[bugfix]: SDPA backend should not restrict attention head sizes

SDPABackend.get_supported_head_sizes() declared [32, 64, 96, 128, 160,
192, 224, 256] -- a list copied verbatim from FlashAttentionBackend in
the v1 refactor (#270). torch's F.scaled_dot_product_attention is
head-size-agnostic (the math backend handles any size; fused kernels
fall back internally), so the list wrongly excluded valid sizes such as
80, the head size of CLIP-L vision encoders (e.g. matrixgame2's image
encoder). Under capability validation (#1494) this produces spurious
'may be incompatible' warnings on the head-size fallback path and would
hard-fail an explicit TORCH_SDPA pin on a head-80 layer.

Declare no restriction (None) instead, matching the capability API's
unrestricted convention, and add CPU tests asserting head-80 SDPA is
allowed, resolves under an explicit TORCH_SDPA pin, and computes
correct attention output.

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

@@ -69,7 +69,7 @@ approval, then upload reviewed accepted baseline records.
 | `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_results` | Yes | One or more local paths or Buildkite artifact URLs for accepted shifted performance JSONs. Prefer normalized `normalized_perf_*.json` artifacts emitted by `compare_baseline.py`. Accept `source_result` as an alias only for a single JSON. |
-| `max_intra_batch_regression` | No | Maximum allowed regression of any source JSON against the source batch median. Default: `PERF_MAX_REGRESSION` if set, otherwise `0.05` (5%). |
+| `max_intra_batch_regression` | No | Maximum allowed regression of any source JSON against the source batch median. Default: `0.05` (5%). |
 | `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:
@@ -148,8 +148,7 @@ For each metric with at least two non-null source values:
 4. Stop if any source record regresses against the batch median by more than
    `max_intra_batch_regression`.
 
-Default `max_intra_batch_regression` to `PERF_MAX_REGRESSION` when set,
-otherwise `0.05`. Print a table with per-source values, batch median, and
+Default `max_intra_batch_regression` to `0.05`. Print a table with per-source values, batch median, and
 worst intra-batch regression.
 
 This check prevents uploading a mixed batch where one JSON is materially
@@ -183,7 +182,7 @@ present, that run is not a valid source for baseline reseeding.
 
 ### 2. Sync and back up existing HF records under /tmp
 
-Use `fastvideo/tests/performance/hf_store.py` helpers directly. Do **not** use
+Use `fastvideo/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.
 
@@ -192,7 +191,7 @@ 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)'
+python -c 'from fastvideo.performance.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 under `/tmp`:
@@ -200,8 +199,8 @@ Then back up only the sanitized model directory under `/tmp`:
 ```bash
 SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
 TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
-MODEL_SAFE=$(PYTHONPATH=fastvideo/tests/performance python - <<'PY'
-from hf_store import sanitize
+MODEL_SAFE=$(python - <<'PY'
+from fastvideo.performance.hf_store import sanitize
 print(sanitize("<model_id>"))
 PY
 )
@@ -235,7 +234,7 @@ first baseline seed. Continue, but report that baseline history was empty.
 Load the last 5 successful records for the target:
 
 ```python
-from hf_store import load_records_for_model
+from fastvideo.performance.hf_store import load_records_for_model
 
 records = load_records_for_model(
     "/tmp/perf-tracking",
@@ -372,7 +371,7 @@ prepared records plus backup on disk.
 Use the shared storage helper so the path and repo type match CI:
 
 ```python
-from hf_store import upload_record
+from fastvideo.performance.hf_store import upload_record
 
 upload_record("<local_record_path>", record, strict=True)
 ```
@@ -460,7 +459,7 @@ directories created for this reseed. Never remove unrelated `/tmp` contents.
   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,
+- `fastvideo/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/wan-t2v-1.3b.json

@@ -1,5 +1,9 @@
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
+  "config_schema_version": 2,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "description": "Wan2.1 T2V 1.3B inference performance",
   "model": {
     "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",

.buildkite/pipeline.yml

@@ -114,6 +114,17 @@ steps:
             limit: 2
       agents:
         queue: "default"
+    - label: ":test_tube: LoRA Extraction Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "lora_extraction"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
     - label: ":test_tube: Training Tests"
       if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "training"
       command: "timeout 90m .buildkite/scripts/pr_test.sh"
@@ -371,6 +382,21 @@ steps:
                   - TEST_TYPE=inference_lora
                 agents:
                   queue: "default"
+            - path:
+                - "scripts/lora_extraction/**"
+                - "fastvideo/tests/lora_extraction/**"
+                - "fastvideo/models/loader/**"
+                - "fastvideo/training/training_utils.py"
+                - "fastvideo/layers/lora/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile"
+              config:
+                command: "timeout 90m .buildkite/scripts/pr_test.sh"
+                label: ":test_tube: LoRA Extraction Tests"
+                env:
+                  - TEST_TYPE=lora_extraction
+                agents:
+                  queue: "default"
             - path:
                 - "fastvideo/**"
                 - "pyproject.toml"
@@ -410,7 +436,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile"
               config:
-                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                command: "timeout 25m .buildkite/scripts/pr_test.sh"
                 label: ":test_tube: LoRA Training Tests"
                 env:
                   - TEST_TYPE=training_lora
@@ -455,6 +481,7 @@ steps:
                 - "fastvideo/layers/**"
                 - "fastvideo/worker/**"
                 - "fastvideo/entrypoints/**"
+                - "fastvideo/performance/**"
                 - "fastvideo/tests/performance/**"
                 - ".buildkite/performance-benchmarks/**"
                 - "pyproject.toml"

.buildkite/scripts/pr_test.sh

@@ -80,6 +80,23 @@ MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUI
 
 POST_RUN_HOOK=""
 
+is_truthy() {
+    case "${1:-}" in
+        1|true|TRUE|yes|YES|on|ON) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+ssim_bootstrap_args() {
+    local title="${PR_TITLE:-}"
+    local message="${BUILDKITE_MESSAGE:-}"
+    if is_truthy "${FASTVIDEO_SSIM_BOOTSTRAP_MODE:-}" \
+        || [[ "$title" == *"[new-model]"* ]] \
+        || [[ "$message" == *"[new-model]"* ]]; then
+        printf ' --bootstrap-mode'
+    fi
+}
+
 upload_performance_artifacts() {
     SHORT_SHA=${BUILDKITE_COMMIT:0:7}
     LOCAL_DIR="downloaded_reports"
@@ -172,7 +189,12 @@ case "$TEST_TYPE" in
         ;;
     "ssim")
         log "Running SSIM tests..."
-        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_SSIM_TEST_FILE::run_ssim_tests"
+        SSIM_BOOTSTRAP_ARGS=$(ssim_bootstrap_args)
+        if [ -n "$SSIM_BOOTSTRAP_ARGS" ]; then
+            log "SSIM bootstrap mode enabled for new-model reference draft generation"
+        fi
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run "
+        MODAL_COMMAND+="$MODAL_SSIM_TEST_FILE::run_ssim_tests$SSIM_BOOTSTRAP_ARGS"
         ;;
     "training")
         log "Running training tests..."

.github/scripts/gate_full_suite.sh

@@ -0,0 +1,133 @@
+#!/usr/bin/env bash
+# Gate the expensive Buildkite full suite on the cheap GitHub checks.
+#
+# Polls the workflow runs for the PR head commit and only exits 0 once the
+# watched cheap workflows (pre-commit, docs build) have succeeded, so the
+# 'ready' label cannot burn ~20 GPU lanes on a head that a cheap check has
+# already doomed.
+#
+# Semantics:
+#   - watched run completed with a bad conclusion  -> exit 1 (fail CLOSED:
+#     no full suite; the next push re-arms via the 'synchronize' trigger)
+#   - watched run cancelled                         -> still pending: the docs
+#     workflow's repo-global 'pages' concurrency group cancels runs superseded
+#     by unrelated pushes, so 'cancelled' is not a verdict on this PR
+#   - watched runs pending                          -> poll until done
+#   - docs run absent                               -> not applicable after a
+#     short grace period ('Deploy Documentation' is path-filtered on PRs)
+#   - pre-commit run absent                         -> keep polling: pre-commit
+#     is never path-filtered, so its absence is always anomalous
+#   - 'ready' label removed while waiting           -> exit 1 (fail CLOSED:
+#     un-labeling is a deliberate maintainer action)
+#   - GitHub API unreachable or timeout             -> exit 0 (fail OPEN,
+#     loud warning: never brick CI on a GitHub outage)
+#
+# Required env: PR_SHA (PR head commit), PR_NUMBER, GITHUB_REPOSITORY, GH_TOKEN.
+set -euo pipefail
+
+: "${PR_SHA:?PR_SHA (PR head commit) is required}"
+: "${PR_NUMBER:?PR_NUMBER (pull request number) is required}"
+: "${GITHUB_REPOSITORY:?GITHUB_REPOSITORY is required}"
+
+# Workflow-level `name:` values that must be green before the full suite
+# may start. "Deploy Documentation" is path-filtered on PRs, so its run may
+# legitimately never exist; pre-commit always runs, so it must appear.
+WATCHED_NAMES='["pre-commit", "Deploy Documentation"]'
+WATCHED_REGEX='^(pre-commit|Deploy Documentation)$'
+POLL_SECS="${POLL_SECS:-20}"
+GRACE_SECS="${GRACE_SECS:-60}"
+MAX_WAIT_SECS="${MAX_WAIT_SECS:-1500}"
+
+# Bound each API call so a hung connection hits the 3-strike fail-open path
+# instead of pinning the loop until the job timeout (which would fail closed
+# on exactly the GitHub-outage case this script is meant to survive).
+if command -v timeout >/dev/null 2>&1; then
+  gh_api() { timeout 30 gh api "$@"; }
+else
+  gh_api() { gh api "$@"; }  # macOS dev boxes; CI always has coreutils timeout
+fi
+
+# The workflow checked the label before starting the gate, but the wait can
+# last ~25 min: re-check once before any exit 0 and fail closed if 'ready'
+# was removed in the meantime. An API error here proceeds (the label was
+# present when the gate started; never brick CI on an outage).
+recheck_ready_label() {
+  local pr_json
+  if pr_json=$(gh_api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" 2>/dev/null); then
+    if ! jq -e '[.labels[]?.name] | index("ready")' <<<"$pr_json" >/dev/null 2>&1; then
+      echo "::error::PR #${PR_NUMBER} no longer has the 'ready' label —" \
+        "NOT triggering the Buildkite full suite. Re-add the label to re-arm."
+      exit 1
+    fi
+  else
+    echo "::warning::Could not re-check the 'ready' label on PR #${PR_NUMBER}; proceeding (it was present when the gate started)."
+  fi
+}
+
+start=$(date +%s)
+api_fails=0
+missing=""
+
+while true; do
+  elapsed=$(( $(date +%s) - start ))
+
+  if runs_json=$(gh_api "repos/${GITHUB_REPOSITORY}/actions/runs?head_sha=${PR_SHA}&per_page=100" 2>/dev/null) \
+     && state=$(jq --arg re "$WATCHED_REGEX" '
+          [.workflow_runs[]? | select(.name // "" | test($re))]
+          | group_by(.name) | map(max_by(.id))
+          | map({name, status, conclusion})' <<<"$runs_json" 2>/dev/null); then
+    api_fails=0
+    echo "t+${elapsed}s watched checks: $(jq -c . <<<"$state")"
+
+    failed=$(jq -r '[.[] | select(.status == "completed"
+          and (.conclusion | IN("success", "skipped", "neutral", "cancelled") | not))]
+        | map(.name) | join(", ")' <<<"$state")
+    if [ -n "$failed" ]; then
+      echo "::error::Cheap check(s) failed on ${PR_SHA}: ${failed}." \
+        "NOT triggering the Buildkite full suite. Push a fix (the 'ready'" \
+        "label re-arms on every push), or re-run the failed check and then" \
+        "re-run this workflow."
+      exit 1
+    fi
+
+    # 'cancelled' counts as pending: wait for a re-run to reach a real verdict
+    # (bounded by MAX_WAIT, then the fail-open below).
+    pending=$(jq '[.[] | select(.status != "completed" or .conclusion == "cancelled")] | length' <<<"$state")
+    missing=$(jq -r --argjson watched "$WATCHED_NAMES" '($watched - map(.name)) | join(", ")' <<<"$state")
+    if [ "$pending" -eq 0 ]; then
+      if [ -z "$missing" ]; then
+        recheck_ready_label
+        echo "All watched cheap checks are green — full suite may proceed."
+        exit 0
+      fi
+      case "$missing" in
+        *pre-commit*)
+          echo "pre-commit run not found for ${PR_SHA} yet; waiting (pre-commit is never path-filtered, so its absence is anomalous)."
+          ;;
+        *)
+          if [ "$elapsed" -ge "$GRACE_SECS" ]; then
+            recheck_ready_label
+            echo "::warning::Watched run(s) never appeared for ${PR_SHA}: ${missing} (path-filtered, likely not applicable). Proceeding on the checks that did run."
+            exit 0
+          fi
+          echo "Waiting up to ${GRACE_SECS}s grace for path-filtered run(s) to appear: ${missing}."
+          ;;
+      esac
+    fi
+  else
+    api_fails=$(( api_fails + 1 ))
+    echo "::warning::GitHub API error querying workflow runs for ${PR_SHA} (attempt ${api_fails}/3)."
+    if [ "$api_fails" -ge 3 ]; then
+      recheck_ready_label
+      echo "::warning::FAILING OPEN: cannot query GitHub check status — triggering the full suite WITHOUT the cheap-check gate."
+      exit 0
+    fi
+  fi
+
+  if [ "$elapsed" -ge "$MAX_WAIT_SECS" ]; then
+    recheck_ready_label
+    echo "::warning::FAILING OPEN: watched checks still pending after $(( MAX_WAIT_SECS / 60 )) min${missing:+ (never appeared: ${missing})} — triggering the full suite anyway."
+    exit 0
+  fi
+  sleep "$POLL_SECS"
+done

.github/scripts/test_gate_full_suite.sh

@@ -0,0 +1,122 @@
+#!/usr/bin/env bash
+# Self-test for gate_full_suite.sh using a mocked `gh`. No network, runs on
+# any dev box: bash .github/scripts/test_gate_full_suite.sh
+set -u
+here=$(cd "$(dirname "$0")" && pwd)
+tmp=$(mktemp -d)
+trap 'rm -rf "$tmp"' EXIT
+
+# Mock gh. Asserts the exact endpoint (including head_sha) it is called
+# with — an endpoint typo in the gate script fails the test rather than
+# silently serving canned data. On the runs endpoint it serves
+# $MOCK_DIR/response_<call#>.json, sticking on the highest existing file,
+# and exits 1 if none exist (simulates a GitHub API outage). On the pulls
+# endpoint it serves $MOCK_DIR/pr.json, defaulting to a 'ready'-labeled PR.
+cat > "$tmp/gh" <<'EOF'
+#!/usr/bin/env bash
+if [ "${1:-}" != "api" ]; then
+  echo "unexpected gh invocation: $*" >> "$MOCK_DIR/endpoint_error"
+  exit 2
+fi
+case "${2:-}" in
+  "repos/o/r/actions/runs?head_sha=deadbeef&per_page=100")
+    n=$(( $(cat "$MOCK_DIR/count" 2>/dev/null || echo 0) + 1 ))
+    echo "$n" > "$MOCK_DIR/count"
+    while [ "$n" -gt 0 ]; do
+      if [ -f "$MOCK_DIR/response_$n.json" ]; then
+        cat "$MOCK_DIR/response_$n.json"
+        exit 0
+      fi
+      n=$(( n - 1 ))
+    done
+    echo "api outage" >&2
+    exit 1
+    ;;
+  "repos/o/r/pulls/42")
+    if [ -f "$MOCK_DIR/pr.json" ]; then
+      cat "$MOCK_DIR/pr.json"
+    else
+      echo '{"labels": [{"name": "ready"}]}'
+    fi
+    ;;
+  *)
+    echo "unexpected gh endpoint: $2" >> "$MOCK_DIR/endpoint_error"
+    exit 2
+    ;;
+esac
+EOF
+chmod +x "$tmp/gh"
+
+PC_OK='{"name": "pre-commit", "id": 1, "status": "completed", "conclusion": "success"}'
+PC_BAD='{"name": "pre-commit", "id": 1, "status": "completed", "conclusion": "failure"}'
+PC_PENDING='{"name": "pre-commit", "id": 1, "status": "in_progress", "conclusion": null}'
+DOCS_OK='{"name": "Deploy Documentation", "id": 2, "status": "completed", "conclusion": "success"}'
+DOCS_BAD='{"name": "Deploy Documentation", "id": 2, "status": "completed", "conclusion": "failure"}'
+DOCS_CANCELLED='{"name": "Deploy Documentation", "id": 2, "status": "completed", "conclusion": "cancelled"}'
+OTHER='{"name": "Trigger Full Suite", "id": 3, "status": "in_progress", "conclusion": null}'
+NULL_NAME='{"name": null, "id": 4, "status": "completed", "conclusion": "failure"}'
+PC_OK_RERUN='{"name": "pre-commit", "id": 5, "status": "completed", "conclusion": "success"}'
+
+fails=0
+want_log=""  # optional: expect() also greps out.log for this regex, then resets
+pr_json=""   # optional: served for the pulls (label re-check) endpoint, then resets
+raw_body=""  # optional: serve responses verbatim instead of wrapping in workflow_runs
+expect() { # <name> <expected-exit> <response json>...
+  local name=$1 want=$2 dir i=1
+  shift 2
+  dir=$(mktemp -d "$tmp/test_XXXXXX")
+  for body in "$@"; do
+    if [ -n "$raw_body" ]; then
+      printf '%s' "$body" > "$dir/response_$i.json"
+    else
+      printf '{"workflow_runs": [%s]}' "$body" > "$dir/response_$i.json"
+    fi
+    i=$(( i + 1 ))
+  done
+  [ -n "$pr_json" ] && printf '%s' "$pr_json" > "$dir/pr.json"
+  ( export PATH="$tmp:$PATH" MOCK_DIR="$dir" PR_SHA=deadbeef PR_NUMBER=42 \
+      GITHUB_REPOSITORY=o/r POLL_SECS=0 GRACE_SECS=1 MAX_WAIT_SECS=3
+    bash "$here/gate_full_suite.sh" > "$dir/out.log" 2>&1 )
+  local rc=$?
+  if [ "$rc" -ne "$want" ]; then
+    echo "FAIL: $name (exit $rc, want $want)"
+    cat "$dir/out.log"
+    fails=1
+  elif [ -f "$dir/endpoint_error" ]; then
+    echo "FAIL: $name (mock gh got an unexpected call)"
+    cat "$dir/endpoint_error"
+    fails=1
+  elif [ -n "$want_log" ] && ! grep -Eq "$want_log" "$dir/out.log"; then
+    echo "FAIL: $name (log does not match: $want_log)"
+    cat "$dir/out.log"
+    fails=1
+  else
+    echo "ok: $name"
+  fi
+  want_log="" pr_json="" raw_body=""
+}
+
+expect "both green -> proceed" 0 "$PC_OK, $DOCS_OK, $OTHER, $NULL_NAME"
+expect "docs build failed -> blocked" 1 "$PC_OK, $DOCS_BAD"
+expect "pre-commit failed -> blocked" 1 "$PC_BAD"
+expect "pending then green -> proceed" 0 "$PC_PENDING" "$PC_OK, $DOCS_OK"
+want_log="never appeared.*Deploy Documentation"
+expect "docs run absent (path-filtered) -> proceed after grace" 0 "$PC_OK"
+expect "API outage -> fail open" 0
+want_log="FAILING OPEN"
+expect "pending past MAX_WAIT -> fail open" 0 "$PC_PENDING"
+want_log="FAILING OPEN"
+expect "unrelated runs only -> no grace, fail open at MAX_WAIT" 0 "$OTHER"
+expect "cancelled docs then green -> proceed" 0 \
+  "$PC_OK, $DOCS_CANCELLED" "$PC_OK, $DOCS_OK"
+want_log="FAILING OPEN"
+expect "cancelled docs forever -> fail open at MAX_WAIT" 0 "$PC_OK, $DOCS_CANCELLED"
+want_log="FAILING OPEN"
+expect "pre-commit absent -> no grace, fail open at MAX_WAIT" 0 "$DOCS_OK"
+expect "duplicate run names -> latest wins" 0 "$PC_BAD, $PC_OK_RERUN, $DOCS_OK"
+raw_body=1
+expect "garbage response body -> fail open" 0 "this is not json"
+pr_json='{"labels": [{"name": "other"}]}'
+expect "ready label removed mid-gate -> blocked" 1 "$PC_OK, $DOCS_OK"
+
+exit "$fails"

.github/workflows/ci-precommit.yml

@@ -1,7 +1,11 @@
 name: pre-commit
 
 on:
-  pull_request:
+  # pull_request_target instead of pull_request: the workflow definition and
+  # the hook config are always taken from the BASE branch, so fork /
+  # first-time-contributor PRs run immediately without a maintainer clicking
+  # "Approve and run". The PR head is checked out as data only.
+  pull_request_target:
     branches: [main]
   workflow_call:
     inputs:
@@ -15,12 +19,25 @@ permissions:
 
 jobs:
   pre-commit:
-    if: github.event_name == 'workflow_call' || github.event.pull_request.draft != true
+    if: github.event.pull_request.draft != true
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
       with:
         ref: ${{ inputs.ref || '' }}
+    # For PR events, lint the PR head — but keep the hook definitions from
+    # the base branch so an untrusted PR cannot alter what gets executed.
+    - name: Save trusted hook config
+      if: github.event_name == 'pull_request_target'
+      run: cp .pre-commit-config.yaml "$RUNNER_TEMP/trusted-pre-commit-config.yaml"
+    - uses: actions/checkout@v4
+      if: github.event_name == 'pull_request_target'
+      with:
+        ref: ${{ github.event.pull_request.head.sha }}
+        persist-credentials: false
+    - name: Restore trusted hook config
+      if: github.event_name == 'pull_request_target'
+      run: cp "$RUNNER_TEMP/trusted-pre-commit-config.yaml" .pre-commit-config.yaml
     - uses: actions/setup-python@v5
       with:
         python-version: "3.12"
@@ -30,3 +47,6 @@ jobs:
     - uses: pre-commit/action@v3.0.1
       with:
         extra_args: --all-files --hook-stage manual
+    # After pre-commit so a self-test failure cannot mask lint failures.
+    - name: Full-suite gate self-test
+      run: bash .github/scripts/test_gate_full_suite.sh

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

@@ -52,6 +52,7 @@ jobs:
             core.setOutput('pr_sha', pr.head.sha);
             core.setOutput('pr_branch', pr.head.ref);
             core.setOutput('pr_number', String(prNumber));
+            core.setOutput('pr_title', pr.title);
 
       - name: Trigger Full Suite
         if: steps.perm.outputs.has_write == 'true'
@@ -60,6 +61,7 @@ jobs:
           PR_SHA: ${{ steps.label.outputs.pr_sha }}
           PR_BRANCH: ${{ steps.label.outputs.pr_branch }}
           PR_NUMBER: ${{ steps.label.outputs.pr_number }}
+          PR_TITLE: ${{ steps.label.outputs.pr_title }}
           BK_ORG: ${{ vars.BUILDKITE_ORG_SLUG }}
           BK_PIPELINE: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -71,6 +73,7 @@ jobs:
               --arg commit "$PR_SHA" \
               --arg branch "$PR_BRANCH" \
               --arg message "Full Suite for PR #${PR_NUMBER} (via /merge)" \
+              --arg pr_title "$PR_TITLE" \
               --argjson pr_id "$PR_NUMBER" \
               '{
                 commit: $commit,
@@ -80,11 +83,12 @@ jobs:
                 pull_request_id: $pr_id,
                 pull_request_base_branch: "main",
                 env: {
-                  TEST_SCOPE: "full",
-                  FULL_SUITE: "true",
-                  PR_NUMBER: ($pr_id | tostring)
-                }
-              }')"
+                    TEST_SCOPE: "full",
+                    FULL_SUITE: "true",
+                    PR_NUMBER: ($pr_id | tostring),
+                    PR_TITLE: $pr_title
+                  }
+                }')"
 
   parse-command:
     if: >-
@@ -125,7 +129,7 @@ jobs:
           set -euo pipefail
           TEST_NAME=$(echo "$COMMENT" | grep -oP '(?<=/test\s)\S+' | head -1 || true)
 
-          VALID="encoder vae transformer kernel unit dreamverse ssim training lora-inference lora-training distillation self-forcing vsa vmoba performance api train-framework eval full fastcheck pre-commit"
+          VALID="encoder vae transformer kernel unit dreamverse ssim training lora-inference lora-training lora-extraction distillation self-forcing vsa vmoba performance api train-framework eval full fastcheck pre-commit"
           if [ -z "$TEST_NAME" ] || ! echo "$VALID" | grep -qw "$TEST_NAME"; then
             echo "Unknown test: '$TEST_NAME'. Valid: $VALID"
             exit 1
@@ -136,6 +140,7 @@ jobs:
             [kernel]=kernel_tests [unit]=unit_test [dreamverse]=dreamverse_app
             [ssim]=ssim [training]=training
             [lora-inference]=inference_lora [lora-training]=training_lora
+            [lora-extraction]=lora_extraction
             [distillation]=distillation_dmd [self-forcing]=self_forcing
             [vsa]=training_vsa [vmoba]=inference_vmoba
             [performance]=performance [api]=api_server
@@ -240,6 +245,7 @@ jobs:
           TEST_SCOPE: ${{ needs.parse-command.outputs.test_scope }}
           FULL_SUITE: ${{ needs.parse-command.outputs.full_suite }}
           TEST_TYPE: ${{ needs.parse-command.outputs.test_type }}
+          PR_TITLE: ${{ github.event.issue.title }}
           BK_ORG: ${{ vars.BUILDKITE_ORG_SLUG }}
           BK_PIPELINE: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -256,6 +262,7 @@ jobs:
               --arg full_suite "$FULL_SUITE" \
               --arg test_type "$TEST_TYPE" \
               --arg pr_number "$PR_NUMBER" \
+              --arg pr_title "$PR_TITLE" \
               '{
                 commit: $commit,
                 branch: $branch,
@@ -265,8 +272,9 @@ jobs:
                 pull_request_base_branch: "main",
                 env: {
                   TEST_SCOPE: $test_scope,
-                  FULL_SUITE: $full_suite,
-                  TEST_TYPE: $test_type,
-                  PR_NUMBER: $pr_number
-                }
-              }')"
+                    FULL_SUITE: $full_suite,
+                    TEST_TYPE: $test_type,
+                    PR_NUMBER: $pr_number,
+                    PR_TITLE: $pr_title
+                  }
+                }')"

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

@@ -7,6 +7,7 @@ on:
 permissions:
   contents: read
   pull-requests: read
+  actions: read
 
 concurrency:
   group: full-suite-${{ github.event.pull_request.number }}
@@ -18,6 +19,8 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'ready')
       || github.event.action == 'synchronize'
     runs-on: ubuntu-latest
+    # Gate below may wait for cheap checks (up to MAX_WAIT_SECS = 25 min).
+    timeout-minutes: 35
     steps:
       - name: Check ready label
         id: check
@@ -49,6 +52,20 @@ jobs:
               "https://api.buildkite.com/v2/organizations/${{ vars.BUILDKITE_ORG_SLUG }}/pipelines/${{ vars.BUILDKITE_PIPELINE_SLUG }}/builds/${build_num}/cancel"
           done
 
+      # Checks out the BASE branch (default for pull_request_target), so PR
+      # authors cannot tamper with the gate script.
+      - name: Checkout gate script
+        if: steps.check.outputs.has_ready == 'true'
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+
+      - name: Wait for pre-commit and docs build
+        if: steps.check.outputs.has_ready == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_SHA: ${{ github.event.pull_request.head.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+        run: bash .github/scripts/gate_full_suite.sh
+
       - name: Trigger Buildkite Full Suite
         if: steps.check.outputs.has_ready == 'true'
         env:
@@ -56,6 +73,7 @@ jobs:
           PR_SHA: ${{ github.event.pull_request.head.sha }}
           PR_BRANCH: ${{ github.event.pull_request.head.ref }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
+          PR_TITLE: ${{ github.event.pull_request.title }}
           BK_ORG: ${{ vars.BUILDKITE_ORG_SLUG }}
           BK_PIPELINE: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -67,6 +85,7 @@ jobs:
               --arg commit "$PR_SHA" \
               --arg branch "$PR_BRANCH" \
               --arg message "Full Suite for PR #${PR_NUMBER}" \
+              --arg pr_title "$PR_TITLE" \
               --argjson pr_id "$PR_NUMBER" \
               '{
                 commit: $commit,
@@ -78,6 +97,7 @@ jobs:
                 env: {
                   TEST_SCOPE: "full",
                   FULL_SUITE: "true",
-                  PR_NUMBER: ($pr_id | tostring)
+                  PR_NUMBER: ($pr_id | tostring),
+                  PR_TITLE: $pr_title
                 }
               }')"

.github/workflows/infra-build-image.yml

@@ -13,12 +13,33 @@ on:
         required: false
         default: false
         type: boolean
+  # Auto-rebuild the CUDA images when their Dockerfile changes on main. The CUDA
+  # matrix is the only lane that builds from docker/Dockerfile, so a path-scoped
+  # push trigger is a sufficient change detector on its own -- no separate
+  # detect-changes/paths-filter job is needed now that there is a single
+  # in-scope Dockerfile. Dreamverse (apps/dreamverse/docker/Dockerfile) and the
+  # rocm Dockerfile stay manual-dispatch only.
+  push:
+    branches: [main]
+    paths:
+      - 'docker/Dockerfile'
 
 
 permissions:
   contents: read
   packages: write
 
+# One static group, no cancellation: every run of this workflow writes the same
+# mutable registry tags (latest, py3.12-latest, ...), so runs must serialize —
+# concurrent push/dispatch runs would race on those tags, and cancelling a run
+# mid-publish can strand the cu126/cu130 tag families at different commits. An
+# in-flight superseded build wastes its runner time, but its tags are then
+# overwritten by the newer queued run. GitHub keeps a single pending run per
+# group: the newest queued run replaces any older queued one.
+concurrency:
+  group: infra-build-image
+  cancel-in-progress: false
+
 jobs:
   # CUDA matrix: Python 3.12 x {12.6.3, 13.0.0} x {amd64, arm64}. Each architecture
   # builds natively and pushes only by digest; publish-cuda-manifests is the sole
@@ -28,7 +49,11 @@ jobs:
   # aliases; 13.0.0/cu130 is published under explicit versioned tags. Flash-attn
   # 2.8.3 comes from the architecture-specific prebuilt releases.
   build-cuda-images:
-    if: ${{ github.event.inputs.build_cuda_matrix == 'true' }}
+    # Runs on a manual dispatch when build_cuda_matrix is set, or automatically
+    # on a push that changed docker/Dockerfile (inputs are null on push). The
+    # repository guard keeps fork syncs from auto-building; manual dispatch
+    # still works in forks.
+    if: ${{ (github.event_name == 'push' && github.repository == 'hao-ai-lab/FastVideo') || github.event.inputs.build_cuda_matrix == 'true' }}
     strategy:
       fail-fast: false
       matrix:
@@ -75,10 +100,10 @@ jobs:
     secrets: inherit
 
   publish-cuda-manifests:
-    # !cancelled(): a failed sibling build leg must not skip the manifests for a
-    # CUDA lane whose own digests all exist; the digest-count check below fails
-    # the incomplete lane loudly instead.
-    if: ${{ !cancelled() && github.event.inputs.build_cuda_matrix == 'true' }}
+    # !cancelled(): publish lanes whose digests exist even if a sibling build
+    # leg failed (the digest-count check fails incomplete lanes); it also
+    # bypasses skipped-needs propagation, hence the explicit skipped check.
+    if: ${{ !cancelled() && needs.build-cuda-images.result != 'skipped' }}
     needs: build-cuda-images
     runs-on: ubuntu-latest
     permissions:

.gitignore

@@ -72,8 +72,7 @@ docs/distillation/examples/
 # Python pickle files
 *.pkl
 
-# Reference videos
-!fastvideo/tests/ssim/reference_videos/**/*.mp4
+# Reference videos (negations must come after the catch-all on line below)
 
 # Static images
 !docs/assets/images/**/*.png
@@ -127,6 +126,8 @@ apps/dreamverse/web/.env.production.local
 .sisyphus/
 openspec/
 fastvideo/tests/ssim/reference_videos/**
+!fastvideo/tests/ssim/reference_videos/**/*.mp4
+!fastvideo/tests/ssim/reference_videos/**/*.png
 
 # Editor logs and local Python version pins (accidentally committed)
 *.nvimlog

apps/dreamverse/docker/Dockerfile

@@ -84,9 +84,12 @@ RUN source /opt/venv/bin/activate \
        FFMPEG_NATIVE_CXX=/usr/bin/g++ \
        bash /opt/FastVideo/apps/dreamverse/scripts/install_native_ffmpeg.sh
 
+# FASTVIDEO_FA4: FA4 (flash_attn.cute) is opt-in; this image installs it via
+# the dreamverse extra and is validated with it, so enable it here.
 ENV FASTVIDEO_DREAMVERSE_HOME=/var/lib/dreamverse \
     STREAM_MODE=av_fmp4 \
     FASTVIDEO_ENABLE_PROMPT_SAFETY=0 \
+    FASTVIDEO_FA4=1 \
     HF_HOME=/root/.cache/huggingface
 
 RUN mkdir -p /var/lib/dreamverse

apps/performance_dashboard/README.md

@@ -12,13 +12,17 @@ Defaults:
 
 - `HF_REPO_ID=FastVideo/performance-tracking`
 - `PERFORMANCE_TRACKING_ROOT=/tmp/fastvideo-perf-dashboard`
-- `PERF_MAX_REGRESSION=0.05`
 
-Records can include source metadata:
+Records can include source metadata and rolling-baseline policy context:
 
 - `run_source`: `pr`, `local`, `scheduled_main`, or `unknown`
 - `baseline_eligible`: only successful scheduled-main records should be true
 - Buildkite metadata such as branch, PR number, build URL, build ID, and job ID
+- `regression_thresholds`: per-metric rolling-baseline percent and absolute
+  floors used for recomputed status context
+
+Dashboard/API metric payloads expose `threshold_exceeded` for raw threshold
+crossings; `regressed` remains the gated CI-failure signal.
 
