N] boolean
+    output: torch.Tensor,  # [B, H, N, Sq, D] (modified in-place)
+    b: int,
+    h: int,
+    N: int,
+    Sq: int,
+    Sk: int,
+    D: int,
+    device: torch.device,
+) -> None:
+    """Run FlashAttention for a single head with its own KV mask."""
+    q_list = []
+    k_list = []
+    v_list = []
+    cu_seqlens_q = [0]
+    cu_seqlens_k = [0]
+    active_blocks = []
+
+    for qb in range(N):
+        selected = mask[qb]
+        sel_idx = selected.nonzero(as_tuple=True)[0]
+
+        if sel_idx.shape[0] == 0:
+            continue
+
+        active_blocks.append(qb)
+        num_kv_tokens = sel_idx.shape[0] * Sk
+
+        # [Sq, D] -> [Sq, 1, D] (single head)
+        q_block = sparse_q_bh[qb].unsqueeze(1)
+        q_list.append(q_block)
+
+        # [num_sel, Sk, D] -> [num_kv_tokens, 1, D]
+        sel_k = k_blocks_bh[sel_idx].reshape(num_kv_tokens, 1, D)
+        sel_v = v_blocks_bh[sel_idx].reshape(num_kv_tokens, 1, D)
+        k_list.append(sel_k)
+        v_list.append(sel_v)
+
+        cu_seqlens_q.append(cu_seqlens_q[-1] + Sq)
+        cu_seqlens_k.append(cu_seqlens_k[-1] + num_kv_tokens)
+
+    if not q_list:
+        return
+
+    flat_q = torch.cat(q_list, dim=0)
+    flat_k = torch.cat(k_list, dim=0)
+    flat_v = torch.cat(v_list, dim=0)
+
+    cu_seqlens_q_t = torch.tensor(cu_seqlens_q, dtype=torch.int32, device=device)
+    cu_seqlens_k_t = torch.tensor(cu_seqlens_k, dtype=torch.int32, device=device)
+
+    max_seqlen_q = Sq
+    max_seqlen_k = int((cu_seqlens_k_t[1:] - cu_seqlens_k_t[:-1]).max().item())
+
+    orig_dtype = flat_q.dtype
+    compute_dtype = orig_dtype
+    if compute_dtype not in (torch.float16, torch.bfloat16):
+        compute_dtype = torch.bfloat16
+        flat_q = flat_q.to(compute_dtype)
+        flat_k = flat_k.to(compute_dtype)
+        flat_v = flat_v.to(compute_dtype)
+
+    flat_out = flash_attn_varlen_func_impl(
+        flat_q,
+        flat_k,
+        flat_v,
+        cu_seqlens_q_t,
+        cu_seqlens_k_t,
+        max_seqlen_q,
+        max_seqlen_k,
+        causal=False,
+    )
+
+    if compute_dtype != orig_dtype:
+        flat_out = flat_out.to(orig_dtype)
+
+    idx = 0
+    for qb in active_blocks:
+        block_out = flat_out[idx:idx + Sq]  # [Sq, 1, D]
+        output[b, h, qb] = block_out.squeeze(1)  # [Sq, D]
+        idx += Sq
+
+
+def _reconstruct_pruned(
+    sparse_output: torch.Tensor,
+    keep_indices: torch.Tensor,
+    block_size: int,
+) -> torch.Tensor:
+    """
+    Scatter sparse output back to full block size.
+    Pruned positions get nearest kept token's output.
+
+    Handles per-batch, per-head indices correctly.
+
+    Args:
+        sparse_output: [B, H, N, keep_size, D]
+        keep_indices:  [B, H, N, keep_size]
+        block_size: original tokens per block
+
+    Returns:
+        full_output: [B, H, N, block_size, D]
+    """
+    B, H, N, keep_size, D = sparse_output.shape
+    device = sparse_output.device
+
+    if keep_size >= block_size:
+        return sparse_output
+
+    full_output = torch.zeros(B, H, N, block_size, D, device=device, dtype=sparse_output.dtype)
+
+    # Scatter kept tokens
+    idx_expand = keep_indices.unsqueeze(-1).expand(-1, -1, -1, -1, D)
+    full_output.scatter_(3, idx_expand, sparse_output)
+
+    # Fill pruned positions with nearest kept token (vectorized)
+    all_pos = torch.arange(block_size, device=device)
+
+    for b in range(B):
+        for h in range(H):
+            for n in range(N):
+                kept = keep_indices[b, h, n]  # [keep_size]
+
+                # Distance from every position to every kept position
+                dists = (all_pos.view(-1, 1) - kept.view(1, -1)).abs()
+                nearest_local_idx = dists.argmin(dim=1)  # [block_size]
+
+                # Identify pruned positions
+                is_pruned = torch.ones(block_size, dtype=torch.bool, device=device)
+                is_pruned[kept] = False
+                pruned_indices = is_pruned.nonzero(as_tuple=True)[0]
+
+                if pruned_indices.numel() > 0:
+                    src_indices = nearest_local_idx[pruned_indices]
+                    full_output[b, h, n, pruned_indices] = sparse_output[b, h, n, src_indices]
+
+    return full_output
+
+
+# ---------------------------------------------------------------------------
+# FastVideo backend classes
+# ---------------------------------------------------------------------------
+
+
+class BSAAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = False
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 128]
+
+    @staticmethod
+    def get_name() -> str:
+        return "BSA_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["BSAAttentionImpl"]:
+        return BSAAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["BSAAttentionMetadata"]:
+        return BSAAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["BSAAttentionMetadataBuilder"]:
+        return BSAAttentionMetadataBuilder
+
+
+@dataclass
+class BSAAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+    dit_seq_shape: tuple[int, int, int]
+    total_seq_length: int
+    num_blocks: int
+    block_size: int
+    tile_partition_indices: torch.LongTensor
+    reverse_tile_partition_indices: torch.LongTensor
+    # BSA-specific config
+    query_keep_ratio: float
+    kv_cumulative_threshold: float
+    min_kv_blocks: int
+
+
+class BSAAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self):
+        pass
+
+    def prepare(self):
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        raw_latent_shape: tuple[int, int, int],
+        patch_size: tuple[int, int, int],
+        device: torch.device,
+        bsa_query_keep_ratio: float = 0.5,
+        bsa_kv_cumulative_threshold: float = 0.9,
+        bsa_min_kv_blocks: int = 4,
+        **kwargs: dict[str, Any],
+    ) -> "BSAAttentionMetadata":
+        # Ensure patching does not drop tokens silently.
+        assert all(r % p == 0 for r, p in zip(raw_latent_shape, patch_size, strict=False)), (
+            "raw_latent_shape must be divisible by patch_size for BSA", )
+
+        dit_seq_shape = (
+            raw_latent_shape[0] // patch_size[0],
+            raw_latent_shape[1] // patch_size[1],
+            raw_latent_shape[2] // patch_size[2],
+        )
+
+        total_seq_length = math.prod(dit_seq_shape)
+        block_size = math.prod(BSA_TILE_SIZE)
+        # Require exact tiling to avoid reshape failures later.
+        assert all(d % t == 0 for d, t in zip(dit_seq_shape, BSA_TILE_SIZE, strict=False)), (
+            "dit_seq_shape must be divisible by BSA_TILE_SIZE", )
+        num_blocks = total_seq_length // block_size
+
+        tile_partition_indices = get_tile_partition_indices(dit_seq_shape, BSA_TILE_SIZE, device)
+        reverse_tile_partition_indices = get_reverse_tile_partition_indices(dit_seq_shape, BSA_TILE_SIZE, device)
+
+        return BSAAttentionMetadata(
+            current_timestep=current_timestep,
+            dit_seq_shape=dit_seq_shape,
+            total_seq_length=total_seq_length,
+            num_blocks=num_blocks,
+            block_size=block_size,
+            tile_partition_indices=tile_partition_indices,
+            reverse_tile_partition_indices=reverse_tile_partition_indices,
+            query_keep_ratio=bsa_query_keep_ratio,
+            kv_cumulative_threshold=bsa_kv_cumulative_threshold,
+            min_kv_blocks=bsa_min_kv_blocks,
+        )
+
+
+class BSAAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.prefix = prefix
+        self.num_heads = num_heads
+        self.head_size = head_size
+        if num_kv_heads is not None and num_kv_heads != num_heads:
+            raise ValueError("BSA backend does not support grouped-query attention")
+        if causal:
+            raise ValueError("BSA backend is bidirectional; causal=True is unsupported")
+        if softmax_scale is not None:
+            expected_scale = 1.0 / math.sqrt(self.head_size)
+            if not math.isclose(softmax_scale, expected_scale, rel_tol=1e-4, abs_tol=1e-5):
+                raise ValueError("softmax_scale must be default (1/sqrt(d)) for BSA")
+        try:
+            sp_group = get_sp_group()
+            self.sp_size = sp_group.world_size
+        except (AssertionError, RuntimeError):
+            self.sp_size = 1
+
+    def preprocess_qkv(
+        self,
+        qkv: torch.Tensor,
+        attn_metadata: BSAAttentionMetadata,
+    ) -> torch.Tensor:
+        """Reorder tokens from raster order to tile-contiguous order."""
+        # qkv: [B, L, num_heads, D]
+        return qkv[:, attn_metadata.tile_partition_indices]
+
+    def postprocess_output(
+        self,
+        output: torch.Tensor,
+        attn_metadata: BSAAttentionMetadata,
+    ) -> torch.Tensor:
+        """Reorder tokens from tile-contiguous order back to raster order."""
+        return output[:, attn_metadata.reverse_tile_partition_indices]
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: BSAAttentionMetadata,
+    ) -> torch.Tensor:
+        """
+        BSA attention forward pass.
+
+        Input tensors are already in tile-contiguous order from preprocess_qkv.
+
+        Args:
+            query: [B, L, num_heads, D] (tile-ordered)
+            key:   [B, L, num_heads, D] (tile-ordered)
+            value: [B, L, num_heads, D] (tile-ordered)
+            attn_metadata: BSA metadata
+
+        Returns:
+            output: [B, L, num_heads, D] (tile-ordered)
+        """
+        B, L, H, D = query.shape
+        block_size = attn_metadata.block_size
+        num_blocks = attn_metadata.num_blocks
+        assert num_blocks * block_size == L, "Sequence length must match tiling"
+
+        # Reshape to [B, H, L, D] for attention computation
+        q = query.transpose(1, 2).contiguous()  # [B, H, L, D]
+        k = key.transpose(1, 2).contiguous()
+        v = value.transpose(1, 2).contiguous()
+
+        # Reshape into blocks: [B, H, num_blocks, block_size, D]
+        q_blocks = q.view(B, H, num_blocks, block_size, D)
+        k_blocks = k.view(B, H, num_blocks, block_size, D)
+        v_blocks = v.view(B, H, num_blocks, block_size, D)
+
+        # --- Query sparsification ---
+        sparse_q, keep_indices, keep_size = _prune_queries(q_blocks, attn_metadata.query_keep_ratio)
+
+        # --- KV block selection ---
+        kv_mask = _select_kv_blocks(
+            sparse_q,
+            k_blocks,
+            attn_metadata.kv_cumulative_threshold,
+            attn_metadata.min_kv_blocks,
+        )
+
+        # --- Sparse attention ---
+        sparse_output = _compute_sparse_attention(sparse_q, k_blocks, v_blocks, kv_mask)
+
+        # --- Reconstruct pruned positions ---
+        full_output = _reconstruct_pruned(sparse_output, keep_indices, block_size)
+
+        # Reshape back: [B, H, num_blocks, block_size, D] -> [B, H, L, D] -> [B, L, H, D]
+        hidden_states = full_output.view(B, H, L, D).transpose(1, 2)
+
+        return hidden_states

fastvideo/configs/models/dits/gen3c.py

@@ -0,0 +1,188 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+
+
+def is_transformer_blocks(n: str, m) -> bool:
+    return "transformer_blocks" in n and str.isdigit(n.split(".")[-1])
+
+
+@dataclass
+class Gen3CArchConfig(DiTArchConfig):
+    """Configuration for GEN3C architecture (VideoExtendGeneralDIT)."""
+
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [is_transformer_blocks])
+
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            # Official GEN3C checkpoint key naming to FastVideo mapping.
+            # The official checkpoint uses nn.Sequential patterns like attn.to_q.0 (Linear)
+            # and attn.to_q.1 (RMSNorm), and layer1/layer2 for MLP.
+            #
+            # Patch embedding: net.x_embedder.proj.1.weight -> patch_embed.proj.weight
+            r"^net\.x_embedder\.proj\.1\.(.*)$": r"patch_embed.proj.\1",
+
+            # Time embedding: net.t_embedder.1.linear_*.weight -> time_embed.t_embedder.linear_*.weight
+            r"^net\.t_embedder\.0\.(.*)$": r"time_embed.time_proj.\1",
+            r"^net\.t_embedder\.1\.linear_1\.(.*)$": r"time_embed.t_embedder.linear_1.\1",
+            r"^net\.t_embedder\.1\.linear_2\.(.*)$": r"time_embed.t_embedder.linear_2.\1",
+
+            # Augment sigma embedding (GEN3C-specific)
+            r"^net\.augment_sigma_embedder\.0\.(.*)$": r"augment_sigma_embed.time_proj.\1",
+            r"^net\.augment_sigma_embedder\.1\.linear_1\.(.*)$": r"augment_sigma_embed.t_embedder.linear_1.\1",
+            r"^net\.augment_sigma_embedder\.1\.linear_2\.(.*)$": r"augment_sigma_embed.t_embedder.linear_2.\1",
+
+            # Affine embedding norm: net.affline_norm.weight -> affine_norm.weight
+            # Note: "affline" is a typo in the official GEN3C checkpoint (should be "affine")
+            r"^net\.affline_norm\.(.*)$": r"affine_norm.\1",
+
+            # Extra positional embeddings (learnable per-axis)
+            r"^net\.extra_pos_embedder\.pos_emb_t$": r"learnable_pos_embed.pos_emb_t",
+            r"^net\.extra_pos_embedder\.pos_emb_h$": r"learnable_pos_embed.pos_emb_h",
+            r"^net\.extra_pos_embedder\.pos_emb_w$": r"learnable_pos_embed.pos_emb_w",
+
+            # Transformer blocks: net.blocks.blockN -> transformer_blocks.N
+            # Official uses: block.attn.to_q.0 (Linear), block.attn.to_q.1 (QK RMSNorm)
+            #
+            # Self-attention (block index 0)
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_q\.0\.(.*)$": r"transformer_blocks.\1.attn1.to_q.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_q\.1\.(.*)$":
+            r"transformer_blocks.\1.attn1.norm_q.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_k\.0\.(.*)$": r"transformer_blocks.\1.attn1.to_k.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_k\.1\.(.*)$":
+            r"transformer_blocks.\1.attn1.norm_k.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_v\.0\.(.*)$": r"transformer_blocks.\1.attn1.to_v.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.block\.attn\.to_out\.0\.(.*)$":
+            r"transformer_blocks.\1.attn1.to_out.\2",
+            # AdaLN modulation for self-attention
+            r"^net\.blocks\.block(\d+)\.blocks\.0\.adaLN_modulation\.(.*)$":
+            r"transformer_blocks.\1.adaln_modulation_self_attn.\2",
+
+            # Cross-attention (block index 1)
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_q\.0\.(.*)$": r"transformer_blocks.\1.attn2.to_q.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_q\.1\.(.*)$":
+            r"transformer_blocks.\1.attn2.norm_q.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_k\.0\.(.*)$": r"transformer_blocks.\1.attn2.to_k.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_k\.1\.(.*)$":
+            r"transformer_blocks.\1.attn2.norm_k.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_v\.0\.(.*)$": r"transformer_blocks.\1.attn2.to_v.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.block\.attn\.to_out\.0\.(.*)$":
+            r"transformer_blocks.\1.attn2.to_out.\2",
+            # AdaLN modulation for cross-attention
+            r"^net\.blocks\.block(\d+)\.blocks\.1\.adaLN_modulation\.(.*)$":
+            r"transformer_blocks.\1.adaln_modulation_cross_attn.\2",
+
+            # MLP (block index 2): layer1 -> fc_in, layer2 -> fc_out
+            r"^net\.blocks\.block(\d+)\.blocks\.2\.block\.layer1\.(.*)$": r"transformer_blocks.\1.mlp.fc_in.\2",
+            r"^net\.blocks\.block(\d+)\.blocks\.2\.block\.layer2\.(.*)$": r"transformer_blocks.\1.mlp.fc_out.\2",
+            # AdaLN modulation for MLP
+            r"^net\.blocks\.block(\d+)\.blocks\.2\.adaLN_modulation\.(.*)$":
+            r"transformer_blocks.\1.adaln_modulation_mlp.\2",
+
+            # Final layer: net.final_layer.linear -> final_layer.proj_out
+            r"^net\.final_layer\.linear\.(.*)$": r"final_layer.proj_out.\1",
+            # Final layer AdaLN: net.final_layer.adaLN_modulation -> final_layer.adaln_modulation
+            r"^net\.final_layer\.adaLN_modulation\.(.*)$": r"final_layer.adaln_modulation.\1",
+
+            # Note: The following keys from official checkpoint are NOT mapped and can be safely ignored:
+            # - net.pos_embedder.* (rope position embeddings computed dynamically)
+            # - net.accum_* keys (training metadata)
+            # - logvar.* (training-only module, not used in inference)
+        })
+
+    lora_param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^transformer_blocks\.(\d+)\.attn1\.to_q\.(.*)$": r"transformer_blocks.\1.attn1.to_q.\2",
+            r"^transformer_blocks\.(\d+)\.attn1\.to_k\.(.*)$": r"transformer_blocks.\1.attn1.to_k.\2",
+            r"^transformer_blocks\.(\d+)\.attn1\.to_v\.(.*)$": r"transformer_blocks.\1.attn1.to_v.\2",
+            r"^transformer_blocks\.(\d+)\.attn1\.to_out\.(.*)$": r"transformer_blocks.\1.attn1.to_out.\2",
+            r"^transformer_blocks\.(\d+)\.attn2\.to_q\.(.*)$": r"transformer_blocks.\1.attn2.to_q.\2",
+            r"^transformer_blocks\.(\d+)\.attn2\.to_k\.(.*)$": r"transformer_blocks.\1.attn2.to_k.\2",
+            r"^transformer_blocks\.(\d+)\.attn2\.to_v\.(.*)$": r"transformer_blocks.\1.attn2.to_v.\2",
+            r"^transformer_blocks\.(\d+)\.attn2\.to_out\.(.*)$": r"transformer_blocks.\1.attn2.to_out.\2",
+            r"^transformer_blocks\.(\d+)\.mlp\.(.*)$": r"transformer_blocks.\1.mlp.\2",
+        })
+
+    # GEN3C architecture parameters
+    # Base VAE latent channels
+    in_channels: int = 16
+    out_channels: int = 16
+
+    # Channels per 3D cache buffer: 16 (warped frame latent) + 16 (warped mask latent)
+    CHANNELS_PER_BUFFER: int = 32
+
+    # Number of 3D cache buffers
+    frame_buffer_max: int = 2
+
+    # Attention configuration (7B model: 32 heads x 128 dim = 4096 hidden)
+    num_attention_heads: int = 32
+    attention_head_dim: int = 128  # 4096 / 32
+    num_layers: int = 28
+    mlp_ratio: float = 4.0
+
+    # Text encoder configuration
+    text_embed_dim: int = 1024
+
+    # AdaLN-LoRA configuration
+    adaln_lora_dim: int = 256
+    use_adaln_lora: bool = True
+
+    # GEN3C-specific: augment sigma embedding for conditioning noise augmentation
+    # Note: The official GEN3C-Cosmos-7B checkpoint was trained without this
+    add_augment_sigma_embedding: bool = False
+
+    # Position embedding configuration
+    max_size: tuple[int, int, int] = (128, 240, 240)  # T, H, W
+    patch_size: tuple[int, int, int] = (1, 2, 2)
+    rope_scale: tuple[float, float, float] = (2.0, 1.0, 1.0)  # T, H, W scaling
+
+    # GEN3C uses learnable positional embeddings in addition to RoPE
+    extra_pos_embed_type: str = "learnable"
+
+    # Padding mask handling
+    concat_padding_mask: bool = True
+
+    # Cross-attention projection (not used in GEN3C 7B)
+    use_crossattn_projection: bool = False
+
+    # RoPE FPS modulation
+    rope_enable_fps_modulation: bool = True
+
+    # QK normalization
+    qk_norm: str = "rms_norm"
+    eps: float = 1e-6
+
+    # Affine embedding normalization
+    affine_emb_norm: bool = True
+
+    # Block format (THWBD for GEN3C compatibility)
+    block_x_format: str = "THWBD"
+
+    exclude_lora_layers: list[str] = field(default_factory=lambda: ["embedder"])
+
+    def __post_init__(self):
+        super().__post_init__()
+        self.out_channels = self.out_channels or self.in_channels
+        self.hidden_size = self.num_attention_heads * self.attention_head_dim
+        self.num_channels_latents = self.in_channels
+
+        # Calculate total input channels for patch embedding:
+        # - in_channels (16): VAE latent
+        # - condition_video_input_mask (1): Binary mask for conditioning frames
+        # - condition_video_pose (frame_buffer_max * 32): 3D cache buffers
+        # - padding_mask (1 if concat_padding_mask): Padding mask
+        self.buffer_channels = self.frame_buffer_max * self.CHANNELS_PER_BUFFER
+        self.total_input_channels = (
+            self.in_channels +  # 16: VAE latent
+            1 +  # 1: condition_video_input_mask
+            self.buffer_channels  # 64: 3D cache buffers (2 * 32)
+        )
+        # padding_mask is added in build_patch_embed if concat_padding_mask=True
+
+
+@dataclass
+class Gen3CVideoConfig(DiTConfig):
+    """Configuration for GEN3C video generation model."""
+    arch_config: DiTArchConfig = field(default_factory=Gen3CArchConfig)
+    prefix: str = "Gen3C"

