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/memory/codebase-map/README.md

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

.agents/skills/index.jsonl

@@ -6,5 +6,3 @@
 {"name": "index-related-work", "description": "Ingest a paper or repository into the related work index", "path": "index-related-work/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "search-related-work", "description": "Query the related work index for relevant papers, repos, or comparisons", "path": "search-related-work/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "seed-ssim-references", "description": "Run a new or updated fastvideo/tests/ssim/ test on Modal, pull generated videos, and upload them to FastVideo/ssim-reference-videos so the test has a regression baseline", "path": "seed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}
-{"name": "reseed-ssim-references", "description": "Re-seed (overwrite) HF reference videos for an existing fastvideo/tests/ssim/ test and a single model id on Modal L40S. Always backs up current refs first, regenerates on Modal, pauses for the user to eyeball before-vs-after, then uploads with --force scoped to --model-id. Sister skill to seed-ssim-references; use when intentional code change has invalidated existing refs", "path": "reseed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}
-{"name": "reseed-performance-baseline", "description": "Re-seed the HF performance-tracking baseline for an intentional runtime, dependency, or environment-caused benchmark shift. Use when performance CI fails because metrics such as latency, throughput, component time, or peak memory changed for an accepted reason and the rolling median baseline must be advanced by replicating one reviewed shifted source result into three success=true records, or five records when explicitly requested", "path": "reseed-performance-baseline/SKILL.md", "status": "draft", "trust": "low"}

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

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

.agents/skills/reseed-performance-baseline/SKILL.md

@@ -1,426 +0,0 @@
----
-name: reseed-performance-baseline
-description: Re-seed the HF performance-tracking baseline for an intentional runtime, dependency, or environment-caused benchmark shift. Use when performance CI fails because metrics such as latency, throughput, component time, or peak memory changed for an accepted reason and the rolling median baseline in FastVideo/performance-tracking must be advanced by replicating one reviewed shifted source result into three success=true records, or five records when explicitly requested.
----
-
-# Re-seed Performance Baseline
-
-## Purpose
-
-Replace or advance the rolling performance baseline for a single
-`(model_id, gpu_type)` pair in the HF dataset
-`FastVideo/performance-tracking`.
-
-Performance comparison uses the median of up to the last 5 successful records
-for the same model and GPU. Failed records are useful audit history, but they
-do not move the future baseline because `compare_baseline.py` loads records
-with `successful_only=True`.
-
-For a 5-record median, one shifted record is not enough to move the median if
-the other four records are from the old runtime. This skill therefore creates
-3 reviewed `success=true` records from one accepted shifted source result by
-default. If the user explicitly asks for a full reset, create 5 records.
-
-These replicated records are an intentional operator-approved baseline reset,
-not independent measurements. Mark them clearly with provenance fields so the
-HF history remains auditable.
-
-Use this skill when a performance test fails for an intentional and reviewed
-reason, such as a torch/runtime/container upgrade that legitimately increases
-peak memory or changes timings. This is the performance equivalent of
-`reseed-ssim-references`: backup first, scope tightly, require explicit human
-approval, then upload reviewed accepted baseline records.
-
-## When to use
-
-- A PR or main run failed the rolling performance comparison by more than the
-  allowed regression threshold, and maintainers agree the shift is caused by
-  an intentional runtime, dependency, hardware image, or benchmark environment
-  change rather than a FastVideo logic regression.
-- One shifted source result has been reviewed and accepted, and the operator
-  wants to replicate it into 3 successful records so the rolling median moves
-  immediately. Use 5 records only when the user explicitly asks to fully reset
-  the last-5 window.
-
-## When not to use
-
-- The benchmark failure might be a real code regression. Fix or investigate
-  the code path first.
-- The fixed benchmark thresholds in
-  `.buildkite/performance-benchmarks/tests/*.json` are too low. Those are a
-  separate gate from the rolling HF baseline and may need a code review change.
-- There is no clear source run, commit, and rationale. Baseline history is a
-  production signal; do not edit it without provenance.
-
-## Inputs
-
-| Parameter | Required | Description |
-|-----------|----------|-------------|
-| `model_id` | Yes | Benchmark id, e.g. `wan-t2v-1.3b-2gpu`. This maps to the HF subdirectory after `sanitize(model_id)`. |
-| `gpu_type` | Yes | Exact GPU device string from the performance record, e.g. the L40S device name emitted by CI. Baselines are GPU-specific. |
-| `source_result` | Yes | Path or Buildkite artifact URL for one accepted shifted performance JSON. Prefer the normalized `normalized_perf_*.json` artifact emitted by `compare_baseline.py`. |
-| `replica_count` | No | Number of success records to create from `source_result`. Default: `3`. Only use `5` if the user explicitly asks for a full reset. |
-| `intent_rationale` | Yes | One-line explanation for why the baseline shift is legitimate. This is written into provenance and should be reused in the PR. |
-
-Hardcoded defaults:
-
-- HF repo: `FastVideo/performance-tracking` (`HF_REPO_ID` override is
-  supported by the code, but use the default unless the user explicitly asks).
-- Local sync root: `/tmp/perf-tracking` or a timestamped local backup under
-  `performance_reseed_backup/`.
-- Baseline window: last 5 `success=true` records for the same
-  `(model_id, gpu_type)`.
-- Default reseed count: 3 replicated `success=true` records from one reviewed
-  source result. Explicit full-reset count: 5.
-
-## Steps
-
-### 1. Validate the target and source result
-
-If `source_result` is a Buildkite artifact URL, download it first into a
-local scratch directory such as `performance_reseed_source/` and use that
-downloaded JSON path for the rest of the workflow. If the agent cannot access
-the artifact because Buildkite authentication is missing, ask the user to
-download the artifact manually and provide the local path.
-
-Prefer the normalized Buildkite artifact emitted by `compare_baseline.py`:
-
-```text
-perf_reports/results/normalized_perf_*.json
-```
-
-That file is already in the HF tracking schema. Load it directly and confirm
-it has the expected baseline fields:
-
-```python
-import json
-
-with open(source_result, encoding="utf-8") as f:
-    record = json.load(f)
-```
-
-If only the older raw `fastvideo/tests/performance/results/perf_*.json`
-artifact is available, normalize it with `compare_baseline.py`'s shared helper
-before continuing. Run this from the repository root with
-`PYTHONPATH=fastvideo/tests/performance` so the script-local `hf_store` import
-resolves the same way it does in CI:
-
-```python
-import json
-from compare_baseline import normalize_performance_result
-
-with open(source_result, encoding="utf-8") as f:
-    record = normalize_performance_result(json.load(f))
-```
-
-The raw-to-normalized helper maps:
-
-- `model_id` comes from `benchmark_id`.
-- `gpu_type` comes from `device`.
-- `memory` comes from `max_peak_memory_mb`.
-- `latency` comes from `avg_generation_time_s`.
-- `throughput` comes from `throughput_fps`.
-- component timings come from the raw `text_encoder_time_s`, `dit_time_s`,
-  and `vae_decode_time_s` fields when present. If an older raw artifact lacks
-  those keys, they normalize to `None`; that source can still reseed latency,
-  throughput, and memory, but it cannot move component-time baselines.
-
-Stop if the normalized record's `model_id` or `gpu_type` does not match the
-requested `model_id` and `gpu_type`.
-
-The source record may have `success: false` when it came from a failed rolling
-baseline comparison. That is expected; only the reviewed reseed replicas become
-new `success: true` baseline records after explicit approval.
-
-Set `replica_count` to `3` by default. Set it to `5` only when the user
-explicitly asks to upload the same shifted source result 5 times for a full
-last-5 reset. Reject other counts unless the user gives a concrete reason.
-
-Check that `HF_API_KEY` is exported. The sync path may be public, but the
-upload path requires write access.
-
-### 1a. How to obtain `source_result` from CI
-
-The performance CI exports normalized source results for failed rolling
-baseline comparisons when `compare_baseline.py` ran. The preferred artifact
-comes from:
-
-```text
-perf_reports/results/normalized_perf_*.json
-```
-
-and is uploaded by Buildkite with the performance reports. The normal operator
-flow is:
-
-1. Open the failed Buildkite performance job.
-2. Download the `normalized_perf_*.json` artifact for the failed benchmark.
-3. Pass the local path or artifact URL as `source_result`.
-
-Do not scrape the Markdown performance summary to reconstruct the JSON. The
-normalized JSON artifact is the source of truth for reseed metrics and
-provenance. If only a raw `fastvideo/tests/performance/results/perf_*.json`
-artifact is present, normalize it with `normalize_performance_result()` before
-continuing. If no JSON artifact is present, the benchmark likely failed before
-writing results, so that run is not a valid source for baseline reseeding.
-
-### 2. Sync and back up existing HF records
-
-Use `fastvideo/tests/performance/hf_store.py` helpers directly. Do **not** use
-`compare_baseline.py` as a sync shortcut; on full main runs it can persist
-records, while this step must only fetch and back up existing history.
-
-The sync command pattern is:
-
-```bash
-export PERFORMANCE_TRACKING_ROOT="${PERFORMANCE_TRACKING_ROOT:-/tmp/perf-tracking}"
-export HF_REPO_ID="${HF_REPO_ID:-FastVideo/performance-tracking}"
-PYTHONPATH=fastvideo/tests/performance python -c 'from hf_store import sync_from_hf; import os; sync_from_hf(os.environ["PERFORMANCE_TRACKING_ROOT"], strict=True)'
-```
-
-Then back up only the sanitized model directory:
-
-```bash
-SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
-TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
-MODEL_SAFE=$(python - <<'PY'
-from fastvideo.tests.performance.hf_store import sanitize
-print(sanitize("<model_id>"))
-PY
-)
-BACKUP_DIR="performance_reseed_backup/${TIMESTAMP}_${SHORT_COMMIT}_${MODEL_SAFE}"
-mkdir -p "$BACKUP_DIR"
-cp -R "${PERFORMANCE_TRACKING_ROOT}/${MODEL_SAFE}" "$BACKUP_DIR/" 2>/dev/null || true
-```
-
-Write provenance next to the backup:
-
-```bash
-cat > "$BACKUP_DIR/PROVENANCE.txt" <<EOF
-model_id: <model_id>
-gpu_type: <gpu_type>
-source_result: <source_result>
-replica_count: <3_or_5>
-head_commit: $(git rev-parse HEAD)
-timestamp_utc: $(date -u +%FT%TZ)
-reason: <intent_rationale>
-EOF
-```
-
-If the backup has no prior records, this is not a destructive reseed; it is a
-first baseline seed. Continue, but report that baseline history was empty.
-
-### 3. Compute old baseline and candidate shift
-
-Load the last 5 successful records for the target:
-
-```python
-from fastvideo.tests.performance.hf_store import load_records_for_model
-
-records = load_records_for_model(
-    "/tmp/perf-tracking",
-    "<model_id>",
-    "<gpu_type>",
-    last_n=5,
-    successful_only=True,
-)
-```
-
-Print a small table showing the source result metrics, the replicated
-candidate median, and the old medians for:
-
-- `latency`
-- `throughput`
-- `memory`
-- `text_encoder_time_s`
-- `dit_time_s`
-- `vae_decode_time_s`
-
-Also print how many successful old records exist. Make clear:
-
-- 1 shifted record only seeds audit history and usually does not move the
-  median.
-- 3 replicated shifted records in a 5-record window move the median
-  immediately.
-- 5 replicated shifted records fully reset the rolling window to the source
-  result's runtime profile.
-- Replicated records are not independent measurements; they are an intentional
-  approved baseline reset and must be labeled that way.
-
-### 4. Confirm intent
-
-Require an explicit confirmation phrase before preparing the upload:
-
-> About to RE-SEED performance baseline for `<model_id>` on `<gpu_type>`.
-> This will upload `<N>` new `success=true` records to
-> `FastVideo/performance-tracking/<sanitize(model_id)>/`.
->
-> Reason: `<intent_rationale>`
-> Source result: `<source_result>`
-> Replica count: `<replica_count>`
-> Note: these records replicate one reviewed measurement to force the rolling
-> median to the accepted runtime profile.
-> HEAD: `<git rev-parse --short=12 HEAD>`
-> Backup: `<BACKUP_DIR>`
->
-> Reply `confirm performance reseed` to proceed, anything else to abort.
-
-Do not continue unless the user types exactly `confirm performance reseed`.
-
-### 5. Create the accepted seed records
-
-Create `replica_count` normalized records from the single source result. Use
-an explicit allowlist; do not copy the raw result JSON wholesale.
-
-Each record must include only these baseline fields plus the reseed provenance
-fields below:
-
-- `model_id`
-- `timestamp`
-- `commit_sha`
-- `gpu_type`
-- `latency`
-- `throughput`
-- `memory`
-- `text_encoder_time_s`
-- `dit_time_s`
-- `vae_decode_time_s`
-- `success: true`
-
-For normalized `normalized_perf_*.json` sources, these fields already exist.
-For older raw `perf_*.json` sources, map the raw fields exactly as
-`normalize_performance_result()` in `compare_baseline.py` does:
-
-| Normalized field | Raw source field |
-|------------------|------------------|
-| `model_id` | `benchmark_id` |
-| `gpu_type` | `device` |
-| `latency` | `avg_generation_time_s` |
-| `throughput` | `throughput_fps` |
-| `memory` | `max_peak_memory_mb` |
-| `text_encoder_time_s` | `text_encoder_time_s` |
-| `dit_time_s` | `dit_time_s` |
-| `vae_decode_time_s` | `vae_decode_time_s` |
-| `commit_sha` | `commit` |
-
-Do not upload raw-only fields such as `model_short_name`, `num_gpus`,
-`num_warmup_runs`, `num_measurement_runs`, `individual_times_s`,
-`individual_peak_memories_mb`, `thresholds`, or `pr_number`.
-
-Optional provenance fields are allowed and useful:
-
-- `baseline_reseed: true`
-- `baseline_reseed_reason`
-- `baseline_reseed_source_result`
-- `baseline_reseed_source_timestamp`
-- `baseline_reseed_replicated_source: true`
-- `baseline_reseed_batch_size`
-- `baseline_reseed_batch_index`
-- `baseline_reseed_operator`
-
-Use a fresh reseed timestamp for each replicated record, not the original
-source result timestamp. This is required because
-`load_records_for_model(..., last_n=5)` keeps the last records after loading
-the model directory; stale filenames/timestamps may not enter the last-5
-window and therefore may not move the median. Preserve the original source
-timestamp in `baseline_reseed_source_timestamp`.
-
-Use the existing filename convention from `_write_tracking_record()`:
-`<sanitize(timestamp)>_<sanitize(commit_sha)>.json` under the sanitized model
-directory, but include a deterministic suffix such as `_reseed_01`,
-`_reseed_02`, and `_reseed_03` before `.json` so the replicated files do not
-overwrite each other. For a 5-record full reset, continue through
-`_reseed_05`.
-
-If the source record already exists on HF with `success=false`, do not edit it
-in place unless the user explicitly asked for an audit-preserving correction.
-Prefer uploading new accepted seed records so failed history remains visible.
-
-### 6. Pause before upload
-
-Print:
-
-- Backup directory path.
-- HF paths that will receive the new records.
-- Old rolling medians.
-- Source metrics, replica count, and candidate median.
-- Rationale.
-
-Ask the user to reply exactly `upload`. Anything else aborts and leaves the
-prepared records plus backup on disk.
-
-### 7. Upload only the scoped records
-
-Use the shared storage helper so the path and repo type match CI:
-
-```python
-from fastvideo.tests.performance.hf_store import upload_record
-
-upload_record("<local_record_path>", record, strict=True)
-```
-
-Run it once per prepared record. Each upload goes to:
-
-```text
-FastVideo/performance-tracking/<sanitize(model_id)>/<record_filename>.json
-```
-
-Never bulk upload the whole tracking root. Never modify another model's
-directory in the same operation.
-
-### 8. Report outcome
-
-Report:
-
-- Uploaded HF paths.
-- Backup directory.
-- Old baseline window count and medians.
-- Source metrics, replica count, and candidate median.
-- Expected effect: 3 replicated shifted records move the 5-record median; 5
-  replicated shifted records fully reset the window to the accepted source
-  result.
-- Any separate threshold changes still needed in
-  `.buildkite/performance-benchmarks/tests/*.json`.
-
-Include the `intent_rationale` in the PR or follow-up comment so reviewers can
-distinguish an accepted baseline shift from a hidden regression.
-
-## Failure modes and handling
-
-- **`HF_API_KEY` unset.** Stop before upload. Do not create an untracked
-  process that appears to have reseeded but never reached HF.
-- **Source result does not match target.** Stop. The wrong benchmark or GPU
-  would poison a separate baseline.
-- **`replica_count` is 5 but the user did not explicitly ask for a full
-  reset.** Stop and use the default count of 3.
-- **The source result is noisy or suspicious.** Stop. Replicating one result
-  amplifies that measurement into the baseline, so it must be reviewed first.
-- **HF sync fails.** Stop for destructive reseeds. A stale or empty sync can
-  make the old baseline look missing.
-- **Candidate still violates fixed thresholds.** Report that this skill only
-  handles the rolling HF baseline; update benchmark JSON thresholds in code
-  review if maintainers accept the new absolute limit.
-- **The user aborts at either confirmation.** Leave the backup and prepared
-  records on disk. Nothing should be uploaded.
-- **A bad seed was uploaded.** Use the backup and HF history to identify the
-  uploaded file, then remove or supersede it with an explicitly reviewed
-  corrective record. Do not silently rewrite unrelated history.
-
-## References
-
-- `.agents/skills/reseed-ssim-references/SKILL.md` — safety pattern for
-  intentional baseline replacement.
-- `fastvideo/tests/performance/compare_baseline.py` — normalization, rolling
-  median comparison, and persistence rules.
-- `fastvideo/tests/performance/hf_store.py` — HF sync, record loading,
-  `sanitize()`, and `upload_record()`.
-- `fastvideo/tests/performance/test_inference_performance.py` — source result
-  JSON schema.
-- `.buildkite/performance-benchmarks/tests/*.json` — fixed absolute benchmark
-  thresholds, separate from rolling baseline comparisons.
-
-## Changelog
-
-| Date | Change |
-|------|--------|
-| 2026-05-03 | Initial version. Sister workflow to `reseed-ssim-references`, scoped to one performance `(model_id, gpu_type)` baseline seed with backup, confirmation, provenance, and `success=true` upload. |
-| 2026-05-03 | Current policy: replicate one approved shifted source result into 3 success records by default, or 5 only when explicitly requested. Add provenance marker for replicated-source reseeds. |

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

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

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

@@ -1,41 +1,26 @@
 ---
 name: seed-ssim-references
-description: Seed HF reference artefacts for a single newly-added SSIM test (pixel `.mp4` for `run_text_to_video_similarity_test`-style tests, or latent `.pt` for `run_text_to_latent_similarity_test`-style tests). Runs the test on Modal L40S, downloads the generated artefacts via `modal volume get`, pauses for the user to verify (visual eyeball for mp4, numerics dump for pt), 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.
+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 Artefacts (mp4 or pt)
+# Seed SSIM Reference Videos
 
 ## Purpose
 
 A brand-new SSIM test in `fastvideo/tests/ssim/` fails forever until its
-reference artefacts exist on the HF dataset
-(`FastVideo/ssim-reference-videos`). The dataset hosts two kinds of artefacts
-side-by-side per `(model_id, backend, prompt)`:
-
-- **`.mp4`** — pixel ground-truth for tests that call
-  `run_text_to_video_similarity_test` / `run_image_to_video_similarity_test`
-  in `inference_similarity_utils.py`. Compared via SSIM.
-- **`.pt`** — pre-VAE latent bundle (fp16 full latent + fp32 slice +
-  metadata + `slice_spec` + `format_version`) for tests that call
-  `run_text_to_latent_similarity_test` in `latent_similarity_utils.py`.
-  Compared via cosine distance on the slice and the full tensor.
-
+reference videos exist on the HF dataset (`FastVideo/ssim-reference-videos`).
 This skill:
 
-1. Detects which artefact type the test produces (pixel vs latent).
-2. Runs the test on Modal's L40S pool to generate the artefacts.
-3. Downloads them to the local repo via `modal volume get`.
-4. Pauses so the user can verify quality:
-   - **mp4**: visual eyeball in a video player.
-   - **pt**: numerics dump (shape, slice stats, NaN/Inf check, metadata).
-5. Uploads only the new test's files to HF, with a guard that refuses to
+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 artefact without crashing (the missing-reference assertion at
-the end is expected). The skill does not re-test locally; it goes straight
-to Modal L40S (which is what CI uses).
+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
 
@@ -84,7 +69,7 @@ Fail fast if the token env var is missing.
 
 ## Steps
 
-### 1. Ask for the test file, then detect artefact type
+### 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`)"*.
@@ -95,22 +80,6 @@ Validate:
 - File defines a `*_MODEL_TO_PARAMS` dict — grep it to extract the set of
   model ids. Those ids drive step 5.
 
-Detect artefact type by inspecting the file's imports / helper call:
-
-- **latent** (`.pt`) — file imports `run_text_to_latent_similarity_test`
-  from `fastvideo.tests.ssim.latent_similarity_utils` (or any other helper
-  that ends with `_latent_similarity_test`).
-- **pixel** (`.mp4`) — file imports
-  `run_text_to_video_similarity_test` / `run_image_to_video_similarity_test`
-  from `fastvideo.tests.ssim.inference_similarity_utils`, OR uses the
-  legacy custom-inline helper pattern (see `test_gamecraft`,
-  `test_longcat`, etc.). Default to pixel when both heuristics fail.
-
-Record `ARTEFACT_TYPE ∈ {pixel, latent}` for use in step 4. Steps 2, 3, 5,
-and 6 are artefact-type-agnostic — `_iter_reference_files`,
-`copy_generated_to_reference`, and `upload_reference_videos` already walk
-both `.mp4` and `.pt` (see `reference_videos_cli.py`).
-
 If either check fails, stop and tell the user what's wrong.
 
 ### 2. Run the test on Modal L40S
@@ -123,19 +92,9 @@ TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
 SUBDIR="${TIMESTAMP}_${SHORT_COMMIT}"
 ```
 
-Then launch the Modal run. The `IMAGE_VERSION` and `BUILDKITE_*` env-prefix
-**must** match what CI exports in `.buildkite/scripts/pr_test.sh`, otherwise
-`fastvideo/tests/modal/ssim_test.py` resolves a different GHCR image tag
-(default is `latest`, CI is `py3.12-latest`) and bakes different values into
-the image's frozen env block (`ssim_test.py:17-18, 38-46`). Mismatched image
-or env produces SSIM drift that doesn't show up until the same commit runs
-in CI.
+Then launch the Modal run:
 
 ```bash
-IMAGE_VERSION="py3.12-latest" \
-BUILDKITE_REPO="$(git config --get remote.origin.url)" \
-BUILDKITE_COMMIT="$(git rev-parse HEAD)" \
-BUILDKITE_PULL_REQUEST="${BUILDKITE_PULL_REQUEST:-false}" \
 modal run fastvideo/tests/modal/ssim_test.py \
     --git-repo="$(git config --get remote.origin.url)" \
     --git-commit="$(git rev-parse HEAD)" \
@@ -147,19 +106,6 @@ modal run fastvideo/tests/modal/ssim_test.py \
     --no-fail-fast
 ```
 
-Env prefix rationale (parity with CI; see `.buildkite/pipeline.yml:1-3` and
-`.buildkite/scripts/pr_test.sh:62-83`):
-- `IMAGE_VERSION=py3.12-latest`: pins the Modal image tag to the same one CI
-  uses. Without this, `ssim_test.py:17` falls back to `latest`, which on
-  GHCR is built from `Dockerfile.python3.10` — different Python, torch, and
-  flash-attn wheel than CI's `py3.12-latest` (`infra-build-image.yml:51-67`,
-  `_template-build-image.yml:65-101`).
-- `BUILDKITE_REPO`/`BUILDKITE_COMMIT`/`BUILDKITE_PULL_REQUEST`: mirror what
-  Buildkite exports. `ssim_test.py:38-46` bakes these into the image's
-  `.env(...)` block; mismatched values can perturb in-container code paths
-  that branch on PR-vs-non-PR. `false` for `BUILDKITE_PULL_REQUEST` matches
-  Buildkite's "non-PR build" sentinel.
-
 Flag rationale:
 - `--skip-reference-download`: no refs exist yet, so conftest must not try to
   pull them.
@@ -197,59 +143,17 @@ get` preserves that trailing `generated_videos/` segment.
 
 ### 4. PAUSE — user reviews quality
 
-Type-aware verification.
-
-**For `ARTEFACT_TYPE = pixel`** — list the downloaded mp4s and ask the user to
-open them in a video player:
+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."
 
-**For `ARTEFACT_TYPE = latent`** — `.pt` files are not human-watchable. Print
-a numerics dump for each `.pt` so the user can sanity-check shape, distribution,
-and metadata:
-
-```python
-import torch
-from pathlib import Path
-ROOT = Path("./generated_videos_modal/default/generated_videos/L40S_reference_videos")
-for p in sorted(ROOT.rglob("*.pt")):
-    d = torch.load(p, map_location="cpu", weights_only=False)
-    s = d["expected_slice"]
-    L = d["latent"].float()
-    print(f"=== {p.relative_to(ROOT)} ===")
-    print(f"  format_version: {d['format_version']}")
-    print(f"  shape:          {d['shape']}")
-    print(f"  dtype_original: {d['dtype_original']}")
-    print(f"  slice_spec:     {d['slice_spec']}")
-    print(f"  slice  shape={tuple(s.shape)}  mean={s.mean():+.4f}  std={s.std():.4f}  min={s.min():+.4f}  max={s.max():+.4f}")
-    print(f"  latent shape={tuple(L.shape)}  mean={L.mean():+.4f}  std={L.std():.4f}  min={L.min():+.4f}  max={L.max():+.4f}")
-    print(f"  finite: latent NaN={torch.isnan(L).any().item()} Inf={torch.isinf(L).any().item()}; "
-          f"slice NaN={torch.isnan(s).any().item()} Inf={torch.isinf(s).any().item()}")
-    print(f"  metadata: {d['metadata']}\n")
-```
-
-Sanity criteria:
-- `format_version == 1` (matches `LATENT_REFERENCE_FORMAT_VERSION`).
-- `shape` matches what the model produces (e.g. LTX-2 distilled =
-  `[1, 128, T_lat, H_lat, W_lat]`; Stable Audio Open 1.0 = `[1, 64, 1024]`).
-- `slice_spec.kind` matches a registered kind (`corner_3x3_first_frame`
-  for video, `audio_first_8_timesteps` for audio).
-- No `NaN`/`Inf`. `mean ≈ 0`, `std ≈ 1` (denoised latents stay close to
-  the initial Gaussian distribution; very wide deviations suggest
-  numerical drift).
-- `metadata.prompt` matches the test's prompt.
-
-Then ask:
-
-> "Numerics look right? 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 artefacts. Single command works for both
-artefact types because `_iter_reference_files` walks `.mp4` and `.pt`:
+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 \
@@ -259,13 +163,12 @@ python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
 ```
 
 (The `--generated-dir` points at the device-folder root inside the
-downloaded tree; `copy-local` walks all `<model>/<backend>/*.{mp4,pt}`
+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 for pixel: `fastvideo/tests/ssim/reference_videos/default/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
-Result for latent: same path with `.pt` extension.
+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
 
@@ -298,54 +201,33 @@ 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. If the user
-  ran `hf auth login` instead of exporting an env var, read the cached
-  token via `huggingface_hub.get_token()` and forward it to Modal as
-  `--hf-api-key="$CACHED_TOKEN"`.
-- **Modal run fails before generation.** No artefacts on the volume — nothing
-  to download. Fix the test locally (`pytest fastvideo/tests/ssim/<test_file>`)
+  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 artefacts (most likely the
-  test crashed before writing, or `REQUIRED_GPUS` exceeded the partition
-  capacity — see Modal logs).
-- **Latent test crashed with FSDP / inference_mode error
-  (`RuntimeError: Inference tensors do not track version counter`).** The
-  test must pass `init_kwargs_override={"use_fsdp_inference": False}` when
-  `sp_size == 1` — see `test_stable_audio_similarity.py` for the pattern.
-  Fix in the test, push, retry.
+  `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 artefacts stay on disk for
+- **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.
-  - For latent: also check `slice_spec.kind` matches the latent rank
-    (`corner_3x3_first_frame` requires 5-D, `audio_first_8_timesteps`
-    requires 3-D); a rank/kind mismatch raises in `_extract_expected_slice`.
 