 Set one of `HF_API_KEY`, `HUGGINGFACE_HUB_TOKEN`, or `HF_TOKEN` if the
 configured dataset repo requires authenticated access:
@@ -89,7 +93,8 @@ Trend charts show metric-specific axes and exact point details on hover/focus:
 - PR number, branch, and Buildkite URL when present
 
 The latest status table uses the stored JSON `success` value. Recomputed
-baseline context is shown separately and does not override stored status.
+baseline context applies each metric's percent and absolute regression floors
+and does not override stored status.
 
 ## API
 

apps/performance_dashboard/frontend/src/App.tsx

@@ -1,6 +1,7 @@
 import { useEffect, useMemo, useState } from "react";
 
-import { fetchSummary, fetchTrends, refreshData, RunSource, SummaryResponse, TrendGroup, TrendPoint } from "./api";
+import { fetchSummary, fetchTrends, refreshData } from "./api";
+import type { CohortValue, RunSource, SummaryResponse, TrendGroup, TrendPoint } from "./api";
 
 const METRIC_KEYS = ["latency", "throughput", "memory", "text_encoder_time_s", "dit_time_s", "vae_decode_time_s"];
 const RUN_SOURCES: Array<{ value: "" | RunSource; label: string }> = [
@@ -109,6 +110,61 @@ function metricLabel(metricKey: string) {
   return METRIC_DEFINITIONS[metricKey]?.label ?? metricKey;
 }
 
+type CohortFields = {
+  model_id: string;
+  gpu_type: string;
+  workload_id: CohortValue;
+  variant_id: CohortValue;
+  benchmark_version: CohortValue;
+  recipe_fingerprint: CohortValue;
+  hardware_profile_id: CohortValue;
+  software_profile_id: CohortValue;
+};
+
+function cohortValue(value: CohortValue) {
+  if (value === null || value === undefined || value === "") {
+    return "legacy";
+  }
+  return String(value);
+}
+
+function shortCohortValue(value: CohortValue) {
+  const text = cohortValue(value);
+  if (text === "legacy" || text.length <= 14) {
+    return text;
+  }
+  return text.slice(0, 12);
+}
+
+function cohortKey(cohort: CohortFields) {
+  return [
+    cohort.model_id,
+    cohort.gpu_type,
+    cohortValue(cohort.workload_id),
+    cohortValue(cohort.variant_id),
+    cohortValue(cohort.benchmark_version),
+    cohortValue(cohort.recipe_fingerprint),
+    cohortValue(cohort.hardware_profile_id),
+    cohortValue(cohort.software_profile_id)
+  ].join("|");
+}
+
+function cohortTitle(cohort: CohortFields) {
+  const workload = cohortValue(cohort.workload_id);
+  const variant = cohortValue(cohort.variant_id);
+  const version = cohortValue(cohort.benchmark_version);
+  const versionLabel = version === "legacy" ? version : `v${version}`;
+  return `${workload} / ${variant} / ${versionLabel}`;
+}
+
+function cohortDetail(cohort: CohortFields) {
+  return [
+    `recipe ${shortCohortValue(cohort.recipe_fingerprint)}`,
+    shortCohortValue(cohort.hardware_profile_id),
+    shortCohortValue(cohort.software_profile_id)
+  ].join(" | ");
+}
+
 function formatMetricValue(metricKey: string, value: number | null | undefined, tooltip = false) {
   const definition = METRIC_DEFINITIONS[metricKey];
   if (!definition) {
@@ -171,7 +227,9 @@ function TrendChart({ group, metricKey }: { group: TrendGroup; metricKey: string
         top: `${(activePoint.y / height) * 100}%`
       }
     : undefined;
-  const ariaLabel = `${metricLabel(metricKey)} trend for ${group.model_id} on ${group.gpu_type}`;
+  const ariaLabel = `${metricLabel(metricKey)} trend for ${group.model_id} on ${group.gpu_type}, ${cohortTitle(
+    group
+  )}`;
 
   return (
     <div className="chart-shell">
@@ -419,7 +477,7 @@ export default function App() {
       <section className="panel">
         <div className="panel-header">
           <h2>Latest Status</h2>
-          <span>{latestRows.length} model/GPU groups</span>
+          <span>{latestRows.length} comparison cohorts</span>
         </div>
         {latestRows.length === 0 ? (
           <div className="empty">No records match the selected filters.</div>
@@ -432,6 +490,7 @@ export default function App() {
                   <th>Recomputed</th>
                   <th>Model</th>
                   <th>GPU</th>
+                  <th>Cohort</th>
                   <th>Commit</th>
                   <th>Source</th>
                   <th>Baseline</th>
@@ -440,11 +499,13 @@ export default function App() {
                   <th>Throughput</th>
                   <th>Memory</th>
                   <th>Worst</th>
+                  <th>Exceeded</th>
+                  <th>Failing</th>
                 </tr>
               </thead>
               <tbody>
                 {latestRows.map((row) => (
-                  <tr key={`${row.model_id}-${row.gpu_type}`}>
+                  <tr key={cohortKey(row)}>
                     <td>
                       <span className={`badge ${row.status}`}>{row.status}</span>
                     </td>
@@ -455,6 +516,12 @@ export default function App() {
                     </td>
                     <td>{row.model_id}</td>
                     <td>{row.gpu_type}</td>
+                    <td>
+                      <div className="cohort-cell">
+                        <strong>{cohortTitle(row)}</strong>
+                        <span>{cohortDetail(row)}</span>
+                      </div>
+                    </td>
                     <td>{shortSha(row.commit_sha)}</td>
                     <td>
                       <span className={`source-badge source-${row.run_source}`}>{runSourceLabel(row.run_source)}</span>
@@ -465,6 +532,12 @@ export default function App() {
                     <td>{formatNumber(row.metrics.throughput?.current, 3)}</td>
                     <td>{formatNumber(row.metrics.memory?.current, 1)}</td>
                     <td>{formatNumber(row.worst_regression_pct, 1)}%</td>
+                    <td>
+                      {row.threshold_exceeded_metrics.length
+                        ? row.threshold_exceeded_metrics.join(", ")
+                        : "none"}
+                    </td>
+                    <td>{row.failing_metrics.length ? row.failing_metrics.join(", ") : "none"}</td>
                   </tr>
                 ))}
               </tbody>
@@ -487,11 +560,13 @@ export default function App() {
           ) : (
             trends.map((group) =>
               METRIC_KEYS.map((metricKey) => (
-                <article className="trend-card" key={`${group.model_id}-${group.gpu_type}-${metricKey}`}>
+                <article className="trend-card" key={`${cohortKey(group)}-${metricKey}`}>
                   <div>
                     <h3>{metricLabel(metricKey)}</h3>
                     <p>
                       {group.model_id} | {group.gpu_type}
+                      <span>{cohortTitle(group)}</span>
+                      <span>{cohortDetail(group)}</span>
                     </p>
                   </div>
                   <TrendChart group={group} metricKey={metricKey} />

apps/performance_dashboard/frontend/src/api.ts

@@ -2,11 +2,28 @@ export type MetricValue = {
   current: number | null;
   baseline: number | null;
   regression_pct: number | null;
+  absolute_delta: number | null;
+  threshold_percent: number;
+  threshold_absolute: number;
+  gated: boolean;
+  threshold_exceeded: boolean;
+  regressed: boolean;
   label: string;
   lower_is_better: boolean;
   precision: number;
 };
 
+export type CohortValue = string | number | null;
+
+export type ComparisonCohort = {
+  workload_id: CohortValue;
+  variant_id: CohortValue;
+  benchmark_version: CohortValue;
+  recipe_fingerprint: CohortValue;
+  hardware_profile_id: CohortValue;
+  software_profile_id: CohortValue;
+};
+
 export type SummaryRow = {
   model_id: string;
   gpu_type: string;
@@ -15,7 +32,8 @@ export type SummaryRow = {
   success: boolean;
   baseline_n: number;
   worst_regression_pct: number | null;
-  regression_threshold_pct: number;
+  threshold_exceeded_metrics: string[];
+  failing_metrics: string[];
   computed_regression_status: "pass" | "fail";
   status: "pass" | "fail";
   run_source: RunSource;
@@ -27,7 +45,7 @@ export type SummaryRow = {
   build_id: string;
   job_id: string;
   metrics: Record<string, MetricValue>;
-};
+} & ComparisonCohort;
 
 export type RunSource = "pr" | "local" | "scheduled_main" | "unknown";
 
@@ -61,13 +79,13 @@ export type TrendPoint = {
   build_id: string;
   job_id: string;
   metrics: Record<string, number | null>;
-};
+} & ComparisonCohort;
 
 export type TrendGroup = {
   model_id: string;
   gpu_type: string;
   points: TrendPoint[];
-};
+} & ComparisonCohort;
 
 export type TrendsResponse = {
   groups: TrendGroup[];

apps/performance_dashboard/frontend/src/styles.css

@@ -149,6 +149,11 @@ h3 {
   font-size: 0.82rem;
 }
 
+.trend-card p {
+  display: grid;
+  gap: 2px;
+}
+
 .stat strong {
   display: block;
   margin-top: 8px;
@@ -186,7 +191,7 @@ h3 {
 
 table {
   width: 100%;
-  min-width: 1120px;
+  min-width: 1260px;
   border-collapse: collapse;
 }
 
@@ -209,6 +214,25 @@ td {
   font-size: 0.9rem;
 }
 
+.cohort-cell {
+  display: grid;
+  gap: 2px;
+}
+
+.cohort-cell strong,
+.trend-card p span {
+  color: #1b2836;
+  font-size: 0.78rem;
+  font-weight: 700;
+}
+
+.cohort-cell span,
+.trend-card p span + span {
+  color: #607080;
+  font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", monospace;
+  font-size: 0.72rem;
+}
+
 .badge {
   display: inline-flex;
   align-items: center;

assets/eval/worldmodel_synthetic_flow_calibration.json

@@ -0,0 +1,22 @@
+{
+  "alpha_yaw": 0.08734091699186919,
+  "alpha_pitch": 0.08169667696275307,
+  "alpha_turn": 5.724587470723463e-17,
+  "beta_fwd": 0.02842768078408099,
+  "beta_strafe": 0.022531015077067108,
+  "focal_length": 457.0,
+  "frame_shape": [
+    352,
+    640
+  ],
+  "calibrated_from": [
+    "1_wasd_only",
+    "camera",
+    "camera4hold_alpha1",
+    "fully_random",
+    "wasdonly_alpha1",
+    "wasd4holdrandview_simple_1key1mouse1"
+  ],
+  "residual_rms": 15.890399609478676,
+  "n_equations": 4125232
+}

docker/Dockerfile

@@ -68,7 +68,7 @@ ARG FLASH_ATTN_WHEEL_RELEASE_ARM64=https://github.com/mjun0812/flash-attention-p
 # cutlass-4.4 `cute.core.ThrMma` API, which crashes on the cutlass-dsl 4.5 that
 # flashinfer/quack pull in. After the wheel install we overlay this cutlass-4.5-safe
 # upstream cute (flash-attn-4) so the image runs FA4 instead of the FA2 fallback.
-ARG FA4_CUTE_REF=940cd9680f3315f2f06b43ab5bea2c2cf2d96806
+ARG FA4_CUTE_REF=82d6441eec5d4dfec120153db2c0145ae855a083
 
 # Provided automatically by BuildKit/buildx (e.g. "amd64" / "arm64") and used to
 # select the prebuilt flash-attn wheel. Empty under a plain `docker build` without
@@ -161,7 +161,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
         uv pip install flash-attn==${FLASH_ATTN_VERSION} --no-build-isolation; \
     fi
 
-# Overlay the cutlass-4.5-safe upstream FA4 cute (FA4_CUTE_REF) over the
+# Overlay the CuTe-DSL-4.6-compatible upstream FA4 cute (FA4_CUTE_REF) over the
 # wheel/source one so the image runs FA4, not the FA2 fallback. This pulls the FA4
 # runtime stack (cutlass-dsl, quack-kernels, apache-tvm-ffi, torch-c-dlpack-ext) --
 # the same deps the [dreamverse] extra already installs in CI; the installed torch
@@ -170,12 +170,14 @@ RUN --mount=type=cache,target=/opt/uv/cache \
 # rmtree clears the wheel's stale cute files first to avoid an install conflict.
 # Then verify both survive so a broken overlay fails the build instead of shipping
 # an FA2-less image. x86 only: the FA4 stack (quack-kernels etc.) is unvalidated on
-# arm64 / GB10 (sm_121), so there we skip the overlay and FA4 falls back to FA2.
+# arm64 / GB10 (sm_121), so there we skip the overlay; FA4 is opt-in
+# (FASTVIDEO_FA4=1) and errors if set without the overlay, so leave it unset on
+# arm64 and the image runs FA3/FA2 as usual.
 RUN --mount=type=cache,target=/opt/uv/cache \
     source $HOME/.local/bin/env && \
     source /opt/venv/bin/activate && \
     if [ "${TARGETARCH}" = "arm64" ]; then \
-        echo "Skipping FA4 cute overlay on arm64 (FA4 stack unvalidated there; FA4 falls back to FA2)"; \
+        echo "Skipping FA4 cute overlay on arm64 (FA4 stack unvalidated there; do not set FASTVIDEO_FA4)"; \
     else \
         python -c "import glob, shutil; [shutil.rmtree(d, ignore_errors=True) for d in glob.glob('/opt/venv/lib/python*/site-packages/flash_attn/cute')]" && \
         uv pip install "flash-attn-4 @ git+https://github.com/Dao-AILab/flash-attention.git@${FA4_CUTE_REF}#subdirectory=flash_attn/cute" && \

docs/contributing/ci_architecture.md

@@ -99,10 +99,18 @@ status.
 Full Suite is also path-filtered. It validates broader behavior before Mergify
 can merge a PR.
 
+A `ready`-labeled PR does not hit Buildkite immediately:
+`ci-trigger-full-suite.yml` first runs `.github/scripts/gate_full_suite.sh`,
+which waits for the cheap Tier-1 checks (pre-commit, docs build) on the PR
+head. A red cheap check blocks the suite (fail closed; the next push re-arms
+it), while a GitHub outage or a >25 min wait lets it run anyway (fail open).
+`/test full` bypasses the gate.
+
 | Buildkite label | `TEST_TYPE` | Main watched paths |
 |---|---|---|
 | SSIM Tests | `ssim` | `fastvideo/**/*.py`, `pyproject.toml`, `docker/Dockerfile` |
 | LoRA Inference Tests | `inference_lora` | LoRA tests, loader, transformer tests, pipelines, LoRA layers |
+| LoRA Extraction Tests | `lora_extraction` | LoRA extraction scripts/tests, loader, training utilities, LoRA layers |
 | Training Tests | `training` | `fastvideo/**`, `pyproject.toml`, `docker/Dockerfile` |
 | Distillation DMD Tests | `distillation_dmd` | `fastvideo/training/*distillation_pipeline.py` |
 | Self-Forcing Tests | `self_forcing` | self-forcing distillation pipeline and tests |
@@ -144,6 +152,7 @@ Valid direct test names:
 | `/test training` | `training` |
 | `/test lora-inference` | `inference_lora` |
 | `/test lora-training` | `training_lora` |
+| `/test lora-extraction` | `lora_extraction` |
 | `/test distillation` | `distillation_dmd` |
 | `/test self-forcing` | `self_forcing` |
 | `/test vsa` | `training_vsa` |

docs/contributing/performance_benchmarks.md

@@ -72,14 +72,18 @@ fastvideo/tests/performance/
     │           writes Markdown summary + (optionally) uploads new records
     ├── dashboard.py
     │       └── builds time-series Plotly HTML from HF history
-    └── hf_store.py               # shared HF I/O + DataFrame helpers
+
+fastvideo/performance/
+    ├── hf_store.py               # shared HF I/O + DataFrame helpers
+    └── metric_policy.py          # shared rolling-baseline threshold policy
 ```
 
 The HF dataset (`FastVideo/performance-tracking` by default) holds one
-normalized JSON per `(model_id, gpu_type, run)` tuple. The rolling baseline is
-the median of the last 5 successful, baseline-eligible records for that
-model+GPU. PR and local records are visible in the dashboard but are not
-baseline eligible.
+normalized JSON per run. For v2 records, the rolling baseline is the median of
+the last 5 successful, baseline-eligible records in the same comparison cohort:
+`model_id`, `gpu_type`, `workload_id`, `variant_id`, `benchmark_version`,
+`recipe_fingerprint`, `hardware_profile_id`, and `software_profile_id`. PR and
+local records are visible in the dashboard but are not baseline eligible.
 
 ## Planned Coverage
 
@@ -92,25 +96,28 @@ and recipe changes instead of treating all records for a model as equivalent.
 
 ## Metrics
 
-Each benchmark records six metrics:
+Each benchmark records six metrics. The rolling-baseline comparator also has a
+per-metric policy with direction, percent threshold, absolute threshold, and a
+`gated` flag.
 
-| Metric | Raw key | Normalized key | Direction |
-|---|---|---|---|
-| End-to-end generation latency | `avg_generation_time_s` | `latency` | Lower is better |
-| Video throughput | `throughput_fps` | `throughput` | Higher is better |
-| Peak GPU memory | `max_peak_memory_mb` | `memory` | Lower is better |
-| Text encoder time | `text_encoder_time_s` | `text_encoder_time_s` | Lower is better |
-| DiT denoising time | `dit_time_s` | `dit_time_s` | Lower is better |
-| VAE decode time | `vae_decode_time_s` | `vae_decode_time_s` | Lower is better |
+| Metric | Raw key | Normalized key | Direction | Default rolling policy |
+|---|---|---|---|---|
+| End-to-end generation latency | `avg_generation_time_s` | `latency` | Lower is better | 8% and 0.5 s |
+| Video throughput | `throughput_fps` | `throughput` | Higher is better | 8% and 0.05 FPS |
+| Peak GPU memory | `max_peak_memory_mb` | `memory` | Lower is better | 5% and 256 MB |
+| Text encoder time | `text_encoder_time_s` | `text_encoder_time_s` | Lower is better | 5% and 0.25 s |
+| DiT denoising time | `dit_time_s` | `dit_time_s` | Lower is better | 5% and 0.25 s |
+| VAE decode time | `vae_decode_time_s` | `vae_decode_time_s` | Lower is better | 5% and 0.25 s |
 
 `test_inference_performance.py` temporarily sets `FASTVIDEO_STAGE_LOGGING=1`
 while it runs so pipeline stage execution times are available in
 `generate_video(...).logging_info`. Stage logs use pipeline-unique keys such as
 `prompt_encoding_stage` so duplicate stage classes do not collide. For
-`PipelineStage` entries, the extractor maps the `stage_class` field:
-`TextEncodingStage` maps to `text_encoder_time_s`, `DenoisingStage` and
-`DmdDenoisingStage` map to `dit_time_s`, and `DecodingStage` maps to
-`vae_decode_time_s`, with a fallback for older logs that used the class name as
+`PipelineStage` entries, shared component stage bases emit a stable
+`component_metric`: text encoding stages map to `text_encoder_time_s`,
+denoising stages and subclasses map to `dit_time_s`, and decoding stages map to
+`vae_decode_time_s`. The extractor falls back to known `stage_class` names for
+older logs that do not include `component_metric` or that used the class name as
 the stage key. Generator-side timings such as `PostDecodeFrameProcessStage`,
 `VideoSaveStage`, and `AudioMuxStage` are intentionally ignored. If a pipeline
 does not report one of the mapped stages, that component metric is stored as
@@ -152,13 +159,29 @@ unrealistic memory growth, and optionally large component-specific slowdowns
 even when the rolling baseline is empty. They are hand-set with generous
 headroom and almost never need touching.
 
-### Rolling baseline (per `(model_id, gpu_type)`)
+### Rolling baseline (per comparison cohort)
 
 `compare_baseline.py` loads the last 5 successful, baseline-eligible records
-for the same `(model_id, gpu_type)` from the HF dataset, computes the median
-for each available metric, and fails if the current run regresses by more than
-`PERF_MAX_REGRESSION` (default 5%). For latency, memory, and component times,
-higher values are regressions. For throughput, lower values are regressions.
+for the same comparison cohort from the HF dataset, computes the median for
+each available metric, and evaluates the current run with the metric's
+rolling regression policy. For v2 records, that cohort is `model_id`,
+`gpu_type`, `workload_id`, `variant_id`, `benchmark_version`,
+`recipe_fingerprint`, `hardware_profile_id`, and `software_profile_id`. For
+latency, memory, and component times, higher values are regressions. For
+throughput, lower values are regressions.
+
+A metric exceeds its rolling threshold when both of these are true:
+
+```text
+percent_delta > threshold_percent
+absolute_delta > threshold_absolute
+```
+
+Gated metrics fail CI when that threshold crossing happens. Set `gated: false`
+for metrics that should remain visible in reports and the dashboard without
+failing CI. Dashboard/API payloads expose `threshold_exceeded` separately from
+`regressed`, where `regressed` means a gated CI failure. Missing or `null`
+metrics are skipped.
 
 This is the **drift detector** — it catches sub-threshold regressions that
 slowly add up. Only scheduled-main successful records are baseline eligible.
@@ -172,6 +195,51 @@ agent skill to advance the rolling median.
 
 ## Schemas
 
+### Benchmark config (`.buildkite/performance-benchmarks/tests/*.json`)
+
+Benchmark configs without `config_schema_version` are treated as legacy v1
+configs and remain loadable. New or migrated configs should use
+`config_schema_version: 2` and include explicit comparable identity fields:
+
+```jsonc
+{
+  "benchmark_id": "wan-t2v-1.3b-2gpu",
+  "config_schema_version": 2,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2
+}
+```
+
+`benchmark_id` is still required in this phase because raw artifact names,
+generated-video directories, normalized record paths, and the current rolling
+baseline comparator still depend on it. The v2 identity fields are config
+metadata that make the measured workload explicit:
+
+| Field | Purpose |
+|---|---|
+| `workload_id` | Stable benchmark family, such as `wan-t2v`. |
+| `variant_id` | Intentional recipe family, including model size and parallelism config, such as `1.3b-sp2`. |
+| `benchmark_version` | Version of the measurement protocol and comparison policy. |
+
+If a config declares `config_schema_version: 2`, loading fails clearly when any
+required v2 identity field is missing. If v2 identity or metadata fields are
+added without `config_schema_version: 2`, loading also fails so partial
+migrations do not silently run as v1 configs. Optional v2 metadata fields
+reserved for follow-up work, such as `metric_threshold_policy` and
+`quality_metadata`, must be JSON objects when present. (`recipe` is emitted
+by the harness and is not config-declarable.)
+
+Recipe fingerprinting, hardware/software profile IDs, exact-identity
+comparison, and dashboard cohort grouping land with this change: v2 records
+compare only within their identity cohort, and a record that opens a NEW
+cohort is marked `baseline_status: "initialized_new_cohort"` (regression
+gating starts once that cohort accumulates history). Legacy v1 configs still
+run and are normalized for reporting, but their records skip rolling-baseline
+comparison entirely (`baseline_status: "skipped_missing_identity"`, never
+baseline eligible); only static thresholds gate them. Metric-specific
+threshold policies and promoted baselines remain separate follow-ups.
+
 ### Raw record (`results/perf_*.json`)
 
 Written by `test_inference_performance.py`. One file per benchmark run.
@@ -179,6 +247,10 @@ Written by `test_inference_performance.py`. One file per benchmark run.
 ```jsonc
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
+  "result_schema_version": 2,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "model_short_name": "Wan2.1-T2V-1.3B-Diffusers",
   "device": "NVIDIA L40S",
   "num_gpus": 2,
@@ -196,12 +268,62 @@ Written by `test_inference_performance.py`. One file per benchmark run.
     "max_dit_time_s": 10.0,
     "max_vae_decode_time_s": 10.0
   },
+  "regression_thresholds": {
+    "latency": {
+      "threshold_percent": 0.10,
+      "threshold_absolute": 1.0,
+      "gated": true
+    }
+  },
   "commit": "<full sha>",
+  "run_source": "pr",
+  "branch": "feature/perf-change",
   "pr_number": "1234",
+  "test_scope": "direct",
+  "build_url": "https://buildkite.example/build",
+  "build_id": "<buildkite-build-id>",
+  "job_id": "<buildkite-job-id>",
   "timestamp": "2026-05-08T22:00:00+00:00",
+  "quality_metadata": { "quality_status": "canonical" },
   "text_encoder_time_s": 2.141,
   "dit_time_s": 8.437,
-  "vae_decode_time_s": 3.208
+  "vae_decode_time_s": 3.208,
+  "recipe": {
+    "recipe_schema_version": 1,
+    "benchmark": {
+      "benchmark_id": "wan-t2v-1.3b-2gpu",
+      "workload_id": "wan-t2v",
+      "variant_id": "1.3b-sp2",
+      "benchmark_version": 2
+    },
+    "model": { "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" },
+    "init_kwargs": { "num_gpus": 2, "sp_size": 2, "tp_size": 1 },
+    "generation_kwargs": { "height": 480, "width": 832, "num_frames": 45 },
+    "inputs": { "prompt_count": 1, "prompt_sha256": ["<measured-prompt-sha256>"] },
+    "attention": { "requested_backend": "FLASH_ATTN", "resolved_backend": "FLASH_ATTN" }
+  },
+  "recipe_fingerprint": "<sha256>",
+  "hardware_profile": {
+    "device_type": "cuda",
+    "gpu_count": 2,
+    "gpus": [{ "name": "NVIDIA L40S", "memory_gb": 48, "compute_capability": "8.9" }],
+    "interconnect": "none_or_partial"
+  },
+  "hardware_profile_id": "hw-<sha256-prefix>",
+  "software_profile": {
+    "python": "3.12",
+    "pytorch": "2.12",
+    "cuda": "13.0",
+    "packages": {
+      "fastvideo_kernel": "0.3.2",
+      "flashinfer": "0.2.11",
+      "nvidia_cutlass_dsl": "4.5.0",
+      "triton": "3.4.1"
+    }
+  },
+  "software_profile_id": "sw-<sha256-prefix>",
+  "environment_metadata": { "env": { "IMAGE_VERSION": "py3.12-cuda13.0.0" } },
+  "environment_fingerprint": "env-<sha256-prefix>"
 }
 ```
 
@@ -213,6 +335,10 @@ result, used as the rolling-baseline source of truth.
 ```jsonc
 {
   "model_id": "wan-t2v-1.3b-2gpu",
+  "result_schema_version": 2,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "timestamp": "2026-05-08T22:00:00+00:00",
   "commit_sha": "<full sha>",
   "gpu_type": "NVIDIA L40S",
@@ -222,34 +348,69 @@ result, used as the rolling-baseline source of truth.
   "text_encoder_time_s": 2.141,
   "dit_time_s": 8.437,
   "vae_decode_time_s": 3.208,
+  "regression_thresholds": {
+    "latency": {
+      "threshold_percent": 0.08,
+      "threshold_absolute": 0.5,
+      "gated": true
+    }
+  },
+  "recipe_fingerprint": "<sha256>",
+  "hardware_profile_id": "hw-<sha256-prefix>",
+  "software_profile_id": "sw-<sha256-prefix>",
+  "environment_fingerprint": "env-<sha256-prefix>",
+  "run_source": "pr",
+  "branch": "feature/perf-change",
+  "pr_number": "1234",
+  "test_scope": "direct",
+  "build_url": "https://buildkite.example/build",
+  "build_id": "<buildkite-build-id>",
+  "job_id": "<buildkite-job-id>",
+  "quality_metadata": { "quality_status": "canonical" },
   "success": true
 }
 ```
 
 ### Compatibility with legacy records
 
-Older records in the HF dataset may not have component timing fields. The
-comparator ignores missing or `null` metrics when computing a median, and the
-dashboard lists skipped plots for metric series that have no non-null values.
-Records missing both `run_source` and `baseline_eligible` are treated as legacy
-successful main/full-suite uploads and remain eligible for rolling baselines.
+Older records in the HF dataset may not have `result_schema_version`,
+component timing fields, or v2 identity/profile fields. Records without
+`result_schema_version` are treated as v1. The comparator ignores missing or
+`null` metrics when computing a median, and the dashboard lists skipped plots
+for metric series that have no non-null values. Records missing both
+`run_source` and `baseline_eligible` are treated as legacy successful
+main/full-suite uploads and remain eligible for rolling baselines.
+Current `perf_*.json` artifacts that lack the v2 comparison identity are
+normalized for reporting but skip rolling-baseline comparison and are not marked
+baseline eligible.
+
+New records compare only against the same `model_id`, `gpu_type`,
+`workload_id`, `variant_id`, `benchmark_version`, `recipe_fingerprint`,
+`hardware_profile_id`, and `software_profile_id` cohort.
+`environment_metadata` and `environment_fingerprint` are audit data and are not
+part of the comparison key.
+The recipe prompt digests describe the prompts actually measured by the
+benchmark run; extra configured prompts are ignored unless the benchmark runner
+executes them.
+Software profile package cohorts keep exact versions for relevant
+attention/kernel packages, including FastVideo kernels, FlashAttention,
+FlashInfer, Cutlass DSL, SageAttention, Triton, and xFormers when installed.
 
 ## Environment variable reference
 
 | Variable | Default | Used by | Purpose |
 |---|---|---|---|
-| `PERF_MAX_REGRESSION` | `0.05` | `compare_baseline.py` | Per-metric regression fraction that fails the build. |
 | `PERFORMANCE_TRACKING_ROOT` | `/tmp/perf-tracking` | `compare_baseline.py`, `dashboard.py` | Local directory the HF dataset is synced to. |
 | `PERF_REPORTS_DIR` | `/root/data/perf_reports` | `compare_baseline.py`, `dashboard.py` | Where the Markdown summary and Plotly HTML get written for Buildkite to pick up. |
-| `HF_REPO_ID` | `FastVideo/performance-tracking` | `hf_store.py` | HF dataset repo holding rolling-baseline records. |
-| `HF_API_KEY`, `HUGGINGFACE_HUB_TOKEN`, `HF_TOKEN` | unset | `hf_store.py` | Required for upload or private dataset reads. |
-| `PERF_RUN_SOURCE` | inferred | `compare_baseline.py` | Source metadata for uploaded records: `pr`, `local`, `scheduled_main`, or `unknown`. |
+| `HF_REPO_ID` | `FastVideo/performance-tracking` | `fastvideo/performance/hf_store.py` | HF dataset repo holding rolling-baseline records. |
+| `HF_API_KEY`, `HUGGINGFACE_HUB_TOKEN`, `HF_TOKEN` | unset | `fastvideo/performance/hf_store.py` | Required for upload or private dataset reads. |
+| `PERF_RUN_SOURCE` | inferred | `compare_baseline.py`, `test_inference_performance.py` | Source metadata for uploaded records: `pr`, `local`, `scheduled_main`, or `unknown`. |
 | `PERF_UPLOAD_POLICY` | `never` | `compare_baseline.py` | Upload policy: `never`, `pass`, or `always`. |
 | `PERF_PYTEST_RC` | unset | `compare_baseline.py` | Static-threshold pytest exit code, used so scheduled-main failures can be uploaded with `success=false`. |
 | `TEST_SCOPE` | unset | `compare_baseline.py` | CI context used to infer scheduled-main runs together with `BUILDKITE_BRANCH=main`. |
 | `BUILDKITE_BRANCH`, `BUILDKITE_COMMIT`, `BUILDKITE_PULL_REQUEST` | unset | `compare_baseline.py`, `test_inference_performance.py` | CI metadata stamped into records. |
 | `DASHBOARD_DAYS` | `30` | `dashboard.py` | Lookback window for the Plotly trend pages. |
-| `PERFORMANCE_TRACKING_SYNC_REUSE_TTL_SECONDS` | `3600` | `hf_store.py` | Freshness window for reusing an existing HF sync when requested by dashboard consumers. |
+| `PERFORMANCE_TRACKING_SYNC_REUSE_TTL_SECONDS` | `3600` | `fastvideo/performance/hf_store.py` | Freshness window for reusing an existing HF sync when requested by dashboard consumers. |
 | `FASTVIDEO_STAGE_LOGGING` | set by the pytest test | `test_inference_performance.py` | Enables pipeline stage timing capture for component metrics during benchmark runs. |
 
 ## CI integration
@@ -260,17 +421,22 @@ point is `fastvideo/tests/modal/pr_test.py:run_performance_tests` and the
 Buildkite artifact upload is in
 `.buildkite/scripts/pr_test.sh:upload_performance_artifacts`.
 
-Each performance build runs pytest first. If that fixed-threshold phase fails,
-`compare_baseline.py` is skipped, so Markdown summaries and normalized JSON
-artifacts are not emitted. The dashboard still runs best-effort for
-observability. When pytest passes, the rolling-baseline phase emits:
+Each performance build runs pytest first. PR and direct runs only continue to
+`compare_baseline.py` when that fixed-threshold phase passes; if pytest fails,
+Markdown summaries and normalized JSON artifacts are not emitted. Scheduled
+main runs set `PERF_UPLOAD_POLICY=always`, so they still run
+`compare_baseline.py` (with `PERF_PYTEST_RC` set) after a fixed-threshold
+failure. Those failed scheduled main runs emit summaries and normalized
+records, upload records with `success=false`, and are excluded from future
+rolling baselines. The dashboard still runs best-effort for observability.
+When the rolling-baseline phase runs, it emits:
 
 * **Markdown summary** — appended to `$GITHUB_STEP_SUMMARY` when that variable
   is set, and written as `perf_<sha>_<ts>.md` for Buildkite upload. Contains a
   per-benchmark row with current vs. baseline values for latency, throughput,
   memory, text encoder time, DiT time, and VAE decode time.
 * **Plotly dashboard** — `dashboard_<sha>_<ts>.html` showing time-series for
-  each metric grouped by `(model_id, gpu_type)`.
+  each metric grouped by comparison cohort.
 * **Normalized records** — `normalized_perf_*.json`, one per benchmark.
   Useful as input to the
   [`reseed-performance-baseline`](https://github.com/hao-ai-lab/FastVideo/blob/main/.agents/skills/reseed-performance-baseline/SKILL.md)
@@ -279,11 +445,16 @@ observability. When pytest passes, the rolling-baseline phase emits:
 ## Adding a new benchmark
 
 1. Drop a new JSON config into
-   `.buildkite/performance-benchmarks/tests/<name>.json`. Required keys:
+   `.buildkite/performance-benchmarks/tests/<name>.json`. New configs should
+   use v2 identity fields:
 
    ```json
    {
      "benchmark_id": "<unique-id>",
+     "config_schema_version": 2,
+     "workload_id": "<stable-workload-id>",
+     "variant_id": "<variant, e.g. 1.3b-sp2>",
+     "benchmark_version": 1,
      "model": { "model_path": "...", "model_short_name": "..." },
      "init_kwargs": { "num_gpus": 1, ... },
      "generation_kwargs": { "num_frames": 45, ... },
@@ -299,9 +470,19 @@ observability. When pytest passes, the rolling-baseline phase emits:
          "max_vae_decode_time_s": 10.0
        },
        "default": { "max_generation_time_s": 120.0, "max_peak_memory_mb": 30000.0 }
+     },
+     "regression_thresholds": {
+       "latency": { "threshold_percent": 0.10, "threshold_absolute": 1.0, "gated": true }
      }
-   }
-   ```
+  }
+
+  ```
+
+   Legacy v1 configs without `config_schema_version` still load, but should not
+   gain v2 identity or metadata fields until they are migrated to
+   `config_schema_version: 2`. For v2 configs, `workload_id`, `variant_id`,
+   and `benchmark_version` are part of the comparison key; benchmark runs
+   fail if any of these identity fields are missing.
 
 2. The pytest test auto-discovers all configs — no test code needed. CI
    picks it up on the next `/test performance` run.
@@ -320,10 +501,17 @@ observability. When pytest passes, the rolling-baseline phase emits:
    a useful fixed gate. The rolling baseline will still track component times
    when static component thresholds are omitted.
 
+6. Omit `regression_thresholds` to use the default rolling-baseline policy, or
+   include only benchmark-specific deviations. Tune these independently from
+   the fixed thresholds when a metric is noisy or should be informational. The
+   fixed `thresholds` block is an absolute pytest ceiling. The
+   `regression_thresholds` block controls rolling-baseline comparisons against
+   recent scheduled-main records.
+
 ## Troubleshooting
 
-**"No baseline for ... Initializing"** — first run for this `(model_id,
-gpu_type)`. Run will pass and (if persisting) seed the first record.
+**"No baseline for ... Initializing"** — first run for this comparison cohort.
+Run will pass and (if persisting) seed the first record.
 
 **Persistent failure right after a torch / kernel / image upgrade** —
 genuine regression *or* baseline drift. Compare the failing normalized record
@@ -336,5 +524,5 @@ pipelines that did not report a mapped component stage.
 
 **Component timing is `null`** — the generated result did not include a mapped
 stage in `logging_info.stages`. Check that the pipeline emits stage logging
-and that the stage name is listed in `STAGE_METRIC_MAP` in
-`test_inference_performance.py`.
+and that the stage emits `component_metric` or is covered by the legacy
+`STAGE_METRIC_MAP` fallback in `test_inference_performance.py`.

docs/contributing/testing.md

@@ -180,6 +180,30 @@ python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
   --device-folder L40S_reference_videos
 ```
 
