[ci] Cherrypick perf regression skills to component time commit

[ci] Add performance baseline reseed skill and failed-run artifacts

.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/scripts/sync-skills.sh

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

.agents/skills/index.jsonl

@@ -5,3 +5,6 @@
 {"name": "evaluate-video-quality", "description": "Evaluate generated video quality using available metrics (SSIM, loss trajectory, caption consistency)", "path": "evaluate-video-quality/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "index-related-work", "description": "Ingest a paper or repository into the related work index", "path": "index-related-work/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "search-related-work", "description": "Query the related work index for relevant papers, repos, or comparisons", "path": "search-related-work/SKILL.md", "status": "draft", "trust": "low"}
+{"name": "seed-ssim-references", "description": "Run a new or updated fastvideo/tests/ssim/ test on Modal, pull generated videos, and upload them to FastVideo/ssim-reference-videos so the test has a regression baseline", "path": "seed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}
+{"name": "reseed-ssim-references", "description": "Re-seed (overwrite) HF reference videos for an existing fastvideo/tests/ssim/ test and a single model id on Modal L40S. Always backs up current refs first, regenerates on Modal, pauses for the user to eyeball before-vs-after, then uploads with --force scoped to --model-id. Sister skill to seed-ssim-references; use when intentional code change has invalidated existing refs", "path": "reseed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}
+{"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

@@ -0,0 +1,426 @@
+---
+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

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

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

@@ -0,0 +1,376 @@
+---
+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.
+---
+
+# Seed SSIM Reference Artefacts (mp4 or pt)
+
+## 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.
+
+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
+   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).
+
+## When to use
+
+- A new `test_*_similarity.py` file has been added in `fastvideo/tests/ssim/`
+  and the HF dataset has no `reference_videos/default/L40S_reference_videos/<model_id>/`
+  subtree for it yet.
+
+## When not to use
+
+- Regular CI runs — once refs exist, `pytest fastvideo/tests/ssim/` downloads
+  them automatically.
+- Re-seeding an existing test. That requires `--force` on the upload step, and
+  is out of scope here; treat as a separate, deliberate operation.
+
+## Inputs
+
+The skill has **one required input**: the path to the new SSIM test file.
+Prompt the user for it if they didn't supply it.
+
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `test_file` | Yes | e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`. The skill's first action is to ask for this if missing. |
+
+Everything else is fixed:
+
+- Modal runner GPU: **L40S** (hardcoded in `fastvideo/tests/modal/ssim_test.py`).
+- Device folder: `L40S_reference_videos`.
+- Quality tier: `default` (the tier CI runs). The `full_quality` tier is not
+  seeded by this skill.
+- HF repo: `FastVideo/ssim-reference-videos` (dataset).
+- Multi-model test files: all model ids in `*_MODEL_TO_PARAMS` are seeded
+  together; the Modal run produces one mp4 per (model, prompt, backend) and
+  the upload scopes by `--model-id`, looping if there is more than one.
+
+## Prerequisites
+
+The user has confirmed:
+
+- `modal` CLI authenticated.
+- `HF_API_KEY` (or `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`) exported with write
+  access to `FastVideo/ssim-reference-videos`.
+- The test file runs locally end-to-end (generates an mp4; SSIM assertion
+  failure due to missing reference is expected and fine).
+
+Fail fast if the token env var is missing.
+
+## Steps
+
+### 1. Ask for the test file, then detect artefact type
+
+If the user didn't name one, ask: *"Which SSIM test file do you want to seed
+references for? (e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`)"*.
+
+Validate:
+
+- Path exists and matches `fastvideo/tests/ssim/test_*_similarity.py`.
+- File defines a `*_MODEL_TO_PARAMS` dict — grep it to extract the set of
+  model ids. Those ids drive step 5.
+
+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
+
+Pick a subdir name so repeated runs don't collide:
+
+```bash
+SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
+TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
+SUBDIR="${TIMESTAMP}_${SHORT_COMMIT}"
+```
+
+Then launch the Modal run. The `IMAGE_VERSION` and `BUILDKITE_*` env-prefix
+**must** match what CI exports in `.buildkite/scripts/pr_test.sh`, otherwise
+`fastvideo/tests/modal/ssim_test.py` resolves a different GHCR image tag
+(default is `latest`, CI is `py3.12-latest`) and bakes different values into
+the image's frozen env block (`ssim_test.py:17-18, 38-46`). Mismatched image
+or env produces SSIM drift that doesn't show up until the same commit runs
+in CI.
+
+```bash
+IMAGE_VERSION="py3.12-latest" \
+BUILDKITE_REPO="$(git config --get remote.origin.url)" \
+BUILDKITE_COMMIT="$(git rev-parse HEAD)" \
+BUILDKITE_PULL_REQUEST="${BUILDKITE_PULL_REQUEST:-false}" \
+modal run fastvideo/tests/modal/ssim_test.py \
+    --git-repo="$(git config --get remote.origin.url)" \
+    --git-commit="$(git rev-parse HEAD)" \
+    --hf-api-key="$HF_API_KEY" \
+    --test-files="<test_file>" \
+    --sync-generated-to-volume \
+    --generated-volume-subdir="$SUBDIR" \
+    --skip-reference-download \
+    --no-fail-fast
+```
+
+Env prefix rationale (parity with CI; see `.buildkite/pipeline.yml:1-3` and
+`.buildkite/scripts/pr_test.sh:62-83`):
+- `IMAGE_VERSION=py3.12-latest`: pins the Modal image tag to the same one CI
+  uses. Without this, `ssim_test.py:17` falls back to `latest`, which on
+  GHCR is built from `Dockerfile.python3.10` — different Python, torch, and
+  flash-attn wheel than CI's `py3.12-latest` (`infra-build-image.yml:51-67`,
+  `_template-build-image.yml:65-101`).
+- `BUILDKITE_REPO`/`BUILDKITE_COMMIT`/`BUILDKITE_PULL_REQUEST`: mirror what
+  Buildkite exports. `ssim_test.py:38-46` bakes these into the image's
+  `.env(...)` block; mismatched values can perturb in-container code paths
+  that branch on PR-vs-non-PR. `false` for `BUILDKITE_PULL_REQUEST` matches
+  Buildkite's "non-PR build" sentinel.
+
+Flag rationale:
+- `--skip-reference-download`: no refs exist yet, so conftest must not try to
+  pull them.
+- `--no-fail-fast`: lets the test finish generation before `_assert_similarity`
+  raises `FileNotFoundError: Reference video folder does not exist`. The
+  expected failure is what we want — the mp4 has already been written.
+- `--sync-generated-to-volume` + `--generated-volume-subdir`: copies the
+  generated mp4s to the `hf-model-weights` Modal volume under
+  `ssim_generated_videos/default/<SUBDIR>/generated_videos/` so we can pull
+  them locally.
+
+The Modal run will end with a nonzero exit (expected) and print a
+`modal volume get hf-model-weights ssim_generated_videos/default/<SUBDIR>/generated_videos ./generated_videos_modal/default`
+command. Capture that `<SUBDIR>` — you need it for step 3.
+
+### 3. Download generated videos locally
+
+```bash
+modal volume get --force hf-model-weights \
+    ssim_generated_videos/default/"$SUBDIR"/generated_videos \
+    ./generated_videos_modal/default
+```
+
+`--force` is required when the parent `./generated_videos_modal/default`
+already exists; without it, `modal volume get` errors with `[Errno 21] Is a
+directory`. Safe to pass on the first run too.
+
+After this, the mp4s live at
+`./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+The extra `generated_videos/` level comes from the volume layout in
+`_sync_generated_videos_to_volume` (`ssim_test.py`) — the command copies
+`<repo>/fastvideo/tests/ssim/generated_videos/<tier>` to
+`ssim_generated_videos/<tier>/<SUBDIR>/generated_videos/`, and `modal volume
+get` preserves that trailing `generated_videos/` segment.
+
+### 4. PAUSE — user reviews quality
+
+Type-aware verification.
+
+**For `ARTEFACT_TYPE = pixel`** — list the downloaded mp4s and ask the user to
+open them in a video player:
+
+> "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`:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --generated-dir ./generated_videos_modal/default/generated_videos/L40S_reference_videos
+```
+
+(The `--generated-dir` points at the device-folder root inside the
+downloaded tree; `copy-local` walks all `<model>/<backend>/*.{mp4,pt}`
+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.
+
+### 6. Upload to HF — scoped per model_id, with overwrite guard
+
+For each `<model_id>`:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py upload \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --model-id "<model_id>"
+```
+
+The upload command:
+
+- Uploads **only** `reference_videos/default/L40S_reference_videos/<model_id>/`.
+- **Refuses** if any file already exists at that path on HF (this is the
+  guard — seeding a new test should never clobber existing refs). To override,
+  the user must re-run with `--force`. If the guard fires, stop and report
+  exactly which files exist; do not silently `--force`.
+
+Reads the HF token from `HF_API_KEY` / `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`.
+
+### 7. Report success
+
+List what was uploaded (paths in repo) and remind the user to push any
+related code changes. Do **not** auto-verify by re-running Modal — the user
+can run `pytest fastvideo/tests/ssim/<test_file>` later to confirm end-to-end;
+it will auto-download the refs they just uploaded.
+
+## Failure modes and how to handle them
+
+- **`HF_API_KEY` unset.** Stop before step 2. The Modal run needs it (passed
+  via `--hf-api-key`), and step 6 needs it for upload. 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>`)
+  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.
+- **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
+  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).
+- 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
+
+- `fastvideo/tests/modal/ssim_test.py` — Modal orchestrator; see
+  `--sync-generated-to-volume`, `--generated-volume-subdir`,
+  `--skip-reference-download`, `--no-fail-fast`.
+- `fastvideo/tests/ssim/reference_videos_cli.py` — `copy-local`, `upload`
+  (with `--model-id`, `--force`), `download`, `ensure` subcommands.
+  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`.
+
+## Changelog
+
+| Date | Change |
+|------|--------|
+| 2026-04-17 | Initial version (Modal sync-to-volume flow). |
+| 2026-04-21 | Rewrite: single-test scope, explicit user-review pause, per-`model_id` upload, HF overwrite guard. Dropped `scripts/seed_ssim.sh`. |
+| 2026-04-21 | Post-first-run fixes: `modal volume get` needs `--force` when parent exists; download tree has an extra `generated_videos/` level so `--generated-dir` must reflect it. |
+| 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

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

.buildkite/scripts/pr_test.sh

@@ -15,8 +15,21 @@ log "Project root: $PROJECT_ROOT"
 # Install Modal if not available
 if ! python3 -m modal --version &> /dev/null; then
     log "Modal not found, installing..."
-    python3 -m pip install modal
-    
+    if ! command -v uv &> /dev/null; then
+        log "uv not found, bootstrapping..."
+        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
+            log "Error: Failed to bootstrap uv via astral.sh installer."
+            exit 1
+        fi
+        export PATH="$HOME/.local/bin:$PATH"
+        if ! command -v uv &> /dev/null; then
+            log "Error: uv still not on PATH after bootstrap."
+            exit 1
+        fi
+    fi
+    # --break-system-packages preserves prior `pip install --user` semantics on PEP 668 agents.
+    uv pip install --system --break-system-packages modal
+
     # Verify installation
     if ! python3 -m modal --version &> /dev/null; then
         log "Error: Failed to install modal. Please install it manually."
@@ -63,7 +76,86 @@ EFFECTIVE_PR=${BUILDKITE_PULL_REQUEST:-false}
 if [ "$EFFECTIVE_PR" = "false" ] && [ -n "${PR_NUMBER:-}" ]; then
     EFFECTIVE_PR=$PR_NUMBER
 fi
-MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR IMAGE_VERSION=$IMAGE_VERSION"
+MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR BUILDKITE_BRANCH=${BUILDKITE_BRANCH:-} TEST_SCOPE=${TEST_SCOPE:-} IMAGE_VERSION=$IMAGE_VERSION"
+
+POST_RUN_HOOK=""
+
+upload_performance_artifacts() {
+    SHORT_SHA=${BUILDKITE_COMMIT:0:7}
+    LOCAL_DIR="downloaded_reports"
+
+    _download_reports() {
+        log "Downloading perf_reports/ from Modal Volume..."
+        mkdir -p "$LOCAL_DIR"
+        if ! modal volume get hf-model-weights "perf_reports/" "$LOCAL_DIR"; then
+            log "Error: Failed to download perf_reports/ from Modal Volume."
+            return 1
+        fi
+    }
+
+    _upload_dashboard() {
+        local target
+        target=$(find "$LOCAL_DIR" -name "dashboard_${SHORT_SHA}_*" | head -n 1)
+        log "TARGET dashboard: '$target'"
+
+        if [ -n "$target" ]; then
+            log "Found dashboard: $target. Uploading to Buildkite..."
+            buildkite-agent artifact upload "$target"
+            buildkite-agent annotate --style info --context "perf-dashboard" < "$target"
+        else
+            log "Warning: Could not find a dashboard file matching $SHORT_SHA"
+        fi
+    }
+
+    _upload_perf_summary() {
+        local target
+        target=$(find "$LOCAL_DIR" -name "perf_${SHORT_SHA}_*" | head -n 1)
+        log "TARGET perf summary: '$target'"
+
+        if [ -n "$target" ]; then
+            log "Found perf summary: $target. Uploading to Buildkite..."
+            buildkite-agent artifact upload "$target"
+            buildkite-agent annotate --style info --context "perf-summary" < "$target"
+        else
+            log "Warning: Could not find a perf summary file matching $SHORT_SHA"
+        fi
+    }
+
+    _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
+}
 
 case "$TEST_TYPE" in
     "encoder")
@@ -124,8 +216,9 @@ case "$TEST_TYPE" in
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_lora_extraction_tests"
         ;;
     "performance")
-        log "Running performance tests..."
+        log "Running performance tests on Modal..."
         MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_performance_tests"
+        POST_RUN_HOOK="upload_performance_artifacts"
         ;;
     "api_server")
         log "Running API server integration tests..."
@@ -147,5 +240,10 @@ else
     log "Error: Modal test failed with exit code: $TEST_EXIT_CODE"
 fi
 
+if [ -n "$POST_RUN_HOOK" ]; then
+    log "Executing post-run hook: $POST_RUN_HOOK"
+    "$POST_RUN_HOOK"
+fi
+
 log "=== Test execution completed with exit code: $TEST_EXIT_CODE ==="
 exit $TEST_EXIT_CODE

.buildkite/scripts/pre_commit.sh

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

.github/mergify.yml

@@ -105,7 +105,7 @@ pull_request_rules:
           - files~=^fastvideo/pipelines/samplers/
           - files~=^fastvideo/entrypoints/
           - files~=^fastvideo/worker/
-          - files~=^fastvideo/configs/sample/
+          - files~=^fastvideo/api/sampling_param
           - files~=^fastvideo/configs/pipelines/
           - files~=^examples/inference/
       - -closed