 ## 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 (pixel SSIM drifts across SKUs; latent cosine has
-  tighter cross-SKU bf16 drift but the configured tolerances assume
-  same-SKU seed → same-SKU verify).
+  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`.
-- Both artefact types share the same Modal flow: the orchestrator sets
-  `--skip-reference-download` + `--no-fail-fast`, runs pytest, the test's
-  helper writes the artefact (`.mp4` via `imageio` for pixel,
-  `save_latent_reference` → `torch.save` for latent) BEFORE the
-  missing-reference assertion raises. `_sync_generated_videos_to_volume` in
-  `ssim_test.py` does a `shutil.copytree` of the whole `generated_videos/`
-  tree, picking up `.mp4`, `.pt`, and the `*_ssim.json` / `*_latent.json`
-  metric files alongside.
 
 ## References
 
@@ -354,17 +236,10 @@ it will auto-download the refs they just uploaded.
   `--skip-reference-download`, `--no-fail-fast`.
 - `fastvideo/tests/ssim/reference_videos_cli.py` — `copy-local`, `upload`
   (with `--model-id`, `--force`), `download`, `ensure` subcommands.
-  Extension allowlist is `REFERENCE_EXTENSIONS = VIDEO_EXTENSIONS +
-  LATENT_EXTENSIONS` (`.pt`).
 - `fastvideo/tests/ssim/README.md` — reference layout, HF repo conventions.
-- `fastvideo/tests/ssim/inference_similarity_utils.py` — pixel helpers
-  (`run_text_to_video_similarity_test`,
-  `run_image_to_video_similarity_test`, `build_init_kwargs`).
-- `fastvideo/tests/ssim/latent_similarity_utils.py` — latent helper
-  (`run_text_to_latent_similarity_test`), slice spec dispatch
-  (`_extract_expected_slice`), reference schema
-  (`save_latent_reference` / `load_latent_reference`),
-  `LATENT_REFERENCE_FORMAT_VERSION`.
+- `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
 
@@ -373,4 +248,3 @@ it will auto-download the refs they just uploaded.
 | 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. |
-| 2026-05-01 | Latent (`*.pt`) artefact support: artefact-type detection in step 1, type-aware verification (visual eyeball for mp4, numerics dump for pt) in step 4, FSDP+inference_mode failure-mode added, design notes for the unified Modal flow. Triggered by PR #1253 (LTX-2 latent migration + Stable Audio latent test). |

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

@@ -29,24 +29,18 @@
     "Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of a bustling street food market. The scene captures a mix of humor and authenticity. Mid-shot framing, vibrant lighting."
   ],
   "run_config": {
-    "num_warmup_runs": 2,
-    "num_measurement_runs": 5,
+    "num_warmup_runs": 1,
+    "num_measurement_runs": 3,
     "required_gpus": 2
   },
   "thresholds": {
     "L40S": {
       "max_generation_time_s": 34.0,
-      "max_peak_memory_mb": 11000.0,
-      "max_text_encoder_time_s": 5.0,
-      "max_dit_time_s": 10.0,
-      "max_vae_decode_time_s": 10.0
+      "max_peak_memory_mb": 11000.0
     },
     "default": {
       "max_generation_time_s": 120.0,
-      "max_peak_memory_mb": 30000.0,
-      "max_text_encoder_time_s": 5.0,
-      "max_dit_time_s": 10.0,
-      "max_vae_decode_time_s": 10.0
+      "max_peak_memory_mb": 30000.0
     }
   }
 }

.buildkite/scripts/pr_test.sh

@@ -15,21 +15,8 @@ log "Project root: $PROJECT_ROOT"
 # Install Modal if not available
 if ! python3 -m modal --version &> /dev/null; then
     log "Modal not found, installing..."
-    if ! command -v uv &> /dev/null; then
-        log "uv not found, bootstrapping..."
-        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
-            log "Error: Failed to bootstrap uv via astral.sh installer."
-            exit 1
-        fi
-        export PATH="$HOME/.local/bin:$PATH"
-        if ! command -v uv &> /dev/null; then
-            log "Error: uv still not on PATH after bootstrap."
-            exit 1
-        fi
-    fi
-    # --break-system-packages preserves prior `pip install --user` semantics on PEP 668 agents.
-    uv pip install --system --break-system-packages modal
-
+    python3 -m pip install modal
+    
     # Verify installation
     if ! python3 -m modal --version &> /dev/null; then
         log "Error: Failed to install modal. Please install it manually."
@@ -76,86 +63,7 @@ EFFECTIVE_PR=${BUILDKITE_PULL_REQUEST:-false}
 if [ "$EFFECTIVE_PR" = "false" ] && [ -n "${PR_NUMBER:-}" ]; then
     EFFECTIVE_PR=$PR_NUMBER
 fi
-MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR BUILDKITE_BRANCH=${BUILDKITE_BRANCH:-} TEST_SCOPE=${TEST_SCOPE:-} IMAGE_VERSION=$IMAGE_VERSION"
-
-POST_RUN_HOOK=""
-
-upload_performance_artifacts() {
-    SHORT_SHA=${BUILDKITE_COMMIT:0:7}
-    LOCAL_DIR="downloaded_reports"
-
-    _download_reports() {
-        log "Downloading perf_reports/ from Modal Volume..."
-        mkdir -p "$LOCAL_DIR"
-        if ! modal volume get hf-model-weights "perf_reports/" "$LOCAL_DIR"; then
-            log "Error: Failed to download perf_reports/ from Modal Volume."
-            return 1
-        fi
-    }
-
-    _upload_dashboard() {
-        local target
-        target=$(find "$LOCAL_DIR" -name "dashboard_${SHORT_SHA}_*" | head -n 1)
-        log "TARGET dashboard: '$target'"
-
-        if [ -n "$target" ]; then
-            log "Found dashboard: $target. Uploading to Buildkite..."
-            buildkite-agent artifact upload "$target"
-            buildkite-agent annotate --style info --context "perf-dashboard" < "$target"
-        else
-            log "Warning: Could not find a dashboard file matching $SHORT_SHA"
-        fi
-    }
-
-    _upload_perf_summary() {
-        local target
-        target=$(find "$LOCAL_DIR" -name "perf_${SHORT_SHA}_*" | head -n 1)
-        log "TARGET perf summary: '$target'"
-
-        if [ -n "$target" ]; then
-            log "Found perf summary: $target. Uploading to Buildkite..."
-            buildkite-agent artifact upload "$target"
-            buildkite-agent annotate --style info --context "perf-summary" < "$target"
-        else
-            log "Warning: Could not find a perf summary file matching $SHORT_SHA"
-        fi
-    }
-
-    _upload_normalized_perf_results() {
-        local found=0
-        while IFS= read -r -d '' target; do
-            found=1
-            log "Found normalized performance result: $target. Uploading to Buildkite..."
-            buildkite-agent artifact upload "$target"
-        done < <(find "$LOCAL_DIR" -path "*/results/normalized_perf_*.json" -print0)
-
-        if [ "$found" -eq 0 ]; then
-            log "No normalized performance result artifacts found. This is expected when the rolling performance comparison did not run."
-        fi
-    }
-
-    _cleanup_modal_volume() {
-        log "Cleaning up perf_reports/ from Modal Volume..."
-        if modal volume rm hf-model-weights "perf_reports/" --recursive; then
-            log "Successfully deleted perf_reports/ from Modal Volume."
-        else
-            log "Warning: Failed to delete perf_reports/ from Modal Volume. Manual cleanup may be required."
-        fi
-    }
-
-    _cleanup_local() {
-        log "Cleaning up local download directory..."
-        rm -rf "$LOCAL_DIR"
-    }
-
-    # --- Main flow ---
-    _download_reports || { _cleanup_local; return 1; }
-    _upload_dashboard
-    _upload_perf_summary
-    _upload_normalized_perf_results
-    _cleanup_modal_volume
-    _cleanup_local
-}
+MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR IMAGE_VERSION=$IMAGE_VERSION"
 
 case "$TEST_TYPE" in
     "encoder")
@@ -216,9 +124,8 @@ case "$TEST_TYPE" in
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_lora_extraction_tests"
         ;;
     "performance")
-        log "Running performance tests on Modal..."
+        log "Running performance tests..."
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_performance_tests"
-        POST_RUN_HOOK="upload_performance_artifacts"
         ;;
     "api_server")
         log "Running API server integration tests..."
@@ -240,10 +147,5 @@ else
     log "Error: Modal test failed with exit code: $TEST_EXIT_CODE"
 fi
 
-if [ -n "$POST_RUN_HOOK" ]; then
-    log "Executing post-run hook: $POST_RUN_HOOK"
-    "$POST_RUN_HOOK"
-fi
-
 log "=== Test execution completed with exit code: $TEST_EXIT_CODE ==="
 exit $TEST_EXIT_CODE

.buildkite/scripts/pre_commit.sh

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

.github/workflows/infra-docs.yml

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

.github/workflows/publish-fastvideo.yml

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

.github/workflows/publish-kernel.yml

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

.pre-commit-config.yaml

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

AGENTS.md

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

benchmarks/fvd/feature_extractors.py

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

benchmarks/fvd/i3d_model.py

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

benchmarks/scripts/run.sh

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

benchmarks/scripts/setup_fvd.sh

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

comfyui/README.md

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

docker/Dockerfile.python3.10

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

docker/Dockerfile.python3.11

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

docker/Dockerfile.python3.12

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

docker/Dockerfile.python3.12.cuda12.9.1

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

docker/rocm7_1.Dockerfile

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

docs/README.md

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

docs/contributing/coding_agents.md

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

docs/contributing/developer_env/runpod.md

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

docs/contributing/overview.md

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

docs/design/inference_schema_parity_inventory.yaml

@@ -306,30 +306,6 @@ surfaces:
         sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
       num_frames_per_block:
         sources: [fastvideo.configs.pipelines.wan.MatrixGameI2V480PConfig]
-      audio_channels:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
-      audio_end_in_s:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
-      audio_start_in_s:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
-      max_audio_duration_s:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
-      sample_size:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
-      sampling_rate:
-        sources:
-          - fastvideo.configs.pipelines.stable_audio.StableAudioT2AConfig
-          - fastvideo.configs.pipelines.stable_audio.StableAudioOpenSmallConfig
     compatibility_only:
       batch_size: "Gen3C inference-only tuning field pending typed batching design."
       gradient_checkpointing: "Gen3C inference-only compatibility field pending typed batching design."
@@ -378,8 +354,6 @@ surfaces:
       return_frames: request.output.return_frames
       return_trajectory_latents: request.runtime.return_trajectory_latents
       return_trajectory_decoded: request.runtime.return_trajectory_decoded
-      continuation_state: request.state
-      return_continuation_state: request.output.return_state
     preset_owned:
       t_thresh: request.stage_overrides.refine.t_thresh
       spatial_refine_only: request.stage_overrides.refine.spatial_refine_only
@@ -404,13 +378,6 @@ surfaces:
       ltx2_stg_scale_audio: request.extensions.ltx2.stg_scale_audio
       ltx2_stg_blocks_video: request.extensions.ltx2.stg_blocks_video
       ltx2_stg_blocks_audio: request.extensions.ltx2.stg_blocks_audio
-      audio_start_in_s: request.extensions.stable_audio.audio_start_in_s
-      audio_end_in_s: request.extensions.stable_audio.audio_end_in_s
-      init_audio: request.extensions.stable_audio.init_audio
-      init_audio_strength: request.extensions.stable_audio.init_audio_strength
-      init_noise_level: request.extensions.stable_audio.init_noise_level
-      inpaint_audio: request.extensions.stable_audio.inpaint_audio
-      inpaint_mask: request.extensions.stable_audio.inpaint_mask
     internal_only:
       data_type: "Derived from the request shape and not a public input."
 

docs/design/server_contracts/streaming.md

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

docs/design/training_architecture.md

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

docs/distillation/dmd.md

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

docs/getting_started/installation.md

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

docs/getting_started/installation/gpu.md

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

docs/getting_started/installation/mps.md

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

docs/inference/gen3c.md

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

docs/inference/optimizations.md

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

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

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

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

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

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

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

examples/inference/basic/basic_gen3c.py

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

examples/inference/basic/basic_stable_audio.py

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

examples/inference/basic/basic_stable_audio_a2a.py

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

examples/inference/basic/basic_stable_audio_inpaint.py

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

examples/inference/basic/basic_stable_audio_small.py

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

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

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

examples/train/configs/overfit_cosmos25_t2w.yaml

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

fastvideo-kernel/CMakeLists.txt

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

fastvideo-kernel/python/fastvideo_kernel/__init__.py

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

fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn.py

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

fastvideo-kernel/python/fastvideo_kernel/ops.py

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

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

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

fastvideo-kernel/python/fastvideo_kernel/vmoba.py

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

fastvideo/AGENTS.md

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

fastvideo/api/compat.py

@@ -17,7 +17,6 @@ from fastvideo.api.request_metadata import (
 )
 from fastvideo.api.schema import (
     CompileConfig,
-    ContinuationState,
     GenerationRequest,
     GeneratorConfig,
     InputConfig,
@@ -27,10 +26,7 @@ from fastvideo.api.schema import (
 )
 from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.pipelines.basic.ltx2.stage_overrides import (
-    refine_preset_override_fields,
-    refine_stage_override_fields,
-)
+from fastvideo.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)}
@@ -44,10 +40,7 @@ _LEGACY_REQUEST_ALIASES = {
 _REQUEST_PIPELINE_OVERRIDE_FIELDS = frozenset({
     "embedded_cfg_scale",
 })
-# torch.compile kwargs that map to first-class CompileConfig fields.
 _COMPILE_TYPED_KEYS = ("backend", "fullgraph", "mode", "dynamic")
-# LTX-2 refine flat kwargs (init + per-request) known to FastVideoArgs.
-_LTX2_REFINE_FLAT_KEYS = (refine_preset_override_fields() | refine_stage_override_fields())
 
 
 def normalize_generator_config(config: GeneratorConfig | Mapping[str, Any], ) -> GeneratorConfig:
@@ -118,10 +111,8 @@ def legacy_from_pretrained_to_config(
             offload["pin_cpu_memory"] = value
         elif key == "enable_torch_compile":
             compile_config["enabled"] = value
-        elif key == "enable_torch_compile_text_encoder":
-            compile_config["text_encoder_enabled"] = value
         elif key == "torch_compile_kwargs":
-            remaining: dict[str, Any] = (dict(deepcopy(value)) if isinstance(value, Mapping) else {})
+            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)
@@ -237,12 +228,6 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         kwargs["workload_type"] = normalized.pipeline.workload_type
     if normalized.pipeline.vae_tiling is not None:
         kwargs["ltx2_vae_tiling"] = normalized.pipeline.vae_tiling
-    if engine.compile.text_encoder_enabled is not None:
-        # ``FastVideoArgs.from_kwargs`` filters to declared fields, so
-        # this is a no-op on the current legacy path. Emit anyway so the
-        # realtime runtime (PR 7.6) — which reads from the kwargs dict
-        # before FastVideoArgs filtering — can pick it up once wired.
-        kwargs["enable_torch_compile_text_encoder"] = (engine.compile.text_encoder_enabled)
 
     quantization = engine.quantization
     if quantization is not None and quantization.text_encoder_quant is not None:
@@ -273,7 +258,7 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
     preset_overrides = deepcopy(normalized.pipeline.preset_overrides)
     refine = preset_overrides.pop("refine", None)
     if isinstance(refine, Mapping):
-        for key in _LTX2_REFINE_FLAT_KEYS:
+        for key in REFINE_FLAT_KEYS:
             if key in refine:
                 kwargs[f"ltx2_refine_{key}"] = refine[key]
     kwargs.update(preset_overrides)
@@ -326,13 +311,10 @@ def request_to_sampling_param(
 ) -> SamplingParam:
     if request.plan is not None:
         raise NotImplementedError("GenerationRequest.plan is not wired into VideoGenerator yet")
+    if request.state is not None:
+        raise NotImplementedError("GenerationRequest.state is not wired into VideoGenerator yet")
 
     sampling_param = SamplingParam.from_pretrained(model_path)
-    if request.state is not None:
-        _validate_continuation_state(request.state)
-        sampling_param.continuation_state = request.state
-    if request.output.return_state:
-        sampling_param.return_continuation_state = True
     updates = explicit_request_updates(request)
 
     for key, value in updates.items():
@@ -375,14 +357,9 @@ def _looks_like_run_or_serve_config(raw: Mapping[str, Any]) -> bool:
 
 
 def _compile_config_to_torch_kwargs(compile_config: CompileConfig, ) -> dict[str, Any]:
-    """Flatten typed ``CompileConfig`` back to a ``torch_compile_kwargs``
-    dict that the legacy ``FastVideoArgs`` path still expects.
-
-    Typed first-class fields (:attr:`backend`, :attr:`fullgraph`,
-    :attr:`mode`, :attr:`dynamic`) are only emitted when the user set
-    them explicitly (non-``None``). ``extras`` is merged on top for any
-    uncommon kwargs.
-    """
+    """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)
@@ -553,37 +530,6 @@ def _serialize_generation_request(request: GenerationRequest) -> dict[str, Any]:
 
 _SCHEMA_DEFAULT_UPDATES = _extract_request_updates(config_to_dict(GenerationRequest()))
 
-_KNOWN_CONTINUATION_KINDS: set[str] = set()
-
-
-def register_continuation_kind(kind: str) -> None:
-    """Register a :class:`ContinuationState.kind` as recognized.
-
-    PR 7 wires the envelope through; per-kind payload deserializers live
-    with each model family (e.g. ``fastvideo.pipelines.basic.ltx2.
-    continuation.LTX2ContinuationState``). The registry lets the
-    public-API compat layer validate the kind early, before the state
-    reaches the pipeline.
-    """
-    if not isinstance(kind, str) or not kind:
-        raise ValueError("ContinuationState kind must be a non-empty string")
-    _KNOWN_CONTINUATION_KINDS.add(kind)
-
-
-def _validate_continuation_state(state: ContinuationState) -> None:
-    if not isinstance(state.kind, str) or not state.kind:
-        raise ValueError("GenerationRequest.state.kind must be a non-empty string; got "
-                         f"{state.kind!r}")
-    if not isinstance(state.payload, Mapping):
-        raise ValueError(f"GenerationRequest.state.payload must be a mapping; got "
-                         f"{type(state.payload).__name__}")
-    if state.kind not in _KNOWN_CONTINUATION_KINDS:
-        known = sorted(_KNOWN_CONTINUATION_KINDS)
-        raise ValueError(f"Unknown ContinuationState kind {state.kind!r}; registered "
-                         f"kinds: {known}. Import the model family that owns this kind "
-                         "(e.g. `import fastvideo.pipelines.basic.ltx2.continuation`) "
-                         "to register it, or drop the state field.")
-
 
 def _fan_out_batched_input_value(
     source_request: GenerationRequest,
@@ -617,7 +563,6 @@ __all__ = [
     "load_generator_config_from_file",
     "normalize_generation_request",
     "normalize_generator_config",
-    "register_continuation_kind",
     "request_to_pipeline_overrides",
     "request_to_sampling_param",
 ]

fastvideo/api/results.py

@@ -15,7 +15,6 @@ class GenerationResult:
     samples: Any | None = None
     frames: Any | None = None
     audio: Any | None = None
-    audio_sample_rate: int | None = None
     size: tuple[int, int, int] | None = None
     generation_time: float | None = None
     logging_info: Any | None = None
@@ -45,7 +44,6 @@ class GenerationResult:
                 "samples",
                 "frames",
                 "audio",
-                "audio_sample_rate",
                 "size",
                 "generation_time",
                 "logging_info",
@@ -64,7 +62,6 @@ class GenerationResult:
             samples=result.get("samples"),
             frames=result.get("frames"),
             audio=result.get("audio"),
-            audio_sample_rate=result.get("audio_sample_rate"),
             size=result.get("size"),
             generation_time=result.get("generation_time"),
             logging_info=result.get("logging_info"),
@@ -83,7 +80,6 @@ class GenerationResult:
             "samples": self.samples,
             "frames": self.frames,
             "audio": self.audio,
-            "audio_sample_rate": self.audio_sample_rate,
             "size": self.size,
             "generation_time": self.generation_time,
             "logging_info": self.logging_info,

fastvideo/api/sampling_param.py

@@ -1,16 +1,11 @@
 # SPDX-License-Identifier: Apache-2.0
-from __future__ import annotations
-
 import copy
 from dataclasses import dataclass, field, fields
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
 from fastvideo.logger import init_logger
 from fastvideo.utils import StoreBoolean
 
-if TYPE_CHECKING:
-    from fastvideo.api.schema import ContinuationState
-
 logger = init_logger(__name__)
 
 
@@ -97,13 +92,9 @@ class SamplingParam:
     movement_distance: float | None = None
     camera_rotation: str | None = None
 
-    # LTX-2 multi-modal CFG and STG.
-    # cfg_scale defaults are 1.0 (CFG off) so ``ForwardBatch.__post_init__``
-    # doesn't force ``do_classifier_free_guidance`` on non-LTX-2 models that
-    # never override these fields. LTX-2 presets that need text-CFG on set
-    # them in their ``defaults`` dict (e.g. ``ltx2_base``).
-    ltx2_cfg_scale_video: float = 1.0
-    ltx2_cfg_scale_audio: float = 1.0
+    # 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
@@ -112,39 +103,6 @@ class SamplingParam:
     ltx2_stg_blocks_video: list[int] = field(default_factory=lambda: [29])
     ltx2_stg_blocks_audio: list[int] = field(default_factory=lambda: [29])
 
-    # Stable Audio (T2A): clip start/end in seconds. Honored by
-    # `StableAudioConditioningStage` + `StableAudioDecodingStage`. Other
-    # families ignore them.
-    audio_start_in_s: float | None = None
-    audio_end_in_s: float | None = None
-
-    # Stable Audio audio-to-audio (variation):
-    #   `init_audio` -- a path or `[B, C, samples]` waveform at the model
-    #                   sample rate; the pipeline encodes it via the VAE
-    #                   and uses it as the starting latent.
-    #   `init_audio_strength` -- 0..1, higher = closer to the reference
-    #                            (matches the convention of Stability's
-    #                            commercial Stable Audio 2.0 UI). 1.0 ~=
-    #                            VAE round-trip, 0.0 ~= plain T2A.
-    #   `init_noise_level` -- legacy raw `sigma_max` override (0.3..500,
-    #                         higher = more freedom). Kept for callers
-    #                         that already use it; prefer `init_audio_strength`.
-    init_audio: Any = None
-    init_audio_strength: float | None = None
-    init_noise_level: float | None = None
-
-    # Stable Audio inpainting (RePaint-style): `inpaint_audio` is the
-    # reference clip, `inpaint_mask` is a [samples] tensor in {0, 1} where
-    # 1 means *keep the reference* and 0 means *regenerate*.
-    inpaint_audio: Any = None
-    inpaint_mask: Any = None
-
-    # Continuation state carried across streaming/multi-segment calls.
-    continuation_state: ContinuationState | None = None
-    # When True, the pipeline returns a ContinuationState on the result so
-    # the caller can resume from the generated segment.
-    return_continuation_state: bool = False
-
     # Misc
     save_video: bool = True
     return_frames: bool = True
@@ -169,7 +127,7 @@ class SamplingParam:
         self.__post_init__()
 
     @classmethod
-    def from_pretrained(cls, model_path: str) -> SamplingParam:
+    def from_pretrained(cls, model_path: str) -> "SamplingParam":
         sampling_param = cls._from_preset(model_path)
         if sampling_param is not None:
             return sampling_param
@@ -185,7 +143,7 @@ class SamplingParam:
     def _from_preset(
         cls,
         model_path: str,
-    ) -> SamplingParam | None:
+    ) -> "SamplingParam | None":
         """Build a SamplingParam from preset defaults.
 