+### SSIM Bootstrap Mode
+
+Normal SSIM runs are strict: if a reference video or latent is missing, the
+test fails. For new-model PRs, CI can run SSIM in bootstrap mode so missing
+references are uploaded as draft artifacts for review instead of immediately
+blocking on a missing canonical reference.
+
+Buildkite enables SSIM bootstrap mode when either condition is true:
+
+- the PR title or Buildkite message contains `[new-model]`;
+- `FASTVIDEO_SSIM_BOOTSTRAP_MODE=1` is set for the Buildkite job.
+
+Bootstrap mode passes `--ssim-bootstrap-mode` to pytest. When a generated
+artifact is available, the test uploads it under the `drafts/...` namespace in
+the SSIM reference repo and marks that case as expected-failed. After reviewing
+the draft, promote it into the canonical reference layout:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py promote-draft \
+  --quality-tier default \
+  --device-folder L40S_reference_videos \
+  --model-id <model_id>
+```
+
 ## CI Integration
 
 FastVideo CI tests are orchestrated by Buildkite and run on Modal GPU

docs/design/inference_schema_parity_inventory.yaml

@@ -191,6 +191,9 @@ surfaces:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig
           - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      color_correction_strength:
+        sources:
+          - fastvideo.configs.pipelines.dreamx_world.DreamXWorld5BARPipelineConfig
       default_camera_rotation:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig
@@ -455,6 +458,8 @@ surfaces:
       guidance_scale: request.sampling.guidance_scale
       guidance_scale_2: request.sampling.guidance_scale_2
       guidance_rescale: request.sampling.guidance_rescale
+      use_embedded_guidance: request.sampling.use_embedded_guidance
+      true_cfg_scale: request.sampling.true_cfg_scale
       boundary_ratio: request.sampling.boundary_ratio
       sigmas: request.sampling.sigmas
       enable_teacache: request.runtime.enable_teacache

docs/distillation/anyflow.md

@@ -0,0 +1,116 @@
+# 🌊 AnyFlow Any-Step Video Distillation
+
+**AnyFlow** ([paper](https://arxiv.org/abs/2605.13724), [project page](https://nvlabs.github.io/AnyFlow/), [official code](https://github.com/NVlabs/AnyFlow), [model weights](https://huggingface.co/collections/nvidia/anyflow)) is an any-step video diffusion framework built on flow maps. A single distilled checkpoint can be evaluated at NFE ∈ {1, 2, 4, 8, 16, 32} without retraining, and quality scales **monotonically** with steps — unlike consistency-based distillation, which often degrades as NFE grows.
+
+The student network ``u_θ(x_t, t, r)`` predicts the *average velocity* from time ``t`` back to time ``r``, so one Euler step is
+
+```
+x_r = x_t - ((t - r) / N) · u_θ(x_t, t, r)
+```
+
+for any ``t > r``.
+
+## 📊 Model Overview
+
+NVIDIA publishes four checkpoints under [`nvidia/anyflow`](https://huggingface.co/collections/nvidia/anyflow):
+
+- `nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers` — bidirectional T2V, Wan2.1 1.3B base
+- `nvidia/AnyFlow-Wan2.1-T2V-14B-Diffusers` — bidirectional T2V, Wan2.1 14B base
+- `nvidia/AnyFlow-FAR-Wan2.1-1.3B-Diffusers` — frame-autoregressive variant, 1.3B
+- `nvidia/AnyFlow-FAR-Wan2.1-14B-Diffusers` — frame-autoregressive variant, 14B
+
+FastVideo currently supports the bidirectional T2V variants for training; the FAR variants can be loaded for inference through the diffusers integration.
+
+## ⚙️ Inference
+
+For inference, load the published checkpoint directly through diffusers; FastVideo's training-side ``WanModel`` config maps the HF AnyFlow ``delta_embedder`` weights onto its internal layout via ``param_names_mapping`` so the same checkpoint can be used as the ``init_from`` for the on-policy YAML below.
+
+## 🧠 Algorithm
+
+Training runs in two stages. Both use the dual-timestep Wan backbone — enabled by ``pipeline.dit_config.r_embedder: true`` in the YAML, which allocates a sibling ``condition_embedder.delta_embedder`` and fuses its embedding with the standard timestep embedding via either an additive or a gated mixer.
+
+### Stage 1 — Pretrain (flow-map central-difference)
+
+Method: ``AnyFlowPretrainMethod`` (``fastvideo/train/methods/distribution_matching/anyflow_pretrain.py``)
+
+For each batch, sample ``(t, r) ∈ [0, 1]`` as ``(max, min)`` of two uniform draws, then:
+
+- a ``diffusion_ratio`` fraction (default 0.5) gets ``r = t`` — recovers plain flow matching;
+- a ``consistency_ratio`` fraction (default 0.25) gets ``r = 0`` — forces consistency to clean data;
+- the remainder is free.
+
+The student forward at ``(t, r)`` is trained against the central-difference target
+
+```
+target = (eps - x_0) - (t - r) · dF/dt
+```
+
+where ``dF/dt`` is estimated from the student's own forward at ``(t ± δ, r)`` with the sample also moved along the flow trajectory by ``v_pred · (δ / N)``. Per-timestep weighting uses ``beta08`` (``w(t) = t · sqrt(1 - t)``, renormalized). A stop-gradient scale-balance keeps the non-diffusion branches' loss magnitude aligned with the diffusion branch.
+
+### Stage 2 — On-policy DMD
+
+Method: ``AnyFlowMethod`` (``fastvideo/train/methods/distribution_matching/anyflow.py``)
+
+Inherits ``DMD2Method``. The student is rolled out for ``student_sample_steps`` Euler-flow steps from pure noise; one randomly-chosen step is gradient-enabled (broadcast from rank 0 so every worker agrees), the rest run under ``torch.no_grad``. With ``use_mean_velocity: true`` (default) the rollout uses ``r = t_next`` at each step, matching AnyFlow's ``WanAnyFlowPipeline.training_rollout``.
+
+The inherited ``_dmd_loss`` (VSD with fake-score critic) consumes the rollout output and the teacher's CFG prediction. The optional pinned ``t_list_override`` lets configs reproduce the paper's hand-tuned 4-step schedule ``[999, 937, 833, 624, 0]``.
+
+## 🚀 Training Scripts
+
+### Stage 1 — pretrain
+
+```bash
+bash examples/train/run.sh \
+    examples/train/configs/distribution_matching/wan/anyflow_pretrain_t2v.yaml
+```
+
+**Key configuration** (in ``examples/train/configs/distribution_matching/wan/anyflow_pretrain_t2v.yaml``):
+
+- Global batch size: 32 (8 GPUs × 4 per-GPU)
+- Learning rate: 5e-5
+- Flow shift: 5.0
+- ``diffusion_ratio`` / ``consistency_ratio``: 0.5 / 0.25
+- ``epsilon`` (finite-difference step): 5 (absolute train-timestep units)
+- ``weight_type``: ``beta08``
+- ``fuse_guidance_scale``: 3.0
+- Training steps: 6000
+
+### Stage 2 — on-policy
+
+```bash
+bash examples/train/run.sh \
+    examples/train/configs/distribution_matching/wan/anyflow_onpolicy_t2v.yaml \
+    --models.student.init_from outputs/wan2.1_anyflow_pretrain/checkpoint-final
+```
+
+(Or point ``models.student.init_from`` directly at ``nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers`` to bootstrap from the paper weights and skip Stage 1.)
+
+**Key configuration**:
+
+- Global batch size: 8 (8 GPUs × 1 per-GPU)
+- Learning rate: 2e-6
+- Flow shift: 5.0
+- ``student_sample_steps``: 4
+- ``t_list_override``: ``[999, 937, 833, 624, 0]``
+- ``use_mean_velocity``: ``true`` (i.e. ``r = t_next`` during rollout)
+- ``real_score_guidance_scale``: 3.0
+- ``generator_update_interval``: 5 (DMD2 alternation)
+- Training steps: 4000
+
+## 🔌 Loading published AnyFlow checkpoints
+
+The HF AnyFlow checkpoints expose ``condition_embedder.delta_embedder.*`` weights that FastVideo internally maps onto its ``condition_embedder.delta_embedder.mlp.*`` layout. This rename happens automatically through the regex in ``WanVideoArchConfig.param_names_mapping`` — no separate adapter is needed. The same regex is a no-op on plain Wan checkpoints (which don't contain any ``delta_embedder`` keys).
+
+Set the YAML's ``pipeline.dit_config.r_embedder: true`` to allocate the ``delta_embedder`` module on the FastVideo side; when initializing from a plain Wan checkpoint the delta weights are deep-copied from ``time_embedder`` (matching AnyFlow's ``setup_flowmap_model()`` behavior).
+
+## 🧭 Note on ``fuse_guidance_scale``
+
+Stage 1 optionally fuses classifier-free guidance into the training target so the resulting checkpoint can be sampled at ``guidance_scale=1.0`` (no extra forward pass at inference time). The transformation is
+
+```
+noise_pred ← (noise_pred - (1 - g) · noise_pred_uncond) / g
+```
+
+with ``g = fuse_guidance_scale``. The negative prompt embedding comes from ``WanModel``'s ``ensure_negative_conditioning()`` — i.e. the dataset's configured ``sampling_param.negative_prompt``. Setting ``fuse_guidance_scale: 1.0`` skips the extra unconditional forward entirely.
+
+The on-policy stage's ``real_score_guidance_scale`` (inherited from DMD2) follows the same parameterization conventions documented in [``dmd.md``](dmd.md#-note-on-real_score_guidance_scale).

docs/inference/optimizations.md

@@ -74,6 +74,23 @@ uv pip install ninja
 python setup.py install
 ```
 
+### Flash Attention 4 (opt-in)
+
+FastVideo never auto-selects FlashAttention-4 (`flash_attn.cute`) just because it
+is installed: its CuTeDSL kernels JIT-compile per shape family and can fail at
+runtime on some GPU/shape combinations. To use FA4, install the pinned
+`flash-attn-4` build (see the `flash-attn-4` source in `pyproject.toml`) and set:
+
+```bash
+export FASTVIDEO_FA4=1
+```
+
+On GPUs below sm90 a capability gate routes to FlashAttention-2 the calls FA4
+cannot serve there: grad-enabled (training) attention (FA4's backward requires
+sm90+) and GQA attention (FA4's `pack_gqa` fails to JIT-compile below sm90).
+On sm90+ both run on FA4. If FA4 is unusable while `FASTVIDEO_FA4=1` is set,
+FastVideo fails loudly instead of silently falling back.
+
 ### FP4 Flash Attention 4 (Blackwell only)
 
 **`FLASH_ATTN`** with **`--nvfp4_fa4`**

docs/inference/support_matrix.md

@@ -58,6 +58,8 @@ pipeline initialization and sampling.
 | FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
 | FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
 | Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ | ⭕ |
+| DreamX-World 5B Cam | `FastVideo/DreamX-World-5B-Cam-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| DreamX-World 5B AR | `FastVideo/DreamX-World-5B-Diffusers` | 704px1280p | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Lucy Edit Dev 5B*** | `decart-ai/Lucy-Edit-Dev` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
 | Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |

docs/training/trackers.md

@@ -0,0 +1,91 @@
+# Training Trackers
+
+FastVideo can send training metrics and validation media to Weights & Biases
+or SwanLab. Tracking runs only on global rank 0, and local tracker files are
+stored under `<output_dir>/tracker`.
+
+## Supported Trackers
+
+| Value | Backend | Installation |
+|-------|---------|--------------|
+| `wandb` | Weights & Biases | Included with FastVideo |
+| `swanlab` | SwanLab | Install the optional `swanlab` dependency |
+| `none` | Disable external tracking | No additional package |
+
+You can enable more than one backend, for example `trackers: [wandb, swanlab]`.
+Metrics and validation media are converted to the artifact type required by
+each backend.
+
+## Install SwanLab
+
+For a published FastVideo installation, install the SwanLab extra:
+
+```bash
+uv pip install "fastvideo[swanlab]"
+```
+
+For an editable source checkout, include the same extra during installation:
+
+```bash
+uv pip install -e ".[swanlab]"
+```
+
+If FastVideo is already installed, you can install the compatible SDK directly:
+
+```bash
+uv pip install "swanlab>=0.6.7"
+```
+
+Authenticate once before starting a training run:
+
+```bash
+swanlab login
+```
+
+See the [SwanLab login documentation](https://docs.swanlab.cn/en/api/cli-swanlab-login.html)
+for non-interactive and self-hosted setups.
+
+## Configure Tracking
+
+Select SwanLab in the YAML config used by the modular training framework:
+
+```yaml
+training:
+  checkpoint:
+    output_dir: outputs/my_run
+  tracker:
+    trackers: [swanlab]
+    project_name: my_project
+    run_name: my_run
+```
+
+To log to both supported services:
+
+```yaml
+training:
+  tracker:
+    trackers: [wandb, swanlab]
+    project_name: my_project
+    run_name: my_run
+```
+
+An empty or omitted `trackers` list selects W&B when `project_name` is set.
+Use an explicit `none` entry to disable external tracking:
+
+```yaml
+training:
+  tracker:
+    trackers: [none]
+```
+
+## Validation Videos
+
+SwanLab currently accepts GIF video artifacts. FastVideo converts validation
+MP4 files and in-memory video arrays to GIF automatically before logging them.
+For video files, FastVideo uses the sampling frame rate supplied by the caller,
+or the source file's frame rate when no value is supplied. In-memory arrays use
+the frame rate supplied by the caller. Both forms fall back to 16 FPS when no
+frame rate is available.
+
+For details about configuring validation callbacks, see
+[Training Infrastructure](train_infra.md#callbacks-pluggable-hooks).

docs/training/train_infra.md

@@ -161,6 +161,21 @@ training:
     decay_interval_steps: 0
 ```
 
+`training.data.data_path` can also mix multiple preprocessed datasets by using a mapping from dataset path to repeat count:
+
+```yaml
+training:
+  data:
+    data_path:
+      data/zeldam2-clean: 1
+      data/multi3d_games: 2
+```
+
+The repeat count duplicates that dataset's parquet file list before shuffling/sampling, so the example above trains with roughly twice as much `multi3d_games` exposure as `zeldam2-clean`. Paths are just suggested locations; use any local path that contains a FastVideo preprocessed parquet dataset.
+
+See [Training Trackers](trackers.md) to configure Weights & Biases or SwanLab,
+including SwanLab installation and authentication.
+
 ### `callbacks` — Pluggable hooks
 
 Callbacks run at specific points in the training loop (before/after optimizer
@@ -323,6 +338,40 @@ Self-Forcing inherits all DMD2 parameters, plus:
 | `enable_gradient_in_rollout` | `true` | Enable backprop through rollout |
 | `start_gradient_frame` | `0` | Frame index where gradients begin |
 
+### Streaming Long Tuning
+
+`StreamingLongTuningMethod` extends Self-Forcing for LongLive-style rollouts. It
+keeps a streaming state, generates overlapping chunks, and trains only the new
+frames while preserving context from earlier chunks.
+
+For the MatrixGame2/Zelda world-model example, self-forcing and long tuning are
+separate runs: first train or load the 1k-step self-forcing checkpoint using
+`examples/train/scenario/worldmodel/zelda/self_forcing_causal_i2v.yaml`,
+then run
+`examples/train/scenario/worldmodel/zelda/streaming_long_tuning_causal_i2v.yaml`
+from that checkpoint for the 3k-step streaming long-tuning stage.
+
+```yaml
+method:
+  _target_: fastvideo.train.methods.distribution_matching.streaming_long_tuning.StreamingLongTuningMethod
+  streaming_chunk_size: 9
+  streaming_max_length: 39
+  streaming_fixed_overlap_latents: 3
+  streaming_reencode_overlap_anchor: true
+  streaming_anchor_inject_k: 1
+  streaming_require_full_blocks: true
+  multi_phased_distill_schedule:
+    - stage: streaming_long
+      start_step: 0
+      end_step: 3000
+      num_latent_t: 39
+      streaming_training: true
+```
+
+See
+`examples/train/scenario/worldmodel/zelda/streaming_long_tuning_causal_i2v.yaml`
+for a complete MatrixGame2/Zelda configuration.
+
 ---
 
 ## Callbacks

examples/inference/basic/basic_dreamx_world.py

@@ -0,0 +1,64 @@
+import os
+
+from fastvideo import VideoGenerator
+
+
+OUTPUT_PATH = os.getenv("DREAMX_WORLD_OUTPUT_PATH", "video_samples_dreamx_world")
+
+
+def _env_int(name: str, default: int) -> int:
+    return int(os.getenv(name, str(default)))
+
+
+def _env_float(name: str, default: float) -> float:
+    return float(os.getenv(name, str(default)))
+
+
+def main():
+    model_name = os.getenv("DREAMX_WORLD_MODEL_DIR", "FastVideo/DreamX-World-5B-Cam-Diffusers")
+    generator = VideoGenerator.from_pretrained(
+        model_name,
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        override_pipeline_cls_name="DreamXWorldPipeline",
+    )
+
+    prompt = os.getenv(
+        "DREAMX_WORLD_PROMPT",
+        "A cinematic first-person drive through a futuristic coastal city at "
+        "sunrise, reflective glass towers, clean streets, soft volumetric light.",
+    )
+    image_path = os.getenv(
+        "DREAMX_WORLD_IMAGE_PATH",
+        "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG",
+    )
+
+    kwargs = {
+        "output_path": OUTPUT_PATH,
+        "save_video": os.getenv("DREAMX_WORLD_SAVE_VIDEO", "1") != "0",
+        "height": _env_int("DREAMX_WORLD_HEIGHT", 480),
+        "width": _env_int("DREAMX_WORLD_WIDTH", 832),
+        "num_frames": _env_int("DREAMX_WORLD_NUM_FRAMES", 161),
+        "num_inference_steps": _env_int("DREAMX_WORLD_STEPS", 30),
+        "guidance_scale": _env_float("DREAMX_WORLD_GUIDANCE", 5.0),
+        "action_list": os.getenv("DREAMX_WORLD_ACTIONS", "w,d,w").split(","),
+        "action_speed_list": [
+            float(value)
+            for value in os.getenv("DREAMX_WORLD_ACTION_SPEEDS", "4.0,2.0,4.0").split(",")
+        ],
+    }
+    if image_path:
+        kwargs["image_path"] = image_path
+
+    try:
+        generator.generate_video(prompt, **kwargs)
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_flux_dev.py

@@ -0,0 +1,140 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import os
+import re
+
+DEFAULT_PROMPTS = [
+    "a photo of a cat",
+    (
+        "a cinematic photo of a red panda wearing a tiny backpack, standing on a "
+        "rainy neon-lit street at night, shallow depth of field, sharp focus, "
+        "35mm, bokeh"
+    ),
+]
+
+
+def _safe_filename(text: str, max_len: int = 100) -> str:
+    """Make a stable, filesystem-friendly filename base."""
+    s = text[:max_len].strip()
+    s = s.replace(os.sep, "_")
+    if os.altsep:
+        s = s.replace(os.altsep, "_")
+    s = re.sub(r"\s+", " ", s)
+    s = re.sub(r"[^A-Za-z0-9 .,_-]", "_", s)
+    s = s.strip(" .")
+    return s or "prompt"
+
+
+def _remove_existing_outputs(out_dir: str, filename_base: str) -> None:
+    """Delete prior outputs so reruns do not get _1, _2 suffixes."""
+    if not os.path.isdir(out_dir):
+        return
+
+    pattern = re.compile(rf"^{re.escape(filename_base)}(_\d+)?\.(mp4|png)$")
+    for fn in os.listdir(out_dir):
+        if pattern.match(fn):
+            with contextlib.suppress(FileNotFoundError):
+                os.remove(os.path.join(out_dir, fn))
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(
+        description="Run FLUX.1-dev text-to-image with FastVideo VideoGenerator.",
+    )
+    p.add_argument(
+        "--model-path",
+        default="official_weights/FLUX.1-dev",
+        help="Local Diffusers checkpoint dir or HF repo id.",
+    )
+    p.add_argument(
+        "--out-dir",
+        "--outdir",
+        default="outputs/flux_dev/samples",
+        help="Directory for saved PNG outputs.",
+    )
+    p.add_argument(
+        "--prompt",
+        action="append",
+        default=None,
+        help="Prompt. Repeat for multiple images.",
+    )
+    p.add_argument(
+        "--backend",
+        default=None,
+        help="Set FASTVIDEO_ATTENTION_BACKEND (e.g. TORCH_SDPA).",
+    )
+    p.add_argument("--seed", type=int, default=42, help="Base seed; each prompt uses seed + index.")
+    p.add_argument("--height", type=int, default=1024, help="Output height.")
+    p.add_argument("--width", type=int, default=1024, help="Output width.")
+    p.add_argument("--steps", type=int, default=28, help="Number of inference steps.")
+    p.add_argument("--guidance", type=float, default=3.5, help="Guidance scale.")
+    p.add_argument("--num-gpus", type=int, default=1, help="GPU count.")
+    return p.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    prompts: list[str] = args.prompt if args.prompt else DEFAULT_PROMPTS
+
+    if args.backend:
+        os.environ["FASTVIDEO_ATTENTION_BACKEND"] = args.backend
+
+    from fastvideo import VideoGenerator
+
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    init_kwargs = {
+        "num_gpus": args.num_gpus,
+        "workload_type": "t2i",
+        "sp_size": 1,
+        "tp_size": 1,
+        "dit_cpu_offload": False,
+        "dit_layerwise_offload": False,
+        "text_encoder_cpu_offload": False,
+        "vae_cpu_offload": False,
+        "image_encoder_cpu_offload": False,
+        "pin_cpu_memory": False,
+        "use_fsdp_inference": False,
+    }
+
+    generator = VideoGenerator.from_pretrained(
+        model_path=args.model_path,
+        **init_kwargs,
+    )
+    try:
+        for i, prompt in enumerate(prompts):
+            seed = args.seed + i
+            filename_base = (
+                f"flux_dev_{i:02d}_seed{seed}_{_safe_filename(prompt, max_len=80)}"
+            )
+            _remove_existing_outputs(args.out_dir, filename_base)
+            output_path = os.path.join(args.out_dir, f"{filename_base}.png")
+            print(f"[flux] prompt_idx={i} seed={seed} output_path={output_path}")
+
+            generation_kwargs = {
+                "output_path": output_path,
+                "height": args.height,
+                "width": args.width,
+                "num_frames": 1,
+                "fps": 1,
+                "num_inference_steps": args.steps,
+                "guidance_scale": args.guidance,
+                "use_embedded_guidance": True,
+                "true_cfg_scale": 1.0,
+                "seed": seed,
+                "save_video": True,
+            }
+
+            generator.generate_video(prompt, **generation_kwargs)
+
+        print(f"[flux] done. outputs written to: {args.out_dir}")
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_glm_image.py

@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Run GLM-Image text-to-image generation through FastVideo.
+
+User story:
+    "I have the HF `zai-org/GLM-Image` checkpoint and want a minimal
+    text-to-image generation command, saved as a PNG."
+"""
+import argparse
+from pathlib import Path
+
+from PIL import Image
+
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    OutputConfig,
+    ParallelismConfig,
+    PipelineSelection,
+    SamplingConfig,
+)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Run GLM-Image text-to-image generation.")
+    parser.add_argument(
+        "--model-path",
+        default="zai-org/GLM-Image",
+        help="HF id or local diffusers-format GLM-Image weights directory.",
+    )
+    parser.add_argument(
+        "--output",
+        default="image_output/landscape.png",
+        help="Output PNG path.",
+    )
+    parser.add_argument(
+        "--prompt",
+        default=("A beautiful landscape photography with rolling hills, "
+                 "a winding river, and a vibrant sunset in the background. "
+                 "Warm golden light, photorealistic style."),
+        help="Text prompt.",
+    )
+    parser.add_argument("--height", type=int, default=1024)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--steps", type=int, default=50)
+    parser.add_argument("--guidance-scale", type=float, default=1.5)
+    parser.add_argument("--seed", type=int, default=1024)
+    parser.add_argument("--num-gpus", type=int, default=1)
+    parser.add_argument("--tp-size", type=int, default=None)
+    parser.add_argument("--sp-size", type=int, default=None)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+
+    output = Path(args.output)
+    output.parent.mkdir(parents=True, exist_ok=True)
+    tp_size = args.tp_size if args.tp_size is not None else (args.num_gpus if args.num_gpus > 1 else 1)
+    sp_size = args.sp_size if args.sp_size is not None else (1 if args.num_gpus > 1 else args.num_gpus)
+
+    # GLM-Image needs trust_remote_code for its AR encoder; offload and the
+    # pipeline class come from the model's registered defaults — don't override.
+    generator_config = GeneratorConfig(
+        model_path=args.model_path,
+        trust_remote_code=True,
+        engine=EngineConfig(
+            num_gpus=args.num_gpus,
+            parallelism=ParallelismConfig(tp_size=tp_size, sp_size=sp_size),
+        ),
+        pipeline=PipelineSelection(workload_type="t2i"),
+    )
+
+    generator = VideoGenerator.from_config(generator_config)
+    try:
+        request = GenerationRequest(
+            prompt=args.prompt,
+            sampling=SamplingConfig(
+                height=args.height,
+                width=args.width,
+                num_frames=1,
+                fps=1,
+                num_inference_steps=args.steps,
+                guidance_scale=args.guidance_scale,
+                seed=args.seed,
+            ),
+            output=OutputConfig(
+                output_path=str(output.parent),
+                save_video=False,
+                return_frames=True,
+            ),
+        )
+        result = generator.generate(request)
+        if isinstance(result, list):
+            result = result[0]
+
+        frames = result.frames
+        if frames is not None and len(frames):
+            Image.fromarray(frames[0]).save(output)
+            print(f"Saved image to {output}")
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_kandinsky5_i2v.py

@@ -0,0 +1,37 @@
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_kandinsky5_i2v"
+
+IMAGE_PATH = "assets/girl.png"
+
+
+def main():
+    generator = VideoGenerator.from_pretrained(
+        "kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers", 
+        # "kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers"
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+        # image_encoder_cpu_offload=False,
+    )
+
+    prompt = (
+        "A woman stands up and walks away"
+    )
+    _ = generator.generate_video(
+        prompt,
+        image_path=IMAGE_PATH,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        height=1024,
+        width=1024,
+        num_frames=121,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_kandinsky5_t2v.py

@@ -0,0 +1,37 @@
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_kandinsky5_t2v"
+
+def main():
+    generator = VideoGenerator.from_pretrained(
+        "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
+        # "kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers"
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+        # image_encoder_cpu_offload=False,
+    )
+
+    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."
+    )
+    _ = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True,height=512, width=768, num_frames=121)
+
+    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.")
+    _ = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, height=512, width=768, num_frames=121)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/edit_glm_image.py