.github/workflows/infra-docs.yml

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

.github/workflows/publish-fastvideo.yml

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

.github/workflows/publish-kernel.yml

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

.pre-commit-config.yaml

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

AGENTS.md

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

benchmarks/fvd/feature_extractors.py

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

benchmarks/fvd/i3d_model.py

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

benchmarks/scripts/run.sh

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

benchmarks/scripts/setup_fvd.sh

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

comfyui/README.md

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

docker/Dockerfile.python3.10

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

docker/Dockerfile.python3.11

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

docker/Dockerfile.python3.12

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

docker/Dockerfile.python3.12.cuda12.9.1

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

docker/rocm7_1.Dockerfile

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

docs/README.md

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

docs/api/fastvideo.md

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

docs/attention/attn_qat/index.md

@@ -1,318 +0,0 @@
-# Attention QAT
-
-Attention QAT in FastVideo covers two related, but different, backends:
-
-- `ATTN_QAT_INFER`: the inference-oriented CUDA kernel path
-- `ATTN_QAT_TRAIN`: the training-oriented Triton attention path
-
-Both are selected with `FASTVIDEO_ATTENTION_BACKEND`, but they are not
-interchangeable. The main practical split is:
-
-- use `ATTN_QAT_INFER` for standalone inference with the dedicated inference
-  kernel
-- use `ATTN_QAT_TRAIN` for finetuning, validation during training, or when you
-  specifically want to reproduce the training-side attention path
-
-## Quick Start
-
-If your goal is "run Wan 2.1 14B with Attention QAT inference weights", this is
-the shortest path:
-
-1. Build the in-repo kernel package so FastVideo can import `attn_qat_infer`.
-2. Download the Wan 2.1 14B QAT checkpoint.
-3. Edit the provided inference example to point at the 14B base model and the
-   downloaded QAT safetensors.
-4. Run the example with `ATTN_QAT_INFER`.
-
-### Step 1. Build the kernel package
-
-Before using either Attention QAT backend, build the in-repo
-`fastvideo-kernel` package from source:
-
-```bash
-git submodule update --init --recursive
-cd fastvideo-kernel
-./build.sh
-```
-
-After a successful build:
-
-- `ATTN_QAT_TRAIN` should be able to import `fastvideo_kernel`
-- `ATTN_QAT_INFER` should be able to import `attn_qat_infer`
-
-`ATTN_QAT_INFER` currently targets the Blackwell CUDA path under
-`fastvideo-kernel/attn_qat_infer/` and requires CUDA 12.8+.
-
-### Step 2. Download the Wan 2.1 14B QAT checkpoint
-
-FastVideo includes a helper script:
-
-- `examples/inference/optimizations/download_14B_qat.sh`
-
-By default it downloads:
-
-- Hugging Face repo: `FastVideo/14B_qat_400`
-- local directory: `checkpoints/14B_qat_400`
-
-Prerequisites:
-
-- `huggingface_hub` installed, for example:
-  `uv pip install huggingface_hub`
-- access to the model repo if it is private or gated:
-  `huggingface-cli login`
-
-Run the downloader:
-
-```bash
-bash examples/inference/optimizations/download_14B_qat.sh
-```
-
-To download into a custom directory:
-
-```bash
-bash examples/inference/optimizations/download_14B_qat.sh /path/to/14B_qat_400
-```
-
-The script prints a ready-to-copy `init_weights_from_safetensors=...` value at
-the end.
-
-### Step 3. Edit the provided inference example
-
-The example to start from is:
-
-- `examples/inference/optimizations/attn_qat_inference_example.py`
-
-Open that file and update these two values:
-
-1. Change the base model from `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` to
-   `Wan-AI/Wan2.1-T2V-14B-Diffusers`
-2. Replace
-   `init_weights_from_safetensors="safetensors_path"` with the directory that
-   contains the downloaded `.safetensors` files
-
-Example:
-
-```python
-import os
-
-from fastvideo import VideoGenerator
-
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"
-
-generator = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
-    num_gpus=1,
-    use_fsdp_inference=True,
-    dit_cpu_offload=False,
-    vae_cpu_offload=False,
-    text_encoder_cpu_offload=True,
-    pin_cpu_memory=False,
-    init_weights_from_safetensors="checkpoints/14B_qat_400",
-)
-```
-
-Important:
-
-- the checked-in example currently uses the `1.3B` base model until you edit it
-- do not load the 14B QAT weights on top of the `1.3B` base model; the weights
-  and model config will not match
-
-### Step 4. Run the inference example
-
-```bash
-python examples/inference/optimizations/attn_qat_inference_example.py
-```
-
-Generated videos are written to `video_samples/` by default.
-
-## Backend Overview
-
-| Backend | Best for | Package requirement | Primary kernel location |
-|---------|----------|---------------------|-------------------------|
-| `ATTN_QAT_TRAIN` | finetuning, training-time validation, reproducing the training path | `fastvideo_kernel` | `fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py` |
-| `ATTN_QAT_INFER` | standalone inference with the dedicated CUDA kernel | `attn_qat_infer` from the in-repo `fastvideo-kernel` checkout | `fastvideo-kernel/attn_qat_infer/` |
-
-FastVideo routes backend selection through:
-
-- `fastvideo/envs.py`
-- `fastvideo/platforms/cuda.py`
-- `fastvideo/attention/backends/attn_qat_train.py`
-- `fastvideo/attention/backends/attn_qat_infer.py`
-
-The legacy training pipeline also contains explicit Attention QAT integration:
-
-- `fastvideo/training/training_pipeline.py`
-
-That pipeline forces generator loading through `ATTN_QAT_TRAIN` when
-`FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN` or `--generator_4bit_attn` is
-enabled.
-
-## Inference Workflows
-
-For standalone inference, prefer `ATTN_QAT_INFER` when the CUDA kernel is
-available. Use `ATTN_QAT_TRAIN` for inference only if you intentionally want to
-exercise the training-side attention path for debugging or parity checks.
-
-### Minimal Python example
-
-```python
-import os
-
-from fastvideo import VideoGenerator
-
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"
-
-generator = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    num_gpus=1,
-)
-
-generator.generate_video(
-    "A cinematic close-up of rain on a neon street at night.",
-    output_path="video_samples",
-    save_video=True,
-)
-```
-
-### Loading custom safetensors during inference
-
-FastVideo supports loading custom transformer weights through
-`init_weights_from_safetensors`.
-
-This value can point to either:
-
-- a directory containing one or more `.safetensors` files
-- a single `.safetensors` file
-
-For Wan 2.1 14B QAT inference, the common pattern is:
-
-```python
-generator = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
-    num_gpus=1,
-    use_fsdp_inference=True,
-    init_weights_from_safetensors="checkpoints/14B_qat_400",
-)
-```
-
-### CLI example
-
-You can also force the backend from the command line:
-
-```bash
-FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER \
-fastvideo generate \
-  --model-path Wan-AI/Wan2.1-T2V-14B-Diffusers \
-  --num-gpus 1 \
-  --sp-size 1 \
-  --tp-size 1 \
-  --height 480 \
-  --width 832 \
-  --num-frames 77 \
-  --num-inference-steps 50 \
-  --guidance-scale 6.0 \
-  --prompt "A cinematic close-up of rain on a neon street at night." \
-  --output-path outputs_video/
-```
-
-If you want to use custom QAT transformer weights from the CLI, pass the same
-custom weight override that the Python API uses:
-
-```bash
-FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER \
-fastvideo generate \
-  --model-path Wan-AI/Wan2.1-T2V-14B-Diffusers \
-  --init-weights-from-safetensors checkpoints/14B_qat_400 \
-  --num-gpus 1 \
-  --output-path outputs_video/ \
-  --prompt "A cinematic close-up of rain on a neon street at night."
-```
-
-## Training Workflows
-
-Today the checked-in Attention QAT training launchers use the legacy training
-pipeline in `fastvideo/training/wan_training_pipeline.py`.
-
-### Ready-made launchers
-
-Use the provided SLURM scripts directly:
-
-```bash
-sbatch examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_attn.sh
-sbatch examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_attn.sh
-```
-
-Both scripts already set:
-
-```bash
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-```
-
-Before launching, update the script-local values that depend on your
-environment:
-
-- `WANDB_API_KEY`
-- `MODEL_PATH`
-- `DATA_DIR`
-- `VALIDATION_DATASET_FILE`
-- output directory and SLURM resource requests
-
-### What the launchers run
-
-The training scripts eventually invoke:
-
-```bash
-torchrun fastvideo/training/wan_training_pipeline.py ...
-```
-
-If you are adapting the workflow to your own cluster or running outside SLURM,
-the main Attention QAT requirement is still:
-
-```bash
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-```
-
-Then launch the normal Wan training pipeline with your preferred `torchrun`
-arguments and training flags.
-
-## Where The Code Lives
-
-Use these paths when you want to trace or modify the Attention QAT flow:
-
-| Location | Purpose |
-|----------|---------|
-| `fastvideo/attention/backends/attn_qat_train.py` | FastVideo wrapper that imports and calls the Triton training kernel |
-| `fastvideo/attention/backends/attn_qat_infer.py` | FastVideo wrapper that imports and calls the inference kernel |
-| `fastvideo-kernel/CMakeLists.txt` | Kernel build definition that compiles the `attn_qat_infer` inference extensions |
-| `fastvideo/platforms/cuda.py` | Chooses the concrete attention backend at runtime |
-| `fastvideo/envs.py` | Documents supported `FASTVIDEO_ATTENTION_BACKEND` values |
-| `fastvideo/training/training_pipeline.py` | Training-time forcing logic for the generator attention backend |
-| `fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py` | Triton implementation for `ATTN_QAT_TRAIN` |
-| `fastvideo-kernel/attn_qat_infer/api.py` | Python API entrypoint for the inference kernel |
-| `fastvideo-kernel/benchmarks/benchmark_*.py` | Kernel-side benchmark scripts for FlashAttn2, SageAttention3, FP4, and comparison plots |
-| `fastvideo-kernel/attn_qat_infer/blackwell/api.cu` | CUDA implementation behind `ATTN_QAT_INFER` |
-| `fastvideo-kernel/tests/test_attn_qat_train.py` | Kernel-level test coverage for the training path |
-| `examples/inference/optimizations/attn_qat_inference_example.py` | Ready-to-edit inference example for custom Attention QAT weights |
-| `examples/inference/optimizations/download_14B_qat.sh` | Helper script for downloading the Wan 2.1 14B QAT checkpoint |
-| `examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_attn.sh` | Ready-to-run Wan 1.3B Attention QAT finetune launcher |
-| `examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_attn.sh` | Ready-to-run Wan 14B Attention QAT finetune launcher |
-
-## Troubleshooting
-
-- If `ATTN_QAT_TRAIN` fails to import, verify that `fastvideo-kernel` built
-  successfully and exposes `fastvideo_kernel`.
-- If `ATTN_QAT_INFER` fails to import, verify that the local build exposes the
-  `attn_qat_infer` package.
-- If the Wan 2.1 14B example fails after you changed only the checkpoint path,
-  make sure you also changed the base model to
-  `Wan-AI/Wan2.1-T2V-14B-Diffusers`.
-- If you hit issues with CPU memory pressure or obscure CUDA argument errors in
-  the example script, try setting `pin_cpu_memory=False`.
-- If you want a known-safe fallback for debugging, use
-  `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`.
-
-## Related Pages
-
-- [Attention Overview](../index.md)
-- [Inference Optimizations](../../inference/optimizations.md)
-- [Debugging](../../utilities/debugging.md)

docs/attention/index.md

@@ -5,8 +5,6 @@ FastVideo provides highly optimized custom attention kernels to accelerate video
 ## Supported Kernels
 
 * **[Video Sparse Attention (VSA)](vsa/index.md)**: Sparse attention mechanism selecting top-k blocks.
-* **[Attention QAT](attn_qat/index.md)**: Dedicated guide for Attention QAT
-  inference, training, checkpoint loading, and troubleshooting.
 * **[Sliding Tile Attention (STA)](sta/index.md)**: STA kernel support is kept in
   `fastvideo-kernel`; full FastVideo STA pipeline workflow is archived in
   `sta_do_not_delete`.

docs/contributing/ci_architecture.md

@@ -173,7 +173,7 @@ Applied by Mergify based on which paths you modified. Multiple scope labels can
 | Label | File paths that trigger it |
 |-------|---------------------------|
 | `scope: training` | `fastvideo/train/`, `fastvideo/training/`, `fastvideo/distillation/`, `examples/train/`, `examples/training/`, `examples/distill/` |
-| `scope: inference` | `fastvideo/pipelines/basic/`, `fastvideo/pipelines/stages/`, `fastvideo/pipelines/samplers/`, `fastvideo/entrypoints/`, `fastvideo/worker/`, `fastvideo/configs/sample/`, `fastvideo/configs/pipelines/`, `examples/inference/` |
+| `scope: inference` | `fastvideo/pipelines/basic/`, `fastvideo/pipelines/stages/`, `fastvideo/pipelines/samplers/`, `fastvideo/entrypoints/`, `fastvideo/worker/`, `fastvideo/api/sampling_param.py`, `fastvideo/configs/pipelines/`, `examples/inference/` |
 | `scope: attention` | `fastvideo/attention/` |
 | `scope: kernel` | `fastvideo-kernel/`, `csrc/` |
 | `scope: data` | `fastvideo/dataset/`, `fastvideo/pipelines/preprocess/`, `examples/preprocessing/` |