fastvideo/configs/models/vaes/__init__.py

@@ -1,6 +1,7 @@
 from fastvideo.configs.models.vaes.cosmosvae import CosmosVAEConfig
 from fastvideo.configs.models.vaes.cosmos2_5vae import Cosmos25VAEConfig
 from fastvideo.configs.models.vaes.gamecraftvae import GameCraftVAEConfig
+from fastvideo.configs.models.vaes.gen3cvae import Gen3CVAEConfig
 from fastvideo.configs.models.vaes.hunyuanvae import HunyuanVAEConfig
 from fastvideo.configs.models.vaes.hunyuan15vae import Hunyuan15VAEConfig
 from fastvideo.configs.models.vaes.ltx2vae import LTX2VAEConfig
@@ -12,6 +13,7 @@ __all__ = [
     "WanVAEConfig",
     "CosmosVAEConfig",
     "Cosmos25VAEConfig",
+    "Gen3CVAEConfig",
     "Hunyuan15VAEConfig",
     "LTX2VAEConfig",
 ]

fastvideo/configs/models/vaes/gen3cvae.py

@@ -0,0 +1,14 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass
+
+from fastvideo.configs.models.vaes.cosmosvae import CosmosVAEConfig
+
+
+@dataclass
+class Gen3CVAEConfig(CosmosVAEConfig):
+    """
+    GEN3C VAE config placeholder.
+
+    GEN3C uses tokenizer-backed VAE loading logic at runtime, but we keep a
+    model-specific config class so pipeline/model configs stay model-scoped.
+    """

fastvideo/configs/pipelines/__init__.py

@@ -5,7 +5,7 @@ from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.hunyuangamecraft import HunyuanGameCraftPipelineConfig
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
-from fastvideo.configs.pipelines.ltx2 import LTX2T2VConfig
+from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
 from fastvideo.registry import get_pipeline_config_cls_from_name
 from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig, WanI2V480PConfig, WanI2V720PConfig,
                                              WanT2V480PConfig, WanT2V720PConfig)

fastvideo/configs/pipelines/gen3c.py

@@ -0,0 +1,171 @@
+# SPDX-License-Identifier: Apache-2.0
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits.gen3c import Gen3CVideoConfig
+from fastvideo.configs.models.encoders import BaseEncoderOutput
+from fastvideo.configs.models.encoders.base import TextEncoderArchConfig
+from fastvideo.configs.models.encoders.t5 import (T5LargeArchConfig, T5LargeConfig)
+from fastvideo.configs.models.vaes import Gen3CVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig
+
+
+@dataclass
+class _Gen3CT5LargeArchConfig(T5LargeArchConfig):
+    """T5 Large arch config that pads inputs to max_length.
+
+    GEN3C requires padded text encoder inputs, while the base 
+    T5 config no longer pads by default after the SP mask 
+    refactor [PR#1142](https://github.com/hao-ai-lab/FastVideo/pull/1142).
+    """
+
+    def __post_init__(self):
+        super().__post_init__()
+        self.tokenizer_kwargs["padding"] = "max_length"
+
+
+@dataclass
+class _Gen3CT5LargeConfig(T5LargeConfig):
+    arch_config: TextEncoderArchConfig = field(default_factory=_Gen3CT5LargeArchConfig)
+    prefix: str = "t5"
+
+
+def t5_large_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    """Postprocess T5 Large text encoder outputs for GEN3C pipeline.
+    
+    Return raw last_hidden_state without truncation/padding.
+    """
+    hidden_state = outputs.last_hidden_state
+
+    if hidden_state is None:
+        raise ValueError("T5 Large outputs missing last_hidden_state")
+
+    nan_count = torch.isnan(hidden_state).sum()
+    if nan_count > 0:
+        hidden_state = hidden_state.masked_fill(torch.isnan(hidden_state), 0.0)
+
+    # Zero out embeddings beyond actual sequence length (vectorized)
+    if outputs.attention_mask is not None:
+        attention_mask = outputs.attention_mask
+        lengths = attention_mask.sum(dim=1)
+        max_len = hidden_state.shape[1]
+        mask = torch.arange(max_len, device=hidden_state.device)[None, :] >= lengths[:, None]
+        hidden_state[mask] = 0.0
+
+    return hidden_state
+
+
+@dataclass
+class Gen3CConfig(PipelineConfig):
+    """Configuration for GEN3C Video Generation Pipeline.
+    
+    GEN3C extends Cosmos with 3D cache for camera-controlled video generation.
+    Key parameters:
+    - frame_buffer_max: Number of 3D cache buffers (default: 2)
+    - noise_aug_strength: Strength of noise augmentation per buffer
+    - filter_points_threshold: Threshold for filtering unreliable depth points
+    """
+
+    dit_config: DiTConfig = field(default_factory=Gen3CVideoConfig)
+
+    vae_config: VAEConfig = field(default_factory=Gen3CVAEConfig)
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(default_factory=lambda: (_Gen3CT5LargeConfig(), ))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
+                                  ...] = field(default_factory=lambda: (t5_large_postprocess_text, ))
+
+    dit_precision: str = "bf16"
+    vae_precision: str = "bf16"
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", ))
+
+    # GEN3C-specific conditioning parameters
+    conditioning_strategy: str = "frame_replace"
+    min_num_conditional_frames: int = 1
+    max_num_conditional_frames: int = 2
+    # Match official GEN3C/Cosmos inference defaults.
+    sigma_conditional: float = 0.001
+    sigma_data: float = 0.5
+    state_ch: int = 16
+    state_t: int = 16  # GEN3C uses 16 latent frames (121 pixel frames)
+    text_encoder_class: str = "T5"
+
+    # Flow matching parameters
+    embedded_cfg_scale: int = 6
+    flow_shift: float = 1.0
+
+    # GEN3C 3D Cache parameters
+    frame_buffer_max: int = 2
+    noise_aug_strength: float = 0.0
+    filter_points_threshold: float = 0.05
+
+    # Depth estimation settings
+    use_moge_depth: bool = True
+    moge_model_name: str = "Ruicheng/moge-vitl"
+    offload_moge_after_depth: bool = True
+
+    # Camera trajectory settings (matching NVIDIA inference defaults)
+    default_trajectory_type: str = "left"
+    default_movement_distance: float = 0.3
+    default_camera_rotation: str = "center_facing"
+
+    # Video generation settings
+    # Match official GEN3C defaults (height=704, width=1280).
+    video_resolution: tuple[int, int] = (704, 1280)  # H, W
+    num_frames: int = 121  # Default number of frames to generate
+
+    # Generation frame rate
+    fps: int = 24
+
+    # Explicit CFG behavior policy:
+    # - "legacy": CFG branch only when guidance_scale > 1.0
+    # - "official_uncond_at_unity": also run uncond branch at guidance_scale == 1.0
+    cfg_behavior: str = "legacy"
+    default_negative_prompt: str = (
+        "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
+        "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
+        "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, "
+        "jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special "
+        "effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and "
+        "flickering. Overall, the video is of poor quality.")
+
+    # Autoregressive generation settings
+    autoregressive_chunk_frames: int = 121  # Frames per chunk
+    autoregressive_overlap_frames: int = 1  # Overlap between chunks
+
+    def __post_init__(self):
+        self.vae_config.load_encoder = True
+        self.vae_config.load_decoder = True
+
+        self._vae_latent_dim = 16
+
+        # Validate frame buffer configuration matches DiT
+        if hasattr(self.dit_config, 'arch_config'):
+            arch_config = self.dit_config.arch_config
+            if (hasattr(arch_config, 'frame_buffer_max') and arch_config.frame_buffer_max != self.frame_buffer_max):
+                raise ValueError(f"frame_buffer_max mismatch: pipeline config has {self.frame_buffer_max}, "
+                                 f"DiT config has {arch_config.frame_buffer_max}")
+
+        allowed_cfg_behavior = {"legacy", "official_uncond_at_unity"}
+        if self.cfg_behavior not in allowed_cfg_behavior:
+            raise ValueError(f"cfg_behavior must be one of {sorted(allowed_cfg_behavior)}, got {self.cfg_behavior!r}")
+
+
+@dataclass
+class Gen3CInferenceConfig(Gen3CConfig):
+    """Configuration for GEN3C inference with optimized defaults."""
+
+    # Use smaller batch sizes for inference
+    batch_size: int = 1
+
+    # Enable gradient checkpointing for memory efficiency
+    gradient_checkpointing: bool = False
+
+    # Inference-specific parameters
+    guidance_scale: float = 1.0
+    num_inference_steps: int = 35
+
+    # Disable noise augmentation during inference
+    noise_aug_strength: float = 0.0

fastvideo/configs/sample/__init__.py

@@ -1,13 +0,0 @@
-from fastvideo.configs.sample.base import SamplingParam
-from fastvideo.configs.sample.hunyuangamecraft import (
-    HunyuanGameCraftSamplingParam,
-    HunyuanGameCraft65FrameSamplingParam,
-    HunyuanGameCraft129FrameSamplingParam,
-)
-
-__all__ = [
-    "SamplingParam",
-    "HunyuanGameCraftSamplingParam",
-    "HunyuanGameCraft65FrameSamplingParam",
-    "HunyuanGameCraft129FrameSamplingParam",
-]

fastvideo/configs/sample/cosmos.py

@@ -1,18 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class Cosmos_Predict2_2B_Video2World_SamplingParam(SamplingParam):
-    # Video parameters
-    height: int = 704
-    width: int = 1280
-    num_frames: int = 93
-    fps: int = 16
-
-    # Denoising stage
-    guidance_scale: float = 7.0
-    negative_prompt: str = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."
-    num_inference_steps: int = 35

fastvideo/configs/sample/cosmos2_5.py

@@ -1,23 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class Cosmos25SamplingParamBase(SamplingParam):
-    height: int = 704
-    width: int = 1280
-    num_frames: int = 77
-    fps: int = 24
-    seed: int = 0
-
-    guidance_scale: float = 7.0
-    negative_prompt: str = (
-        "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
-        "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
-        "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, "
-        "low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, "
-        "unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. "
-        "Overall, the video is of poor quality.")
-    num_inference_steps: int = 35

fastvideo/configs/sample/hunyuan.py

@@ -1,21 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class HunyuanSamplingParam(SamplingParam):
-    num_inference_steps: int = 50
-
-    num_frames: int = 125
-    height: int = 720
-    width: int = 1280
-    fps: int = 24
-
-    guidance_scale: float = 1.0
-
-
-@dataclass
-class FastHunyuanSamplingParam(HunyuanSamplingParam):
-    num_inference_steps: int = 6

fastvideo/configs/sample/hunyuan15.py

@@ -1,55 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import numpy as np
-from dataclasses import dataclass, field
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class Hunyuan15_480P_SamplingParam(SamplingParam):
-    num_inference_steps: int = 50
-
-    num_frames: int = 121
-    height: int = 480
-    width: int = 848
-    fps: int = 24
-
-    guidance_scale: float = 6.0
-    sigmas: list[float] | None = field(default_factory=lambda: list(np.linspace(1.0, 0.0, 50 + 1)[:-1]))
-
-    negative_prompt: str = ""
-
-    def __post_init__(self):
-        super().__post_init__()
-        self.sigmas = list(np.linspace(1.0, 0.0, self.num_inference_steps + 1)[:-1])
-
-
-@dataclass
-class Hunyuan15_480P_StepDistilled_I2V_SamplingParam(Hunyuan15_480P_SamplingParam):
-    num_inference_steps: int = 12
-
-    height: int = 720
-    width: int = 1280
-    guidance_scale: float = 1.0
-
-
-@dataclass
-class Hunyuan15_720P_SamplingParam(Hunyuan15_480P_SamplingParam):
-    height: int = 720
-    width: int = 1280
-
-
-@dataclass
-class Hunyuan15_720P_Distilled_I2V_SamplingParam(Hunyuan15_720P_SamplingParam):
-    guidance_scale: float = 1.0
-
-
-@dataclass
-class Hunyuan15_SR_1080P_SamplingParam(Hunyuan15_480P_SamplingParam):
-    height_sr: int = 1072
-    width_sr: int = 1920
-
-    num_inference_steps: int = 12
-    num_inference_steps_sr: int = 8
-
-    guidance_scale: float = 1.0

fastvideo/configs/sample/hunyuangamecraft.py

@@ -1,92 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""
-Sampling parameters for HunyuanGameCraft video generation.
-
-GameCraft generates game-like videos with camera/action control.
-Default parameters are based on the official implementation.
-"""
-from dataclasses import dataclass
-from typing import Any
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class HunyuanGameCraftSamplingParam(SamplingParam):
-    """Sampling parameters for HunyuanGameCraft video generation.
-    
-    Supports camera/action conditioning via:
-    - camera_trajectory: Plücker coordinates for camera motion
-    - action_list: List of actions (e.g., ["forward", "left", "right"])
-    - action_speed_list: Speed multipliers for each action
-    
-    Default resolution is 704x1280 (same as HunyuanVideo).
-    Default frame count is 33 video frames -> 9 latent frames.
-    """
-
-    # Number of denoising steps
-    num_inference_steps: int = 50
-
-    # Video dimensions
-    # 33 video frames -> 9 latent frames (4x temporal compression)
-    num_frames: int = 33
-    height: int = 704
-    width: int = 1280
-    fps: int = 24
-
-    # Guidance scale - official GameCraft uses CFG with guidance_scale=6.0
-    guidance_scale: float = 6.0
-
-    # Negative prompt for CFG (empty string = unconditional)
-    negative_prompt: str = ""
-
-    # Camera/Action conditioning
-    # Camera states as Plücker coordinates [B, T_video, 6, H, W]
-    camera_states: Any | None = None
-
-    # Camera trajectory file/identifier (alternative to camera_states)
-    camera_trajectory: str | None = None
-
-    # Action list for camera motion (e.g., ["forward", "left"])
-    action_list: list[str] | None = None
-
-    # Speed multipliers for each action
-    action_speed_list: list[float] | None = None
-
-    # History frame conditioning (for autoregressive generation)
-    # Ground truth latents for conditioning [B, 16, T, H, W]
-    gt_latents: Any | None = None
-
-    # Mask for conditioning (1=use gt, 0=generate) [B, 1, T, H, W]
-    conditioning_mask: Any | None = None
-
-    # Number of conditioning frames (for autoregressive) - maps to num_cond_frames
-    num_cond_frames: int = 0
-
-    def __post_init__(self) -> None:
-        super().__post_init__()
-        # Validate action lists
-        if (self.action_list is not None and self.action_speed_list is not None
-                and len(self.action_list) != len(self.action_speed_list)):
-            raise ValueError(f"action_list length ({len(self.action_list)}) must match "
-                             f"action_speed_list length ({len(self.action_speed_list)})")
-
-
-@dataclass
-class HunyuanGameCraft65FrameSamplingParam(HunyuanGameCraftSamplingParam):
-    """Sampling parameters for 65-frame GameCraft generation.
-    
-    65 video frames -> 17 latent frames (with first frame as key frame).
-    This is useful for longer video generation.
-    """
-    num_frames: int = 65
-
-
-@dataclass
-class HunyuanGameCraft129FrameSamplingParam(HunyuanGameCraftSamplingParam):
-    """Sampling parameters for 129-frame GameCraft generation.
-    
-    129 video frames -> 33 latent frames.
-    This is the maximum supported by the official implementation.
-    """
-    num_frames: int = 129

fastvideo/configs/sample/hyworld.py

@@ -1,25 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass, field
-
-from fastvideo.configs.sample.base import SamplingParam
-import numpy as np
-
-
-@dataclass
-class HYWorld_SamplingParam(SamplingParam):
-    num_inference_steps: int = 50
-
-    num_frames: int = 125
-    height: int = 480
-    width: int = 832
-    fps: int = 24
-
-    # Camera trajectory: pose string (e.g., 'w-31' means generating [1 + 31] latents) or JSON file path
-    pose: str = 'w-31'
-
-    guidance_scale: float = 6.0
-    prompt_attention_mask: list = field(default_factory=list)
-    negative_attention_mask: list = field(default_factory=list)
-    sigmas: list[float] | None = field(default_factory=lambda: list(np.linspace(1.0, 0.0, 50 + 1)[:-1]))
-
-    negative_prompt: str = ""

fastvideo/configs/sample/lingbotworld.py

@@ -1,20 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
-from fastvideo.configs.sample.wan import Wan2_2_I2V_A14B_SamplingParam
-
-
-@dataclass
-class LingBotWorld_SamplingParam(Wan2_2_I2V_A14B_SamplingParam):
-    guidance_scale: float = 5.0  # high_noise
-    guidance_scale_2: float = 5.0  # low_noise
-    num_inference_steps: int = 70
-    boundary_ratio: float | None = 0.947
-    negative_prompt: str | None = ("画面突变，色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，"
-                                   "最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，"
-                                   "畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走，"
-                                   "镜头晃动，画面闪烁，模糊，噪点，水印，签名，文字，变形，扭曲，液化，不合逻辑的结构，卡顿，"
-                                   "PPT幻灯片感，过暗，欠曝，低对比度，霓虹灯光感，过度锐化，3D渲染感，人物，行人，游客，身体，"
-                                   "皮肤，肢体，面部特征，汽车，电线")
-    fps: int = 16
-    # NOTE(will): default boundary timestep is tracked by PipelineConfig, but
-    # can be overridden during sampling