@@ -0,0 +1,120 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Run GLM-Image image-to-image (edit) generation through FastVideo.
+
+User story:
+    "I have the HF `zai-org/GLM-Image` checkpoint and a condition image, and
+    want a minimal edit command (text + image -> edited image), saved as a PNG."
+
+GLM-Image is a single unified pipeline: passing a condition image switches it
+from text-to-image to the edit path (the condition enters the DiT via a KV-cache
+write pass), so the generator config is identical to `basic_glm_image.py` — the
+`inputs.pil_image` on the request is what selects the edit mode.
+"""
+import argparse
+from pathlib import Path
+
+from PIL import Image
+
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    InputConfig,
+    OutputConfig,
+    ParallelismConfig,
+    PipelineSelection,
+    SamplingConfig,
+)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Run GLM-Image image-to-image (edit) generation.")
+    parser.add_argument(
+        "--model-path",
+        default="zai-org/GLM-Image",
+        help="HF id or local diffusers-format GLM-Image weights directory.",
+    )
+    parser.add_argument(
+        "--image",
+        default="assets/images/couple.jpg",
+        help="Condition image to edit.",
+    )
+    parser.add_argument(
+        "--output",
+        default="image_output/edited.png",
+        help="Output PNG path.",
+    )
+    parser.add_argument(
+        "--prompt",
+        default="Change the background to a snowy mountain landscape at golden hour.",
+        help="Edit instruction.",
+    )
+    parser.add_argument("--height", type=int, default=1024)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--steps", type=int, default=50)
+    parser.add_argument("--guidance-scale", type=float, default=1.5)
+    parser.add_argument("--seed", type=int, default=1024)
+    parser.add_argument("--num-gpus", type=int, default=1)
+    parser.add_argument("--tp-size", type=int, default=None)
+    parser.add_argument("--sp-size", type=int, default=None)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+
+    output = Path(args.output)
+    output.parent.mkdir(parents=True, exist_ok=True)
+    condition = Image.open(args.image).convert("RGB")
+    tp_size = args.tp_size if args.tp_size is not None else (args.num_gpus if args.num_gpus > 1 else 1)
+    sp_size = args.sp_size if args.sp_size is not None else (1 if args.num_gpus > 1 else args.num_gpus)
+
+    # GLM-Image needs trust_remote_code for its AR encoder; offload and the
+    # pipeline class come from the model's registered defaults — don't override.
+    # The pipeline is registered as t2i; passing inputs.pil_image below switches
+    # it to the edit path.
+    generator_config = GeneratorConfig(
+        model_path=args.model_path,
+        trust_remote_code=True,
+        engine=EngineConfig(
+            num_gpus=args.num_gpus,
+            parallelism=ParallelismConfig(tp_size=tp_size, sp_size=sp_size),
+        ),
+        pipeline=PipelineSelection(workload_type="t2i"),
+    )
+
+    generator = VideoGenerator.from_config(generator_config)
+    try:
+        request = GenerationRequest(
+            prompt=args.prompt,
+            inputs=InputConfig(pil_image=condition),
+            sampling=SamplingConfig(
+                height=args.height,
+                width=args.width,
+                num_frames=1,
+                fps=1,
+                num_inference_steps=args.steps,
+                guidance_scale=args.guidance_scale,
+                seed=args.seed,
+            ),
+            output=OutputConfig(
+                output_path=str(output.parent),
+                save_video=False,
+                return_frames=True,
+            ),
+        )
+        result = generator.generate(request)
+        if isinstance(result, list):
+            result = result[0]
+
+        frames = result.frames
+        if frames is not None and len(frames):
+            Image.fromarray(frames[0]).save(output)
+            print(f"Saved image to {output}")
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/ltx2_3/README.md

@@ -0,0 +1,69 @@
+# LTX-2.3 distilled inference configs
+
+Ready-to-run `fastvideo generate` run configs for the LTX-2.3
+distilled model (`FastVideo/LTX-2.3-Distilled-Diffusers`), covering both
+workloads (t2v / i2v), both two-stage step schedules (`5+2`, `8+3` = denoise
++ refine), and four resolutions.
+
+```bash
+fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_1280x832.yaml
+```
+
+Each config is self-contained (no preset registry needed): the two-stage
+refine is wired via `generator.pipeline.preset_overrides.refine`, and the
+base sampling knobs live under `request.sampling`. The refine upsampler
+auto-resolves from the model's `spatial_upscaler`.
+
+## Configs
+
+| workload | schedule | resolution (HxW) | file |
+|---|---|---|---|
+| t2v | 5+2 | 1280x832 | `t2v_5s2_1280x832.yaml` |
+| t2v | 5+2 | 1024x1536 | `t2v_5s2_1024x1536.yaml` |
+| t2v | 5+2 | 768x1280 | `t2v_5s2_768x1280.yaml` |
+| t2v | 5+2 | 512x768 | `t2v_5s2_512x768.yaml` |
+| t2v | 8+3 | 1280x832 | `t2v_8s3_1280x832.yaml` |
+| t2v | 8+3 | 1024x1536 | `t2v_8s3_1024x1536.yaml` |
+| t2v | 8+3 | 768x1280 | `t2v_8s3_768x1280.yaml` |
+| t2v | 8+3 | 512x768 | `t2v_8s3_512x768.yaml` |
+| i2v | 5+2 | 1280x832 | `i2v_5s2_1280x832.yaml` |
+| i2v | 5+2 | 1024x1536 | `i2v_5s2_1024x1536.yaml` |
+| i2v | 5+2 | 768x1280 | `i2v_5s2_768x1280.yaml` |
+| i2v | 5+2 | 512x768 | `i2v_5s2_512x768.yaml` |
+| i2v | 8+3 | 1280x832 | `i2v_8s3_1280x832.yaml` |
+| i2v | 8+3 | 1024x1536 | `i2v_8s3_1024x1536.yaml` |
+| i2v | 8+3 | 768x1280 | `i2v_8s3_768x1280.yaml` |
+| i2v | 8+3 | 512x768 | `i2v_8s3_512x768.yaml` |
+
+## Overriding without editing a file
+
+Dotted overrides (prefixes `generator.` / `request.`) let you tweak any field:
+
+```bash
+# swap prompt
+fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_1280x832.yaml \
+  --request.prompt "a red fox running through fresh snow"
+
+# change output path / gpu count
+fastvideo generate --config examples/inference/ltx2_3/t2v_5s2_512x768.yaml \
+  --request.output.output_path outputs/preview.mp4 \
+  --generator.engine.num_gpus 4
+```
+
+## i2v
+
+The `i2v_*` configs take a first-frame image via
+`request.extensions.ltx2_images` (`[[path, frame_offset, weight]]`). Edit the
+path in the file, or override it:
+
+```bash
+fastvideo generate --config examples/inference/ltx2_3/i2v_8s3_1280x832.yaml \
+  --request.extensions.ltx2_images '[["/data/portrait.jpg", 0, 1.0]]'
+```
+
+## Schedules
+
+`5+2` is the fast preview schedule; `8+3` is the higher-quality distilled
+recipe. Refine (`preset_overrides.refine.num_inference_steps`) only accepts 2
+or 3 steps.
+

examples/inference/ltx2_3/i2v_5s2_1024x1536.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 5+2 two-stage at 1024x1536.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_5s2_1024x1536.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 1024
+    width: 1536
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_5s2_1024x1536.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_5s2_1280x832.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 5+2 two-stage at 1280x832.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_5s2_1280x832.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 1280
+    width: 832
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_5s2_1280x832.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_5s2_512x768.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 5+2 two-stage at 512x768.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_5s2_512x768.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 512
+    width: 768
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_5s2_512x768.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_5s2_768x1280.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 5+2 two-stage at 768x1280.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_5s2_768x1280.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 768
+    width: 1280
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_5s2_768x1280.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_8s3_1024x1536.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 8+3 two-stage at 1024x1536.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_8s3_1024x1536.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 1024
+    width: 1536
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_8s3_1024x1536.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_8s3_1280x832.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 8+3 two-stage at 1280x832.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_8s3_1280x832.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 1280
+    width: 832
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_8s3_1280x832.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_8s3_512x768.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 8+3 two-stage at 512x768.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_8s3_512x768.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 512
+    width: 768
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_8s3_512x768.mp4
+    save_video: true

examples/inference/ltx2_3/i2v_8s3_768x1280.yaml

@@ -0,0 +1,38 @@
+# LTX-2.3 distilled i2v — 8+3 two-stage at 768x1280.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/i2v_8s3_768x1280.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: i2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    The subject slowly turns toward the camera with a soft, natural expression, hair and clothing swaying gently, shallow depth of field, subtle cinematic motion.
+  sampling:
+    height: 768
+    width: 1280
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  # i2v conditioning — replace the path with your own first-frame image.
+  extensions:
+    ltx2_images:
+      - ["/path/to/your/first_frame.jpg", 0, 1.0]
+    ltx2_image_crf: 0.0
+  output:
+    output_path: outputs/ltx2_3_i2v_8s3_768x1280.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_5s2_1024x1536.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 5+2 two-stage at 1024x1536.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_5s2_1024x1536.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 1024
+    width: 1536
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  output:
+    output_path: outputs/ltx2_3_t2v_5s2_1024x1536.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_5s2_1280x832.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 5+2 two-stage at 1280x832.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_5s2_1280x832.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 1280
+    width: 832
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  output:
+    output_path: outputs/ltx2_3_t2v_5s2_1280x832.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_5s2_512x768.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 5+2 two-stage at 512x768.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_5s2_512x768.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 512
+    width: 768
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  output:
+    output_path: outputs/ltx2_3_t2v_5s2_512x768.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_5s2_768x1280.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 5+2 two-stage at 768x1280.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_5s2_768x1280.yaml
+#
+# Stage 1 denoises for 5 steps at half resolution; the latents are then
+# spatially upsampled and refined for 2 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 2
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 768
+    width: 1280
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 5
+  output:
+    output_path: outputs/ltx2_3_t2v_5s2_768x1280.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_8s3_1024x1536.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 8+3 two-stage at 1024x1536.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_1024x1536.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 1024
+    width: 1536
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  output:
+    output_path: outputs/ltx2_3_t2v_8s3_1024x1536.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_8s3_1280x832.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 8+3 two-stage at 1280x832.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_1280x832.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 1280
+    width: 832
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  output:
+    output_path: outputs/ltx2_3_t2v_8s3_1280x832.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_8s3_512x768.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 8+3 two-stage at 512x768.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_512x768.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 512
+    width: 768
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  output:
+    output_path: outputs/ltx2_3_t2v_8s3_512x768.mp4
+    save_video: true

examples/inference/ltx2_3/t2v_8s3_768x1280.yaml

@@ -0,0 +1,33 @@
+# LTX-2.3 distilled t2v — 8+3 two-stage at 768x1280.
+#
+# Run:
+#   fastvideo generate --config examples/inference/ltx2_3/t2v_8s3_768x1280.yaml
+#
+# Stage 1 denoises for 8 steps at half resolution; the latents are then
+# spatially upsampled and refined for 3 steps (refine only supports 2 or 3).
+# The refine upsampler auto-resolves from the model's `spatial_upscaler`.
+generator:
+  model_path: FastVideo/LTX-2.3-Distilled-Diffusers
+  engine:
+    num_gpus: 1
+  pipeline:
+    workload_type: t2v
+    preset_overrides:
+      refine:
+        enabled: true
+        num_inference_steps: 3
+        guidance_scale: 1.0
+        add_noise: true
+request:
+  prompt: >-
+    A cinematic drone shot flying over dramatic coastal cliffs at sunrise, golden light spilling across the water, gentle waves breaking on the rocks below, ultra-detailed, smooth camera motion.
+  sampling:
+    height: 768
+    width: 1280
+    num_frames: 121
+    fps: 24
+    guidance_scale: 1.0
+    num_inference_steps: 8
+  output:
+    output_path: outputs/ltx2_3_t2v_8s3_768x1280.mp4
+    save_video: true

examples/train/configs/consistency_model/wan/causal_cd_t2v.yaml

@@ -0,0 +1,78 @@
+# Causal Consistency Distillation: Wan 2.1 T2V 1.3B Causal
+#
+# ODE-data-free distillation. A frozen teacher takes a single CFG Euler step;
+# the student matches an EMA copy of itself at the next timestep, all under
+# clean-history teacher forcing.
+#
+# All three roles initialize from the SAME checkpoint (the teacher-forcing
+# AR-diffusion model). Point init_from at that checkpoint for a real run.
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanCausalModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: true
+  teacher:
+    _target_: fastvideo.train.models.wan.WanCausalModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: false
+  ema:
+    _target_: fastvideo.train.models.wan.WanCausalModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: false
+
+method:
+  _target_: fastvideo.train.methods.consistency_model.causal_cd.CausalConsistencyDistillationMethod
+  discrete_cd_N: 48
+  guidance_scale: 3.0
+  ema_decay: 0.99
+  ema_start_step: 200
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/Wan-Syn_77x448x832_600k
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 18
+    num_height: 448
+    num_width: 832
+    num_frames: 69
+
+  optimizer:
+    learning_rate: 2.0e-6
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 3000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/wan2.1_causal_cd
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 3
+
+  tracker:
+    project_name: distillation_wan_r
+    run_name: wan2.1_causal_cd_shift5
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    max_grad_norm: 1.0
+
+pipeline:
+  flow_shift: 5

examples/train/configs/distribution_matching/wan/anyflow_onpolicy_t2v.yaml

@@ -0,0 +1,105 @@
+# AnyFlow on-policy DMD — Wan 2.1 T2V 1.3B.
+#
+# Stage 2 of the AnyFlow two-stage recipe. Continues from the pretrain
+# checkpoint; refines the student via DMD2 with a multi-step Euler-flow
+# rollout from pure noise. Teacher provides the real score, critic
+# learns the fake score; both inherited from DMD2Method.
+#
+# Replace <PATH_TO_PRETRAIN_CKPT> with the output of the pretrain stage,
+# or with the NVIDIA-released checkpoint
+# nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers to bootstrap directly from
+# the paper weights (the delta_embedder rename is handled by the
+# param_names_mapping in WanVideoArchConfig).
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanModel
+    init_from: <PATH_TO_PRETRAIN_CKPT>
+    trainable: true
+  teacher:
+    _target_: fastvideo.train.models.wan.WanModel
+    init_from: Wan-AI/Wan2.1-T2V-14B-Diffusers
+    trainable: false
+    disable_custom_init_weights: true
+  critic:
+    _target_: fastvideo.train.models.wan.WanModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: true
+    disable_custom_init_weights: true
+
+method:
+  _target_: fastvideo.train.methods.distribution_matching.anyflow.AnyFlowMethod
+  rollout_mode: simulate
+  generator_update_interval: 5
+  real_score_guidance_scale: 3.0
+  dmd_denoising_steps: [999, 937, 833, 624]
+  warp_denoising_step: false
+
+  # AnyFlow rollout knobs.
+  student_sample_steps: 4
+  use_mean_velocity: true
+  t_list_override: [999.0, 937.0, 833.0, 624.0, 0.0]
+  dmd_score_r_value: 0.0  # DMD scoring conditioning is at r=0 (consistency target).
+
+  # Critic optimizer (DMD2 inherited).
+  fake_score_learning_rate: 8.0e-6
+  fake_score_betas: [0.0, 0.999]
+  fake_score_lr_scheduler: constant
+
+  attn_kind: vsa
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/preprocessed
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 21
+    num_height: 480
+    num_width: 832
+    num_frames: 81
+
+  optimizer:
+    learning_rate: 2.0e-6
+    betas: [0.0, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 4000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/wan2.1_anyflow_onpolicy
+    training_state_checkpointing_steps: 500
+    checkpoints_total_limit: 3
+    resume_from_checkpoint: latest
+
+  tracker:
+    project_name: anyflow-wan
+    run_name: wan2.1_t2v_anyflow_onpolicy
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+
+pipeline:
+  flow_shift: 5.0
+  dit_config:
+    r_embedder: true
+    r_embedder_fusion: gated
+    r_embedder_gate_value: 0.25
+    r_embedder_deltatime_type: r

examples/train/configs/distribution_matching/wan/anyflow_pretrain_t2v.yaml

@@ -0,0 +1,83 @@
+# AnyFlow pretrain (flow-map central-difference) — Wan 2.1 T2V 1.3B.
+#
+# Stage 1 of the AnyFlow two-stage recipe. Trains the dual-timestep
+# u_θ(x_t, t, r) on the central-difference target so the same checkpoint
+# can be sampled at arbitrary NFE in the on-policy stage.
+#
+# Initialize from base Wan 2.1 T2V 1.3B. No teacher or critic at this
+# stage; AnyFlowPretrainMethod owns a single student + one optimizer.
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: true
+
+method:
+  _target_: fastvideo.train.methods.distribution_matching.anyflow_pretrain.AnyFlowPretrainMethod
+  diffusion_ratio: 0.5
+  consistency_ratio: 0.25
+  epsilon: 5            # finite-difference step in absolute train-timestep units
+  weight_type: beta08   # per-timestep loss weight = t * sqrt(1 - t), renormalized
+  fuse_guidance_scale: 3.0
+  # shift is taken from pipeline.flow_shift below.
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/preprocessed
+    dataloader_num_workers: 4
+    train_batch_size: 4
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 21
+    num_height: 480
+    num_width: 832
+    num_frames: 81
+
+  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: 6000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/wan2.1_anyflow_pretrain
+    training_state_checkpointing_steps: 500
+    checkpoints_total_limit: 3
+    resume_from_checkpoint: latest
+
+  tracker:
+    project_name: anyflow-wan
+    run_name: wan2.1_t2v_anyflow_pretrain
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+
+pipeline:
+  flow_shift: 5.0
+  dit_config:
+    # Enable AnyFlow dual-timestep conditioning. The student loads from
+    # base Wan 2.1 — its checkpoint has no delta_embedder weights, so they
+    # get initialized identically to time_embedder via deep-copy in
+    # WanTimeTextImageEmbedding.__init__.
+    r_embedder: true
+    r_embedder_fusion: gated
+    r_embedder_gate_value: 0.25
+    r_embedder_deltatime_type: r

examples/train/configs/distribution_matching/wan/self_forcing_causal_t2v.yaml

@@ -100,9 +100,6 @@ callbacks:
     sampling_steps: [4]
     sampling_timesteps: [1000, 750, 500, 250]
     num_frames: 81
-    # Validation/inference uses standard CFG in both clean and Self-Forcing,
-    # so this directly matches Self-Forcing guidance_scale=3.0.
-    guidance_scale: 3.0
 
 pipeline:
   flow_shift: 5

examples/train/configs/fine_tuning/wan/dfsft_causal_t2v_framewise.yaml

@@ -0,0 +1,73 @@
+# DFSFT (Diffusion-Forcing SFT), frame-wise: Wan 2.1 T2V 1.3B Causal
+#
+# - Student: trainable causal Wan model with a block size of 1 frame
+# - Training: each frame gets its own independent noise level (frame-wise
+#   diffusion forcing), versus the chunk-wise variant that shares one noise
+#   level across num_frames_per_block frames.
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanCausalModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: true
+    num_frames_per_block: 1
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.dfsft.DiffusionForcingSFTMethod
+  chunk_size: 1
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/Wan-Syn_77x448x832_600k
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 18
+    num_height: 448
+    num_width: 832
+    num_frames: 69
+
+  optimizer:
+    learning_rate: 2.0e-6
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 4000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/wan2.1_causal_dfsft_framewise
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 3
+
+  tracker:
+    project_name: distillation_wan_r
+    run_name: wan2.1_causal_dfsft_framewise_shift5_gauss_weight
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    max_grad_norm: 1.0
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.wan.wan_causal_pipeline.WanCausalPipeline
+    dataset_file: examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/validation_4.json
+    every_steps: 50
+    sampling_steps: [40]
+    guidance_scale: 6.0
+    num_frames: 69
+
+pipeline:
+  flow_shift: 5

examples/train/configs/fine_tuning/wan/tfsft_causal_t2v.yaml

@@ -0,0 +1,72 @@
+# TFSFT (Teacher-Forcing SFT): Wan 2.1 T2V 1.3B Causal
+#
+# - Student: trainable causal Wan model
+# - Training: inhomogeneous timesteps per chunk, but the causal transformer
+#   denoises the current block while attending to *clean* history (clean_x),
+#   not its own noisy rollout.
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanCausalModel
+    init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
+    trainable: true
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.tfsft.TeacherForcingSFTMethod
+  chunk_size: 3
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/Wan-Syn_77x448x832_600k
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 18
+    num_height: 448
+    num_width: 832
+    num_frames: 69
+
+  optimizer:
+    learning_rate: 2.0e-6
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 4000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/wan2.1_causal_tfsft
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 3
+
+  tracker:
+    project_name: distillation_wan_r
+    run_name: wan2.1_causal_tfsft_shift5_gauss_weight
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    max_grad_norm: 1.0
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.wan.wan_causal_pipeline.WanCausalPipeline
+    dataset_file: examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/validation_4.json
+    every_steps: 50
+    sampling_steps: [40]
+    guidance_scale: 6.0
+    num_frames: 69
+
+pipeline:
+  flow_shift: 5

examples/train/scenario/worldmodel/README.md

@@ -1,32 +1,119 @@
 # World-Model: Matrix-Game 2.0 I2V
 
-Three training scenarios for the Matrix-Game 2.0 I2V world model on the
-new YAML-driven trainer (`fastvideo/train/entrypoint/train.py`).
+Training scenarios for the Matrix-Game 2.0 I2V world model on Solaris (Minecraft)
+data and Zelda data, using the new YAML-driven trainer
+(`fastvideo/train/entrypoint/train.py`).
+
+## Solaris Configs
 
 | Config | Method | Student | Notes |
 |---|---|---|---|
-| `finetune_i2v.yaml` | `FineTuneMethod` | `MatrixGame2Model` (bidirectional) | Multi-step SFT from `mg_bidirectional_Solaris`. |
-| `dfsft_causal_i2v.yaml` | `DiffusionForcingSFTMethod` | `MatrixGame2CausalModel` | Diffusion-Forcing SFT with chunkwise timesteps. |
-| `self_forcing_causal_i2v.yaml` | `SelfForcingMethod` | `MatrixGame2CausalModel` | DMD/Self-Forcing distillation; teacher = bidirectional, critic = bidirectional. |
+| `solaris/finetune_i2v.yaml` | `FineTuneMethod` | `MatrixGame2Model` (bidirectional) | Multi-step SFT from `mg_bidirectional_Solaris`. |
+| `solaris/dfsft_causal_i2v.yaml` | `DiffusionForcingSFTMethod` | `MatrixGame2CausalModel` | Diffusion-Forcing SFT with chunkwise timesteps. |
+| `solaris/self_forcing_causal_i2v.yaml` | `SelfForcingMethod` | `MatrixGame2CausalModel` | Matrix-Game 2.0 DMD/Self-Forcing distillation; teacher = bidirectional, critic = bidirectional. |
+
+## Zelda Configs
+
+| Config | Method | Student | Notes |
+|---|---|---|---|
+| `zelda/finetune_i2v.yaml` | `FineTuneMethod` | `MatrixGame2Model` (bidirectional) | Zelda bidirectional I2V finetuning from `FastVideo/Matrix-Game-2.0-Base-Diffusers`. Uses 33-frame clips and Zelda validation with action overlays. |
+| `zelda/dfsft_causal_i2v.yaml` | `DiffusionForcingSFTMethod` | `MatrixGame2CausalModel` | Zelda causal Diffusion-Forcing SFT from `mignonjia/mg_bidirectional_zelda`. Uses the same Zelda data, resolution, optimizer, and validation defaults as the Zelda finetune config. |
+| `zelda/self_forcing_causal_i2v.yaml` | `SelfForcingMethod` | `MatrixGame2CausalModel` | Zelda DMD/Self-Forcing distillation; student init = `mignonjia/mg_causal_zelda`, teacher = bidirectional (`mignonjia/mg_bidirectional_zelda`), critic = bidirectional. |
+| `zelda/streaming_long_tuning_causal_i2v.yaml` | `StreamingLongTuningMethod` | `MatrixGame2CausalModel` | LongLive-style streaming long tuning from the 1k-step Zelda self-forcing checkpoint. |
+
+Zelda world-model distillation is a two-run workflow: first run
+`zelda/self_forcing_causal_i2v.yaml` to train or load the 1k-step
+self-forcing checkpoint (`mignonjia/mg_sf_distilled_zelda_1k_steps`), then run
+`zelda/streaming_long_tuning_causal_i2v.yaml` for the 3k-step streaming
+long-tuning stage. The long-tuning YAML starts from that 1k-step checkpoint; it
+does not run the short self-forcing stage inside the same config.
+
+## Zelda Training Data
+
+The Zelda training configs use `data/zeldam2-clean` as a suggested local path.
+Download the dataset from Hugging Face before running those configs:
+
+```bash
+python scripts/huggingface/download_hf.py \
+    --repo_id mignonjia/zeldam2-clean \
+    --local_dir data/zeldam2-clean \
+    --repo_type dataset
+```
+
+You can store the dataset elsewhere; update `training.data.data_path` in the
+YAML to point at that location.
+
+## Multi3D Training Data
+
+`zelda/finetune_i2v.yaml` includes an optional, commented-out Multi3D entry.
+Enable it only when you want to mix Zelda with multi-game data from
+`data/multi3d_games`. You can store this dataset anywhere; before enabling it,
+update the matching commented `training.data.data_path` key in the YAML to the
+correct location.
+
+To mix datasets in a training YAML, set `training.data.data_path` to a
+path-to-repeat-count mapping. For example, `zelda/finetune_i2v.yaml` can use
+`data/zeldam2-clean: 1` and `# data/multi3d_games: 10`; uncommenting the
+Multi3D entry repeats the multi-game parquet list ten times before training
+samples are shuffled.
+
+## World Model Validation Data
+
+The Zelda validation configs expect a small public validation bundle under
+`data/zelda_validation_data`.
+
+Download it from Hugging Face before running the Zelda scenarios:
+
+```bash
+python scripts/huggingface/download_hf.py \
+    --repo_id mignonjia/zelda_validation_data \
+    --local_dir data/zelda_validation_data \
+    --repo_type dataset
+```
+
+The bundle contains `validation_zelda.json`, `images/`, and `actions/`.
+The Zelda configs point
+`callbacks.validation.dataset_file` at
+`data/zelda_validation_data/validation_zelda.json`.
 
 ## Usage
 
+### Solaris
+
 ```bash
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/finetune_i2v.yaml
+    examples/train/scenario/worldmodel/solaris/finetune_i2v.yaml
 
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/dfsft_causal_i2v.yaml
+    examples/train/scenario/worldmodel/solaris/dfsft_causal_i2v.yaml
 
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/self_forcing_causal_i2v.yaml
+    examples/train/scenario/worldmodel/solaris/self_forcing_causal_i2v.yaml
+```
+
+### Zelda
+
+```bash
+# Finetuning / DFSFT
+bash examples/train/run.sh \
+    examples/train/scenario/worldmodel/zelda/finetune_i2v.yaml
+
+bash examples/train/run.sh \
+    examples/train/scenario/worldmodel/zelda/dfsft_causal_i2v.yaml
+
+# Distillation / long tuning
+bash examples/train/run.sh \
+    examples/train/scenario/worldmodel/zelda/self_forcing_causal_i2v.yaml
+
+bash examples/train/run.sh \
+    examples/train/scenario/worldmodel/zelda/streaming_long_tuning_causal_i2v.yaml
 ```
 
 Override any field on the command line:
 
 ```bash
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/dfsft_causal_i2v.yaml \
+    examples/train/scenario/worldmodel/solaris/dfsft_causal_i2v.yaml \
     --training.distributed.num_gpus 8 \
     --training.optimizer.learning_rate 1e-5
 ```

examples/train/scenario/worldmodel/solaris/self_forcing_causal_i2v.yaml

@@ -97,4 +97,4 @@ callbacks:
     guidance_scale: 6.0
 
 pipeline:
-  flow_shift: 5
+  flow_shift: 5

examples/train/scenario/worldmodel/zelda/dfsft_causal_i2v.yaml

@@ -0,0 +1,94 @@
+# Diffusion-Forcing SFT: Zelda world model I2V Causal
+
+models:
+  student:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2_causal.MatrixGame2CausalModel
+    init_from: mignonjia/mg_bidirectional_zelda
+    trainable: true
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.dfsft.DiffusionForcingSFTMethod
+  chunk_size: 3
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path:
+      data/zeldam2-clean: 1
+    dataloader_num_workers: 1
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 42
+    num_latent_t: 9
+    num_height: 480
+    num_width: 832
+    num_frames: 33
+
+  optimizer:
+    learning_rate: 2.0e-5
+    betas: [0.9, 0.95]
+    weight_decay: 0.0
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 60000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/matrixgame_finetune/checkpoints/zelda_causal_dfsft
+    training_state_checkpointing_steps: 5000
+    checkpoints_total_limit: 3
+
+  tracker:
+    entity: hapo-exp
+    project_name: mg_1.3b_zelda
+    run_name: zelda_causal_dfsft
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+  dit_precision: fp32
+
+callbacks:
+  grad_clip:
+    max_grad_norm: 1.0
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.matrixgame2.matrixgame2_causal_dmd_pipeline.MatrixGame2CausalDMDPipeline
+    dataset_file: data/zelda_validation_data/validation_zelda.json
+    every_steps: 200
+    sampling_steps: [40]
+    sampling_timesteps: [1000, 975, 950, 925, 900, 875, 850, 825, 800, 775,
+                         750, 725, 700, 675, 650, 625, 600, 575, 550, 525,
+                         500, 475, 450, 425, 400, 375, 350, 325, 300, 275,
+                         250, 225, 200, 175, 150, 125, 100, 75, 50, 25]
+    num_frames: 33
+    overlay_actions: true
+    guidance_scale: 6.0
+    metrics:
+      enabled: true
+      names:
+        - vbench.imaging_quality
+        - vbench.aesthetic_quality
+        - vbench.temporal_flickering
+        - vbench.motion_smoothness
+        - vbench.subject_consistency
+        - vbench.background_consistency
+        - vbench.dynamic_degree
+        - optical_flow.synthetic_optical_flow
+      calibration_path: assets/eval/worldmodel_synthetic_flow_calibration.json
+      skip_missing_deps: true
+      strict: false
+      unload_after_validation: true
+
+pipeline:
+  flow_shift: 5
+  dit_config:
+    local_attn_size: 6
+    sink_size: 1

examples/train/scenario/worldmodel/zelda/finetune_i2v.yaml

@@ -0,0 +1,88 @@
+# Matrix-Game 2.0 Zelda + multi-game I2V finetune.
+
+models:
+  student:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2.MatrixGame2Model
+    init_from: FastVideo/Matrix-Game-2.0-Base-Diffusers
+    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: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path:
+      data/zeldam2-clean: 1
+      # data/multi3d_games: 10
+    dataloader_num_workers: 1
+    train_batch_size: 1
+    training_cfg_rate: 0.0  # unused for MatrixGame2 I2V; no text_embedding CFG dropout
+    seed: 42
+    num_latent_t: 9
+    num_height: 480
+    num_width: 832
+    num_frames: 33
+
+  optimizer:
+    learning_rate: 2.0e-5
+    betas: [0.9, 0.95]
+    weight_decay: 0.0
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 60000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/matrixgame_finetune/checkpoints/zelda_with_mg_init
+    training_state_checkpointing_steps: 5000
+    checkpoints_total_limit: 3
+
+  tracker:
+    project_name: mg_1.3b_zelda
+    run_name: zelda_with_mg_init
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+  dit_precision: fp32
+
+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.matrixgame2.matrixgame2_i2v_pipeline.MatrixGame2I2VPipeline
+    dataset_file: data/zelda_validation_data/validation_zelda.json
+    every_steps: 200
+    sampling_steps: [40]
+    num_frames: 33
+    overlay_actions: true
+    guidance_scale: 6.0
+    metrics:
+      enabled: true
+      names:
+        - vbench.imaging_quality
+        - vbench.aesthetic_quality
+        - vbench.temporal_flickering
+        - vbench.motion_smoothness
+        - vbench.subject_consistency
+        - vbench.background_consistency
+        - vbench.dynamic_degree
+        - optical_flow.synthetic_optical_flow
+      calibration_path: assets/eval/worldmodel_synthetic_flow_calibration.json
+      skip_missing_deps: true
+      strict: false
+      unload_after_validation: true
+
+pipeline:
+  flow_shift: 5

examples/train/scenario/worldmodel/zelda/self_forcing_causal_i2v.yaml

@@ -0,0 +1,120 @@
+# Self-forcing distillation: Zelda world model I2V Causal
+
+models:
+  student:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2_causal.MatrixGame2CausalModel
+    init_from: mignonjia/mg_causal_zelda
+    trainable: true
+  teacher:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2.MatrixGame2Model
+    init_from: mignonjia/mg_bidirectional_zelda
+    trainable: false
+    disable_custom_init_weights: true
+  critic:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2.MatrixGame2Model
+    init_from: mignonjia/mg_bidirectional_zelda
+    trainable: true
+    disable_custom_init_weights: true
+
+method:
+  _target_: fastvideo.train.methods.distribution_matching.self_forcing.SelfForcingMethod
+  rollout_mode: simulate
+  generator_update_interval: 5
+  dmd_denoising_steps: [1000, 750, 500, 250]
+  warp_denoising_step: true
+  min_timestep_ratio: 0.02
+  max_timestep_ratio: 0.98
+
+  chunk_size: 3
+  student_sample_type: sde
+  same_step_across_blocks: true
+  last_step_only: false
+  context_noise: 0.0
+  enable_gradient_in_rollout: true
+  start_gradient_frame: 0
+
+  # Critic optimizer
+  fake_score_learning_rate: 3.0e-7
+  fake_score_betas: [0.9, 0.95]
+  fake_score_lr_scheduler: constant
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/zeldam2-clean
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1001
+    num_latent_t: 9
+    num_height: 480
+    num_width: 832
+    num_frames: 33
+
+  optimizer:
+    learning_rate: 3.0e-6
+    betas: [0.9, 0.95]
+    weight_decay: 0.0
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 1000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/zelda_causal_self_forcing
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 1
+
+  tracker:
+    project_name: wangame_sf
+    run_name: mg2_self_forcing_9_latents
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+  dit_precision: fp32
+
+callbacks:
+  ema:
+    decay: 0.99
+    start_iter: 200
+  grad_clip:
+    max_grad_norm: 1.0
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.matrixgame2.matrixgame2_causal_dmd_pipeline.MatrixGame2CausalDMDPipeline
+    dataset_file: data/zelda_validation_data/validation_zelda.json
+    every_steps: 100
+    sampling_steps: [4]
+    sampling_timesteps: [1000, 750, 500, 250]
+    num_frames: 153
+    overlay_actions: true
+    keyboard_value_scale: 1.0
+    metrics:
+      enabled: true
+      names:
+        - vbench.imaging_quality
+        - vbench.aesthetic_quality
+        - vbench.temporal_flickering
+        - vbench.motion_smoothness
+        - vbench.subject_consistency
+        - vbench.background_consistency
+        - vbench.dynamic_degree
+        - optical_flow.synthetic_optical_flow
+      calibration_path: assets/eval/worldmodel_synthetic_flow_calibration.json
+      skip_missing_deps: true
+      strict: false
+      unload_after_validation: true
+
+pipeline:
+  flow_shift: 5
+  dit_config:
+    local_attn_size: 6
+    sink_size: 1