docs/contributing/coding_agents.md

@@ -44,7 +44,7 @@ FastVideo maps a Diffusers-style repo into a pipeline like:
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
   weight name translation.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
-- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipeline logic built from stages.
 - `model_index.json`: the HF repo entrypoint that maps component names to
   classes and weight files.
@@ -55,7 +55,7 @@ Minimal usage example (based on `examples/inference/basic/basic.py`):
 
 ```python
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
 generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
@@ -296,8 +296,10 @@ Action:
 
 - Add or reuse a numerical parity test that loads the official model and the
   FastVideo model and compares outputs.
-- See examples in `tests/local_tests/` (e.g., `tests/local_tests/upsamplers/`)
-  and the commands in `tests/local_tests/README.md`.
+- See examples in `tests/local_tests/` organized by model family
+  (e.g., `tests/local_tests/sd35/`, `tests/local_tests/ltx2/`,
+  `tests/local_tests/stable_audio/`) and the navigation index in
+  `tests/local_tests/README.md`.
 - If there are discrepancies, add opt‑in logging to both models and compare
   activation summaries (layer output sums, per‑stage logs).
 - First align the loaded weights (validate `param_names_mapping`).
@@ -319,7 +321,8 @@ Purpose:
 
 - `fastvideo/configs/pipelines/` describes pipeline wiring and model module
   names.
-- `fastvideo/configs/sample/` defines default runtime parameters.
+- `fastvideo/api/sampling_param.py` defines runtime sampling parameters.
+  Defaults come from profiles in `fastvideo/pipelines/basic/<family>/profiles.py`.
 
 Action:
 
@@ -347,7 +350,8 @@ Purpose:
 
 Action:
 
-- Add a pipeline parity test under `tests/local_tests/pipelines/`.
+- Add a pipeline parity test under `tests/local_tests/<family>/`
+  (e.g., `tests/local_tests/<family>/test_<family>_pipeline_parity.py`).
 - See the [Testing Guide](testing.md) for test conventions.
 
 ### 7) Add user‑facing examples
@@ -474,7 +478,7 @@ FastVideo integration.
 3. Pipeline wiring.
    - Pipeline: `fastvideo/pipelines/basic/wan/wan_pipeline.py`
    - Pipeline config: `fastvideo/configs/pipelines/wan.py`
-   - Sampling defaults: `fastvideo/configs/sample/wan.py`
+   - Sampling defaults: `fastvideo/pipelines/basic/wan/profiles.py`
 
 4. Minimal example.
    - Script: `examples/inference/basic/basic.py`

docs/contributing/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

@@ -1,7 +1,7 @@
 status_definitions:
   kept: "Public field remains on a public adapter surface with the same meaning."
   moved: "Public field remains supported but normalizes into a different nested path."
-  profile_owned: "Public field remains supported only through a model/profile-specific surface."
+  preset_owned: "Public field remains supported only through a model/preset-specific surface."
   compatibility_only: "Legacy public field remains adapter-only during migration and is not part of the canonical typed schema."
   private_only: "Field should only be handled by private adapters and is not a public FastVideo compatibility promise."
   internal_only: "Field is runtime/config plumbing and should not be part of the new public typed inference API."
@@ -29,24 +29,23 @@ surfaces:
       vae_cpu_offload: generator.engine.offload.vae
       pin_cpu_memory: generator.engine.offload.pin_cpu_memory
       enable_torch_compile: generator.engine.compile.enabled
-      torch_compile_kwargs: generator.engine.compile.kwargs
+      torch_compile_kwargs: generator.engine.compile.backend,fullgraph,mode,dynamic,extras
       disable_autocast: generator.engine.disable_autocast
       enable_stage_verification: generator.engine.enable_stage_verification
       prompt_txt: request.inputs.prompt_path
       override_text_encoder_safetensors: generator.pipeline.components.text_encoder_weights
       override_text_encoder_quant: generator.engine.quantization.text_encoder_quant
-      transformer_quant: generator.engine.quantization.transformer_quant
       override_transformer_cls_name: generator.pipeline.components.override_transformer_cls_name
       init_weights_from_safetensors: generator.pipeline.components.transformer_weights
       init_weights_from_safetensors_2: generator.pipeline.components.transformer_2_weights
       override_pipeline_cls_name: generator.pipeline.components.override_pipeline_cls_name
       boundary_ratio: request.sampling.boundary_ratio
-    profile_owned:
-      ltx2_vae_tiling: generator.pipeline.profile_overrides.ltx2.vae_tiling
-      ltx2_vae_spatial_tile_size_in_pixels: generator.pipeline.profile_overrides.ltx2.vae.spatial_tile_size_in_pixels
-      ltx2_vae_spatial_tile_overlap_in_pixels: generator.pipeline.profile_overrides.ltx2.vae.spatial_tile_overlap_in_pixels
-      ltx2_vae_temporal_tile_size_in_frames: generator.pipeline.profile_overrides.ltx2.vae.temporal_tile_size_in_frames
-      ltx2_vae_temporal_tile_overlap_in_frames: generator.pipeline.profile_overrides.ltx2.vae.temporal_tile_overlap_in_frames
+      ltx2_vae_tiling: generator.pipeline.vae_tiling
+    preset_owned:
+      ltx2_vae_spatial_tile_size_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_size_in_pixels
+      ltx2_vae_spatial_tile_overlap_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_overlap_in_pixels
+      ltx2_vae_temporal_tile_size_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_size_in_frames
+      ltx2_vae_temporal_tile_overlap_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_overlap_in_frames
       ltx2_initial_latent_path: request.extensions.ltx2.initial_latent_path
     compatibility_only:
       mode: "Legacy multi-mode FastVideoArgs switch; typed inference config should not expose execution mode."
@@ -70,16 +69,16 @@ surfaces:
   pipeline_config_base:
     moved:
       pipeline_config_path: generator.pipeline.components.pipeline_config_path
-    profile_owned:
-      embedded_cfg_scale: generator.pipeline.profile_overrides.embedded_cfg_scale
-      flow_shift: generator.pipeline.profile_overrides.flow_shift
-      flow_shift_sr: generator.pipeline.profile_overrides.flow_shift_sr
-      is_causal: generator.pipeline.profile_overrides.is_causal
-      vae_tiling: generator.pipeline.profile_overrides.vae_tiling
-      vae_sp: generator.pipeline.profile_overrides.vae_sp
-      dmd_denoising_steps: generator.pipeline.profile_overrides.dmd_denoising_steps
-      ti2v_task: generator.pipeline.profile_overrides.ti2v_task
-      boundary_ratio: generator.pipeline.profile_overrides.boundary_ratio
+    preset_owned:
+      embedded_cfg_scale: generator.pipeline.preset_overrides.embedded_cfg_scale
+      flow_shift: generator.pipeline.preset_overrides.flow_shift
+      flow_shift_sr: generator.pipeline.preset_overrides.flow_shift_sr
+      is_causal: generator.pipeline.preset_overrides.is_causal
+      vae_tiling: generator.pipeline.preset_overrides.vae_tiling
+      vae_sp: generator.pipeline.preset_overrides.vae_sp
+      dmd_denoising_steps: generator.pipeline.preset_overrides.dmd_denoising_steps
+      ti2v_task: generator.pipeline.preset_overrides.ti2v_task
+      boundary_ratio: generator.pipeline.preset_overrides.boundary_ratio
     compatibility_only:
       model_path: "Redundant with generator.model_path."
       disable_autocast: "Duplicated by generator.engine.disable_autocast during migration."
@@ -98,7 +97,7 @@ surfaces:
       postprocess_text_funcs: "Internal text postprocessing hooks."
 
   pipeline_config_extensions:
-    profile_owned:
+    preset_owned:
       conditioning_strategy:
         sources:
           - fastvideo.configs.pipelines.cosmos.CosmosConfig
@@ -307,11 +306,35 @@ 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."
-      guidance_scale: "Gen3C pipeline-level default pending profile/default-request cleanup."
-      num_inference_steps: "Gen3C pipeline-level default pending profile/default-request cleanup."
+      guidance_scale: "Gen3C pipeline-level default pending preset/default-request cleanup."
+      num_inference_steps: "Gen3C pipeline-level default pending preset/default-request cleanup."
     internal_only:
       audio_decoder_config: "Legacy internal component config object."
       audio_decoder_precision: "Precision override pending dedicated component precision design."
@@ -355,88 +378,43 @@ surfaces:
       return_frames: request.output.return_frames
       return_trajectory_latents: request.runtime.return_trajectory_latents
       return_trajectory_decoded: request.runtime.return_trajectory_decoded
-    profile_owned:
+      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
       num_cond_frames: request.stage_overrides.refine.num_cond_frames
       trajectory_type: request.extensions.gen3c.trajectory_type
       movement_distance: request.extensions.gen3c.movement_distance
       camera_rotation: request.extensions.gen3c.camera_rotation
+      prompt_attention_mask: request.extensions.hyworld.prompt_attention_mask
+      negative_attention_mask: request.extensions.hyworld.negative_attention_mask
+      camera_states: request.extensions.hunyuangamecraft.camera_states
+      camera_trajectory: request.extensions.hunyuangamecraft.camera_trajectory
+      action_list: request.extensions.hunyuangamecraft.action_list
+      action_speed_list: request.extensions.hunyuangamecraft.action_speed_list
+      gt_latents: request.extensions.hunyuangamecraft.gt_latents
+      conditioning_mask: request.extensions.hunyuangamecraft.conditioning_mask
+      ltx2_cfg_scale_video: request.extensions.ltx2.cfg_scale_video
+      ltx2_cfg_scale_audio: request.extensions.ltx2.cfg_scale_audio
+      ltx2_modality_scale_video: request.extensions.ltx2.modality_scale_video
+      ltx2_modality_scale_audio: request.extensions.ltx2.modality_scale_audio
+      ltx2_rescale_scale: request.extensions.ltx2.rescale_scale
+      ltx2_stg_scale_video: request.extensions.ltx2.stg_scale_video
+      ltx2_stg_scale_audio: request.extensions.ltx2.stg_scale_audio
+      ltx2_stg_blocks_video: request.extensions.ltx2.stg_blocks_video
+      ltx2_stg_blocks_audio: request.extensions.ltx2.stg_blocks_audio
+      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."
 
-  sampling_param_extensions:
-    moved: {}
-    profile_owned:
-      action_list:
-        target: request.extensions.hunyuangamecraft.action_list
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      action_speed_list:
-        target: request.extensions.hunyuangamecraft.action_speed_list
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      camera_states:
-        target: request.extensions.hunyuangamecraft.camera_states
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      camera_trajectory:
-        target: request.extensions.hunyuangamecraft.camera_trajectory
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      conditioning_mask:
-        target: request.extensions.hunyuangamecraft.conditioning_mask
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      gt_latents:
-        target: request.extensions.hunyuangamecraft.gt_latents
-        sources:
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraftSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft65FrameSamplingParam
-          - fastvideo.configs.sample.hunyuangamecraft.HunyuanGameCraft129FrameSamplingParam
-      prompt_attention_mask:
-        target: request.extensions.hyworld.prompt_attention_mask
-        sources: [fastvideo.configs.sample.hyworld.HYWorld_SamplingParam]
-      negative_attention_mask:
-        target: request.extensions.hyworld.negative_attention_mask
-        sources: [fastvideo.configs.sample.hyworld.HYWorld_SamplingParam]
-      ltx2_cfg_scale_audio:
-        target: request.extensions.ltx2.cfg_scale_audio
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_cfg_scale_video:
-        target: request.extensions.ltx2.cfg_scale_video
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_modality_scale_audio:
-        target: request.extensions.ltx2.modality_scale_audio
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_modality_scale_video:
-        target: request.extensions.ltx2.modality_scale_video
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_rescale_scale:
-        target: request.extensions.ltx2.rescale_scale
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_stg_blocks_audio:
-        target: request.extensions.ltx2.stg_blocks_audio
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_stg_blocks_video:
-        target: request.extensions.ltx2.stg_blocks_video
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_stg_scale_audio:
-        target: request.extensions.ltx2.stg_scale_audio
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
-      ltx2_stg_scale_video:
-        target: request.extensions.ltx2.stg_scale_video
-        sources: [fastvideo.configs.sample.ltx2.LTX2BaseSamplingParam]
+  sampling_param_extensions: {}
 
   openai_image_request:
     kept:
@@ -489,232 +467,14 @@ cli:
   notes:
     - "CLI parity is checked against the actual generate/serve parser dest sets."
     - "The inventory tracks parser dest names, excluding argparse's implicit help action."
+    - "The refactored inference CLI is config-only: subcommands expose only --config, and any additional CLI input must use dotted override paths."
   generate:
     explicit_local_fields:
       - config
     expected_dests:
-      - VSA_sparsity
-      - boundary_ratio
-      - bsa_cdf_threshold
-      - bsa_chunk_k
-      - bsa_chunk_q
-      - bsa_sparsity
       - config
-      - disable_autocast
-      - dist_timeout
-      - distributed_executor_backend
-      - dit_config.prefix
-      - dit_config.quant_config
-      - dit_cpu_offload
-      - dit_layerwise_offload
-      - dit_precision
-      - dmd_denoising_steps
-      - embedded_cfg_scale
-      - enable_bsa
-      - enable_stage_verification
-      - enable_torch_compile
-      - flow_shift
-      - fps
-      - guidance_rescale
-      - guidance_scale
-      - height
-      - hsdp_replicate_dim
-      - hsdp_shard_dim
-      - image_encoder_cpu_offload
-      - image_encoder_precision
-      - image_path
-      - inference_mode
-      - init_weights_from_safetensors
-      - init_weights_from_safetensors_2
-      - lora_nickname
-      - lora_path
-      - lora_target_modules
-      - ltx2_initial_latent_path
-      - ltx2_vae_spatial_tile_overlap_in_pixels
-      - ltx2_vae_spatial_tile_size_in_pixels
-      - ltx2_vae_temporal_tile_overlap_in_frames
-      - ltx2_vae_temporal_tile_size_in_frames
-      - ltx2_vae_tiling
-      - master_port
-      - moba_config_path
-      - mode
-      - model_path
-      - negative_prompt
-      - num_cond_frames
-      - num_frames
-      - num_gpus
-      - num_inference_steps
-      - num_videos_per_prompt
-      - output_path
-      - output_type
-      - output_video_name
-      - override_pipeline_cls_name
-      - override_text_encoder_quant
-      - override_text_encoder_safetensors
-      - override_transformer_cls_name
-      - pin_cpu_memory
-      - pipeline_config_path
-      - preprocess.dataloader_num_workers
-      - preprocess.dataset_output_dir
-      - preprocess.dataset_path
-      - preprocess.dataset_type
-      - preprocess.do_temporal_sample
-      - preprocess.drop_short_ratio
-      - preprocess.flush_frequency
-      - preprocess.max_height
-      - preprocess.max_width
-      - preprocess.model_path
-      - preprocess.num_frames
-      - preprocess.preprocess_video_batch_size
-      - preprocess.samples_per_file
-      - preprocess.seed
-      - preprocess.speed_factor
-      - preprocess.train_fps
-      - preprocess.training_cfg_rate
-      - preprocess.video_length_tolerance_range
-      - preprocess.video_loader_type
-      - preprocess.with_audio
-      - prompt
-      - prompt_path
-      - prompt_txt
-      - refine_from
-      - return_frames
-      - return_trajectory_decoded
-      - return_trajectory_latents
-      - revision
-      - save_video
-      - seed
-      - sp_size
-      - spatial_refine_only
-      - t_thresh
-      - text_encoder_configs
-      - text_encoder_cpu_offload
-      - text_encoder_precisions
-      - torch_compile_kwargs
-      - transformer_quant
-      - tp_size
-      - trust_remote_code
-      - use_fsdp_inference
-      - vae_config.blend_num_frames
-      - vae_config.load_decoder
-      - vae_config.load_encoder
-      - vae_config.tile_sample_min_height
-      - vae_config.tile_sample_min_num_frames
-      - vae_config.tile_sample_min_width
-      - vae_config.tile_sample_stride_height
-      - vae_config.tile_sample_stride_num_frames
-      - vae_config.tile_sample_stride_width
-      - vae_config.use_parallel_tiling
-      - vae_config.use_temporal_tiling
-      - vae_config.use_tiling
-      - vae_cpu_offload
-      - vae_precision
-      - vae_sp
-      - vae_tiling
-      - video_path
-      - width
-      - workload_type
   serve:
     explicit_local_fields:
       - config
-      - host
-      - output_dir
-      - port
     expected_dests:
-      - VSA_sparsity
-      - bsa_cdf_threshold
-      - bsa_chunk_k
-      - bsa_chunk_q
-      - bsa_sparsity
       - config
-      - disable_autocast
-      - dist_timeout
-      - distributed_executor_backend
-      - dit_config.prefix
-      - dit_config.quant_config
-      - dit_cpu_offload
-      - dit_layerwise_offload
-      - dit_precision
-      - dmd_denoising_steps
-      - embedded_cfg_scale
-      - enable_bsa
-      - enable_stage_verification
-      - enable_torch_compile
-      - flow_shift
-      - host
-      - hsdp_replicate_dim
-      - hsdp_shard_dim
-      - image_encoder_cpu_offload
-      - image_encoder_precision
-      - inference_mode
-      - init_weights_from_safetensors
-      - init_weights_from_safetensors_2
-      - lora_nickname
-      - lora_path
-      - lora_target_modules
-      - ltx2_initial_latent_path
-      - ltx2_vae_spatial_tile_overlap_in_pixels
-      - ltx2_vae_spatial_tile_size_in_pixels
-      - ltx2_vae_temporal_tile_overlap_in_frames
-      - ltx2_vae_temporal_tile_size_in_frames
-      - ltx2_vae_tiling
-      - master_port
-      - mode
-      - model_path
-      - num_gpus
-      - output_dir
-      - output_type
-      - override_pipeline_cls_name
-      - override_text_encoder_quant
-      - override_text_encoder_safetensors
-      - override_transformer_cls_name
-      - pin_cpu_memory
-      - pipeline_config_path
-      - port
-      - preprocess.dataloader_num_workers
-      - preprocess.dataset_output_dir
-      - preprocess.dataset_path
-      - preprocess.dataset_type
-      - preprocess.do_temporal_sample
-      - preprocess.drop_short_ratio
-      - preprocess.flush_frequency
-      - preprocess.max_height
-      - preprocess.max_width
-      - preprocess.model_path
-      - preprocess.num_frames
-      - preprocess.preprocess_video_batch_size
-      - preprocess.samples_per_file
-      - preprocess.seed
-      - preprocess.speed_factor
-      - preprocess.train_fps
-      - preprocess.training_cfg_rate
-      - preprocess.video_length_tolerance_range
-      - preprocess.video_loader_type
-      - preprocess.with_audio
-      - prompt_txt
-      - revision
-      - sp_size
-      - text_encoder_cpu_offload
-      - text_encoder_precisions
-      - torch_compile_kwargs
-      - transformer_quant
-      - tp_size
-      - trust_remote_code
-      - use_fsdp_inference
-      - vae_config.blend_num_frames
-      - vae_config.load_decoder
-      - vae_config.load_encoder
-      - vae_config.tile_sample_min_height
-      - vae_config.tile_sample_min_num_frames
-      - vae_config.tile_sample_min_width
-      - vae_config.tile_sample_stride_height
-      - vae_config.tile_sample_stride_num_frames
-      - vae_config.tile_sample_stride_width
-      - vae_config.use_parallel_tiling
-      - vae_config.use_temporal_tiling
-      - vae_config.use_tiling
-      - vae_cpu_offload
-      - vae_precision
-      - vae_sp
-      - vae_tiling
-      - workload_type

docs/design/overview.md

@@ -12,7 +12,7 @@ FastVideo maps a Diffusers-style repo into a pipeline like this:
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
   weight name translation.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
-- `fastvideo/configs/sample/*`: default runtime sampling parameters.
+- `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipelines.
 - `fastvideo/pipelines/stages/*`: reusable pipeline stages.
 - `fastvideo/models/loader/*`: component loaders for Diffusers-style repos.
@@ -26,7 +26,7 @@ Minimal usage (from `examples/inference/basic/basic.py`):
 
 ```python
 from fastvideo import VideoGenerator
-from fastvideo.configs.sample import SamplingParam
+from fastvideo.api.sampling_param import SamplingParam
 
 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
 generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
@@ -49,8 +49,9 @@ runtime parameters consistent:
 - `fastvideo/configs/models/`: architecture definitions, layer shapes, and
   `param_names_mapping` rules for key renaming.
 - `fastvideo/configs/pipelines/`: pipeline wiring and required components.
-- `fastvideo/configs/sample/`: default sampling parameters (steps, frames,
-  guidance scale, resolution, fps).
+- `fastvideo/api/sampling_param.py`: sampling parameters (steps, frames,
+  guidance scale, resolution, fps). Defaults come from profiles in
+  `fastvideo/pipelines/basic/<family>/profiles.py`.
 - `fastvideo/registry.py`: unified registry for pipeline config + sampling
   defaults and model metadata resolution, defined via explicit
   `register_configs(...)` blocks (no separate dict registries).
@@ -142,7 +143,7 @@ How this maps to FastVideo:
 - `T5TokenizerFast` -> loaded via HF in `fastvideo/models/loader/`
 - `UniPCMultistepScheduler` -> loaded via Diffusers scheduler utilities
 - Pipeline defaults -> `fastvideo/configs/pipelines/wan.py`
-- Sampling defaults -> `fastvideo/configs/sample/wan.py`
+- Sampling defaults -> `fastvideo/pipelines/basic/wan/profiles.py`
 
 ## Pipeline system
 
@@ -167,11 +168,6 @@ How this maps to FastVideo:
 
 - Attention backends live in `fastvideo/attention/` and can be selected via
   `FASTVIDEO_ATTENTION_BACKEND`.
-- SageAttention3 is split into two selectable backends:
-  `SAGE_ATTN_THREE` for the regular upstream package and
-  `ATTN_QAT_INFER` for the FastVideoKernel-backed inference variant.
-- `ATTN_QAT_TRAIN` is a separate FastVideoKernel Triton backend for the QAT attention
-  path.
 - `LocalAttention` is used for cross-attention and most attention layers.
 - `DistributedAttention` is used for full-sequence self-attention in the DiT.
 - Tensor-parallel layers live in `fastvideo/layers/`.

docs/design/server_contracts/streaming.md

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

docs/design/training_architecture.md

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

docs/distillation/dmd.md

@@ -16,7 +16,8 @@ Both models are trained on **61×448×832** resolution but support generating vi
 First install [VSA](../attention/vsa/index.md). Set `MODEL_BASE` to your own model path and run:
 
 ```bash
-bash scripts/inference/v1_inference_wan_dmd.sh
+FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN \
+  fastvideo generate --config scripts/inference/inference_wan_VSA_DMD_1_3B.yaml
 ```
 
 ## 🗂️ Dataset
@@ -85,3 +86,25 @@ sbatch examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B.sh
 - Learning rate: 2e-5
 - Training steps: 3000 (~12 hours)
 - HSDP shard dim: 1
+
+## 🧭 Note on `real_score_guidance_scale`
+
+The teacher CFG used inside the DMD loss follows the DMD2 reference
+implementation and uses the parameterization
+
+```
+x = x_cond + w * (x_cond - x_uncond)
+```
+
+rather than the Ho & Salimans form `x_uncond + w * (x_cond - x_uncond)`. The
+two are mathematically equivalent up to a constant offset:
+
+| `real_score_guidance_scale` (`w`) | Equivalent standard CFG (`w + 1`) | Output                |
+|-----------------------------------|-----------------------------------|-----------------------|
+| `-1`                              | `0`                               | unconditional         |
+| `0`                               | `1`                               | conditional           |
+| `3.5` (default)                   | `4.5`                             | strong guidance       |
+
+So `real_score_guidance_scale` should be read as the **extra** guidance
+strength added on top of the conditional prediction. When porting values
+from a paper that uses the Ho & Salimans form, subtract 1.

docs/getting_started/installation.md

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

docs/getting_started/installation/gpu.md

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

docs/getting_started/installation/mps.md

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

docs/getting_started/v1_api.md

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

docs/inference/architecture.md

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

docs/inference/cli.md

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

docs/inference/configuration.md

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

docs/inference/gen3c.md

@@ -19,7 +19,7 @@
 - Install MoGe:
 
 ```bash
-pip install git+https://github.com/microsoft/MoGe.git
+uv pip install git+https://github.com/microsoft/MoGe.git
 ```
 
 - If you hit `ImportError: libGL.so.1` (common on Ubuntu/headless nodes), you can try installing OpenCV runtime libs:
@@ -89,7 +89,7 @@ GEN3C defaults in FastVideo:
 
 These values are defined in:
 
-- `fastvideo/configs/sample/gen3c.py`
+- `fastvideo/pipelines/basic/gen3c/profiles.py`
 - `fastvideo/configs/pipelines/gen3c.py`
 
 and align with the official GEN3C inference defaults in:

docs/inference/inference_quick_start.md

@@ -107,8 +107,6 @@ If you encounter CUDA out of memory errors:
   (single GPU) or `use_fsdp_inference=True` (multi-GPU)
 - Try a smaller model or use distilled versions
 - Use `num_gpus` > 1 if multiple GPUs are available
-- Try enabling FSDP inference with `use_fsdp_inference=True` (may slow down generation)
-- Try enabling DiT layerwise offload with `dit_layerwise_offload=True` (now only a few models support this, but may introduce less overhead than FSDP)
 
 ### Slow Generation
 

docs/inference/optimizations.md

@@ -21,8 +21,6 @@ This page describes the various options for speeding up generation times in Fast
 - Video Sparse Attention: `FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN`
 - Sage Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN`
 - Sage Attention 3: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN_THREE`
-- Attn QAT Infer: `FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER`
-- Attn QAT Train: `FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN`
 - Video MoBA Attention: `FASTVIDEO_ATTENTION_BACKEND=VMOBA_ATTN`
 - Sparse Linear Attention: `FASTVIDEO_ATTENTION_BACKEND=SLA_ATTN`
 - SageSLA Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_SLA_ATTN`
@@ -56,7 +54,7 @@ FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 We recommend always installing [Flash Attention 2](https://github.com/Dao-AILab/flash-attention):
 
 ```bash
-pip install flash-attn==2.7.4.post1 --no-build-isolation
+uv pip install flash-attn==2.7.4.post1 --no-build-isolation
 ```
 