fastvideo/configs/sample/ltx2.py

@@ -1,69 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass, field
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class LTX2BaseSamplingParam(SamplingParam):
-    """Default sampling parameters for LTX-2 base one-stage T2V.
-
-    Values follow the official LTX-2 one-stage defaults.
-    Multi-modal CFG params are read by ``LTX2DenoisingStage``.
-    """
-
-    seed: int = 10
-    num_frames: int = 121
-    height: int = 512
-    width: int = 768
-    fps: int = 24
-    num_inference_steps: int = 40
-    guidance_scale: float = 3.0
-    # Copied/following official LTX-2 DEFAULT_NEGATIVE_PROMPT.
-    negative_prompt: str = ("blurry, out of focus, overexposed, underexposed, low contrast, "
-                            "washed out colors, excessive noise, grainy texture, poor lighting, "
-                            "flickering, motion blur, distorted proportions, unnatural skin "
-                            "tones, deformed facial features, asymmetrical face, missing facial "
-                            "features, extra limbs, disfigured hands, wrong hand count, "
-                            "artifacts around text, inconsistent perspective, camera shake, "
-                            "incorrect depth of field, background too sharp, background clutter, "
-                            "distracting reflections, harsh shadows, inconsistent lighting "
-                            "direction, color banding, cartoonish rendering, 3D CGI look, "
-                            "unrealistic materials, uncanny valley effect, incorrect ethnicity, "
-                            "wrong gender, exaggerated expressions, wrong gaze direction, "
-                            "mismatched lip sync, silent or muted audio, distorted voice, "
-                            "robotic voice, echo, background noise, off-sync audio, incorrect "
-                            "dialogue, added dialogue, repetitive speech, jittery movement, "
-                            "awkward pauses, incorrect timing, unnatural transitions, "
-                            "inconsistent framing, tilted camera, flat lighting, inconsistent "
-                            "tone, cinematic oversaturation, stylized filters, or AI artifacts.")
-    # Official LTX-2 multi-modal CFG defaults.
-    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
-    # STG (Spatio-Temporal Guidance) defaults from official LTX-2.
-    ltx2_stg_scale_video: float = 1.0
-    ltx2_stg_scale_audio: float = 1.0
-    ltx2_stg_blocks_video: list[int] = field(default_factory=lambda: [29])
-    ltx2_stg_blocks_audio: list[int] = field(default_factory=lambda: [29])
-
-
-@dataclass
-class LTX2DistilledSamplingParam(SamplingParam):
-    """Default sampling parameters for LTX-2 distilled one-stage T2V."""
-
-    seed: int = 10
-    num_frames: int = 121
-    height: int = 1024
-    width: int = 1536
-    fps: int = 24
-    num_inference_steps: int = 8
-    guidance_scale: float = 1.0
-    # No default negative_prompt for distilled models
-    negative_prompt: str = ""
-
-
-# Backward compatibility alias.
-LTX2SamplingParam = LTX2DistilledSamplingParam

fastvideo/configs/sample/sd35.py

@@ -1,25 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from __future__ import annotations
-
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class SD35SamplingParam(SamplingParam):
-
-    prompt: str | None = "a photo of a cat"
-    negative_prompt: str = ""
-
-    num_videos_per_prompt: int = 1
-    seed: int = 0
-
-    num_frames: int = 1
-    height: int = 512
-    width: int = 512
-    fps: int = 1
-
-    num_inference_steps: int = 28
-    guidance_scale: float = 6.0

fastvideo/configs/sample/turbodiffusion.py

@@ -1,73 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""
-TurboDiffusion sampling parameters.
-
-TurboDiffusion uses RCM (recurrent Consistency Model) scheduler for
-1-4 step video generation with no classifier-free guidance.
-"""
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class TurboDiffusionT2V_1_3B_SamplingParam(SamplingParam):
-    """Sampling parameters for TurboDiffusion T2V 1.3B model.
-    
-    Uses 4-step RCM sampling with guidance_scale=1.0 (no CFG).
-    """
-    # Video parameters
-    height: int = 480
-    width: int = 832
-    num_frames: int = 81
-    fps: int = 16
-
-    # Denoising stage - TurboDiffusion uses 1-4 steps with no CFG
-    guidance_scale: float = 1.0
-    num_inference_steps: int = 4
-
-    # No negative prompt needed for TurboDiffusion (no CFG)
-    negative_prompt: str | None = None
-
-
-@dataclass
-class TurboDiffusionT2V_14B_SamplingParam(SamplingParam):
-    """Sampling parameters for TurboDiffusion T2V 14B model.
-    
-    Uses 4-step RCM sampling with guidance_scale=1.0 (no CFG).
-    """
-    # Video parameters (720p for 14B)
-    height: int = 720
-    width: int = 1280
-    num_frames: int = 81
-    fps: int = 16
-
-    # Denoising stage - TurboDiffusion uses 1-4 steps with no CFG
-    guidance_scale: float = 1.0
-    num_inference_steps: int = 4
-
-    # No negative prompt needed for TurboDiffusion (no CFG)
-    negative_prompt: str | None = None
-
-
-@dataclass
-class TurboDiffusionI2V_A14B_SamplingParam(SamplingParam):
-    """Sampling parameters for TurboDiffusion I2V A14B model.
-    
-    Uses 4-step RCM sampling with dual-model switching (high/low noise).
-    """
-    # Video parameters (720p for A14B I2V)
-    height: int = 720
-    width: int = 1280
-    num_frames: int = 81
-    fps: int = 16
-
-    # Denoising stage - TurboDiffusion uses 1-4 steps with no CFG
-    guidance_scale: float = 1.0
-    num_inference_steps: int = 4
-
-    # Note: boundary_ratio is set in the pipeline config (TurboDiffusionI2VConfig),
-    # not here. This keeps sampling params and pipeline config separate.
-
-    # No negative prompt needed for TurboDiffusion (no CFG)
-    negative_prompt: str | None = None

fastvideo/configs/sample/wan.py

@@ -1,154 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from dataclasses import dataclass
-
-from fastvideo.configs.sample.base import SamplingParam
-
-
-@dataclass
-class WanT2V_1_3B_SamplingParam(SamplingParam):
-    # Video parameters
-    height: int = 480
-    width: int = 832
-    num_frames: int = 81
-    fps: int = 16
-
-    # Denoising stage
-    guidance_scale: float = 3.0
-    negative_prompt: str = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
-    num_inference_steps: int = 50
-
-
-@dataclass
-class WanT2V_14B_SamplingParam(SamplingParam):
-    # Video parameters
-    height: int = 720
-    width: int = 1280
-    num_frames: int = 81
-    fps: int = 16
-
-    # Denoising stage
-    guidance_scale: float = 5.0
-    negative_prompt: str = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
-    num_inference_steps: int = 50
-
-
-@dataclass
-class WanI2V_14B_480P_SamplingParam(WanT2V_1_3B_SamplingParam):
-    # Denoising stage
-    guidance_scale: float = 5.0
-    num_inference_steps: int = 40
-
-
-@dataclass
-class WanI2V_14B_720P_SamplingParam(WanT2V_14B_SamplingParam):
-    # Denoising stage
-    guidance_scale: float = 5.0
-    num_inference_steps: int = 40
-
-
-@dataclass
-class FastWanT2V480P_SamplingParam(WanT2V_1_3B_SamplingParam):
-    # DMD parameters
-    # dmd_denoising_steps: list[int] | None = field(default_factory=lambda: [1000, 757, 522])
-    num_inference_steps: int = 3
-    num_frames: int = 61
-    height: int = 448
-    width: int = 832
-    fps: int = 16
-
-
-# =============================================
-# ============= Wan2.1 Fun Models =============
-# =============================================
-@dataclass
-class Wan2_1_Fun_1_3B_InP_SamplingParam(SamplingParam):
-    """Sampling parameters for Wan2.1 Fun 1.3B InP model."""
-    height: int = 480
-    width: int = 832
-    num_frames: int = 81
-    fps: int = 16
-    negative_prompt: str | None = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
-    guidance_scale: float = 6.0
-    num_inference_steps: int = 50
-
-
-@dataclass
-class Wan2_1_Fun_1_3B_Control_SamplingParam(SamplingParam):
-    fps: int = 16
-    num_frames: int = 49
-    height: int = 832
-    width: int = 480
-    guidance_scale: float = 6.0
-
-
-# =============================================
-# ============= Wan2.2 TI2V Models =============
-# =============================================
-@dataclass
-class Wan2_2_Base_SamplingParam(SamplingParam):
-    """Sampling parameters for Wan2.2 TI2V 5B model."""
-    negative_prompt: str | None = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
-
-
-@dataclass
-class Wan2_2_TI2V_5B_SamplingParam(Wan2_2_Base_SamplingParam):
-    """Sampling parameters for Wan2.2 TI2V 5B model."""
-    height: int = 704
-    width: int = 1280
-    num_frames: int = 121
-    fps: int = 24
-    guidance_scale: float = 5.0
-    num_inference_steps: int = 50
-
-
-@dataclass
-class Wan2_2_T2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
-    guidance_scale: float = 4.0  # high_noise
-    guidance_scale_2: float = 3.0  # low_noise
-    num_inference_steps: int = 40
-    fps: int = 16
-    # NOTE(will): default boundary timestep is tracked by PipelineConfig, but
-    # can be overridden during sampling
-
-
-@dataclass
-class Wan2_2_I2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
-    guidance_scale: float = 3.5  # high_noise
-    guidance_scale_2: float = 3.5  # low_noise
-    num_inference_steps: int = 40
-    fps: int = 16
-    # NOTE(will): default boundary timestep is tracked by PipelineConfig, but
-    # can be overridden during sampling
-
-
-@dataclass
-class Wan2_2_Fun_A14B_Control_SamplingParam(Wan2_1_Fun_1_3B_Control_SamplingParam):
-    num_frames: int = 81
-
-
-# =============================================
-# ============= Causal Self-Forcing =============
-# =============================================
-@dataclass
-class SelfForcingWan2_1_T2V_1_3B_480P_SamplingParam(Wan2_1_Fun_1_3B_InP_SamplingParam):
-    pass
-
-
-@dataclass
-class SelfForcingWan2_2_T2V_A14B_480P_SamplingParam(Wan2_2_T2V_A14B_SamplingParam):
-    num_inference_steps: int = 8
-    num_frames: int = 81
-    height: int = 448
-    width: int = 832
-    fps: int = 16
-
-
-@dataclass
-class MatrixGame2_SamplingParam(SamplingParam):
-    height: int = 352
-    width: int = 640
-    num_frames: int = 57
-    fps: int = 25
-    guidance_scale: float = 1.0
-    num_inference_steps: int = 3
-    negative_prompt: str | None = None

fastvideo/entrypoints/cli/bench_serving.py

@@ -7,7 +7,7 @@ Example usage:
     # launch a server and benchmark on it
 
     # T2V or T2I or any other multimodal generation model
-    fastvideo serve --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers --port 8000
+    fastvideo serve --config serve.yaml
 
     # benchmark it and make sure the port is the same as the server's port
     fastvideo bench --dataset vbench --num-prompts 20 --port 8000

fastvideo/entrypoints/cli/generate.py

@@ -2,19 +2,17 @@
 # adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/entrypoints/cli/serve.py
 
 import argparse
-import dataclasses
 import os
 from typing import cast
 
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample.base import SamplingParam
 from fastvideo.entrypoints.cli.cli_types import CLISubcommand
-from fastvideo.entrypoints.cli.utils import RaiseNotImplementedAction
-from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.entrypoints.cli.inference_config import build_generate_run_config
 from fastvideo.logger import init_logger
 from fastvideo.utils import FlexibleArgumentParser
 
 logger = init_logger(__name__)
+_VALIDATED_RUN_CONFIG_ATTR = "_fastvideo_validated_run_config"
 
 
 class GenerateSubcommand(CLISubcommand):
@@ -23,89 +21,47 @@ class GenerateSubcommand(CLISubcommand):
     def __init__(self) -> None:
         self.name = "generate"
         super().__init__()
-        self.init_arg_names = self._get_init_arg_names()
-        self.generation_arg_names = self._get_generation_arg_names()
-
-    def _get_init_arg_names(self) -> list[str]:
-        """Get names of arguments for VideoGenerator initialization"""
-        return ["num_gpus", "tp_size", "sp_size", "model_path"]
-
-    def _get_generation_arg_names(self) -> list[str]:
-        """Get names of arguments for generate_video method"""
-        return [field.name for field in dataclasses.fields(SamplingParam)]
 
     def cmd(self, args: argparse.Namespace) -> None:
-        excluded_args = ['subparser', 'config', 'dispatch_function']
+        run_config = getattr(args, _VALIDATED_RUN_CONFIG_ATTR, None)
+        if run_config is None:
+            run_config = build_generate_run_config(
+                args,
+                overrides=getattr(args, "_unknown", None),
+            )
+        logger.info("CLI generate config: %s", run_config)
 
-        provided_args = {}
-        for k, v in vars(args).items():
-            if (k not in excluded_args and v is not None and hasattr(args, '_provided') and k in args._provided):
-                provided_args[k] = v
-
-        if 'model_path' in vars(args) and args.model_path is not None:
-            provided_args['model_path'] = args.model_path
-
-        if 'prompt' in vars(args) and args.prompt is not None:
-            provided_args['prompt'] = args.prompt
-
-        merged_args = {**provided_args}
-
-        logger.info('CLI Args: %s', merged_args)
-
-        if 'model_path' not in merged_args or not merged_args['model_path']:
-            raise ValueError("model_path must be provided either in config file or via --model-path")
-
-        # Check if either prompt or prompt_txt is provided
-        has_prompt = 'prompt' in merged_args and merged_args['prompt']
-        has_prompt_txt = 'prompt_txt' in merged_args and merged_args['prompt_txt']
-
-        if not (has_prompt or has_prompt_txt):
-            raise ValueError("Either prompt or prompt_txt must be provided")
-
-        if has_prompt and has_prompt_txt:
-            raise ValueError("Cannot provide both 'prompt' and 'prompt_txt'. Use only one of them.")
-
-        init_args = {k: v for k, v in merged_args.items() if k not in self.generation_arg_names}
-        generation_args = {k: v for k, v in merged_args.items() if k in self.generation_arg_names}
-        generation_args.setdefault("return_frames", False)
-
-        model_path = init_args.pop('model_path')
-        prompt = generation_args.pop('prompt', None)
-
-        generator = VideoGenerator.from_pretrained(model_path=model_path, **init_args)
-
-        # Call generate_video - it handles both single and batch modes
-        generator.generate_video(prompt=prompt, **generation_args)
+        generator = VideoGenerator.from_config(run_config.generator)
+        generator.generate(run_config.request)
 
     def validate(self, args: argparse.Namespace) -> None:
         """Validate the arguments for this command"""
-        if args.num_gpus is not None and args.num_gpus <= 0:
-            raise ValueError("Number of gpus must be positive")
-
-        if args.config and not os.path.exists(args.config):
+        if not args.config:
+            raise ValueError("fastvideo generate requires --config PATH; use a nested "
+                             "run config plus optional dotted overrides")
+        if not os.path.exists(args.config):
             raise ValueError(f"Config file not found: {args.config}")
+        setattr(
+            args,
+            _VALIDATED_RUN_CONFIG_ATTR,
+            build_generate_run_config(
+                args,
+                overrides=getattr(args, "_unknown", None),
+            ),
+        )
 
     def subparser_init(self, subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
         generate_parser = subparsers.add_parser(
             "generate",
             help="Run inference on a model",
-            usage="fastvideo generate (--model-path MODEL_PATH_OR_ID --prompt PROMPT) | --config CONFIG_FILE [OPTIONS]")
+            usage="fastvideo generate --config RUN_CONFIG [--dotted.override VALUE]")
 
         generate_parser.add_argument(
             "--config",
             type=str,
             default='',
             required=False,
-            help="Read CLI options from a config JSON or YAML file. If provided, --model-path and --prompt are optional."
-        )
-
-        generate_parser = FastVideoArgs.add_cli_args(generate_parser)
-        generate_parser = SamplingParam.add_cli_args(generate_parser)
-
-        generate_parser.add_argument(
-            "--text-encoder-configs",
-            action=RaiseNotImplementedAction,
-            help="JSON array of text encoder configurations (NOT YET IMPLEMENTED)",
+            help="Path to a nested run config JSON or YAML file. Required.",
         )
 
         return cast(FlexibleArgumentParser, generate_parser)

fastvideo/entrypoints/cli/inference_config.py

@@ -0,0 +1,111 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import argparse
+from collections.abc import Mapping
+from copy import deepcopy
+from typing import Any
+
+from fastvideo.api.overrides import apply_overrides, parse_cli_overrides
+from fastvideo.api.parser import load_raw_config, parse_config
+from fastvideo.api.schema import RunConfig, ServeConfig
+
+_GENERATE_OVERRIDE_PREFIXES = ("generator.", "request.")
+_SERVE_OVERRIDE_PREFIXES = (
+    "generator.",
+    "server.",
+    "default_request.",
+)
+
+
+def build_generate_run_config(
+    args: argparse.Namespace,
+    overrides: list[str] | None = None,
+) -> RunConfig:
+    raw = _load_nested_config(getattr(args, "config", None))
+    raw.setdefault("request", {})
+    raw = _apply_dotted_overrides(
+        raw,
+        overrides,
+        allowed_prefixes=_GENERATE_OVERRIDE_PREFIXES,
+    )
+    _ensure_generate_cli_defaults(raw)
+    config = parse_config(RunConfig, raw)
+    _validate_num_gpus(config.generator.engine.num_gpus)
+    _validate_generate_prompt_sources(config)
+    return config
+
+
+def build_serve_config(
+    args: argparse.Namespace,
+    overrides: list[str] | None = None,
+) -> ServeConfig:
+    raw = _load_nested_config(getattr(args, "config", None))
+    raw.setdefault("server", {})
+    raw.setdefault("default_request", {})
+    raw = _apply_dotted_overrides(
+        raw,
+        overrides,
+        allowed_prefixes=_SERVE_OVERRIDE_PREFIXES,
+    )
+    config = parse_config(ServeConfig, raw)
+    _validate_num_gpus(config.generator.engine.num_gpus)
+    return config
+
+
+def _load_nested_config(path: str | None) -> dict[str, Any]:
+    if not path:
+        raise ValueError("Inference CLI requires --config PATH; use a nested config file "
+                         "plus optional dotted overrides")
+
+    raw = load_raw_config(path)
+    if not isinstance(raw.get("generator"), Mapping):
+        raise ValueError("Inference config must use the nested schema with a top-level "
+                         "'generator' mapping")
+    return deepcopy(dict(raw))
+
+
+def _apply_dotted_overrides(
+    raw: Mapping[str, Any],
+    overrides: list[str] | None,
+    *,
+    allowed_prefixes: tuple[str, ...],
+) -> dict[str, Any]:
+    if not overrides:
+        return deepcopy(dict(raw))
+
+    parsed = parse_cli_overrides(overrides)
+    for key in parsed:
+        if "." not in key:
+            raise ValueError("CLI overrides must use dotted config paths like "
+                             "--request.sampling.seed 42")
+        if not key.startswith(allowed_prefixes):
+            allowed = ", ".join(allowed_prefixes)
+            raise ValueError(f"Unsupported override path {key!r}. Allowed prefixes: {allowed}")
+    return apply_overrides(raw, parsed)
+
+
+def _ensure_generate_cli_defaults(raw: dict[str, Any]) -> None:
+    request = raw.setdefault("request", {})
+    output = request.setdefault("output", {})
+    output.setdefault("return_frames", False)
+
+
+def _validate_generate_prompt_sources(config: RunConfig) -> None:
+    has_prompt = config.request.prompt is not None
+    has_prompt_path = config.request.inputs.prompt_path is not None
+    if not (has_prompt or has_prompt_path):
+        raise ValueError("Either request.prompt or request.inputs.prompt_path must be provided")
+    if has_prompt and has_prompt_path:
+        raise ValueError("Cannot provide both request.prompt and request.inputs.prompt_path")
+
+
+def _validate_num_gpus(num_gpus: int) -> None:
+    if num_gpus <= 0:
+        raise ValueError(f"generator.engine.num_gpus must be > 0; got {num_gpus}")
+
+
+__all__ = [
+    "build_generate_run_config",
+    "build_serve_config",
+]

fastvideo/entrypoints/cli/main.py

@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/entrypoints/cli/main.py
-
 from fastvideo.entrypoints.cli.cli_types import CLISubcommand
 from fastvideo.entrypoints.cli.generate import cmd_init as generate_cmd_init
 from fastvideo.utils import FlexibleArgumentParser
@@ -27,14 +26,17 @@ def main() -> None:
     for cmd in cmd_init():
         cmd.subparser_init(subparsers).set_defaults(dispatch_function=cmd.cmd)
         cmds[cmd.name] = cmd
-    args = parser.parse_args()
+
+    args, unknown = parser.parse_known_args()
+    if unknown and args.subparser not in {"generate", "serve"}:
+        parser.error(f"unrecognized arguments: {' '.join(unknown)}")
+    args._unknown = unknown
     if args.subparser in cmds:
         cmds[args.subparser].validate(args)
-
-    if hasattr(args, "dispatch_function"):
         args.dispatch_function(args)
-    else:
-        parser.print_help()
+        return
+
+    parser.print_help()
 
 
 if __name__ == "__main__":

fastvideo/entrypoints/cli/serve.py

@@ -2,14 +2,17 @@
 # adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/entrypoints/cli/serve.py
 
 import argparse
+import os
 from typing import cast
 
+from fastvideo.api.compat import generator_config_to_fastvideo_args
 from fastvideo.entrypoints.cli.cli_types import CLISubcommand
-from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.entrypoints.cli.inference_config import build_serve_config
 from fastvideo.logger import init_logger
 from fastvideo.utils import FlexibleArgumentParser
 
 logger = init_logger(__name__)
+_VALIDATED_SERVE_CONFIG_ATTR = "_fastvideo_validated_serve_config"
 
 
 class ServeSubcommand(CLISubcommand):
@@ -20,94 +23,69 @@ class ServeSubcommand(CLISubcommand):
         super().__init__()
 
     def cmd(self, args: argparse.Namespace) -> None:
-        excluded_args = {
-            "subparser",
-            "config",
-            "dispatch_function",
-            "host",
-            "port",
-            "output_dir",
-        }
+        serve_config = getattr(args, _VALIDATED_SERVE_CONFIG_ATTR, None)
+        if serve_config is None:
+            serve_config = build_serve_config(
+                args,
+                overrides=getattr(args, "_unknown", None),
+            )
 