examples/train/scenario/worldmodel/zelda/streaming_long_tuning_causal_i2v.yaml

@@ -0,0 +1,140 @@
+# MatrixGame2 I2V LongLive-style streaming distillation: Zelda world model I2V Causal
+# Student init from self forcing after 1k steps
+
+models:
+  student:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2_causal.MatrixGame2CausalModel
+    init_from: mignonjia/mg_sf_distilled_zelda_1k_steps 
+    trainable: true
+    # transformer_override_safetensor: outputs/matrixgame_dmd/checkpoints/mg_zelda_sf_m2/checkpoint-1000_weight_only/ema/generator_ema.safetensors
+  teacher:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2.MatrixGame2Model
+    init_from: mignonjia/mg_bidirectional_zelda
+    trainable: false
+    disable_custom_init_weights: true
+  critic:
+    _target_: fastvideo.train.models.matrixgame2.matrixgame2.MatrixGame2Model
+    init_from: mignonjia/mg_bidirectional_zelda
+    trainable: true
+    disable_custom_init_weights: true
+
+method:
+  _target_: fastvideo.train.methods.distribution_matching.streaming_long_tuning.StreamingLongTuningMethod
+  rollout_mode: simulate
+  generator_update_interval: 5
+  dmd_denoising_steps: [1000, 750, 500, 250]
+  warp_denoising_step: true
+  min_timestep_ratio: 0.02
+  max_timestep_ratio: 0.98
+
+  chunk_size: 3
+  student_sample_type: sde
+  same_step_across_blocks: true
+  last_step_only: false
+  context_noise: 0.0
+  enable_gradient_in_rollout: true
+  start_gradient_frame: 0
+
+  streaming_training: true
+  streaming_chunk_size: 9
+  streaming_max_length: 39
+  streaming_fixed_overlap_latents: 3
+  streaming_reencode_overlap_anchor: true
+  streaming_anchor_inject_k: 1
+  streaming_require_full_blocks: true
+
+  multi_phased_distill_schedule:
+    - stage: streaming_long
+      start_step: 0
+      end_step: 3000
+      num_latent_t: 39
+      streaming_training: true
+      streaming_chunk_size: 9
+      streaming_max_length: 39
+      streaming_fixed_overlap_latents: 3
+
+  # Critic optimizer
+  fake_score_learning_rate: 3.0e-7
+  fake_score_betas: [0.9, 0.95]
+  fake_score_lr_scheduler: constant
+
+training:
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/zeldam2-clean
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1001
+    num_latent_t: 39
+    num_height: 480
+    num_width: 832
+    num_frames: 153
+
+  optimizer:
+    learning_rate: 3.0e-6
+    betas: [0.9, 0.95]
+    weight_decay: 0.0
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 3000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/zelda_causal_long_tuning
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 2
+
+  tracker:
+    project_name: wangame_sf
+    run_name: mg2_39only_streaming_long
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+  dit_precision: fp32
+
+callbacks:
+  ema:
+    decay: 0.99
+    start_iter: 200
+  grad_clip:
+    max_grad_norm: 1.0
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.matrixgame2.matrixgame2_causal_dmd_pipeline.MatrixGame2CausalDMDPipeline
+    dataset_file: data/zelda_validation_data/validation_zelda.json
+    every_steps: 100
+    sampling_steps: [4]
+    sampling_timesteps: [1000, 750, 500, 250]
+    num_frames: 153
+    overlay_actions: true
+    keyboard_value_scale: 1.0
+    metrics:
+      enabled: true
+      names:
+        - vbench.imaging_quality
+        - vbench.aesthetic_quality
+        - vbench.temporal_flickering
+        - vbench.motion_smoothness
+        - vbench.subject_consistency
+        - vbench.background_consistency
+        - vbench.dynamic_degree
+        - optical_flow.synthetic_optical_flow
+      calibration_path: assets/eval/worldmodel_synthetic_flow_calibration.json
+      skip_missing_deps: true
+      strict: false
+      unload_after_validation: true
+
+pipeline:
+  flow_shift: 5
+  dit_config:
+    local_attn_size: 6
+    sink_size: 1

fastvideo-kernel/CMakeLists.txt

@@ -12,6 +12,9 @@ set(_FASTVIDEO_USER_CUDA_ARCH "${CMAKE_CUDA_ARCHITECTURES}")
 if(NOT DEFINED GPU_BACKEND AND DEFINED ENV{GPU_BACKEND})
     set(GPU_BACKEND "$ENV{GPU_BACKEND}")
 endif()
+if(NOT GPU_BACKEND)
+    set(GPU_BACKEND "CUDA")
+endif()
 
 if(GPU_BACKEND STREQUAL "ROCM")
     enable_language(HIP)
@@ -50,7 +53,16 @@ if(NOT GPU_BACKEND STREQUAL "ROCM")
         if(_FASTVIDEO_USER_CUDA_ARCH)
             # Caller pinned -DCMAKE_CUDA_ARCHITECTURES (which torch ignores); translate it
             # to the TORCH_CUDA_ARCH_LIST spelling: "121" -> "12.1", "90a" -> "9.0a".
+            # Only numeric spellings translate; keywords like "native"/"all" would
+            # otherwise be mangled into nonsense ("nativ.e").
             foreach(_fv_arch IN LISTS _FASTVIDEO_USER_CUDA_ARCH)
+                if(NOT _fv_arch MATCHES "^[0-9]+[af]?$")
+                    message(FATAL_ERROR
+                        "fastvideo-kernel: CMAKE_CUDA_ARCHITECTURES='${_fv_arch}' is not "
+                        "supported. Use a numeric arch (e.g. 90a, 121), set "
+                        "TORCH_CUDA_ARCH_LIST directly (e.g. 9.0a), or unset both to "
+                        "auto-detect from the visible GPU.")
+                endif()
                 string(REGEX MATCH "[af]$" _fv_suffix "${_fv_arch}")
                 string(REGEX REPLACE "[af]$" "" _fv_num "${_fv_arch}")
                 string(REGEX REPLACE "(.)$" ".\\1" _fv_num "${_fv_num}")    # dot before the last digit
@@ -173,6 +185,14 @@ else()
     endif()
 endif()
 
+# ThunderKittens headers don't compile on aarch64 hosts: plain char is unsigned
+# there, and tk's base_types.cuh brace-initializes signed-char vector members
+# from char (narrowing error). Skip TK until upstream is aarch64-clean.
+if(ENABLE_TK_KERNELS AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64)$")
+    message(STATUS "ThunderKittens kernels: forced OFF on ${CMAKE_SYSTEM_PROCESSOR} (tk headers are not aarch64-clean)")
+    set(ENABLE_TK_KERNELS OFF)
+endif()
+
 if(ENABLE_TK_KERNELS)
     message(STATUS "ThunderKittens kernels: ENABLED")
 else()
@@ -263,11 +283,6 @@ set(CUDA_FLAGS
     "--expt-relaxed-constexpr"
     "-Xcompiler=-fno-strict-aliasing"
     "-Xcompiler=-fPIC"
-    # ARM/aarch64 defaults `char` to unsigned, but ThunderKittens headers assume the
-    # x86 signed-char behavior (else base_types.cuh hits "narrowing conversion from
-    # char to signed char"). Force signed char so TK compiles on Grace Hopper; this
-    # is a no-op on x86_64, where char is already signed.
-    "-Xcompiler=-fsigned-char"
     "-DTORCH_COMPILE"
     "-Xnvlink=--verbose"
     "-Xptxas=--verbose"
@@ -426,3 +441,14 @@ if(ENABLE_ATTN_QAT_INFER)
     install(TARGETS fp4attn_cuda LIBRARY DESTINATION .)
     install(TARGETS fp4quant_cuda LIBRARY DESTINATION .)
 endif()
+
+# One-look answer to "what is this build producing?" — kept last so it is the
+# final thing configure prints. The per-kernel matrix lives in README.md.
+message(STATUS "============== fastvideo-kernel build summary ==============")
+message(STATUS "host / backend:       ${CMAKE_SYSTEM_PROCESSOR} / ${GPU_BACKEND}")
+message(STATUS "TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST}")
+message(STATUS "fastvideo_kernel_ops: ON  (turbodiffusion int8-gemm/quant/rmsnorm/layernorm, all listed archs)")
+message(STATUS "  + TK sta/block_sparse (sm_90a only):          ${ENABLE_TK_KERNELS}")
+message(STATUS "fp4attn/fp4quant (sm_120a only, CUDA >= 12.8):  ${ENABLE_ATTN_QAT_INFER}")
+message(STATUS "Triton fallbacks ship in python/fastvideo_kernel/triton_kernels regardless.")
+message(STATUS "============================================================")

fastvideo-kernel/README.md

@@ -2,6 +2,42 @@
 
 CUDA kernels for FastVideo video generation.
 
+## Kernel inventory
+
+Compiled CUDA extensions (CMake, see the build summary printed at the end of every configure):
+
+| Extension | Kernels | Sources | GPU arch | Build gate |
+|---|---|---|---|---|
+| `fastvideo_kernel._C.fastvideo_kernel_ops` | TurboDiffusion INT8 GEMM, quant, RMSNorm, LayerNorm | `csrc/turbodiffusion/` | every arch in `TORCH_CUDA_ARCH_LIST` | always built |
+| same extension, optional part | ThunderKittens sliding-tile attention (`sta_fwd`) and VSA block-sparse (`block_sparse_fwd/bwd`) | `csrc/attention/*_h100.cu` | Hopper `sm_90a` only | `FASTVIDEO_KERNEL_BUILD_TK` (AUTO = ON iff `9.0a` is in the arch list; always OFF on aarch64 hosts — TK headers don't compile there) |
+| `fp4attn_cuda`, `fp4quant_cuda` | FP4 attention + quantization ("attn_qat_infer", modified SageAttention3) | `attn_qat_infer/` | consumer Blackwell `sm_120a` only, CUDA ≥ 12.8 | `FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER` (AUTO = ON iff `12.0a` is in the arch list) |
+
+Runtime-JIT kernels (no build step, ship in every wheel/image):
+
+| Kernels | Where | Used when |
+|---|---|---|
+| Triton: STA, VSA block-sparse, SLA, fused compress+topk, FP4 QAT training, quant/norm utils | `python/fastvideo_kernel/triton_kernels/` | automatic fallback when the matching C++ op is absent (`ops.py`, `turbodiffusion_ops.py`) |
+| FA4 CuTe-DSL block-sparse (VSA-256 fastpath on `sm_100`) | `block_sparse_attn_cute_fwd.py` | optional `flash_attn.cute` dependency, see below |
+| VMoBA `moba_attn_varlen` | `vmoba.py` | wraps flash-attn varlen |
+
+## What gets built where, and when
+
+| Surface | Trigger | Leg | `TORCH_CUDA_ARCH_LIST` | TK | FP4 |
+|---|---|---|---|---|---|
+| PyPI wheels (`.github/workflows/publish-kernel.yml`) | version bump in `fastvideo-kernel/pyproject.toml` on main, or manual dispatch | x86_64 cu126 | `9.0a` | ON | — (CUDA < 12.8) |
+| | | x86_64 cu130 | `9.0a;12.0a` | ON | ON |
+| | | aarch64 cu130 | `10.0a;12.0a` | — | ON |
+| Docker images `ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev` (`.github/workflows/infra-build-image.yml`) | `docker/Dockerfile` changes on main, or manual dispatch | amd64 cuda12.6.3 + cuda13.0.0 | `9.0a` | ON | — |
+| | | arm64 cuda12.6.3 (GH200) | `9.0a` | — (aarch64) | — |
+| | | arm64 cuda13.0.0 (GB10 / DGX Spark) | `12.1` | — | — |
+| Local `./build.sh` | manual | probes the visible GPU via torch | detected | ON iff sm_90 (non-aarch64 host) | ON iff sm_120 |
+
+Notes:
+
+- No Docker image ships the FP4 kernels; only the x86_64/aarch64 cu130 wheels do.
+- On arm64 images (GH200 included) STA/VSA run on the Triton fallbacks, since TK never builds on aarch64.
+- Kernel tests run on Buildkite GPU CI for PRs touching `fastvideo-kernel/**` (see `.buildkite/pipeline.yml`).
+
 ## Installation
 
 ### Standard Installation (Local Development)

fastvideo-kernel/build.sh

@@ -46,6 +46,23 @@ fi
 if git rev-parse --git-dir >/dev/null 2>&1; then
     git submodule update --init --recursive include/cutlass include/tk
 fi
+# Fail fast with a clear message if the headers are still missing (e.g. a
+# Docker context that excluded .git AND the submodule contents) instead of
+# dying later in a wall of nvcc include errors. CUTLASS is consumed by the
+# always-built turbodiffusion sources, so it is a hard error; ThunderKittens
+# only feeds the TK-gated Hopper kernels, so a missing tree just warns (the
+# TK gate resolves later, and non-SM90/ROCm builds never touch it).
+if [ ! -d include/cutlass/include ]; then
+    echo "ERROR: include/cutlass/include is missing. Outside a git checkout the" >&2
+    echo "       CUTLASS sources must already be present (run" >&2
+    echo "       'git submodule update --init --recursive include/cutlass include/tk'" >&2
+    echo "       in the source checkout, or include them in the build context)." >&2
+    exit 1
+fi
+if [ ! -d include/tk/include ]; then
+    echo "WARNING: include/tk/include is missing; ThunderKittens (Hopper sm_90a)" >&2
+    echo "         kernels cannot be built. Fine for non-SM90/ROCm targets." >&2
+fi
 
 # Install build dependencies
 uv pip install scikit-build-core cmake ninja

fastvideo/api/flux.py

@@ -0,0 +1,27 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from fastvideo.api.sampling_param import SamplingParam
+
+
+@dataclass
+class FluxSamplingParam(SamplingParam):
+
+    prompt: str | None = "a photo of a cat"
+    negative_prompt: str = ""
+
+    num_videos_per_prompt: int = 1
+    seed: int = 0
+
+    num_frames: int = 1
+    height: int = 1024
+    width: int = 1024
+    fps: int = 1
+
+    num_inference_steps: int = 28
+    guidance_scale: float = 3.5
+    use_embedded_guidance: bool = True
+    true_cfg_scale: float = 1.0

fastvideo/api/sampling_param.py

@@ -90,6 +90,10 @@ class SamplingParam:
     num_inference_steps_sr: int = 50
     guidance_scale: float = 1.0
     guidance_scale_2: float | None = None
+    # Embedded guidance (FLUX): do not treat ``guidance_scale > 1`` as classic CFG.
+    use_embedded_guidance: bool = False
+    # Diffusers-style true CFG for FLUX when > 1 (requires negative prompt encoding).
+    true_cfg_scale: float = 1.0
     guidance_rescale: float = 0.0
     boundary_ratio: float | None = None
     sigmas: list[float] | None = None
@@ -325,6 +329,18 @@ class SamplingParam:
             default=SamplingParam.guidance_rescale,
             help="Guidance rescale factor",
         )
+        parser.add_argument(
+            "--use-embedded-guidance",
+            action="store_true",
+            default=SamplingParam.use_embedded_guidance,
+            help="Use embedded guidance scale (FLUX-style) instead of classic CFG",
+        )
+        parser.add_argument(
+            "--true-cfg-scale",
+            type=float,
+            default=SamplingParam.true_cfg_scale,
+            help="True CFG scale for FLUX when > 1 (requires negative prompt encoding)",
+        )
         parser.add_argument(
             "--boundary-ratio",
             type=float,

fastvideo/api/schema.py

@@ -150,6 +150,7 @@ class SamplingConfig:
     guidance_scale_2: float | None = None
     guidance_rescale: float = 0.0
     true_cfg_scale: float | None = None
+    use_embedded_guidance: bool | None = None
     boundary_ratio: float | None = None
     sigmas: list[float] | None = None
 

fastvideo/attention/backends/flash_attn.py

@@ -5,99 +5,10 @@ import torch
 import torch.nn.functional as F
 from dataclasses import dataclass
 
-try:
-    from fastvideo.attention.utils.flash_attn_cute import flash_attn_func
-
-    fa_version = "4"
-except ImportError:
-    try:
-        from flash_attn_interface import flash_attn_func as flash_attn_3_func
-
-        # flash_attn 3 no longer have a different API, see following commit:
-        # https://github.com/Dao-AILab/flash-attention/commit/ed209409acedbb2379f870bbd03abce31a7a51b7
-        flash_attn_func = flash_attn_3_func
-        fa_version = "3"
-    except ImportError:
-        from flash_attn import flash_attn_func as flash_attn_2_func
-        flash_attn_func = flash_attn_2_func
-        fa_version = "2"
-
-# torch.compile traceability: the FA4/cute path (fa_version=="4") is
-# already a registered torch.library custom op, so dynamo treats it as a
-# graph node. The external FA2/FA3 `flash_attn_func` is NOT — dynamo
-# breaks the graph at the call site (observed: wanvideo.py self-attn,
-# once per layer every step), which fragments the compiled region and
-# blocks CUDA-graph capture. Wrap the FA2/FA3 default call in a custom
-# op (mirrors the FP4 `flash_attn_cute` template) so it becomes an
-# opaque-but-traceable node. The kernel still runs eager inside the op
-# (correct — flash-attn must run eager); only dynamo's treatment of the
-# boundary changes, so numerics are unchanged (SSIM-gate to confirm).
-if fa_version in ("2", "3"):
-    _fa_default = flash_attn_func
-
-    # Scope: this op covers exactly the q/k/v + softmax_scale + causal
-    # call shape used by FlashAttentionImpl.forward's default branch
-    # (see `flash_attn_func_compilable(...)` call site below). The
-    # masked/no-pad and varlen / cross-attn paths use different
-    # entry points (`flash_attn_no_pad`, `flash_attn_varlen_*`) which
-    # are intentionally out of scope for this PR — wrapping them is a
-    # natural follow-up. The wrapper's signature is the contract: any
-    # extra kwarg (dropout_p, window_size, alibi_slopes, deterministic,
-    # return_attn_probs, ...) raises TypeError at the call site, so
-    # silent loss of kwargs is not a failure mode.
-    @torch.library.custom_op(
-        "fastvideo::_flash_attn_default_forward",
-        mutates_args=(),
-        device_types="cuda",
-    )
-    def _flash_attn_default_forward(
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        softmax_scale: float | None,
-        causal: bool,
-    ) -> torch.Tensor:
-        return _fa_default(q, k, v, softmax_scale=softmax_scale, causal=causal)
-
-    @torch.library.register_fake("fastvideo::_flash_attn_default_forward")
-    def _flash_attn_default_forward_fake(
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        softmax_scale: float | None,
-        causal: bool,
-    ) -> torch.Tensor:
-        del softmax_scale, causal
-        # FA2/FA3 default path: [batch, seqlen_q, nheads, head_dim_v],
-        # same dtype/device as q (head dim taken from v).
-        return q.new_empty(q.shape[0], q.shape[1], q.shape[2], v.shape[-1])
-
-    def flash_attn_func_compilable(q, k, v, softmax_scale=None, causal=False):
-        # Autograd carve-out. The custom op above registers a forward + fake
-        # kernel but NO backward (register_autograd), so it is opaque to
-        # autograd. Inference runs under no_grad / inference_mode and routes
-        # through the traceable custom op — that is the torch.compile win, and
-        # the only path this PR claims. Training backprops through attention,
-        # so route grad-enabled calls to the original FA2/FA3 `flash_attn_func`
-        # (itself an autograd.Function, so backward is correct) at the cost of a
-        # dynamo graph break on the training path — i.e. pre-PR behavior, no
-        # regression. Full autograd parity for the custom op (mirroring the FP4
-        # cute template) is a tracked follow-up.
-        if torch.is_grad_enabled() and (q.requires_grad or k.requires_grad or v.requires_grad):
-            return _fa_default(q, k, v, softmax_scale=softmax_scale, causal=causal)
-        return torch.ops.fastvideo._flash_attn_default_forward(q, k, v, softmax_scale, causal)
-elif fa_version == "4":
-    # FA4 path: `flash_attn_func` is already a torch.library custom op
-    # (registered in `fastvideo.attention.utils.flash_attn_cute`), so a
-    # passthrough is enough — no extra registration needed.
-    def flash_attn_func_compilable(q, k, v, softmax_scale=None, causal=False):
-        return flash_attn_func(q, k, v, softmax_scale=softmax_scale, causal=causal)
-else:
-    # Defensive: the probe above only ever sets fa_version to "2", "3",
-    # or "4"; an unexpected value means an import/probe regression and
-    # we want a loud error at import, not a silent NameError later.
-    raise RuntimeError(f"Unsupported FlashAttention version: {fa_version!r} — expected "
-                       f"'2', '3', or '4' from the import probe above.")
+from fastvideo.attention.utils.flash_attn_default import (
+    fa_version,
+    flash_attn_func_compilable,
+)
 
 from fastvideo.attention.backends.abstract import (
     AttentionBackend,
@@ -108,8 +19,6 @@ from fastvideo.attention.backends.abstract import (
 from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
-
-_WARNED_NON_FA_DTYPE = False
 logger.info("Using FlashAttention-%s backend", fa_version)
 
 # FP4 FA4 support: quantize Q/K to NVFP4 E2M1 for block-scaled MMA on Blackwell.
@@ -271,12 +180,8 @@ class FlashAttentionImpl(AttentionImpl):
         # SP). Cast through bf16 and restore, matching TORCH_SDPA's tolerance.
         orig_dtype = query.dtype
         if orig_dtype not in (torch.float16, torch.bfloat16):
-            global _WARNED_NON_FA_DTYPE
-            if not _WARNED_NON_FA_DTYPE:
-                _WARNED_NON_FA_DTYPE = True
-                logger.warning(
-                    "FLASH_ATTN received %s inputs; casting to bfloat16 for the "
-                    "kernel and restoring on output.", orig_dtype)
+            logger.warning_once(f"FLASH_ATTN received {orig_dtype} inputs; casting to "
+                                f"bfloat16 for the kernel and restoring on output.")
             query = query.to(torch.bfloat16)
             key = key.to(torch.bfloat16)
             value = value.to(torch.bfloat16)
@@ -293,9 +198,17 @@ class FlashAttentionImpl(AttentionImpl):
         attn_metadata: FlashAttnMetadata,
     ):
         if (attn_metadata is not None and hasattr(attn_metadata, "attn_mask") and attn_metadata.attn_mask is not None):
+            # Route through the *_compilable wrappers so dynamo sees one
+            # traceable node for each masked entry point (the unpad/pad
+            # bookkeeping runs eager inside the custom op). On FA2 these
+            # wrappers go through ops with full register_autograd, so
+            # training also backprops through the op (no graph break on
+            # the training path); on FA3/FA4 they carve out to the
+            # autograd.Function for grad-enabled calls — see
+            # fastvideo/attention/utils/flash_attn_no_pad.py.
             from fastvideo.attention.utils.flash_attn_no_pad import (
-                flash_attn_no_pad,
-                flash_attn_varlen_qk_no_pad,
+                flash_attn_no_pad_compilable as flash_attn_no_pad,
+                flash_attn_varlen_qk_no_pad_compilable as flash_attn_varlen_qk_no_pad,
             )
 
             attn_mask = attn_metadata.attn_mask
@@ -322,7 +235,11 @@ class FlashAttentionImpl(AttentionImpl):
                 )
 
             qkv = torch.stack([query, key, value], dim=2)
-            attn_mask_padded = F.pad(attn_mask, (qkv.shape[1] - attn_mask.shape[1], 0), value=True)
+            key_padding_mask = _key_padding_mask_from_attn_mask(attn_mask, attn_mask.shape[-1]).to(device=query.device)
+            if key_padding_mask.shape[-1] > qkv.shape[1]:
+                raise ValueError("Invalid key padding mask length for FLASH_ATTN: "
+                                 f"expected at most {qkv.shape[1]}, got {key_padding_mask.shape[-1]}")
+            attn_mask_padded = F.pad(key_padding_mask, (qkv.shape[1] - key_padding_mask.shape[-1], 0), value=True)
             output = flash_attn_no_pad(qkv, attn_mask_padded, causal=False, dropout_p=0, softmax_scale=None)
         elif self.nvfp4_fa4:
             output = self._forward_nvfp4(query, key, value)

fastvideo/attention/backends/nabla.py

@@ -0,0 +1,147 @@
+# SPDX-License-Identifier: Apache-2.0
+"""NABLA block-sparse flex-attention backend (Kandinsky5 "nabla" checkpoints).
+
+The block mask is data-dependent: nablaT_v2 mean-pools 64-token blocks of the
+fractal-ordered sequence, thresholds the softmaxed block map, and ORs it with a
+precomputed spatio-temporal-window (STA) mask carried on the attention
+metadata. The mask spans the full sequence, so this backend does not support
+sequence parallelism — use it via LocalAttention only.
+"""
+
+import math
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+
+try:
+    from torch.nn.attention.flex_attention import BlockMask, flex_attention
+    flex_attention = torch.compile(flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs")
+    CAN_USE_FLEX_ATTN = True
+except ImportError:
+    CAN_USE_FLEX_ATTN = False
+
+from fastvideo.attention.backends.abstract import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+
+
+def nablaT_v2(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    sta: torch.Tensor,
+    thr: float = 0.9,
+) -> "BlockMask":
+    q = q.transpose(1, 2).contiguous()
+    k = k.transpose(1, 2).contiguous()
+
+    # Map estimation
+    B, h, S, D = q.shape
+    s1 = S // 64
+    qa = q.reshape(B, h, s1, 64, D).mean(-2)
+    ka = k.reshape(B, h, s1, 64, D).mean(-2).transpose(-2, -1)
+    map = qa @ ka
+
+    map = torch.softmax(map / math.sqrt(D), dim=-1)
+    # Map binarization
+    vals, inds = map.sort(-1)
+    cvals = vals.cumsum_(-1)
+    mask = (cvals >= 1 - thr).int()
+    mask = mask.gather(-1, inds.argsort(-1))
+
+    mask = torch.logical_or(mask, sta)
+
+    # BlockMask creation
+    kv_nb = mask.sum(-1).to(torch.int32)
+    kv_inds = mask.argsort(dim=-1, descending=True).to(torch.int32)
+    return BlockMask.from_kv_blocks(torch.zeros_like(kv_nb), kv_inds, kv_nb, kv_inds, BLOCK_SIZE=64, mask_mod=None)
+
+
+class NablaAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "NABLA_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["NablaAttentionImpl"]:
+        return NablaAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["NablaAttentionMetadata"]:
+        return NablaAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["NablaAttentionMetadataBuilder"]:
+        return NablaAttentionMetadataBuilder
+
+
+@dataclass
+class NablaAttentionMetadata(AttentionMetadata):
+    # Block-level STA window mask [1, 1, S/64, S/64], precomputed once per run.
+    sta_mask: torch.Tensor = None  # type: ignore[assignment]
+    # Cumulative-probability threshold for block-map binarization.
+    P: float = 0.9
+    visual_shape: tuple[int, int, int] = (0, 0, 0)
+
+
+class NablaAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self) -> None:
+        pass
+
+    def prepare(self) -> None:
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        sta_mask: torch.Tensor,
+        P: float,
+        visual_shape: tuple[int, int, int],
+        **kwargs: Any,
+    ) -> NablaAttentionMetadata:
+        return NablaAttentionMetadata(
+            current_timestep=current_timestep,
+            sta_mask=sta_mask,
+            P=P,
+            visual_shape=visual_shape,
+        )
+
+
+class NablaAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        softmax_scale: float,
+        causal: bool = False,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        if not CAN_USE_FLEX_ATTN:
+            raise RuntimeError("NABLA attention requires torch.nn.attention.flex_attention, "
+                               "which is unavailable in this PyTorch build.")
+        if causal:
+            raise ValueError("NABLA attention does not support causal masking.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: NablaAttentionMetadata,
+    ) -> torch.Tensor:
+        # q/k/v: [B, S, heads, head_dim], fractal-ordered by the model; S % 64 == 0.
+        block_mask = nablaT_v2(query, key, attn_metadata.sta_mask, thr=attn_metadata.P)
+        return flex_attention(
+            query=query.transpose(1, 2),
+            key=key.transpose(1, 2),
+            value=value.transpose(1, 2),
+            block_mask=block_mask,
+        ).transpose(1, 2)

fastvideo/attention/backends/sdpa.py

@@ -2,6 +2,7 @@
 
 import torch
 from dataclasses import dataclass
+from torch.nn import functional as F
 from fastvideo.attention.backends.abstract import (  # FlashAttentionMetadata,
     AttentionBackend, AttentionImpl, AttentionMetadata, AttentionMetadataBuilder)
 from fastvideo.logger import init_logger
@@ -14,12 +15,18 @@ class SDPABackend(AttentionBackend):
     accept_output_buffer: bool = True
 
     @staticmethod
-    def get_supported_head_sizes() -> list[int]:
-        return [32, 64, 96, 128, 160, 192, 224, 256]
+    def get_supported_head_sizes() -> list[int] | None:
+        # torch.nn.functional.scaled_dot_product_attention is head-size
+        # agnostic: the math backend handles any size and the fused kernels
+        # fall back internally when they cannot. The previous list
+        # ([32, 64, ..., 256]) was copied from FlashAttentionBackend in the
+        # v1 refactor (#270) and wrongly excluded sizes such as 80 (e.g.
+        # CLIP vision encoders). None means "no restriction".
+        return None
 
     @staticmethod
     def get_name() -> str:
-        return "SDPA"
+        return "TORCH_SDPA"
 
     @staticmethod
     def get_impl_cls() -> type["SDPAImpl"]:
@@ -49,9 +56,51 @@ class SDPAMetadataBuilder(AttentionMetadataBuilder):
             current_timestep: int,
             attn_mask: torch.Tensor,
     ) -> SDPAMetadata:
+        # Store the mask exactly as passed. The metadata is cross-backend:
+        # call sites (HYWorld, HunyuanVideo15) build SDPAMetadata while the
+        # layer's selector may pick FLASH_ATTN, and the shared convention for
+        # padding masks is the tokenizer-style 2D [batch, key_len]. Any
+        # reshaping for torch.sdpa happens inside the SDPA impl
+        # (_normalize_attn_mask_for_sdpa).
         return SDPAMetadata(current_timestep=current_timestep, attn_mask=attn_mask)
 
 
+def _normalize_attn_mask_for_sdpa(
+    attn_mask: torch.Tensor | None,
+    query: torch.Tensor,
+    key: torch.Tensor,
+) -> torch.Tensor | None:
+    if attn_mask is None:
+        return None
+
+    attn_mask = attn_mask.to(device=query.device)
+    # F.scaled_dot_product_attention only accepts bool or float masks;
+    # tokenizers commonly produce int64 0/1 padding masks.
+    if attn_mask.dtype != torch.bool and not attn_mask.dtype.is_floating_point:
+        attn_mask = attn_mask != 0
+
+    key_len = key.shape[-2]
+    if attn_mask.shape[-1] > key_len:
+        raise ValueError("Invalid attention mask length for SDPA: "
+                         f"expected at most {key_len}, got {attn_mask.shape[-1]}")
+    if attn_mask.shape[-1] < key_len:
+        # Front-pad as "attend": double-stream layouts (HYWorld) prepend
+        # non-text tokens the tokenizer mask does not cover.
+        valid_value = True if attn_mask.dtype == torch.bool else 0.0
+        attn_mask = F.pad(attn_mask, (key_len - attn_mask.shape[-1], 0), value=valid_value)
+
+    if attn_mask.dim() == 2:
+        # In-tree producers pass 2D [batch, key_len] padding masks; lift to a
+        # broadcastable [batch, 1, 1, key_len] here so torch.sdpa does not
+        # reinterpret 2D as its documented [query_len, key_len] broadcast.
+        return attn_mask[:, None, None, :]
+    if attn_mask.dim() == 3:
+        return attn_mask[:, None, :, :]
+    if attn_mask.dim() == 4:
+        return attn_mask
+    raise ValueError(f"Unsupported attention mask shape for SDPA: {attn_mask.shape}")
+
+
 class SDPAImpl(AttentionImpl):
 
     def __init__(
@@ -82,6 +131,7 @@ class SDPAImpl(AttentionImpl):
 
         attn_mask = attn_metadata.attn_mask if (attn_metadata is not None
                                                 and hasattr(attn_metadata, "attn_mask")) else None
+        attn_mask = _normalize_attn_mask_for_sdpa(attn_mask, query, key)
         attn_kwargs = {
             "attn_mask": attn_mask,
             "dropout_p": self.dropout,

fastvideo/attention/layer.py

@@ -252,6 +252,7 @@ class LocalAttention(nn.Module):
                  causal: bool = False,
                  supported_attention_backends: tuple[AttentionBackendEnum, ...]
                  | None = None,
+                 default_backend: AttentionBackendEnum | None = None,
                  **extra_impl_args) -> None:
         super().__init__()
         if softmax_scale is None:
@@ -262,7 +263,10 @@ class LocalAttention(nn.Module):
             num_kv_heads = num_heads
 
         dtype = get_compute_dtype()
-        attn_backend = get_attn_backend(head_size, dtype, supported_attention_backends=supported_attention_backends)
+        attn_backend = get_attn_backend(head_size,
+                                        dtype,
+                                        supported_attention_backends=supported_attention_backends,
+                                        default_backend=default_backend)
         impl_cls = attn_backend.get_impl_cls()
         self.attn_impl = impl_cls(num_heads=num_heads,
                                   head_size=head_size,

fastvideo/attention/selector.py

@@ -84,8 +84,9 @@ def get_attn_backend(
     dtype: torch.dtype,
     supported_attention_backends: tuple[AttentionBackendEnum, ...]
     | None = None,
+    default_backend: AttentionBackendEnum | None = None,
 ) -> type[AttentionBackend]:
-    return _cached_get_attn_backend(head_size, dtype, supported_attention_backends)
+    return _cached_get_attn_backend(head_size, dtype, supported_attention_backends, default_backend)
 