         Returns ``None`` when no preset is configured for

fastvideo/api/schema.py

@@ -41,12 +41,6 @@ class CompileConfig:
     """
 
     enabled: bool = False
-    text_encoder_enabled: bool | None = None
-    """Whether ``torch.compile`` is applied to the text encoder. ``None``
-    keeps the runtime default. The public ``FastVideoArgs`` adapter does
-    not yet consume this flag; reserved so the realtime runtime upstream
-    (PR 7.6) has a typed home for its ``enable_torch_compile_text_encoder``
-    kwarg without routing through ``pipeline.experimental``."""
     backend: str | None = None
     fullgraph: bool | None = None
     mode: str | None = None

fastvideo/attention/AGENTS.md

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

fastvideo/attention/backends/flash_attn.py

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

fastvideo/attention/backends/sla.py

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

fastvideo/configs/AGENTS.md

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

fastvideo/configs/models/base.py

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

fastvideo/configs/models/dits/__init__.py

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

fastvideo/configs/models/dits/cosmos.py

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

fastvideo/configs/models/dits/stable_audio.py

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

fastvideo/configs/models/encoders/__init__.py

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

fastvideo/configs/models/encoders/stable_audio_conditioner.py

@@ -1,80 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Config for the Stable Audio Open 1.0 multi-conditioner.
-
-The conditioner bundles three sub-conditioners — a T5 text encoder
-(prompt) and two NumberConditioners (`seconds_start` / `seconds_total`)
-— into the (cross_attn_cond, cross_attn_mask, global_embed) triple the
-DiT consumes. The architecture is fully specified by the official
-`stable_audio_tools` `MultiConditioner` config; the constants here
-mirror that.
-"""
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-
-from fastvideo.configs.models.base import ArchConfig
-from fastvideo.configs.models.encoders.base import (EncoderArchConfig, EncoderConfig)
-
-
-def _default_configs() -> list[dict]:
-    """Default = `stable-audio-open-1.0`'s three sub-conditioners."""
-    return [
-        {
-            "id": "prompt",
-            "type": "t5",
-            "config": {
-                "t5_model_name": "t5-base",
-                "max_length": 128
-            }
-        },
-        {
-            "id": "seconds_start",
-            "type": "number",
-            "config": {
-                "min_val": 0,
-                "max_val": 512
-            }
-        },
-        {
-            "id": "seconds_total",
-            "type": "number",
-            "config": {
-                "min_val": 0,
-                "max_val": 512
-            }
-        },
-    ]
-
-
-@dataclass
-class StableAudioConditionerArchConfig(EncoderArchConfig):
-    architectures: list[str] = field(default_factory=lambda: ["StableAudioMultiConditioner"])
-
-    # Shared embedding width across all sub-conditioners (T5 last-hidden
-    # dim and NumberEmbedder feature dim both = `cond_dim`).
-    cond_dim: int = 768
-
-    # Sub-conditioner identifiers. Order in `cross_attention_cond_ids`
-    # is the concat order for the cross-attn token sequence; order in
-    # `global_cond_ids` is the concat order for the global FiLM-style
-    # embedding.
-    cross_attention_cond_ids: tuple[str, ...] = ("prompt", "seconds_start", "seconds_total")
-    global_cond_ids: tuple[str, ...] = ("seconds_start", "seconds_total")
-
-    # Per-sub-conditioner spec list (mirrors upstream
-    # `model_config.json.model.conditioning.configs`). Each entry is
-    # `{"id": ..., "type": "t5"|"number", "config": {...}}`. The default
-    # matches `stable-audio-open-1.0`; SA-small overrides via the
-    # `conditioner/config.json` shipped in the converted repo.
-    configs: list = field(default_factory=_default_configs)
-
-    # Match official `stable_audio_tools/models/conditioners.py:334`:
-    # T5 is loaded directly in fp16.
-    t5_dtype: str = "float16"
-
-
-@dataclass
-class StableAudioConditionerConfig(EncoderConfig):
-    arch_config: ArchConfig = field(default_factory=StableAudioConditionerArchConfig)
-
-    prefix: str = "stable_audio_conditioner"

fastvideo/configs/models/encoders/t5.py

@@ -41,14 +41,6 @@ class T5ArchConfig(TextEncoderArchConfig):
     text_len: int = 512
     dtype: str | None = None
     gradient_checkpointing: bool = False
-    # Extra fields present in upstream HF T5Config but unused by FastVideo's
-    # encoder. Declared here so `update_model_arch` doesn't reject them when
-    # loading repos like `stabilityai/stable-audio-open-1.0` that ship the
-    # full HF config.
-    n_positions: int = 512
-    decoder_start_token_id: int = 0
-    output_past: bool = True
-    task_specific_params: dict | None = None
     stacked_params_mapping: list[tuple[str, str, str]] = field(default_factory=lambda: [
         # (param_name, shard_name, shard_id)
         (".qkv_proj", ".q", "q"),

fastvideo/configs/models/vaes/__init__.py

@@ -5,7 +5,6 @@ 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
-from fastvideo.configs.models.vaes.oobleck import OobleckVAEArchConfig, OobleckVAEConfig
 from fastvideo.configs.models.vaes.wanvae import WanVAEConfig
 
 __all__ = [
@@ -17,6 +16,4 @@ __all__ = [
     "Gen3CVAEConfig",
     "Hunyuan15VAEConfig",
     "LTX2VAEConfig",
-    "OobleckVAEArchConfig",
-    "OobleckVAEConfig",
 ]

fastvideo/configs/models/vaes/oobleck.py

@@ -1,68 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Config for the Stable Audio Open 1.0 "Oobleck" VAE.
-
-Mirrors the per-channel `vae/config.json` shipped in
-`stabilityai/stable-audio-open-1.0` 1:1 (see
-`fastvideo/models/vaes/oobleck.py::OobleckVAE.from_pretrained`, which
-constructs the VAE from these fields). Inherits the FastVideo VAEConfig
-base so the standard `load_encoder` / `load_decoder` flags + tiling
-knobs apply.
-
-Naming: the VAE architecture is officially "Oobleck" (per Stability
-AI's stable-audio-tools) — the surrounding model family is "Stable
-Audio Open 1.0". This config is named after the architecture
-(`OobleckVAEConfig`) since the same VAE is shared across Stable Audio
-checkpoints; downstream pipelines reference it by its arch name, not
-by a host-pipeline name.
-"""
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-
-from fastvideo.configs.models.vaes.base import VAEArchConfig, VAEConfig
-
-
-@dataclass
-class OobleckVAEArchConfig(VAEArchConfig):
-    """Stable Audio Open 1.0 VAE architecture constants."""
-
-    architectures: list[str] = field(default_factory=lambda: ["AutoencoderOobleck"])
-
-    # From stabilityai/stable-audio-open-1.0/vae/config.json.
-    encoder_hidden_size: int = 128
-    downsampling_ratios: list[int] = field(default_factory=lambda: [2, 4, 4, 8, 8])
-    channel_multiples: list[int] = field(default_factory=lambda: [1, 2, 4, 8, 16])
-    decoder_channels: int = 128
-    decoder_input_channels: int = 64
-    audio_channels: int = 2  # stereo
-    sampling_rate: int = 44100
-
-
-@dataclass
-class OobleckVAEConfig(VAEConfig):
-    """FastVideo VAE config wrapping the Oobleck arch.
-
-    Audio VAEs don't use the temporal/spatial tiling defaults that the
-    base VAEConfig is shaped for (those exist for video VAEs); they are
-    retained but irrelevant for audio.
-    """
-
-    arch_config: VAEArchConfig = field(default_factory=OobleckVAEArchConfig)
-
-    # Audio is 1-D, so the video-VAE tiling defaults are inert. Disable
-    # them so callers don't accidentally trip on tile-stride math built
-    # for spatial tensors.
-    use_tiling: bool = False
-    use_temporal_tiling: bool = False
-    use_parallel_tiling: bool = False
-
-    # Where the FastVideo loader / pipeline-glue wrapper should fetch
-    # weights from when no local path is supplied. Gated repo — caller's
-    # HF token must have accepted terms on
-    # https://huggingface.co/stabilityai/stable-audio-open-1.0.
-    pretrained_path: str = "stabilityai/stable-audio-open-1.0"
-    pretrained_subfolder: str = "vae"
-    # Match official `stable_audio_tools`: VAE runs in fp16 (the
-    # `pretransform.model_half` path in
-    # `stable_audio_tools/models/pretransforms.py`).
-    pretrained_dtype: str = "float16"

fastvideo/configs/pipelines/longcat.py

@@ -11,34 +11,10 @@ import torch
 from fastvideo.configs.models import DiTConfig, VAEConfig
 from fastvideo.configs.models.dits.base import DiTArchConfig
 from fastvideo.configs.models.encoders import BaseEncoderOutput, T5Config
-from fastvideo.configs.models.encoders.base import TextEncoderArchConfig
-from fastvideo.configs.models.encoders.t5 import T5ArchConfig
 from fastvideo.configs.models.vaes import WanVAEConfig
 from fastvideo.configs.pipelines.base import PipelineConfig
 
 
-@dataclass
-class LongCatT5ArchConfig(T5ArchConfig):
-    """T5 arch that pads tokenizer output to ``max_length``.
-
-    LongCat's denoising stage concatenates positive and negative
-    attention masks along the batch dimension for CFG, which requires
-    uniform seq length. The shared :class:`T5ArchConfig` dropped the
-    ``"padding": "max_length"`` tokenizer kwarg so other DiTs could run
-    with variable-length masks; LongCat still needs the uniform
-    contract.
-    """
-
-    def __post_init__(self) -> None:
-        super().__post_init__()
-        self.tokenizer_kwargs["padding"] = "max_length"
-
-
-@dataclass
-class LongCatT5Config(T5Config):
-    arch_config: TextEncoderArchConfig = field(default_factory=LongCatT5ArchConfig)
-
-
 @dataclass
 class LongCatDiTArchConfig(DiTArchConfig):
     """Extended DiTArchConfig with LongCat-specific fields."""
@@ -127,9 +103,8 @@ class LongCatT2V480PConfig(PipelineConfig):
     vae_precision: str = "bf16"
     text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", ))
 
-    # UMT5 uses T5-like config; postprocess pads to 512. LongCatT5Config
-    # restores ``padding="max_length"`` for the CFG concat contract.
-    text_encoder_configs: tuple[T5Config, ...] = field(default_factory=lambda: (LongCatT5Config(), ))
+    # Text encoding (UMT5 uses T5-like config; postprocess to fixed 512)
+    text_encoder_configs: tuple[T5Config, ...] = field(default_factory=lambda: (T5Config(), ))
     preprocess_text_funcs: tuple[Callable[[str], str], ...] = field(default_factory=lambda: (longcat_preprocess_text, ))
     postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
                                   ...] = field(default_factory=lambda: (umt5_postprocess_text, ))

fastvideo/configs/pipelines/stable_audio.py

@@ -1,67 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""`PipelineConfig` for Stable Audio Open 1.0."""
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-
-from fastvideo.configs.models import DiTConfig, VAEConfig
-from fastvideo.configs.models.dits import StableAudioConfig
-from fastvideo.configs.models.vaes import OobleckVAEConfig
-from fastvideo.configs.pipelines.base import PipelineConfig
-
-
-@dataclass
-class StableAudioT2AConfig(PipelineConfig):
-    """Stable Audio Open 1.0 pipeline config."""
-
-    dit_config: DiTConfig = field(default_factory=StableAudioConfig)
-    # Standard `TransformerLoader` reads `dit_precision`; default in
-    # `PipelineConfig` is bf16, but we want fp16 to match official.
-    dit_precision: str = "fp16"
-
-    vae_config: VAEConfig = field(default_factory=OobleckVAEConfig)
-    vae_tiling: bool = False
-    vae_sp: bool = False
-
-    # `StableAudioMultiConditioner` owns its own T5; zero out the
-    # parent's text-encoder slots so the length-equality validator passes.
-    text_encoder_configs: tuple = field(default_factory=tuple)
-    preprocess_text_funcs: tuple = field(default_factory=tuple)
-    postprocess_text_funcs: tuple = field(default_factory=tuple)
-
-    num_inference_steps: int = 100
-    guidance_scale: float = 7.0
-    audio_end_in_s: float = 10.0  # short-clip default
-    audio_start_in_s: float = 0.0
-    sampling_rate: int = 44100
-    audio_channels: int = 2
-    # Stable Audio Open 1.0 was trained at a fixed 2,097,152-sample
-    # window (= 2097152 / 44100 ≈ 47.55s). Anything past this is
-    # silently truncated by the post-decode slice — validate up-front.
-    sample_size: int = 2097152
-    max_audio_duration_s: float = 2097152 / 44100
-
-    # Match the official `stable_audio_tools` defaults (`model_half=True`
-    # in `run_gradio.py`), which loads the DiT, VAE, and T5 in fp16 and
-    # wraps T5 forward in `autocast(fp16)`. fp16 is also a hard
-    # requirement for FlashAttention-2 / FA-3.
-    precision: str = "fp16"
-    vae_precision: str = "fp16"
-    text_encoder_precisions: tuple[str, ...] = field(default_factory=tuple)
-
-    def __post_init__(self) -> None:
-        # A2A needs encode; load both halves for either path.
-        self.vae_config.load_encoder = True
-        self.vae_config.load_decoder = True
-
-
-@dataclass
-class StableAudioOpenSmallConfig(StableAudioT2AConfig):
-    """`stable-audio-open-small` overrides: shorter training window
-    (524288 samples ≈ 11.89s @ 44.1 kHz) and a faster default sampler
-    config carried by the small preset.
-    """
-
-    sample_size: int = 524288
-    max_audio_duration_s: float = 524288 / 44100
-    audio_end_in_s: float = 6.0  # short-clip default suitable for the small window

fastvideo/entrypoints/cli/main.py

@@ -3,8 +3,6 @@
 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
-from fastvideo.entrypoints.cli.router_serve import (
-    cmd_init as router_serve_cmd_init, )
 from fastvideo.entrypoints.cli.serve import cmd_init as serve_cmd_init
 from fastvideo.entrypoints.cli.bench import cmd_init as bench_cmd_init
 
@@ -14,7 +12,6 @@ def cmd_init() -> list[CLISubcommand]:
     commands = []
     commands.extend(generate_cmd_init())
     commands.extend(serve_cmd_init())
-    commands.extend(router_serve_cmd_init())
     commands.extend(bench_cmd_init())
     return commands
 

fastvideo/entrypoints/cli/router_serve.py

@@ -1,115 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""``fastvideo router-serve`` CLI subcommand.
-
-Launches the streaming router from a YAML config. Separate from
-``fastvideo serve`` because the router is an orthogonal process: it
-fronts one or more running servers rather than hosting a generator
-itself.
-"""
-from __future__ import annotations
-
-import argparse
-import os
-from typing import cast
-
-from fastvideo.api.parser import load_raw_config
-from fastvideo.entrypoints.cli.cli_types import CLISubcommand
-from fastvideo.entrypoints.streaming.router.config import (
-    ReplicaEndpoint,
-    RouterConfig,
-)
-from fastvideo.logger import init_logger
-from fastvideo.utils import FlexibleArgumentParser
-
-logger = init_logger(__name__)
-
-
-class RouterServeSubcommand(CLISubcommand):
-    """Start the multi-replica WebSocket router."""
-
-    def __init__(self) -> None:
-        self.name = "router-serve"
-        super().__init__()
-
-    def cmd(self, args: argparse.Namespace) -> None:
-        config = _load_router_config(args.config)
-        logger.info(
-            "router listening on %s:%d (%d replicas, %d primary)",
-            config.host,
-            config.port,
-            len(config.replicas),
-            sum(1 for r in config.replicas if r.primary),
-        )
-        from fastvideo.entrypoints.streaming.router.main import run_router
-
-        run_router(config)
-
-    def validate(self, args: argparse.Namespace) -> None:
-        if not args.config:
-            raise ValueError("fastvideo router-serve requires --config PATH")
-        if not os.path.exists(args.config):
-            raise ValueError(f"Router config file not found: {args.config}")
-
-    def subparser_init(
-        self,
-        subparsers: argparse._SubParsersAction,
-    ) -> FlexibleArgumentParser:
-        parser = subparsers.add_parser(
-            "router-serve",
-            help="Start the streaming router (multi-replica load balancer)",
-            usage="fastvideo router-serve --config ROUTER_CONFIG",
-        )
-        parser.add_argument(
-            "--config",
-            type=str,
-            default="",
-            required=False,
-            help="Path to a YAML/JSON router config. Required.",
-        )
-        return cast(FlexibleArgumentParser, parser)
-
-
-def _load_router_config(path: str) -> RouterConfig:
-    raw = load_raw_config(path)
-    router_raw = raw.get("router") if isinstance(raw, dict) else None
-    if not isinstance(router_raw, dict):
-        raise ValueError(f"Router config {path!r} must have a top-level `router:` block")
-
-    replicas_raw = router_raw.get("replicas", [])
-    if not isinstance(replicas_raw, list):
-        raise ValueError(f"router.replicas must be a list, got {type(replicas_raw).__name__}")
-    replicas = []
-    for i, r in enumerate(replicas_raw):
-        if not isinstance(r, dict):
-            raise ValueError(f"router.replicas[{i}] must be a mapping, got {type(r).__name__}")
-        url = r.get("url")
-        if not url:
-            raise ValueError(f"router.replicas[{i}] is missing required key 'url'")
-        replicas.append(
-            ReplicaEndpoint(
-                url=url,
-                name=r.get("name"),
-                primary=bool(r.get("primary", False)),
-                weight=float(r.get("weight", 1.0)),
-            ))
-    if not replicas:
-        raise ValueError("Router config must list at least one replica under `router.replicas`")
-
-    health_check = router_raw.get("health_check") or {}
-    return RouterConfig(
-        host=str(router_raw.get("host", "0.0.0.0")),
-        port=int(router_raw.get("port", 9000)),
-        replicas=replicas,
-        health_check_path=str(health_check.get("path", "/health")),
-        health_check_interval_seconds=float(health_check.get("interval_seconds", 5.0)),
-        health_check_timeout_seconds=float(health_check.get("timeout_seconds", 2.0)),
-        failure_threshold=int(health_check.get("failure_threshold", 3)),
-        recovery_threshold=int(health_check.get("recovery_threshold", 2)),
-    )
-
-
-def cmd_init() -> list[CLISubcommand]:
-    return [RouterServeSubcommand()]
-
-
-__all__ = ["RouterServeSubcommand", "cmd_init"]

fastvideo/entrypoints/streaming/__init__.py

@@ -1,65 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-from fastvideo.entrypoints.streaming.server import build_app, run_server
-from fastvideo.entrypoints.streaming.session import (
-    Session,
-    SessionManager,
-    SessionState,
-)
-from fastvideo.entrypoints.streaming.session_store import (
-    BlobStore,
-    InMemoryBlobStore,
-    InMemorySessionStore,
-    SessionStore,
-)
-from fastvideo.entrypoints.streaming.gpu_pool import (
-    GpuPool,
-    InProcessGpuPool,
-    PoolAcquireTimeout,
-    SubprocessGpuPool,
-)
-from fastvideo.entrypoints.streaming.mock_server import (
-    MockGenerator,
-    build_mock_app,
-)
-from fastvideo.entrypoints.streaming.prompt import (
-    LLMProvider,
-    PromptEnhancer,
-)
-from fastvideo.entrypoints.streaming.prompt.safety import (
-    PromptSafetyFilter,
-    SafetyDecision,
-)
-from fastvideo.entrypoints.streaming.session_logger import (
-    SessionLogEvent,
-    SessionLogger,
-)
-from fastvideo.entrypoints.streaming.stream import (
-    FragmentedMP4Chunk,
-    FragmentedMP4Encoder,
-)
+from fastvideo.entrypoints.streaming.server import run_server
 
-__all__ = [
-    "BlobStore",
-    "FragmentedMP4Chunk",
-    "FragmentedMP4Encoder",
-    "GpuPool",
-    "InMemoryBlobStore",
-    "InMemorySessionStore",
-    "InProcessGpuPool",
-    "LLMProvider",
-    "MockGenerator",
-    "PoolAcquireTimeout",
-    "PromptEnhancer",
-    "PromptSafetyFilter",
-    "SafetyDecision",
-    "SessionLogEvent",
-    "SessionLogger",
-    "build_mock_app",
-    "Session",
-    "SessionManager",
-    "SessionState",
-    "SessionStore",
-    "SubprocessGpuPool",
-    "build_app",
-    "run_server",
-]
+__all__ = ["run_server"]