-        provided: set[str] = getattr(args, '_provided', set())
-        cli_kwargs = {}
-        for k, v in vars(args).items():
-            if k in excluded_args:
-                continue
-            if k == '_provided':
-                continue
-            if k in provided and v is not None:
-                cli_kwargs[k] = v
+        logger.info("CLI serve config: %s", serve_config)
 
-        if 'model_path' not in cli_kwargs and args.model_path is not None:
-            cli_kwargs['model_path'] = args.model_path
-
-        if not cli_kwargs.get('model_path'):
-            raise ValueError("model_path must be provided via --model-path")
+        # A `streaming:` block selects the WebSocket/Dynamo runtime;
+        # its deps stay out of REST-only deployments via lazy import.
+        if serve_config.streaming is not None:
+            from fastvideo.entrypoints.streaming.server import (
+                run_server as run_streaming_server, )
+            run_streaming_server(serve_config)
+            return
 
         from fastvideo.entrypoints.openai.api_server import (
-            DEFAULT_HOST,
-            DEFAULT_OUTPUT_DIR,
-            DEFAULT_PORT,
-            run_server,
+            run_server, )
+
+        logger.info(
+            "Server will listen on %s:%d",
+            serve_config.server.host,
+            serve_config.server.port,
         )
 
-        host = getattr(args, "host", DEFAULT_HOST)
-        port = getattr(args, "port", DEFAULT_PORT)
-        output_dir = getattr(args, "output_dir", DEFAULT_OUTPUT_DIR)
-
-        logger.info("CLI serve args: %s", cli_kwargs)
-        logger.info("Server will listen on %s:%d", host, port)
-
-        fastvideo_args = FastVideoArgs.from_kwargs(**cli_kwargs)
-        run_server(fastvideo_args, host=host, port=port, output_dir=output_dir)
+        fastvideo_args = generator_config_to_fastvideo_args(serve_config.generator)
+        run_server(
+            fastvideo_args,
+            host=serve_config.server.host,
+            port=serve_config.server.port,
+            output_dir=serve_config.server.output_dir,
+            default_request=serve_config.default_request,
+        )
 
     def validate(self, args: argparse.Namespace) -> None:
-        if args.num_gpus is not None and args.num_gpus <= 0:
-            raise ValueError("Number of gpus must be positive")
-
-    def subparser_init(self, subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
-        from fastvideo.entrypoints.openai.api_server import (
-            DEFAULT_HOST,
-            DEFAULT_OUTPUT_DIR,
-            DEFAULT_PORT,
+        if not args.config:
+            raise ValueError("fastvideo serve requires --config PATH; use a nested "
+                             "serve config plus optional dotted overrides")
+        if not os.path.exists(args.config):
+            raise ValueError(f"Config file not found: {args.config}")
+        setattr(
+            args,
+            _VALIDATED_SERVE_CONFIG_ATTR,
+            build_serve_config(
+                args,
+                overrides=getattr(args, "_unknown", None),
+            ),
         )
 
+    def subparser_init(self, subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
         serve_parser = subparsers.add_parser(
             "serve",
             help="Start an OpenAI-compatible HTTP server",
-            usage=("fastvideo serve --model-path MODEL_PATH_OR_ID "
-                   "[--host HOST] [--port PORT] [OPTIONS]"),
-        )
-
-        serve_parser.add_argument(
-            "--host",
-            type=str,
-            default=DEFAULT_HOST,
-            help=f"Host to bind the server to (default: {DEFAULT_HOST})",
-        )
-        serve_parser.add_argument(
-            "--port",
-            type=int,
-            default=DEFAULT_PORT,
-            help=f"Port to listen on (default: {DEFAULT_PORT})",
-        )
-        serve_parser.add_argument(
-            "--output-dir",
-            type=str,
-            default=DEFAULT_OUTPUT_DIR,
-            help=("Directory for generated outputs "
-                  f"(default: {DEFAULT_OUTPUT_DIR})"),
+            usage="fastvideo serve --config SERVE_CONFIG [--dotted.override VALUE]",
         )
         serve_parser.add_argument(
             "--config",
             type=str,
             default="",
             required=False,
-            help="Read CLI options from a config JSON or YAML file.",
+            help="Path to a nested config JSON or YAML file. Required.",
         )
-
-        serve_parser = FastVideoArgs.add_cli_args(serve_parser)
         return cast(FlexibleArgumentParser, serve_parser)
 
 

fastvideo/entrypoints/openai/api_server.py

@@ -8,6 +8,8 @@ import uvicorn
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 
+from fastvideo.api.presets import validate_preset_selection
+from fastvideo.api.schema import GenerationRequest
 from fastvideo.entrypoints.openai.state import (
     DEFAULT_OUTPUT_DIR,
     clear_state,
@@ -16,6 +18,7 @@ from fastvideo.entrypoints.openai.state import (
 from fastvideo.entrypoints.video_generator import VideoGenerator
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
+from fastvideo.registry import get_preset_selection
 
 logger = init_logger(__name__)
 
@@ -23,17 +26,40 @@ DEFAULT_HOST = "0.0.0.0"
 DEFAULT_PORT = 8000
 
 
+def _validate_default_request_against_preset(
+    default_request: GenerationRequest,
+    model_path: str,
+) -> None:
+    """Validate ``default_request.stage_overrides`` against the model's preset.
+
+    Called once at server startup from :func:`run_server`. The
+    ``default_request`` is static server config, so validation results are
+    invariant across requests — there's no reason to re-run per request.
+    """
+    if not default_request.stage_overrides:
+        return
+    preset_name, model_family = get_preset_selection(model_path)
+    if preset_name is None or model_family is None:
+        return
+    validate_preset_selection(
+        preset_name,
+        model_family,
+        stage_overrides=default_request.stage_overrides,
+    )
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     """Load model on startup, clean up on shutdown"""
     args: FastVideoArgs = app.state.fastvideo_args
     output_dir: str = app.state.output_dir
+    default_request: GenerationRequest | None = getattr(app.state, "default_request", None)
 
     logger.info("Loading model from %s ...", args.model_path)
     generator = VideoGenerator.from_fastvideo_args(args)
     logger.info("Model loaded successfully.")
 
-    set_state(generator, args, output_dir)
+    set_state(generator, args, output_dir, default_request=default_request)
 
     yield  # server is running
 
@@ -46,6 +72,7 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
 def create_app(
     fastvideo_args: FastVideoArgs,
     output_dir: str = DEFAULT_OUTPUT_DIR,
+    default_request: GenerationRequest | None = None,
 ) -> FastAPI:
     """Build the FastAPI application with all routers mounted"""
 
@@ -56,6 +83,7 @@ def create_app(
     )
     app.state.fastvideo_args = fastvideo_args
     app.state.output_dir = output_dir
+    app.state.default_request = default_request
 
     app.add_middleware(
         CORSMiddleware,
@@ -108,9 +136,17 @@ def run_server(
     host: str = DEFAULT_HOST,
     port: int = DEFAULT_PORT,
     output_dir: str = DEFAULT_OUTPUT_DIR,
+    default_request: GenerationRequest | None = None,
 ):
     """Create the app and run it with uvicorn"""
-    app = create_app(fastvideo_args, output_dir=output_dir)
+    if default_request is not None:
+        _validate_default_request_against_preset(default_request, fastvideo_args.model_path)
+
+    app = create_app(
+        fastvideo_args,
+        output_dir=output_dir,
+        default_request=default_request,
+    )
 
     logger.info("Starting FastVideo server on %s:%d", host, port)
     logger.info("Model: %s", fastvideo_args.model_path)

fastvideo/entrypoints/openai/state.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
+    from fastvideo.api.schema import GenerationRequest
     from fastvideo.entrypoints.video_generator import VideoGenerator
     from fastvideo.fastvideo_args import FastVideoArgs
 
@@ -18,6 +19,7 @@ DEFAULT_OUTPUT_DIR = "outputs"
 _generator: VideoGenerator | None = None
 _fastvideo_args: FastVideoArgs | None = None
 _output_dir: str = DEFAULT_OUTPUT_DIR
+_default_request: GenerationRequest | None = None
 
 
 def get_generator() -> VideoGenerator:
@@ -37,20 +39,28 @@ def get_output_dir() -> str:
     return _output_dir
 
 
+def get_default_request() -> GenerationRequest | None:
+    """Return the ServeConfig.default_request set at startup, if any."""
+    return _default_request
+
+
 def set_state(
     generator: VideoGenerator,
     fastvideo_args: FastVideoArgs,
     output_dir: str,
+    default_request: GenerationRequest | None = None,
 ) -> None:
     """Set all server state at once (called from lifespan)."""
-    global _generator, _fastvideo_args, _output_dir
+    global _generator, _fastvideo_args, _output_dir, _default_request
     _generator = generator
     _fastvideo_args = fastvideo_args
     _output_dir = output_dir
+    _default_request = default_request
 
 
 def clear_state() -> None:
     """Clear server state on shutdown."""
-    global _generator, _fastvideo_args
+    global _generator, _fastvideo_args, _default_request
     _generator = None
     _fastvideo_args = None
+    _default_request = None

fastvideo/entrypoints/openai/video_api.py

@@ -19,7 +19,10 @@ from fastapi import (
 )
 from fastapi.responses import FileResponse
 
+from fastvideo.api.compat import explicit_request_updates
+from fastvideo.api.schema import GenerationRequest
 from fastvideo.entrypoints.openai.state import (
+    get_default_request,
     get_generator,
     get_output_dir,
     get_server_args,
@@ -42,49 +45,73 @@ logger = init_logger(__name__)
 router = APIRouter(prefix="/v1/videos", tags=["videos"])
 
 
-def _build_generation_kwargs(request_id: str, req: VideoGenerationsRequest) -> dict[str, Any]:
+def _build_generation_kwargs(
+    request_id: str,
+    req: VideoGenerationsRequest,
+    default_request: GenerationRequest | None = None,
+) -> dict[str, Any]:
+    """Build a flat kwargs dict for ``generator.generate_video``.
 
+    Precedence (highest to lowest):
+      1. Request body — only fields the client explicitly sent
+         (``req.model_fields_set``, Pydantic v2).
+      2. ``default_request`` — only fields the operator explicitly set in
+         the serve YAML, projected via ``explicit_request_updates``. Schema
+         defaults on the dataclass are *not* treated as defaults here.
+      3. Hardcoded fallback (e.g. ``fps=24`` when neither side set it).
+
+    Why gate on ``model_fields_set`` / explicit paths? Both the request
+    Pydantic model and the ``GenerationRequest`` dataclass carry schema
+    defaults (e.g. ``seed=1024``, ``num_frames=125``). Without the gate
+    those would masquerade as intent and shadow the other side — the
+    gate preserves "operator pinned it" vs. "dataclass happened to have
+    that default."
+    """
     kwargs: dict[str, Any] = {}
+    if default_request is not None:
+        kwargs.update(explicit_request_updates(default_request))
+
+    body_set = req.model_fields_set
     kwargs["prompt"] = req.prompt
 
-    # Resolution
-    if req.size:
+    if "size" in body_set and req.size:
         w, h = parse_size(req.size)
         if w is not None and h is not None:
             kwargs["width"] = w
             kwargs["height"] = h
 
-    # Frame count / duration
-    fps = req.fps if req.fps is not None else 24
-    kwargs["fps"] = fps
+    if "fps" in body_set and req.fps is not None:
+        kwargs["fps"] = req.fps
 
-    if req.num_frames is not None:
+    if "num_frames" in body_set and req.num_frames is not None:
         kwargs["num_frames"] = req.num_frames
-    elif req.seconds is not None:
+    elif "seconds" in body_set and req.seconds is not None:
+        fps = kwargs.get("fps", 24)
         kwargs["num_frames"] = fps * req.seconds
 
-    # Sampling parameters
-    if req.seed is not None:
+    if "seed" in body_set and req.seed is not None:
         kwargs["seed"] = req.seed
-    if req.num_inference_steps is not None:
+    if ("num_inference_steps" in body_set and req.num_inference_steps is not None):
         kwargs["num_inference_steps"] = req.num_inference_steps
-    if req.guidance_scale is not None:
+    if "guidance_scale" in body_set and req.guidance_scale is not None:
         kwargs["guidance_scale"] = req.guidance_scale
-    if req.guidance_scale_2 is not None:
+    if "guidance_scale_2" in body_set and req.guidance_scale_2 is not None:
         kwargs["guidance_scale_2"] = req.guidance_scale_2
-    if req.negative_prompt is not None:
+    if "negative_prompt" in body_set and req.negative_prompt is not None:
         kwargs["negative_prompt"] = req.negative_prompt
-    if req.enable_teacache:
+    if "enable_teacache" in body_set and req.enable_teacache:
         kwargs["enable_teacache"] = True
-    if req.true_cfg_scale is not None:
+    if "true_cfg_scale" in body_set and req.true_cfg_scale is not None:
         kwargs["true_cfg_scale"] = req.true_cfg_scale
 
-    # Image-to-video input
-    if req.input_reference is not None:
+    if "input_reference" in body_set and req.input_reference is not None:
         kwargs["image_path"] = req.input_reference
 
-    # Output path
-    output_dir = req.output_path or os.path.join(get_output_dir(), "videos")
+    kwargs.setdefault("fps", 24)
+
+    default_output_path = kwargs.pop("output_path", None)
+    body_output_dir = req.output_path if "output_path" in body_set else None
+    output_dir = body_output_dir or default_output_path or os.path.join(get_output_dir(), "videos")
     os.makedirs(output_dir, exist_ok=True)
     kwargs["output_path"] = os.path.join(output_dir, f"{request_id}.mp4")
     kwargs["save_video"] = True
@@ -272,7 +299,12 @@ async def create_video(
 
     logger.info("Video generation request %s: prompt=%s", request_id, req.prompt[:100])
 
-    gen_kwargs = _build_generation_kwargs(request_id, req)
+    # default_request was validated at server startup (run_server) and is
+    # read-only on the request hot path — _build_generation_kwargs and
+    # explicit_request_updates only read, so no per-request deepcopy needed.
+    default_request = get_default_request()
+
+    gen_kwargs = _build_generation_kwargs(request_id, req, default_request=default_request)
     job = _make_video_job(request_id, req, gen_kwargs)
     await VIDEO_STORE.upsert(request_id, job)
 

fastvideo/entrypoints/streaming/__init__.py

@@ -0,0 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
+from fastvideo.entrypoints.streaming.server import run_server
+
+__all__ = ["run_server"]

fastvideo/entrypoints/streaming/server.py

@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from fastvideo.api.schema import ServeConfig
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def run_server(serve_config: ServeConfig) -> None:
+    """Launch the streaming (WebSocket / Dynamo) server."""
+    if serve_config.streaming is None:
+        raise ValueError("ServeConfig.streaming must be set to launch the streaming server; "
+                         "got None. Add a `streaming:` block to your serve config.")
+    raise NotImplementedError("streaming server is not implemented yet")

fastvideo/entrypoints/streaming_generator.py

@@ -8,7 +8,7 @@ import torch
 import torchvision
 from einops import rearrange
 
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.entrypoints.video_generator import VideoGenerator
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger

fastvideo/entrypoints/video_generator.py

@@ -8,8 +8,12 @@ diffusion models.
 
 import os
 import re
+import shutil
 import threading
 import time
+import tempfile
+import warnings
+from collections.abc import Mapping
 from copy import deepcopy
 from typing import Any
 
@@ -18,10 +22,20 @@ import numpy as np
 import torch
 import torchvision
 from einops import rearrange
-import shutil
-import tempfile
 
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.compat import (
+    expand_request_prompt_batch,
+    generator_config_to_fastvideo_args,
+    legacy_from_pretrained_to_config,
+    load_generator_config_from_file,
+    normalize_generation_request,
+    normalize_generator_config,
+    request_to_pipeline_overrides,
+    request_to_sampling_param,
+)
+from fastvideo.api.results import GenerationResult
+from fastvideo.api.schema import GenerationRequest, GeneratorConfig
+from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
 from fastvideo.pipelines import ForwardBatch
@@ -30,6 +44,29 @@ from fastvideo.worker.executor import Executor
 
 logger = init_logger(__name__)
 
+_FROM_PRETRAINED_CONVENIENCE_KWARGS = frozenset({
+    "num_gpus",
+    "revision",
+    "trust_remote_code",
+    "distributed_executor_backend",
+    "tp_size",
+    "sp_size",
+    "hsdp_replicate_dim",
+    "hsdp_shard_dim",
+    "dist_timeout",
+    "use_fsdp_inference",
+    "disable_autocast",
+    "enable_stage_verification",
+    "dit_cpu_offload",
+    "dit_layerwise_offload",
+    "text_encoder_cpu_offload",
+    "image_encoder_cpu_offload",
+    "vae_cpu_offload",
+    "pin_cpu_memory",
+    "enable_torch_compile",
+    "torch_compile_kwargs",
+})
+
 
 def _infer_latent_batch_size(batch: ForwardBatch) -> int:
     if isinstance(batch.prompt, list):
@@ -52,19 +89,33 @@ class VideoGenerator:
     customization options, similar to popular frameworks like HF Diffusers.
     """
 
-    def __init__(self, fastvideo_args: FastVideoArgs, executor_class: type[Executor], log_stats: bool):
+    def __init__(
+        self,
+        fastvideo_args: FastVideoArgs,
+        executor_class: type[Executor],
+        log_stats: bool,
+        *,
+        log_queue=None,
+    ):
         """
         Initialize the video generator.
-        
+
         Args:
             fastvideo_args: The inference arguments
             executor_class: The executor class to use for inference
+            log_stats: Whether to log statistics
+            log_queue: Optional multiprocessing.Queue to forward worker logs to
         """
+        self.config: GeneratorConfig | None = None
         self.fastvideo_args = fastvideo_args
-        self.executor = executor_class(fastvideo_args)
+        self.executor = executor_class(fastvideo_args, log_queue=log_queue)
 
     @classmethod
-    def from_pretrained(cls, model_path: str, **kwargs) -> "VideoGenerator":
+    def from_pretrained(
+        cls,
+        model_path: str | GeneratorConfig | Mapping[str, Any] | None = None,
+        **kwargs,
+    ) -> "VideoGenerator":
         """
         Create a video generator from a pretrained model.
         