 
 @cache
@@ -94,6 +95,7 @@ def _cached_get_attn_backend(
     dtype: torch.dtype,
     supported_attention_backends: tuple[AttentionBackendEnum, ...]
     | None = None,
+    default_backend: AttentionBackendEnum | None = None,
 ) -> type[AttentionBackend]:
     # Check whether a particular choice of backend was
     # previously forced.
@@ -112,6 +114,12 @@ def _cached_get_attn_backend(
         if backend_by_env_var is not None:
             selected_backend = backend_name_to_enum(backend_by_env_var)
 
+    # Layer-level default (e.g. a checkpoint that requires a specific sparse
+    # backend). Lower precedence than the global force and the env var, so
+    # users can still override it.
+    if selected_backend is None and default_backend is not None:
+        selected_backend = default_backend
+
     # get device-specific attn_backend
     from fastvideo.platforms import current_platform
 

fastvideo/attention/utils/flash_attn_cute.py

@@ -4,10 +4,9 @@ import functools
 from collections.abc import Callable
 
 import torch
-from flash_attn import flash_attn_func as _flash_attn_2_func
-from flash_attn import flash_attn_varlen_func as _flash_attn_2_varlen_func
 
 from fastvideo.logger import init_logger
+from fastvideo.platforms import current_platform
 
 logger = init_logger(__name__)
 
@@ -16,7 +15,8 @@ if torch.cuda.is_available():
         from flash_attn.cute.interface import _flash_attn_bwd, _flash_attn_fwd
     except ImportError:
         # flash_attn.cute (FA4) is simply not installed -- expected on builds
-        # without it; callers fall back to FA3/FA2 quietly.
+        # without it; callers handle the ImportError (the FASTVIDEO_FA4 gate in
+        # flash_attn.py raises, the FP4 probe treats FA4 as unavailable).
         raise
     except Exception as e:
         # flash_attn.cute IS installed but failed to import -- almost always an
@@ -24,23 +24,57 @@ if torch.cuda.is_available():
         # 'cutlass.cute.core' has no attribute 'ThrMma'" (an AttributeError, not
         # ImportError). This is fixable by pinning a compatible
         # nvidia-cutlass-dsl, so warn loudly, then re-raise as ImportError so
-        # callers fall back to FA3/FA2 instead of crashing worker init.
+        # callers can handle it uniformly.
         logger.warning(
-            "flash_attn.cute (FA4) is installed but failed to import (%r); "
-            "falling back to FA3/FA2. This is usually an nvidia-cutlass-dsl "
-            "version mismatch -- pin a compatible nvidia-cutlass-dsl to "
-            "restore FA4.", e)
+            "flash_attn.cute (FA4) is installed but failed to import (%r). "
+            "This is usually an nvidia-cutlass-dsl version mismatch -- pin a "
+            "compatible nvidia-cutlass-dsl to restore FA4.", e)
         raise ImportError(f"flash_attn.cute (FA4) import failed: {e!r}") from e
 else:
     # This error will be caught in flash_attn.py or flash_attn_no_pad.py
     raise ImportError("flash_attn.cute is only available on CUDA devices; this error must be handled internally")
 
+try:
+    # FA2 serves the calls FA4 cute cannot on pre-sm90 GPUs (backward, GQA).
+    # Optional so FA4-only installs can still import this module.
+    from flash_attn import flash_attn_func as _flash_attn_2_func
+    from flash_attn import flash_attn_varlen_func as _flash_attn_2_varlen_func
+except ImportError:
+    _flash_attn_2_func = None
+    _flash_attn_2_varlen_func = None
+
 
 def _check_dropout(dropout_p: float) -> None:
     if dropout_p != 0.0:
         raise NotImplementedError(f"flash_attn.cute does not support dropout (got dropout_p={dropout_p})")
 
 
+@functools.cache
+def _sm90_or_newer() -> bool:
+    return current_platform.has_device_capability(90)
+
+
+def _use_fa2(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> bool:
+    if _sm90_or_newer():
+        return False
+    # Pre-sm90 FA4 cute limitations, both served by FA2 (deterministic
+    # capability gate, not a runtime fallback):
+    #   * the backward asserts sm90+ (L40S/sm_89 dies on its arch check);
+    #   * GQA fails CuTeDSL JIT in pack_gqa ("ValueError: Operation creation
+    #     failed", observed on sm_89 with HunyuanGameCraft/LTX2).
+    if q.shape[-2] != k.shape[-2]:
+        return True
+    return torch.is_grad_enabled() and any(t.requires_grad for t in (q, k, v))
+
+
+def _fa2_or_raise(fa2_func: Callable | None) -> Callable:
+    if fa2_func is None:
+        raise RuntimeError("this attention call cannot run on FA4 cute below sm90 (its backward and "
+                           "GQA support require sm90+) and flash-attn 2, which serves it there, is "
+                           "not installed.")
+    return fa2_func
+
+
 @torch.library.custom_op(
     "fastvideo::_flash_attn_cute_forward",
     mutates_args=(),
@@ -243,70 +277,6 @@ torch.library.register_autograd(
 )
 
 
-# FA4's CuTeDSL kernels JIT-compile per shape family, and some configurations
-# fail MLIR op creation at runtime even though the import succeeded (observed:
-# GQA models on sm_89 dying in pack_gqa with "ValueError: Operation creation
-# failed"). Degrade to FA2 once, process-wide, instead of crashing inference.
-class _FA4Policy:
-    """Per-call gate for the FA4 cute fast path, with FA2 as the fallback.
-
-    FA4 is skipped when:
-      * a previous call failed at runtime -- CuTeDSL JIT compilation is
-        shape-dependent, so the first failure disables FA4 for the rest of
-        the process instead of retrying a broken JIT on every call; or
-      * the call needs autograd -- FA4's backward asserts sm90+ (L40S/sm_89
-        dies on its arch check) and is unvalidated for training in this repo
-        (its lse is not even allocated through our inference-shaped custom
-        op), so training keeps the pre-FA4 behavior: FA2 on every device.
-    """
-
-    def __init__(self) -> None:
-        self.broken = False
-
-    def use_fa4(self, *tensors: torch.Tensor) -> bool:
-        if self.broken:
-            return False
-        return not (torch.is_grad_enabled() and any(t.requires_grad for t in tensors))
-
-    def mark_broken(self, error: Exception) -> None:
-        if not self.broken:
-            self.broken = True
-            logger.warning(
-                "flash_attn.cute (FA4) failed at runtime (%r); falling back "
-                "to FA2 for the rest of this process.", error)
-
-
-_FA4 = _FA4Policy()
-
-
-def _with_fa2_fallback(fa2_func: Callable) -> Callable:
-    """Pair an FA4 cute wrapper with its FA2 twin of the same signature.
-
-    The decorated body runs only when ``_FA4`` allows it; otherwise (or after
-    the first FA4 runtime failure) the call is served by ``fa2_func``.
-    ``NotImplementedError`` is a contract error (e.g. dropout), not a JIT
-    failure, so it propagates without disabling FA4.
-    """
-
-    def decorator(fa4_func: Callable) -> Callable:
-
-        @functools.wraps(fa4_func)
-        def wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, *args, **kwargs) -> torch.Tensor:
-            if _FA4.use_fa4(q, k, v):
-                try:
-                    return fa4_func(q, k, v, *args, **kwargs)
-                except NotImplementedError:
-                    raise
-                except Exception as e:  # CuTeDSL compile errors surface as ValueError
-                    _FA4.mark_broken(e)
-            return fa2_func(q, k, v, *args, **kwargs)
-
-        return wrapper
-
-    return decorator
-
-
-@_with_fa2_fallback(_flash_attn_2_func)
 def flash_attn_func(
     q: torch.Tensor,
     k: torch.Tensor,
@@ -317,6 +287,16 @@ def flash_attn_func(
     deterministic: bool = False,
 ) -> torch.Tensor:
     """Only returns the output, not the lse."""
+    if _use_fa2(q, k, v):
+        return _fa2_or_raise(_flash_attn_2_func)(
+            q,
+            k,
+            v,
+            dropout_p=dropout_p,
+            softmax_scale=softmax_scale,
+            causal=causal,
+            deterministic=deterministic,
+        )
     _check_dropout(dropout_p)
     out, _ = torch.ops.fastvideo._flash_attn_cute_forward(q, k, v, softmax_scale, causal, deterministic)
     return out
@@ -392,7 +372,6 @@ def flash_attn_fp4_func(
     return torch.ops.fastvideo._flash_attn_cute_fp4_forward(q, k, v, sfq, sfk, softmax_scale, causal)
 
 
-@_with_fa2_fallback(_flash_attn_2_varlen_func)
 def flash_attn_varlen_func(
     q: torch.Tensor,
     k: torch.Tensor,
@@ -407,6 +386,20 @@ def flash_attn_varlen_func(
     deterministic: bool = False,
 ) -> torch.Tensor:
     """Only returns the output, not the lse."""
+    if _use_fa2(q, k, v):
+        return _fa2_or_raise(_flash_attn_2_varlen_func)(
+            q,
+            k,
+            v,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            dropout_p=dropout_p,
+            softmax_scale=softmax_scale,
+            causal=causal,
+            deterministic=deterministic,
+        )
     _check_dropout(dropout_p)
     out, _ = torch.ops.fastvideo._flash_attn_cute_varlen_forward(
         q,

fastvideo/attention/utils/flash_attn_default.py

@@ -0,0 +1,251 @@
+# SPDX-License-Identifier: Apache-2.0
+"""torch.compile-traceable wrapper for the FA2/FA3/FA4 default attention path.
+
+The FA4/cute path (`fa_version == "4"`) is already a registered
+`torch.library.custom_op` in `fastvideo.attention.utils.flash_attn_cute`, so
+dynamo treats it as a graph node. The external FA2/FA3 ``flash_attn_func`` is
+NOT — dynamo breaks the graph at the call site (observed: wanvideo.py
+self-attn, once per layer every step), which fragments the compiled region
+and blocks CUDA-graph capture. Wrap the FA2/FA3 default call in a custom op
+(mirrors the FP4 `flash_attn_cute` template) so it becomes an
+opaque-but-traceable node. The kernel still runs eager inside the op
+(correct — flash-attn must run eager); only dynamo's treatment of the
+boundary changes, so numerics are unchanged (SSIM-gated).
+
+Autograd: FA2 has full ``register_autograd`` parity — the custom op's
+backward calls flash_attn's ``_flash_attn_backward`` directly, so training
+backprops *through* the op (no graph break on the training path either).
+FA3 currently keeps the no-backward + carve-out pattern from PR #1373
+because FA3's private backward signature wants validation on a real Hopper
+box (gated on Kuan-Hao's Modal FA3 setup PR). Once that lands the FA3 path
+can mirror FA2.
+
+Lives in `attention/utils/` (sibling of `flash_attn_cute.py` and
+`flash_attn_no_pad.py`) so it can be imported by any backend that wants the
+traceable FA default call without pulling in backend dispatch logic. The
+backend (`attention/backends/flash_attn.py`) just imports
+`flash_attn_func_compilable` and `fa_version` from here.
+"""
+
+import importlib.util
+
+import torch
+
+from fastvideo import envs
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+# Pick the same backend the rest of FastVideo picked for `flash_attn_func`
+# (FA4/cute → FA3 → FA2). Mirror the precedence used in
+# `attention/utils/flash_attn_no_pad.py` so the two probes always agree.
+#
+# FA4 (flash_attn.cute) is explicit opt-in via FASTVIDEO_FA4=1: its CuTeDSL
+# kernels JIT-compile per shape family and can fail at runtime on some
+# arch/shape combinations, so it is never auto-selected just because it is
+# installed.
+if envs.FASTVIDEO_FA4:
+    try:
+        from fastvideo.attention.utils.flash_attn_cute import flash_attn_func
+    except ImportError as e:
+        raise RuntimeError(f"FASTVIDEO_FA4=1 but flash_attn.cute (FA4) is not usable ({e}); "
+                           "fix the FA4 install (see the flash-attn-4 pin in pyproject.toml) "
+                           "or unset FASTVIDEO_FA4.") from e
+    fa_version = "4"
+else:
+    try:
+        from flash_attn_interface import flash_attn_func as flash_attn_3_func
+
+        # flash_attn 3 no longer has a different API, see following commit:
+        # https://github.com/Dao-AILab/flash-attention/commit/ed209409acedbb2379f870bbd03abce31a7a51b7
+        flash_attn_func = flash_attn_3_func
+        fa_version = "3"
+    except ImportError:
+        from flash_attn import flash_attn_func as flash_attn_2_func
+        flash_attn_func = flash_attn_2_func
+        fa_version = "2"
+    try:
+        if importlib.util.find_spec("flash_attn.cute") is not None:
+            logger.info("flash_attn.cute (FA4) is installed but not enabled; "
+                        "set FASTVIDEO_FA4=1 to use it for inference.")
+    except ImportError:
+        pass
+
+if fa_version == "2":
+    # Scope: this op covers exactly the q/k/v + softmax_scale + causal call
+    # shape used by FlashAttentionImpl.forward's default branch (see
+    # `flash_attn_func_compilable(...)` call site in
+    # `attention/backends/flash_attn.py`). The masked/no-pad and varlen /
+    # cross-attn paths use different entry points
+    # (`flash_attn_no_pad`, `flash_attn_varlen_*`) which live in
+    # `attention/utils/flash_attn_no_pad.py`. The wrapper's signature is the
+    # contract: any extra kwarg (dropout_p, window_size, alibi_slopes,
+    # deterministic, return_attn_probs, ...) raises TypeError at the call
+    # site, so silent loss of kwargs is not a failure mode.
+    from flash_attn.flash_attn_interface import _flash_attn_backward as _fa2_backward
+    _fa_default = flash_attn_func
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_default_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_default_forward(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        softmax_scale: float | None,
+        causal: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # `return_attn_probs=True` asks FA2 to also return softmax_lse +
+        # S_dmask. We need softmax_lse to feed the backward; S_dmask is the
+        # dropout mask (always None here since dropout_p is fixed at 0).
+        out, softmax_lse, _ = _fa_default(q, k, v, softmax_scale=softmax_scale, causal=causal, return_attn_probs=True)
+        return out, softmax_lse
+
+    @torch.library.register_fake("fastvideo::_flash_attn_default_forward")
+    def _flash_attn_default_forward_fake(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        softmax_scale: float | None,
+        causal: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        del softmax_scale, causal
+        # FA2 default path: out = [batch, seqlen_q, nheads, head_dim_v],
+        # softmax_lse = [batch, nheads, seqlen_q], fp32 regardless of q dtype.
+        b, sq, hq = q.shape[0], q.shape[1], q.shape[2]
+        out = q.new_empty(b, sq, hq, v.shape[-1])
+        lse = q.new_empty(b, hq, sq, dtype=torch.float32)
+        return out, lse
+
+    def _flash_attn_default_setup_context(ctx, inputs, output):
+        q, k, v, softmax_scale, causal = inputs
+        out, lse = output
+        ctx.save_for_backward(q, k, v, out, lse)
+        # `lse` is an auxiliary output we save to feed FA2's backward; nobody
+        # should differentiate through it. Mark it non-differentiable so
+        # autograd errors loudly if a caller wires it into a loss, rather
+        # than silently producing zero/None grads through the `del grad_lse`
+        # in our backward.
+        ctx.mark_non_differentiable(lse)
+        # FA2's *forward* substitutes `1 / sqrt(head_dim)` for `softmax_scale=None`
+        # internally; FA2's *backward* (`_flash_attn_backward`) demands a concrete
+        # float in its C++ schema and rejects None at the binding boundary. Resolve
+        # the default here so the value saved on ctx (and passed to backward) is
+        # always a real float — matches what FA2's own autograd.Function does.
+        if softmax_scale is None:
+            softmax_scale = q.shape[-1]**-0.5
+        ctx.softmax_scale = softmax_scale
+        ctx.causal = causal
+
+    def _flash_attn_default_backward(ctx, grad_out, grad_lse):
+        # We only differentiate `out`; softmax_lse is saved-for-backward, not
+        # a real differentiable output. (Mirrors the FP4 cute template.)
+        del grad_lse
+        q, k, v, out, lse = ctx.saved_tensors
+        dq = torch.empty_like(q)
+        dk = torch.empty_like(k)
+        dv = torch.empty_like(v)
+        # FA2's `_flash_attn_backward` writes into dq/dk/dv in place. The
+        # extra kwargs (window_size_*, softcap, alibi_slopes, deterministic,
+        # rng_state) are pinned to the same defaults the forward wrapper
+        # uses — flash-attn==2.8.1 (the version FastVideo pins) requires
+        # all of them explicitly. `rng_state=None` is correct for our
+        # `dropout_p=0` configuration.
+        _fa2_backward(
+            grad_out,
+            q,
+            k,
+            v,
+            out,
+            lse,
+            dq,
+            dk,
+            dv,
+            dropout_p=0.0,
+            softmax_scale=ctx.softmax_scale,
+            causal=ctx.causal,
+            window_size_left=-1,
+            window_size_right=-1,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=False,
+            rng_state=None,
+        )
+        return dq, dk, dv, None, None
+
+    torch.library.register_autograd(
+        "fastvideo::_flash_attn_default_forward",
+        _flash_attn_default_backward,
+        setup_context=_flash_attn_default_setup_context,
+    )
+
+    def flash_attn_func_compilable(q, k, v, softmax_scale=None, causal=False):
+        # Backward is registered: autograd flows through the op (training
+        # path is also traceable; no carve-out needed). Public API matches
+        # `flash_attn_func` — returns just `out`; we drop the saved-for-
+        # backward `lse` here so callers see the original single-tensor
+        # contract.
+        out, _ = torch.ops.fastvideo._flash_attn_default_forward(q, k, v, softmax_scale, causal)
+        return out
+elif fa_version == "3":
+    # FA3 path: same forward+fake custom op as the original PR #1373, with
+    # the autograd carve-out kept. The full backward (mirroring the FA2 leg
+    # above) wants a Hopper box for grad-check validation, which we don't
+    # have until Kuan-Hao's Modal FA3 setup PR lands. Until then this keeps
+    # inference traceable + training correct (via the original
+    # autograd.Function path + a pre-PR-style graph break on training).
+    _fa_default = flash_attn_func
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_default_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_default_forward(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        softmax_scale: float | None,
+        causal: bool,
+    ) -> torch.Tensor:
+        return _fa_default(q, k, v, softmax_scale=softmax_scale, causal=causal)
+
+    @torch.library.register_fake("fastvideo::_flash_attn_default_forward")
+    def _flash_attn_default_forward_fake(
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        softmax_scale: float | None,
+        causal: bool,
+    ) -> torch.Tensor:
+        del softmax_scale, causal
+        return q.new_empty(q.shape[0], q.shape[1], q.shape[2], v.shape[-1])
+
+    def flash_attn_func_compilable(q, k, v, softmax_scale=None, causal=False):
+        # Autograd carve-out. The custom op above registers a forward + fake
+        # kernel but NO backward (register_autograd), so it is opaque to
+        # autograd. Inference runs under no_grad / inference_mode and routes
+        # through the traceable custom op — that is the torch.compile win, and
+        # the only path this PR claims. Training backprops through attention,
+        # so route grad-enabled calls to the original FA2/FA3 `flash_attn_func`
+        # (itself an autograd.Function, so backward is correct) at the cost of a
+        # dynamo graph break on the training path — i.e. pre-PR behavior, no
+        # regression. Full autograd parity for the custom op (mirroring the FP4
+        # cute template) is a tracked follow-up.
+        if torch.is_grad_enabled() and (q.requires_grad or k.requires_grad or v.requires_grad):
+            return _fa_default(q, k, v, softmax_scale=softmax_scale, causal=causal)
+        return torch.ops.fastvideo._flash_attn_default_forward(q, k, v, softmax_scale, causal)
+elif fa_version == "4":
+    # FA4 path: `flash_attn_func` is already a torch.library custom op
+    # (registered in `fastvideo.attention.utils.flash_attn_cute`), so a
+    # passthrough is enough — no extra registration needed.
+    def flash_attn_func_compilable(q, k, v, softmax_scale=None, causal=False):
+        return flash_attn_func(q, k, v, softmax_scale=softmax_scale, causal=causal)
+else:
+    # Defensive: the probe above only ever sets fa_version to "2", "3",
+    # or "4"; an unexpected value means an import/probe regression and
+    # we want a loud error at import, not a silent NameError later.
+    raise RuntimeError(f"Unsupported FlashAttention version: {fa_version!r} — expected "
+                       f"'2', '3', or '4' from the import probe above.")

fastvideo/attention/utils/flash_attn_no_pad.py

@@ -21,27 +21,46 @@ from einops import rearrange
 from flash_attn import flash_attn_varlen_qkvpacked_func
 from flash_attn.bert_padding import pad_input, unpad_input
 
+from fastvideo import envs
 
-def _resolve_flash_attn_varlen_func() -> Any:
-    try:
-        from fastvideo.attention.utils.flash_attn_cute import (
-            flash_attn_varlen_func as flash_attn_varlen_func_cute, )
 
-        return flash_attn_varlen_func_cute
-    except ImportError:
+def _resolve_flash_attn_varlen_func() -> tuple[Any, str]:
+    if envs.FASTVIDEO_FA4:
+        # FA4 cute is explicit opt-in (see fastvideo/attention/backends/
+        # flash_attn.py); with FASTVIDEO_FA4=1 an unimportable FA4 build must
+        # fail loudly here rather than fall through to FA3/FA2. RuntimeError,
+        # not ImportError: importers like bsa_attn.py treat ImportError as
+        # "flash-attn not installed" and silently degrade to reference kernels.
         try:
-            from flash_attn_interface import (
-                flash_attn_varlen_func as flash_attn_varlen_func_interface, )
+            from fastvideo.attention.utils.flash_attn_cute import (
+                flash_attn_varlen_func as flash_attn_varlen_func_cute, )
+        except ImportError as e:
+            raise RuntimeError(f"FASTVIDEO_FA4=1 but flash_attn.cute (FA4) is not usable ({e}); "
+                               "fix the FA4 install (see the flash-attn-4 pin in pyproject.toml) "
+                               "or unset FASTVIDEO_FA4.") from e
 
-            return flash_attn_varlen_func_interface
-        except ImportError:
-            from flash_attn import (
-                flash_attn_varlen_func as flash_attn_varlen_func_flash, )
+        return flash_attn_varlen_func_cute, "4"
+    try:
+        from flash_attn_interface import (
+            flash_attn_varlen_func as flash_attn_varlen_func_interface, )
 
-            return flash_attn_varlen_func_flash
+        return flash_attn_varlen_func_interface, "3"
+    except ImportError:
+        from flash_attn import (
+            flash_attn_varlen_func as flash_attn_varlen_func_flash, )
+
+        return flash_attn_varlen_func_flash, "2"
 
 
-flash_attn_varlen_func_impl = _resolve_flash_attn_varlen_func()
+flash_attn_varlen_func_impl, _FA_VARLEN_VERSION = _resolve_flash_attn_varlen_func()
+
+# FA2-only: the private varlen backward we register against the custom ops
+# below. FA3 / FA4 have different private signatures and validation paths
+# (Hopper / Blackwell boxes) — those legs keep the autograd carve-out
+# pattern from PR #1373 until their setup PRs land.
+if _FA_VARLEN_VERSION == "2":
+    from flash_attn.flash_attn_interface import (
+        _flash_attn_varlen_backward as _fa2_varlen_backward, )
 