fastvideo/entrypoints/streaming/gpu_pool.py

@@ -1,542 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""GPU pool manager for the streaming server.
-
-Replaces the single-generator path in PR 7.5 with a typed pool
-abstraction. Three implementations ship here:
-
-* :class:`InProcessGpuPool` — one in-process ``VideoGenerator``; used
-  by tests and single-GPU dev deployments.
-* :class:`SubprocessGpuPool` — one ``multiprocessing.Process`` per
-  GPU, each running :func:`worker_main` against a ``GeneratorConfig``.
-  Jobs are dispatched via ``multiprocessing.Queue``.
-* :class:`GpuPool` (abstract) — the interface both use.
-
-Session-to-GPU binding lives in the pool so continuation state stays
-on the GPU that generated the previous segment (matching the internal
-``gpu_pool.py``'s per-GPU cache behavior). Cross-GPU handoff is
-supported via :class:`SessionStore` snapshot + hydrate, which
-serializes the state before the migration and rehydrates it on the
-new worker.
-
-Typed config: workers start from a :class:`GeneratorConfig` (no flat
-LTX-2 kwargs), satisfying the PR 6 + PR 7 contracts that the public
-surface doesn't reintroduce the legacy kwarg bag.
-"""
-from __future__ import annotations
-
-import asyncio
-import multiprocessing as mp
-import queue
-import threading
-import time
-import uuid
-from abc import ABC, abstractmethod
-from concurrent.futures import Future
-from dataclasses import dataclass, field
-from typing import Any, Protocol
-
-from fastvideo.api.schema import (
-    GeneratorConfig,
-    GenerationRequest,
-    GpuPoolConfig,
-    WarmupConfig,
-)
-from fastvideo.entrypoints.streaming.session_store import (
-    InMemorySessionStore,
-    SessionStore,
-)
-from fastvideo.entrypoints.streaming.worker import worker_main
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-# ---------------------------------------------------------------------------
-# Public interface
-# ---------------------------------------------------------------------------
-
-
-class _GeneratorLike(Protocol):
-    """Subset the pool calls on a worker-side generator."""
-
-    def generate(self, request: GenerationRequest) -> Any:
-        ...
-
-
-@dataclass
-class PoolAssignment:
-    """The worker a session is currently bound to."""
-
-    gpu_id: int
-    worker_id: str
-    pinned_at: float = field(default_factory=time.monotonic)
-
-
-class GpuPool(ABC):
-    """Abstract GPU pool.
-
-    ``acquire`` binds a session to a worker and holds that binding
-    across segments so continuation state can stay hot. ``run`` submits
-    a single ``GenerationRequest`` for a bound session.
-
-    Acquire / release are independent of run — a session can run many
-    segments on one acquired worker, and must release on disconnect.
-    """
-
-    @abstractmethod
-    async def acquire(
-        self,
-        session_id: str,
-        *,
-        timeout: float | None = None,
-    ) -> PoolAssignment:
-        ...
-
-    @abstractmethod
-    async def run(
-        self,
-        session_id: str,
-        request: GenerationRequest,
-    ) -> Any:
-        ...
-
-    @abstractmethod
-    async def release(self, session_id: str) -> None:
-        ...
-
-    @abstractmethod
-    async def shutdown(self) -> None:
-        ...
-
-    @abstractmethod
-    def health(self) -> PoolHealth:
-        ...
-
-
-@dataclass
-class PoolHealth:
-    total_workers: int
-    available_workers: int
-    active_sessions: int
-    queued_sessions: int = 0
-
-
-class PoolAcquireTimeout(RuntimeError):
-    """Raised when ``acquire`` times out waiting for a free worker."""
-
-
-# ---------------------------------------------------------------------------
-# In-process implementation (single-worker, test / dev)
-# ---------------------------------------------------------------------------
-
-
-class InProcessGpuPool(GpuPool):
-    """Single-process pool backed by one :class:`_GeneratorLike`.
-
-    This is what PR 7.5's server uses by default; PR 7.6 adds the real
-    ``SubprocessGpuPool`` alternative but keeps this one for tests and
-    small deployments.
-    """
-
-    def __init__(
-        self,
-        generator: _GeneratorLike,
-        *,
-        gpu_id: int = 0,
-        session_store: SessionStore | None = None,
-    ) -> None:
-        self._generator = generator
-        self._gpu_id = gpu_id
-        self._worker_id = f"inproc-{uuid.uuid4().hex[:6]}"
-        self._session_store = session_store or InMemorySessionStore()
-        self._active: dict[str, PoolAssignment] = {}
-        self._lock = asyncio.Lock()
-        self._gen_lock = asyncio.Lock()
-
-    async def acquire(
-        self,
-        session_id: str,
-        *,
-        timeout: float | None = None,
-    ) -> PoolAssignment:
-        async with self._lock:
-            existing = self._active.get(session_id)
-            if existing is not None:
-                return existing
-            assignment = PoolAssignment(gpu_id=self._gpu_id, worker_id=self._worker_id)
-            self._active[session_id] = assignment
-            return assignment
-
-    async def run(
-        self,
-        session_id: str,
-        request: GenerationRequest,
-    ) -> Any:
-        if session_id not in self._active:
-            raise RuntimeError(f"session {session_id!r} is not acquired on this pool")
-        # Serialize generator access so one GPU runs one request at a
-        # time, matching the internal gpu_pool's per-GPU lock.
-        async with self._gen_lock:
-            loop = asyncio.get_running_loop()
-            return await loop.run_in_executor(None, self._generator.generate, request)
-
-    async def release(self, session_id: str) -> None:
-        async with self._lock:
-            self._active.pop(session_id, None)
-
-    async def shutdown(self) -> None:
-        self._active.clear()
-
-    def health(self) -> PoolHealth:
-        return PoolHealth(
-            total_workers=1,
-            available_workers=1 if not self._active else 0,
-            active_sessions=len(self._active),
-        )
-
-
-# ---------------------------------------------------------------------------
-# Subprocess implementation (multi-worker, real deployment)
-# ---------------------------------------------------------------------------
-
-
-@dataclass
-class _WorkerHandle:
-    process: Any  # mp.Process or compatible handle with is_alive / join / kill
-    job_queue: mp.Queue
-    result_queue: mp.Queue
-    gpu_id: int
-    worker_id: str
-    ready: threading.Event
-    # ``ready`` flips on either successful boot or boot failure so the
-    # parent stops waiting; ``boot_ok`` is set only on a real ready
-    # acknowledgement and is what gates pool admission.
-    boot_ok: threading.Event
-    shutdown_event: Any  # mp.Event is a factory, not a type — Any keeps mypy sane
-
-
-@dataclass
-class _PendingJob:
-    job_id: str
-    future: Future
-    session_id: str
-    worker_id: str
-
-
-class SubprocessGpuPool(GpuPool):
-    """One ``multiprocessing.Process`` per GPU.
-
-    Each worker boots :class:`fastvideo.VideoGenerator` from a typed
-    :class:`GeneratorConfig` inside the child process (post-
-    ``CUDA_VISIBLE_DEVICES`` setup) and consumes jobs from an mp Queue.
-
-    This is the production shape: the parent process stays CPU-only, and
-    GPU state never crosses process boundaries. Continuation state is
-    serialized through :class:`SessionStore` for cross-GPU handoff.
-
-    PR 7.6 ships this as an opt-in; PR 7.5's in-process pool remains the
-    default until nightly runs validate the subprocess path.
-    """
-
-    def __init__(
-        self,
-        generator_config: GeneratorConfig,
-        *,
-        pool_config: GpuPoolConfig,
-        warmup_config: WarmupConfig | None = None,
-        session_store: SessionStore | None = None,
-        worker_factory: WorkerFactory | None = None,
-    ) -> None:
-        self._generator_config = generator_config
-        self._pool_config = pool_config
-        self._warmup_config = warmup_config or WarmupConfig()
-        self._session_store = session_store or InMemorySessionStore()
-        self._worker_factory = worker_factory or _default_worker_factory
-        self._workers: list[_WorkerHandle] = []
-        self._available: asyncio.Queue[int] = asyncio.Queue()
-        self._assignments: dict[str, PoolAssignment] = {}
-        self._worker_by_id: dict[str, _WorkerHandle] = {}
-        self._pending: dict[str, _PendingJob] = {}
-        self._lock = asyncio.Lock()
-        self._result_reader_tasks: list[asyncio.Task] = []
-
-    async def start(self) -> None:
-        """Spawn worker processes and wait for each to report ready."""
-        num_workers = self._pool_config.num_workers or 1
-        for gpu_id in range(num_workers):
-            handle = self._worker_factory(
-                gpu_id=gpu_id,
-                generator_config=self._generator_config,
-                warmup_config=self._warmup_config,
-            )
-            self._workers.append(handle)
-            self._worker_by_id[handle.worker_id] = handle
-
-        # Wait for each worker's ready event in a thread to avoid
-        # blocking the event loop.
-        loop = asyncio.get_running_loop()
-        await asyncio.gather(*[
-            loop.run_in_executor(None, handle.ready.wait, self._warmup_config.timeout_seconds)
-            for handle in self._workers
-        ])
-
-        # Start background result readers — one task per worker
-        # drains its result queue and resolves futures in _pending.
-        for handle in self._workers:
-            task = asyncio.create_task(self._drain_results(handle))
-            self._result_reader_tasks.append(task)
-
-        # Only admit workers that successfully booted. Anything that
-        # failed boot (timeout, crash, error sentinel) stays out of the
-        # available queue so we never assign a session to it.
-        for idx, handle in enumerate(self._workers):
-            if handle.boot_ok.is_set():
-                await self._available.put(idx)
-            else:
-                logger.error(
-                    "pool: worker %s failed to boot; skipping",
-                    handle.worker_id,
-                )
-
-    async def acquire(
-        self,
-        session_id: str,
-        *,
-        timeout: float | None = None,
-    ) -> PoolAssignment:
-        async with self._lock:
-            existing = self._assignments.get(session_id)
-            if existing is not None:
-                return existing
-        try:
-            idx = await asyncio.wait_for(self._available.get(), timeout=timeout)
-        except asyncio.TimeoutError as exc:
-            raise PoolAcquireTimeout(f"no worker available after {timeout}s") from exc
-        handle = self._workers[idx]
-        assignment = PoolAssignment(gpu_id=handle.gpu_id, worker_id=handle.worker_id)
-        async with self._lock:
-            self._assignments[session_id] = assignment
-        return assignment
-
-    async def run(
-        self,
-        session_id: str,
-        request: GenerationRequest,
-    ) -> Any:
-        assignment = self._assignments.get(session_id)
-        if assignment is None:
-            raise RuntimeError(f"session {session_id!r} not acquired on this pool")
-        handle = self._worker_by_id[assignment.worker_id]
-        job_id = uuid.uuid4().hex
-        future: Future = Future()
-        self._pending[job_id] = _PendingJob(
-            job_id=job_id,
-            future=future,
-            session_id=session_id,
-            worker_id=handle.worker_id,
-        )
-        # mp.Queue.put can block if the underlying pipe buffer is full;
-        # offload to a thread so the event loop keeps serving other
-        # sessions. If the put itself fails, drop the pending entry so
-        # _drain_results doesn't dangle a future forever.
-        loop = asyncio.get_running_loop()
-        try:
-            await loop.run_in_executor(
-                None,
-                handle.job_queue.put,
-                {
-                    "job_id": job_id,
-                    "request": request
-                },
-            )
-        except Exception:
-            self._pending.pop(job_id, None)
-            raise
-        return await asyncio.wrap_future(future)
-
-    async def release(self, session_id: str) -> None:
-        async with self._lock:
-            assignment = self._assignments.pop(session_id, None)
-        if assignment is None:
-            return
-        idx = next((i for i, h in enumerate(self._workers) if h.worker_id == assignment.worker_id), None)
-        if idx is None:
-            return
-        # Don't return a dead worker to the pool; otherwise the next
-        # acquire will hand a session to a process that can't run jobs.
-        if not self._workers[idx].process.is_alive():
-            logger.warning(
-                "pool: worker %s died; not returning to available queue",
-                self._workers[idx].worker_id,
-            )
-            return
-        await self._available.put(idx)
-
-    async def shutdown(self) -> None:
-        loop = asyncio.get_running_loop()
-
-        # Signal all workers in parallel; .put may block on a full pipe,
-        # so off-load it the same way run() does.
-        async def _signal(handle: _WorkerHandle) -> None:
-            try:
-                handle.shutdown_event.set()
-                await loop.run_in_executor(None, handle.job_queue.put, None)
-            except Exception:  # pragma: no cover - best-effort cleanup
-                pass
-
-        await asyncio.gather(*(_signal(h) for h in self._workers))
-        # Join in parallel so total shutdown is bounded by the slowest
-        # worker, not the sum of all timeouts.
-        await asyncio.gather(*(loop.run_in_executor(None, handle.process.join, 5.0) for handle in self._workers))
-        for handle in self._workers:
-            if handle.process.is_alive():
-                handle.process.kill()
-        for task in self._result_reader_tasks:
-            task.cancel()
-        self._result_reader_tasks.clear()
-        self._workers.clear()
-        self._worker_by_id.clear()
-
-    def health(self) -> PoolHealth:
-        return PoolHealth(
-            total_workers=len(self._workers),
-            available_workers=self._available.qsize(),
-            active_sessions=len(self._assignments),
-        )
-
-    async def _drain_results(self, handle: _WorkerHandle) -> None:
-        loop = asyncio.get_running_loop()
-        try:
-            while not handle.shutdown_event.is_set():
-                try:
-                    msg = await loop.run_in_executor(None, _safe_queue_get, handle.result_queue, 0.5)
-                except Exception:
-                    logger.exception("pool: worker %s result reader failed", handle.worker_id)
-                    return
-                if msg is None:
-                    continue
-                job_id = msg.get("job_id")
-                if job_id is None:
-                    continue
-                pending = self._pending.pop(job_id, None)
-                if pending is None:
-                    continue
-                if msg.get("kind") == "error":
-                    pending.future.set_exception(RuntimeError(msg["error"]))
-                else:
-                    pending.future.set_result(msg.get("result"))
-        finally:
-            # If we exit for any reason — shutdown, exception, cancel —
-            # surface that to any in-flight jobs on this worker so their
-            # await never hangs on a future no one will resolve.
-            for jid in [jid for jid, job in self._pending.items() if job.worker_id == handle.worker_id]:
-                pending = self._pending.pop(jid, None)
-                if pending is not None and not pending.future.done():
-                    pending.future.set_exception(
-                        RuntimeError(f"worker {handle.worker_id} result reader exited "
-                                     "with pending jobs"))
-
-
-def _safe_queue_get(q: mp.Queue, timeout: float) -> Any | None:
-    try:
-        return q.get(timeout=timeout)
-    except queue.Empty:
-        return None
-
-
-# ---------------------------------------------------------------------------
-# Worker process
-# ---------------------------------------------------------------------------
-
-
-class WorkerFactory(Protocol):
-
-    def __call__(
-        self,
-        *,
-        gpu_id: int,
-        generator_config: GeneratorConfig,
-        warmup_config: WarmupConfig,
-    ) -> _WorkerHandle:
-        ...
-
-
-def _default_worker_factory(
-    *,
-    gpu_id: int,
-    generator_config: GeneratorConfig,
-    warmup_config: WarmupConfig,
-) -> _WorkerHandle:
-    """Spawn a real multiprocessing worker.
-
-    The child process calls :func:`worker_main` which constructs a
-    :class:`VideoGenerator` from ``generator_config`` and runs a
-    blocking job loop. The ``ready`` event flips after the warmup
-    request completes.
-    """
-    ctx = mp.get_context("spawn")
-    job_queue: mp.Queue = ctx.Queue()
-    result_queue: mp.Queue = ctx.Queue()
-    ready = threading.Event()
-    boot_ok = threading.Event()
-    shutdown_event = ctx.Event()
-    worker_id = f"gpu{gpu_id}-{uuid.uuid4().hex[:6]}"
-    process = ctx.Process(
-        target=worker_main,
-        kwargs={
-            "gpu_id": gpu_id,
-            "worker_id": worker_id,
-            "generator_config": generator_config,
-            "warmup_config": warmup_config,
-            "job_queue": job_queue,
-            "result_queue": result_queue,
-            "shutdown_event": shutdown_event,
-        },
-        daemon=False,
-    )
-    process.start()
-
-    # Block the parent-side ``ready`` flag until the worker posts a
-    # ready acknowledgement on the result queue. We drain that single
-    # sentinel here; subsequent results belong to jobs. ``boot_ok``
-    # only flips on a real ready; on error we set ``ready`` to unblock
-    # the parent's wait but leave ``boot_ok`` clear so the pool keeps
-    # the worker out of the available queue.
-    def _await_ready() -> None:
-        while not shutdown_event.is_set():
-            try:
-                msg = result_queue.get(timeout=1.0)
-            except queue.Empty:
-                continue
-            if isinstance(msg, dict) and msg.get("kind") == "ready":
-                boot_ok.set()
-                ready.set()
-                return
-            if isinstance(msg, dict) and msg.get("kind") == "error":
-                logger.error("pool: worker %s failed to boot: %s", worker_id, msg.get("error"))
-                ready.set()
-                return
-
-    threading.Thread(target=_await_ready, daemon=True).start()
-
-    return _WorkerHandle(
-        process=process,
-        job_queue=job_queue,
-        result_queue=result_queue,
-        gpu_id=gpu_id,
-        worker_id=worker_id,
-        ready=ready,
-        boot_ok=boot_ok,
-        shutdown_event=shutdown_event,
-    )
-
-
-__all__ = [
-    "GpuPool",
-    "InProcessGpuPool",
-    "PoolAcquireTimeout",
-    "PoolAssignment",
-    "PoolHealth",
-    "SubprocessGpuPool",
-    "WorkerFactory",
-    "worker_main",
-]

fastvideo/entrypoints/streaming/mock_server.py

@@ -1,122 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Mock streaming server — a frontend dev aid.
-
-Boots the same FastAPI app the real streaming server uses, but backs
-it with :class:`InProcessGpuPool` wrapping a synthetic generator that
-emits pre-baked RGB frames. No GPU or model weights required.
-
-Use cases:
-
-* Frontend development without a real model loaded.
-* Integration tests that exercise the WS protocol end-to-end.
-* Reproducing protocol bugs locally.
-
-Launch: ``python -m fastvideo.entrypoints.streaming.mock_server``.
-"""
-from __future__ import annotations
-
-import argparse
-import time
-from dataclasses import dataclass
-from typing import Any
-
-import numpy as np
-
-from fastvideo.api.schema import (
-    ContinuationState,
-    GenerationRequest,
-    GeneratorConfig,
-    SamplingConfig,
-    ServeConfig,
-    StreamingConfig,
-)
-from fastvideo.entrypoints.streaming.server import build_app
-
-
-@dataclass
-class MockGenerator:
-    """Generator stand-in that returns synthetic gradient frames.
-
-    Each call produces one segment worth of frames whose pixels vary by
-    a constant derived from the request seed and segment index. Latency
-    is configurable via ``sleep_ms`` so the caller can exercise slow-
-    generate scenarios without spinning a GPU.
-    """
-
-    sleep_ms: float = 0.0
-
-    def generate(self, request: GenerationRequest) -> dict[str, Any]:
-        if self.sleep_ms:
-            time.sleep(self.sleep_ms / 1000.0)
-        width = max(16, request.sampling.width)
-        height = max(16, request.sampling.height)
-        num_frames = max(1, request.sampling.num_frames)
-        frames = [_gradient_frame(height, width, idx, seed=request.sampling.seed) for idx in range(num_frames)]
-        state = ContinuationState(
-            kind="ltx2.v1",
-            payload={
-                "schema_version": 1,
-                "segment_index": 0,
-                "source_prompt": request.prompt,
-            },
-        )
-        return {
-            "frames": frames,
-            "audio_sample_rate": 24000,
-            "state": state,
-        }
-
-
-def _gradient_frame(height: int, width: int, idx: int, *, seed: int) -> np.ndarray:
-    base = (idx * 17 + seed * 3) % 256
-    row = np.linspace(base, (base + 64) % 256, width, dtype=np.uint8)
-    frame = np.tile(row, (height, 1))
-    stacked = np.stack([frame, np.roll(frame, 8, axis=1), np.roll(frame, 16, axis=1)], axis=-1)
-    return stacked.astype(np.uint8)
-
-
-def build_mock_app(*, sleep_ms: float = 0.0):
-    """Build a FastAPI app backed by :class:`MockGenerator`."""
-    serve_config = ServeConfig(
-        generator=GeneratorConfig(model_path="/models/mock"),
-        streaming=StreamingConfig(
-            session_timeout_seconds=120,
-            generation_segment_cap=6,
-        ),
-    )
-    serve_config.default_request.sampling = SamplingConfig(
-        num_frames=24,
-        height=256,
-        width=256,
-        fps=24,
-        num_inference_steps=1,
-    )
-    return build_app(serve_config, MockGenerator(sleep_ms=sleep_ms))
-
-
-def main() -> None:  # pragma: no cover - CLI entry
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--host", default="127.0.0.1")
-    parser.add_argument("--port", type=int, default=8000)
-    parser.add_argument(
-        "--sleep-ms",
-        type=float,
-        default=0.0,
-        help="Per-segment artificial latency for testing slow paths",
-    )
-    args = parser.parse_args()
-
-    import uvicorn
-
-    app = build_mock_app(sleep_ms=args.sleep_ms)
-    uvicorn.run(app, host=args.host, port=args.port)
-
-
-__all__ = [
-    "MockGenerator",
-    "build_mock_app",
-    "main",
-]
-
-if __name__ == "__main__":  # pragma: no cover - CLI entry
-    main()

fastvideo/entrypoints/streaming/prompt/__init__.py

@@ -1,36 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Prompt pipeline for the streaming server.
-
-* :mod:`providers` — LLM backend abstraction + built-in adapters
-* :mod:`enhancer` — provider-agnostic enhance / auto-extend / rewrite
-  operations on top of the provider layer
-
-All of this is optional; the streaming server runs fine without it
-(PR 7.5's skeleton never invokes the enhancer). When the operator
-enables ``ServeConfig.streaming.prompt.enabled``, the server routes
-each ``session_init_v2`` curated prompt through ``enhance`` before the
-first segment.
-"""
-from fastvideo.entrypoints.streaming.prompt.enhancer import (
-    PromptEnhancer,
-    PromptOperation,
-)
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMMessage,
-    LLMProvider,
-    LLMProviderError,
-    LLMRequest,
-    LLMResponse,
-    LLMTimeoutError,
-)
-
-__all__ = [
-    "LLMMessage",
-    "LLMProvider",
-    "LLMProviderError",
-    "LLMRequest",
-    "LLMResponse",
-    "LLMTimeoutError",
-    "PromptEnhancer",
-    "PromptOperation",
-]

fastvideo/entrypoints/streaming/prompt/enhancer.py

@@ -1,197 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Provider-agnostic prompt orchestration for the streaming server.
-
-Three operations the streaming server needs:
-
-* ``enhance`` — polish a user prompt (add cinematic detail, fix syntax)
-* ``auto_extend`` — generate a follow-on prompt for loop generation
-* ``rewrite`` — rewrite a seed prompt for a user-directed rewrite flow
-
-All three share the same orchestration: pick a provider in priority
-order, submit an ``LLMRequest``, fall back to the next provider on
-retryable errors, and surface a structured :class:`LLMResponse` back
-to the caller.
-
-System prompts are loaded from ``system_prompt_dir`` on construction
-and can be hot-reloaded via :meth:`PromptEnhancer.reload_system_prompts`.
-The streaming server's management endpoint calls that method in
-response to a ``rewrite_seed_prompts_started`` frame.
-"""
-from __future__ import annotations
-
-import enum
-import os
-from collections.abc import Sequence
-from dataclasses import dataclass, replace
-
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMMessage,
-    LLMProvider,
-    LLMProviderError,
-    LLMRequest,
-    LLMResponse,
-)
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-
-class PromptOperation(enum.Enum):
-    ENHANCE = "enhance"
-    AUTO_EXTEND = "auto_extend"
-    REWRITE = "rewrite"
-
-
-@dataclass
-class _SystemPrompts:
-    enhance: str
-    auto_extend: str
-    rewrite: str
-
-
-_DEFAULT_SYSTEM_PROMPTS = _SystemPrompts(
-    enhance=("You are a prompt enhancer for cinematic video generation. Given "
-             "a user prompt, produce an enhanced prompt that is more vivid, "
-             "specific, and concrete. Keep the subject intact; add lighting, "
-             "camera, and motion detail. Reply with just the enhanced prompt."),
-    auto_extend=("You are a video continuation assistant. Given the current "
-                 "sequence of prompts, produce one new prompt that naturally "
-                 "continues the sequence. Reply with just the next prompt."),
-    rewrite=("You are a creative prompt rewriter. Given a seed prompt, produce "
-             "a set of alternative prompts that explore different angles, "
-             "styles, and moods. Reply with one prompt per line."),
-)
-
-
-class PromptEnhancer:
-    """Orchestrates prompt operations across a priority-ordered provider
-    list with structured fallback + hot-reloadable system prompts.
-
-    Usage::
-
-        enhancer = PromptEnhancer(
-            providers=[CerebrasProvider(), GroqProvider()],
-            model="gpt-oss-120b",
-            system_prompt_dir="/etc/fastvideo/prompts",
-        )
-        response = await enhancer.enhance("a fox running through snow")
-    """
-
-    def __init__(
-        self,
-        *,
-        providers: Sequence[LLMProvider],
-        model: str,
-        timeout_ms: int = 20000,
-        temperature: float = 0.7,
-        max_tokens: int | None = 256,
-        system_prompt_dir: str | None = None,
-    ) -> None:
-        if not providers:
-            raise ValueError("PromptEnhancer requires at least one LLMProvider")
-        self._providers = list(providers)
-        self._model = model
-        self._timeout_ms = timeout_ms
-        self._temperature = temperature
-        self._max_tokens = max_tokens
-        self._system_prompt_dir = system_prompt_dir
-        self._system_prompts = self._load_system_prompts()
-
-    @property
-    def providers(self) -> list[LLMProvider]:
-        return list(self._providers)
-
-    def register_provider(self, provider: LLMProvider, *, priority: int = -1) -> None:
-        """Insert an additional provider. ``priority=0`` makes it primary;
-        ``priority=-1`` (default) appends as a fallback."""
-        if priority < 0:
-            self._providers.append(provider)
-        else:
-            self._providers.insert(priority, provider)
-
-    def reload_system_prompts(self) -> None:
-        """Re-read the system prompt files from ``system_prompt_dir``.
-
-        The streaming server exposes this via a management endpoint so
-        operators can iterate on prompt templates without restarting
-        workers.
-        """
-        self._system_prompts = self._load_system_prompts()
-        logger.info("prompt enhancer: reloaded system prompts from %s", self._system_prompt_dir or "defaults")
-
-    async def enhance(self, prompt: str) -> LLMResponse:
-        return await self._run(
-            PromptOperation.ENHANCE,
-            system=self._system_prompts.enhance,
-            user=prompt,
-        )
-
-    async def auto_extend(self, prior_prompts: Sequence[str]) -> LLMResponse:
-        user = "\n".join(prior_prompts)
-        return await self._run(
-            PromptOperation.AUTO_EXTEND,
-            system=self._system_prompts.auto_extend,
-            user=user,
-        )
-
-    async def rewrite(self, seed_prompt: str) -> LLMResponse:
-        return await self._run(
-            PromptOperation.REWRITE,
-            system=self._system_prompts.rewrite,
-            user=seed_prompt,
-        )
-
-    async def _run(
-        self,
-        operation: PromptOperation,
-        *,
-        system: str,
-        user: str,
-    ) -> LLMResponse:
-        request = LLMRequest(
-            messages=[
-                LLMMessage(role="system", content=system),
-                LLMMessage(role="user", content=user),
-            ],
-            model=self._model,
-            max_tokens=self._max_tokens,
-            temperature=self._temperature,
-            timeout_ms=self._timeout_ms,
-        )
-        last_error: LLMProviderError | None = None
-        for idx, provider in enumerate(self._providers):
-            try:
-                response = await provider.complete(request)
-                if idx > 0:
-                    # Mark the fallback flag without losing any other
-                    # response fields the provider populated.
-                    response = replace(response, fallback_used=True)
-                return response
-            except LLMProviderError as exc:
-                logger.warning("prompt %s: provider %s failed: %s; trying next", operation.value, provider.name, exc)
-                last_error = exc
-                if not exc.retryable:
-                    break
-        assert last_error is not None
-        raise last_error
-
-    def _load_system_prompts(self) -> _SystemPrompts:
-        if not self._system_prompt_dir:
-            return _DEFAULT_SYSTEM_PROMPTS
-        return _SystemPrompts(
-            enhance=_read_prompt(self._system_prompt_dir, "enhance.txt", _DEFAULT_SYSTEM_PROMPTS.enhance),
-            auto_extend=_read_prompt(self._system_prompt_dir, "auto_extend.txt", _DEFAULT_SYSTEM_PROMPTS.auto_extend),
-            rewrite=_read_prompt(self._system_prompt_dir, "rewrite.txt", _DEFAULT_SYSTEM_PROMPTS.rewrite),
-        )
-
-
-def _read_prompt(dirname: str, filename: str, default: str) -> str:
-    path = os.path.join(dirname, filename)
-    if not os.path.exists(path):
-        return default
-    with open(path, encoding="utf-8") as f:
-        content = f.read().strip()
-    return content or default
-
-
-__all__ = ["PromptEnhancer", "PromptOperation"]

fastvideo/entrypoints/streaming/prompt/providers/__init__.py

@@ -1,24 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""LLM provider implementations used by the prompt enhancer."""
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMMessage,
-    LLMProvider,
-    LLMProviderError,
-    LLMRequest,
-    LLMResponse,
-    LLMTimeoutError,
-)
-from fastvideo.entrypoints.streaming.prompt.providers.cerebras import (
-    CerebrasProvider, )
-from fastvideo.entrypoints.streaming.prompt.providers.groq import GroqProvider
-
-__all__ = [
-    "CerebrasProvider",
-    "GroqProvider",
-    "LLMMessage",
-    "LLMProvider",
-    "LLMProviderError",
-    "LLMRequest",
-    "LLMResponse",
-    "LLMTimeoutError",
-]