@@ -77,21 +128,84 @@ class VideoGenerator:
             The created video generator
 
         Priority level: Default pipeline config < User's pipeline config < User's kwargs
-        """
-        # If users also provide some kwargs, it will override the FastVideoArgs and PipelineConfig.
-        kwargs['model_path'] = model_path
-        fastvideo_args = FastVideoArgs.from_kwargs(**kwargs)
 
-        return cls.from_fastvideo_args(fastvideo_args)
+        Stable convenience kwargs remain supported here for common engine and
+        offload settings. Advanced model- or pipeline-specific options should
+        move to VideoGenerator.from_config(...).
+        """
+        log_queue = kwargs.pop("log_queue", None)
+        typed_config = kwargs.pop("config", None)
+        if typed_config is not None:
+            if model_path is not None:
+                raise TypeError("Pass either model_path or config to from_pretrained, not both")
+            if kwargs:
+                unexpected = ", ".join(sorted(kwargs))
+                raise TypeError(f"Unexpected keyword arguments with config: {unexpected}")
+            return cls.from_config(typed_config, log_queue=log_queue)
+
+        if isinstance(model_path, GeneratorConfig | Mapping):
+            if kwargs:
+                unexpected = ", ".join(sorted(kwargs))
+                raise TypeError(f"Unexpected keyword arguments with typed config: {unexpected}")
+            return cls.from_config(model_path, log_queue=log_queue)
+
+        if model_path is None:
+            raise TypeError("model_path or config is required")
+
+        legacy_only_kwargs = sorted(set(kwargs) - _FROM_PRETRAINED_CONVENIENCE_KWARGS)
+        if legacy_only_kwargs:
+            warnings.warn(
+                "VideoGenerator.from_pretrained(...) received legacy-only kwargs "
+                f"({', '.join(legacy_only_kwargs)}); prefer VideoGenerator.from_config(...) "
+                "for advanced configuration.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+        return cls.from_config(
+            legacy_from_pretrained_to_config(model_path, kwargs),
+            log_queue=log_queue,
+        )
 
     @classmethod
-    def from_fastvideo_args(cls, fastvideo_args: FastVideoArgs) -> "VideoGenerator":
+    def from_config(
+        cls,
+        config: GeneratorConfig | Mapping[str, Any],
+        *,
+        log_queue=None,
+    ) -> "VideoGenerator":
+        normalized = normalize_generator_config(config)
+        fastvideo_args = generator_config_to_fastvideo_args(normalized)
+        generator = cls.from_fastvideo_args(fastvideo_args, log_queue=log_queue)
+        generator.config = normalized
+        return generator
+
+    @classmethod
+    def from_file(
+        cls,
+        path: str,
+        overrides: list[str] | Mapping[str, Any] | None = None,
+        *,
+        log_queue=None,
+    ) -> "VideoGenerator":
+        return cls.from_config(
+            load_generator_config_from_file(path, overrides=overrides),
+            log_queue=log_queue,
+        )
+
+    @classmethod
+    def from_fastvideo_args(
+        cls,
+        fastvideo_args: FastVideoArgs,
+        *,
+        log_queue=None,
+    ) -> "VideoGenerator":
         """
         Create a video generator with the specified arguments.
-        
+
         Args:
             fastvideo_args: The inference arguments
-                
+            log_queue: Optional multiprocessing.Queue to forward worker logs to
+
         Returns:
             The created video generator
         """
@@ -103,8 +217,40 @@ class VideoGenerator:
             fastvideo_args=fastvideo_args,
             executor_class=executor_class,
             log_stats=False,  # TODO: implement
+            log_queue=log_queue,
         )
 
+    def generate(
+        self,
+        request: GenerationRequest | Mapping[str, Any],
+        *,
+        log_queue=None,
+    ) -> GenerationResult | list[GenerationResult]:
+        """
+        Generate video or image outputs from a typed inference request.
+
+        Args:
+            request: A `GenerationRequest` instance or a mapping that can be
+                parsed into one. This is the primary public inference
+                entrypoint for the typed API.
+            log_queue: Optional multiprocessing.Queue to forward worker logs to
+                during this request.
+
+        Returns:
+            A `GenerationResult` for single-request generation, or a list of
+            `GenerationResult` objects when the request expands into multiple
+            prompts.
+        """
+        normalized_request = normalize_generation_request(request)
+        if log_queue:
+            self.executor.set_log_queue(log_queue)
+
+        try:
+            return self._generate_request_impl(normalized_request)
+        finally:
+            if log_queue:
+                self.executor.clear_log_queue()
+
     def generate_video(
         self,
         prompt: str | None = None,
@@ -140,9 +286,93 @@ class VideoGenerator:
             A metadata dictionary for single-prompt generation, or a list of
             metadata dictionaries for prompt-file batch generation.
         """
+        log_queue = kwargs.pop("log_queue", None)
+        warnings.warn(
+            "VideoGenerator.generate_video(...) is deprecated; use "
+            "VideoGenerator.generate(request=...) instead.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        if log_queue:
+            self.executor.set_log_queue(log_queue)
+
+        try:
+            return self._generate_video_impl(
+                prompt=prompt,
+                sampling_param=sampling_param,
+                mouse_cond=mouse_cond,
+                keyboard_cond=keyboard_cond,
+                grid_sizes=grid_sizes,
+                **kwargs,
+            )
+        finally:
+            if log_queue:
+                self.executor.clear_log_queue()
+
+    def _generate_request_impl(
+        self,
+        request: GenerationRequest,
+    ) -> GenerationResult | list[GenerationResult]:
+        if isinstance(request.prompt, list):
+            if request.inputs.prompt_path is not None:
+                raise ValueError("request.prompt list cannot be combined with request.inputs.prompt_path")
+            results: list[GenerationResult] = []
+            for index, single_request in enumerate(expand_request_prompt_batch(request)):
+                prompt = single_request.prompt
+                wrapped = self._generate_single_request(single_request)
+                if isinstance(wrapped, list):
+                    results.extend(wrapped)
+                    continue
+                wrapped.prompt_index = index
+                if wrapped.prompt is None:
+                    wrapped.prompt = prompt
+                results.append(wrapped)
+            return results
+
+        return self._generate_single_request(request)
+
+    def _generate_single_request(
+        self,
+        request: GenerationRequest,
+    ) -> GenerationResult | list[GenerationResult]:
+        fastvideo_args = self.fastvideo_args
+        pipeline_overrides = request_to_pipeline_overrides(request)
+        if pipeline_overrides:
+            fastvideo_args = deepcopy(self.fastvideo_args)
+            for key, value in pipeline_overrides.items():
+                if not hasattr(fastvideo_args.pipeline_config, key):
+                    raise ValueError(f"Request field {key!r} is not supported by pipeline config overrides")
+                setattr(fastvideo_args.pipeline_config, key, deepcopy(value))
+
+        sampling_param = request_to_sampling_param(
+            request,
+            model_path=self.fastvideo_args.model_path,
+        )
+        result = self._generate_video_impl(
+            prompt=request.prompt,
+            sampling_param=sampling_param,
+            fastvideo_args=fastvideo_args,
+        )
+        return self._wrap_legacy_result(result)
+
+    def _generate_video_impl(
+        self,
+        prompt: str | None = None,
+        sampling_param: SamplingParam | None = None,
+        mouse_cond: torch.Tensor | None = None,
+        keyboard_cond: torch.Tensor | None = None,
+        grid_sizes: tuple[int, int, int] | list[int] | torch.Tensor
+        | None = None,
+        fastvideo_args: FastVideoArgs | None = None,
+        **kwargs,
+    ) -> dict[str, Any] | list[np.ndarray] | list[dict[str, Any]]:
+        """Internal implementation of generate_video."""
+        if fastvideo_args is None:
+            fastvideo_args = self.fastvideo_args
+
         # Handle batch processing from text file
         if sampling_param is None:
-            sampling_param = SamplingParam.from_pretrained(self.fastvideo_args.model_path)
+            sampling_param = SamplingParam.from_pretrained(fastvideo_args.model_path)
 
         # Add action control inputs to kwargs if provided
         if mouse_cond is not None:
@@ -154,9 +384,9 @@ class VideoGenerator:
 
         sampling_param.update(kwargs)
 
-        if self.fastvideo_args.prompt_txt is not None or sampling_param.prompt_path is not None:
-            prompt_txt_path = sampling_param.prompt_path or self.fastvideo_args.prompt_txt
-            if not os.path.exists(prompt_txt_path):
+        if fastvideo_args.prompt_txt is not None or sampling_param.prompt_path is not None:
+            prompt_txt_path = sampling_param.prompt_path or fastvideo_args.prompt_txt
+            if not prompt_txt_path or not os.path.exists(prompt_txt_path):
                 raise FileNotFoundError(f"Prompt text file not found: {prompt_txt_path}")
 
             # Read prompts from file
@@ -175,7 +405,12 @@ class VideoGenerator:
                     # Generate video for this prompt using the same logic below
                     output_path = self._prepare_output_path(sampling_param.output_path, batch_prompt)
                     kwargs["output_path"] = output_path
-                    result = self._generate_single_video(prompt=batch_prompt, sampling_param=sampling_param, **kwargs)
+                    result = self._generate_single_video(
+                        prompt=batch_prompt,
+                        sampling_param=sampling_param,
+                        fastvideo_args=fastvideo_args,
+                        **kwargs,
+                    )
 
                     # Add prompt info to result
                     result["prompt_index"] = i
@@ -196,7 +431,12 @@ class VideoGenerator:
             raise ValueError("Either prompt or prompt_txt must be provided")
         output_path = self._prepare_output_path(sampling_param.output_path, prompt)
         kwargs["output_path"] = output_path
-        return self._generate_single_video(prompt=prompt, sampling_param=sampling_param, **kwargs)
+        return self._generate_single_video(
+            prompt=prompt,
+            sampling_param=sampling_param,
+            fastvideo_args=fastvideo_args,
+            **kwargs,
+        )
 
     def _is_image_workload(self) -> bool:
         """Return True when the workload produces a single image (t2i, i2i …)."""
@@ -280,11 +520,12 @@ class VideoGenerator:
         self,
         prompt: str,
         sampling_param: SamplingParam | None = None,
+        fastvideo_args: FastVideoArgs | None = None,
         **kwargs,
     ) -> dict[str, Any]:
         """Internal method for single video generation"""
-        # Create a copy of inference args to avoid modifying the original
-        fastvideo_args = self.fastvideo_args
+        if fastvideo_args is None:
+            fastvideo_args = self.fastvideo_args
 
         # Validate inputs
         if not isinstance(prompt, str):
@@ -427,6 +668,20 @@ class VideoGenerator:
 
         return result
 
+    @staticmethod
+    def _wrap_legacy_result(
+        result: dict[str, Any] | list[dict[str, Any]], ) -> GenerationResult | list[GenerationResult]:
+        if isinstance(result, list):
+            return [GenerationResult.from_legacy_result(item) for item in result]
+        return GenerationResult.from_legacy_result(result)
+
+    @staticmethod
+    def _unwrap_typed_result(
+        result: GenerationResult | list[GenerationResult], ) -> dict[str, Any] | list[dict[str, Any]]:
+        if isinstance(result, list):
+            return [item.to_legacy_dict() for item in result]
+        return result.to_legacy_dict()
+
     @staticmethod
     def _mux_audio(
         video_path: str,

fastvideo/models/loader/component_loader.py

@@ -588,6 +588,33 @@ class TokenizerLoader(ComponentLoader):
 class VAELoader(ComponentLoader):
     """Loader for VAE."""
 
+    @staticmethod
+    def _find_gen3c_tokenizer_checkpoint(model_path: str) -> str | None:
+        """Locate tokenizer-backed VAE checkpoint used by GEN3C integration."""
+        candidates = [
+            os.path.join(model_path, "tokenizer.pth"),
+            os.path.join(os.path.dirname(model_path), "tokenizer",
+                         "tokenizer.pth"),
+        ]
+        for candidate in candidates:
+            if os.path.exists(candidate):
+                return candidate
+        return None
+
+    @staticmethod
+    def _find_gen3c_jit_tokenizer_dir(model_path: str) -> str | None:
+        """Locate official tokenizer JIT assets (encoder/decoder/mean_std)."""
+        candidates = [
+            model_path,
+            os.path.join(os.path.dirname(model_path), "tokenizer"),
+        ]
+        required = ("encoder.jit", "decoder.jit", "mean_std.pt")
+        for directory in candidates:
+            if all(os.path.exists(os.path.join(directory, name))
+                   for name in required):
+                return directory
+        return None
+
     def load(self, model_path: str, fastvideo_args: FastVideoArgs):
         """Load the VAE based on the model path, and inference args."""
         config = get_diffusers_config(model=model_path)
@@ -614,8 +641,68 @@ class VAELoader(ComponentLoader):
             if fastvideo_args.pipeline_config.vae_precision
             else torch.bfloat16
         ):
+            pipeline_name = fastvideo_args.pipeline_config.__class__.__name__
+            is_gen3c = pipeline_name.startswith("Gen3C")
+            is_cosmos25 = pipeline_name == "Cosmos25Config"
+
+            # GEN3C: prefer tokenizer-backed VAE checkpoint when available.
+            # This aligns latent conditioning with the GEN3C temporal contract.
+            if is_gen3c and class_name in (
+                    "AutoencoderKLWan", "AutoencoderKLGen3CTokenizer"):
+                from fastvideo.models.vaes.gen3c_tokenizer_vae import (
+                    AutoencoderKLGen3CTokenizer)
+
+                dtype = PRECISION_TO_TYPE[
+                    fastvideo_args.pipeline_config.vae_precision]
+                num_frames = int(
+                    getattr(fastvideo_args.pipeline_config, "num_frames", 121))
+                state_t = int(
+                    getattr(fastvideo_args.pipeline_config, "state_t", 16))
+                if state_t > 1 and num_frames > 1:
+                    target_temporal = max(1,
+                                          (num_frames - 1) // (state_t - 1))
+                else:
+                    target_temporal = 8
+
+                jit_dir = self._find_gen3c_jit_tokenizer_dir(model_path)
+                if jit_dir is not None:
+                    vae = AutoencoderKLGen3CTokenizer.from_jit_tokenizer(
+                        jit_dir,
+                        device=target_device,
+                        dtype=dtype,
+                        target_temporal_compression=target_temporal,
+                        pixel_chunk_duration=num_frames,
+                    )
+                    logger.info(
+                        "Loaded GEN3C tokenizer VAE from JIT assets in %s (target temporal compression=%d)",
+                        jit_dir,
+                        target_temporal,
+                    )
+                    return vae.eval()
+
+                tokenizer_ckpt = self._find_gen3c_tokenizer_checkpoint(
+                    model_path)
+                if tokenizer_ckpt is not None:
+                    vae = AutoencoderKLGen3CTokenizer.from_tokenizer_checkpoint(
+                        tokenizer_ckpt,
+                        device=target_device,
+                        dtype=dtype,
+                        target_temporal_compression=target_temporal,
+                        pixel_chunk_duration=num_frames,
+                    )
+                    logger.info(
+                        "Loaded GEN3C tokenizer VAE from %s (target temporal compression=%d)",
+                        tokenizer_ckpt,
+                        target_temporal,
+                    )
+                    return vae.eval()
+                logger.warning(
+                    "GEN3C tokenizer VAE checkpoint not found near %s; falling back to configured class %s.",
+                    model_path,
+                    class_name,
+                )
+
             # Cosmos2.5 uses a Wan2.1 VAE stored as `tokenizer.safetensors` under the VAE folder.
-            is_cosmos25 = fastvideo_args.pipeline_config.__class__.__name__ == "Cosmos25Config"
             if class_name == "AutoencoderKLWan" and is_cosmos25:
                 from fastvideo.models.vaes.cosmos25wanvae import Cosmos25WanVAE
 

fastvideo/models/registry.py

@@ -40,8 +40,8 @@ _TEXT_TO_VIDEO_DIT_MODELS = {
     "LTX2Transformer3DModel": ("dits", "ltx2", "LTX2Transformer3DModel"),
     "SD3Transformer2DModel": ("dits", "sd3", "SD3Transformer2DModel"),
     "LingBotWorldTransformer3DModel": ("dits", "lingbotworld", "LingBotWorldTransformer3DModel"),
-    "Kandinsky5Transformer3DModel":
-    ("dits", "kandinsky5", "Kandinsky5Transformer3DModel"),
+    "Gen3CTransformer3DModel": ("dits", "gen3c", "Gen3CTransformer3DModel"),
+    "Kandinsky5Transformer3DModel": ("dits", "kandinsky5", "Kandinsky5Transformer3DModel"),
 }
 