 And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#flashattention-3-beta-release) by compiling it from source (takes about 10 minutes for me):
@@ -65,7 +63,7 @@ And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://git
 git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention
 
 cd hopper
-pip install ninja
+uv pip install ninja
 python setup.py install
 ```
 
@@ -100,19 +98,11 @@ To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please co
 ```bash
 git clone https://github.com/thu-ml/SageAttention.git
 cd sageattention
-python setup.py install  # or pip install -e .
+python setup.py install  # or uv pip install -e .
 ```
 
 ### Sage Attention 3
 
-FastVideo now exposes two SageAttention3-compatible backends with distinct
-environment variable values:
-
-- `SAGE_ATTN_THREE`: the regular upstream SageAttention3 backend imported from
-  the `sageattn3` package.
-- `ATTN_QAT_INFER`: the inference CUDA-kernel backend imported from the
-  in-repo `attn_qat_infer` package.
-
 **`SAGE_ATTN_THREE`**
 
 [SageAttention 3](https://github.com/thu-ml/SageAttention/tree/main/sageattention3_blackwell) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.
@@ -127,53 +117,6 @@ Note that Sage Attention 3 requires `python>=3.13`, `torch>=2.8.0`, `CUDA >=12.8
 
 To use Sage Attention 3 in FastVideo, follow the `README.md` in the linked repository to install the package from source.
 
-### Attn QAT Infer
-
-**`ATTN_QAT_INFER`**
-
-This backend uses the `attn_qat_infer` implementation that lives in the
-`fastvideo-kernel` repository alongside the `fastvideo_kernel` Triton kernels.
-Use this backend when you want to run the dedicated FP4 inference CUDA kernel
-directly during inference.
-
-For the full Attention QAT guide, including Wan 2.1 14B checkpoint download,
-example editing steps, training launchers, and troubleshooting, see
-[Attention QAT](../attention/attn_qat/index.md).
-
-This backend currently assumes access to the in-repo `fastvideo-kernel`
-checkout or an equivalent editable/source install that exposes:
-
-- `attn_qat_infer`
-
-Example:
-
-```python
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"
-```
-
-### QAT Attention
-
-**`ATTN_QAT_TRAIN`**
-
-This backend uses the FastVideoKernel Triton attention implementation from
-`fastvideo_kernel.triton_kernels.attn_qat_train`. Use it when you specifically
-want the training-oriented Triton attention path rather than the
-`attn_qat_infer` CUDA kernel path.
-
-The dedicated [Attention QAT](../attention/attn_qat/index.md) page covers when
-to use `ATTN_QAT_TRAIN` versus `ATTN_QAT_INFER`, the ready-made training
-launchers, and the end-to-end Wan 2.1 14B inference workflow.
-
-This backend currently assumes access to an install that exposes:
-
-- `fastvideo_kernel`
-
-Example:
-
-```python
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_TRAIN"
-```
-
 ### V-MoBA / SLA / SageSLA
 
 These backends are model-specific and require the corresponding kernels and

docs/utilities/debugging.md

@@ -27,8 +27,7 @@ Useful variables:
 - `FASTVIDEO_LOGGING_LEVEL`: `DEBUG`, `INFO`, `WARNING`, `ERROR`
 - `FASTVIDEO_STAGE_LOGGING`: print per-stage timings during pipeline execution
 - `FASTVIDEO_ATTENTION_BACKEND`: force an attention backend (for example
-  `TORCH_SDPA`, `FLASH_ATTN`, `SAGE_ATTN_THREE`, or
-  `ATTN_QAT_INFER`, or `ATTN_QAT_TRAIN`)
+  `TORCH_SDPA` or `FLASH_ATTN`)
 
 ## Common Failure Modes
 
@@ -53,11 +52,7 @@ If forcing a backend fails, verify optional dependencies are installed:
 - `VIDEO_SPARSE_ATTN`: `fastvideo-kernel`
 - `SLIDING_TILE_ATTN`: STA legacy workflow in
   `sta_do_not_delete` + `fastvideo-kernel`
-- `SAGE_ATTN`: SageAttention package
-- `SAGE_ATTN_THREE`: upstream `sageattn3` package
-- `ATTN_QAT_INFER`: `fastvideo-kernel` checkout/source install that exposes
-  `attn_qat_infer`
-- `ATTN_QAT_TRAIN`: `fastvideo-kernel` install exposing `fastvideo_kernel`
+- `SAGE_ATTN` / `SAGE_ATTN_THREE`: SageAttention packages
 
 As a fallback, use:
 

examples/distill/SFWan2.2-A14B/distill_dmd.sh

@@ -22,7 +22,7 @@ export NODE_RANK=$SLURM_PROCID
 nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
 export MASTER_ADDR=${nodes[0]}
 export TOKENIZERS_PARALLELISM=false
-export WANDB_API_KEY=YOUR_WANDB_API_KEY
+export WANDB_API_KEY="2f25ad37933894dbf0966c838c0b8494987f9f2f"
 # export WANDB_API_KEY='your_wandb_api_key_here'
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online

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

@@ -4,12 +4,12 @@ These are end-to-end example scripts for distilling Wan2.1 T2V 1.3B model using
 ### 0. Make sure you have installed VSA
 