fastvideo/entrypoints/streaming/prompt/providers/_openai_compat.py

@@ -1,101 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Shared HTTP path for OpenAI-compatible ``/chat/completions`` providers.
-
-Cerebras and Groq both expose the OpenAI chat-completions schema, so
-the request shape, error mapping, and response decoding are identical
-between them. This module centralizes that logic; the per-provider
-modules stay thin (just defaults + env var wiring).
-"""
-from __future__ import annotations
-
-import time
-
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMProviderError,
-    LLMRequest,
-    LLMResponse,
-    LLMTimeoutError,
-)
-
-
-async def complete_openai_compatible(
-    *,
-    api_key: str | None,
-    api_key_hint: str,
-    base_url: str,
-    provider_name: str,
-    request: LLMRequest,
-) -> LLMResponse:
-    """Issue a chat-completions call and decode the OpenAI response."""
-    if not api_key:
-        raise LLMProviderError(
-            f"{provider_name} provider requires {api_key_hint} "
-            "(or explicit api_key=...)",
-            retryable=False,
-        )
-    try:
-        import httpx
-    except ImportError as exc:  # pragma: no cover - optional dep
-        raise LLMProviderError(
-            f"{provider_name} provider requires httpx; install httpx",
-            retryable=False,
-        ) from exc
-
-    timeout_s = (request.timeout_ms or 20000) / 1000.0
-    t0 = time.perf_counter()
-    try:
-        async with httpx.AsyncClient(timeout=timeout_s) as client:
-            response = await client.post(
-                f"{base_url}/chat/completions",
-                headers={
-                    "Authorization": f"Bearer {api_key}",
-                    "Content-Type": "application/json",
-                },
-                json={
-                    "model": request.model,
-                    "messages": [{
-                        "role": m.role,
-                        "content": m.content
-                    } for m in request.messages],
-                    "max_tokens": request.max_tokens,
-                    "temperature": request.temperature,
-                },
-            )
-    except httpx.TimeoutException as exc:
-        raise LLMTimeoutError(f"{provider_name} timed out after {timeout_s}s") from exc
-    except httpx.HTTPError as exc:
-        raise LLMProviderError(f"{provider_name} HTTP error: {exc}") from exc
-
-    if response.status_code >= 400:
-        # 5xx and 429 (rate-limit) are retryable: another provider may
-        # succeed. 4xx (auth, bad-request, etc.) are client errors —
-        # the enhancer should stop fallback traversal.
-        retryable = (response.status_code >= 500 or response.status_code == 429)
-        raise LLMProviderError(
-            f"{provider_name} returned {response.status_code}: "
-            f"{response.text[:200]}",
-            retryable=retryable,
-        )
-
-    try:
-        data = response.json()
-    except Exception as exc:
-        # Non-JSON body usually means a proxy / load-balancer error
-        # page; leave it retryable so a fallback provider can try.
-        raise LLMProviderError(f"{provider_name} returned non-JSON body: {exc}") from exc
-
-    choices = data.get("choices") or []
-    if not choices:
-        raise LLMProviderError(f"{provider_name} returned no choices")
-    content = choices[0].get("message", {}).get("content") or ""
-
-    latency_ms = (time.perf_counter() - t0) * 1000.0
-    return LLMResponse(
-        content=content.strip(),
-        provider=provider_name,
-        model=request.model,
-        latency_ms=latency_ms,
-    )
-
-
-__all__ = ["complete_openai_compatible"]

fastvideo/entrypoints/streaming/prompt/providers/base.py

@@ -1,85 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""LLM provider protocol + DTOs used by the prompt enhancer.
-
-Third-party users add a new provider by implementing
-:class:`LLMProvider` and registering it with a prompt enhancer
-instance. The shipped providers live in sibling modules
-(``cerebras.py``, ``groq.py``) and each is ~100-200 LOC — the
-provider layer is intentionally thin so the enhancer stays
-provider-agnostic.
-"""
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Literal, Protocol, runtime_checkable
-
-
-@dataclass
-class LLMMessage:
-    role: Literal["system", "user", "assistant"]
-    content: str
-
-
-@dataclass
-class LLMRequest:
-    messages: list[LLMMessage]
-    model: str
-    max_tokens: int | None = None
-    temperature: float | None = None
-    timeout_ms: int | None = None
-
-
-@dataclass
-class LLMResponse:
-    content: str
-    provider: str
-    model: str
-    latency_ms: float
-    fallback_used: bool = False
-
-
-class LLMProviderError(RuntimeError):
-    """Raised when an LLM provider fails a request.
-
-    ``retryable`` controls whether the enhancer falls back to the next
-    provider. It is settable per-instance so the same exception type
-    can describe retryable transport errors (5xx, 429) and
-    non-retryable client errors (4xx auth/bad-request) without forcing
-    a separate subclass for every status family.
-    """
-
-    def __init__(self, message: str, *, retryable: bool = True) -> None:
-        super().__init__(message)
-        self.retryable = retryable
-
-
-class LLMTimeoutError(LLMProviderError):
-    """Raised when an LLM provider times out — always retryable."""
-
-    def __init__(self, message: str) -> None:
-        super().__init__(message, retryable=True)
-
-
-@runtime_checkable
-class LLMProvider(Protocol):
-    """Provider interface every LLM adapter implements.
-
-    Providers are async-first because every built-in implementation
-    talks to an HTTP API. Synchronous providers can wrap their call in
-    ``asyncio.to_thread`` internally.
-    """
-
-    name: str
-
-    async def complete(self, request: LLMRequest) -> LLMResponse:
-        ...
-
-
-__all__ = [
-    "LLMMessage",
-    "LLMProvider",
-    "LLMProviderError",
-    "LLMRequest",
-    "LLMResponse",
-    "LLMTimeoutError",
-]

fastvideo/entrypoints/streaming/prompt/providers/cerebras.py

@@ -1,44 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Cerebras LLM provider (OpenAI-compatible chat endpoint)."""
-from __future__ import annotations
-
-import os
-from dataclasses import dataclass
-
-from fastvideo.entrypoints.streaming.prompt.providers._openai_compat import (
-    complete_openai_compatible, )
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMRequest,
-    LLMResponse,
-)
-
-_DEFAULT_BASE_URL = "https://api.cerebras.ai/v1"
-_API_KEY_ENV = "CEREBRAS_API_KEY"
-
-
-@dataclass
-class CerebrasProvider:
-    """Cerebras inference adapter.
-
-    ``api_key`` falls back to ``CEREBRAS_API_KEY`` when unset.
-    """
-
-    api_key: str | None = None
-    base_url: str = _DEFAULT_BASE_URL
-    name: str = "cerebras"
-
-    def __post_init__(self) -> None:
-        if self.api_key is None:
-            self.api_key = os.environ.get(_API_KEY_ENV)
-
-    async def complete(self, request: LLMRequest) -> LLMResponse:
-        return await complete_openai_compatible(
-            api_key=self.api_key,
-            api_key_hint=_API_KEY_ENV,
-            base_url=self.base_url,
-            provider_name=self.name,
-            request=request,
-        )
-
-
-__all__ = ["CerebrasProvider"]

fastvideo/entrypoints/streaming/prompt/providers/groq.py

@@ -1,46 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Groq LLM provider (OpenAI-compatible chat endpoint)."""
-from __future__ import annotations
-
-import os
-from dataclasses import dataclass
-
-from fastvideo.entrypoints.streaming.prompt.providers._openai_compat import (
-    complete_openai_compatible, )
-from fastvideo.entrypoints.streaming.prompt.providers.base import (
-    LLMRequest,
-    LLMResponse,
-)
-
-_DEFAULT_BASE_URL = "https://api.groq.com/openai/v1"
-_API_KEY_ENV = "GROQ_API_KEY"
-
-
-@dataclass
-class GroqProvider:
-    """Groq inference adapter.
-
-    Identical wire format to :class:`CerebrasProvider`; both go through
-    :func:`complete_openai_compatible`. The two providers differ only
-    in base URL, env var, and model id conventions.
-    """
-
-    api_key: str | None = None
-    base_url: str = _DEFAULT_BASE_URL
-    name: str = "groq"
-
-    def __post_init__(self) -> None:
-        if self.api_key is None:
-            self.api_key = os.environ.get(_API_KEY_ENV)
-
-    async def complete(self, request: LLMRequest) -> LLMResponse:
-        return await complete_openai_compatible(
-            api_key=self.api_key,
-            api_key_hint=_API_KEY_ENV,
-            base_url=self.base_url,
-            provider_name=self.name,
-            request=request,
-        )
-
-
-__all__ = ["GroqProvider"]

fastvideo/entrypoints/streaming/prompt/rewrite.py

@@ -1,82 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Rewrite payload builder.
-
-The UI's "rewrite seed prompts" flow asks the enhancer to produce a
-batch of alternative prompts given one seed. This module packages the
-seed + options into the payload the enhancer expects and unpacks the
-response back into a typed :class:`RewriteResult`.
-
-Separating this from :mod:`enhancer` keeps the enhancer provider-
-agnostic; anything UI-specific (how many alternatives to request, how
-to split the response, temperature) lives here.
-"""
-from __future__ import annotations
-
-import re
-from dataclasses import dataclass
-
-from fastvideo.entrypoints.streaming.prompt.enhancer import PromptEnhancer
-
-_LEADING_MARKER_RE = re.compile(r"^(?:[-*•]\s*|\d+\s*[.)]\s*)+")
-
-
-@dataclass
-class RewriteOptions:
-    count: int = 3
-    """Number of alternative prompts to request."""
-    temperature: float | None = None
-
-
-@dataclass
-class RewriteResult:
-    seed_prompt: str
-    alternatives: list[str]
-    provider: str
-    model: str
-    latency_ms: float
-    fallback_used: bool = False
-
-
-async def build_rewrite(
-    enhancer: PromptEnhancer,
-    seed_prompt: str,
-    *,
-    options: RewriteOptions | None = None,
-) -> RewriteResult:
-    """Run a rewrite op through the enhancer and return a typed result."""
-    if not seed_prompt.strip():
-        raise ValueError("rewrite seed prompt must be non-empty")
-    options = options or RewriteOptions()
-    response = await enhancer.rewrite(seed_prompt)
-    alternatives = _split_response(response.content, limit=options.count)
-    return RewriteResult(
-        seed_prompt=seed_prompt,
-        alternatives=alternatives,
-        provider=response.provider,
-        model=response.model,
-        latency_ms=response.latency_ms,
-        fallback_used=response.fallback_used,
-    )
-
-
-def _split_response(content: str, *, limit: int) -> list[str]:
-    """Split the LLM response into discrete prompt candidates.
-
-    The shipped system prompt instructs the model to emit one prompt
-    per line; this function is forgiving about numbered lists or
-    leading bullets so user-supplied system prompts don't break it.
-    """
-    lines = [line.strip() for line in content.splitlines() if line.strip()]
-    cleaned: list[str] = []
-    for line in lines:
-        stripped = _LEADING_MARKER_RE.sub("", line).strip()
-        if stripped:
-            cleaned.append(stripped)
-    return cleaned[:max(1, limit)]
-
-
-__all__ = [
-    "RewriteOptions",
-    "RewriteResult",
-    "build_rewrite",
-]

fastvideo/entrypoints/streaming/prompt/safety.py

@@ -1,146 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Optional prompt safety filter.
-
-Uses a fastText classifier to score prompts against a banned-content
-rubric. Only loaded when ``ServeConfig.streaming.safety.enabled`` is
-True and fastText is installed — users who don't need it see no
-runtime cost.
-
-Install: ``pip install fastvideo[prompt-safety]`` (ships fasttext as an
-optional extra) or install fasttext directly.
-"""
-from __future__ import annotations
-
-import enum
-import threading
-from dataclasses import dataclass
-from typing import Any
-
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-
-class SafetyDecision(enum.Enum):
-    ALLOW = "allow"
-    BLOCK = "block"
-    UNAVAILABLE = "unavailable"
-    """Returned when the classifier can't run (not configured, fastText
-    missing). Safety is opt-in; the server treats ``UNAVAILABLE`` as
-    ``ALLOW`` but logs it so operators know the filter is off."""
-
-
-@dataclass
-class SafetyResult:
-    prompt: str
-    decision: SafetyDecision
-    score: float = 0.0
-    label: str | None = None
-    reason: str | None = None
-
-
-class PromptSafetyFilter:
-    """Minimal fastText-backed prompt safety filter.
-
-    Loads the classifier lazily on first use so the streaming server
-    can construct the filter eagerly at startup without paying the
-    model-load cost when safety is disabled.
-    """
-
-    def __init__(
-        self,
-        *,
-        classifier_path: str | None,
-        enabled: bool = True,
-        block_threshold: float = 0.5,
-    ) -> None:
-        self._classifier_path = classifier_path
-        self._enabled = enabled
-        self._block_threshold = block_threshold
-        self._model: Any | None = None
-        self._load_attempted = False
-        self._load_lock = threading.Lock()
-
-    @property
-    def enabled(self) -> bool:
-        return self._enabled and self._classifier_path is not None
-
-    def classify(self, prompt: str) -> SafetyResult:
-        if not self.enabled:
-            return SafetyResult(
-                prompt=prompt,
-                decision=SafetyDecision.UNAVAILABLE,
-                reason="safety filter not enabled",
-            )
-        model = self._ensure_loaded()
-        if model is None:
-            return SafetyResult(
-                prompt=prompt,
-                decision=SafetyDecision.UNAVAILABLE,
-                reason="fastText model unavailable",
-            )
-        try:
-            labels, probs = model.predict(prompt.replace("\n", " "), k=1)
-        except Exception as exc:  # pragma: no cover - defensive
-            logger.warning("safety: classifier failed: %s", exc)
-            return SafetyResult(
-                prompt=prompt,
-                decision=SafetyDecision.UNAVAILABLE,
-                reason=f"classifier error: {exc}",
-            )
-        label = labels[0].removeprefix("__label__") if labels else None
-        score = float(probs[0]) if len(probs) else 0.0
-        decision = (SafetyDecision.BLOCK if
-                    (label == "unsafe" and score >= self._block_threshold) else SafetyDecision.ALLOW)
-        return SafetyResult(
-            prompt=prompt,
-            decision=decision,
-            score=score,
-            label=label,
-        )
-
-    def _ensure_loaded(self) -> Any | None:
-        if self._model is not None:
-            return self._model
-        if self._load_attempted:
-            return None
-        with self._load_lock:
-            if self._model is not None:
-                return self._model
-            if self._load_attempted:
-                return None
-            self._load_attempted = True
-            if self._classifier_path is None:
-                return None
-            try:
-                import fasttext  # type: ignore[import-not-found]
-            except ImportError:
-                logger.warning("safety: fasttext not installed; safety filter disabled. "
-                               "Install fastvideo[prompt-safety] to enable.")
-                return None
-            try:
-                self._model = fasttext.load_model(self._classifier_path)
-            except Exception as exc:  # pragma: no cover - requires real model
-                logger.warning("safety: failed to load %s: %s", self._classifier_path, exc)
-                return None
-            return self._model
-
-
-def first_blocked(
-    filter_: PromptSafetyFilter,
-    prompts: list[str],
-) -> SafetyResult | None:
-    """Return the first prompt the filter blocks, or ``None``."""
-    for prompt in prompts:
-        result = filter_.classify(prompt)
-        if result.decision is SafetyDecision.BLOCK:
-            return result
-    return None
-
-
-__all__ = [
-    "PromptSafetyFilter",
-    "SafetyDecision",
-    "SafetyResult",
-    "first_blocked",
-]

fastvideo/entrypoints/streaming/protocol.py

@@ -1,252 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""JSON WebSocket protocol schemas for the streaming server.
-
-Every control message shares the envelope ``{"type": <str>, ...}``.
-Pydantic models live here so the server can parse / validate incoming
-frames and emit well-typed outgoing frames without hand-rolled dicts.
-
-The message catalogue matches the contract in
-``docs/design/server_contracts/streaming.md``; additions must land in
-both places in the same PR.
-"""
-from __future__ import annotations
-
-from typing import Annotated, Any, Literal, Union
-
-from pydantic import BaseModel, ConfigDict, Field
-
-# ---------------------------------------------------------------------------
-# Client → server
-# ---------------------------------------------------------------------------
-
-
-class SessionInitV2(BaseModel):
-    """Opening frame the client sends after the WebSocket handshake."""
-
-    model_config = ConfigDict(extra="allow")
-
-    type: Literal["session_init_v2"]
-    client_id: str | None = None
-    preset: str | None = None
-    preset_label: str | None = None
-    curated_prompts: list[str] = Field(default_factory=list)
-    initial_image: dict[str, Any] | None = None
-    enhancement_enabled: bool = False
-    auto_extension_enabled: bool = False
-    loop_generation_enabled: bool = False
-    single_clip_mode: bool = False
-    stream_mode: Literal["av_fmp4", "legacy_jpeg"] = "av_fmp4"
-    continuation_state: dict[str, Any] | None = None
-    """Optional ``{kind, payload}`` dict; hydrated into
-    :class:`fastvideo.api.ContinuationState` server-side."""
-
-
-class SegmentPromptSource(BaseModel):
-    """Request a new segment using a specific prompt."""
-
-    type: Literal["segment_prompt_source"]
-    prompt: str
-    negative_prompt: str | None = None
-    source: Literal["curated", "enhanced", "user", "auto_extension"] = "user"
-    seed: int | None = None
-    num_inference_steps: int | None = None
-    guidance_scale: float | None = None
-
-
-class SeedPromptsUpdated(BaseModel):
-    type: Literal["seed_prompts_updated"]
-    seed_prompts: list[str] = Field(default_factory=list)
-
-
-class EnhancementUpdated(BaseModel):
-    type: Literal["enhancement_updated"]
-    enabled: bool
-
-
-class AutoExtensionUpdated(BaseModel):
-    type: Literal["auto_extension_updated"]
-    enabled: bool
-
-
-class LoopGenerationUpdated(BaseModel):
-    type: Literal["loop_generation_updated"]
-    enabled: bool
-
-
-class GenerationPausedUpdated(BaseModel):
-    type: Literal["generation_paused_updated"]
-    paused: bool
-
-
-class SnapshotState(BaseModel):
-    """Request the current ``ContinuationState`` for export."""
-
-    type: Literal["snapshot_state"]
-
-
-ClientMessage = Annotated[
-    Union[  # noqa: UP007 - Annotated requires Union for discriminator
-        SessionInitV2,
-        SegmentPromptSource,
-        SeedPromptsUpdated,
-        EnhancementUpdated,
-        AutoExtensionUpdated,
-        LoopGenerationUpdated,
-        GenerationPausedUpdated,
-        SnapshotState,
-    ],
-    Field(discriminator="type"),
-]
-
-# ---------------------------------------------------------------------------
-# Server → client
-# ---------------------------------------------------------------------------
-
-
-class QueueStatus(BaseModel):
-    type: Literal["queue_status"] = "queue_status"
-    position: int
-    queue_depth: int
-
-
-class GpuAssigned(BaseModel):
-    type: Literal["gpu_assigned"] = "gpu_assigned"
-    gpu_id: int
-    session_timeout: int
-
-
-class Ltx2StreamStart(BaseModel):
-    type: Literal["ltx2_stream_start"] = "ltx2_stream_start"
-    preset: str | None = None
-    width: int
-    height: int
-    fps: int
-    num_frames: int
-
-
-class Ltx2SegmentStart(BaseModel):
-    type: Literal["ltx2_segment_start"] = "ltx2_segment_start"
-    segment_idx: int
-    prompt: str
-    total_steps: int
-
-
-class StepComplete(BaseModel):
-    type: Literal["step_complete"] = "step_complete"
-    segment_idx: int
-    step: int
-    total_steps: int
-    stage: str = "denoise"
-
-
-class MediaInit(BaseModel):
-    """Descriptor for the fMP4 initialization segment that follows."""
-
-    type: Literal["media_init"] = "media_init"
-    segment_idx: int
-    mime: str = "video/mp4; codecs=\"avc1.64001f, mp4a.40.2\""
-    stream_id: str
-    mode: Literal["av_fmp4"] = "av_fmp4"
-
-
-class MediaSegmentComplete(BaseModel):
-    type: Literal["media_segment_complete"] = "media_segment_complete"
-    segment_idx: int
-    stream_id: str
-    chunks: int
-    duration_ms: float | None = None
-    pts_base_ms: float | None = None
-
-
-class Ltx2SegmentComplete(BaseModel):
-    type: Literal["ltx2_segment_complete"] = "ltx2_segment_complete"
-    segment_idx: int
-    generation_time_ms: float
-    e2e_latency_ms: float | None = None
-
-
-class Ltx2StreamComplete(BaseModel):
-    type: Literal["ltx2_stream_complete"] = "ltx2_stream_complete"
-    reason: Literal["segment_cap", "stop_requested", "error"] = "stop_requested"
-
-
-class SessionTimeout(BaseModel):
-    type: Literal["session_timeout"] = "session_timeout"
-    timeout_seconds: int
-
-
-class ContinuationStateSnapshot(BaseModel):
-    type: Literal["continuation_state_snapshot"] = "continuation_state_snapshot"
-    state: dict[str, Any]
-    """``{kind, payload}`` dict matching
-    :class:`fastvideo.api.ContinuationState`."""
-
-
-class ErrorMessage(BaseModel):
-    type: Literal["error"] = "error"
-    code: Literal[
-        "session_rejected",
-        "invalid_message",
-        "preset_mismatch",
-        "gpu_unavailable",
-        "worker_failed",
-        "upstream_timeout",
-        "internal_error",
-    ] = "internal_error"
-    message: str
-    retryable: bool = False
-
-
-ServerMessage = Union[  # noqa: UP007 - pydantic Union handling
-    QueueStatus,
-    GpuAssigned,
-    Ltx2StreamStart,
-    Ltx2SegmentStart,
-    StepComplete,
-    MediaInit,
-    MediaSegmentComplete,
-    Ltx2SegmentComplete,
-    Ltx2StreamComplete,
-    SessionTimeout,
-    ContinuationStateSnapshot,
-    ErrorMessage,
-]
-
-
-def parse_client_message(raw: dict[str, Any]) -> ClientMessage:
-    """Parse an incoming WebSocket dict into a typed client message.
-
-    Unknown ``type`` values raise :class:`pydantic.ValidationError`; the
-    server handler turns that into an ``error`` frame with
-    ``code="invalid_message"``.
-    """
-    from pydantic import TypeAdapter
-
-    return TypeAdapter(ClientMessage).validate_python(raw)
-
-
-__all__ = [
-    "AutoExtensionUpdated",
-    "ClientMessage",
-    "ContinuationStateSnapshot",
-    "EnhancementUpdated",
-    "ErrorMessage",
-    "GenerationPausedUpdated",
-    "GpuAssigned",
-    "Ltx2SegmentComplete",
-    "Ltx2SegmentStart",
-    "Ltx2StreamComplete",
-    "Ltx2StreamStart",
-    "LoopGenerationUpdated",
-    "MediaInit",
-    "MediaSegmentComplete",
-    "QueueStatus",
-    "SeedPromptsUpdated",
-    "SegmentPromptSource",
-    "ServerMessage",
-    "SessionInitV2",
-    "SessionTimeout",
-    "SnapshotState",
-    "StepComplete",
-    "parse_client_message",
-]

fastvideo/entrypoints/streaming/router/__init__.py

@@ -1,27 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Multi-replica load balancer + WebSocket proxy for the streaming server.
-
-Sits in front of one-or-more streaming-server replicas and forwards
-WebSocket sessions to a healthy primary, with failover to secondaries.
-Kept in-repo under ``fastvideo/entrypoints/streaming/router/`` per the
-PR plan's default; the alternative (separate package) is an open
-question deferred to review.
-"""
-from fastvideo.entrypoints.streaming.router.registry import (
-    Replica,
-    ReplicaHealth,
-    ReplicaRegistry,
-    ReplicaStatus,
-)
-from fastvideo.entrypoints.streaming.router.config import RouterConfig
-from fastvideo.entrypoints.streaming.router.main import build_router_app, run_router
-
-__all__ = [
-    "Replica",
-    "ReplicaHealth",
-    "ReplicaRegistry",
-    "ReplicaStatus",
-    "RouterConfig",
-    "build_router_app",
-    "run_router",
-]