 
 def flash_attn_no_pad(
@@ -180,3 +199,473 @@ def flash_attn_varlen_qk_no_pad(
         h=nheads,
     )
     return output
+
+
+# ---------------------------------------------------------------------------
+# torch.compile traceability + register_autograd parity for the masked /
+# varlen attention paths.
+#
+# Wraps the two entry points `FlashAttentionImpl.forward` calls
+# (`flash_attn_no_pad`, `flash_attn_varlen_qk_no_pad`) as
+# `torch.library.custom_op`s so dynamo sees one traceable node — the
+# internal unpad / pad bookkeeping (data-dependent `nnz` shapes) runs
+# eager inside the op, and the op's outputs are the statically-shaped
+# padded tensors. This mirrors the FA2 default-path wrapper in
+# `fastvideo/attention/backends/flash_attn.py`.
+#
+# Autograd: on FA2 we register a real backward (`register_autograd`)
+# that calls FA2's `_flash_attn_varlen_backward` on the unpadded form
+# — re-unpadding the saved padded tensors using the saved mask. The
+# `softmax_lse` from the varlen forward is naturally unpadded
+# (`[nheads, total_q]`); we pad it to `[batch, nheads, seqlen]` on
+# the way out (statically shaped) and re-unpad in backward. So
+# training backprops *through* the op (no graph break on the training
+# path either).
+#
+# FA3 / FA4 keep the autograd carve-out pattern from PR #1373: the
+# custom op has forward + fake only, and `*_compilable` falls back to
+# the original autograd.Function for grad-enabled calls. Those legs
+# are gated on Hopper-class / Blackwell-class boxes for backward
+# validation and ship as separate follow-ups.
+
+if _FA_VARLEN_VERSION == "2":
+    # ---------- masked self-attention: flash_attn_no_pad (FA2) ----------
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_no_pad_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_no_pad_forward(
+        qkv: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        b, s, _three, h, d = qkv.shape
+        x = rearrange(qkv, "b s three h d -> b s (three h d)")
+        x_unpad, indices, cu_seqlens, max_s, _ = unpad_input(x, key_padding_mask)
+        x_unpad = rearrange(x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=h)
+        out_unpad, lse_unpad, _ = flash_attn_varlen_qkvpacked_func(x_unpad,
+                                                                   cu_seqlens,
+                                                                   max_s,
+                                                                   dropout_p,
+                                                                   softmax_scale=softmax_scale,
+                                                                   causal=causal,
+                                                                   deterministic=deterministic,
+                                                                   return_attn_probs=True)
+        # Pad out: [nnz, h, d] -> [b, s, h, d]
+        out_padded = rearrange(pad_input(rearrange(out_unpad, "nnz h d -> nnz (h d)"), indices, b, s),
+                               "b s (h d) -> b s h d",
+                               h=h)
+        # Pad lse: FA2 varlen returns [nheads, total_q]. Transpose to [total_q,
+        # nheads], pad to [b, s, nheads], permute to [b, nheads, s] — statically
+        # shaped so register_fake matches.
+        lse_padded = pad_input(lse_unpad.t().contiguous(), indices, b, s).permute(0, 2, 1).contiguous()
+        return out_padded, lse_padded
+
+    @torch.library.register_fake("fastvideo::_flash_attn_no_pad_forward")
+    def _flash_attn_no_pad_forward_fake(
+        qkv: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        del key_padding_mask, causal, dropout_p, softmax_scale, deterministic
+        b, s, _three, h, d = qkv.shape
+        out = qkv.new_empty(b, s, h, d)
+        lse = qkv.new_empty(b, h, s, dtype=torch.float32)
+        return out, lse
+
+    def _flash_attn_no_pad_setup_context(ctx, inputs, output):
+        qkv, key_padding_mask, causal, dropout_p, softmax_scale, deterministic = inputs
+        out, lse = output
+        ctx.save_for_backward(qkv, out, lse, key_padding_mask)
+        # Auxiliary output, not differentiable — see default-path note.
+        ctx.mark_non_differentiable(lse)
+        # FA2's varlen backward requires a concrete float for softmax_scale.
+        if softmax_scale is None:
+            softmax_scale = qkv.shape[-1]**-0.5  # head_dim from qkv's last dim
+        ctx.softmax_scale = softmax_scale
+        ctx.causal = causal
+        ctx.dropout_p = dropout_p
+        ctx.deterministic = deterministic
+
+    def _flash_attn_no_pad_backward(ctx, grad_out, grad_lse):
+        # lse is saved-for-backward, not differentiated.
+        del grad_lse
+        qkv, out_padded, lse_padded, key_padding_mask = ctx.saved_tensors
+        b, s, _three, h, d = qkv.shape
+
+        # One `unpad_input` call (on qkv) gives us indices + cu_seqlens + max_s;
+        # reuse those for out / dout / lse below via direct indexing instead
+        # of redundant `unpad_input` calls (each of which would re-run
+        # `nonzero` + `cumsum` + a `.max().item()` GPU→CPU sync).
+        x = rearrange(qkv, "b s three h d -> b s (three h d)")
+        x_unpad, indices, cu_seqlens, max_s, _ = unpad_input(x, key_padding_mask)
+        x_unpad = rearrange(x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=h)
+        q_unpad, k_unpad, v_unpad = (t.contiguous() for t in x_unpad.unbind(dim=1))
+
+        # Direct-index variants reuse `indices` (computed above).
+        out_unpad = out_padded.flatten(0, 1)[indices].view(-1, h, d).contiguous()
+        dout_unpad = grad_out.flatten(0, 1)[indices].view(-1, h, d).contiguous()
+        # lse_padded [b, h, s] -> [b, s, h] -> [nnz, h] -> [h, nnz].
+        lse_unpad = lse_padded.permute(0, 2, 1).contiguous().flatten(0, 1)[indices].t().contiguous()
+
+        dq_unpad = torch.empty_like(q_unpad)
+        dk_unpad = torch.empty_like(k_unpad)
+        dv_unpad = torch.empty_like(v_unpad)
+        _fa2_varlen_backward(
+            dout_unpad,
+            q_unpad,
+            k_unpad,
+            v_unpad,
+            out_unpad,
+            lse_unpad,
+            dq_unpad,
+            dk_unpad,
+            dv_unpad,
+            cu_seqlens_q=cu_seqlens,
+            cu_seqlens_k=cu_seqlens,
+            max_seqlen_q=max_s,
+            max_seqlen_k=max_s,
+            dropout_p=ctx.dropout_p,
+            softmax_scale=ctx.softmax_scale,
+            causal=ctx.causal,
+            window_size_left=-1,
+            window_size_right=-1,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=ctx.deterministic,
+            rng_state=None,
+        )
+
+        # Re-pad each grad and stack into dqkv.
+        def _repad(dt_unpad: torch.Tensor) -> torch.Tensor:
+            padded = pad_input(rearrange(dt_unpad, "nnz h d -> nnz (h d)"), indices, b, s)
+            return rearrange(padded, "b s (h d) -> b s h d", h=h)
+
+        dqkv = torch.stack([_repad(dq_unpad), _repad(dk_unpad), _repad(dv_unpad)], dim=2)
+        # 6 inputs total: qkv, key_padding_mask, causal, dropout_p, softmax_scale, deterministic.
+        return dqkv, None, None, None, None, None
+
+    torch.library.register_autograd(
+        "fastvideo::_flash_attn_no_pad_forward",
+        _flash_attn_no_pad_backward,
+        setup_context=_flash_attn_no_pad_setup_context,
+    )
+
+    # ---------- cross-attention: flash_attn_varlen_qk_no_pad (FA2) ----------
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_varlen_qk_no_pad_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_varlen_qk_no_pad_forward(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        query_padding_mask: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        b, sq, h, d = query.shape
+        q_unpad, q_indices, cu_seqlens_q, max_seqlen_q, _ = unpad_input(rearrange(query, "b s h d -> b s (h d)"),
+                                                                        query_padding_mask)
+        k_unpad, _, cu_seqlens_k, max_seqlen_k, _ = unpad_input(rearrange(key, "b s h d -> b s (h d)"),
+                                                                key_padding_mask)
+        v_unpad, _, _, _, _ = unpad_input(rearrange(value, "b s h d -> b s (h d)"), key_padding_mask)
+        q_unpad = rearrange(q_unpad, "nnz (h d) -> nnz h d", h=h)
+        k_unpad = rearrange(k_unpad, "nnz (h d) -> nnz h d", h=h)
+        v_unpad = rearrange(v_unpad, "nnz (h d) -> nnz h d", h=h)
+        out_unpad, lse_unpad, _ = flash_attn_varlen_func_impl(q_unpad,
+                                                              k_unpad,
+                                                              v_unpad,
+                                                              cu_seqlens_q,
+                                                              cu_seqlens_k,
+                                                              max_seqlen_q,
+                                                              max_seqlen_k,
+                                                              dropout_p=dropout_p,
+                                                              softmax_scale=softmax_scale,
+                                                              causal=causal,
+                                                              deterministic=deterministic,
+                                                              return_attn_probs=True)
+        # Pad out: [nnz_q, h, d] -> [b, sq, h, d]
+        out_padded = rearrange(pad_input(rearrange(out_unpad, "nnz h d -> nnz (h d)"), q_indices, b, sq),
+                               "b s (h d) -> b s h d",
+                               h=h)
+        # Pad lse: [h, nnz_q] -> [b, h, sq]
+        lse_padded = pad_input(lse_unpad.t().contiguous(), q_indices, b, sq).permute(0, 2, 1).contiguous()
+        return out_padded, lse_padded
+
+    @torch.library.register_fake("fastvideo::_flash_attn_varlen_qk_no_pad_forward")
+    def _flash_attn_varlen_qk_no_pad_forward_fake(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        query_padding_mask: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        del key, query_padding_mask, key_padding_mask
+        del causal, dropout_p, softmax_scale, deterministic
+        b, sq, h, _ = query.shape
+        # `out`'s head_dim comes from value (d_v), matching the real forward's
+        # out_padded ([b, sq, h, d_v]); it can differ from query's d_q.
+        out = query.new_empty(b, sq, h, value.shape[-1])
+        lse = query.new_empty(b, h, sq, dtype=torch.float32)
+        return out, lse
+
+    def _flash_attn_varlen_qk_no_pad_setup_context(ctx, inputs, output):
+        (query, key, value, query_padding_mask, key_padding_mask, causal, dropout_p, softmax_scale,
+         deterministic) = inputs
+        out, lse = output
+        ctx.save_for_backward(query, key, value, out, lse, query_padding_mask, key_padding_mask)
+        # Auxiliary output, not differentiable — see default-path note.
+        ctx.mark_non_differentiable(lse)
+        if softmax_scale is None:
+            softmax_scale = query.shape[-1]**-0.5
+        ctx.softmax_scale = softmax_scale
+        ctx.causal = causal
+        ctx.dropout_p = dropout_p
+        ctx.deterministic = deterministic
+
+    def _flash_attn_varlen_qk_no_pad_backward(ctx, grad_out, grad_lse):
+        del grad_lse
+        (query, key, value, out_padded, lse_padded, query_padding_mask, key_padding_mask) = ctx.saved_tensors
+        b, sq, h, d = query.shape
+        sk = key.shape[1]
+
+        # One `unpad_input` call per distinct mask; reuse the returned
+        # indices via direct indexing for everything else that shares
+        # the same mask (v with k_mask; out/dout/lse with q_mask; the
+        # final repad of dk/dv also reuses k_indices). Avoids ~4
+        # redundant `unpad_input` calls + their GPU→CPU `.max().item()`
+        # syncs.
+        q_unpad, q_indices, cu_seqlens_q, max_seqlen_q, _ = unpad_input(rearrange(query, "b s h d -> b s (h d)"),
+                                                                        query_padding_mask)
+        k_unpad, k_indices, cu_seqlens_k, max_seqlen_k, _ = unpad_input(rearrange(key, "b s h d -> b s (h d)"),
+                                                                        key_padding_mask)
+        q_unpad = rearrange(q_unpad, "nnz (h d) -> nnz h d", h=h).contiguous()
+        k_unpad = rearrange(k_unpad, "nnz (h d) -> nnz h d", h=h).contiguous()
+        v_unpad = value.flatten(0, 1)[k_indices].view(-1, h, d).contiguous()
+
+        # out / dout / lse follow q's shape, so index with q_indices.
+        out_unpad = out_padded.flatten(0, 1)[q_indices].view(-1, h, d).contiguous()
+        dout_unpad = grad_out.flatten(0, 1)[q_indices].view(-1, h, d).contiguous()
+        lse_unpad = lse_padded.permute(0, 2, 1).contiguous().flatten(0, 1)[q_indices].t().contiguous()
+
+        dq_unpad = torch.empty_like(q_unpad)
+        dk_unpad = torch.empty_like(k_unpad)
+        dv_unpad = torch.empty_like(v_unpad)
+        _fa2_varlen_backward(
+            dout_unpad,
+            q_unpad,
+            k_unpad,
+            v_unpad,
+            out_unpad,
+            lse_unpad,
+            dq_unpad,
+            dk_unpad,
+            dv_unpad,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_q,
+            max_seqlen_k=max_seqlen_k,
+            dropout_p=ctx.dropout_p,
+            softmax_scale=ctx.softmax_scale,
+            causal=ctx.causal,
+            window_size_left=-1,
+            window_size_right=-1,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=ctx.deterministic,
+            rng_state=None,
+        )
+
+        # k_indices is already available from the unpad_input above —
+        # no need to recompute it for the dk/dv repad.
+        def _repad(dt_unpad: torch.Tensor, indices: torch.Tensor, batch: int, seqlen: int) -> torch.Tensor:
+            padded = pad_input(rearrange(dt_unpad, "nnz h d -> nnz (h d)"), indices, batch, seqlen)
+            return rearrange(padded, "b s (h d) -> b s h d", h=h)
+
+        dq_padded = _repad(dq_unpad, q_indices, b, sq)
+        dk_padded = _repad(dk_unpad, k_indices, b, sk)
+        dv_padded = _repad(dv_unpad, k_indices, b, sk)
+        # 9 inputs total: query, key, value, q_mask, k_mask, causal, dropout_p,
+        # softmax_scale, deterministic.
+        return dq_padded, dk_padded, dv_padded, None, None, None, None, None, None
+
+    torch.library.register_autograd(
+        "fastvideo::_flash_attn_varlen_qk_no_pad_forward",
+        _flash_attn_varlen_qk_no_pad_backward,
+        setup_context=_flash_attn_varlen_qk_no_pad_setup_context,
+    )
+
+    # ---------- public dispatchers (FA2: autograd flows through the op) -----
+
+
+    def flash_attn_no_pad_compilable(qkv,
+                                     key_padding_mask,
+                                     causal=False,
+                                     dropout_p=0.0,
+                                     softmax_scale=None,
+                                     deterministic=False):
+        """dynamo-traceable wrapper around ``flash_attn_no_pad`` (registered op,
+        full register_autograd on FA2 — both inference and training go through
+        the op, no graph break on either)."""
+        out, _ = torch.ops.fastvideo._flash_attn_no_pad_forward(qkv, key_padding_mask, causal, dropout_p, softmax_scale,
+                                                                deterministic)
+        return out
+
+    def flash_attn_varlen_qk_no_pad_compilable(query,
+                                               key,
+                                               value,
+                                               query_padding_mask,
+                                               key_padding_mask,
+                                               causal=False,
+                                               dropout_p=0.0,
+                                               softmax_scale=None,
+                                               deterministic=False):
+        """dynamo-traceable wrapper around ``flash_attn_varlen_qk_no_pad`` (registered
+        op, full register_autograd on FA2)."""
+        out, _ = torch.ops.fastvideo._flash_attn_varlen_qk_no_pad_forward(query, key, value, query_padding_mask,
+                                                                          key_padding_mask, causal, dropout_p,
+                                                                          softmax_scale, deterministic)
+        return out
+
+else:
+    # ---------- FA3 / FA4: carve-out (forward+fake only, no real backward) ---
+    # Same pattern as the parked varlen-extension and the FA3 default leg in
+    # `fastvideo/attention/backends/flash_attn.py`. Real backward for these
+    # versions is a follow-up gated on Hopper / Blackwell box validation.
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_no_pad_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_no_pad_forward(
+        qkv: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> torch.Tensor:
+        return flash_attn_no_pad(  # type: ignore[no-untyped-call]
+            qkv,
+            key_padding_mask,
+            causal=causal,
+            dropout_p=dropout_p,
+            softmax_scale=softmax_scale,
+            deterministic=deterministic)
+
+    @torch.library.register_fake("fastvideo::_flash_attn_no_pad_forward")
+    def _flash_attn_no_pad_forward_fake(
+        qkv: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> torch.Tensor:
+        del key_padding_mask, causal, dropout_p, softmax_scale, deterministic
+        b, s, _three, h, d = qkv.shape
+        return qkv.new_empty(b, s, h, d)
+
+    @torch.library.custom_op(
+        "fastvideo::_flash_attn_varlen_qk_no_pad_forward",
+        mutates_args=(),
+        device_types="cuda",
+    )
+    def _flash_attn_varlen_qk_no_pad_forward(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        query_padding_mask: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> torch.Tensor:
+        return flash_attn_varlen_qk_no_pad(  # type: ignore[no-untyped-call]
+            query,
+            key,
+            value,
+            query_padding_mask=query_padding_mask,
+            key_padding_mask=key_padding_mask,
+            causal=causal,
+            dropout_p=dropout_p,
+            softmax_scale=softmax_scale,
+            deterministic=deterministic)
+
+    @torch.library.register_fake("fastvideo::_flash_attn_varlen_qk_no_pad_forward")
+    def _flash_attn_varlen_qk_no_pad_forward_fake(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        query_padding_mask: torch.Tensor,
+        key_padding_mask: torch.Tensor,
+        causal: bool,
+        dropout_p: float,
+        softmax_scale: float | None,
+        deterministic: bool,
+    ) -> torch.Tensor:
+        del key, query_padding_mask, key_padding_mask
+        del causal, dropout_p, softmax_scale, deterministic
+        b, sq, h, _ = query.shape
+        # `out`'s head_dim comes from value (d_v), matching the real forward's
+        # output ([b, sq, h, d_v]); it can differ from query's d_q.
+        return query.new_empty(b, sq, h, value.shape[-1])
+
+    def flash_attn_no_pad_compilable(qkv,
+                                     key_padding_mask,
+                                     causal=False,
+                                     dropout_p=0.0,
+                                     softmax_scale=None,
+                                     deterministic=False):
+        if torch.is_grad_enabled() and qkv.requires_grad:
+            return flash_attn_no_pad(qkv,
+                                     key_padding_mask,
+                                     causal=causal,
+                                     dropout_p=dropout_p,
+                                     softmax_scale=softmax_scale,
+                                     deterministic=deterministic)
+        return torch.ops.fastvideo._flash_attn_no_pad_forward(qkv, key_padding_mask, causal, dropout_p, softmax_scale,
+                                                              deterministic)
+
+    def flash_attn_varlen_qk_no_pad_compilable(query,
+                                               key,
+                                               value,
+                                               query_padding_mask,
+                                               key_padding_mask,
+                                               causal=False,
+                                               dropout_p=0.0,
+                                               softmax_scale=None,
+                                               deterministic=False):
+        if torch.is_grad_enabled() and (query.requires_grad or key.requires_grad or value.requires_grad):
+            return flash_attn_varlen_qk_no_pad(query,
+                                               key,
+                                               value,
+                                               query_padding_mask=query_padding_mask,
+                                               key_padding_mask=key_padding_mask,
+                                               causal=causal,
+                                               dropout_p=dropout_p,
+                                               softmax_scale=softmax_scale,
+                                               deterministic=deterministic)
+        return torch.ops.fastvideo._flash_attn_varlen_qk_no_pad_forward(query, key, value, query_padding_mask,
+                                                                        key_padding_mask, causal, dropout_p,
+                                                                        softmax_scale, deterministic)

fastvideo/configs/models/dits/__init__.py

@@ -1,6 +1,9 @@
 from fastvideo.configs.models.dits.cosmos import CosmosVideoConfig
 from fastvideo.configs.models.dits.cosmos2_5 import Cosmos25VideoConfig
+from fastvideo.configs.models.dits.dreamx_world import DreamXWorldARConfig, DreamXWorldConfig
+from fastvideo.configs.models.dits.flux import FluxDiTConfig
 from fastvideo.configs.models.dits.flux_2 import Flux2Config
+from fastvideo.configs.models.dits.glm_image import GlmImageDiTConfig
 from fastvideo.configs.models.dits.hunyuangamecraft import HunyuanGameCraftConfig
 from fastvideo.configs.models.dits.hunyuanvideo import HunyuanVideoConfig
 from fastvideo.configs.models.dits.hunyuanvideo15 import HunyuanVideo15Config
@@ -13,7 +16,8 @@ from fastvideo.configs.models.dits.hyworld import HYWorldConfig
 from fastvideo.configs.models.dits.kandinsky5 import Kandinsky5VideoConfig
 
 __all__ = [
-    "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "CosmosVideoConfig",
-    "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig",
-    "MagiHumanVideoConfig", "StableAudioConfig", "Flux2Config"
+    "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "DreamXWorldConfig",
+    "DreamXWorldARConfig", "CosmosVideoConfig", "Cosmos25VideoConfig", "FluxDiTConfig", "Flux2Config",
+    "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig", "MagiHumanVideoConfig",
+    "StableAudioConfig", "GlmImageDiTConfig"
 ]

fastvideo/configs/models/dits/dreamx_world.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+from fastvideo.configs.models.dits.wanvideo import WanVideoArchConfig
+
+
+@dataclass
+class DreamXWorldArchConfig(WanVideoArchConfig):
+    """DreamX-World DiT config with camera PRoPE control fields."""
+
+    add_control_adapter: bool = True
+    cam_method: str | None = "prope"
+    attn_compress: int = 1
+    cam_self_attn_layers: tuple[int, ...] | None = None
+
+
+@dataclass
+class DreamXWorldConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=DreamXWorldArchConfig)
+
+    prefix: str = "Wan"
+
+
+@dataclass
+class DreamXWorldARArchConfig(DreamXWorldArchConfig):
+    """DreamX-World-5B autoregressive causal DiT config."""
+
+    model_type: str = "ti2v"
+    patch_size: tuple[int, int, int] = (1, 2, 2)
+    text_len: int = 512
+    text_dim: int = 4096
+    freq_dim: int = 256
+    attn_compress: int = 4
+    cam_self_attn_layers: tuple[int, ...] | None = tuple(range(30))
+    local_attn_size: int = 12
+    sink_size: int = 3
+    num_frames_per_block: int = 3
+    rope_cache_policy: str = "block_relativistic"
+    # The official AR checkpoint (AMAP-ML/DreamX-World ``model.safetensors``)
+    # already uses FastVideo's native key names and the converter copies the
+    # tensors verbatim, so every rule is an identity. The rules enumerate the
+    # full state-dict surface of ``DreamXWorldARTransformer3DModel`` (norm1 /
+    # norm2 / head.norm are affine-free and have no parameters).
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^patch_embedding\.(.*)$": r"patch_embedding.\1",
+            r"^text_embedding\.([02])\.(.*)$": r"text_embedding.\1.\2",
+            r"^time_embedding\.([02])\.(.*)$": r"time_embedding.\1.\2",
+            r"^time_projection\.1\.(.*)$": r"time_projection.1.\1",
+            r"^blocks\.(\d+)\.self_attn\.(q|k|v|o)\.(.*)$": r"blocks.\1.self_attn.\2.\3",
+            r"^blocks\.(\d+)\.self_attn\.norm_(q|k)\.weight$": r"blocks.\1.self_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.cross_attn\.(q|k|v|o)\.(.*)$": r"blocks.\1.cross_attn.\2.\3",
+            r"^blocks\.(\d+)\.cross_attn\.norm_(q|k)\.weight$": r"blocks.\1.cross_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.cam_self_attn\.(q_proj|k_proj|v_proj|out_proj)\.(.*)$": r"blocks.\1.cam_self_attn.\2.\3",
+            r"^blocks\.(\d+)\.cam_self_attn\.norm_(q|k)\.weight$": r"blocks.\1.cam_self_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.norm3\.(.*)$": r"blocks.\1.norm3.\2",
+            r"^blocks\.(\d+)\.ffn\.([02])\.(.*)$": r"blocks.\1.ffn.\2.\3",
+            r"^blocks\.(\d+)\.modulation$": r"blocks.\1.modulation",
+            r"^head\.head\.(.*)$": r"head.head.\1",
+            r"^head\.modulation$": r"head.modulation",
+        })
+    reverse_param_names_mapping: dict = field(default_factory=dict)
+    lora_param_names_mapping: dict = field(default_factory=dict)
+
+
+@dataclass
+class DreamXWorldARConfig(DreamXWorldConfig):
+    arch_config: DiTArchConfig = field(default_factory=DreamXWorldARArchConfig)
+
+    prefix: str = "Wan"

fastvideo/configs/models/dits/flux.py

@@ -0,0 +1,27 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+
+
+@dataclass
+class FluxTransformer2DArchConfig(DiTArchConfig):
+
+    patch_size: int = 1
+    in_channels: int = 64
+    out_channels: int | None = None
+    num_layers: int = 19
+    num_single_layers: int = 38
+    attention_head_dim: int = 128
+    num_attention_heads: int = 24
+    joint_attention_dim: int = 4096
+    pooled_projection_dim: int = 768
+    guidance_embeds: bool = True
+    axes_dims_rope: tuple[int, int, int] = (16, 56, 56)
+
+
+@dataclass
+class FluxDiTConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=FluxTransformer2DArchConfig)
+    prefix: str = "flux"

fastvideo/configs/models/dits/glm_image.py

@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+
+
+def is_blocks(n: str, m) -> bool:
+    return "transformer_blocks" in n and str.isdigit(n.split(".")[-1])
+
+
+@dataclass
+class GlmImageDiTArchConfig(DiTArchConfig):
+
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [is_blocks])
+
+    hidden_size: int = 4096
+    num_attention_heads: int = 32
+    attention_head_dim: int = 128
+    in_channels: int = 16
+    out_channels: int = 16
+    num_layers: int = 30
+
+    text_embed_dim: int = 1472
+    time_embed_dim: int = 512
+    condition_dim: int = 256
+
+    prior_vq_quantizer_codebook_size: int = 16384
+
+    patch_size: int = 2
+
+    max_height: int = 2048
+    max_width: int = 2048
+
+    qk_norm: str = "layer_norm"
+    eps: float = 1e-5
+
+    exclude_lora_layers: list[str] = field(
+        default_factory=lambda: ["image_projector", "glyph_projector", "prior_token_embedding"])
+
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^glyph_projector\.net\.0\.proj\.(.*)$": r"glyph_projector.fc_in.\1",
+            r"^glyph_projector\.net\.2\.(.*)$": r"glyph_projector.fc_out.\1",
+            r"^prior_projector\.net\.0\.proj\.(.*)$": r"prior_projector.fc_in.\1",
+            r"^prior_projector\.net\.2\.(.*)$": r"prior_projector.fc_out.\1",
+            r"^transformer_blocks\.(\d+)\.ff\.net\.0\.proj\.(.*)$": r"transformer_blocks.\1.ff.fc_in.\2",
+            r"^transformer_blocks\.(\d+)\.ff\.net\.2\.(.*)$": r"transformer_blocks.\1.ff.fc_out.\2",
+        })
+
+    reverse_param_names_mapping: dict = field(default_factory=dict)
+    lora_param_names_mapping: dict = field(default_factory=dict)
+
+    def __post_init__(self):
+        super().__post_init__()
+        self.num_channels_latents = self.out_channels
+
+
+@dataclass
+class GlmImageDiTConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=GlmImageDiTArchConfig)
+    prefix: str = "GlmImage"

fastvideo/configs/models/dits/kandinsky5.py

@@ -2,14 +2,24 @@
 from dataclasses import dataclass, field
 
 from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+from fastvideo.platforms import AttentionBackendEnum
+
+
+def _is_kandinsky5_transformer_block(n: str, m) -> bool:
+    return ("text_transformer_blocks" in n or "visual_transformer_blocks" in n) and n.split(".")[-1].isdigit()
 
 
 @dataclass
 class Kandinsky5ArchConfig(DiTArchConfig):
-    _fsdp_shard_conditions: list = field(default_factory=lambda: [
-        lambda n, m:
-        ("text_transformer_blocks" in n or "visual_transformer_blocks" in n) and n.split(".")[-1].isdigit()
-    ])
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [_is_kandinsky5_transformer_block])
+
+    # NABLA block-sparse attention for attention_type="nabla" checkpoints, plus
+    # the dense backends every DiT supports.
+    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.NABLA_ATTN,
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+    )
 
     # Native FastVideo implementation uses the same parameter names as diffusers
     # except FFN internals: Diffusers FFN uses `in_layer/out_layer`, while

fastvideo/configs/models/dits/wanvideo.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
+from typing import Literal
 
 from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
 
@@ -19,6 +20,11 @@ class WanVideoArchConfig(DiTArchConfig):
             r"^condition_embedder\.text_embedder\.linear_2\.(.*)$": r"condition_embedder.text_embedder.fc_out.\1",
             r"^condition_embedder\.time_embedder\.linear_1\.(.*)$": r"condition_embedder.time_embedder.mlp.fc_in.\1",
             r"^condition_embedder\.time_embedder\.linear_2\.(.*)$": r"condition_embedder.time_embedder.mlp.fc_out.\1",
+            # AnyFlow dual-timestep checkpoints expose delta_embedder weights with the
+            # same internal layout as time_embedder. The regex is harmless on plain
+            # Wan checkpoints (no delta_embedder keys to match).
+            r"^condition_embedder\.delta_embedder\.linear_1\.(.*)$": r"condition_embedder.delta_embedder.mlp.fc_in.\1",
+            r"^condition_embedder\.delta_embedder\.linear_2\.(.*)$": r"condition_embedder.delta_embedder.mlp.fc_out.\1",
             r"^condition_embedder\.time_proj\.(.*)$": r"condition_embedder.time_modulation.linear.\1",
             r"^condition_embedder\.image_embedder\.ff\.net\.0\.proj\.(.*)$":
             r"condition_embedder.image_embedder.ff.fc_in.\1",
@@ -86,6 +92,14 @@ class WanVideoArchConfig(DiTArchConfig):
     # "relativistic" keeps long rollouts in-distribution; a no-op unless sink_size > 0 and local_attn_size > 0.
     rope_cache_policy: str = "absolute"
 
+    # AnyFlow dual-timestep conditioning. Defaults preserve bit-identity with
+    # the legacy single-timestep forward (no delta_embedder allocated, no
+    # extra computation on the embedder forward path).
+    r_embedder: bool = False
+    r_embedder_fusion: Literal["additive", "gated"] = "additive"
+    r_embedder_gate_value: float = 0.25
+    r_embedder_deltatime_type: Literal["r", "t-r"] = "r"
+
     def __post_init__(self):
         super().__post_init__()
         self.out_channels = self.out_channels or self.in_channels

fastvideo/configs/models/encoders/base.py

@@ -43,11 +43,16 @@ class TextEncoderArchConfig(EncoderArchConfig):
     require_processor: bool = False
 
     def __post_init__(self) -> None:
-        self.tokenizer_kwargs = {
+        # update_model_arch re-runs __post_init__ after pipeline configs may
+        # have customized tokenizer_kwargs (e.g. kandinsky5/gen3c/longcat set
+        # "padding"); rebuilding the dict here would silently wipe those
+        # customizations, so only fill in defaults for keys not already set.
+        defaults = {
             "truncation": True,
             "max_length": self.text_len,
             "return_tensors": "pt",
         }
+        self.tokenizer_kwargs = defaults | self.tokenizer_kwargs
 
 
 @dataclass

fastvideo/configs/models/vaes/__init__.py

@@ -2,6 +2,7 @@ from fastvideo.configs.models.vaes.cosmosvae import CosmosVAEConfig
 from fastvideo.configs.models.vaes.cosmos2_5vae import Cosmos25VAEConfig
 from fastvideo.configs.models.vaes.gamecraftvae import GameCraftVAEConfig
 from fastvideo.configs.models.vaes.gen3cvae import Gen3CVAEConfig
+from fastvideo.configs.models.vaes.glm_image import GlmImageVAEConfig
 from fastvideo.configs.models.vaes.hunyuanvae import HunyuanVAEConfig
 from fastvideo.configs.models.vaes.hunyuan15vae import Hunyuan15VAEConfig
 from fastvideo.configs.models.vaes.ltx2vae import LTX2VAEConfig
@@ -21,4 +22,5 @@ __all__ = [
     "OobleckVAEArchConfig",
     "OobleckVAEConfig",
     "Flux2VAEConfig",
+    "GlmImageVAEConfig",
 ]

fastvideo/configs/models/vaes/glm_image.py

@@ -0,0 +1,95 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.vaes.autoencoder_kl import (AutoencoderKLArchConfig, AutoencoderKLVAEConfig)
+
+_GLM_IMAGE_LATENTS_MEAN: tuple[float, ...] = (
+    -0.2080078125,
+    1.875,
+    -0.470703125,
+    -1.265625,
+    -1.421875,
+    0.77734375,
+    -0.3671875,
+    -0.9453125,
+    0.318359375,
+    0.7734375,
+    -0.1884765625,
+    -0.022216796875,
+    -0.220703125,
+    -1.59375,
+    -0.81640625,
+    -0.255859375,
+)
+_GLM_IMAGE_LATENTS_STD: tuple[float, ...] = (
+    3.0625,
+    2.203125,
+    2.265625,
+    4.84375,
+    2.5,
+    3.9375,
+    2.203125,
+    3.03125,
+    2.1875,
+    2.046875,
+    2.71875,
+    2.390625,
+    2.390625,
+    2.453125,
+    2.25,
+    2.15625,
+)
+
+
+@dataclass
+class GlmImageVAEArchConfig(AutoencoderKLArchConfig):
+    act_fn: str = "silu"
+    block_out_channels: tuple[int, ...] = (128, 512, 1024, 1024)
+    down_block_types: tuple[str, ...] = (
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+    )
+    up_block_types: tuple[str, ...] = (
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+    )
+    force_upcast: bool = True
+    in_channels: int = 3
+    out_channels: int = 3
+    latent_channels: int = 16
+    latents_mean: tuple[float, ...] = _GLM_IMAGE_LATENTS_MEAN
+    latents_std: tuple[float, ...] = _GLM_IMAGE_LATENTS_STD
+    layers_per_block: int = 3
+    mid_block_add_attention: bool = False
+    norm_num_groups: int = 32
+    sample_size: int = 1024
+    scaling_factor: float = 0.18215
+    shift_factor: float | None = None
+    use_quant_conv: bool = False
+    use_post_quant_conv: bool = False
+
+    temporal_compression_ratio: int = 1
+    spatial_compression_ratio: int = 8
+
+
+@dataclass
+class GlmImageVAEConfig(AutoencoderKLVAEConfig):
+    arch_config: GlmImageVAEArchConfig = field(default_factory=GlmImageVAEArchConfig)
+
+    use_tiling: bool = True
+    use_temporal_tiling: bool = False
+    use_parallel_tiling: bool = False
+
+    tile_sample_min_height: int = 512
+    tile_sample_min_width: int = 512
+    tile_sample_stride_height: int = 384
+    tile_sample_stride_width: int = 384
+
+    load_encoder: bool = True
+    load_decoder: bool = True

fastvideo/configs/pipelines/__init__.py

@@ -1,10 +1,12 @@
 from fastvideo.configs.pipelines.base import PipelineConfig
 from fastvideo.configs.pipelines.cosmos import CosmosConfig
 from fastvideo.configs.pipelines.cosmos2_5 import Cosmos25Config
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BARPipelineConfig, DreamXWorld5BCamPipelineConfig
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.hunyuangamecraft import HunyuanGameCraftPipelineConfig
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
+from fastvideo.configs.pipelines.kandinsky5 import Kandinsky5I2VConfig, Kandinsky5T2VConfig
 from fastvideo.configs.pipelines.matrixgame2 import MatrixGame2I2V480PConfig
 from fastvideo.configs.pipelines.matrixgame3 import MatrixGame3I2V720PConfig
 from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
@@ -16,5 +18,6 @@ __all__ = [
     "HunyuanConfig", "FastHunyuanConfig", "HunyuanGameCraftPipelineConfig", "PipelineConfig", "Hunyuan15T2V480PConfig",
     "Hunyuan15T2V720PConfig", "WanT2V480PConfig", "WanI2V480PConfig", "WanT2V720PConfig", "WanI2V720PConfig",
     "SelfForcingWanT2V480PConfig", "LucyEditDevConfig", "CosmosConfig", "Cosmos25Config", "LTX2T2VConfig",
-    "HYWorldConfig", "MatrixGame2I2V480PConfig", "MatrixGame3I2V720PConfig", "get_pipeline_config_cls_from_name"
+    "DreamXWorld5BCamPipelineConfig", "DreamXWorld5BARPipelineConfig", "HYWorldConfig", "Kandinsky5T2VConfig",
+    "Kandinsky5I2VConfig", "MatrixGame2I2V480PConfig", "MatrixGame3I2V720PConfig", "get_pipeline_config_cls_from_name"
 ]

fastvideo/configs/pipelines/dreamx_world.py

@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World-5B-Cam FastVideo model configuration helpers."""
+
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits.dreamx_world import (DreamXWorldARArchConfig, DreamXWorldARConfig,
+                                                        DreamXWorldArchConfig, DreamXWorldConfig)
+from fastvideo.configs.models.encoders import BaseEncoderOutput, T5Config
+from fastvideo.configs.models.encoders.t5 import T5ArchConfig
+from fastvideo.configs.models.vaes import WanVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.configs.pipelines.wan import LucyEditDevConfig, t5_postprocess_text
+
+
+def make_dreamx_world_5b_cam_dit_config() -> DreamXWorldConfig:
+    """Return the DreamX-World DiT config matching DreamX-World-5B-Cam."""
+    return DreamXWorldConfig(arch_config=DreamXWorldArchConfig(
+        num_attention_heads=24,
+        attention_head_dim=128,
+        in_channels=48,
+        out_channels=48,
+        ffn_dim=14336,
+        num_layers=30,
+        cross_attn_norm=True,
+        qk_norm="rms_norm_across_heads",
+        add_control_adapter=True,
+        cam_method="prope",
+        attn_compress=1,
+        cam_self_attn_layers=None,
+    ))
+
+
+def make_dreamx_world_5b_ar_dit_config() -> DreamXWorldARConfig:
+    """Return the DreamX-World-5B autoregressive causal DiT config."""
+    return DreamXWorldARConfig(arch_config=DreamXWorldARArchConfig(
+        model_type="ti2v",
+        num_attention_heads=24,
+        attention_head_dim=128,
+        in_channels=48,
+        out_channels=48,
+        ffn_dim=14336,
+        num_layers=30,
+        cross_attn_norm=True,
+        qk_norm=True,
+        add_control_adapter=True,
+        cam_method="prope",
+        attn_compress=4,
+        cam_self_attn_layers=tuple(range(30)),
+        local_attn_size=12,
+        sink_size=3,
+        num_frames_per_block=3,
+    ))
+
+
+def make_dreamx_world_5b_cam_vae_config() -> WanVAEConfig:
+    """Return the Wan2.2 48-channel VAE config used by DreamX-World-5B-Cam."""
+    return LucyEditDevConfig().vae_config
+
+
+def make_dreamx_world_5b_cam_text_encoder_config() -> T5Config:
+    """Return the UMT5-XXL text encoder config used by DreamX-World-5B-Cam."""
+    return T5Config(
+        arch_config=T5ArchConfig(
+            vocab_size=256384,
+            d_model=4096,
+            d_kv=64,
+            d_ff=10240,
+            num_layers=24,
+            num_decoder_layers=None,
+            num_heads=64,
+            relative_attention_num_buckets=32,
+            dropout_rate=0.0,
+            text_len=512,
+            feed_forward_proj="gelu",
+            is_encoder_decoder=False,
+        ),
+        prefix="umt5",
+    )
+
+
+@dataclass
+class DreamXWorld5BCamPipelineConfig(PipelineConfig):
+    """Pipeline config for the first-scope DreamX-World-5B-Cam mode."""
+
+    dit_config: DiTConfig = field(default_factory=make_dreamx_world_5b_cam_dit_config)
+    vae_config: VAEConfig = field(default_factory=make_dreamx_world_5b_cam_vae_config)
+    text_encoder_configs: tuple[EncoderConfig,
+                                ...] = field(default_factory=lambda: (make_dreamx_world_5b_cam_text_encoder_config(), ))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
+                                  ...] = field(default_factory=lambda: (t5_postprocess_text, ))
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", ))
+    flow_shift: float | None = 3.0
+    ti2v_task: bool = True
+    expand_timesteps: bool = True
+    vae_tiling: bool = False
+    vae_sp: bool = False
+    vae_precision: str = "fp32"
+    vae_decode_precision: str | None = "bf16"
+    dit_precision: str = "bf16"
+
+    def __post_init__(self) -> None:
+        self.vae_config.load_encoder = True
+        self.vae_config.load_decoder = True
+        self.dit_config.expand_timesteps = self.expand_timesteps
+
+
+@dataclass
+class DreamXWorld5BARPipelineConfig(DreamXWorld5BCamPipelineConfig):
+    """Pipeline config for DreamX-World-5B autoregressive forcing."""
+
+    dit_config: DiTConfig = field(default_factory=make_dreamx_world_5b_ar_dit_config)
+    flow_shift: float | None = 5.0
+    ti2v_task: bool = True
+    is_causal: bool = True
+    dmd_denoising_steps: tuple[int, ...] = (1000, 750, 500, 250)
+    warp_denoising_step: bool = True
+    context_noise: float = 0.1
+    num_frames_per_block: int = 3
+    color_correction_strength: float = 1.0
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.dit_config.expand_timesteps = True