 _IMAGE_TO_VIDEO_DIT_MODELS = {
@@ -82,6 +82,9 @@ _VAE_MODELS = {
     "AutoencoderKLHunyuanVideo15": ("vaes", "hunyuan15vae", "AutoencoderKLHunyuanVideo15"),
     "AutoencoderKLWan": ("vaes", "wanvae", "AutoencoderKLWan"),
     "AutoencoderKL": ("vaes", "autoencoder_kl", "AutoencoderKL"),
+    "AutoencoderKLGen3CTokenizer":
+    ("vaes", "gen3c_tokenizer_vae", "AutoencoderKLGen3CTokenizer"),
+    "AutoencoderKLStepvideo": ("vaes", "stepvideovae", "AutoencoderKLStepvideo"),
     "CausalVideoAutoencoder": ("vaes", "ltx2vae", "LTX2CausalVideoAutoencoder"),
 }
 

fastvideo/models/vaes/gen3c_tokenizer_vae.py

@@ -0,0 +1,366 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+GEN3C tokenizer-backed VAE adapter.
+
+This wrapper loads the available tokenizer checkpoint (`tokenizer.pth`) and
+adapts it to GEN3C's latent-time contract (T=16 for 121 output frames).
+
+Why this exists:
+- The converted GEN3C bundle includes tokenizer-style VAE weights, not a
+  standard diffusers Wan VAE contract.
+- GEN3C diffusion expects 8x temporal compression (121 -> 16), while the
+  available tokenizer checkpoint follows a 4x temporal path.
+
+To bridge this at inference time, we:
+- keep the inner tokenizer model as-is,
+- downsample encoded latent time from inner-T to target-T for DiT input,
+- upsample generated latent time back to inner-T before decoding.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+import torch.nn as nn
+
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class _TensorLatentDist:
+    """Minimal distribution-like wrapper used by pipeline stages."""
+
+    mean: torch.Tensor
+
+    def mode(self) -> torch.Tensor:
+        return self.mean
+
+    def sample(self, generator: Any | None = None) -> torch.Tensor:
+        _ = generator
+        return self.mean
+
+
+class _JITGen3CTokenizerInner(nn.Module):
+    """Minimal wrapper around official tokenizer JIT encoder/decoder exports."""
+
+    def __init__(
+        self,
+        *,
+        encoder_path: str,
+        decoder_path: str,
+        mean_std_path: str,
+        dtype: torch.dtype,
+        device: torch.device,
+        latent_channels: int = 16,
+        latent_chunk_duration: int = 16,
+    ) -> None:
+        super().__init__()
+        self._dtype = dtype
+        self._forced_bf16 = False
+        self.encoder = torch.jit.load(encoder_path, map_location=device).eval().to(
+            device=device, dtype=dtype)
+        self.decoder = torch.jit.load(decoder_path, map_location=device).eval().to(
+            device=device, dtype=dtype)
+
+        latent_mean, latent_std = torch.load(mean_std_path, map_location="cpu")
+        latent_mean = latent_mean.view(latent_channels, -1)[:, :latent_chunk_duration]
+        latent_std = latent_std.view(latent_channels, -1)[:, :latent_chunk_duration]
+
+        self.register_buffer(
+            "_latent_mean",
+            latent_mean.to(torch.float32).view(1, latent_channels,
+                                               latent_chunk_duration, 1, 1),
+            persistent=False,
+        )
+        self.register_buffer(
+            "_latent_std",
+            latent_std.to(torch.float32).view(1, latent_channels,
+                                              latent_chunk_duration, 1, 1),
+            persistent=False,
+        )
+
+    def _match_stats(self, like: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        mean = self._latent_mean.to(device=like.device, dtype=like.dtype)
+        std = self._latent_std.to(device=like.device, dtype=like.dtype)
+        t = like.shape[2]
+        if mean.shape[2] == t:
+            return mean, std
+        if t < mean.shape[2]:
+            return mean[:, :, :t], std[:, :, :t]
+        # fallback for non-default lengths
+        mean = torch.nn.functional.interpolate(
+            mean, size=(t, 1, 1), mode="trilinear", align_corners=False)
+        std = torch.nn.functional.interpolate(
+            std, size=(t, 1, 1), mode="trilinear", align_corners=False)
+        return mean, std
+
+    @staticmethod
+    def _module_dtype_device(module: torch.nn.Module) -> tuple[torch.dtype, torch.device]:
+        for param in module.parameters():
+            return param.dtype, param.device
+        for buf in module.buffers():
+            return buf.dtype, buf.device
+        raise RuntimeError("Tokenizer JIT module has no parameters/buffers to infer dtype/device.")
+
+    def _coerce_modules_to_bf16(self) -> None:
+        if self._forced_bf16:
+            return
+        self.encoder = self.encoder.to(dtype=torch.bfloat16)
+        self.decoder = self.decoder.to(dtype=torch.bfloat16)
+        self._dtype = torch.bfloat16
+        self._forced_bf16 = True
+        logger.warning(
+            "GEN3C tokenizer JIT hit fp16/bf16 mismatch; coercing tokenizer encoder/decoder to bf16."
+        )
+
+    def encode(self, x: torch.Tensor) -> _TensorLatentDist:
+        enc_dtype, enc_device = self._module_dtype_device(self.encoder)
+        x_in = x.to(device=enc_device, dtype=enc_dtype)
+        try:
+            with torch.autocast(device_type=enc_device.type, enabled=False):
+                z = self.encoder(x_in)
+        except RuntimeError as e:
+            err = str(e)
+            mismatch_tokens = (
+                "Input type (CUDABFloat16Type) and weight type (torch.cuda.HalfTensor)",
+                "Input type (torch.cuda.HalfTensor) and weight type (CUDABFloat16Type)",
+            )
+            if any(token in err for token in mismatch_tokens):
+                self._coerce_modules_to_bf16()
+                enc_dtype, enc_device = self._module_dtype_device(self.encoder)
+                x_in = x.to(device=enc_device, dtype=enc_dtype)
+                with torch.autocast(device_type=enc_device.type, enabled=False):
+                    z = self.encoder(x_in)
+            else:
+                raise
+        if isinstance(z, tuple):
+            z = z[0]
+        z = z.to(dtype=x.dtype, device=x.device)
+        mean, std = self._match_stats(z)
+        return _TensorLatentDist((z - mean) / std)
+
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        mean, std = self._match_stats(z)
+        dec_dtype, dec_device = self._module_dtype_device(self.decoder)
+        z_in = (z * std + mean).to(device=dec_device, dtype=dec_dtype)
+        with torch.autocast(device_type=dec_device.type, enabled=False):
+            x = self.decoder(z_in)
+        if isinstance(x, tuple):
+            x = x[0]
+        return x.to(dtype=z.dtype, device=z.device)
+
+
+class AutoencoderKLGen3CTokenizer(nn.Module):
+    """
+    GEN3C VAE wrapper with temporal contract adaptation.
+
+    Interface contract:
+    - `encode(x)` returns normalized latents in the *target* temporal layout.
+    - `decode(z)` expects normalized latents in the *target* temporal layout.
+    """
+
+    handles_latent_norm: bool = True
+    handles_latent_denorm: bool = True
+
+    def __init__(
+        self,
+        inner: nn.Module,
+        *,
+        target_temporal_compression: int = 8,
+        inner_temporal_compression: int = 4,
+        spatial_compression_factor: int = 8,
+        pixel_chunk_duration: int = 121,
+    ) -> None:
+        super().__init__()
+        self.inner = inner
+        self.config = getattr(inner, "config", None)
+        self._target_temporal_compression = int(target_temporal_compression)
+        self._inner_temporal_compression = int(inner_temporal_compression)
+        self._spatial_compression_factor = int(spatial_compression_factor)
+        self._pixel_chunk_duration = int(pixel_chunk_duration)
+
+    @staticmethod
+    def _extract_latents(encoder_output: Any) -> torch.Tensor:
+        if hasattr(encoder_output, "latent_dist"):
+            dist = encoder_output.latent_dist
+            if hasattr(dist, "mode"):
+                return dist.mode()
+            if hasattr(dist, "mean"):
+                return dist.mean
+            return dist.sample()
+        if hasattr(encoder_output, "mode"):
+            return encoder_output.mode()
+        if hasattr(encoder_output, "latents"):
+            return encoder_output.latents
+        if hasattr(encoder_output, "sample"):
+            return encoder_output.sample()
+        if isinstance(encoder_output, torch.Tensor):
+            return encoder_output
+        raise TypeError(f"Unsupported encoder output type: {type(encoder_output)}")
+
+    def _inner_to_target_time(self, z_inner: torch.Tensor) -> torch.Tensor:
+        if z_inner.shape[2] <= 1:
+            return z_inner
+
+        # Common GEN3C case: inner=4x, target=8x => keep every other latent frame.
+        if self._target_temporal_compression == 2 * self._inner_temporal_compression:
+            return z_inner[:, :, 0::2, :, :].contiguous()
+
+        # Generic fallback: keep boundary latents and sample uniformly.
+        t_inner = z_inner.shape[2]
+        t_target = 1 + (t_inner - 1) * self._inner_temporal_compression // self._target_temporal_compression
+        idx = torch.linspace(0, t_inner - 1, t_target, device=z_inner.device)
+        idx = idx.round().long()
+        return z_inner.index_select(2, idx).contiguous()
+
+    def _target_to_inner_time(self, z_target: torch.Tensor) -> torch.Tensor:
+        if z_target.shape[2] <= 1:
+            return z_target
+
+        # Common GEN3C case: inner=4x, target=8x => insert midpoint frames.
+        if self._target_temporal_compression == 2 * self._inner_temporal_compression:
+            b, c, t, h, w = z_target.shape
+            t_inner = 2 * t - 1
+            out = torch.empty(
+                b, c, t_inner, h, w, device=z_target.device, dtype=z_target.dtype)
+            out[:, :, 0::2, :, :] = z_target
+            out[:, :, 1::2, :, :] = 0.5 * (
+                z_target[:, :, :-1, :, :] + z_target[:, :, 1:, :, :]
+            )
+            return out.contiguous()
+
+        # Generic fallback: linear index interpolation in time.
+        t_target = z_target.shape[2]
+        t_inner = 1 + (t_target - 1) * self._target_temporal_compression // self._inner_temporal_compression
+        idx = torch.linspace(0, t_target - 1, t_inner, device=z_target.device)
+        idx0 = idx.floor().long()
+        idx1 = idx.ceil().long().clamp_max(t_target - 1)
+        frac = (idx - idx0).view(1, 1, -1, 1, 1)
+        z0 = z_target.index_select(2, idx0)
+        z1 = z_target.index_select(2, idx1)
+        return (z0 * (1.0 - frac) + z1 * frac).contiguous()
+
+    def encode(self, x: torch.Tensor) -> _TensorLatentDist:
+        z_inner = self._extract_latents(self.inner.encode(x))
+        z_target = self._inner_to_target_time(z_inner)
+        return _TensorLatentDist(z_target)
+
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        z_inner = self._target_to_inner_time(z)
+        out = self.inner.decode(z_inner)
+        return out.sample if hasattr(out, "sample") else out
+
+    def enable_tiling(self) -> None:
+        if hasattr(self.inner, "enable_tiling"):
+            self.inner.enable_tiling()
+
+    def disable_tiling(self) -> None:
+        if hasattr(self.inner, "disable_tiling"):
+            self.inner.disable_tiling()
+
+    def get_latent_num_frames(self, num_pixel_frames: int) -> int:
+        num_pixel_frames = int(num_pixel_frames)
+        if num_pixel_frames <= 1:
+            return 1
+        return 1 + (num_pixel_frames - 1) // self._target_temporal_compression
+
+    def get_pixel_num_frames(self, num_latent_frames: int) -> int:
+        num_latent_frames = int(num_latent_frames)
+        if num_latent_frames <= 1:
+            return 1
+        return (num_latent_frames - 1) * self._target_temporal_compression + 1
+
+    @property
+    def spatial_compression_factor(self) -> int:
+        return self._spatial_compression_factor
+
+    @property
+    def temporal_compression_factor(self) -> int:
+        return self._target_temporal_compression
+
+    @property
+    def temporal_compression_ratio(self) -> int:
+        return self._target_temporal_compression
+
+    @property
+    def pixel_chunk_duration(self) -> int:
+        return self._pixel_chunk_duration
+
+    @property
+    def latent_chunk_duration(self) -> int:
+        return self.get_latent_num_frames(self._pixel_chunk_duration)
+
+    @classmethod
+    def from_tokenizer_checkpoint(
+        cls,
+        checkpoint_path: str,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        target_temporal_compression: int = 8,
+        pixel_chunk_duration: int = 121,
+    ) -> "AutoencoderKLGen3CTokenizer":
+        from fastvideo.models.vaes.cosmos25wanvae import Cosmos25WanVAE
+
+        inner = Cosmos25WanVAE(device=device, dtype=dtype)
+        loaded = torch.load(checkpoint_path, map_location="cpu")
+        if isinstance(loaded, dict):
+            for key in ("state_dict", "model", "ema", "model_state_dict"):
+                if key in loaded and isinstance(loaded[key], dict):
+                    loaded = loaded[key]
+                    break
+        missing, unexpected = inner.load_state_dict(loaded, strict=False)
+        if missing:
+            logger.warning(
+                "GEN3C tokenizer VAE missing keys (%d). Example: %s",
+                len(missing),
+                missing[:5],
+            )
+        if unexpected:
+            logger.warning(
+                "GEN3C tokenizer VAE unexpected keys (%d). Example: %s",
+                len(unexpected),
+                unexpected[:5],
+            )
+        return cls(
+            inner,
+            target_temporal_compression=target_temporal_compression,
+            inner_temporal_compression=4,
+            spatial_compression_factor=8,
+            pixel_chunk_duration=pixel_chunk_duration,
+        )
+
+    @classmethod
+    def from_jit_tokenizer(
+        cls,
+        tokenizer_dir: str,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        target_temporal_compression: int = 8,
+        pixel_chunk_duration: int = 121,
+    ) -> "AutoencoderKLGen3CTokenizer":
+        encoder_path = f"{tokenizer_dir}/encoder.jit"
+        decoder_path = f"{tokenizer_dir}/decoder.jit"
+        mean_std_path = f"{tokenizer_dir}/mean_std.pt"
+        inner = _JITGen3CTokenizerInner(
+            encoder_path=encoder_path,
+            decoder_path=decoder_path,
+            mean_std_path=mean_std_path,
+            dtype=dtype,
+            device=device,
+            latent_channels=16,
+            latent_chunk_duration=1 + (pixel_chunk_duration - 1) //
+            target_temporal_compression,
+        )
+        return cls(
+            inner,
+            target_temporal_compression=target_temporal_compression,
+            inner_temporal_compression=target_temporal_compression,
+            spatial_compression_factor=8,
+            pixel_chunk_duration=pixel_chunk_duration,
+        )

fastvideo/pipelines/basic/cosmos/presets.py

@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Cosmos model family pipeline presets.
+
+Covers both Cosmos Predict2 and Cosmos Predict2.5, which share the
+same pipeline directory but have distinct model families.
+"""
+from fastvideo.api.presets import InferencePreset, PresetStageSpec
+
+_DENOISE_STAGE = PresetStageSpec(
+    name="denoise",
+    kind="denoising",
+    description="Main denoising pass",
+    allowed_overrides=frozenset({
+        "num_inference_steps",
+        "guidance_scale",
+    }),
+)
+
+# -------------------------------------------------------------------
+# Cosmos Predict2
+# -------------------------------------------------------------------
+
+_COSMOS_NEGATIVE_PROMPT = ("The video captures a series of frames showing ugly scenes, "
+                           "static with no motion, motion blur, over-saturation, shaky "
+                           "footage, low resolution, grainy texture, pixelated images, "
+                           "poorly lit areas, underexposed and overexposed scenes, poor "
+                           "color balance, washed out colors, choppy sequences, jerky "
+                           "movements, low frame rate, artifacting, color banding, "
+                           "unnatural transitions, outdated special effects, fake elements, "
+                           "unconvincing visuals, poorly edited content, jump cuts, visual "
+                           "noise, and flickering. Overall, the video is of poor quality.")
+
+COSMOS_PREDICT2_2B = InferencePreset(
+    name="cosmos_predict2_2b",
+    version=1,
+    model_family="cosmos",
+    description="Cosmos Predict2 2B Video2World",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 704,
+        "width": 1280,
+        "num_frames": 93,
+        "fps": 16,
+        "guidance_scale": 7.0,
+        "num_inference_steps": 35,
+        "negative_prompt": _COSMOS_NEGATIVE_PROMPT,
+    },
+)
+
+# -------------------------------------------------------------------
+# Cosmos Predict2.5
+# -------------------------------------------------------------------
+
+_COSMOS25_NEGATIVE_PROMPT = ("The video captures a series of frames showing ugly scenes, "
+                             "static with no motion, motion blur, over-saturation, shaky "
+                             "footage, low resolution, grainy texture, pixelated images, "
+                             "poorly lit areas, underexposed and overexposed scenes, poor "
+                             "color balance, washed out colors, choppy sequences, jerky "
+                             "movements, low frame rate, artifacting, color banding, "
+                             "unnatural transitions, outdated special effects, fake elements, "
+                             "unconvincing visuals, poorly edited content, jump cuts, visual "
+                             "noise, and flickering. Overall, the video is of poor quality.")
+
+COSMOS25_PREDICT2_2B = InferencePreset(
+    name="cosmos25_predict2_2b",
+    version=1,
+    model_family="cosmos25",
+    description="Cosmos Predict2.5 2B",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "seed": 0,
+        "height": 704,
+        "width": 1280,
+        "num_frames": 77,
+        "fps": 24,
+        "guidance_scale": 7.0,
+        "num_inference_steps": 35,
+        "negative_prompt": _COSMOS25_NEGATIVE_PROMPT,
+    },
+)
+
+ALL_PRESETS = (COSMOS_PREDICT2_2B, COSMOS25_PREDICT2_2B)

fastvideo/pipelines/basic/gamecraft/presets.py

@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: Apache-2.0
+"""HunyuanGameCraft model family pipeline presets."""
+from fastvideo.api.presets import InferencePreset, PresetStageSpec
+
+_DENOISE_STAGE = PresetStageSpec(
+    name="denoise",
+    kind="denoising",
+    description="Action-controlled denoising pass",
+    allowed_overrides=frozenset({
+        "num_inference_steps",
+        "guidance_scale",
+    }),
+)
+
+GAMECRAFT_I2V = InferencePreset(
+    name="gamecraft_i2v",
+    version=1,
+    model_family="gamecraft",
+    description="HunyuanGameCraft I2V at 704x1280",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 704,
+        "width": 1280,
+        "num_frames": 33,
+        "fps": 24,
+        "guidance_scale": 6.0,
+        "num_inference_steps": 50,
+        "negative_prompt": "",
+    },
+)
+
+ALL_PRESETS = (GAMECRAFT_I2V, )

fastvideo/pipelines/basic/gen3c/__init__.py

@@ -0,0 +1,36 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+GEN3C is a 3D-informed world-consistent video generation model with precise camera control.
+"""
+
+from fastvideo.pipelines.basic.gen3c.cache_3d import (
+    Cache3DBase,
+    Cache3DBuffer,
+    forward_warp,
+    unproject_points,
+    project_points,
+)
+from fastvideo.pipelines.basic.gen3c.gen3c_pipeline import (
+    Gen3CPipeline,
+    Gen3CConditioningStage,
+    Gen3CDenoisingStage,
+    Gen3CLatentPreparationStage,
+)
+from fastvideo.pipelines.basic.gen3c.camera_utils import (
+    generate_camera_trajectory, )
+
+__all__ = [
+    # 3D Cache
+    "Cache3DBase",
+    "Cache3DBuffer",
+    "forward_warp",
+    "unproject_points",
+    "project_points",
+    # Camera
+    "generate_camera_trajectory",
+    # Pipeline
+    "Gen3CPipeline",
+    "Gen3CConditioningStage",
+    "Gen3CDenoisingStage",
+    "Gen3CLatentPreparationStage",
+]