fastvideo/entrypoints/streaming/router/config.py

@@ -1,88 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Typed router configuration."""
-from __future__ import annotations
-
-from dataclasses import dataclass, field
-from urllib.parse import urlparse
-
-
-@dataclass
-class ReplicaEndpoint:
-    """One backend replica the router can route to."""
-
-    url: str
-    """HTTP base URL, e.g. ``http://host:8000``. WebSocket URL is
-    derived automatically by replacing the scheme."""
-    name: str | None = None
-    primary: bool = False
-    """``True`` = prefer this replica over others in steady state."""
-    weight: float = 1.0
-
-
-@dataclass
-class RouterConfig:
-    """Typed router config loaded from a YAML file.
-
-    Example::
-
-        router:
-          host: 0.0.0.0
-          port: 9000
-          replicas:
-            - url: http://streamer-a:8000
-              primary: true
-            - url: http://streamer-b:8000
-          health_check:
-            path: /health
-            interval_seconds: 5
-            failure_threshold: 3
-
-    Validation runs in ``__post_init__``: empty replicas, non-positive
-    intervals/timeouts, thresholds < 1, non-http(s) URLs, and more than
-    one primary all raise ``ValueError`` so misconfigurations surface at
-    load time rather than as confusing runtime failures.
-    """
-
-    host: str = "0.0.0.0"
-    port: int = 9000
-    replicas: list[ReplicaEndpoint] = field(default_factory=list)
-    health_check_path: str = "/health"
-    health_check_interval_seconds: float = 5.0
-    health_check_timeout_seconds: float = 2.0
-    failure_threshold: int = 3
-    recovery_threshold: int = 2
-
-    def __post_init__(self) -> None:
-        if not self.replicas:
-            raise ValueError("RouterConfig.replicas must list at least one replica")
-        if self.health_check_interval_seconds <= 0:
-            raise ValueError(f"health_check_interval_seconds must be > 0, got {self.health_check_interval_seconds}")
-        if self.health_check_timeout_seconds <= 0:
-            raise ValueError(f"health_check_timeout_seconds must be > 0, got {self.health_check_timeout_seconds}")
-        if self.failure_threshold < 1:
-            raise ValueError(f"failure_threshold must be >= 1, got {self.failure_threshold}")
-        if self.recovery_threshold < 1:
-            raise ValueError(f"recovery_threshold must be >= 1, got {self.recovery_threshold}")
-        seen_urls: set[str] = set()
-        for replica in self.replicas:
-            if not replica.url.startswith(("http://", "https://")):
-                raise ValueError(f"ReplicaEndpoint.url must start with http:// or https://, got {replica.url!r}")
-            parsed = urlparse(replica.url)
-            if parsed.path not in ("", "/"):
-                raise ValueError(f"ReplicaEndpoint.url must be a base host[:port] URL without a path; "
-                                 f"got {replica.url!r} with path {parsed.path!r}. The router appends "
-                                 "`/health` and `/v1/stream` itself.")
-            if parsed.query or parsed.fragment:
-                raise ValueError(f"ReplicaEndpoint.url must not include query/fragment; got {replica.url!r}")
-            if replica.url in seen_urls:
-                raise ValueError(f"Duplicate ReplicaEndpoint.url {replica.url!r}; "
-                                 "router selection keys by URL so duplicates would silently collapse")
-            seen_urls.add(replica.url)
-        primaries = sum(1 for r in self.replicas if r.primary)
-        if primaries > 1:
-            raise ValueError(f"RouterConfig allows at most one primary replica; got {primaries}. "
-                             "Multi-primary load distribution is deferred — promote one replica to "
-                             "primary and treat the rest as secondaries.")
-
-
-__all__ = ["ReplicaEndpoint", "RouterConfig"]

fastvideo/entrypoints/streaming/router/main.py

@@ -1,218 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Router FastAPI entry point.
-
-Exposes the same ``/v1/stream`` WebSocket path the backend servers do,
-accepts a client, picks a healthy replica from the registry, and
-proxies frames bidirectionally.
-
-PR 7.9 ships the minimum-viable shape: explicit replica list, single
-primary, JSON + binary passthrough in both directions, and a
-``/status`` endpoint for operators. Sticky-session routing (so a
-reconnect lands on the same backend) is left for a follow-up.
-"""
-from __future__ import annotations
-
-import asyncio
-import contextlib
-from dataclasses import dataclass
-
-from fastapi import FastAPI, WebSocket, WebSocketDisconnect
-from fastapi.responses import JSONResponse
-
-from fastvideo.entrypoints.streaming.router.config import RouterConfig
-from fastvideo.entrypoints.streaming.router.registry import (
-    ReplicaRegistry,
-    run_health_check_loop,
-)
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-
-@dataclass
-class _RouterState:
-    config: RouterConfig
-    registry: ReplicaRegistry
-    stop_event: asyncio.Event
-    health_task: asyncio.Task | None = None
-
-
-def build_router_app(
-    config: RouterConfig,
-    *,
-    registry: ReplicaRegistry | None = None,
-) -> FastAPI:
-    """Build the router FastAPI app.
-
-    ``registry`` can be injected for tests; defaults to one built from
-    ``config.replicas``.
-    """
-    registry = registry or ReplicaRegistry(config.replicas)
-    state = _RouterState(
-        config=config,
-        registry=registry,
-        stop_event=asyncio.Event(),
-    )
-
-    @contextlib.asynccontextmanager
-    async def _lifespan(_app: FastAPI):
-        state.health_task = asyncio.create_task(
-            run_health_check_loop(
-                registry=state.registry,
-                config=state.config,
-                stop_event=state.stop_event,
-            ))
-        try:
-            yield
-        finally:
-            state.stop_event.set()
-            if state.health_task is not None:
-                with contextlib.suppress(asyncio.CancelledError):
-                    await state.health_task
-
-    app = FastAPI(title="FastVideo Streaming Router", lifespan=_lifespan)
-
-    @app.get("/status")
-    async def _status() -> JSONResponse:
-        return JSONResponse({
-            "replicas": [{
-                "url": r.url,
-                "primary": r.primary,
-                "status": r.health.status.value,
-                "last_ok_at": r.health.last_ok_at,
-                "last_latency_ms": r.health.last_latency_ms,
-                "consecutive_failures": r.health.consecutive_failures,
-            } for r in state.registry.all()],
-        })
-
-    @app.websocket("/v1/stream")
-    async def _proxy(websocket: WebSocket) -> None:
-        await websocket.accept()
-        replica = state.registry.select()
-        if replica is None:
-            await websocket.send_json({
-                "type": "error",
-                "code": "gpu_unavailable",
-                "message": "router: no healthy replica available",
-                "retryable": True,
-            })
-            await websocket.close(code=1013, reason="no_healthy_replica")
-            return
-
-        ws_url = _websocket_url_for(replica.url)
-        try:
-            await _bridge_session(websocket, ws_url)
-        except WebSocketDisconnect:
-            logger.info("router: client disconnected")
-        except Exception as exc:
-            logger.exception("router: bridge failed: %s", exc)
-            with contextlib.suppress(RuntimeError):
-                await websocket.send_json({
-                    "type": "error",
-                    "code": "worker_failed",
-                    "message": f"router bridge failed: {exc}",
-                    "retryable": True,
-                })
-            with contextlib.suppress(RuntimeError):
-                await websocket.close(code=1011)
-
-    app.state.router_state = state
-    return app
-
-
-def run_router(config: RouterConfig) -> None:  # pragma: no cover - CLI
-    import uvicorn
-
-    app = build_router_app(config)
-    uvicorn.run(app, host=config.host, port=config.port)
-
-
-async def _bridge_session(
-    client_ws: WebSocket,
-    backend_ws_url: str,
-) -> None:
-    """Connect to backend and shuttle messages in both directions.
-
-    Uses ``websockets`` for the backend side; imported lazily to keep
-    the router's import graph small for users who only want the server.
-
-    Cancellation: when either direction completes (client disconnect,
-    backend close, exception), the other is cancelled explicitly and
-    both are drained before returning. Unexpected exceptions from the
-    direction that completed first are re-raised; normal disconnect
-    paths (``WebSocketDisconnect``, ``ConnectionClosed``,
-    ``CancelledError``) are swallowed.
-    """
-    try:
-        import websockets
-    except ImportError as exc:  # pragma: no cover - optional extra
-        raise RuntimeError("router requires the `websockets` package for backend proxying") from exc
-
-    async with websockets.connect(backend_ws_url + "/v1/stream") as backend_ws:
-        c2b = asyncio.create_task(_forward_client_to_backend(client_ws, backend_ws))
-        b2c = asyncio.create_task(_forward_backend_to_client(backend_ws, client_ws))
-        try:
-            done, _pending = await asyncio.wait(
-                {c2b, b2c},
-                return_when=asyncio.FIRST_COMPLETED,
-            )
-        finally:
-            for task in (c2b, b2c):
-                if not task.done():
-                    task.cancel()
-            await asyncio.gather(c2b, b2c, return_exceptions=True)
-        for task in done:
-            task_exc = task.exception()
-            if task_exc is not None and not _is_normal_disconnect(task_exc):
-                raise task_exc
-
-
-def _is_normal_disconnect(exc: BaseException) -> bool:
-    """Whether ``exc`` is a routine WebSocket teardown vs a real bridge fault."""
-    if isinstance(exc, asyncio.CancelledError | WebSocketDisconnect):
-        return True
-    name = type(exc).__name__
-    # websockets.exceptions.ConnectionClosed{,OK,Error} all subclass
-    # WebSocketException; check by name to avoid the lazy-import dance.
-    return name.startswith("ConnectionClosed")
-
-
-async def _forward_client_to_backend(client_ws: WebSocket, backend_ws) -> None:
-    try:
-        while True:
-            msg = await client_ws.receive()
-            if msg.get("type") == "websocket.disconnect":
-                break
-            if "text" in msg and msg["text"] is not None:
-                await backend_ws.send(msg["text"])
-            elif "bytes" in msg and msg["bytes"] is not None:
-                await backend_ws.send(msg["bytes"])
-    finally:
-        with contextlib.suppress(Exception):
-            await backend_ws.close()
-
-
-async def _forward_backend_to_client(backend_ws, client_ws: WebSocket) -> None:
-    try:
-        async for frame in backend_ws:
-            if isinstance(frame, bytes):
-                await client_ws.send_bytes(frame)
-            else:
-                await client_ws.send_text(frame)
-    finally:
-        with contextlib.suppress(Exception):
-            await client_ws.close()
-
-
-def _websocket_url_for(http_url: str) -> str:
-    if http_url.startswith("https://"):
-        return "wss://" + http_url[len("https://"):]
-    if http_url.startswith("http://"):
-        return "ws://" + http_url[len("http://"):]
-    return http_url
-
-
-__all__ = [
-    "build_router_app",
-    "run_router",
-]

fastvideo/entrypoints/streaming/router/registry.py

@@ -1,268 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Replica registry + health-check loop.
-
-The registry tracks the set of known backend replicas and their live
-health. The router consults it for "pick a backend for this session"
-decisions and a background task updates it from periodic HTTP probes.
-
-State machine per replica::
-
-    HEALTHY ──(N consecutive failures)──▶ UNHEALTHY
-       ▲                                     │
-       └──────(M consecutive successes)──────┘
-
-Where N = :attr:`RouterConfig.failure_threshold` and
-M = :attr:`RouterConfig.recovery_threshold`.
-"""
-from __future__ import annotations
-
-import asyncio
-import contextlib
-import enum
-import time
-from collections.abc import AsyncIterator, Awaitable, Callable
-from dataclasses import dataclass, field
-from typing import Any
-
-from fastvideo.entrypoints.streaming.router.config import (
-    ReplicaEndpoint,
-    RouterConfig,
-)
-from fastvideo.logger import init_logger
-
-HttpProbe = Any
-"""Structural alias for health-probe callables. Concrete signature is
-``async def __call__(url: str, *, timeout: float) -> tuple[float,
-str | None]``; typing.Callable cannot express keyword-only parameters,
-so duck-typing is the pragmatic compromise."""
-
-logger = init_logger(__name__)
-
-
-class ReplicaStatus(enum.Enum):
-    UNKNOWN = "unknown"
-    HEALTHY = "healthy"
-    UNHEALTHY = "unhealthy"
-
-
-@dataclass
-class ReplicaHealth:
-    status: ReplicaStatus = ReplicaStatus.UNKNOWN
-    last_ok_at: float | None = None
-    last_failure_at: float | None = None
-    consecutive_failures: int = 0
-    consecutive_successes: int = 0
-    last_latency_ms: float | None = None
-
-
-@dataclass
-class Replica:
-    endpoint: ReplicaEndpoint
-    health: ReplicaHealth = field(default_factory=ReplicaHealth)
-
-    @property
-    def url(self) -> str:
-        return self.endpoint.url
-
-    @property
-    def primary(self) -> bool:
-        return self.endpoint.primary
-
-    @property
-    def is_healthy(self) -> bool:
-        return self.health.status is ReplicaStatus.HEALTHY
-
-
-class ReplicaRegistry:
-    """Stateful map of replica URL → :class:`Replica`.
-
-    Selection favors primary replicas when healthy; otherwise the first
-    healthy non-primary is returned. When none are healthy, the
-    registry returns ``None`` so the router can reject incoming
-    sessions with ``gpu_unavailable``.
-    """
-
-    def __init__(self, replicas: list[ReplicaEndpoint]) -> None:
-        if not replicas:
-            raise ValueError("ReplicaRegistry requires at least one replica")
-        self._replicas: dict[str, Replica] = {endpoint.url: Replica(endpoint=endpoint) for endpoint in replicas}
-        self._lock = asyncio.Lock()
-
-    def all(self) -> list[Replica]:
-        return list(self._replicas.values())
-
-    def get(self, url: str) -> Replica | None:
-        return self._replicas.get(url)
-
-    def primaries(self) -> list[Replica]:
-        return [r for r in self._replicas.values() if r.primary]
-
-    def select(self) -> Replica | None:
-        """Pick the best healthy replica.
-
-        Priority order:
-
-        1. The first healthy primary (insertion order).
-        2. The first healthy non-primary (insertion order).
-        3. ``None`` when nothing is healthy.
-
-        This MVP picks the first match within each tier; it does NOT
-        load-balance across multiple healthy replicas of the same tier.
-        Round-robin and weighted distribution are deferred until a real
-        N-way active deployment exists.
-        """
-        healthy_primaries = [r for r in self._replicas.values() if r.primary and r.is_healthy]
-        if healthy_primaries:
-            return healthy_primaries[0]
-        healthy = [r for r in self._replicas.values() if r.is_healthy]
-        if healthy:
-            return healthy[0]
-        return None
-
-    async def record_success(
-        self,
-        replica: Replica,
-        *,
-        recovery_threshold: int,
-        latency_ms: float,
-    ) -> None:
-        async with self._lock:
-            h = replica.health
-            h.last_ok_at = time.time()
-            h.last_latency_ms = latency_ms
-            h.consecutive_failures = 0
-            h.consecutive_successes += 1
-            # State machine: UNKNOWN -> HEALTHY is immediate; only the
-            # UNHEALTHY -> HEALTHY transition is gated by recovery_threshold.
-            if h.status is ReplicaStatus.UNKNOWN:
-                logger.info("router: replica %s initial probe ok, marking HEALTHY", replica.url)
-                h.status = ReplicaStatus.HEALTHY
-                h.consecutive_successes = 0
-            elif (h.status is ReplicaStatus.UNHEALTHY and h.consecutive_successes >= recovery_threshold):
-                logger.info("router: replica %s recovered to HEALTHY after %d successes", replica.url,
-                            h.consecutive_successes)
-                h.status = ReplicaStatus.HEALTHY
-                h.consecutive_successes = 0
-
-    async def record_failure(
-        self,
-        replica: Replica,
-        *,
-        failure_threshold: int,
-        reason: str,
-    ) -> None:
-        async with self._lock:
-            h = replica.health
-            h.last_failure_at = time.time()
-            h.consecutive_successes = 0
-            h.consecutive_failures += 1
-            if (h.status is not ReplicaStatus.UNHEALTHY and h.consecutive_failures >= failure_threshold):
-                logger.warning("router: replica %s marked UNHEALTHY after %d failures: %s", replica.url,
-                               h.consecutive_failures, reason)
-                h.status = ReplicaStatus.UNHEALTHY
-
-
-async def run_health_check_loop(
-    registry: ReplicaRegistry,
-    config: RouterConfig,
-    *,
-    stop_event: asyncio.Event,
-    http_get: HttpProbe | None = None,
-) -> None:
-    """Poll all replicas' health endpoints in parallel on a fixed interval.
-
-    ``http_get`` is pluggable so unit tests can inject a deterministic
-    probe without hitting the network. The default builds a single
-    ``httpx.AsyncClient`` shared across the loop's lifetime so the
-    common case (steady polling against a stable replica set) reuses
-    TCP/TLS connections instead of paying handshake cost per probe.
-
-    Probes within one polling cycle run concurrently via ``asyncio.gather``
-    so a slow replica doesn't push the cycle past
-    ``health_check_interval_seconds``.
-    """
-    if http_get is not None:
-        await _run_loop(registry, config, stop_event, http_get)
-        return
-    async with _build_default_probe(config) as probe:
-        await _run_loop(registry, config, stop_event, probe)
-
-
-async def _run_loop(
-    registry: ReplicaRegistry,
-    config: RouterConfig,
-    stop_event: asyncio.Event,
-    http_get: Callable[..., Awaitable[tuple[float, str | None]]],
-) -> None:
-    while not stop_event.is_set():
-        replicas = registry.all()
-        results = await asyncio.gather(
-            *[
-                http_get(replica.url + config.health_check_path, timeout=config.health_check_timeout_seconds)
-                for replica in replicas
-            ],
-            return_exceptions=True,
-        )
-        for replica, result in zip(replicas, results, strict=True):
-            if isinstance(result, BaseException):
-                await registry.record_failure(
-                    replica,
-                    failure_threshold=config.failure_threshold,
-                    reason=f"{type(result).__name__}: {result}",
-                )
-                continue
-            status_ms, error = result
-            if error is None:
-                await registry.record_success(
-                    replica,
-                    recovery_threshold=config.recovery_threshold,
-                    latency_ms=status_ms,
-                )
-            else:
-                await registry.record_failure(
-                    replica,
-                    failure_threshold=config.failure_threshold,
-                    reason=error,
-                )
-        try:
-            await asyncio.wait_for(
-                stop_event.wait(),
-                timeout=config.health_check_interval_seconds,
-            )
-        except asyncio.TimeoutError:
-            continue
-
-
-@contextlib.asynccontextmanager
-async def _build_default_probe(
-    config: RouterConfig, ) -> AsyncIterator[Callable[..., Awaitable[tuple[float, str | None]]]]:
-    try:
-        import httpx
-    except ImportError as exc:  # pragma: no cover - optional extra
-        raise RuntimeError("router health checks require httpx; install with "
-                           "`pip install fastvideo[streaming]` or `pip install httpx`") from exc
-
-    async with httpx.AsyncClient(timeout=config.health_check_timeout_seconds) as client:
-
-        async def probe(url: str, *, timeout: float) -> tuple[float, str | None]:
-            start = time.perf_counter()
-            try:
-                response = await client.get(url, timeout=timeout)
-            except Exception as exc:
-                return 0.0, f"{type(exc).__name__}: {exc}"
-            latency_ms = (time.perf_counter() - start) * 1000.0
-            if response.status_code >= 400:
-                return latency_ms, f"HTTP {response.status_code}"
-            return latency_ms, None
-
-        yield probe
-
-
-__all__ = [
-    "HttpProbe",
-    "Replica",
-    "ReplicaHealth",
-    "ReplicaRegistry",
-    "ReplicaStatus",
-    "run_health_check_loop",
-]

fastvideo/entrypoints/streaming/server.py

@@ -1,560 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
-"""Single-generator FastAPI + WebSocket streaming server."""
 from __future__ import annotations
 
-import asyncio
-import contextlib
-import os
-import time
-from dataclasses import dataclass
-from typing import Any, Protocol
-
-from fastapi import FastAPI, WebSocket, WebSocketDisconnect
-from fastapi.responses import JSONResponse
-
-from fastvideo.api.schema import (
-    ContinuationState,
-    GenerationRequest,
-    InputConfig,
-    OutputConfig,
-    SamplingConfig,
-    ServeConfig,
-)
-from fastvideo.entrypoints.streaming.protocol import (
-    AutoExtensionUpdated,
-    ContinuationStateSnapshot,
-    EnhancementUpdated,
-    ErrorMessage,
-    GenerationPausedUpdated,
-    GpuAssigned,
-    LoopGenerationUpdated,
-    Ltx2SegmentComplete,
-    Ltx2SegmentStart,
-    Ltx2StreamComplete,
-    Ltx2StreamStart,
-    MediaInit,
-    MediaSegmentComplete,
-    QueueStatus,
-    SeedPromptsUpdated,
-    SegmentPromptSource,
-    SessionInitV2,
-    SnapshotState,
-    StepComplete,
-    parse_client_message,
-)
-from fastvideo.entrypoints.streaming.session import (
-    InvalidSessionTransition,
-    Session,
-    SessionManager,
-    SessionRejected,
-    SessionState,
-)
-from fastvideo.entrypoints.streaming.session_init_image import (
-    persist_session_init_image, )
-from fastvideo.entrypoints.streaming.gpu_pool import (
-    GpuPool,
-    InProcessGpuPool,
-    PoolAcquireTimeout,
-)
-from fastvideo.entrypoints.streaming.session_store import (
-    InMemorySessionStore,
-    SessionStore,
-)
-from fastvideo.entrypoints.streaming.stream import FragmentedMP4Encoder
+from fastvideo.api.schema import ServeConfig
 from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
 
-# RFC 6455 WebSocket close codes used by the server.
-_WS_CLOSE_UNSUPPORTED_DATA = 1003
-_WS_CLOSE_TRY_AGAIN_LATER = 1013
 