fastvideo/configs/pipelines/flux.py

@@ -0,0 +1,74 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import EncoderConfig
+from fastvideo.configs.models.dits.flux import FluxDiTConfig
+from fastvideo.configs.models.encoders import (
+    BaseEncoderOutput,
+    CLIPTextConfig,
+    T5LargeConfig,
+)
+from fastvideo.configs.models.vaes.autoencoder_kl import AutoencoderKLVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, preprocess_text
+
+
+def _flux_clip_pooled_postprocess(outputs: BaseEncoderOutput) -> torch.Tensor:
+    """CLIP branch for FLUX: Diffusers uses pooled prompt embeddings only."""
+    if outputs.pooler_output is None:
+        raise RuntimeError(
+            "FLUX CLIP conditioning requires pooler_output. Ensure the CLIP text encoder returns pooled features.")
+    return outputs.pooler_output
+
+
+def _flux_t5_sequence_postprocess(outputs: BaseEncoderOutput) -> torch.Tensor:
+    if outputs.last_hidden_state is None:
+        raise RuntimeError("FLUX T5 conditioning requires last_hidden_state.")
+    return outputs.last_hidden_state
+
+
+@dataclass
+class FluxPipelineConfig(PipelineConfig):
+    """Pipeline layout for Diffusers FLUX.1-dev (CLIP + T5 + packed DiT + FlowMatch)."""
+
+    scheduler_arch: str = "FlowMatchEulerDiscreteScheduler"
+    transformer_arch: str = "FluxTransformer2DModel"
+    vae_arch: str = "AutoencoderKL"
+    text_encoder_archs: tuple[str, ...] = ("CLIPTextModel", "T5EncoderModel")
+    tokenizer_archs: tuple[str, ...] = ("CLIPTokenizer", "T5TokenizerFast")
+
+    dit_config: FluxDiTConfig = field(default_factory=FluxDiTConfig)
+    vae_config: AutoencoderKLVAEConfig = field(default_factory=AutoencoderKLVAEConfig)
+
+    embedded_cfg_scale: float = 3.5
+    flow_shift: float | None = None
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(default_factory=lambda: (CLIPTextConfig(), T5LargeConfig()))
+    preprocess_text_funcs: tuple[Callable[[str], str],
+                                 ...] = field(default_factory=lambda: (preprocess_text, preprocess_text))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor], ...] = field(
+        default_factory=lambda: (_flux_clip_pooled_postprocess, _flux_t5_sequence_postprocess))
+
+    dit_precision: str = "bf16"
+    vae_precision: str = "fp32"
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("fp32", "bf16"))
+
+    def __post_init__(self) -> None:
+        te_cfgs = list(self.text_encoder_configs)
+        if len(te_cfgs) >= 1:
+            te_cfgs[0].tokenizer_kwargs.setdefault("padding", "max_length")
+            te_cfgs[0].tokenizer_kwargs.setdefault("max_length", 77)
+            te_cfgs[0].tokenizer_kwargs.setdefault("truncation", True)
+            te_cfgs[0].tokenizer_kwargs.setdefault("return_tensors", "pt")
+        if len(te_cfgs) >= 2:
+            cap = 512
+            te_cfgs[1].tokenizer_kwargs["max_length"] = min(int(te_cfgs[1].tokenizer_kwargs.get("max_length", cap)),
+                                                            cap)
+            te_cfgs[1].tokenizer_kwargs.setdefault("padding", "max_length")
+            te_cfgs[1].tokenizer_kwargs.setdefault("truncation", True)
+            te_cfgs[1].tokenizer_kwargs.setdefault("return_tensors", "pt")

fastvideo/configs/pipelines/glm_image.py

@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: Apache-2.0
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits.glm_image import GlmImageDiTConfig
+from fastvideo.configs.models.encoders import BaseEncoderOutput, T5Config
+from fastvideo.configs.models.vaes.glm_image import GlmImageVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig
+
+
+def glm_image_t5_postprocess(outputs: BaseEncoderOutput) -> torch.Tensor:
+    mask: torch.Tensor = outputs.attention_mask
+    hidden_state: torch.Tensor = outputs.last_hidden_state
+    seq_lens = mask.gt(0).sum(dim=1).long()
+
+    assert torch.isnan(hidden_state).sum() == 0, "T5 hidden states contain NaN"
+
+    max_len = 512
+    prompt_embeds = [u[:min(v, max_len)] for u, v in zip(hidden_state, seq_lens, strict=True)]
+    prompt_embeds_tensor: torch.Tensor = torch.stack(
+        [torch.cat([u, u.new_zeros(max_len - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0)
+
+    return prompt_embeds_tensor
+
+
+@dataclass
+class GlmImageConfig(PipelineConfig):
+
+    dit_config: DiTConfig = field(default_factory=GlmImageDiTConfig)
+    dit_precision: str = "bf16"
+
+    vae_config: VAEConfig = field(default_factory=GlmImageVAEConfig)
+    vae_precision: str = "fp32"
+    vae_tiling: bool = True
+    vae_sp: bool = False
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(default_factory=lambda: (T5Config(), ))
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("fp32", ))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
+                                  ...] = field(default_factory=lambda: (glm_image_t5_postprocess, ))
+
+    flow_shift: float | None = 1.0
+    embedded_cfg_scale: float = 7.5

fastvideo/configs/pipelines/kandinsky5.py

@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits import Kandinsky5VideoConfig
+from fastvideo.configs.models.encoders import BaseEncoderOutput, CLIPTextConfig
+from fastvideo.configs.models.encoders.reason1 import Reason1Config
+from fastvideo.configs.models.vaes import HunyuanVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, preprocess_text
+
+# Byte-exact copy of the upstream Kandinsky5/diffusers template, including the
+# "promt"/"scren" typos: the checkpoints were trained with this exact system  # codespell:ignore promt,scren
+# prompt, and ENCODE_START_IDX below is the tokenized length of everything
+# before the user prompt. Fixing the typos shifts user content to index 127
+# and mis-conditions every generation.
+KANDINSKY5_PROMPT_TEMPLATE = "\n".join([
+    "<|im_start|>system\nYou are a promt engineer. Describe the video in detail.",  # codespell:ignore promt
+    "Describe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.",
+    "Describe the location of the video, main characters or objects and their action.",
+    "Describe the dynamism of the video and presented actions.",
+    "Name the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.",  # codespell:ignore scren
+    "Describe the visual effects, postprocessing and transitions if they are presented in the video.",
+    "Pay attention to the order of key actions shown in the scene.<|im_end|>",
+    "<|im_start|>user\n{}<|im_end|>",
+])
+KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX = 129
+
+
+def kandinsky5_qwen_preprocess_text(prompt: str) -> str:
+    if not prompt.strip():
+        prompt = "."
+    return KANDINSKY5_PROMPT_TEMPLATE.format(prompt)
+
+
+def kandinsky5_qwen_postprocess_text(outputs: BaseEncoderOutput,
+                                     mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    if outputs.hidden_states is None:
+        raise RuntimeError("Kandinsky5 Qwen prompt embeddings require hidden_states.")
+    hidden_states = outputs.hidden_states[-1]
+    prompt_embeds = hidden_states[:, KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX:]
+    mask = mask[:, KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX:]
+    if prompt_embeds.shape[1] == 0:
+        prompt_embeds = hidden_states[:, -1:]
+        mask = torch.ones((mask.shape[0], 1), dtype=mask.dtype, device=mask.device)
+    return prompt_embeds, mask
+
+
+def kandinsky5_clip_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    if outputs.pooler_output is None:
+        raise RuntimeError("Kandinsky5 CLIP pooled output is required.")
+    return outputs.pooler_output
+
+
+@dataclass
+class Kandinsky5T2VConfig(PipelineConfig):
+    """Kandinsky-5.0 Lite text-to-video pipeline configuration."""
+
+    dit_config: DiTConfig = field(default_factory=Kandinsky5VideoConfig)
+    vae_config: VAEConfig = field(default_factory=HunyuanVAEConfig)
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(default_factory=lambda: (Reason1Config(), CLIPTextConfig()))
+    preprocess_text_funcs: tuple[Callable[[str], Any], ...] = field(
+        default_factory=lambda: (kandinsky5_qwen_preprocess_text, preprocess_text))
+    postprocess_text_funcs: tuple[Callable[..., Any], ...] = field(
+        default_factory=lambda: (kandinsky5_qwen_postprocess_text, kandinsky5_clip_postprocess_text))
+
+    dit_precision: str = "bf16"
+    vae_precision: str = "bf16"
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", "bf16"))
+    text_encoder_max_lengths: tuple[int, ...] = field(
+        default_factory=lambda: (KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX + 512, 77))
+
+    flow_shift: float | None = 5.0
+    vae_tiling: bool = True
+
+    def __post_init__(self) -> None:
+        if len(self.text_encoder_configs) != 2:
+            raise ValueError(f"Kandinsky5 pipeline requires exactly 2 text encoders (qwen and clip), "
+                             f"but got {len(self.text_encoder_configs)} encoder(s).")
+        if len(self.text_encoder_precisions) != 2:
+            raise ValueError("Kandinsky5 pipeline requires exactly 2 text encoder precisions, "
+                             f"but got {len(self.text_encoder_precisions)}.")
+        if len(self.text_encoder_max_lengths) != 2:
+            raise ValueError("Kandinsky5 pipeline requires exactly 2 text encoder max lengths, "
+                             f"but got {len(self.text_encoder_max_lengths)}.")
+
+        self.vae_config.load_encoder = False
+        self.vae_config.load_decoder = True
+
+        qwen_cfg = self.text_encoder_configs[0]
+        qwen_cfg.arch_config.output_hidden_states = True
+        qwen_cfg.arch_config.tokenizer_kwargs.update({
+            "padding": True,
+            "truncation": True,
+            "return_tensors": "pt",
+        })
+
+        clip_cfg = self.text_encoder_configs[1]
+        clip_cfg.arch_config.tokenizer_kwargs.update({
+            "padding": "max_length",
+            "max_length": 77,
+            "truncation": True,
+            "add_special_tokens": True,
+            "return_tensors": "pt",
+        })
+
+
+@dataclass
+class Kandinsky5I2VConfig(Kandinsky5T2VConfig):
+    """Kandinsky-5.0 image-to-video pipeline configuration."""
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        # I2V needs the VAE encoder to encode the conditioning image.
+        self.vae_config.load_encoder = True

fastvideo/dataset/parquet_dataset_map_style.py

@@ -2,6 +2,7 @@
 import os
 import pickle
 import random
+from collections.abc import Sequence
 from typing import Any
 
 import pyarrow as pa
@@ -94,8 +95,56 @@ class DP_SP_BatchSampler(Sampler[list[int]]):
         return len(self.sp_group_local_indices) // self.batch_size
 
 
-def get_parquet_files_and_length(path: str):
-    dataset_root = os.path.realpath(os.path.expanduser(path))
+def _parse_data_path_specs(path: str | Sequence[str] | dict[str, int]) -> list[tuple[str, int]]:
+    """Parse one or more dataset roots with old-framework repeat counts."""
+    if isinstance(path, dict):
+        return [
+            (str(root), int(repeat))
+            for root, repeat in path.items()
+            if int(repeat) > 0
+        ]
+    if isinstance(path, Sequence) and not isinstance(path, str):
+        return [(str(root), 1) for root in path]
+
+    specs: list[tuple[str, int]] = []
+    for part in str(path).split(","):
+        part = part.strip()
+        if not part:
+            continue
+        if ":" in part:
+            dir_path, count_str = part.rsplit(":", 1)
+            count = int(count_str)
+        else:
+            dir_path, count = part, 1
+        if count > 0:
+            specs.append((dir_path.strip(), count))
+    return specs
+
+
+def get_parquet_files_and_length(path: str | Sequence[str] | dict[str, int]):
+    specs = _parse_data_path_specs(path)
+    if len(specs) != 1 or specs[0][1] != 1:
+        all_file_names: list[str] = []
+        all_lengths: list[int] = []
+        for root, repeat in specs:
+            file_names, lengths = get_parquet_files_and_length(str(root))
+            for _ in range(repeat):
+                all_file_names.extend(file_names)
+                all_lengths.extend(lengths)
+        if not all_file_names:
+            raise FileNotFoundError(
+                "No parquet files found under dataset paths: "
+                f"{path}. "
+                "Please verify these paths point to preprocessed parquet data."
+            )
+        file_lengths = sorted(
+            zip(all_file_names, all_lengths, strict=True),
+            key=lambda x: x[0],
+        )
+        file_names_sorted, lengths_sorted = zip(*file_lengths, strict=True)
+        return file_names_sorted, lengths_sorted
+
+    dataset_root = os.path.realpath(os.path.expanduser(specs[0][0]))
     # Check if cached info exists
     cache_dir = os.path.join(dataset_root, "map_style_cache")
     cache_file = os.path.join(cache_dir, "file_info.pkl")
@@ -268,7 +317,7 @@ class LatentsParquetMapStyleDataset(Dataset):
 
     def __init__(
         self,
-        path: str,
+        path: str | Sequence[str] | dict[str, int],
         batch_size: int,
         parquet_schema: pa.Schema,
         cfg_rate: float = 0.0,
@@ -358,7 +407,6 @@ class LatentsParquetMapStyleDataset(Dataset):
     def __len__(self):
         return sum(self.lengths)
 
-
 # ────────────────────────────────────────────────────────────────────────────
 # 3.  Loader helper – everything else stays just like your original trainer
 # ────────────────────────────────────────────────────────────────────────────

fastvideo/distributed/parallel_state.py

@@ -23,6 +23,7 @@ If you only need to use the distributed environment without model parallelism,
  you can skip the model parallel initialization and destruction steps.
 """
 import contextlib
+import gc
 import os
 import pickle
 import weakref
@@ -1006,6 +1007,13 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
     if shutdown_ray:
         import ray  # Lazy import Ray
         ray.shutdown()
+    # Actually free GPU memory, as the name promises. FSDP-wrapped modules sit
+    # in reference cycles (module <-> FSDP state <-> hooks), so dropping the
+    # last user reference does not free their parameters until a gc pass runs.
+    # Without this, back-to-back model-load tests in one pytest session OOM.
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
 
 
 def get_same_node_ranks(pg: ProcessGroup | StatelessProcessGroup, source_rank: int = 0) -> list[int]:

fastvideo/envs.py

@@ -20,6 +20,7 @@ if TYPE_CHECKING:
     FASTVIDEO_LOGGING_CONFIG_PATH: str | None = None
     FASTVIDEO_TRACE_FUNCTION: int = 0
     FASTVIDEO_ATTENTION_BACKEND: str | None = None
+    FASTVIDEO_FA4: bool = False
     FASTVIDEO_WORKER_MULTIPROC_METHOD: str = "spawn"
     FASTVIDEO_TARGET_DEVICE: str = "cuda"
     MAX_JOBS: str | None = None
@@ -207,9 +208,18 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
     # - "SAGE_ATTN": use Sage Attention
     # - "SAGE_ATTN_THREE": use Sage Attention 3
+    # FLASH_ATTN uses FlashAttention-3/2; to run FlashAttention-4 set
+    # FASTVIDEO_FA4=1 as well (see below).
     "FASTVIDEO_ATTENTION_BACKEND":
     lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),
 
+    # If set (=1), the FLASH_ATTN backend uses FlashAttention-4
+    # (flash_attn.cute). FA4 is opt-in and never auto-selected just because it
+    # is installed. Below sm90, grad-enabled and GQA calls are routed to FA2
+    # (FA4's backward asserts sm90+ and its pack_gqa fails to JIT there).
+    "FASTVIDEO_FA4":
+    lambda: os.getenv("FASTVIDEO_FA4", "0") != "0",
+
     # Use dedicated multiprocess context for workers.
     "FASTVIDEO_WORKER_MULTIPROC_METHOD":
     lambda: os.getenv("FASTVIDEO_WORKER_MULTIPROC_METHOD", "spawn"),

fastvideo/eval/README.md

@@ -122,6 +122,44 @@ to run a subset of the Evaluator's registered metrics on this batch
 (useful for scoring different corpora with different metric subsets in
 successive calls without burning model loads).
 
+### Training-time validation metrics
+
+The modular trainer can score validation videos during training through
+`callbacks.validation.metrics`. Each distributed rank that writes local
+validation videos also runs its local metrics on `cuda:<local_rank>`;
+rank 0 only merges scalar summaries and logs artifacts.
+
+```yaml
+callbacks:
+  validation:
+    pipeline_target: fastvideo.pipelines.basic.matrixgame2.matrixgame2_causal_dmd_pipeline.MatrixGame2CausalDMDPipeline
+    dataset_file: examples/distill/MatrixGame2.0/validation.json
+    sampling_steps: [4]
+    metrics:
+      enabled: true
+      names:
+        - vbench.imaging_quality
+        - vbench.aesthetic_quality
+        - optical_flow.synthetic_optical_flow
+        - common.fvd
+      skip_missing_deps: true
+      strict: false
+      unload_after_validation: true
+
+      # `synthetic_optical_flow` also needs actions in the validation
+      # manifest (`action_path`) plus a repo-local calibration file.
+      calibration_path: assets/eval/worldmodel_synthetic_flow_calibration.json
+```
+
+Metric summaries are written under
+`<output_dir>/eval/step_<step>/inference_steps_<n>_rank_<rank>.json` and
+scalar means are logged with the `metrics/validation/...` prefix.
+Install `fastvideo[eval]` for optical-flow dependencies such as
+`ptlflow`; VBench metrics use the pinned submodule under
+`fastvideo/third_party/eval/vbench`. FVD uses `ref_video` entries from
+the validation manifest when present, or the standard
+`FASTVIDEO_FVD_REF_FEATURES` / eval-cache reference feature path.
+
 ### CLI
 
 ```bash
@@ -350,5 +388,3 @@ sweep several baselines into a table, see
 
 - **MIND** metrics. Depend on a separate `vipe` upstream submodule.
 - **VBench-2.0**. Sibling vbench2 package; needs its own port.
-- **Training-time eval callback** (`EvalCallback`) and the
-  `RolloutEvaluator` helper.

fastvideo/eval/metrics/optical_flow/synthetic_optical_flow/_thirdperson.py

@@ -63,7 +63,7 @@ class ThirdPersonCalibration:
     beta_fwd: float
     beta_strafe: float
     focal_length: float
-    r_z: float
+    r_z: float = 0.0
     r_y: float = 0.0
     init_pitch: float = 0.0
     notes: str = ""

fastvideo/layers/quantization/nvfp4_config.py

@@ -261,6 +261,22 @@ def _mm_fp4(
     )
 
 
+def _coerce_fp4_input_dtype(x: torch.Tensor) -> torch.Tensor:
+    """Coerce an activation to a dtype the FP4 linear accepts.
+
+    The pre-attention norm can emit fp32 (e.g. in eager mode, without the
+    torch.compile fusion that keeps it bf16). The FP4 linear emits bf16
+    regardless (see _mm_fp4 out dtype), so cast fp32 -> bf16 rather than
+    failing, matching the sibling fastvideo/layers/fp4linear.py. Non-floating
+    inputs (e.g. int/bool) are a genuine error and are rejected fast.
+    """
+    if not x.is_floating_point():
+        raise TypeError(f"fp4 linear expects floating-point inputs, got {x.dtype}")
+    if x.dtype not in (torch.bfloat16, torch.float16):
+        x = x.to(torch.bfloat16)
+    return x
+
+
 class NVFP4QuantizeMethod(QuantizeMethodBase):
 
     def __init__(self, layer_prefix: str = ""):
@@ -285,8 +301,7 @@ class NVFP4QuantizeMethod(QuantizeMethodBase):
 
     def quantize_input(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         SfLayout, _, _ = _require_flashinfer()
-        assert x.dtype == torch.bfloat16 or x.dtype == torch.float16, (
-            f"only allow bf16/fp16 inputs to fp4 linear, got {x.dtype}")
+        x = _coerce_fp4_input_dtype(x)
         x_2d = x.view(-1, x.shape[-1])
         x_fp4, x_scale = _nvfp4_quantize(
             x_2d,
@@ -332,8 +347,7 @@ class NVFP4QuantizeMethod(QuantizeMethodBase):
             if x_scale.dim() > 2:
                 x_scale = x_scale.view(-1, x_scale.shape[-1])
         else:
-            assert x.dtype == torch.bfloat16 or x.dtype == torch.float16, (
-                f"only allow bf16/fp16 inputs to fp4 linear, got {x.dtype}")
+            x = _coerce_fp4_input_dtype(x)
             x = x.view(-1, x.shape[-1])
             x_global_sf = self.x_global_sf
             x_fp4, x_scale = _nvfp4_quantize(

fastvideo/layers/rotary_embedding.py

@@ -295,6 +295,7 @@ def get_1d_rotary_pos_embed(
     interpolation_factor: float = 1.0,
     dtype: torch.dtype = torch.float32,
     use_real: bool = True,
+    freqs_dtype: torch.dtype | None = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """
     Precompute the frequency tensor for complex exponential (cis) with given dimensions.
@@ -319,12 +320,16 @@ def get_1d_rotary_pos_embed(
     if isinstance(pos, int):
         pos = torch.arange(pos).float()
 
+    # freqs_dtype is an alias for dtype (Diffusers-compatible calling convention).
+    if freqs_dtype is not None:
+        dtype = freqs_dtype
+
     # proposed by reddit user bloc97, to rescale rotary embeddings to longer sequence length without fine-tuning
     # has some connection to NTK literature
     if theta_rescale_factor != 1.0:
         theta *= theta_rescale_factor**(dim / (dim - 2))
 
-    freqs = 1.0 / (theta**(torch.arange(0, dim, 2)[:(dim // 2)].to(dtype) / dim))  # [D/2]
+    freqs = 1.0 / (theta**(torch.arange(0, dim, 2, device=pos.device)[:(dim // 2)].to(dtype) / dim))  # [D/2]
     freqs = torch.outer(pos * interpolation_factor, freqs)  # [S, D/2]
     freqs_cos = freqs.cos()  # [S, D/2]
     freqs_sin = freqs.sin()  # [S, D/2]
@@ -445,6 +450,21 @@ def get_nd_rotary_pos_embed(
     return cos, sin
 
 
+_ROTARY_POS_EMBED_CACHE: dict[tuple, tuple[torch.Tensor, torch.Tensor]] = {}
+# Bound the table cache so long-running servers / causal models (which vary
+# start_frame per frame) cannot grow it without limit; entries are large float64
+# tensors. Least-recently-used eviction keeps the active resolution(s) hot while
+# capping memory.
+_ROTARY_POS_EMBED_CACHE_MAXSIZE = 16
+
+
+def _hashable(value: Any) -> Any:
+    """Return a hashable view of a scalar or sequence for use in a cache key."""
+    if isinstance(value, list | tuple):
+        return tuple(value)
+    return value
+
+
 def get_rotary_pos_embed(
     rope_sizes,
     hidden_size,
@@ -495,6 +515,31 @@ def get_rotary_pos_embed(
         sp_rank = 0
         sp_world_size = 1
 
+    # Memoize on every output-affecting argument; the table is constant across
+    # denoising steps, so this avoids recomputing the float64 cos/sin tables.
+    cache_key = (
+        _hashable(rope_sizes),
+        tuple(rope_dim_list),
+        rope_theta,
+        _hashable(theta_rescale_factor),
+        _hashable(interpolation_factor),
+        shard_dim,
+        sp_rank,
+        sp_world_size,
+        dtype,
+        start_frame,
+        use_real,
+    )
+    cached = _ROTARY_POS_EMBED_CACHE.get(cache_key)
+    if cached is not None:
+        # Move to most-recently-used position so the active table is not evicted
+        # when several resolutions / buckets share the process (LRU recency).
+        # Pop with a default: a concurrent eviction between the get() above and
+        # here would otherwise raise KeyError on the hit path.
+        if _ROTARY_POS_EMBED_CACHE.pop(cache_key, None) is not None:
+            _ROTARY_POS_EMBED_CACHE[cache_key] = cached
+        return cached
+
     freqs_cos, freqs_sin = get_nd_rotary_pos_embed(
         rope_dim_list,
         rope_sizes,
@@ -508,6 +553,14 @@ def get_rotary_pos_embed(
         start_frame=start_frame,
         use_real=use_real,
     )
+    # The returned tensors are shared cache entries: callers must never mutate
+    # them in place. Note .to(device) is an identity alias when the tensor is
+    # already on the target device (e.g. CPU runs), so it does NOT guarantee a
+    # copy — treat the tables as read-only and copy before any in-place op.
+    # Reached only on a miss, so evict the least-recently-used entry at capacity.
+    if len(_ROTARY_POS_EMBED_CACHE) >= _ROTARY_POS_EMBED_CACHE_MAXSIZE:
+        _ROTARY_POS_EMBED_CACHE.pop(next(iter(_ROTARY_POS_EMBED_CACHE)))
+    _ROTARY_POS_EMBED_CACHE[cache_key] = (freqs_cos, freqs_sin)
     return freqs_cos, freqs_sin
 
 

fastvideo/models/dits/causal_wanvideo.py

@@ -323,9 +323,9 @@ class CausalWanTransformerBlock(nn.Module):
         value, _ = self.to_v(norm_hidden_states)
 
         if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
         if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)
 
         query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
         key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -437,6 +437,7 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         # Causal-specific
         self.block_mask = None
+        self.teacher_forcing_block_mask = None
         self.num_frame_per_block = config.arch_config.num_frames_per_block
         assert self.num_frame_per_block <= 3
         self.independent_first_frame = False
@@ -500,6 +501,70 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         return block_mask
 
+    @staticmethod
+    def _prepare_teacher_forcing_mask(
+        device: torch.device | str, num_frames: int = 21,
+        frame_seqlen: int = 1560, num_frame_per_block=1, local_attn_size=-1
+    ) -> BlockMask:
+        """Attention mask for the teacher-forcing ``[clean | noisy]`` sequence.
+
+        A noisy token attends to its own block plus the clean context of all
+        strictly previous blocks; clean tokens are block-wise causal.
+        """
+        if local_attn_size != -1:
+            raise NotImplementedError(
+                f"Teacher forcing ignores local_attn_size={local_attn_size}: "
+                "unlike the block-wise causal mask, this mask always attends "
+                "to the full clean context. Windowed teacher forcing is not "
+                "implemented; use local_attn_size=-1 for teacher-forcing "
+                "training.")
+        total_length = num_frames * frame_seqlen * 2
+        padded_length = math.ceil(total_length / 128) * 128 - total_length
+
+        clean_ends = num_frames * frame_seqlen
+        context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
+        noise_context_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
+        noise_context_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
+        noise_noise_starts = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
+        noise_noise_ends = torch.zeros(total_length + padded_length, device=device, dtype=torch.long)
+
+        attention_block_size = frame_seqlen * num_frame_per_block
+        frame_indices = torch.arange(
+            start=0, end=num_frames * frame_seqlen,
+            step=attention_block_size, device=device, dtype=torch.long
+        )
+        for start in frame_indices:
+            context_ends[start:start + attention_block_size] = start + attention_block_size
+
+        noisy_image_start_list = torch.arange(
+            num_frames * frame_seqlen, total_length,
+            step=attention_block_size, device=device, dtype=torch.long
+        )
+        noisy_image_end_list = noisy_image_start_list + attention_block_size
+        for block_index, (start, end) in enumerate(zip(noisy_image_start_list, noisy_image_end_list)):
+            noise_noise_starts[start:end] = start
+            noise_noise_ends[start:end] = end
+            noise_context_ends[start:end] = block_index * attention_block_size
+
+        def attention_mask(b, h, q_idx, kv_idx):
+            clean_mask = (q_idx < clean_ends) & (kv_idx < context_ends[q_idx])
+            c1 = (kv_idx < noise_noise_ends[q_idx]) & (kv_idx >= noise_noise_starts[q_idx])
+            c2 = (kv_idx < noise_context_ends[q_idx]) & (kv_idx >= noise_context_starts[q_idx])
+            noise_mask = (q_idx >= clean_ends) & (c1 | c2)
+            eye_mask = q_idx == kv_idx
+            return eye_mask | clean_mask | noise_mask
+
+        block_mask = create_block_mask(
+            attention_mask, B=None, H=None,
+            Q_LEN=total_length + padded_length, KV_LEN=total_length + padded_length,
+            _compile=False, device=device)
+
+        if not dist.is_initialized() or dist.get_rank() == 0:
+            print(f" cache a teacher-forcing mask with block size of {num_frame_per_block} frames")
+            print(block_mask)
+
+        return block_mask
+
     def _forward_inference(
                 self,
                 hidden_states: torch.Tensor,
@@ -600,7 +665,8 @@ class CausalWanTransformer3DModel(BaseDiT):
             else:
                 causal_kwargs = {
                     "kv_cache": kv_cache[block_index],
-                    "crossattn_cache": crossattn_cache[block_index],
+                    "crossattn_cache": (crossattn_cache[block_index]
+                                        if crossattn_cache is not None else None),
                     "current_start": current_start,
                     "cache_start": cache_start,
                     "block_mask": self.block_mask,
@@ -628,9 +694,12 @@ class CausalWanTransformer3DModel(BaseDiT):
                 encoder_hidden_states_image: torch.Tensor | list[torch.Tensor]
                 | None = None,
                 start_frame: int = 0,
+                clean_x: torch.Tensor | None = None,
+                aug_t: torch.Tensor | None = None,
                 **kwargs) -> torch.Tensor:
 
         orig_dtype = hidden_states.dtype
+        teacher_forcing = clean_x is not None
         if not isinstance(encoder_hidden_states, torch.Tensor):
             encoder_hidden_states = encoder_hidden_states[0]
         if isinstance(encoder_hidden_states_image,
@@ -663,15 +732,26 @@ class CausalWanTransformer3DModel(BaseDiT):
         freqs_cis = (freqs_cos,
                      freqs_sin) if freqs_cos is not None else None
 
-        # Construct blockwise causal attn mask
-        if self.block_mask is None:
-            self.block_mask = self._prepare_blockwise_causal_attn_mask(
-                device=hidden_states.device,
-                num_frames=num_frames,
-                frame_seqlen=post_patch_height * post_patch_width,
-                num_frame_per_block=self.num_frame_per_block,
-                local_attn_size=self.local_attn_size
-            )
+        if teacher_forcing:
+            if self.teacher_forcing_block_mask is None:
+                self.teacher_forcing_block_mask = self._prepare_teacher_forcing_mask(
+                    device=hidden_states.device,
+                    num_frames=num_frames,
+                    frame_seqlen=post_patch_height * post_patch_width,
+                    num_frame_per_block=self.num_frame_per_block,
+                    local_attn_size=self.local_attn_size,
+                )
+            block_mask = self.teacher_forcing_block_mask
+        else:
+            if self.block_mask is None:
+                self.block_mask = self._prepare_blockwise_causal_attn_mask(
+                    device=hidden_states.device,
+                    num_frames=num_frames,
+                    frame_seqlen=post_patch_height * post_patch_width,
+                    num_frame_per_block=self.num_frame_per_block,
+                    local_attn_size=self.local_attn_size
+                )
+            block_mask = self.block_mask
 
         hidden_states = self.patch_embedding(hidden_states)
         grid_sizes = torch.stack(
@@ -679,6 +759,7 @@ class CausalWanTransformer3DModel(BaseDiT):
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
 
         encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states.new_zeros(1, self.text_len - encoder_hidden_states.size(1), encoder_hidden_states.size(2))], dim=1)
+        encoder_hidden_states_text = encoder_hidden_states
 
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
                         timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
@@ -694,18 +775,35 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         assert encoder_hidden_states.dtype == orig_dtype
 
+        if teacher_forcing:
+            # Tile RoPE/modulation so clean frame i and noisy frame i share a position.
+            clean_tokens = self.patch_embedding(clean_x).flatten(2).transpose(1, 2)
+            hidden_states = torch.cat([clean_tokens, hidden_states], dim=1)
+            if aug_t is None:
+                aug_t = torch.zeros_like(timestep)
+            _, timestep_proj_clean, _, _ = self.condition_embedder(
+                aug_t.flatten(), encoder_hidden_states_text, None)
+            timestep_proj_clean = timestep_proj_clean.unflatten(
+                1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
+            timestep_proj = torch.cat([timestep_proj_clean, timestep_proj], dim=1)
+            freqs_cis = (torch.cat([freqs_cos, freqs_cos], dim=0),
+                         torch.cat([freqs_sin, freqs_sin], dim=0))
+
         # 4. Transformer blocks
         if torch.is_grad_enabled() and self.gradient_checkpointing:
             for block in self.blocks:
                 hidden_states = self._gradient_checkpointing_func(
                     block, hidden_states, encoder_hidden_states,
                     timestep_proj, freqs_cis,
-                    block_mask=self.block_mask)
+                    block_mask=block_mask)
         else:
             for block in self.blocks:
                 hidden_states = block(hidden_states, encoder_hidden_states,
                                         timestep_proj, freqs_cis,
-                                        block_mask=self.block_mask)
+                                        block_mask=block_mask)
+
+        if teacher_forcing:
+            hidden_states = hidden_states[:, hidden_states.shape[1] // 2:]
 
         # 5. Output norm, projection & unpatchify
         temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)