fastvideo/pipelines/basic/gen3c/cache_3d.py

@@ -0,0 +1,720 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+This module implements the 3D cache system for GEN3C video generation with camera control.
+The cache maintains a point cloud representation of the scene, enabling:
+- Unprojecting depth maps to 3D world points
+- Forward warping rendered views to new camera poses
+- Managing multiple frame buffers for temporal consistency
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+
+def inverse_with_conversion(mtx: torch.Tensor) -> torch.Tensor:
+    """Compute matrix inverse with float32 conversion for numerical stability."""
+    return torch.linalg.inv(mtx.to(torch.float32)).to(mtx.dtype)
+
+
+def create_grid(b: int, h: int, w: int, device: str = "cpu", dtype: torch.dtype = torch.float32) -> torch.Tensor:
+    """
+    Create a dense grid of (x, y) coordinates of shape (b, 2, h, w).
+    
+    Args:
+        b: Batch size
+        h: Height
+        w: Width
+        device: Device for tensor creation
+        dtype: Data type for tensor
+        
+    Returns:
+        Grid tensor of shape (b, 2, h, w)
+    """
+    x = torch.arange(0, w, device=device, dtype=dtype).view(1, 1, 1, w).expand(b, 1, h, w)
+    y = torch.arange(0, h, device=device, dtype=dtype).view(1, 1, h, 1).expand(b, 1, h, w)
+    return torch.cat([x, y], dim=1)
+
+
+def unproject_points(
+    depth: torch.Tensor,
+    w2c: torch.Tensor,
+    intrinsic: torch.Tensor,
+    is_depth: bool = True,
+    mask: torch.Tensor | None = None,
+) -> torch.Tensor:
+    """
+    Unproject depth map to 3D world points.
+    
+    Args:
+        depth: (b, 1, h, w) depth map
+        w2c: (b, 4, 4) world-to-camera transformation matrix
+        intrinsic: (b, 3, 3) camera intrinsic matrix
+        is_depth: If True, depth is z-depth; if False, depth is distance to camera
+        mask: Optional (b, h, w) or (b, 1, h, w) mask for valid pixels
+        
+    Returns:
+        world_points: (b, h, w, 3) 3D world coordinates
+    """
+    b, _, h, w = depth.shape
+    device = depth.device
+    dtype = depth.dtype
+
+    if mask is None:
+        mask = depth > 0
+    if mask.dim() == depth.dim() and mask.shape[1] == 1:
+        mask = mask[:, 0]
+
+    idx = torch.nonzero(mask)
+    if idx.numel() == 0:
+        return torch.zeros((b, h, w, 3), device=device, dtype=dtype)
+
+    b_idx, y_idx, x_idx = idx[:, 0], idx[:, 1], idx[:, 2]
+
+    intrinsic_inv = inverse_with_conversion(intrinsic)  # (b, 3, 3)
+
+    x_valid = x_idx.to(dtype)
+    y_valid = y_idx.to(dtype)
+    ones = torch.ones_like(x_valid)
+    pos = torch.stack([x_valid, y_valid, ones], dim=1).unsqueeze(-1)  # (N, 3, 1)
+
+    intrinsic_inv_valid = intrinsic_inv[b_idx]  # (N, 3, 3)
+    unnormalized_pos = torch.matmul(intrinsic_inv_valid, pos)  # (N, 3, 1)
+
+    depth_valid = depth[b_idx, 0, y_idx, x_idx].view(-1, 1, 1)
+    if is_depth:
+        world_points_cam = depth_valid * unnormalized_pos
+    else:
+        norm_val = torch.norm(unnormalized_pos, dim=1, keepdim=True)
+        direction = unnormalized_pos / (norm_val + 1e-8)
+        world_points_cam = depth_valid * direction
+
+    ones_h = torch.ones((world_points_cam.shape[0], 1, 1), device=device, dtype=dtype)
+    world_points_homo = torch.cat([world_points_cam, ones_h], dim=1)  # (N, 4, 1)
+
+    trans = inverse_with_conversion(w2c)  # (b, 4, 4)
+    trans_valid = trans[b_idx]  # (N, 4, 4)
+    world_points_transformed = torch.matmul(trans_valid, world_points_homo)  # (N, 4, 1)
+    sparse_points = world_points_transformed[:, :3, 0]  # (N, 3)
+
+    out_points = torch.zeros((b, h, w, 3), device=device, dtype=dtype)
+    out_points[b_idx, y_idx, x_idx, :] = sparse_points
+    return out_points
+
+
+def project_points(
+    world_points: torch.Tensor,
+    w2c: torch.Tensor,
+    intrinsic: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Project 3D world points to 2D pixel coordinates.
+    
+    Args:
+        world_points: (b, h, w, 3) 3D world coordinates
+        w2c: (b, 4, 4) world-to-camera transformation matrix
+        intrinsic: (b, 3, 3) camera intrinsic matrix
+        
+    Returns:
+        projected_points: (b, h, w, 3, 1) projected 2D coordinates (x, y, z)
+    """
+    world_points = world_points.unsqueeze(-1)  # (b, h, w, 3, 1)
+    b, h, w, _, _ = world_points.shape
+
+    ones_4d = torch.ones((b, h, w, 1, 1), device=world_points.device, dtype=world_points.dtype)
+    world_points_homo = torch.cat([world_points, ones_4d], dim=3)  # (b, h, w, 4, 1)
+
+    trans_4d = w2c[:, None, None]  # (b, 1, 1, 4, 4)
+    camera_points_homo = torch.matmul(trans_4d, world_points_homo)  # (b, h, w, 4, 1)
+
+    camera_points = camera_points_homo[:, :, :, :3]  # (b, h, w, 3, 1)
+    intrinsic_4d = intrinsic[:, None, None]  # (b, 1, 1, 3, 3)
+    projected_points = torch.matmul(intrinsic_4d, camera_points)  # (b, h, w, 3, 1)
+
+    return projected_points
+
+
+def bilinear_splatting(
+    frame1: torch.Tensor,
+    mask1: torch.Tensor | None,
+    depth1: torch.Tensor,
+    flow12: torch.Tensor,
+    flow12_mask: torch.Tensor | None = None,
+    is_image: bool = False,
+    depth_weight_scale: float = 50.0,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Bilinear splatting for forward warping.
+    
+    Args:
+        frame1: (b, c, h, w) source frame
+        mask1: (b, 1, h, w) valid pixel mask (1 for known, 0 for unknown)
+        depth1: (b, 1, h, w) depth map
+        flow12: (b, 2, h, w) optical flow from frame1 to frame2
+        flow12_mask: (b, 1, h, w) flow validity mask
+        is_image: If True, output will be clipped to (-1, 1) range
+        depth_weight_scale: Scale factor for depth weighting
+        
+    Returns:
+        warped_frame2: (b, c, h, w) warped frame
+        mask2: (b, 1, h, w) validity mask for warped frame
+    """
+    b, c, h, w = frame1.shape
+    device = frame1.device
+    dtype = frame1.dtype
+
+    if mask1 is None:
+        mask1 = torch.ones(size=(b, 1, h, w), device=device, dtype=dtype)
+    if flow12_mask is None:
+        flow12_mask = torch.ones(size=(b, 1, h, w), device=device, dtype=dtype)
+
+    grid = create_grid(b, h, w, device=device, dtype=dtype)
+    trans_pos = flow12 + grid
+
+    trans_pos_offset = trans_pos + 1
+    trans_pos_floor = torch.floor(trans_pos_offset).long()
+    trans_pos_ceil = torch.ceil(trans_pos_offset).long()
+
+    trans_pos_offset = torch.stack(
+        [torch.clamp(trans_pos_offset[:, 0], min=0, max=w + 1),
+         torch.clamp(trans_pos_offset[:, 1], min=0, max=h + 1)],
+        dim=1)
+    trans_pos_floor = torch.stack(
+        [torch.clamp(trans_pos_floor[:, 0], min=0, max=w + 1),
+         torch.clamp(trans_pos_floor[:, 1], min=0, max=h + 1)],
+        dim=1)
+    trans_pos_ceil = torch.stack(
+        [torch.clamp(trans_pos_ceil[:, 0], min=0, max=w + 1),
+         torch.clamp(trans_pos_ceil[:, 1], min=0, max=h + 1)],
+        dim=1)
+
+    # Bilinear weights
+    prox_weight_nw = (1 - (trans_pos_offset[:, 1:2] - trans_pos_floor[:, 1:2])) * \
+                     (1 - (trans_pos_offset[:, 0:1] - trans_pos_floor[:, 0:1]))
+    prox_weight_sw = (1 - (trans_pos_ceil[:, 1:2] - trans_pos_offset[:, 1:2])) * \
+                     (1 - (trans_pos_offset[:, 0:1] - trans_pos_floor[:, 0:1]))
+    prox_weight_ne = (1 - (trans_pos_offset[:, 1:2] - trans_pos_floor[:, 1:2])) * \
+                     (1 - (trans_pos_ceil[:, 0:1] - trans_pos_offset[:, 0:1]))
+    prox_weight_se = (1 - (trans_pos_ceil[:, 1:2] - trans_pos_offset[:, 1:2])) * \
+                     (1 - (trans_pos_ceil[:, 0:1] - trans_pos_offset[:, 0:1]))
+
+    # Depth weighting for occlusion handling
+    clamped_depth1 = torch.clamp(depth1, min=0)
+    log_depth1 = torch.log1p(clamped_depth1)
+    exponent = log_depth1 / (log_depth1.max() + 1e-7) * depth_weight_scale
+    max_exponent = 80.0 if dtype in [torch.float32, torch.bfloat16] else 10.0
+    clamped_exponent = torch.clamp(exponent, max=max_exponent)
+    depth_weights = torch.exp(clamped_exponent) + 1e-7
+
+    weight_nw = torch.moveaxis(prox_weight_nw * mask1 * flow12_mask / depth_weights, [0, 1, 2, 3], [0, 3, 1, 2])
+    weight_sw = torch.moveaxis(prox_weight_sw * mask1 * flow12_mask / depth_weights, [0, 1, 2, 3], [0, 3, 1, 2])
+    weight_ne = torch.moveaxis(prox_weight_ne * mask1 * flow12_mask / depth_weights, [0, 1, 2, 3], [0, 3, 1, 2])
+    weight_se = torch.moveaxis(prox_weight_se * mask1 * flow12_mask / depth_weights, [0, 1, 2, 3], [0, 3, 1, 2])
+
+    warped_frame = torch.zeros(size=(b, h + 2, w + 2, c), dtype=dtype, device=device)
+    warped_weights = torch.zeros(size=(b, h + 2, w + 2, 1), dtype=dtype, device=device)
+
+    frame1_cl = torch.moveaxis(frame1, [0, 1, 2, 3], [0, 3, 1, 2])
+    batch_indices = torch.arange(b, device=device, dtype=torch.long)[:, None, None]
+
+    warped_frame.index_put_((batch_indices, trans_pos_floor[:, 1], trans_pos_floor[:, 0]),
+                            frame1_cl * weight_nw,
+                            accumulate=True)
+    warped_frame.index_put_((batch_indices, trans_pos_ceil[:, 1], trans_pos_floor[:, 0]),
+                            frame1_cl * weight_sw,
+                            accumulate=True)
+    warped_frame.index_put_((batch_indices, trans_pos_floor[:, 1], trans_pos_ceil[:, 0]),
+                            frame1_cl * weight_ne,
+                            accumulate=True)
+    warped_frame.index_put_((batch_indices, trans_pos_ceil[:, 1], trans_pos_ceil[:, 0]),
+                            frame1_cl * weight_se,
+                            accumulate=True)
+
+    warped_weights.index_put_((batch_indices, trans_pos_floor[:, 1], trans_pos_floor[:, 0]), weight_nw, accumulate=True)
+    warped_weights.index_put_((batch_indices, trans_pos_ceil[:, 1], trans_pos_floor[:, 0]), weight_sw, accumulate=True)
+    warped_weights.index_put_((batch_indices, trans_pos_floor[:, 1], trans_pos_ceil[:, 0]), weight_ne, accumulate=True)
+    warped_weights.index_put_((batch_indices, trans_pos_ceil[:, 1], trans_pos_ceil[:, 0]), weight_se, accumulate=True)
+
+    warped_frame_cf = torch.moveaxis(warped_frame, [0, 1, 2, 3], [0, 2, 3, 1])
+    warped_weights_cf = torch.moveaxis(warped_weights, [0, 1, 2, 3], [0, 2, 3, 1])
+    cropped_warped_frame = warped_frame_cf[:, :, 1:-1, 1:-1]
+    cropped_weights = warped_weights_cf[:, :, 1:-1, 1:-1]
+    cropped_weights = torch.nan_to_num(cropped_weights, nan=1000.0)
+
+    mask = cropped_weights > 0
+    zero_value = -1 if is_image else 0
+    zero_tensor = torch.tensor(zero_value, dtype=frame1.dtype, device=frame1.device)
+    warped_frame2 = torch.where(mask, cropped_warped_frame / cropped_weights, zero_tensor)
+    mask2 = mask.to(frame1)
+
+    if is_image:
+        warped_frame2 = torch.clamp(warped_frame2, min=-1, max=1)
+
+    return warped_frame2, mask2
+
+
+def forward_warp(
+    frame1: torch.Tensor,
+    mask1: torch.Tensor | None,
+    depth1: torch.Tensor | None,
+    transformation1: torch.Tensor | None,
+    transformation2: torch.Tensor,
+    intrinsic1: torch.Tensor | None,
+    intrinsic2: torch.Tensor | None,
+    is_image: bool = True,
+    is_depth: bool = True,
+    render_depth: bool = False,
+    world_points1: torch.Tensor | None = None,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None, torch.Tensor]:
+    """
+    Forward warp frame1 to a new view defined by transformation2.
+    
+    Args:
+        frame1: (b, c, h, w) source frame in range [-1, 1] for images
+        mask1: (b, 1, h, w) valid pixel mask
+        depth1: (b, 1, h, w) depth map (required if world_points1 is None)
+        transformation1: (b, 4, 4) source camera w2c (required if depth1 is provided)
+        transformation2: (b, 4, 4) target camera w2c
+        intrinsic1: (b, 3, 3) source camera intrinsics
+        intrinsic2: (b, 3, 3) target camera intrinsics
+        is_image: If True, output will be clipped to (-1, 1)
+        is_depth: If True, depth1 is z-depth; if False, it's distance
+        render_depth: If True, also return the warped depth map
+        world_points1: (b, h, w, 3) pre-computed world points (alternative to depth1)
+        
+    Returns:
+        warped_frame2: (b, c, h, w) warped frame
+        mask2: (b, 1, h, w) validity mask
+        warped_depth2: (b, h, w) warped depth (if render_depth=True)
+        flow12: (b, 2, h, w) optical flow
+    """
+    device = frame1.device
+    b, c, h, w = frame1.shape
+    dtype = frame1.dtype
+
+    if mask1 is None:
+        mask1 = torch.ones(size=(b, 1, h, w), device=device, dtype=dtype)
+    if intrinsic2 is None:
+        assert intrinsic1 is not None
+        intrinsic2 = intrinsic1.clone()
+
+    if world_points1 is not None:
+        # Use pre-computed world points
+        assert world_points1.shape == (b, h, w, 3)
+        trans_points1 = project_points(world_points1, transformation2, intrinsic2)
+    else:
+        # Compute from depth
+        assert depth1 is not None and transformation1 is not None
+        assert depth1.shape == (b, 1, h, w)
+
+        depth1 = torch.nan_to_num(depth1, nan=1e4)
+        depth1 = torch.clamp(depth1, min=0, max=1e4)
+
+        # Unproject to world, then project to target view
+        world_points1 = unproject_points(depth1, transformation1, intrinsic1, is_depth=is_depth)
+        trans_points1 = project_points(world_points1, transformation2, intrinsic2)
+
+    # Filter points behind camera
+    mask1 = mask1 * (trans_points1[:, :, :, 2, 0].unsqueeze(1) > 0)
+    trans_coordinates = trans_points1[:, :, :, :2, 0] / (trans_points1[:, :, :, 2:3, 0] + 1e-7)
+    trans_coordinates = trans_coordinates.permute(0, 3, 1, 2)  # b, 2, h, w
+    trans_depth1 = trans_points1[:, :, :, 2, 0].unsqueeze(1)
+
+    grid = create_grid(b, h, w, device=device, dtype=dtype)
+    flow12 = trans_coordinates - grid
+
+    warped_frame2, mask2 = bilinear_splatting(frame1, mask1, trans_depth1, flow12, None, is_image=is_image)
+
+    warped_depth2 = None
+    if render_depth:
+        warped_depth2 = bilinear_splatting(trans_depth1, mask1, trans_depth1, flow12, None, is_image=False)[0][:, 0]
+
+    return warped_frame2, mask2, warped_depth2, flow12
+
+
+def reliable_depth_mask_range_batch(
+    depth: torch.Tensor,
+    window_size: int = 5,
+    ratio_thresh: float = 0.05,
+    eps: float = 1e-6,
+) -> torch.Tensor:
+    """
+    Compute a mask for reliable depth values based on local variation.
+    
+    Args:
+        depth: (b, h, w) or (b, 1, h, w) depth map
+        window_size: Size of the local window (must be odd)
+        ratio_thresh: Threshold for depth variation ratio
+        eps: Small epsilon for numerical stability
+        
+    Returns:
+        reliable_mask: Boolean mask where True indicates reliable depth
+    """
+    assert window_size % 2 == 1, "Window size must be odd."
+
+    if depth.dim() == 3:
+        depth_unsq = depth.unsqueeze(1)
+    elif depth.dim() == 4:
+        depth_unsq = depth
+    else:
+        raise ValueError("depth tensor must be of shape (b, h, w) or (b, 1, h, w)")
+
+    local_max = F.max_pool2d(depth_unsq, kernel_size=window_size, stride=1, padding=window_size // 2)
+    local_min = -F.max_pool2d(-depth_unsq, kernel_size=window_size, stride=1, padding=window_size // 2)
+    local_mean = F.avg_pool2d(depth_unsq, kernel_size=window_size, stride=1, padding=window_size // 2)
+
+    ratio = (local_max - local_min) / (local_mean + eps)
+    reliable_mask = (ratio < ratio_thresh) & (depth_unsq > 0)
+
+    return reliable_mask
+
+
+class Cache3DBase:
+    """
+    Base class for 3D cache management.
+    
+    The cache maintains:
+    - input_image: RGB images stored in the cache
+    - input_points: 3D world coordinates for each pixel
+    - input_mask: Validity mask for each pixel
+    """
+
+    def __init__(
+        self,
+        input_image: torch.Tensor,
+        input_depth: torch.Tensor,
+        input_w2c: torch.Tensor,
+        input_intrinsics: torch.Tensor,
+        input_mask: torch.Tensor | None = None,
+        input_format: list[str] | None = None,
+        input_points: torch.Tensor | None = None,
+        weight_dtype: torch.dtype = torch.float32,
+        is_depth: bool = True,
+        device: str = "cuda",
+        filter_points_threshold: float = 1.0,
+    ):
+        """
+        Initialize the 3D cache.
+        
+        Args:
+            input_image: Input image tensor with varying dimensions
+            input_depth: Depth map tensor
+            input_w2c: World-to-camera transformation matrix
+            input_intrinsics: Camera intrinsic matrix
+            input_mask: Optional validity mask
+            input_format: Dimension labels for input_image (e.g., ['B', 'C', 'H', 'W'])
+            input_points: Pre-computed 3D world points (alternative to depth)
+            weight_dtype: Data type for computations
+            is_depth: If True, input_depth is z-depth; if False, it's distance
+            device: Computation device
+            filter_points_threshold: Threshold for filtering unreliable depth
+        """
+        self.weight_dtype = weight_dtype
+        self.is_depth = is_depth
+        self.device = device
+        self.filter_points_threshold = filter_points_threshold
+
+        if input_format is None:
+            assert input_image.dim() == 4
+            input_format = ["B", "C", "H", "W"]
+
+        # Map dimension names to indices
+        format_to_indices = {dim: idx for idx, dim in enumerate(input_format)}
+        input_shape = input_image.shape
+
+        if input_mask is not None:
+            input_image = torch.cat([input_image, input_mask], dim=format_to_indices.get("C"))
+
+        # Extract dimensions
+        B = input_shape[format_to_indices.get("B", 0)] if "B" in format_to_indices else 1
+        F = input_shape[format_to_indices.get("F", 0)] if "F" in format_to_indices else 1
+        N = input_shape[format_to_indices.get("N", 0)] if "N" in format_to_indices else 1
+        V = input_shape[format_to_indices.get("V", 0)] if "V" in format_to_indices else 1
+        H = input_shape[format_to_indices.get("H", 0)] if "H" in format_to_indices else None
+        W = input_shape[format_to_indices.get("W", 0)] if "W" in format_to_indices else None
+
+        # Reorder dimensions to B x F x N x V x C x H x W
+        desired_dims = ["B", "F", "N", "V", "C", "H", "W"]
+        permute_order: list[int | None] = []
+        for dim in desired_dims:
+            idx = format_to_indices.get(dim)
+            permute_order.append(idx)
+
+        permute_indices = [idx for idx in permute_order if idx is not None]
+        input_image = input_image.permute(*permute_indices)
+
+        for i, idx in enumerate(permute_order):
+            if idx is None:
+                input_image = input_image.unsqueeze(i)
+
+        # Now input_image has shape B x F x N x V x C x H x W
+        if input_mask is not None:
+            self.input_image, self.input_mask = input_image[:, :, :, :, :3], input_image[:, :, :, :, 3:]
+            self.input_mask = self.input_mask.to("cpu")
+        else:
+            self.input_mask = None
+            self.input_image = input_image
+        self.input_image = self.input_image.to(weight_dtype).to("cpu")
+
+        # Compute 3D world points
+        if input_points is not None:
+            self.input_points = input_points.reshape(B, F, N, V, H, W, 3).to("cpu")
+            self.input_depth = None
+        else:
+            input_depth = torch.nan_to_num(input_depth, nan=100)
+            input_depth = torch.clamp(input_depth, min=0, max=100)
+            if weight_dtype == torch.float16:
+                input_depth = torch.clamp(input_depth, max=70)
+
+            self.input_points = (unproject_points(
+                input_depth.reshape(-1, 1, H, W),
+                input_w2c.reshape(-1, 4, 4),
+                input_intrinsics.reshape(-1, 3, 3),
+                is_depth=self.is_depth,
+            ).to(weight_dtype).reshape(B, F, N, V, H, W, 3).to("cpu"))
+            self.input_depth = input_depth
+
+        # Filter unreliable depth
+        if self.filter_points_threshold < 1.0 and input_depth is not None:
+            input_depth = input_depth.reshape(-1, 1, H, W)
+            depth_mask = reliable_depth_mask_range_batch(input_depth,
+                                                         ratio_thresh=self.filter_points_threshold).reshape(
+                                                             B, F, N, V, 1, H, W)
+            if self.input_mask is None:
+                self.input_mask = depth_mask.to("cpu")
+            else:
+                self.input_mask = self.input_mask * depth_mask.to(self.input_mask.device)
+
+    def update_cache(self, **kwargs):
+        """Update the cache with new frames. To be implemented by subclasses."""
+        raise NotImplementedError
+
+    def input_frame_count(self) -> int:
+        """Return the number of frames in the cache."""
+        return self.input_image.shape[1]
+
+    def render_cache(
+        self,
+        target_w2cs: torch.Tensor,
+        target_intrinsics: torch.Tensor,
+        render_depth: bool = False,
+        start_frame_idx: int = 0,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Render the cached 3D points from new camera viewpoints.
+        
+        Args:
+            target_w2cs: (b, F_target, 4, 4) target camera transformations
+            target_intrinsics: (b, F_target, 3, 3) target camera intrinsics
+            render_depth: If True, return depth instead of RGB
+            start_frame_idx: Starting frame index in the cache
+            
+        Returns:
+            pixels: (b, F_target, N, c, h, w) rendered images or depth
+            masks: (b, F_target, N, 1, h, w) validity masks
+        """
+        bs, F_target, _, _ = target_w2cs.shape
+        B, F, N, V, C, H, W = self.input_image.shape
+        assert bs == B
+
+        target_w2cs = target_w2cs.reshape(B, F_target, 1, 4, 4).expand(B, F_target, N, 4, 4).reshape(-1, 4, 4)
+        target_intrinsics = target_intrinsics.reshape(B, F_target, 1, 3, 3).expand(B, F_target, N, 3,
+                                                                                   3).reshape(-1, 3, 3)
+
+        # Prepare inputs
+        first_images = rearrange(
+            self.input_image[:, start_frame_idx:start_frame_idx + F_target].expand(B, F_target, N, V, C, H, W),
+            "B F N V C H W -> (B F N) V C H W")
+        first_points = rearrange(
+            self.input_points[:, start_frame_idx:start_frame_idx + F_target].expand(B, F_target, N, V, H, W, 3),
+            "B F N V H W C -> (B F N) V H W C")
+        first_masks = rearrange(
+            self.input_mask[:, start_frame_idx:start_frame_idx + F_target].expand(B, F_target, N, V, 1, H, W),
+            "B F N V C H W -> (B F N) V C H W") if self.input_mask is not None else None
+
+        # Process in chunks for memory efficiency
+        if first_images.shape[1] == 1:
+            warp_chunk_size = 2
+            rendered_warp_images = []
+            rendered_warp_masks = []
+            rendered_warp_depth = []
+
+            first_images = first_images.squeeze(1)
+            first_points = first_points.squeeze(1)
+            first_masks = first_masks.squeeze(1) if first_masks is not None else None
+
+            for i in range(0, first_images.shape[0], warp_chunk_size):
+                with torch.no_grad():
+                    imgs_chunk = first_images[i:i + warp_chunk_size].to(self.device, non_blocking=True)
+                    pts_chunk = first_points[i:i + warp_chunk_size].to(self.device, non_blocking=True)
+                    masks_chunk = (first_masks[i:i + warp_chunk_size].to(self.device, non_blocking=True)
+                                   if first_masks is not None else None)
+
+                    (
+                        rendered_warp_images_chunk,
+                        rendered_warp_masks_chunk,
+                        rendered_warp_depth_chunk,
+                        _,
+                    ) = forward_warp(
+                        imgs_chunk,
+                        mask1=masks_chunk,
+                        depth1=None,
+                        transformation1=None,
+                        transformation2=target_w2cs[i:i + warp_chunk_size],
+                        intrinsic1=target_intrinsics[i:i + warp_chunk_size],
+                        intrinsic2=target_intrinsics[i:i + warp_chunk_size],
+                        render_depth=render_depth,
+                        world_points1=pts_chunk,
+                    )
+
+                    rendered_warp_images.append(rendered_warp_images_chunk.to("cpu"))
+                    rendered_warp_masks.append(rendered_warp_masks_chunk.to("cpu"))
+                    if render_depth:
+                        rendered_warp_depth.append(rendered_warp_depth_chunk.to("cpu"))
+
+                    del imgs_chunk, pts_chunk, masks_chunk
+                    torch.cuda.empty_cache()
+
+            rendered_warp_images = torch.cat(rendered_warp_images, dim=0)
+            rendered_warp_masks = torch.cat(rendered_warp_masks, dim=0)
+            if render_depth:
+                rendered_warp_depth = torch.cat(rendered_warp_depth, dim=0)
+        else:
+            raise NotImplementedError("Multi-view rendering not yet supported")
+
+        pixels = rearrange(rendered_warp_images, "(b f n) c h w -> b f n c h w", b=bs, f=F_target, n=N)
+        masks = rearrange(rendered_warp_masks, "(b f n) c h w -> b f n c h w", b=bs, f=F_target, n=N)
+
+        if render_depth:
+            pixels = rearrange(rendered_warp_depth, "(b f n) h w -> b f n h w", b=bs, f=F_target, n=N)
+
+        return pixels.to(self.device), masks.to(self.device)
+
+
+class Cache3DBuffer(Cache3DBase):
+    """
+    3D cache with frame buffer support.
+    
+    This class manages multiple frame buffers for temporal consistency
+    and supports noise augmentation for training stability.
+    """
+
+    def __init__(
+        self,
+        frame_buffer_max: int = 2,
+        noise_aug_strength: float = 0.0,
+        generator: torch.Generator | None = None,
+        **kwargs,
+    ):
+        """
+        Initialize the buffered 3D cache.
+        
+        Args:
+            frame_buffer_max: Maximum number of frames to buffer
+            noise_aug_strength: Strength of noise augmentation per buffer
+            generator: Random generator for reproducibility
+            **kwargs: Arguments passed to Cache3DBase
+        """
+        super().__init__(**kwargs)
+        self.frame_buffer_max = frame_buffer_max
+        self.noise_aug_strength = noise_aug_strength
+        self.generator = generator
+
+    def update_cache(
+        self,
+        new_image: torch.Tensor,
+        new_depth: torch.Tensor,
+        new_w2c: torch.Tensor,
+        new_mask: torch.Tensor | None = None,
+        new_intrinsics: torch.Tensor | None = None,
+    ):
+        """
+        Update the cache with a new frame.
+        
+        Args:
+            new_image: (B, C, H, W) new RGB image
+            new_depth: (B, 1, H, W) new depth map
+            new_w2c: (B, 4, 4) new world-to-camera transformation
+            new_mask: Optional (B, 1, H, W) validity mask
+            new_intrinsics: (B, 3, 3) camera intrinsics (optional)
+        """
+        new_image = new_image.to(self.weight_dtype).to(self.device)
+        new_depth = new_depth.to(self.weight_dtype).to(self.device)
+        new_w2c = new_w2c.to(self.weight_dtype).to(self.device)
+        if new_intrinsics is not None:
+            new_intrinsics = new_intrinsics.to(self.weight_dtype).to(self.device)
+
+        new_depth = torch.nan_to_num(new_depth, nan=1e4)
+        new_depth = torch.clamp(new_depth, min=0, max=1e4)
+
+        B, F, N, V, C, H, W = self.input_image.shape
+
+        # Compute new 3D points
+        new_points = unproject_points(new_depth, new_w2c, new_intrinsics, is_depth=self.is_depth).cpu()
+        new_image = new_image.cpu()
+
+        if self.filter_points_threshold < 1.0:
+            new_depth = new_depth.reshape(-1, 1, H, W)
+            depth_mask = reliable_depth_mask_range_batch(new_depth,
+                                                         ratio_thresh=self.filter_points_threshold).reshape(B, 1, H, W)
+            new_mask = depth_mask.to("cpu") if new_mask is None else new_mask * depth_mask.to(new_mask.device)
+        if new_mask is not None:
+            new_mask = new_mask.cpu()
+
+        # Update buffer (newest frame first)
+        if self.frame_buffer_max > 1:
+            if self.input_image.shape[2] < self.frame_buffer_max:
+                self.input_image = torch.cat([new_image[:, None, None, None], self.input_image], 2)
+                self.input_points = torch.cat([new_points[:, None, None, None], self.input_points], 2)
+                if self.input_mask is not None:
+                    self.input_mask = torch.cat([new_mask[:, None, None, None], self.input_mask], 2)
+            else:
+                self.input_image[:, :, 0] = new_image[:, None, None]
+                self.input_points[:, :, 0] = new_points[:, None, None]
+                if self.input_mask is not None:
+                    self.input_mask[:, :, 0] = new_mask[:, None, None]
+        else:
+            self.input_image = new_image[:, None, None, None]
+            self.input_points = new_points[:, None, None, None]
+
+    def render_cache(
+        self,
+        target_w2cs: torch.Tensor,
+        target_intrinsics: torch.Tensor,
+        render_depth: bool = False,
+        start_frame_idx: int = 0,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Render the cache with optional noise augmentation.
+        
+        Args:
+            target_w2cs: (b, F_target, 4, 4) target camera transformations
+            target_intrinsics: (b, F_target, 3, 3) target camera intrinsics
+            render_depth: If True, return depth instead of RGB
+            start_frame_idx: Starting frame index (must be 0 for this class)
+            
+        Returns:
+            pixels: (b, F_target, N, c, h, w) rendered images
+            masks: (b, F_target, N, 1, h, w) validity masks
+        """
+        assert start_frame_idx == 0, "start_frame_idx must be 0 for Cache3DBuffer"
+
+        output_device = target_w2cs.device
+        target_w2cs = target_w2cs.to(self.weight_dtype).to(self.device)
+        target_intrinsics = target_intrinsics.to(self.weight_dtype).to(self.device)
+
+        pixels, masks = super().render_cache(target_w2cs, target_intrinsics, render_depth)
+
+        pixels = pixels.to(output_device)
+        masks = masks.to(output_device)
+
+        # Apply noise augmentation (stronger for older buffers)
+        if not render_depth and self.noise_aug_strength > 0:
+            noise = torch.randn(pixels.shape, generator=self.generator, device=pixels.device, dtype=pixels.dtype)
+            per_buffer_noise = (torch.arange(start=pixels.shape[2] - 1, end=-1, step=-1, device=pixels.device) *
+                                self.noise_aug_strength)
+            pixels = pixels + noise * per_buffer_noise.reshape(1, 1, -1, 1, 1, 1)
+
+        return pixels, masks