-
-class _GeneratorProto(Protocol):
-    """Subset of :class:`fastvideo.VideoGenerator` the server calls."""
-
-    def generate(self, request: GenerationRequest) -> Any:
-        ...
-
-
-@dataclass
-class ServerState:
-    serve_config: ServeConfig
-    pool: GpuPool
-    sessions: SessionManager
-    session_store: SessionStore
-
-
-def build_app(
-    serve_config: ServeConfig,
-    generator: _GeneratorProto | None = None,
-    *,
-    pool: GpuPool | None = None,
-    session_store: SessionStore | None = None,
-) -> FastAPI:
-    """Build the FastAPI app used by :func:`run_server`.
-
-    Exposed so tests can drive the WebSocket endpoint in-process via
-    ``starlette.testclient.TestClient(app).websocket_connect(...)``.
-
-    Exactly one of ``generator`` (backed by :class:`InProcessGpuPool`)
-    or ``pool`` (for the subprocess-backed production shape) must be
-    given.
-    """
-    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.")
-    if (generator is None) == (pool is None):
-        raise ValueError("build_app requires exactly one of `generator` or `pool`")
-
-    store = session_store or InMemorySessionStore()
-    if pool is None:
-        assert generator is not None
-        pool = InProcessGpuPool(generator, session_store=store)
-
-    sessions = SessionManager(
-        segment_cap=serve_config.streaming.generation_segment_cap,
-        session_timeout_seconds=serve_config.streaming.session_timeout_seconds,
-    )
-    state = ServerState(
-        serve_config=serve_config,
-        pool=pool,
-        sessions=sessions,
-        session_store=store,
-    )
-
-    app = FastAPI(title="FastVideo Streaming")
-
-    @app.get("/health")
-    async def _health() -> JSONResponse:
-        return JSONResponse({
-            "status": "ok",
-            "sessions": len(state.sessions),
-            "stream_mode": state.serve_config.streaming.stream_mode,
-        })
-
-    @app.websocket("/v1/stream")
-    async def _stream(websocket: WebSocket) -> None:
-        await websocket.accept()
-        try:
-            session = state.sessions.create()
-        except SessionRejected as exc:
-            await _send_error(websocket, "session_rejected", str(exc), retryable=False)
-            await websocket.close(code=_WS_CLOSE_TRY_AGAIN_LATER, reason="session_rejected")
-            return
-
-        try:
-            await _handle_session(websocket, session, state)
-        except WebSocketDisconnect:
-            logger.info("session %s: client disconnected", session.id[:8])
-        except Exception:  # pragma: no cover - defensive catch-all
-            logger.exception("session %s: unhandled error", session.id[:8])
-            with contextlib.suppress(InvalidSessionTransition):
-                session.transition(SessionState.ERROR)
-        finally:
-            with contextlib.suppress(Exception):
-                await state.pool.release(session.id)
-            _cleanup_session(session, state)
-
-    app.state.server_state = state
-    return app
-
-
-def run_server(serve_config: ServeConfig, *, generator: _GeneratorProto | None = None) -> None:
-    """Launch the streaming server.
-
-    Boots a :class:`fastvideo.VideoGenerator` from
-    ``serve_config.generator`` unless ``generator`` is provided, then
-    serves ``build_app(...)`` via uvicorn.
-    """
+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.")
-
-    import uvicorn
-
-    if generator is None:
-        from fastvideo import VideoGenerator  # lazy to avoid boot cost
-
-        generator = VideoGenerator.from_pretrained(config=serve_config.generator)
-    app = build_app(serve_config, generator)
-    uvicorn.run(
-        app,
-        host=serve_config.server.host,
-        port=serve_config.server.port,
-    )
-
-
-async def _handle_session(
-    websocket: WebSocket,
-    session: Session,
-    state: ServerState,
-) -> None:
-    init = await _read_init_message(websocket, session, state)
-    if init is None:
-        return
-
-    await _apply_session_init(session, init, state)
-    await _send_json(websocket, QueueStatus(position=0, queue_depth=0))
-    session.transition(SessionState.GPU_BINDING)
-    try:
-        assignment = await state.pool.acquire(
-            session.id,
-            timeout=float(state.sessions.session_timeout_seconds),
-        )
-    except PoolAcquireTimeout as exc:
-        await _send_error(websocket, "gpu_unavailable", str(exc), retryable=True)
-        with contextlib.suppress(InvalidSessionTransition):
-            session.transition(SessionState.TIMEOUT)
-        return
-    session.gpu_id = assignment.gpu_id
-    await _send_json(websocket,
-                     GpuAssigned(
-                         gpu_id=assignment.gpu_id,
-                         session_timeout=state.sessions.session_timeout_seconds,
-                     ))
-    session.transition(SessionState.ACTIVE)
-    await _send_json(websocket, _build_stream_start(session, state))
-
-    try:
-        await _run_segment_loop(websocket, session, state)
-    finally:
-        with contextlib.suppress(RuntimeError):
-            await _send_json(websocket, Ltx2StreamComplete(reason="stop_requested"))
-
-
-async def _read_init_message(
-    websocket: WebSocket,
-    session: Session,
-    state: ServerState,
-) -> SessionInitV2 | None:
-    try:
-        raw = await asyncio.wait_for(
-            websocket.receive_json(),
-            timeout=state.sessions.session_timeout_seconds,
-        )
-    except asyncio.TimeoutError:
-        logger.info("session %s: init timeout", session.id[:8])
-        with contextlib.suppress(InvalidSessionTransition):
-            session.transition(SessionState.TIMEOUT)
-        return None
-    except WebSocketDisconnect:
-        return None
-    try:
-        parsed = parse_client_message(raw)
-    except Exception as exc:
-        await _reject_init(websocket, session, f"opening frame failed validation: {exc}", "invalid_init")
-        return None
-    if not isinstance(parsed, SessionInitV2):
-        await _reject_init(websocket, session, "first frame must be session_init_v2", "expected_session_init_v2")
-        return None
-    return parsed
-
-
-async def _reject_init(
-    websocket: WebSocket,
-    session: Session,
-    message: str,
-    close_reason: str,
-) -> None:
-    await _send_error(websocket, "invalid_message", message, retryable=False)
-    await websocket.close(code=_WS_CLOSE_UNSUPPORTED_DATA, reason=close_reason)
-    with contextlib.suppress(InvalidSessionTransition):
-        session.transition(SessionState.REJECTED)
-
-
-async def _apply_session_init(
-    session: Session,
-    init: SessionInitV2,
-    state: ServerState,
-) -> None:
-    session.client_id = init.client_id
-    session.preset = init.preset
-    session.preset_label = init.preset_label
-    session.curated_prompts = list(init.curated_prompts)
-    session.enhancement_enabled = init.enhancement_enabled
-    session.auto_extension_enabled = init.auto_extension_enabled
-    session.loop_generation_enabled = init.loop_generation_enabled
-    session.single_clip_mode = init.single_clip_mode
-    session.stream_mode = init.stream_mode
-
-    if init.initial_image is not None:
-        # Decode + disk write off the event loop; payload is up to 32 MiB.
-        image = await asyncio.to_thread(persist_session_init_image, init.initial_image)
-        if image is not None:
-            session.metadata["session_init_image"] = image.path
-
-    if init.continuation_state is not None:
-        session.continuation_state = _coerce_state(init.continuation_state)
-        if session.continuation_state is not None:
-            state.session_store.store(session.id, session.continuation_state)
-
-
-async def _run_segment_loop(
-    websocket: WebSocket,
-    session: Session,
-    state: ServerState,
-) -> None:
-    cap = state.sessions.segment_cap
-    while True:
-        if session.segment_cap_reached(cap):
-            logger.info("session %s: segment cap (%d) reached", session.id[:8], cap)
-            return
-
-        try:
-            raw = await asyncio.wait_for(
-                websocket.receive_json(),
-                timeout=state.sessions.session_timeout_seconds,
-            )
-        except asyncio.TimeoutError:
-            logger.info("session %s: idle timeout", session.id[:8])
-            with contextlib.suppress(InvalidSessionTransition):
-                session.transition(SessionState.TIMEOUT)
-            return
-        except WebSocketDisconnect:
-            return
-        session.touch()
-
-        try:
-            parsed = parse_client_message(raw)
-        except Exception as exc:
-            await _send_error(websocket, "invalid_message", str(exc), retryable=True)
-            continue
-
-        if isinstance(parsed, SnapshotState):
-            snap = state.session_store.snapshot(session.id)
-            if snap is None:
-                await _send_error(websocket,
-                                  "internal_error",
-                                  "no continuation state available for session",
-                                  retryable=False)
-                continue
-            await _send_json(websocket, ContinuationStateSnapshot(state={"kind": snap.kind, "payload": snap.payload}, ))
-            continue
-
-        if isinstance(parsed, SegmentPromptSource):
-            await _run_segment(websocket, session, state, parsed)
-            continue
-
-        # Silently ignore unknown-but-valid types (additive-evolution
-        # rule in streaming.md).
-        _apply_toggle(session, parsed)
-
-
-async def _run_segment(
-    websocket: WebSocket,
-    session: Session,
-    state: ServerState,
-    message: SegmentPromptSource,
-) -> None:
-    request = _build_generation_request(session, message, state)
-    segment_idx = session.segment_idx
-    await _send_json(
-        websocket,
-        Ltx2SegmentStart(
-            segment_idx=segment_idx,
-            prompt=message.prompt,
-            total_steps=request.sampling.num_inference_steps,
-        ))
-
-    start = time.perf_counter()
-    # TODO: pool.run() runs to completion even if the client disconnects
-    # mid-segment. Real cancellation needs the generate_async API.
-    try:
-        result = await state.pool.run(session.id, request)
-    except Exception as exc:
-        logger.exception("session %s: pool.run failed", session.id[:8])
-        await _send_error(websocket, "worker_failed", f"pool.run failed: {exc}", retryable=True)
-        with contextlib.suppress(InvalidSessionTransition):
-            session.transition(SessionState.ERROR)
-        return
-    elapsed_ms = (time.perf_counter() - start) * 1000.0
-
-    frames = _extract_frames(result)
-    if not frames:
-        await _send_error(websocket, "worker_failed", "generator returned no frames", retryable=True)
-        with contextlib.suppress(InvalidSessionTransition):
-            session.transition(SessionState.ERROR)
-        return
-
-    # Synchronous generator call has no per-step hook; emit one
-    # terminal StepComplete so observability wiring still sees the
-    # segment finish.
-    total = request.sampling.num_inference_steps
-    await _send_json(websocket, StepComplete(
-        segment_idx=segment_idx,
-        step=total,
-        total_steps=total,
-        stage="denoise",
-    ))
-
-    encoder = FragmentedMP4Encoder(
-        width=request.sampling.width,
-        height=request.sampling.height,
-        fps=request.sampling.fps,
-        segment_idx=segment_idx,
-    )
-    chunks_relayed = 0
-    async with encoder:
-        init_sent = False
-        async for chunk in encoder.encode(frames):
-            if chunk.kind == "init":
-                await _send_json(websocket, MediaInit(
-                    segment_idx=segment_idx,
-                    stream_id=chunk.stream_id,
-                ))
-                init_sent = True
-            await websocket.send_bytes(chunk.data)
-            if init_sent and chunk.kind == "media":
-                chunks_relayed += 1
-
-    await _send_json(
-        websocket,
-        MediaSegmentComplete(
-            segment_idx=segment_idx,
-            stream_id=encoder.stream_id,
-            chunks=chunks_relayed,
-            duration_ms=float(request.sampling.num_frames) / request.sampling.fps * 1000.0,
-        ))
-
-    new_state = _extract_state(result)
-    if new_state is not None:
-        session.continuation_state = new_state
-        state.session_store.store(session.id, new_state)
-
-    session.segment_idx += 1
-    with contextlib.suppress(InvalidSessionTransition):
-        session.transition(SessionState.ACTIVE)
-
-    await _send_json(
-        websocket,
-        Ltx2SegmentComplete(
-            segment_idx=segment_idx,
-            generation_time_ms=elapsed_ms,
-            e2e_latency_ms=elapsed_ms,
-        ))
-
-
-def _build_stream_start(
-    session: Session,
-    state: ServerState,
-) -> Ltx2StreamStart:
-    default = state.serve_config.default_request
-    return Ltx2StreamStart(
-        preset=session.preset,
-        width=default.sampling.width,
-        height=default.sampling.height,
-        fps=default.sampling.fps,
-        num_frames=default.sampling.num_frames,
-    )
-
-
-def _build_generation_request(
-    session: Session,
-    message: SegmentPromptSource,
-    state: ServerState,
-) -> GenerationRequest:
-    # Start from the operator-pinned default_request to pick up the
-    # preset-selected sampling knobs; override with per-message values.
-    base = state.serve_config.default_request
-    sampling_kwargs: dict[str, Any] = {
-        "num_videos_per_prompt":
-        base.sampling.num_videos_per_prompt,
-        "seed":
-        message.seed if message.seed is not None else base.sampling.seed,
-        "num_frames":
-        base.sampling.num_frames,
-        "height":
-        base.sampling.height,
-        "width":
-        base.sampling.width,
-        "fps":
-        base.sampling.fps,
-        "num_inference_steps":
-        (message.num_inference_steps if message.num_inference_steps is not None else base.sampling.num_inference_steps),
-        "guidance_scale":
-        (message.guidance_scale if message.guidance_scale is not None else base.sampling.guidance_scale),
-    }
-    request = GenerationRequest(
-        prompt=message.prompt,
-        negative_prompt=message.negative_prompt or base.negative_prompt,
-        inputs=InputConfig(image_path=session.metadata.get("session_init_image"), ),
-        sampling=SamplingConfig(**sampling_kwargs),
-        output=OutputConfig(save_video=False, return_frames=True, return_state=True),
-        state=session.continuation_state,
-    )
-    return request
-
-
-def _coerce_state(raw: dict[str, Any]) -> ContinuationState | None:
-    kind = raw.get("kind")
-    payload = raw.get("payload")
-    if not isinstance(kind, str) or not isinstance(payload, dict):
-        return None
-    return ContinuationState(kind=kind, payload=payload)
-
-
-def _apply_toggle(session: Session, message: Any) -> None:
-    if isinstance(message, EnhancementUpdated):
-        session.enhancement_enabled = message.enabled
-    elif isinstance(message, AutoExtensionUpdated):
-        session.auto_extension_enabled = message.enabled
-    elif isinstance(message, LoopGenerationUpdated):
-        session.loop_generation_enabled = message.enabled
-    elif isinstance(message, GenerationPausedUpdated):
-        session.generation_paused = message.paused
-    elif isinstance(message, SeedPromptsUpdated):
-        session.curated_prompts = list(message.seed_prompts)
-
-
-def _extract_frames(result: Any) -> list:
-    if hasattr(result, "frames"):
-        return list(result.frames or [])
-    if isinstance(result, dict):
-        return list(result.get("frames") or [])
-    return []
-
-
-def _extract_state(result: Any) -> ContinuationState | None:
-    state = getattr(result, "state", None)
-    if state is None and isinstance(result, dict):
-        state = result.get("state")
-    if isinstance(state, ContinuationState):
-        return state
-    if isinstance(state, dict):
-        return _coerce_state(state)
-    return None
-
-
-async def _send_json(websocket: WebSocket, message: Any) -> None:
-    payload = (message.model_dump(mode="json", exclude_none=True) if hasattr(message, "model_dump") else message)
-    await websocket.send_json(payload)
-
-
-async def _send_error(
-    websocket: WebSocket,
-    code: str,
-    message: str,
-    *,
-    retryable: bool,
-) -> None:
-    await _send_json(
-        websocket,
-        ErrorMessage(code=code, message=message, retryable=retryable),
-    )
-
-
-def _cleanup_session(session: Session, state: ServerState) -> None:
-    state.sessions.close(session.id)
-    state.session_store.drop(session.id)
-    init_image_path = session.metadata.get("session_init_image")
-    if isinstance(init_image_path, str):
-        with contextlib.suppress(FileNotFoundError):
-            os.unlink(init_image_path)
-
-
-__all__ = [
-    "ServerState",
-    "build_app",
-    "run_server",
-]
+    raise NotImplementedError("streaming server is not implemented yet")

fastvideo/entrypoints/streaming/session.py

@@ -1,214 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Per-connection session lifecycle for the streaming server.
-
-Each WebSocket opens exactly one :class:`Session`. :class:`SessionManager`
-enforces the ``generation_segment_cap`` and ``session_timeout_seconds``
-budgets from :class:`fastvideo.api.StreamingConfig`.
-"""
-from __future__ import annotations
-
-import enum
-import time
-import uuid
-from dataclasses import dataclass, field
-from typing import Any
-
-from fastvideo.api.schema import ContinuationState
-
-
-class SessionState(enum.Enum):
-    """State-machine positions for a streaming session.
-
-    Transitions are server-owned. See
-    ``docs/design/server_contracts/streaming.md`` for the full diagram.
-    """
-
-    INITIALIZING = "initializing"
-    QUEUED = "queued"
-    GPU_BINDING = "gpu_binding"
-    ACTIVE = "active"
-    COMPLETE = "complete"
-    ERROR = "error"
-    TIMEOUT = "timeout"
-    REJECTED = "rejected"
-
-
-_VALID_TRANSITIONS: dict[SessionState, frozenset[SessionState]] = {
-    SessionState.INITIALIZING:
-    frozenset({
-        SessionState.QUEUED,
-        SessionState.GPU_BINDING,
-        SessionState.REJECTED,
-        SessionState.ERROR,
-    }),
-    SessionState.QUEUED:
-    frozenset({
-        SessionState.GPU_BINDING,
-        SessionState.ERROR,
-        SessionState.TIMEOUT,
-        SessionState.REJECTED,
-    }),
-    SessionState.GPU_BINDING:
-    frozenset({
-        SessionState.ACTIVE,
-        SessionState.ERROR,
-        SessionState.TIMEOUT,
-    }),
-    SessionState.ACTIVE:
-    frozenset({
-        SessionState.ACTIVE,
-        SessionState.COMPLETE,
-        SessionState.ERROR,
-        SessionState.TIMEOUT,
-    }),
-    SessionState.COMPLETE:
-    frozenset(),
-    SessionState.ERROR:
-    frozenset(),
-    SessionState.TIMEOUT:
-    frozenset(),
-    SessionState.REJECTED:
-    frozenset(),
-}
-
-
-class InvalidSessionTransition(RuntimeError):
-    """Raised when a session is asked to transition along an illegal edge."""
-
-
-@dataclass
-class Session:
-    id: str = field(default_factory=lambda: uuid.uuid4().hex)
-    state: SessionState = SessionState.INITIALIZING
-    created_at: float = field(default_factory=time.monotonic)
-    last_activity: float = field(default_factory=time.monotonic)
-
-    client_id: str | None = None
-    preset: str | None = None
-    preset_label: str | None = None
-
-    curated_prompts: list[str] = field(default_factory=list)
-
-    segment_idx: int = 0
-
-    enhancement_enabled: bool = False
-    auto_extension_enabled: bool = False
-    loop_generation_enabled: bool = False
-    single_clip_mode: bool = False
-    generation_paused: bool = False
-
-    stream_mode: str = "av_fmp4"
-    gpu_id: int | None = None
-
-    continuation_state: ContinuationState | None = None
-
-    metadata: dict[str, Any] = field(default_factory=dict)
-
-    def transition(self, target: SessionState) -> None:
-        """Move to ``target`` if the edge is allowed.
-
-        Raises :class:`InvalidSessionTransition` on illegal moves. The
-        self-loop on ``ACTIVE`` is legal so the server can re-assert
-        ACTIVE on segment completion without special casing.
-        """
-        allowed = _VALID_TRANSITIONS.get(self.state, frozenset())
-        if target not in allowed and target is not self.state:
-            raise InvalidSessionTransition(f"{self.state.value} -> {target.value} is not a valid "
-                                           f"session transition")
-        self.state = target
-        self.last_activity = time.monotonic()
-
-    def touch(self) -> None:
-        self.last_activity = time.monotonic()
-
-    def is_active(self) -> bool:
-        return self.state is SessionState.ACTIVE
-
-    def segment_cap_reached(self, cap: int) -> bool:
-        return self.segment_idx >= cap
-
-
-class SessionManager:
-    """Registers sessions and enforces per-server session limits."""
-
-    def __init__(
-        self,
-        *,
-        segment_cap: int,
-        session_timeout_seconds: int,
-        max_sessions: int = 1,
-    ) -> None:
-        self._segment_cap = segment_cap
-        self._session_timeout_seconds = session_timeout_seconds
-        self._max_sessions = max_sessions
-        self._sessions: dict[str, Session] = {}
-
-    @property
-    def segment_cap(self) -> int:
-        return self._segment_cap
-
-    @property
-    def session_timeout_seconds(self) -> int:
-        return self._session_timeout_seconds
-
-    def create(self) -> Session:
-        if len(self._sessions) >= self._max_sessions:
-            raise SessionRejected(f"max sessions reached ({self._max_sessions})")
-        session = Session()
-        self._sessions[session.id] = session
-        return session
-
-    def get(self, session_id: str) -> Session | None:
-        return self._sessions.get(session_id)
-
-    def close(self, session_id: str) -> None:
-        self._sessions.pop(session_id, None)
-
-    def __contains__(self, session_id: str) -> bool:
-        return session_id in self._sessions
-
-    def __len__(self) -> int:
-        return len(self._sessions)
-
-    def active_sessions(self) -> list[Session]:
-        return [s for s in self._sessions.values() if s.is_active()]
-
-    def reap_timed_out(self, now: float | None = None) -> list[str]:
-        """Return the ids of sessions that have exceeded the idle timeout.
-
-        The caller is responsible for actually closing them — this
-        method only *identifies* dead sessions so the server can emit
-        ``session_timeout`` frames before dropping the WebSocket.
-
-        TODO: unused until a background driver calls it. Per-connection
-        idle enforcement currently happens via asyncio.wait_for on
-        receive_json; this helper catches sessions stuck before any
-        receive (e.g. future QUEUED state) and is expected to be wired
-        into the GPU-pool reaper.
-        """
-        now = now if now is not None else time.monotonic()
-        dead: list[str] = []
-        for sid, session in self._sessions.items():
-            if session.state in {
-                    SessionState.COMPLETE,
-                    SessionState.ERROR,
-                    SessionState.TIMEOUT,
-                    SessionState.REJECTED,
-            }:
-                continue
-            if now - session.last_activity > self._session_timeout_seconds:
-                dead.append(sid)
-        return dead
-
-
-class SessionRejected(RuntimeError):
-    """Raised when session creation fails (queue full, auth, etc.)."""
-
-
-__all__ = [
-    "InvalidSessionTransition",
-    "Session",
-    "SessionManager",
-    "SessionRejected",
-    "SessionState",
-]

fastvideo/entrypoints/streaming/session_init_image.py

@@ -1,103 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Persist the initial-image blob attached to a streaming session."""
-from __future__ import annotations
-
-import base64
-import binascii
-import contextlib
-import os
-import tempfile
-from dataclasses import dataclass
-from typing import Any
-
-_ACCEPTED_MIMES = {
-    "image/png": ".png",
-    "image/jpeg": ".jpg",
-    "image/jpg": ".jpg",
-    "image/webp": ".webp",
-}
-
-_MAX_IMAGE_BYTES = 32 * 1024 * 1024  # 32 MiB cap
-
-
-@dataclass(frozen=True)
-class SessionInitImage:
-    """Location of the persisted init image.
-
-    Callers pass ``path`` to ``InputConfig.image_path``; ``display_name``
-    is only used for logs.
-    """
-
-    path: str
-    display_name: str
-    mime: str
-
-
-def persist_session_init_image(
-    payload: Any,
-    *,
-    output_dir: str | None = None,
-) -> SessionInitImage | None:
-    """Decode a client init-image blob and persist it to disk.
-
-    ``payload`` shape (matches the internal UI protocol)::
-
-        {
-            "mime": "image/png",
-            "name": "ref.png",
-            "data": "<base64 bytes>",
-        }
-
-    Returns ``None`` when ``payload`` is falsy (no init image). Raises
-    :class:`ValueError` on schema / size / decode errors so the caller
-    can surface a user-facing ``error`` frame.
-    """
-    if not payload:
-        return None
-    if not isinstance(payload, dict):
-        raise ValueError("session init image must be an object")
-
-    mime = payload.get("mime")
-    if mime not in _ACCEPTED_MIMES:
-        raise ValueError(f"session init image mime {mime!r} is not one of "
-                         f"{sorted(_ACCEPTED_MIMES)}")
-    data_b64 = payload.get("data")
-    if not isinstance(data_b64, str):
-        raise ValueError("session init image data must be a base64 string")
-    try:
-        data = base64.b64decode(data_b64, validate=True)
-    except (binascii.Error, ValueError) as exc:
-        raise ValueError(f"session init image data is not valid base64: {exc}") from exc
-    if len(data) > _MAX_IMAGE_BYTES:
-        raise ValueError(f"session init image is {len(data)} bytes; limit is "
-                         f"{_MAX_IMAGE_BYTES}")
-    if len(data) == 0:
-        raise ValueError("session init image data is empty")
-
-    ext = _ACCEPTED_MIMES[mime]
-    display_name = _sanitize_display_name(payload.get("name")) or f"init{ext}"
-    fd, path = tempfile.mkstemp(prefix="fastvideo-init-", suffix=ext, dir=output_dir)
-    try:
-        with os.fdopen(fd, "wb") as f:
-            f.write(data)
-    except Exception:
-        with contextlib.suppress(FileNotFoundError):
-            os.unlink(path)
-        raise
-    return SessionInitImage(path=path, display_name=display_name, mime=mime)
-
-
-def _sanitize_display_name(name: Any) -> str | None:
-    if not isinstance(name, str):
-        return None
-    name = name.strip()
-    if not name:
-        return None
-    # Strip any path components — we only keep the leaf for logging.
-    return os.path.basename(name)
-
-
-__all__ = [
-    "SessionInitImage",
-    "persist_session_init_image",
-]

fastvideo/entrypoints/streaming/session_logger.py

@@ -1,113 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Per-session JSONL event logger.
-
-Each session gets its own JSONL file under the configured log root so
-post-hoc analytics (enhancer latency, GPU assignment, segment timings)
-can be recovered without a tracing backend. The internal UI uses this
-format; keeping the same shape makes log tooling portable.
-"""
-from __future__ import annotations
-
-import contextlib
-import json
-import os
-import re
-import threading
-import time
-from dataclasses import dataclass, field
-from typing import Any, TextIO
-
-_FILENAME_SANITIZE_RE = re.compile(r"[^A-Za-z0-9._-]")
-
-
-@dataclass
-class SessionLogEvent:
-    """One line in the session JSONL file."""
-
-    session_id: str
-    event: str
-    payload: dict[str, Any] = field(default_factory=dict)
-    ts: float = field(default_factory=time.time)
-
-
-class SessionLogger:
-    """Append-only JSONL logger keyed by session id.
-
-    Thread-safe; the server may be writing from multiple asyncio tasks
-    (fMP4 encoder thread + control-frame handler) for the same session.
-    """
-
-    def __init__(self, log_dir: str | None) -> None:
-        self._log_dir = log_dir
-        self._files: dict[str, TextIO] = {}
-        self._locks: dict[str, threading.Lock] = {}
-        self._registry_lock = threading.Lock()
-        self._ensure_dir()
-
-    def log(self, event: SessionLogEvent) -> None:
-        if self._log_dir is None:
-            return
-        opened = self._get_file(event.session_id)
-        if opened is None:
-            return
-        handle, lock = opened
-        line = json.dumps({
-            "session_id": event.session_id,
-            "event": event.event,
-            "ts": event.ts,
-            "payload": event.payload,
-        })
-        with lock, contextlib.suppress(ValueError):
-            handle.write(line + "\n")
-            handle.flush()
-
-    def close(self, session_id: str) -> None:
-        with self._registry_lock:
-            handle = self._files.pop(session_id, None)
-            lock = self._locks.pop(session_id, None)
-        if handle is None or lock is None:
-            return
-        with lock, contextlib.suppress(Exception):
-            handle.close()
-
-    def close_all(self) -> None:
-        with self._registry_lock:
-            sids = list(self._files)
-        for sid in sids:
-            self.close(sid)
-
-    def _ensure_dir(self) -> None:
-        if self._log_dir is None:
-            return
-        os.makedirs(self._log_dir, exist_ok=True)
-
-    def _get_file(self, session_id: str) -> tuple[TextIO, threading.Lock] | None:
-        if self._log_dir is None:
-            return None
-        with self._registry_lock:
-            handle = self._files.get(session_id)
-            lock = self._locks.get(session_id)
-            if handle is not None and lock is not None:
-                return handle, lock
-            # Defense-in-depth: session_id is server-generated UUID today,
-            # but sanitize against path traversal in case future code paths
-            # allow client-supplied ids.
-            safe_id = _FILENAME_SANITIZE_RE.sub("_", session_id) or "unknown"
-            path = os.path.join(
-                self._log_dir,
-                f"session-{safe_id}.jsonl",
-            )
-            try:
-                handle = open(path, "a", encoding="utf-8")  # noqa: SIM115
-            except OSError:
-                return None
-            lock = threading.Lock()
-            self._files[session_id] = handle
-            self._locks[session_id] = lock
-            return handle, lock
-
-
-__all__ = [
-    "SessionLogEvent",
-    "SessionLogger",
-]