 ```bash
-pip install vsa
+uv pip install vsa
 ```
 
 ### 1. Download dataset:
 ```bash
-bash examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh
+bash examples/distill/Wan-Syn-480P/download_dataset.sh
 ```
 
 ### 2. Configure and run distillation:

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
-mkdir -p data
-python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "data/Wan-Syn_77x448x832_600k" --repo_type "dataset"
+
+python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"

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

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

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

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

examples/inference/basic/basic.py

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

examples/inference/basic/basic_cosmos2_5_i2w.py

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

examples/inference/basic/basic_cosmos2_5_t2w.py

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

examples/inference/basic/basic_cosmos2_5_v2w.py

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

examples/inference/basic/basic_dmd.py

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

examples/inference/basic/basic_gen3c.py

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

examples/inference/basic/basic_hy15.py

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

examples/inference/basic/basic_hy15_1080p.py

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

examples/inference/basic/basic_lingbotworld_base_cam.py

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

examples/inference/basic/basic_mps.py

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

examples/inference/basic/basic_ray.py

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

examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py

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

examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py

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

examples/inference/basic/basic_stable_audio.py

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

examples/inference/basic/basic_stable_audio_a2a.py

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

examples/inference/basic/basic_stable_audio_inpaint.py

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

examples/inference/basic/basic_stable_audio_small.py

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

examples/inference/basic/basic_wan2_2.py

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

examples/inference/basic/basic_wan2_2_Fun.py

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

examples/inference/basic/basic_wan2_2_i2v.py

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

examples/inference/gradio/local/gradio_local_demo.py

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

examples/inference/gradio/serving/gradio_frontend.py

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

examples/inference/gradio/serving/ray_serve_backend.py

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

examples/inference/lora/wan_lora_inference.py

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

examples/inference/lora/wan_lora_inference_from_ckpt.py

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

examples/inference/optimizations/README.md

@@ -1,79 +1,5 @@
 # Optimization Examples
 
-## Wan 2.1 QAT Attention 14B Inference
-
-Use these files for Wan 2.1 14B inference with the `ATTN_QAT_INFER` backend:
-
-- `examples/inference/optimizations/download_14B_qat.sh`
-- `examples/inference/optimizations/attn_qat_inference_example.py`
-
-### 1. Download the 14B QAT checkpoint
-
-The helper script downloads the QAT safetensors from
-`FastVideo/14B_qat_400` into `checkpoints/14B_qat_400` by default.
-
-Prerequisites:
-
-- `huggingface_hub` installed, for example: `uv pip install huggingface_hub`
-- access to the model repo if it is private or gated: `huggingface-cli login`
-
-Run:
-
 ```bash
-bash examples/inference/optimizations/download_14B_qat.sh
+python examples/inference/optimizations/attention_example.py
 ```
-
-To download into a custom directory, pass it as the first argument:
-
-```bash
-bash examples/inference/optimizations/download_14B_qat.sh /path/to/14B_qat_400
-```
-
-### 2. Edit the inference example for Wan 2.1 14B
-
-Open `examples/inference/optimizations/attn_qat_inference_example.py` and
-update these two values:
-
-1. Change the base model from `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` to
-   `Wan-AI/Wan2.1-T2V-14B-Diffusers`.
-2. Replace the placeholder
-   `init_weights_from_safetensors="safetensors_path"` with the directory that
-   contains the downloaded `.safetensors` files.
-
-Example:
-
-```python
-generator = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
-    num_gpus=1,
-    use_fsdp_inference=True,
-    dit_cpu_offload=False,
-    vae_cpu_offload=False,
-    text_encoder_cpu_offload=True,
-    pin_cpu_memory=False,
-    init_weights_from_safetensors="checkpoints/14B_qat_400",
-)
-```
-
-The script already sets:
-
-```python
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"
-```
-
-### 3. Run the example
-
-```bash
-python examples/inference/optimizations/attn_qat_inference_example.py
-```
-
-The generated videos are written to `video_samples/` by default.
-
-### Notes
-
-- `ATTN_QAT_INFER` requires the in-repo `fastvideo-kernel` build to expose the
-  `attn_qat_infer` package.
-- If you have not built the kernel yet, run `cd fastvideo-kernel && ./build.sh`
-  first.
-- If you keep the example on the `1.3B` base model while loading the 14B QAT
-  weights, the model/config will not match.

examples/inference/optimizations/attn_qat_inference_example.py

@@ -1,54 +0,0 @@
-from fastvideo import VideoGenerator
-import os
-from pathlib import Path
-# from fastvideo.configs.sample import SamplingParam
-
-OUTPUT_PATH = "video_samples"
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"
-
-CHECKPOINT_PATH = Path(__file__).parent.parent.parent 
-
-def main():
-    # FastVideo will automatically use the optimal default arguments for the
-    # model.
-    # If a local path is provided, FastVideo will make a best effort
-    # attempt to identify the optimal arguments.
-    generator = VideoGenerator.from_pretrained(
-        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-        # FastVideo will automatically handle distributed setup
-        num_gpus=1,
-        use_fsdp_inference=True,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=False, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
-        # image_encoder_cpu_offload=False,
-        # Load custom weights from checkpoint 
-        init_weights_from_safetensors="safetensors_path"
-    )
-
-    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
-    # sampling_param.num_frames = 45
-    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
-    # Generate videos with the same simple API, regardless of GPU count
-    prompt = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-        "wide with interest. The playful yet serene atmosphere is complemented by soft "
-        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
-    )
-    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True)
-    # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
-
-    # Generate another video with a different prompt, without reloading the
-    # model!
-    prompt2 = (
-        "A majestic lion strides across the golden savanna, its powerful frame "
-        "glistening under the warm afternoon sun. The tall grass ripples gently in "
-        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-        "cinematic.")
-    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)
-
-
-if __name__ == "__main__":
-    main()

examples/inference/optimizations/download_14B_qat.sh

@@ -1,58 +0,0 @@
-#!/usr/bin/env bash
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../../.." && pwd)"
-
-HF_REPO_ID="${HF_REPO_ID:-FastVideo/14B_qat_400}"
-HF_REVISION="${HF_REVISION:-main}"
-LOCAL_DIR="${1:-${REPO_ROOT}/checkpoints/14B_qat_400}"
-PYTHON_BIN="${PYTHON:-python}"
-
-if ! command -v "${PYTHON_BIN}" >/dev/null 2>&1; then
-    echo "Python executable not found: ${PYTHON_BIN}" >&2
-    exit 1
-fi
-
-if ! "${PYTHON_BIN}" -c "import huggingface_hub" >/dev/null 2>&1; then
-    echo "Missing dependency: huggingface_hub" >&2
-    echo "Install it with: uv pip install huggingface_hub" >&2
-    exit 1
-fi
-
-mkdir -p "${LOCAL_DIR}"
-
-echo "Downloading ${HF_REPO_ID}@${HF_REVISION}"
-echo "Local directory: ${LOCAL_DIR}"
-
-"${PYTHON_BIN}" -c '
-import argparse
-from huggingface_hub import snapshot_download
-
-parser = argparse.ArgumentParser()
-parser.add_argument("--repo-id", required=True)
-parser.add_argument("--revision", required=True)
-parser.add_argument("--local-dir", required=True)
-args = parser.parse_args()
-
-snapshot_download(
-    repo_id=args.repo_id,
-    revision=args.revision,
-    repo_type="model",
-    local_dir=args.local_dir,
-    local_dir_use_symlinks=False,
-    resume_download=True,
-)
-' \
-    --repo-id "${HF_REPO_ID}" \
-    --revision "${HF_REVISION}" \
-    --local-dir "${LOCAL_DIR}"
-
-echo
-echo "Download complete."
-echo "Use this in your inference script:"
-echo "init_weights_from_safetensors=\"${LOCAL_DIR}\""
-echo
-echo "If the repo is private or gated, make sure you are logged in with:"
-echo "huggingface-cli login"

examples/inference/optimizations/quantization_example.py

@@ -1,88 +0,0 @@
-import torch
-from fastvideo import VideoGenerator
-from fastvideo.configs.pipelines.base import PipelineConfig
-
-OUTPUT_PATH = "video_samples"
-
-
-def main():
-    print("=== FP4 Quantization Video Generation Example ===")
-
-    if not torch.cuda.is_available():
-        print("Warning: CUDA not available. FP4 quantization requires GPU.")
-        return
-    
-    gpu_capability = torch.cuda.get_device_capability()
-    if gpu_capability[0] < 9:  # H100 and newer
-        print(f"Warning: GPU capability {gpu_capability} may not support FP4. Recommended: 9.0+")
-    
-    print(f"GPU: {torch.cuda.get_device_name()}")
-    print(f"GPU Capability: {gpu_capability}")
-    
-    model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-    # model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
-    pipeline_config = PipelineConfig.from_pretrained(model_id)
-    pipeline_config.dit_precision = "bf16"
-
-    print("\nLoading model with FP4 quantization...")
-    generator = VideoGenerator.from_pretrained(
-        model_id,
-        pipeline_config=pipeline_config,
-        num_gpus=1,
-        use_fsdp_inference=True,
-        transformer_quant="fp4",
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=False,
-    )
-    
-    print("FP4 configuration applied. Generating videos...")
-    
-    print("\n=== Generating Video with FP4 Quantization ===")
-    
-    prompt1 = (
-        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
-        "wide with interest. The playful yet serene atmosphere is complemented by soft "
-        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
-    )
-    
-    print(f"Prompt: {prompt1}")
-    print("Generating video...")
-    
-    try:
-        video1 = generator.generate_video(
-            prompt1,
-            output_path=OUTPUT_PATH,
-            save_video=True,
-        )
-        print("✓ First video generated successfully with FP4 quantization!")
-        
-        # # Generate a second video to show the model can be reused
-        prompt2 = (
-            "A majestic lion strides across the golden savanna, its powerful frame "
-            "glistening under the warm afternoon sun. The tall grass ripples gently in "
-            "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-            "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-            "cinematic."
-        )
-        
-        print(f"\nGenerating second video...")
-        print(f"Prompt: {prompt2}")
-        
-        video2 = generator.generate_video(
-            prompt2,
-            output_path=OUTPUT_PATH,
-            save_video=True,
-        )
-        print("✓ Second video generated successfully with FP4 quantization!")
-        
-    except Exception as e:
-        print(f"Error during video generation: {e}")
-        return
-    
-    print(f"Videos saved to: {OUTPUT_PATH}")
-
-
-if __name__ == "__main__":
-    main() 

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

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

examples/train/configs/overfit_cosmos25_t2w.yaml

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

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh

@@ -1,47 +1,27 @@
 #!/bin/bash
-#SBATCH --job-name=wan_t2v_1.3B_finetune
-#SBATCH --partition=all
-#SBATCH --nodes=1
-#SBATCH --gres=gpu:4
-#SBATCH --ntasks-per-node=1
-#SBATCH --output=logs/wan_t2v_1.3B_finetune.out
-#SBATCH --error=logs/wan_t2v_1.3B_finetune.err
-
-source .venv/bin/activate
-
 
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
-# export TRITON_PRINT_AUTOTUNING=1  # to print the best config
-export WANDB_API_KEY=YOUR_WANDB_API_KEY
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR=data/Wan-Syn_77x448x832_600k
-VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json"
-NUM_GPUS=1
+DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+NUM_GPUS=4
 # export CUDA_VISIBLE_DEVICES=4,5
 
-set -euo pipefail
-
-# ---- torchrun rendezvous (multi-node) ----
-# Launch ONE torchrun per node (via srun) and let torchrun spawn 4 workers per node.
-MASTER_ADDR="$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)"
-MASTER_PORT="${MASTER_PORT:-29500}"
-export MASTER_ADDR MASTER_PORT
-
 
 # Training arguments
 training_args=(
-  --tracker_project_name "wan_t2v_finetune_qat"
-  --output_dir "checkpoints/wan_t2v_finetune_1.3B_77"
-  --max_train_steps 4000
+  --tracker_project_name "wan_t2v_finetune"
+  --output_dir "checkpoints/wan_t2v_finetune"
+  --max_train_steps 5000
   --train_batch_size 1
   --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
+  --gradient_accumulation_steps 8
   --num_latent_t 20
-  --num_height 448  
+  --num_height 480
   --num_width 832
   --num_frames 77
   --enable_gradient_checkpointing_type "full"
@@ -50,7 +30,7 @@ training_args=(
 # Parallel arguments
 parallel_args=(
   --num_gpus $NUM_GPUS 
-  --sp_size 1
+  --sp_size $NUM_GPUS 
   --tp_size 1
   --hsdp_replicate_dim 1
   --hsdp_shard_dim $NUM_GPUS
@@ -65,7 +45,7 @@ model_args=(
 # Dataset arguments
 dataset_args=(
   --data_path $DATA_DIR
-  --dataloader_num_workers 4
+  --dataloader_num_workers 1
 )
 
 # Validation arguments
@@ -74,16 +54,16 @@ validation_args=(
   --validation_dataset_file $VALIDATION_DATASET_FILE
   --validation_steps 200
   --validation_sampling_steps "50" 
-  --validation_guidance_scale "5.0"
+  --validation_guidance_scale "3.0"
 )
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-6
+  --learning_rate 5e-5
   --mixed_precision "bf16"
   --weight_only_checkpointing_steps 1000
   --training_state_checkpointing_steps 1000
-  --weight_decay 0.01
+  --weight_decay 1e-4
   --max_grad_norm 1.0
 )
 
@@ -92,24 +72,23 @@ miscellaneous_args=(
   --inference_mode False
   --checkpoints_total_limit 3
   --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
   --dit_precision "fp32"
+  --num_euler_timesteps 50
   --ema_start_step 0
-  --flow_shift 5
-  --seed 1000
+  --enable_gradient_checkpointing_type "full"
+  # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500"
 )
 
-srun --nodes="$SLURM_NNODES" --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \
-  torchrun \
-    --nnodes "$SLURM_NNODES" \
-    --nproc_per_node 4 \
-    --rdzv_backend c10d \
-    --rdzv_endpoint "${MASTER_ADDR}:${MASTER_PORT}" \
-    --rdzv_id "$SLURM_JOB_ID" \
-      fastvideo/training/wan_training_pipeline.py \
-      "${parallel_args[@]}" \
-      "${model_args[@]}" \
-      "${dataset_args[@]}" \
-      "${training_args[@]}" \
-      "${optimizer_args[@]}" \
-      "${validation_args[@]}" \
-      "${miscellaneous_args[@]}"
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_qat_attn.sh

@@ -1,124 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_16
-#SBATCH --partition=all
-#SBATCH --nodes=4
-#SBATCH --gres=gpu:4
-#SBATCH --ntasks-per-node=1
-#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_16.out
-#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_16.err
-
-source .venv/bin/activate
-
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-
-# export TRITON_PRINT_AUTOTUNING=1  # to print the best config
-export WANDB_API_KEY=YOUR_WANDB_API_KEY
-# Use node-local Triton cache to avoid stale file handle errors on shared filesystems
-export TRITON_CACHE_DIR="/tmp/triton_cache_${SLURM_JOB_ID}_${SLURM_NODEID}"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR=YOUR_DATA_DIR
-VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json"
-NUM_GPUS=16
-# export CUDA_VISIBLE_DEVICES=4,5
-
-set -euo pipefail
-
-# ---- torchrun rendezvous (multi-node) ----
-# 1. Get the hostname of the first node (Master)
-nodes=( $( scontrol show hostnames $SLURM_JOB_NODELIST ) )
-nodes_array=($nodes)
-head_node=${nodes_array[0]}
-MASTER_ADDR=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address)
-MASTER_PORT=29500
-
-# 2. Get the node count automatically
-NNODES=$SLURM_NNODES
-GPUS_PER_NODE=$SLURM_GPUS_ON_NODE
-NUM_GPUS=$((NNODES * GPUS_PER_NODE))
-
-echo "MASTER_ADDR=$MASTER_ADDR MASTER_PORT=$MASTER_PORT NNODES=$NNODES"
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_finetune_qat"
-  --output_dir "checkpoints/wan_1.3B_t2v_finetune_qat"
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 20
-  --num_height 448  
-  --num_width 832
-  --num_frames 77
-  --enable_gradient_checkpointing_type "full" # if OOM enable this
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS 
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation 
-  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 200
-  --validation_sampling_steps "50" 
-  --validation_guidance_scale "5.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-6
-  --mixed_precision "bf16"
-  --weight_only_checkpointing_steps 200
-  --training_state_checkpointing_steps 200
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --dit_precision "fp32"
-  --ema_start_step 0
-  --flow_shift 1
-  --seed 1000
-)
-
-srun torchrun \
-    --nnodes $NNODES \
-    --nproc_per_node $GPUS_PER_NODE \
-    --node_rank $SLURM_PROCID \
-    --rdzv_backend c10d \
-    --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
-      fastvideo/training/wan_training_pipeline.py \
-      "${parallel_args[@]}" \
-      "${model_args[@]}" \
-      "${dataset_args[@]}" \
-      "${training_args[@]}" \
-      "${optimizer_args[@]}" \
-      "${validation_args[@]}" \
-      "${miscellaneous_args[@]}"

examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json

@@ -1,131 +1,31 @@
 {
   "data": [
     {
-      "caption": "In the video, a woman is elegantly showcasing her earrings, bringing attention to their intricate design with a gentle touch of her fingers. She is bathed in ambient purple and pink lighting, which casts a soft glow on her delicate features and enhances the vivid tones of her lipstick and eye makeup. Her hair is styled to frame her face smoothly, emphasizing the contours of her jawline and cheekbones. The background features a blurred neon light, adding an artistic and modern touch to the overall aesthetic.",
-      "video_path": "Fashion/mixkit-face-of-an-elegant-and-captivating-woman-41914_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 50,
+      "height": 480,
       "width": 832,
       "num_frames": 77
     },
     {
-      "caption": "In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.",
-      "video_path": "Man/mixkit-a-rancher-riding-a-horse-at-sunset-1143_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
+      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 50,
+      "height": 480,
       "width": 832,
       "num_frames": 77
     },
     {
-      "caption": "In a dimly lit, eerie setting, a mysterious pink bottle labeled \"Authentic 100% organic POISON\" sits prominently in the foreground, casting a menacing aura. The bottle is accentuated by green fog, which swirls lightly around it, enhancing its sinister allure. Behind it, a shadowy golden bottle adorned with a spider emblem subtly emerges, adding an extra layer of mystery to the scene. Dim candles provide faint, flickering light, which complements the dark atmosphere, making the setting ideal for an illusion of hidden dangers.",
-      "video_path": "smoke/mixkit-poison-in-halloween-ritual-33879_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
+      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 50,
+      "height": 480,
       "width": 832,
       "num_frames": 77
-    },
-    {