fastvideo/pipelines/basic/gen3c/camera_utils.py

@@ -0,0 +1,203 @@
+# SPDX-License-Identifier: Apache-2.0
+# Ported from NVIDIA GEN3C: cosmos_predict1/diffusion/inference/camera_utils.py
+"""Camera trajectory generation utilities for GEN3C 3D cache conditioning."""
+
+import math
+
+import torch
+
+
+def apply_transformation(Bx4x4: torch.Tensor, another_matrix: torch.Tensor) -> torch.Tensor:
+    """Apply batch transformation to a matrix."""
+    B = Bx4x4.shape[0]
+    if another_matrix.dim() == 2:
+        another_matrix = another_matrix.unsqueeze(0).expand(B, -1, -1)
+    return torch.bmm(Bx4x4, another_matrix)
+
+
+def look_at_matrix(camera_pos: torch.Tensor, target: torch.Tensor, invert_pos: bool = True) -> torch.Tensor:
+    """Create a 4x4 look-at view matrix pointing camera toward target."""
+    forward = (target - camera_pos).float()
+    forward = forward / torch.norm(forward)
+
+    up = torch.tensor([0.0, 1.0, 0.0], device=camera_pos.device)
+    right = torch.cross(up, forward)
+    right = right / torch.norm(right)
+    up = torch.cross(forward, right)
+
+    look_at = torch.eye(4, device=camera_pos.device)
+    look_at[0, :3] = right
+    look_at[1, :3] = up
+    look_at[2, :3] = forward
+    look_at[:3, 3] = (-camera_pos) if invert_pos else camera_pos
+
+    return look_at
+
+
+def create_horizontal_trajectory(
+    world_to_camera_matrix: torch.Tensor,
+    center_depth: float,
+    positive: bool = True,
+    n_steps: int = 13,
+    distance: float = 0.1,
+    device: str = "cuda",
+    axis: str = "x",
+    camera_rotation: str = "center_facing",
+) -> torch.Tensor:
+    """Create a linear camera trajectory along a specified axis."""
+    look_at_target = torch.tensor([0.0, 0.0, center_depth]).to(device)
+    trajectory = []
+    initial_camera_pos = torch.tensor([0, 0, 0], device=device, dtype=torch.float32)
+
+    translation_positions = []
+    for i in range(n_steps):
+        offset = i * distance * center_depth / n_steps * (1 if positive else -1)
+        if axis == "x":
+            pos = torch.tensor([offset, 0, 0], device=device)
+        elif axis == "y":
+            pos = torch.tensor([0, offset, 0], device=device)
+        elif axis == "z":
+            pos = torch.tensor([0, 0, offset], device=device)
+        else:
+            raise ValueError(f"Axis should be x, y or z, got {axis}")
+        translation_positions.append(pos)
+
+    for pos in translation_positions:
+        camera_pos = initial_camera_pos + pos
+        if camera_rotation == "trajectory_aligned":
+            _look_at = look_at_target + pos * 2
+        elif camera_rotation == "center_facing":
+            _look_at = look_at_target
+        elif camera_rotation == "no_rotation":
+            _look_at = look_at_target + pos
+        else:
+            raise ValueError(f"camera_rotation should be center_facing, trajectory_aligned, "
+                             f"or no_rotation, got {camera_rotation}")
+        view_matrix = look_at_matrix(camera_pos, _look_at)
+        trajectory.append(view_matrix)
+
+    trajectory = torch.stack(trajectory)
+    return apply_transformation(trajectory, world_to_camera_matrix)
+
+
+def create_spiral_trajectory(
+    world_to_camera_matrix: torch.Tensor,
+    center_depth: float,
+    radius_x: float = 0.03,
+    radius_y: float = 0.02,
+    radius_z: float = 0.0,
+    positive: bool = True,
+    camera_rotation: str = "center_facing",
+    n_steps: int = 13,
+    device: str = "cuda",
+    start_from_zero: bool = True,
+    num_circles: int = 1,
+) -> torch.Tensor:
+    """Create a spiral/circular camera trajectory."""
+    look_at_target = torch.tensor([0.0, 0.0, center_depth]).to(device)
+    trajectory = []
+    initial_camera_pos = torch.tensor([0, 0, 0], device=device, dtype=torch.float32)
+
+    theta_max = 2 * math.pi * num_circles
+    spiral_positions = []
+
+    for i in range(n_steps):
+        theta = theta_max * i / (n_steps - 1)
+        if start_from_zero:
+            x = radius_x * (math.cos(theta) - 1) * (1 if positive else -1) * center_depth
+        else:
+            x = radius_x * math.cos(theta) * center_depth
+        y = radius_y * math.sin(theta) * center_depth
+        z = radius_z * math.sin(theta) * center_depth
+        spiral_positions.append(torch.tensor([x, y, z], device=device))
+
+    for pos in spiral_positions:
+        camera_pos = initial_camera_pos + pos
+        if camera_rotation == "center_facing":
+            view_matrix = look_at_matrix(camera_pos, look_at_target)
+        elif camera_rotation == "trajectory_aligned":
+            view_matrix = look_at_matrix(camera_pos, look_at_target + pos * 2)
+        elif camera_rotation == "no_rotation":
+            view_matrix = look_at_matrix(camera_pos, look_at_target + pos)
+        else:
+            raise ValueError(f"camera_rotation should be center_facing, trajectory_aligned, "
+                             f"or no_rotation, got {camera_rotation}")
+        trajectory.append(view_matrix)
+
+    trajectory = torch.stack(trajectory)
+    return apply_transformation(trajectory, world_to_camera_matrix)
+
+
+def generate_camera_trajectory(
+    trajectory_type: str,
+    initial_w2c: torch.Tensor,
+    initial_intrinsics: torch.Tensor,
+    num_frames: int,
+    movement_distance: float,
+    camera_rotation: str = "center_facing",
+    center_depth: float = 1.0,
+    device: str = "cuda",
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Generate camera trajectory for GEN3C video generation.
+
+    Args:
+        trajectory_type: One of "left", "right", "up", "down", "zoom_in",
+            "zoom_out", "clockwise", "counterclockwise".
+        initial_w2c: Initial world-to-camera matrix (4, 4).
+        initial_intrinsics: Camera intrinsics matrix (3, 3).
+        num_frames: Number of frames in the trajectory.
+        movement_distance: Distance factor for camera movement.
+        camera_rotation: "center_facing", "no_rotation", or "trajectory_aligned".
+        center_depth: Depth of the scene center point.
+        device: Computation device.
+
+    Returns:
+        generated_w2cs: (1, num_frames, 4, 4) world-to-camera matrices.
+        generated_intrinsics: (1, num_frames, 3, 3) camera intrinsics.
+    """
+    if trajectory_type in ["clockwise", "counterclockwise"]:
+        new_w2cs_seq = create_spiral_trajectory(
+            world_to_camera_matrix=initial_w2c,
+            center_depth=center_depth,
+            n_steps=num_frames,
+            positive=trajectory_type == "clockwise",
+            device=device,
+            camera_rotation=camera_rotation,
+            radius_x=movement_distance,
+            radius_y=movement_distance,
+        )
+    elif trajectory_type == "none":
+        # Static camera - repeat identity
+        new_w2cs_seq = initial_w2c.unsqueeze(0).expand(num_frames, -1, -1)
+    else:
+        axis_map = {
+            "left": (False, "x"),
+            "right": (True, "x"),
+            "up": (False, "y"),
+            "down": (True, "y"),
+            "zoom_in": (True, "z"),
+            "zoom_out": (False, "z"),
+        }
+        if trajectory_type not in axis_map:
+            raise ValueError(f"Unsupported trajectory type: {trajectory_type}")
+        positive, axis = axis_map[trajectory_type]
+
+        new_w2cs_seq = create_horizontal_trajectory(
+            world_to_camera_matrix=initial_w2c,
+            center_depth=center_depth,
+            n_steps=num_frames,
+            positive=positive,
+            axis=axis,
+            distance=movement_distance,
+            device=device,
+            camera_rotation=camera_rotation,
+        )
+
+    generated_w2cs = new_w2cs_seq.unsqueeze(0)  # (1, num_frames, 4, 4)
+    if initial_intrinsics.dim() == 2:
+        generated_intrinsics = initial_intrinsics.unsqueeze(0).unsqueeze(0).repeat(1, num_frames, 1, 1)
+    else:
+        generated_intrinsics = initial_intrinsics.unsqueeze(0)
+
+    return generated_w2cs, generated_intrinsics