fastvideo/entrypoints/streaming/session_store.py

@@ -1,206 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Session state store for the FastVideo streaming server.
-
-The streaming server keeps continuation state (decoded frames + audio
-latents from the previous segment) server-side so the client doesn't
-re-upload multi-megabyte tensors each WebSocket message. Two operations
-are needed:
-
-* ``snapshot(session_id) -> ContinuationState`` — serialize the current
-  state so it can be exported (e.g. over HTTP) or migrated to a
-  different server.
-* ``hydrate(state) -> session_id`` — load a previously serialized state
-  into a new session (for resume-after-disconnect flows).
-
-The store is an ABC with an :class:`InMemorySessionStore` default; Redis
-or other backends can drop in without touching the pipeline.
-
-Large tensor payloads (video frames, audio latents) are kept out of the
-JSON payload via an accompanying :class:`BlobStore`. Both stores share a
-process today; they are separate types so that a future implementation
-can put blobs on S3 while keeping session metadata in Redis.
-"""
-from __future__ import annotations
-
-import threading
-import uuid
-from abc import ABC, abstractmethod
-from collections.abc import Iterator
-from dataclasses import dataclass
-
-from fastvideo.api.schema import ContinuationState
-
-
-class BlobStore(ABC):
-    """Opaque byte-blob storage keyed by id.
-
-    A :class:`ContinuationState` payload can reference large tensors
-    stored in a :class:`BlobStore` rather than inlining them, so the
-    JSON payload stays small when the state travels over the wire.
-    """
-
-    @abstractmethod
-    def put(self, data: bytes, *, mime: str = "application/octet-stream") -> str:
-        """Store ``data`` and return a blob id for later retrieval."""
-
-    @abstractmethod
-    def get(self, blob_id: str) -> bytes:
-        """Load a previously stored blob. Raises ``KeyError`` if absent."""
-
-    @abstractmethod
-    def drop(self, blob_id: str) -> None:
-        """Remove a blob. Missing ids are a no-op."""
-
-    @abstractmethod
-    def __contains__(self, blob_id: str) -> bool:
-        ...
-
-
-@dataclass(frozen=True)
-class _BlobRecord:
-    data: bytes
-    mime: str
-
-
-class InMemoryBlobStore(BlobStore):
-    """Thread-safe in-memory :class:`BlobStore` for single-process servers.
-
-    No eviction policy — callers are responsible for calling
-    :meth:`drop` when a blob's owning state is replaced or a session
-    ends. A redis- or filesystem-backed :class:`BlobStore` should
-    replace this when the streaming server lands as a real service
-    (PR 7.5+).
-    """
-
-    def __init__(self) -> None:
-        self._blobs: dict[str, _BlobRecord] = {}
-        self._lock = threading.Lock()
-
-    def put(self, data: bytes, *, mime: str = "application/octet-stream") -> str:
-        blob_id = uuid.uuid4().hex
-        with self._lock:
-            self._blobs[blob_id] = _BlobRecord(data=data, mime=mime)
-        return blob_id
-
-    def get(self, blob_id: str) -> bytes:
-        with self._lock:
-            record = self._blobs.get(blob_id)
-        if record is None:
-            raise KeyError(f"Unknown blob id: {blob_id}")
-        return record.data
-
-    def drop(self, blob_id: str) -> None:
-        with self._lock:
-            self._blobs.pop(blob_id, None)
-
-    def __contains__(self, blob_id: str) -> bool:
-        with self._lock:
-            return blob_id in self._blobs
-
-    def __len__(self) -> int:
-        with self._lock:
-            return len(self._blobs)
-
-
-class SessionStore(ABC):
-    """Keyed store for per-session continuation state.
-
-    Implementations own the session-id → state mapping. The streaming
-    server calls :meth:`store` after each segment and :meth:`snapshot`
-    when a client explicitly asks for an exportable state handle.
-    """
-
-    @abstractmethod
-    def store(self, session_id: str, state: ContinuationState) -> None:
-        """Persist ``state`` for ``session_id``, replacing any prior value."""
-
-    @abstractmethod
-    def snapshot(self, session_id: str) -> ContinuationState | None:
-        """Return the current state for ``session_id`` (or ``None``)."""
-
-    @abstractmethod
-    def hydrate(
-        self,
-        state: ContinuationState,
-        *,
-        session_id: str | None = None,
-    ) -> str:
-        """Install ``state`` as the starting point for a session.
-
-        When ``session_id`` is ``None`` the store allocates a fresh id
-        (UUID4); when provided the store uses it verbatim, overwriting
-        any prior state at that id.
-        """
-
-    @abstractmethod
-    def drop(self, session_id: str) -> None:
-        """Forget a session. Missing ids are a no-op."""
-
-    @abstractmethod
-    def __contains__(self, session_id: str) -> bool:
-        ...
-
-    @abstractmethod
-    def __iter__(self) -> Iterator[str]:
-        ...
-
-
-class InMemorySessionStore(SessionStore):
-    """Thread-safe in-memory :class:`SessionStore`.
-
-    Default implementation used by single-process deployments; a future
-    Redis-backed store can be dropped in without changes to the server.
-
-    No eviction / TTL / bounded capacity — sessions only leave via
-    :meth:`drop`. The live streaming server (PR 7.5+) is responsible
-    for bounding growth and for dropping any :class:`BlobStore` blobs
-    referenced by a state when that state is replaced or a session
-    ends; this class does not know about blobs.
-    """
-
-    def __init__(self) -> None:
-        self._sessions: dict[str, ContinuationState] = {}
-        self._lock = threading.Lock()
-
-    def store(self, session_id: str, state: ContinuationState) -> None:
-        with self._lock:
-            self._sessions[session_id] = state
-
-    def snapshot(self, session_id: str) -> ContinuationState | None:
-        with self._lock:
-            return self._sessions.get(session_id)
-
-    def hydrate(
-        self,
-        state: ContinuationState,
-        *,
-        session_id: str | None = None,
-    ) -> str:
-        sid = session_id or uuid.uuid4().hex
-        with self._lock:
-            self._sessions[sid] = state
-        return sid
-
-    def drop(self, session_id: str) -> None:
-        with self._lock:
-            self._sessions.pop(session_id, None)
-
-    def __contains__(self, session_id: str) -> bool:
-        with self._lock:
-            return session_id in self._sessions
-
-    def __iter__(self) -> Iterator[str]:
-        with self._lock:
-            return iter(list(self._sessions))
-
-    def __len__(self) -> int:
-        with self._lock:
-            return len(self._sessions)
-
-
-__all__ = [
-    "BlobStore",
-    "InMemoryBlobStore",
-    "InMemorySessionStore",
-    "SessionStore",
-]

fastvideo/entrypoints/streaming/stream.py

@@ -1,213 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""fMP4 stream encoder used by the streaming server.
-
-The client's Media Source Extensions player needs a continuous fMP4
-byte stream: first an *initialization segment* (``ftyp`` + ``moov``),
-then one or more *media segments* (``moof`` + ``mdat``). We pipe raw
-RGB frames into an ffmpeg subprocess configured for fragmented output
-via ``-movflags empty_moov+default_base_moof+frag_keyframe+faststart``
-and stream the bytes back out.
-"""
-from __future__ import annotations
-
-import asyncio
-import contextlib
-import subprocess
-import uuid
-from collections.abc import AsyncIterator
-from dataclasses import dataclass
-from typing import TYPE_CHECKING, Literal
-
-if TYPE_CHECKING:
-    import numpy as np
-
-
-@dataclass
-class FragmentedMP4Chunk:
-    """A single fMP4 byte chunk emitted by :class:`FragmentedMP4Encoder`.
-
-    ``kind`` identifies whether the chunk is the init segment (must be
-    fed into the client's ``SourceBuffer`` first) or a media fragment.
-    """
-
-    kind: Literal["init", "media"]
-    data: bytes
-    stream_id: str
-    segment_idx: int
-
-
-class FragmentedMP4Encoder:
-    """Stream RGB frames in, fMP4 chunks out.
-
-    One encoder covers one segment. The server creates a new encoder
-    per :class:`ltx2_segment_start`` boundary so each segment becomes
-    one media fragment the client can append independently.
-
-    Example::
-
-        encoder = FragmentedMP4Encoder(width=1024, height=576, fps=24,
-                                        segment_idx=0)
-        async with encoder:
-            async for chunk in encoder.encode(frames):
-                await websocket.send_bytes(chunk.data)
-    """
-
-    def __init__(
-        self,
-        *,
-        width: int,
-        height: int,
-        fps: int,
-        segment_idx: int,
-        stream_id: str | None = None,
-        ffmpeg_path: str = "ffmpeg",
-        preset: str = "ultrafast",
-        pixel_format_out: str = "yuv420p",
-        extra_args: list[str] | None = None,
-    ) -> None:
-        self.width = width
-        self.height = height
-        self.fps = fps
-        self.segment_idx = segment_idx
-        self.stream_id = stream_id or uuid.uuid4().hex
-        self._ffmpeg_path = ffmpeg_path
-        self._preset = preset
-        self._pixel_format_out = pixel_format_out
-        self._extra_args = list(extra_args or [])
-        self._proc: subprocess.Popen | None = None
-        self._init_emitted = False
-
-    async def __aenter__(self) -> FragmentedMP4Encoder:
-        self._spawn()
-        return self
-
-    async def __aexit__(self, exc_type, exc, tb) -> None:
-        await self.close()
-
-    def _spawn(self) -> None:
-        args = [
-            self._ffmpeg_path,
-            "-hide_banner",
-            "-loglevel",
-            "error",
-            "-f",
-            "rawvideo",
-            "-pix_fmt",
-            "rgb24",
-            "-s",
-            f"{self.width}x{self.height}",
-            "-r",
-            str(self.fps),
-            "-i",
-            "-",
-            "-c:v",
-            "libx264",
-            "-preset",
-            self._preset,
-            "-tune",
-            "zerolatency",
-            "-pix_fmt",
-            self._pixel_format_out,
-            "-movflags",
-            "empty_moov+default_base_moof+frag_keyframe+faststart",
-            "-f",
-            "mp4",
-            *self._extra_args,
-            "-",
-        ]
-        # stderr → DEVNULL: with -loglevel error on, the only thing
-        # stderr would carry is unsolicited warnings. Piping without a
-        # reader deadlocks ffmpeg once the pipe buffer (~64 KiB) fills.
-        self._proc = subprocess.Popen(  # noqa: S603
-            args,
-            stdin=subprocess.PIPE,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.DEVNULL,
-            bufsize=0,
-        )
-
-    async def encode(
-        self,
-        frames: list[np.ndarray] | AsyncIterator[np.ndarray],
-    ) -> AsyncIterator[FragmentedMP4Chunk]:
-        """Feed frames into ffmpeg and yield fMP4 chunks as they appear."""
-        if self._proc is None:
-            self._spawn()
-        assert self._proc is not None and self._proc.stdin is not None
-        proc = self._proc
-
-        loop = asyncio.get_running_loop()
-
-        async def _writer() -> None:
-            try:
-                if hasattr(frames, "__aiter__"):
-                    async for frame in frames:  # type: ignore[union-attr]
-                        await loop.run_in_executor(None, _write_frame, proc.stdin, frame)
-                else:
-                    for frame in frames:  # type: ignore[assignment]
-                        await loop.run_in_executor(None, _write_frame, proc.stdin, frame)
-            finally:
-                with contextlib.suppress(BrokenPipeError):
-                    proc.stdin.close()
-
-        writer_task = asyncio.create_task(_writer())
-        try:
-            reader = proc.stdout
-            assert reader is not None
-            # Read in reasonably-sized chunks; MSE tolerates any size
-            # but we don't want to starve the event loop.
-            chunk_size = 64 * 1024
-            while True:
-                data = await loop.run_in_executor(None, reader.read, chunk_size)
-                if not data:
-                    break
-                kind: Literal["init", "media"] = "init" if not self._init_emitted else "media"
-                self._init_emitted = True
-                yield FragmentedMP4Chunk(
-                    kind=kind,
-                    data=bytes(data),
-                    stream_id=self.stream_id,
-                    segment_idx=self.segment_idx,
-                )
-        finally:
-            await writer_task
-
-    async def close(self) -> None:
-        if self._proc is None:
-            return
-        proc = self._proc
-        self._proc = None
-        try:
-            if proc.stdin and not proc.stdin.closed:
-                proc.stdin.close()
-        except BrokenPipeError:
-            pass
-        loop = asyncio.get_running_loop()
-        try:
-            await asyncio.wait_for(
-                loop.run_in_executor(None, proc.wait),
-                timeout=5.0,
-            )
-        except asyncio.TimeoutError:
-            proc.kill()
-            await loop.run_in_executor(None, proc.wait)
-
-
-def _write_frame(stdin, frame: np.ndarray) -> None:
-    import numpy as np
-
-    if not isinstance(frame, np.ndarray):
-        raise TypeError("fMP4 encoder frames must be numpy.ndarray")
-    if frame.dtype != np.uint8:
-        frame = frame.astype(np.uint8)
-    if frame.ndim != 3 or frame.shape[-1] != 3:
-        raise ValueError("fMP4 encoder frames must be HxWx3 uint8 RGB; got "
-                         f"shape={frame.shape}, dtype={frame.dtype}")
-    with contextlib.suppress(BrokenPipeError):
-        stdin.write(frame.tobytes())
-
-
-__all__ = [
-    "FragmentedMP4Chunk",
-    "FragmentedMP4Encoder",
-]

fastvideo/entrypoints/streaming/worker.py

@@ -1,133 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Per-GPU worker subprocess entry for :class:`SubprocessGpuPool`.
-
-The pool manages binding, lifecycle, and message dispatch in the parent
-process. The worker constructs its :class:`VideoGenerator` from a typed
-:class:`GeneratorConfig`, runs the two-segment warmup so both
-initial-segment and continuation-branch compile graphs are hot, and
-then loops on the job queue.
-"""
-from __future__ import annotations
-
-import multiprocessing as mp
-import queue
-from typing import Any
-
-from fastvideo.api.schema import (
-    GeneratorConfig,
-    GenerationRequest,
-    InputConfig,
-    OutputConfig,
-    SamplingConfig,
-    WarmupConfig,
-)
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
-
-# Synthetic warmup dimensions: small enough to keep boot fast, big enough
-# to exercise the real shape-dependent compile paths. Keep in sync with
-# WarmupConfig if these become user-tunable.
-_WARMUP_NUM_FRAMES = 8
-_WARMUP_HEIGHT = 256
-_WARMUP_WIDTH = 256
-_WARMUP_NUM_INFERENCE_STEPS = 1
-
-
-def worker_main(
-    *,
-    gpu_id: int,
-    worker_id: str,
-    generator_config: GeneratorConfig,
-    warmup_config: WarmupConfig,
-    job_queue: mp.Queue,
-    result_queue: mp.Queue,
-    shutdown_event: Any,
-) -> None:  # pragma: no cover - exercised via integration only
-    """Per-worker subprocess entry.
-
-    Runs inside the child spawned by ``SubprocessGpuPool``. Blocking
-    ``VideoGenerator`` construction + generation happens here, not in
-    the parent's event loop.
-    """
-    import os
-
-    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
-    try:
-        from fastvideo import VideoGenerator
-
-        generator = VideoGenerator.from_pretrained(config=generator_config)
-        if warmup_config.enabled:
-            _warmup_worker(generator, warmup_config)
-        result_queue.put({"kind": "ready", "worker_id": worker_id})
-    except Exception as exc:
-        result_queue.put({"kind": "error", "error": repr(exc)})
-        return
-
-    while not shutdown_event.is_set():
-        try:
-            item = job_queue.get(timeout=0.5)
-        except queue.Empty:
-            continue
-        if item is None:
-            break
-        job_id = item["job_id"]
-        request = item["request"]
-        try:
-            result = generator.generate(request)
-            result_queue.put({
-                "kind": "result",
-                "job_id": job_id,
-                "result": result,
-            })
-        except Exception as exc:
-            result_queue.put({
-                "kind": "error",
-                "job_id": job_id,
-                "error": repr(exc),
-            })
-
-
-def _warmup_worker(
-    generator: Any,
-    warmup_config: WarmupConfig,
-) -> None:
-    """Run two synthetic generations so both compile branches are primed.
-
-    Segment 1 is a fresh start (no continuation state) and exercises
-    the initial-segment graph. Segment 2 feeds segment 1's continuation
-    state back in so the conditioning branch is also compiled before
-    the first user request lands.
-    """
-    sampling = SamplingConfig(
-        num_frames=_WARMUP_NUM_FRAMES,
-        height=_WARMUP_HEIGHT,
-        width=_WARMUP_WIDTH,
-        num_inference_steps=_WARMUP_NUM_INFERENCE_STEPS,
-    )
-    seg1 = GenerationRequest(
-        prompt=warmup_config.prompt,
-        sampling=sampling,
-        inputs=InputConfig(),
-        output=OutputConfig(save_video=False, return_frames=False, return_state=True),
-    )
-    seg1_result = generator.generate(seg1)
-
-    seg2 = GenerationRequest(
-        prompt=warmup_config.prompt,
-        sampling=sampling,
-        inputs=InputConfig(),
-        output=OutputConfig(save_video=False, return_frames=False),
-        state=_extract_continuation_state(seg1_result),
-    )
-    generator.generate(seg2)
-
-
-def _extract_continuation_state(result: Any) -> Any:
-    state = getattr(result, "state", None)
-    if state is None and isinstance(result, dict):
-        state = result.get("state")
-    return state
-
-
-__all__ = ["worker_main"]

fastvideo/entrypoints/video_generator.py

@@ -65,7 +65,6 @@ _FROM_PRETRAINED_CONVENIENCE_KWARGS = frozenset({
     "pin_cpu_memory",
     "enable_torch_compile",
     "torch_compile_kwargs",
-    "output_type",
 })
 
 
@@ -602,20 +601,10 @@ class VideoGenerator:
         thread = threading.Thread(target=execute_forward_thread)
         thread.start()
         latent_batch_size = _infer_latent_batch_size(batch)
-        # When ``output_type == "latent"`` the forward output has latent
-        # shape (e.g. ``[B, C_latent, T_latent, H_latent, W_latent]``)
-        # rather than the pre-allocation's pixel shape. Skip the pinned
-        # ~50 MB buffer entirely; we always fall through to the
-        # ``samples = output_batch.output.cpu()`` branch below in that
-        # mode. ``skip_pixel_prealloc`` also gates the slow-path warning.
-        skip_pixel_prealloc = fastvideo_args.output_type == "latent"
-        if skip_pixel_prealloc:
-            samples = torch.empty(0, device='cpu')
-        else:
-            samples = torch.empty(
-                (latent_batch_size, 3, sampling_param.num_frames, sampling_param.height, sampling_param.width),
-                device='cpu',
-                pin_memory=fastvideo_args.pin_cpu_memory)
+        samples = torch.empty(
+            (latent_batch_size, 3, sampling_param.num_frames, sampling_param.height, sampling_param.width),
+            device='cpu',
+            pin_memory=fastvideo_args.pin_cpu_memory)
         thread.join()
 
         if thread_error["error"] is not None:
@@ -630,61 +619,30 @@ class VideoGenerator:
         if output_batch.output.shape == samples.shape:
             samples.copy_(output_batch.output)
         else:
-            if not skip_pixel_prealloc:
-                logger.warning("Output shape %s does not match expected shape %s; use slow path",
-                               output_batch.output.shape, samples.shape)
+            logger.warning("Output shape %s does not match expected shape %s; use slow path", output_batch.output.shape,
+                           samples.shape)
             samples = output_batch.output.cpu()
         logging_info = output_batch.logging_info
 
         gen_time = time.perf_counter() - start_time
         logger.info("Generated successfully in %.2f seconds", gen_time)
 
-        # Three mutually-exclusive output modes determine whether (a) we
-        # build an RGB frame buffer and (b) what file we write to disk:
-        #
-        #   1. `output_type == "latent"` — VAE is bypassed in DecodingStage
-        #      and `samples` holds raw latents (arbitrary channel count).
-        #      The RGB grid / uint8 / mp4 / png pipeline below cannot
-        #      consume those, so we skip it entirely and let callers work
-        #      with the latent tensor directly via `result["samples"]`.
-        #   2. Audio-only workload — `samples` is a 1×3×1×8×8 placeholder
-        #      no caller will use; skip the grid loop and save a `.wav`.
-        #   3. Pixel video / image — the historical happy path.
-        is_latent_output = fastvideo_args.output_type == "latent"
-        audio_only = bool(output_batch.extra.get("audio_only"))
+        # Process outputs
+        videos = rearrange(samples, "b c t h w -> t b c h w")
+        frames = []
+        for x in videos:
+            x = torchvision.utils.make_grid(x, nrow=6)
+            x = x.permute(1, 2, 0).squeeze(-1)
+            x = (x * 255).to(torch.uint8)
+            frames.append(x.cpu().numpy())
 
-        frames: list[np.ndarray] | None
-        if is_latent_output or audio_only:
-            frames = None if is_latent_output else []
-        else:
-            videos = rearrange(samples, "b c t h w -> t b c h w")
-            frames = []
-            for x in videos:
-                x = torchvision.utils.make_grid(x, nrow=6)
-                x = x.permute(1, 2, 0).squeeze(-1)
-                x = (x * 255).to(torch.uint8)
-                frames.append(x.cpu().numpy())
-
-        save_to_disk = batch.save_video and not is_latent_output
-        if save_to_disk:
-            if audio_only:
-                # Audio-only workload: write a standalone .wav rather than
-                # muxing the audio into a placeholder mp4 (which forces
-                # ffmpeg to round 8x8 placeholder frames up to 16x16).
-                output_path = self._rewrite_extension(output_path, ".wav")
-                self._write_pcm_wav(
-                    output_path,
-                    output_batch.extra["audio"],
-                    int(output_batch.extra["audio_sample_rate"]),
-                )
-                logger.info("Saved audio to %s", output_path)
-            elif self._is_image_workload():
+        # Save output if requested
+        if batch.save_video:
+            if self._is_image_workload():
                 # Image workloads (t2i, i2i, …): save the first frame as PNG.
-                assert frames is not None  # implied by save_to_disk and not audio_only
                 imageio.imwrite(output_path, frames[0])
                 logger.info("Saved image to %s", output_path)
             else:
-                assert frames is not None  # implied by save_to_disk and not audio_only
                 imageio.mimsave(output_path, frames, fps=batch.fps, format="mp4")
                 logger.info("Saved video to %s", output_path)
                 audio = output_batch.extra.get("audio")
@@ -697,18 +655,14 @@ class VideoGenerator:
             "prompts": prompt,
             "samples": samples if batch.return_frames else None,
             "frames": frames if batch.return_frames else None,
-            # Audio is the primary output for audio workloads — return it
-            # whenever the pipeline produced one, regardless of
-            # `return_frames` (which gates the video-shaped buffers).
-            "audio": output_batch.extra.get("audio"),
-            "audio_sample_rate": output_batch.extra.get("audio_sample_rate"),
+            "audio": output_batch.extra.get("audio") if batch.return_frames else None,
             "size": (target_height, target_width, batch.num_frames),
             "generation_time": gen_time,
             "logging_info": logging_info,
             "trajectory": output_batch.trajectory_latents,
             "trajectory_timesteps": output_batch.trajectory_timesteps,
             "trajectory_decoded": output_batch.trajectory_decoded,
-            "video_path": output_path if save_to_disk else None,
+            "video_path": output_path if batch.save_video else None,
             "peak_memory_mb": output_batch.extra.get("peak_memory_mb"),
         }
 
@@ -729,55 +683,7 @@ class VideoGenerator:
         return result.to_legacy_dict()
 
     @staticmethod
-    def _rewrite_extension(path: str, new_ext: str) -> str:
-        root, old_ext = os.path.splitext(path)
-        new_path = root + new_ext
-        if old_ext and old_ext.lower() != new_ext.lower():
-            logger.info("Rewriting output extension %s -> %s.", old_ext, new_ext)
-        return new_path
-
-    @staticmethod
-    def _audio_to_int16(audio: torch.Tensor | np.ndarray, ) -> tuple[np.ndarray, int]:
-        """Normalize `[samples]` / `[samples, channels]` / `[channels,
-        samples]` audio in roughly [-1, 1] to a `(int16 [samples,
-        channels], num_channels)` pair. Raises `ValueError` for shapes
-        we can't classify.
-        """
-        if torch.is_tensor(audio):
-            audio_np = audio.detach().cpu().float().numpy()
-        else:
-            audio_np = np.asarray(audio, dtype=np.float32)
-        if audio_np.ndim == 1:
-            audio_np = audio_np[:, None]
-        elif audio_np.ndim == 2:
-            if audio_np.shape[0] <= 8 and audio_np.shape[1] > audio_np.shape[0]:
-                audio_np = audio_np.T
-        else:
-            raise ValueError(f"Unexpected audio shape {audio_np.shape}.")
-        audio_np = np.clip(audio_np, -1.0, 1.0)
-        audio_int16 = (audio_np * 32767.0).astype(np.int16)
-        return audio_int16, audio_int16.shape[1]
-
-    @classmethod
-    def _write_pcm_wav(
-        cls,
-        wav_path: str,
-        audio: torch.Tensor | np.ndarray,
-        sample_rate: int,
-    ) -> int:
-        """Write 16-bit PCM WAV; returns the channel count."""
-        import wave
-        audio_int16, num_channels = cls._audio_to_int16(audio)
-        with wave.open(wav_path, "wb") as f:
-            f.setnchannels(num_channels)
-            f.setsampwidth(2)
-            f.setframerate(sample_rate)
-            f.writeframes(audio_int16.tobytes())
-        return num_channels
-
-    @classmethod
     def _mux_audio(
-        cls,
         video_path: str,
         audio: torch.Tensor | np.ndarray,
         sample_rate: int,
@@ -787,16 +693,40 @@ class VideoGenerator:
             import av
         except ImportError:
             logger.warning("PyAV not installed; cannot mux audio. "
-                           "Install with: uv pip install av")
+                           "Install with: pip install av")
             return False
 
+        if torch.is_tensor(audio):
+            audio_np = audio.detach().cpu().float().numpy()
+        else:
+            audio_np = np.asarray(audio, dtype=np.float32)
+
+        if audio_np.ndim == 1:
+            audio_np = audio_np[:, None]
+        elif audio_np.ndim == 2:
+            if audio_np.shape[0] <= 8 and audio_np.shape[1] > audio_np.shape[0]:
+                audio_np = audio_np.T
+        else:
+            logger.warning("Unexpected audio shape %s; skipping mux.", audio_np.shape)
+            return False
+
+        audio_np = np.clip(audio_np, -1.0, 1.0)
+        audio_int16 = (audio_np * 32767.0).astype(np.int16)
+        num_channels = audio_int16.shape[1]
+        layout = "stereo" if num_channels == 2 else "mono"
+
         try:
+            import wave
             with tempfile.TemporaryDirectory() as tmpdir:
                 out_path = os.path.join(tmpdir, "muxed.mp4")
                 wav_path = os.path.join(tmpdir, "audio.wav")
 
-                num_channels = cls._write_pcm_wav(wav_path, audio, sample_rate)
-                layout = "stereo" if num_channels == 2 else "mono"
+                # Write audio to WAV file
+                with wave.open(wav_path, "wb") as wav_file:
+                    wav_file.setnchannels(num_channels)
+                    wav_file.setsampwidth(2)
+                    wav_file.setframerate(sample_rate)
+                    wav_file.writeframes(audio_int16.tobytes())
 
                 # Open input video and audio
                 input_video = av.open(video_path)