-      "caption": "The video opens with a tranquil scene in the heart of a dense forest, emphasizing two large, textured tree trunks in the foreground framing the view. Sunlight filters through the canopy above, casting intricate patterns of light and shadow on the trees and the ground. Between the tree trunks, a clear view of a calm, muddy river unfolds, its surface shimmering under the gentle sunlight. The riverbank is decorated with a variety of small bushes and vibrant foliage, subtly transitioning into the deep greens of tall, leafy plants. In the background, the dense forest looms, filled with dark, towering trees, their branches intertwining to form an intricate canopy. The scene is bathed in the soft glow of the sun, creating a serene and picturesque setting. Occasional sunbeams pierce through the foliage, adding a magical aura to the landscape. The vibrant reds and oranges of the smaller plants add contrast, bringing warmth to the earthy tones of the scenery. Overall, this harmonious blend of natural elements creates a peaceful and idyllic forest setting.",
-      "video_path": "forest/mixkit-view-of-a-river-between-two-old-trees-560_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the video, a martial artist dressed in a traditional white uniform with a black belt demonstrates a series of precise movements against a stark black background. The individual gracefully transitions between stances, embodying a sense of focused discipline and control. Each motion is executed with a deliberate pace, showcasing the fluidity of martial arts techniques. The soft lighting creates subtle highlights on the uniform, adding depth to the figure as it moves. The practitioner begins with an open-hand pose, feet firmly grounded, gradually shifting to a powerful forward punch. The fluidity of the sequence displays a mastery of balance and poise. Every trajectory of the limbs is precise and deliberate, capturing the elegance and strength of martial arts. The serene, isolated setting enhances the intensity and concentration of the practitioner. This visual presentation is an elegant interplay of motion and stillness, displaying the art form's discipline and grace.",
-      "video_path": "Man/mixkit-a-young-man-practicing-his-karate-moves-49635_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A tranquil coastal scene unfolds with a drone's aerial view capturing a serene beach landscape. The camera glides over a quiet stretch of sandy shoreline, where gentle waves kiss the shore under a clear blue sky. Nestled amidst lush palm trees are a series of traditional thatched-roof huts, their earthy tones blending harmoniously with the natural surroundings. The sandy beach stretches endlessly, bordered by the rhythmic dance of ocean waves on one side and verdant greenery on the other. A pair of white umbrellas is set up on the sand, suggesting a place to relax and enjoy the sun. In the distance, two small human figures can be seen walking leisurely along the water's edge, leaving faint footprints behind them. The scene exudes a calm and inviting atmosphere, with the soft rustle of palm leaves and the whisper of the ocean breeze almost audible. The overall composition is a captivating blend of nature's tranquility and architectural simplicity. This picturesque setting invites viewers to imagine themselves steps away from this idyllic coastal escape.",
-      "video_path": "beach/mixkit-sunny-beach-in-a-dynamic-shot-from-a-drone-44383_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A lone figure stands on a large, moss-covered rock, surrounded by the soft rush of a nearby stream. The figure is wearing white sneakers and shorts, with a plaid shirt that hangs loosely in the breeze. The lighting creates dramatic shadows, enhancing the textures of the rock and the subtle movement of the water below. In the background, a waterfall cascades into the stream, completing this tranquil and serene nature scene.",
-      "video_path": "forest/mixkit-woman-standing-in-front-of-waterfall-559_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In an industrial setting, a person leans casually against a railing, exuding a sense of confidence and composure. They are wearing a striking outfit, consisting of a vibrant, patterned jacket over a simple white crop top, creating a bold contrast. The atmosphere is infused with warm, ambient lighting that casts soft shadows on the concrete walls and metallic surfaces. Intricate wiring and pipes form an intricate backdrop, enhancing the urban aesthetic. Their relaxed posture and direct, engaging gaze suggest a sense of ease in this industrial environment. This scene encapsulates a blend of modern fashion and gritty, urban architecture, creating a visually compelling narrative.",
-      "video_path": "Fashion/mixkit-portrait-of-a-hipster-woman-walking-down-a-stairs-1297_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A man is energetically stretching in an open-air setting, surrounded by rows of vibrant red seats that suggest an amphitheater or outdoor venue. He wears a sleeveless black shirt layered with a hooded vest, emphasizing his athletic build as he engages in a warm-up routine. Behind him, the striking modern architecture of the building features geometric panels, with large sections of glass and overlapping metallic beams creating a dynamic backdrop. The scene captures the contrast between his focused movements and the static, bold design of the structure, while the surrounding greenery adds a touch of nature to the environment. The overall atmosphere is one of preparation and anticipation, with the man appearing determined and ready for an upcoming event or performance.",
-      "video_path": "Sport/mixkit-man-doing-arm-stretches-595_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young woman is seated on the floor in front of a plush, beige tufted couch, fully engrossed in sorting through a stack of papers. Her dark hair falls loosely past her shoulders, and she wears a green plaid shirt, contributing to the casual yet focused atmosphere. She gently places the papers onto a small round white table, occasionally lifting individual sheets to examine them more closely. Her expression shifts subtly, reflecting concentration and contemplation as she processes the information on the pages. Two small, round nested tables hold her documents, along with a small plant in a gray pot, adding a touch of greenery to the scene. The background features a dark paneled wall, creating a contrasting backdrop for the light-colored furniture. The setting is tranquil and organized, the couch and tables arranged symmetrically, conveying a sense of harmony. A calculator rests on the smaller table, hinting at a task involving calculations or budgeting.",
-      "video_path": "Woman/mixkit-frustrated-woman-throws-paperwork-on-the-floor-4526_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A heavily rusted metal gate stands firmly locked, with two vertical bars joined by a thick, old chain that loops elegantly around them. The chain's texture is coarse and rugged, its surface reflecting varying shades of orange and brown, indicative of years exposed to the elements. At the heart of the chain, a black iron padlock, slightly worn yet imposing, secures the gate, its curves and edges smooth against the aged links. The gate's metalwork is outlined by a backdrop of soft, blurred greenery, suggesting a serene and isolated location beyond the barrier. Tall trees rise in the distance, their trunks and leaves creating a lush, forest-like setting that contrasts with the gate's severe rust. A pathway leads away from the gate, its surface uneven with patches of moss and weathered stone visible in the soft focus, inviting yet inaccessible. The ambiance is quiet and mysterious, with a sense of abandonment hanging subtly in the air, evoking curiosity about what lies beyond. Shadows play across the gate, cast by branches swaying gently in the breeze, adding to the dynamic interaction of light and texture. This scene, rich in detail and atmosphere, captures the viewer's imagination, evoking both the allure of the forbidden and the beauty of decay.",
-      "video_path": "forest/mixkit-rusty-fence-with-a-chain-of-a-property-in-nature-5294_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a serene and softly lit yoga studio, three individuals engage in a yoga session, each performing an upward-facing stretch. The central figure is a woman with shoulder-length brown hair, dressed in a light cropped top and green leggings, her posture reflecting grace and concentration. To her right, another participant, a woman in a purple outfit, mirrors the pose with equal poise. On her left, a person with a bun focuses intently, supported slightly by yoga blocks beneath their hands. The warm-colored wooden floor contrasts soothingly with the soft pastel mural on the back wall, featuring an abstract design and partial visage of a serene face. Natural light floods the space from a large window on the right, where lush greens peek through, adding an element of tranquility. In the corner of the room, a collection of meditation instruments, including a gong and a Buddha statue, subtly frame the peaceful setting. The mood is calm yet focused, as all three participants are deeply engaged in their practice. The scene combines elements of balance, harmony, and a shared journey towards mindfulness. This depiction captures the essence of a yoga session that blends personal growth with collective experience.",
-      "video_path": "People/mixkit-small-group-of-people-doing-yoga-together-43730_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the deep blue expanse of the ocean, two dolphins glide effortlessly, their sleek bodies reflecting the sunlight filtering through the water. The prominent shadows and caustics create a shimmering effect on their skin, capturing the beauty of their natural habitat. Each dolphin moves with a fluid grace, occasionally interacting with gentle nudges, showcasing their playful and social nature. The scene is vibrant and dynamic, with the clear blue background accentuating the dolphins' movements, making it an ideal subject for AI recreation.",
-      "video_path": "sea/mixkit-dolphins-underwater-4133_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the video, a young woman stands against a vibrant graffiti-covered wall, deeply engrossed in her smartphone. Her expression reflects a mix of focus and subtle satisfaction as she interacts with the screen. She wears a black floral-patterned top, which contrasts with the bright, abstract shapes and bold colors of the mural behind her. As she continues to engage with her phone, a series of like count notifications appear on the screen, indicating a growing online appreciation. The wall behind her features a striking mix of geometric and organic shapes, including swirls of teal, orange, and black, with large humanoid figures in a pop-art style. Her long, light-brown hair frames her face, adding a calm, composed aura amidst the lively backdrop. The video captures a blend of contemporary digital interaction and expressive urban art, creating a dynamic yet harmonious scene.",
-      "video_path": "Girl/mixkit-girl-looking-at-the-likes-in-her-post-4914_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young mother and her baby sit comfortably on a bed, surrounded by an inviting, cozy atmosphere. The woman, wearing a sleeveless top and jeans, is gently engaging with the baby, who is dressed in an adorable animal-print onesie. The child is seated on the bed with colorful toys scattered around, including a plush toy and a board book. The warm glow from a hanging lamp casts a soft light on them, enhancing the serene environment. Pillows are propped up against the headboard, providing a cushioned backdrop as the mother leans slightly over to interact with the baby. A small bottle is visible beside her, suggesting a nurturing setting. Her hand gestures animatedly as she holds up a soft, white cushion with red and blue accents, likely stimulating the baby\u2019s curiosity. Their shared moment is filled with affection and joy, a perfect snapshot of familial bonding.",
-      "video_path": "Baby/mixkit-loving-mother-and-her-baby-playing-with-soft-toys-49966_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young girl with long brown hair sits at a round wooden table, engrossed in working on her laptop. The laptop screen is a vivid green, suggesting a green screen effect is in use. To her left, a doll dressed in a yellow and white outfit is casually laid on top of some books, adding a playful and innocent touch to the scene. The setting is cozy, with sheer curtains in the background allowing soft natural light to spill into the room. The girl's posture and focused attention on the laptop suggest she is either playing a game or learning something new. This serene and domestic atmosphere is complemented by the slight blur of a dark couch in the foreground, framing the focused activity of the child.",
-      "video_path": "Girl/mixkit-little-girl-doing-homework-on-a-laptop-4757_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 448,
-      "width": 832,
-      "num_frames": 77
-  }  ]
+    }
+  ]
 }

examples/training/finetune/wan_t2v_14B/finetune_t2v_qat_attn.sh

@@ -1,123 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=wan_t2v_1.3B_finetune_qat_16
-#SBATCH --partition=main
-#SBATCH --nodes=4
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=logs/wan_t2v_1.3B_finetune_qat_16.out
-#SBATCH --error=logs/wan_t2v_1.3B_finetune_qat_16.err
-#SBATCH --exclusive
-
-source ~/conda/miniconda/bin/activate
-conda activate matthew-fv
-
-# Basic Info
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-# export TRITON_PRINT_AUTOTUNING=1  # to print the best config
-export WANDB_API_KEY=YOUR_WANDB_API_KEY
-MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
-DATA_DIR=YOUR_DATA_DIR
-VALIDATION_DATASET_FILE="examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json"
-NUM_GPUS_PER_NODE=8
-TOTAL_GPUS=$((NUM_GPUS_PER_NODE * SLURM_JOB_NUM_NODES))
-# export CUDA_VISIBLE_DEVICES=4,5
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_finetune_qat"
-  --output_dir "checkpoints/wan_14B_t2v_finetune_qat"
-  --max_train_steps 4000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 20
-  --num_height 768 
-  --num_width 1280
-  --num_frames 77
-  --enable_gradient_checkpointing_type "full" # if OOM enable this
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $TOTAL_GPUS 
-  --sp_size 4
-  --tp_size 1
-  --hsdp_replicate_dim 4
-  --hsdp_shard_dim 8
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR"
-  --dataloader_num_workers 4
-)
-
-# Validation arguments
-validation_args=(
-  # --log_validation 
-  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 200
-  --validation_sampling_steps "50" 
-  --validation_guidance_scale "5.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-6
-  --mixed_precision "bf16"
-  --weight_only_checkpointing_steps 200
-  --training_state_checkpointing_steps 200
-  --weight_decay 0.01
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --dit_precision "fp32"
-  --ema_start_step 0
-  --flow_shift 5
-  --seed 1000
-)
-
-srun torchrun \
---nnodes $SLURM_JOB_NUM_NODES \
---nproc_per_node $NUM_GPUS_PER_NODE \
---node_rank $SLURM_PROCID \
---rdzv_backend=c10d \
---rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/wan_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/finetune/wan_t2v_14B/validation.json

@@ -1,131 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "In the video, a woman is elegantly showcasing her earrings, bringing attention to their intricate design with a gentle touch of her fingers. She is bathed in ambient purple and pink lighting, which casts a soft glow on her delicate features and enhances the vivid tones of her lipstick and eye makeup. Her hair is styled to frame her face smoothly, emphasizing the contours of her jawline and cheekbones. The background features a blurred neon light, adding an artistic and modern touch to the overall aesthetic.",
-      "video_path": "Fashion/mixkit-face-of-an-elegant-and-captivating-woman-41914_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the video, a lone rider guides a majestic horse across an expansive, open field as the sun sets in the background. The rider, dressed in a classic blue shirt and wide-brimmed hat, sits confidently in the saddle, silhouetted against the warm glow of the evening sky. The horse moves gracefully, its mane and tail flowing with each step, creating a sense of harmony between horse and rider. Surrounding the pair, towering trees form a natural border, their leaves gently rustling in the breeze. The shadows lengthen on the ground, accentuating the serene and timeless feel of the scene. The distant hills and wooden fences frame the horizon, adding depth to the tranquil landscape. A few horses graze peacefully in the background, blending into the pastoral setting. The overall ambiance evokes a sense of calmness and quietude, capturing a perfect moment in the golden light of dusk.",
-      "video_path": "Man/mixkit-a-rancher-riding-a-horse-at-sunset-1143_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a dimly lit, eerie setting, a mysterious pink bottle labeled \"Authentic 100% organic POISON\" sits prominently in the foreground, casting a menacing aura. The bottle is accentuated by green fog, which swirls lightly around it, enhancing its sinister allure. Behind it, a shadowy golden bottle adorned with a spider emblem subtly emerges, adding an extra layer of mystery to the scene. Dim candles provide faint, flickering light, which complements the dark atmosphere, making the setting ideal for an illusion of hidden dangers.",
-      "video_path": "smoke/mixkit-poison-in-halloween-ritual-33879_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video opens with a tranquil scene in the heart of a dense forest, emphasizing two large, textured tree trunks in the foreground framing the view. Sunlight filters through the canopy above, casting intricate patterns of light and shadow on the trees and the ground. Between the tree trunks, a clear view of a calm, muddy river unfolds, its surface shimmering under the gentle sunlight. The riverbank is decorated with a variety of small bushes and vibrant foliage, subtly transitioning into the deep greens of tall, leafy plants. In the background, the dense forest looms, filled with dark, towering trees, their branches intertwining to form an intricate canopy. The scene is bathed in the soft glow of the sun, creating a serene and picturesque setting. Occasional sunbeams pierce through the foliage, adding a magical aura to the landscape. The vibrant reds and oranges of the smaller plants add contrast, bringing warmth to the earthy tones of the scenery. Overall, this harmonious blend of natural elements creates a peaceful and idyllic forest setting.",
-      "video_path": "forest/mixkit-view-of-a-river-between-two-old-trees-560_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the video, a martial artist dressed in a traditional white uniform with a black belt demonstrates a series of precise movements against a stark black background. The individual gracefully transitions between stances, embodying a sense of focused discipline and control. Each motion is executed with a deliberate pace, showcasing the fluidity of martial arts techniques. The soft lighting creates subtle highlights on the uniform, adding depth to the figure as it moves. The practitioner begins with an open-hand pose, feet firmly grounded, gradually shifting to a powerful forward punch. The fluidity of the sequence displays a mastery of balance and poise. Every trajectory of the limbs is precise and deliberate, capturing the elegance and strength of martial arts. The serene, isolated setting enhances the intensity and concentration of the practitioner. This visual presentation is an elegant interplay of motion and stillness, displaying the art form's discipline and grace.",
-      "video_path": "Man/mixkit-a-young-man-practicing-his-karate-moves-49635_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A tranquil coastal scene unfolds with a drone's aerial view capturing a serene beach landscape. The camera glides over a quiet stretch of sandy shoreline, where gentle waves kiss the shore under a clear blue sky. Nestled amidst lush palm trees are a series of traditional thatched-roof huts, their earthy tones blending harmoniously with the natural surroundings. The sandy beach stretches endlessly, bordered by the rhythmic dance of ocean waves on one side and verdant greenery on the other. A pair of white umbrellas is set up on the sand, suggesting a place to relax and enjoy the sun. In the distance, two small human figures can be seen walking leisurely along the water's edge, leaving faint footprints behind them. The scene exudes a calm and inviting atmosphere, with the soft rustle of palm leaves and the whisper of the ocean breeze almost audible. The overall composition is a captivating blend of nature's tranquility and architectural simplicity. This picturesque setting invites viewers to imagine themselves steps away from this idyllic coastal escape.",
-      "video_path": "beach/mixkit-sunny-beach-in-a-dynamic-shot-from-a-drone-44383_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A lone figure stands on a large, moss-covered rock, surrounded by the soft rush of a nearby stream. The figure is wearing white sneakers and shorts, with a plaid shirt that hangs loosely in the breeze. The lighting creates dramatic shadows, enhancing the textures of the rock and the subtle movement of the water below. In the background, a waterfall cascades into the stream, completing this tranquil and serene nature scene.",
-      "video_path": "forest/mixkit-woman-standing-in-front-of-waterfall-559_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In an industrial setting, a person leans casually against a railing, exuding a sense of confidence and composure. They are wearing a striking outfit, consisting of a vibrant, patterned jacket over a simple white crop top, creating a bold contrast. The atmosphere is infused with warm, ambient lighting that casts soft shadows on the concrete walls and metallic surfaces. Intricate wiring and pipes form an intricate backdrop, enhancing the urban aesthetic. Their relaxed posture and direct, engaging gaze suggest a sense of ease in this industrial environment. This scene encapsulates a blend of modern fashion and gritty, urban architecture, creating a visually compelling narrative.",
-      "video_path": "Fashion/mixkit-portrait-of-a-hipster-woman-walking-down-a-stairs-1297_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A man is energetically stretching in an open-air setting, surrounded by rows of vibrant red seats that suggest an amphitheater or outdoor venue. He wears a sleeveless black shirt layered with a hooded vest, emphasizing his athletic build as he engages in a warm-up routine. Behind him, the striking modern architecture of the building features geometric panels, with large sections of glass and overlapping metallic beams creating a dynamic backdrop. The scene captures the contrast between his focused movements and the static, bold design of the structure, while the surrounding greenery adds a touch of nature to the environment. The overall atmosphere is one of preparation and anticipation, with the man appearing determined and ready for an upcoming event or performance.",
-      "video_path": "Sport/mixkit-man-doing-arm-stretches-595_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young woman is seated on the floor in front of a plush, beige tufted couch, fully engrossed in sorting through a stack of papers. Her dark hair falls loosely past her shoulders, and she wears a green plaid shirt, contributing to the casual yet focused atmosphere. She gently places the papers onto a small round white table, occasionally lifting individual sheets to examine them more closely. Her expression shifts subtly, reflecting concentration and contemplation as she processes the information on the pages. Two small, round nested tables hold her documents, along with a small plant in a gray pot, adding a touch of greenery to the scene. The background features a dark paneled wall, creating a contrasting backdrop for the light-colored furniture. The setting is tranquil and organized, the couch and tables arranged symmetrically, conveying a sense of harmony. A calculator rests on the smaller table, hinting at a task involving calculations or budgeting.",
-      "video_path": "Woman/mixkit-frustrated-woman-throws-paperwork-on-the-floor-4526_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A heavily rusted metal gate stands firmly locked, with two vertical bars joined by a thick, old chain that loops elegantly around them. The chain's texture is coarse and rugged, its surface reflecting varying shades of orange and brown, indicative of years exposed to the elements. At the heart of the chain, a black iron padlock, slightly worn yet imposing, secures the gate, its curves and edges smooth against the aged links. The gate's metalwork is outlined by a backdrop of soft, blurred greenery, suggesting a serene and isolated location beyond the barrier. Tall trees rise in the distance, their trunks and leaves creating a lush, forest-like setting that contrasts with the gate's severe rust. A pathway leads away from the gate, its surface uneven with patches of moss and weathered stone visible in the soft focus, inviting yet inaccessible. The ambiance is quiet and mysterious, with a sense of abandonment hanging subtly in the air, evoking curiosity about what lies beyond. Shadows play across the gate, cast by branches swaying gently in the breeze, adding to the dynamic interaction of light and texture. This scene, rich in detail and atmosphere, captures the viewer's imagination, evoking both the allure of the forbidden and the beauty of decay.",
-      "video_path": "forest/mixkit-rusty-fence-with-a-chain-of-a-property-in-nature-5294_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a serene and softly lit yoga studio, three individuals engage in a yoga session, each performing an upward-facing stretch. The central figure is a woman with shoulder-length brown hair, dressed in a light cropped top and green leggings, her posture reflecting grace and concentration. To her right, another participant, a woman in a purple outfit, mirrors the pose with equal poise. On her left, a person with a bun focuses intently, supported slightly by yoga blocks beneath their hands. The warm-colored wooden floor contrasts soothingly with the soft pastel mural on the back wall, featuring an abstract design and partial visage of a serene face. Natural light floods the space from a large window on the right, where lush greens peek through, adding an element of tranquility. In the corner of the room, a collection of meditation instruments, including a gong and a Buddha statue, subtly frame the peaceful setting. The mood is calm yet focused, as all three participants are deeply engaged in their practice. The scene combines elements of balance, harmony, and a shared journey towards mindfulness. This depiction captures the essence of a yoga session that blends personal growth with collective experience.",
-      "video_path": "People/mixkit-small-group-of-people-doing-yoga-together-43730_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the deep blue expanse of the ocean, two dolphins glide effortlessly, their sleek bodies reflecting the sunlight filtering through the water. The prominent shadows and caustics create a shimmering effect on their skin, capturing the beauty of their natural habitat. Each dolphin moves with a fluid grace, occasionally interacting with gentle nudges, showcasing their playful and social nature. The scene is vibrant and dynamic, with the clear blue background accentuating the dolphins' movements, making it an ideal subject for AI recreation.",
-      "video_path": "sea/mixkit-dolphins-underwater-4133_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "In the video, a young woman stands against a vibrant graffiti-covered wall, deeply engrossed in her smartphone. Her expression reflects a mix of focus and subtle satisfaction as she interacts with the screen. She wears a black floral-patterned top, which contrasts with the bright, abstract shapes and bold colors of the mural behind her. As she continues to engage with her phone, a series of like count notifications appear on the screen, indicating a growing online appreciation. The wall behind her features a striking mix of geometric and organic shapes, including swirls of teal, orange, and black, with large humanoid figures in a pop-art style. Her long, light-brown hair frames her face, adding a calm, composed aura amidst the lively backdrop. The video captures a blend of contemporary digital interaction and expressive urban art, creating a dynamic yet harmonious scene.",
-      "video_path": "Girl/mixkit-girl-looking-at-the-likes-in-her-post-4914_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young mother and her baby sit comfortably on a bed, surrounded by an inviting, cozy atmosphere. The woman, wearing a sleeveless top and jeans, is gently engaging with the baby, who is dressed in an adorable animal-print onesie. The child is seated on the bed with colorful toys scattered around, including a plush toy and a board book. The warm glow from a hanging lamp casts a soft light on them, enhancing the serene environment. Pillows are propped up against the headboard, providing a cushioned backdrop as the mother leans slightly over to interact with the baby. A small bottle is visible beside her, suggesting a nurturing setting. Her hand gestures animatedly as she holds up a soft, white cushion with red and blue accents, likely stimulating the baby\u2019s curiosity. Their shared moment is filled with affection and joy, a perfect snapshot of familial bonding.",
-      "video_path": "Baby/mixkit-loving-mother-and-her-baby-playing-with-soft-toys-49966_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-    },
-    {
-      "caption": "A young girl with long brown hair sits at a round wooden table, engrossed in working on her laptop. The laptop screen is a vivid green, suggesting a green screen effect is in use. To her left, a doll dressed in a yellow and white outfit is casually laid on top of some books, adding a playful and innocent touch to the scene. The setting is cozy, with sheer curtains in the background allowing soft natural light to spill into the room. The girl's posture and focused attention on the laptop suggest she is either playing a game or learning something new. This serene and domestic atmosphere is complemented by the slight blur of a dark couch in the foreground, framing the focused activity of the child.",
-      "video_path": "Girl/mixkit-little-girl-doing-homework-on-a-laptop-4757_clip_1.mp4",
-      "num_inference_steps":  50,
-      "height": 768,
-      "width": 1280,
-      "num_frames": 77
-  }  ]
-}

fastvideo-kernel/CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.26 FATAL_ERROR)
 project(fastvideo-kernel LANGUAGES CXX)
 
-# Prefer environment variable (used by CI or pip install git+repo_addr) if CMake var is not explicitly set.
+# Prefer environment variable (used by CI or uv pip install git+repo_addr) if CMake var is not explicitly set.
 if(NOT DEFINED GPU_BACKEND AND DEFINED ENV{GPU_BACKEND})
     set(GPU_BACKEND "$ENV{GPU_BACKEND}")
 endif()
@@ -12,18 +12,6 @@ else()
     enable_language(CUDA)
     # Ensure CUDA toolkit targets (CUDA::cudart, CUDA::cuda_driver, etc.) are available.
     find_package(CUDAToolkit REQUIRED)
-    if(NOT DEFINED CUDA_TOOLKIT_ROOT_DIR)
-        if(DEFINED CUDAToolkit_ROOT)
-            set(CUDA_TOOLKIT_ROOT_DIR "${CUDAToolkit_ROOT}" CACHE PATH
-                "CUDA toolkit root directory" FORCE)
-        elseif(DEFINED ENV{CUDAToolkit_ROOT})
-            set(CUDA_TOOLKIT_ROOT_DIR "$ENV{CUDAToolkit_ROOT}" CACHE PATH
-                "CUDA toolkit root directory" FORCE)
-        elseif(DEFINED ENV{CUDA_HOME})
-            set(CUDA_TOOLKIT_ROOT_DIR "$ENV{CUDA_HOME}" CACHE PATH
-                "CUDA toolkit root directory" FORCE)
-        endif()
-    endif()
 endif()
 
 # Import common utils if needed, but we keep it simple for now
@@ -31,46 +19,13 @@ endif()
 # Find Python and Torch
 find_package(Python COMPONENTS Interpreter Development.Module REQUIRED)
 
-# Locate the installed torch package without importing it. This keeps CMake
-# configure working even on nodes where CUDA runtime libraries are not yet on
-# the dynamic loader path.
+# Robustly find Torch include paths using Python
 execute_process(
-    COMMAND "${Python_EXECUTABLE}" -c "import sysconfig; print(sysconfig.get_path('platlib'))"
-    OUTPUT_VARIABLE PYTHON_PLATLIB
+    COMMAND "${Python_EXECUTABLE}" -c "import torch; from torch.utils.cpp_extension import include_paths; print(';'.join(include_paths()))"
+    OUTPUT_VARIABLE TORCH_INCLUDE_PATHS
     OUTPUT_STRIP_TRAILING_WHITESPACE
 )
-execute_process(
-    COMMAND "${Python_EXECUTABLE}" -c "import sysconfig; print(sysconfig.get_path('purelib'))"
-    OUTPUT_VARIABLE PYTHON_PURELIB
-    OUTPUT_STRIP_TRAILING_WHITESPACE
-)
-
-set(TORCH_PYTHON_PACKAGE_DIR "")
-foreach(_candidate
-    "${PYTHON_PLATLIB}/torch"
-    "${PYTHON_PURELIB}/torch"
-)
-    if(EXISTS "${_candidate}")
-        set(TORCH_PYTHON_PACKAGE_DIR "${_candidate}")
-        break()
-    endif()
-endforeach()
-
-if(NOT TORCH_PYTHON_PACKAGE_DIR)
-    message(FATAL_ERROR "Could not locate the installed torch Python package.")
-endif()
-
-list(APPEND TORCH_INCLUDE_DIRS
-    "${TORCH_PYTHON_PACKAGE_DIR}/include"
-    "${TORCH_PYTHON_PACKAGE_DIR}/include/torch/csrc/api/include"
-)
-
-if(NOT Torch_DIR)
-    set(_TORCH_CONFIG_DIR "${TORCH_PYTHON_PACKAGE_DIR}/share/cmake/Torch")
-    if(EXISTS "${_TORCH_CONFIG_DIR}/TorchConfig.cmake")
-        set(Torch_DIR "${_TORCH_CONFIG_DIR}" CACHE PATH "Path to Torch CMake config" FORCE)
-    endif()
-endif()
+list(APPEND TORCH_INCLUDE_DIRS ${TORCH_INCLUDE_PATHS})
 
 # Find Torch package (still useful for libraries)
 find_package(Torch REQUIRED)
@@ -95,21 +50,6 @@ include_directories(
 set(FASTVIDEO_KERNEL_BUILD_TK "AUTO" CACHE STRING "Build ThunderKittens kernels: AUTO/ON/OFF")
 set_property(CACHE FASTVIDEO_KERNEL_BUILD_TK PROPERTY STRINGS AUTO ON OFF)
 
-set(_FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER_DEFAULT "AUTO")
-if(DEFINED FASTVIDEO_KERNEL_BUILD_MODIFIED_SAGE3 AND NOT DEFINED CACHE{FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER})
-    set(_FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER_DEFAULT "${FASTVIDEO_KERNEL_BUILD_MODIFIED_SAGE3}")
-endif()
-
-set(FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER "${_FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER_DEFAULT}" CACHE STRING
-    "Build attn_qat_infer Blackwell inference kernels: AUTO/ON/OFF")
-set_property(CACHE FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER PROPERTY STRINGS AUTO ON OFF)
-
-if(DEFINED FASTVIDEO_KERNEL_BUILD_MODIFIED_SAGE3)
-    message(DEPRECATION
-        "FASTVIDEO_KERNEL_BUILD_MODIFIED_SAGE3 is deprecated. "
-        "Use FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER instead.")
-endif()
-
 # Prefer environment variable (used by CI) if CMake var is not explicitly set.
 if(NOT DEFINED TORCH_CUDA_ARCH_LIST AND DEFINED ENV{TORCH_CUDA_ARCH_LIST})
     set(TORCH_CUDA_ARCH_LIST "$ENV{TORCH_CUDA_ARCH_LIST}")
@@ -117,7 +57,6 @@ endif()
 
 message(STATUS "TORCH_CUDA_ARCH_LIST (cmake/env): ${TORCH_CUDA_ARCH_LIST}")
 message(STATUS "FASTVIDEO_KERNEL_BUILD_TK: ${FASTVIDEO_KERNEL_BUILD_TK}")
-message(STATUS "FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER: ${FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER}")
 
 set(ENABLE_TK_KERNELS OFF)
 if(FASTVIDEO_KERNEL_BUILD_TK STREQUAL "ON")
@@ -152,54 +91,6 @@ else()
     message(STATUS "ThunderKittens kernels: DISABLED (will use Triton fallbacks at runtime)")
 endif()
 
-set(ENABLE_ATTN_QAT_INFER OFF)
-if(GPU_BACKEND STREQUAL "ROCM")
-    message(STATUS "attn_qat_infer kernels: DISABLED (ROCm build)")
-else()
-    set(_WANTS_ATTN_QAT_INFER OFF)
-    if(FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER STREQUAL "ON")
-        set(_WANTS_ATTN_QAT_INFER ON)
-    elseif(FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER STREQUAL "AUTO")
-        if(TORCH_CUDA_ARCH_LIST)
-            string(REGEX MATCH
-                "(^|[; ,])((12\\.0a)|(120a)|(sm_120a))([; ,]|$)"
-                _HAS_120A "${TORCH_CUDA_ARCH_LIST}")
-            if(_HAS_120A)
-                set(_WANTS_ATTN_QAT_INFER ON)
-            endif()
-        else()
-            execute_process(
-                COMMAND "${Python_EXECUTABLE}" -c
-                "import torch; print('1' if (torch.cuda.is_available() and torch.version.cuda and torch.cuda.get_device_capability()[0] >= 12) else '0')"
-                OUTPUT_VARIABLE _LOCAL_HAS_BLACKWELL
-                OUTPUT_STRIP_TRAILING_WHITESPACE
-                ERROR_QUIET
-            )
-            if(_LOCAL_HAS_BLACKWELL STREQUAL "1")
-                set(_WANTS_ATTN_QAT_INFER ON)
-            endif()
-        endif()
-    endif()
-
-    if(_WANTS_ATTN_QAT_INFER)
-        if(CUDAToolkit_VERSION VERSION_LESS 12.8)
-            message(WARNING
-                "attn_qat_infer kernels require CUDA Toolkit 12.8+. "
-                "Skipping because CUDAToolkit_VERSION=${CUDAToolkit_VERSION}.")
-        else()
-            set(ENABLE_ATTN_QAT_INFER ON)
-        endif()
-    endif()
-
-    if(ENABLE_ATTN_QAT_INFER)
-        message(STATUS "attn_qat_infer kernels: ENABLED")
-    else()
-        message(STATUS
-            "attn_qat_infer kernels: DISABLED "
-            "(requires CUDA 12.8+ and Blackwell sm_120a)")
-    endif()
-endif()
-
 # Always try to build the extension if CUDA is available, but conditionally add sources/flags
 set(BUILD_CXX_KERNELS ON)
 
@@ -270,15 +161,12 @@ if(BUILD_CXX_KERNELS)
 
     # Also link against libtorch_python to satisfy Python-binding symbols
     # (e.g., torch::PyWarningHandler) required by torch/extension.h.
-    file(GLOB TORCH_PYTHON_LIBRARY_CANDIDATES
-        "${TORCH_PYTHON_PACKAGE_DIR}/lib/libtorch_python*"
+    execute_process(
+        COMMAND "${Python_EXECUTABLE}" -c "import torch; from pathlib import Path; p=Path(torch.__file__).parent/'lib'; m=sorted(p.glob('libtorch_python*')); print(str(m[0]) if m else '')"
+        OUTPUT_VARIABLE TORCH_PYTHON_LIBRARY_PATH
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
     )
-    list(LENGTH TORCH_PYTHON_LIBRARY_CANDIDATES _TORCH_PYTHON_LIBRARY_COUNT)
-    if(_TORCH_PYTHON_LIBRARY_COUNT GREATER 0)
-        list(GET TORCH_PYTHON_LIBRARY_CANDIDATES 0 TORCH_PYTHON_LIBRARY_PATH)
-    else()
-        set(TORCH_PYTHON_LIBRARY_PATH "")
-    endif()
     if(TORCH_PYTHON_LIBRARY_PATH)
         message(STATUS "TORCH_PYTHON_LIBRARY_PATH: ${TORCH_PYTHON_LIBRARY_PATH}")
         target_link_libraries(fastvideo_kernel_ops PRIVATE "${TORCH_PYTHON_LIBRARY_PATH}")
@@ -295,73 +183,3 @@ if(BUILD_CXX_KERNELS)
     install(TARGETS fastvideo_kernel_ops LIBRARY DESTINATION fastvideo_kernel/_C)
 endif()
 
-if(ENABLE_ATTN_QAT_INFER)
-    set(ATTN_QAT_INFER_DIR ${CMAKE_SOURCE_DIR}/attn_qat_infer)
-    set(ATTN_QAT_INFER_INCLUDE_DIRS
-        ${ATTN_QAT_INFER_DIR}
-        ${CMAKE_SOURCE_DIR}/include/cutlass/include
-        ${CMAKE_SOURCE_DIR}/include/cutlass/tools/util/include
-        ${TORCH_INCLUDE_DIRS}
-    )
-    set(ATTN_QAT_INFER_CUDA_FLAGS
-        "-O3"
-        "-std=c++17"
-        "-U__CUDA_NO_HALF_OPERATORS__"
-        "-U__CUDA_NO_HALF_CONVERSIONS__"
-        "-U__CUDA_NO_BFLOAT16_OPERATORS__"
-        "-U__CUDA_NO_BFLOAT16_CONVERSIONS__"
-        "-U__CUDA_NO_BFLOAT162_OPERATORS__"
-        "-U__CUDA_NO_BFLOAT162_CONVERSIONS__"
-        "--expt-relaxed-constexpr"
-        "--expt-extended-lambda"
-        "--use_fast_math"
-        "--ptxas-options=--verbose,--warn-on-local-memory-usage"
-        "-lineinfo"
-        "-DCUTLASS_DEBUG_TRACE_LEVEL=0"
-        "-DNDEBUG"
-        "-DQBLKSIZE=128"
-        "-DKBLKSIZE=128"
-        "-DCTA256"
-        "-DDQINRMEM"
-    )
-
-    Python_add_library(fp4attn_cuda MODULE WITH_SOABI
-        attn_qat_infer/blackwell/api.cu
-    )
-    target_include_directories(fp4attn_cuda PRIVATE ${ATTN_QAT_INFER_INCLUDE_DIRS})
-    target_compile_definitions(fp4attn_cuda PRIVATE TORCH_EXTENSION_NAME=fp4attn_cuda)
-    target_compile_options(fp4attn_cuda PRIVATE
-        $<$<COMPILE_LANGUAGE:CXX>:-O3 -std=c++17>
-        $<$<COMPILE_LANGUAGE:CUDA>:${ATTN_QAT_INFER_CUDA_FLAGS}>
-    )
-    set_target_properties(fp4attn_cuda PROPERTIES
-        CUDA_ARCHITECTURES "120a"
-        CXX_STANDARD 17
-        CUDA_STANDARD 17
-    )
-    target_link_libraries(fp4attn_cuda PRIVATE ${TORCH_LIBRARIES} CUDA::cudart CUDA::cuda_driver)
-
-    Python_add_library(fp4quant_cuda MODULE WITH_SOABI
-        attn_qat_infer/quantization/fp4_quantization_4d.cu
-    )
-    target_include_directories(fp4quant_cuda PRIVATE ${ATTN_QAT_INFER_INCLUDE_DIRS})
-    target_compile_definitions(fp4quant_cuda PRIVATE TORCH_EXTENSION_NAME=fp4quant_cuda)
-    target_compile_options(fp4quant_cuda PRIVATE
-        $<$<COMPILE_LANGUAGE:CXX>:-O3 -std=c++17>
-        $<$<COMPILE_LANGUAGE:CUDA>:${ATTN_QAT_INFER_CUDA_FLAGS}>
-    )
-    set_target_properties(fp4quant_cuda PROPERTIES
-        CUDA_ARCHITECTURES "120a"
-        CXX_STANDARD 17
-        CUDA_STANDARD 17
-    )
-    target_link_libraries(fp4quant_cuda PRIVATE ${TORCH_LIBRARIES} CUDA::cudart CUDA::cuda_driver)
-
-    if(TORCH_PYTHON_LIBRARY_PATH)
-        target_link_libraries(fp4attn_cuda PRIVATE "${TORCH_PYTHON_LIBRARY_PATH}")
-        target_link_libraries(fp4quant_cuda PRIVATE "${TORCH_PYTHON_LIBRARY_PATH}")
-    endif()
-
-    install(TARGETS fp4attn_cuda LIBRARY DESTINATION .)
-    install(TARGETS fp4quant_cuda LIBRARY DESTINATION .)
-endif()

fastvideo-kernel/MANIFEST.in

@@ -2,6 +2,5 @@ include LICENSE
 include README.md
 include pyproject.toml
 recursive-include python/fastvideo_kernel *.py
-recursive-include attn_qat_infer *.py *.cu *.cuh *.cpp *.h
 recursive-include csrc *.cu *.cuh *.cpp *.h
 recursive-include include/tk *.cu *.cuh *.cpp *.h *.src

fastvideo-kernel/README.md

@@ -20,11 +20,6 @@ cd fastvideo-kernel
 ./build.sh
 ```
 
-On supported Blackwell environments, the same install also packages
-`attn_qat_infer` and builds its `fp4attn_cuda` / `fp4quant_cuda`
-extensions directly from `fastvideo-kernel/attn_qat_infer/`. This path
-requires CUDA Toolkit 12.8+ and targets `sm_120a`.
-
 ### Rocm Build
 If you are in a rocm environment without the compilation toolchaine of CUDA.
 
@@ -63,18 +58,6 @@ cd fastvideo-kernel
 python benchmarks/bench_vsa.py --batch_size 1 --num_heads 16 --head_dim 128 --q_seq_lens 49152 --topk 64
 ```
 
-### Attn QAT Attention Benchmarks
-
-The Attn QAT microbenchmarks now live alongside the kernel package:
-
-```bash
-cd fastvideo-kernel
-python benchmarks/benchmark_flashattn2.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128
-python benchmarks/benchmark_sageattn3.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128
-python benchmarks/benchmark_blockscaled_fp4_attn.py --batch-size 1 --num-heads 16 --seq-len 4096 --head-dim 128
-python benchmarks/benchmark_combined.py --output benchmark_attention.png
-```
-
 ### TurboDiffusion Kernels
 
 This package also includes kernels from [TurboDiffusion](https://github.com/thu-ml/TurboDiffusion), including INT8 GEMM, Quantization, RMSNorm and LayerNorm.

fastvideo-kernel/attn_qat_infer/__init__.py

@@ -1,16 +0,0 @@
-"""
-Copyright (c) 2025 by SageAttention team.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-from .api import sageattn_blackwell

fastvideo-kernel/attn_qat_infer/api.py

@@ -1,185 +0,0 @@
-# Modified from the original SageATtention3 code
-"""
-Copyright (c) 2025 by SageAttention team.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-import torch
-import triton
-import triton.language as tl
-import torch.nn.functional as F
-from typing import Tuple
-from torch.nn.functional import scaled_dot_product_attention as sdpa
-import fp4attn_cuda
-import fp4quant_cuda
-
-# Centralized block size configuration for sageattn_blackwell kernels
-# These should match the values in fastvideo/attention/backends/sageattn/blackwell/block_config.h
-BLOCK_M = 128  # Block size for M dimension (query sequence length)
-BLOCK_N = 128  # Block size for N dimension (key/value sequence length)
-
-
-@triton.jit
-def group_mean_kernel(
-    q_ptr,          
-    q_out_ptr,      
-    qm_out_ptr,     
-    B, H, L, D: tl.constexpr,    
-    stride_qb, stride_qh, stride_ql, stride_qd,  
-    stride_qmb, stride_qmh, stride_qml, stride_qmd,  
-    GROUP_SIZE: tl.constexpr
-):
-    pid_b = tl.program_id(0)
-    pid_h = tl.program_id(1)
-    pid_group = tl.program_id(2)
-    
-    group_start = pid_group * GROUP_SIZE
-    offsets = group_start + tl.arange(0, GROUP_SIZE)
-    
-    q_offsets = pid_b * stride_qb + pid_h * stride_qh + offsets[:, None] * stride_ql + tl.arange(0, D)[None, :] * stride_qd
-    q_group = tl.load(q_ptr + q_offsets)
-    
-    qm_group = tl.sum(q_group, axis=0) / GROUP_SIZE
-    
-    q_group = q_group - qm_group
-    tl.store(q_out_ptr + q_offsets, q_group)
-
-    qm_offset = pid_b * stride_qmb + pid_h * stride_qmh + pid_group * stride_qml + tl.arange(0, D) * stride_qmd
-    tl.store(qm_out_ptr + qm_offset, qm_group)
-
-
-def triton_group_mean(q: torch.Tensor):
-    B, H, L, D = q.shape
-    GROUP_SIZE = BLOCK_M
-    num_groups = L // GROUP_SIZE
-    
-    q_out = torch.empty_like(q)  # [B, H, L, D]
-    qm = torch.empty(B, H, num_groups, D, device=q.device, dtype=q.dtype) 
-    
-    grid = (B, H, num_groups)
-    
-    group_mean_kernel[grid](
-        q, q_out, qm,
-        B, H, L, D,
-        q.stride(0), q.stride(1), q.stride(2), q.stride(3),
-        qm.stride(0), qm.stride(1), qm.stride(2), qm.stride(3),
-        GROUP_SIZE=GROUP_SIZE
-    )
-    return q_out, qm
-
-
-def preprocess_qkv(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, per_block_mean: bool = True, enable_smoothing_q: bool = False, enable_smoothing_k: bool = False):
-
-    def pad_to_block_size(x):
-        L = x.size(2)
-        pad_len = (BLOCK_M - L % BLOCK_M) % BLOCK_M
-        if pad_len == 0:
-            return x.contiguous()
-        return F.pad(x, (0, 0, 0, pad_len), value=0).contiguous()
-    
-    if enable_smoothing_k:
-        k -= k.mean(dim=-2, keepdim=True)  
-    q, k, v = map(lambda x: pad_to_block_size(x), [q, k, v])
-    if per_block_mean and enable_smoothing_q:
-        q, qm = triton_group_mean(q)
-    elif enable_smoothing_q:
-        qm = q.mean(dim=-2, keepdim=True)
-        q = q - qm
-    if enable_smoothing_q:
-        delta_s = torch.matmul(qm, k.transpose(-2, -1)).to(torch.float32).contiguous()
-    else:  # used to disable q smoothing
-        B, H, L, D = q.shape
-        delta_s = torch.zeros((B, H, L // BLOCK_M, k.shape[2]), device=q.device, dtype=torch.float32)
-        
-    return q, k, v, delta_s
-
-def scale_and_quant_fp4(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    assert x.ndim == 4
-    B, H, N, D = x.shape
-    packed_fp4 = torch.empty((B, H, N, D // 2), device=x.device, dtype=torch.uint8)
-    fp8_scale = torch.empty((B, H, N, D // 16), device=x.device, dtype=torch.float8_e4m3fn)
-    fp4quant_cuda.scaled_fp4_quant(x, packed_fp4, fp8_scale, 1)
-    return packed_fp4, fp8_scale
-
-def scale_and_quant_fp4_permute(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    assert x.ndim == 4
-    B, H, N, D = x.shape
-    packed_fp4 = torch.empty((B, H, N, D // 2), device=x.device, dtype=torch.uint8)
-    fp8_scale = torch.empty((B, H, N, D // 16), device=x.device, dtype=torch.float8_e4m3fn)
-    fp4quant_cuda.scaled_fp4_quant_permute(x, packed_fp4, fp8_scale, 1)
-    return packed_fp4, fp8_scale
-
-def scale_and_quant_fp4_transpose(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    assert x.ndim == 4
-    B, H, N, D = x.shape
-    packed_fp4 = torch.empty((B, H, D, N // 2), device=x.device, dtype=torch.uint8)
-    fp8_scale = torch.empty((B, H, D, N // 16), device=x.device, dtype=torch.float8_e4m3fn)
-    fp4quant_cuda.scaled_fp4_quant_trans(x, packed_fp4, fp8_scale, 1)
-    return packed_fp4, fp8_scale
-
-def blockscaled_fp4_attn(qlist: Tuple, 
-                         klist: Tuple,
-                         vlist: Tuple,
-                         delta_s: torch.Tensor,
-                         KL: int,
-                         is_causal: bool = False, 
-                         per_block_mean: bool = True,
-                         is_bf16: bool = True,
-                         single_level_p_quant: bool = False
-                        ):
-    softmax_scale = (qlist[0].shape[-1] * 2) ** (-0.5)
-    return fp4attn_cuda.fwd(qlist[0], klist[0], vlist[0], qlist[1], klist[1], vlist[1], delta_s, KL, None, softmax_scale, is_causal, per_block_mean, is_bf16, single_level_p_quant)
-
-
-def sageattn_blackwell(q, k, v, attn_mask = None, is_causal = False, per_block_mean = True, single_level_p_quant = True, **kwargs):
-    """
-    SageAttention3 Blackwell kernel for FP4 attention.
-    
-    Args:
-        q: Query tensor [B, H, L, D]
-        k: Key tensor [B, H, L, D]
-        v: Value tensor [B, H, L, D]
-        attn_mask: Attention mask (not used)
-        is_causal: Whether to use causal masking
-        per_block_mean: Whether to use per-block mean for Q smoothing
-        single_level_p_quant: If True, use single-level quantization: s_P2, P̂_2 = φ(P̃) directly
-                              (standard per-block FP4 quantization like V, no s_P1).
-                              If False (default), use two-level quantization:
-                              s_P1 = rowmax(P̃)/(448×6), then s_P2, P̂_2 = φ(P̃/s_P1).
-        **kwargs: Additional arguments (ignored)
-    
-    Returns:
-        Output tensor [B, H, L, D]
-    """
-    if q.size(-1) >= 256:
-        print(f"Unsupported Headdim {q.size(-1)}")
-        return sdpa(q, k, v, is_causal = is_causal)
-    QL = q.size(2)
-    KL = k.size(2)
-    is_bf16 = q.dtype == torch.bfloat16
-    q, k, v, delta_s = preprocess_qkv(q, k, v, per_block_mean)
-    qlist_from_cuda = scale_and_quant_fp4(q)
-    klist_from_cuda = scale_and_quant_fp4_permute(k)
-    vlist_from_cuda = scale_and_quant_fp4_transpose(v)
-    o_fp4 = blockscaled_fp4_attn(
-    qlist_from_cuda,
-    klist_from_cuda, 
-    vlist_from_cuda,
-    delta_s,
-    KL,
-    is_causal,
-    per_block_mean,
-    is_bf16,
-    single_level_p_quant
-    )[0][:, :, :QL, :].contiguous()
-    return o_fp4

fastvideo-kernel/attn_qat_infer/blackwell/__init__.py

@@ -1 +0,0 @@
-__version__ = "3.0.0.b1"

fastvideo-kernel/attn_qat_infer/blackwell/api.cu

@@ -1,346 +0,0 @@
-// Modified from the original SageAttention3 code
-/*
- * Copyright (c) 2025 by SageAttention team.
- * 
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// Include these 2 headers instead of torch/extension.h since we don't need all of the torch headers.
-#include <torch/python.h>
-#include <torch/nn/functional.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include <cutlass/numeric_types.h>
-
-#include "params.h"
-#include "launch.h"
-#include "static_switch.h"
-#include "block_config.h"
-
-#define CHECK_DEVICE(x) TORCH_CHECK(x.is_cuda(), #x " must be on CUDA")
-#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
-#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
-
-
-void set_params_fprop(Flash_fwd_params &params,
-                      // sizes
-                      const size_t b,
-                      const size_t seqlen_q,
-                      const size_t seqlen_k,
-                      const size_t unpadded_seqlen_k,
-                      const size_t seqlen_q_rounded,
-                      const size_t seqlen_k_rounded,
-                      const size_t h,
-                      const size_t h_k,
-                      const size_t d,
-                      const size_t d_rounded,
-                      // device pointers
-                      const at::Tensor q,
-                      const at::Tensor k,
-                      const at::Tensor v,
-                      const at::Tensor delta_s,
-                      at::Tensor out,
-                      const at::Tensor sfq,
-                      const at::Tensor sfk,
-                      const at::Tensor sfv,
-                      void *cu_seqlens_q_d,
-                      void *cu_seqlens_k_d,
-                      void *seqused_k,
-                      void *p_d,
-                      void *softmax_lse_d,
-                      float p_dropout,
-                      float softmax_scale,
-                      int window_size_left,
-                      int window_size_right,
-                      bool per_block_mean,
-                      bool is_bf16,
-                      bool single_level_p_quant=false,
-                      bool seqlenq_ngroups_swapped=false) {
-
-    // Reset the parameters
-    params = {};
-    // Set the pointers and strides.
-    params.q_ptr = q.data_ptr();
-    params.k_ptr = k.data_ptr();
-    params.v_ptr = v.data_ptr();
-    params.delta_s_ptr = delta_s.data_ptr();
-    params.sfq_ptr = sfq.data_ptr();
-    params.sfk_ptr = sfk.data_ptr();
-    params.sfv_ptr = sfv.data_ptr();
-    
-    // All stride are in elements, not bytes.
-    params.q_row_stride = q.stride(-2) * 2;
-    params.k_row_stride = k.stride(-2) * 2;
-    params.v_row_stride = v.stride(-2) * 2;;
-    params.q_head_stride = q.stride(-3) * 2;
-    params.k_head_stride = k.stride(-3) * 2;
-    params.v_head_stride = v.stride(-3) * 2; // for packed q k v
-
-    params.ds_row_stride = delta_s.stride(-2);
-    params.ds_head_stride = delta_s.stride(-3);
-    
-    params.sfq_row_stride = sfq.stride(-2);
-    params.sfk_row_stride = sfk.stride(-2);
-    params.sfv_row_stride = sfv.stride(-2);
-    params.sfq_head_stride = sfq.stride(-3);
-    params.sfk_head_stride = sfk.stride(-3);
-    params.sfv_head_stride = sfv.stride(-3);
-    params.o_ptr = out.data_ptr();
-    params.o_row_stride = out.stride(-2);
-    params.o_head_stride = out.stride(-3);
-
-    if (cu_seqlens_q_d == nullptr) {
-        params.q_batch_stride = q.stride(0) * 2;
-        params.k_batch_stride = k.stride(0) * 2;
-        params.v_batch_stride = v.stride(0) * 2;
-        params.ds_batch_stride = delta_s.stride(0);
-        params.sfq_batch_stride = sfq.stride(0);
-        params.sfk_batch_stride = sfk.stride(0);
-        params.sfv_batch_stride = sfv.stride(0);
-        params.o_batch_stride = out.stride(0);
-        if (seqlenq_ngroups_swapped) {
-             params.q_batch_stride *= seqlen_q;
-             params.o_batch_stride *= seqlen_q;
-        }
-    }
-
-    params.cu_seqlens_q = static_cast<int *>(cu_seqlens_q_d);
-    params.cu_seqlens_k = static_cast<int *>(cu_seqlens_k_d);
-    params.seqused_k = static_cast<int *>(seqused_k);
-
-    // P = softmax(QK^T)
-    params.p_ptr = p_d;
-
-    // Softmax sum
-    params.softmax_lse_ptr = softmax_lse_d;
-
-    // Set the dimensions.
-    params.b = b;
-    params.h = h;
-    params.h_k = h_k;
-    params.h_h_k_ratio = h / h_k;
-    params.seqlen_q = seqlen_q;
-    params.seqlen_k = seqlen_k;
-    params.unpadded_seqlen_k = unpadded_seqlen_k;
-    params.seqlen_q_rounded = seqlen_q_rounded;
-    params.seqlen_k_rounded = seqlen_k_rounded;
-    params.d = d;
-    params.d_rounded = d_rounded;
-
-    params.head_divmod = cutlass::FastDivmod(int(h));
-
-    // Set the different scale values.
-    params.scale_softmax = softmax_scale;
-    params.scale_softmax_log2 = softmax_scale * M_LOG2E;
-    __half scale_softmax_log2_half = __float2half(params.scale_softmax_log2);
-    __half2 scale_softmax_log2_half2 = __half2(scale_softmax_log2_half, scale_softmax_log2_half);
-    params.scale_softmax_log2_half2 = reinterpret_cast<uint32_t&>(scale_softmax_log2_half2);
-
-    // Set this to probability of keeping an element to simplify things.
-    params.p_dropout = 1.f - p_dropout;
-    // Convert p from float to int so we don't have to convert the random uint to float to compare.
-    // [Minor] We want to round down since when we do the comparison we use <= instead of <
-    // params.p_dropout_in_uint = uint32_t(std::floor(params.p_dropout * 4294967295.0));
-    // params.p_dropout_in_uint16_t = uint16_t(std::floor(params.p_dropout * 65535.0));
-    params.p_dropout_in_uint8_t = uint8_t(std::floor(params.p_dropout * 255.0));
-    params.rp_dropout = 1.f / params.p_dropout;
-    params.scale_softmax_rp_dropout = params.rp_dropout * params.scale_softmax;
-    TORCH_CHECK(p_dropout < 1.f);
-    #ifdef FLASHATTENTION_DISABLE_DROPOUT
-        TORCH_CHECK(p_dropout == 0.0f, "This flash attention build does not support dropout.");
-    #endif
-
-    // Causal is the special case where window_size_right == 0 and window_size_left < 0.
-    // Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
-    params.is_causal = window_size_left < 0 && window_size_right == 0;
-    params.per_block_mean = per_block_mean;
-    if (per_block_mean) {
-        params.seqlen_s = seqlen_q;
-    } else {
-        params.seqlen_s = flash::BLOCK_M; // size of BLOCK_M
-    }
-    if (window_size_left < 0 && window_size_right >= 0) { window_size_left = seqlen_k; }
-    if (window_size_left >= 0 && window_size_right < 0) { window_size_right = seqlen_k; }
-    params.window_size_left = window_size_left;
-    params.window_size_right = window_size_right;
-
-    #ifdef FLASHATTENTION_DISABLE_LOCAL
-        TORCH_CHECK(params.is_causal || (window_size_left < 0 && window_size_right < 0),
-            "This flash attention build does not support local attention.");
-    #endif
-
-    params.is_seqlens_k_cumulative = true;
-    params.is_bf16 = is_bf16;
-    params.single_level_p_quant = single_level_p_quant;
-    #ifdef FLASHATTENTION_DISABLE_UNEVEN_K
-        TORCH_CHECK(d == d_rounded, "This flash attention build does not support headdim not being a multiple of 32.");
-    #endif
-}
-
-template<bool IsBF16>
-void run_mha_fwd_dispatch_dtype(Flash_fwd_params &params, cudaStream_t stream) {
-    using OType = std::conditional_t<IsBF16, cutlass::bfloat16_t, cutlass::half_t>;
-    if (params.d == 64) {
-        run_mha_fwd_<cutlass::nv_float4_t<cutlass::float_e2m1_t>, 64, OType>(params, stream);
-    } else if (params.d == 128) {
-        run_mha_fwd_<cutlass::nv_float4_t<cutlass::float_e2m1_t>, 128, OType>(params, stream);
-    }
-}
-
-void run_mha_fwd(Flash_fwd_params &params, cudaStream_t stream, bool force_split_kernel = false) {
-    BOOL_SWITCH(params.is_bf16, IsBF16, ([&] {
-        run_mha_fwd_dispatch_dtype<IsBF16>(params, stream);
-    }));
-}
-
-std::vector<at::Tensor>
-mha_fwd(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head_size // 2)
-        const at::Tensor &k,         // batch_size x seqlen_k x num_heads_k x (head_size // 2)
-        const at::Tensor &v,         // batch_size x seqlen_k x num_heads_k x (head_size // 2)
-        const at::Tensor &sfq,
-        const at::Tensor &sfk,
-        const at::Tensor &sfv,
-        const at::Tensor &delta_s,
-        int unpadded_k,
-        c10::optional<at::Tensor> &out_,             // batch_size x seqlen_q x num_heads x head_size
-        const float softmax_scale,
-        bool is_causal, 
-        bool per_block_mean,
-        bool is_bf16,
-        bool single_level_p_quant=false  // If true, use only per-row scale s_P2 (no per-block s_P1)
-    ) {
-
-    auto dprops = at::cuda::getCurrentDeviceProperties();
-    bool is_sm120 = dprops->major == 12 && dprops->minor == 0;
-    TORCH_CHECK(is_sm120, "only supports Blackwell GPUs or newer.");
-
-    auto q_dtype = q.dtype();
-    auto sfq_dtype = sfq.dtype();
-    TORCH_CHECK(q_dtype == torch::kUInt8, "q dtype must be uint8");
-    TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
-    TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
-    CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
-
-    TORCH_CHECK(sfq_dtype == torch::kFloat8_e4m3fn, "q dtype must be uint8");
-    TORCH_CHECK(sfk.dtype() == sfq_dtype, "query and key must have the same dtype");
-    TORCH_CHECK(sfv.dtype() == sfq_dtype, "query and value must have the same dtype");
-    CHECK_DEVICE(sfq); CHECK_DEVICE(sfk); CHECK_DEVICE(sfv);
-    
-    TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
-    TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
-    TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
-    TORCH_CHECK(delta_s.stride(-1) == 1, "Input tensor must have contiguous last dimension");
-
-    TORCH_CHECK(q.is_contiguous(), "Input tensor must be contiguous");
-    TORCH_CHECK(k.is_contiguous(), "Input tensor must be contiguous");
-    TORCH_CHECK(v.is_contiguous(), "Input tensor must be contiguous");
-
-    const auto sizes = q.sizes();
-    auto opts = q.options();
-    const int batch_size = sizes[0];
-    int seqlen_q = sizes[2];
-    int num_heads = sizes[1];
-    const int head_size_og = sizes[3];
-    const int unpacked_head_size = head_size_og * 2;
-    const int seqlen_k = k.size(2);
-    const int num_heads_k = k.size(1);
-    
-    TORCH_CHECK(batch_size > 0, "batch size must be postive");
-    TORCH_CHECK(unpacked_head_size <= 256, "FlashAttention forward only supports head dimension at most 256");
-    TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
-    TORCH_CHECK(num_heads == num_heads_k, "We do not support MQA/GQA yet");
-
-    TORCH_CHECK(unpacked_head_size == 64 || unpacked_head_size == 128 || unpacked_head_size == 256, "Only support head size 64, 128, and 256 for now");
-
-    CHECK_SHAPE(q, batch_size, num_heads, seqlen_q, head_size_og);
-    CHECK_SHAPE(k, batch_size, num_heads_k, seqlen_k, head_size_og);
-    CHECK_SHAPE(v, batch_size, num_heads_k, unpacked_head_size, seqlen_k/2);
-    // CHECK_SHAPE(delta_s, batch_size, num_heads, seqlen_q / 128, seqlen_k);
-    // CHECK_SHAPE(sfq, batch_size, seqlen_q, num_heads, unpacked_head_size);
-    // CHECK_SHAPE(sfk, batch_size, seqlen_k, num_heads_k, unpacked_head_size);
-    // CHECK_SHAPE(sfv, batch_size, unpacked_head_size, num_heads_k, seqlen_k);
-    TORCH_CHECK(unpacked_head_size % 8 == 0, "head_size must be a multiple of 8");
-    
-    auto dtype = is_bf16 ? at::ScalarType::BFloat16 : at::ScalarType::Half;
-    at::Tensor out = torch::empty({batch_size, num_heads, seqlen_q, unpacked_head_size}, opts.dtype(dtype));
-
-    auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
-    // const int head_size = round_multiple(head_size_og, 8);
-    // const int head_size_rounded = round_multiple(head_size, 32);
-    const int seqlen_q_rounded = round_multiple(seqlen_q, flash::BLOCK_M);
-    const int seqlen_k_rounded = round_multiple(seqlen_k, flash::BLOCK_N);
-
-    // Otherwise the kernel will be launched from cuda:0 device
-    // Cast to char to avoid compiler warning about narrowing
-    at::cuda::CUDAGuard device_guard{(char)q.get_device()};
-
-    
-
-    auto softmax_lse = torch::empty({batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
-    at::Tensor p;
-
-    Flash_fwd_params params;
-    set_params_fprop(params,
-                     batch_size,
-                     seqlen_q, seqlen_k, unpadded_k,
-                     seqlen_q_rounded, seqlen_k_rounded,
-                     num_heads, num_heads_k,
-                     unpacked_head_size, unpacked_head_size,
-                     q, k, v, delta_s, out, 
-                     sfq, sfk, sfv,
-                     /*cu_seqlens_q_d=*/nullptr,
-                     /*cu_seqlens_k_d=*/nullptr,
-                     /*seqused_k=*/nullptr,
-                     nullptr,
-                     softmax_lse.data_ptr(),
-                     /*p_dropout=*/0.f,
-                     softmax_scale,
-                     /*window_size_left=*/-1,
-                     /*window_size_right=*/is_causal ? 0 : -1,
-                     per_block_mean,
-                     is_bf16,
-                     single_level_p_quant
-                    );
-    // TODO: 132 sm count?
-    auto tile_count_semaphore = is_causal ? torch::full({1}, 132, opts.dtype(torch::kInt32)) : torch::empty({1}, opts.dtype(torch::kInt32));
-    params.tile_count_semaphore = tile_count_semaphore.data_ptr<int>();
-
-    if (seqlen_k > 0) {
-        auto stream = at::cuda::getCurrentCUDAStream().stream();
-        run_mha_fwd(params, stream);
-    } else {
-        // If seqlen_k == 0, then we have an empty tensor. We need to set the output to 0.
-        out.zero_();
-        softmax_lse.fill_(std::numeric_limits<float>::infinity());
-    }
-
-    // at::Tensor out_padded = out;
-    // if (head_size_og % 8 != 0) {
-    //     out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
-    //     if (out_.has_value()) { out_.value().copy_(out); }
-    // }
-
-    // return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p};
-    // cudaDeviceSynchronize();
-    // auto err = cudaGetLastError();
-    // printf("%s\n", cudaGetErrorString(err));
-    return {out, softmax_lse};
-}
-
-
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-    m.doc() = "FlashAttention";
-    m.def("fwd", &mha_fwd, "Forward pass");
-}

fastvideo-kernel/attn_qat_infer/blackwell/block_config.h

@@ -1,28 +0,0 @@
-/*
- * Copyright (c) 2025 by SageAttention team.
- * 
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#pragma once
-
-// Centralized block size configuration for sageattn_blackwell kernels
-// Block sizes for M and N dimensions
-namespace flash {
-    // Block size for M dimension (query sequence length)
-    static constexpr int BLOCK_M = 128;
-    
-    // Block size for N dimension (key/value sequence length)
-    static constexpr int BLOCK_N = 128;
-}
-