[bugfix]: keep _rms_norm_dispatch output in the input dtype under autocast

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

@@ -1,23 +1,20 @@
 ---
 name: reseed-performance-baseline
-description: Re-seed the HF performance-tracking baseline for an intentional runtime, dependency, environment-caused benchmark shift, or reviewed v2 calibration using one or more reviewed normalized performance JSONs. Use when performance CI fails because metrics such as latency, throughput, component time, or peak memory changed for an accepted reason and the rolling median baseline in FastVideo/performance-tracking must be advanced, or when a new v2 exact comparable identity needs its first approved baseline. The workflow backs up existing history under /tmp, validates all source JSONs for the same legacy (model_id, gpu_type) target or the same v2 exact identity, rejects internally inconsistent source batches, uploads one success=true baseline record per accepted source JSON, and offers to clean local temp state after a successful upload.
+description: Re-seed the HF performance-tracking baseline for an intentional runtime, dependency, or environment-caused benchmark shift using one or more reviewed normalized performance JSONs. Use when performance CI fails because metrics such as latency, throughput, component time, or peak memory changed for an accepted reason and the rolling median baseline in FastVideo/performance-tracking must be advanced from a consistent batch of reviewed source results. The workflow backs up existing history under /tmp, validates all source JSONs for the same (model_id, gpu_type), rejects internally inconsistent source batches, uploads one success=true reseed record per accepted source JSON, and offers to clean local temp state after a successful upload.
 ---
 
 # Re-seed Performance Baseline
 
 ## Purpose
 
-Replace or advance the rolling performance baseline in the HF dataset
-`FastVideo/performance-tracking`. Legacy targets are scoped by
-`(model_id, gpu_type)`. V2 targets are scoped by exact comparable identity:
-`workload_id`, `variant_id`, `benchmark_version`, `hardware_profile_id`,
-`software_profile_id`, and `recipe_fingerprint`.
+Replace or advance the rolling performance baseline for a single
+`(model_id, gpu_type)` pair in the HF dataset
+`FastVideo/performance-tracking`.
 
-Performance comparison uses the median of up to the last 5 successful,
-baseline-eligible records for the same target. Failed or calibration-only
-records are useful audit history, but they do not move the future baseline
-because `compare_baseline.py` loads records with `successful_only=True` and
-`baseline_eligible_only=True`.
+Performance comparison uses the median of up to the last 5 successful records
+for the same model and GPU. Failed records are useful audit history, but they
+do not move the future baseline because `compare_baseline.py` loads records
+with `successful_only=True`.
 
 This skill now reseeds from a reviewed batch of one or more source performance
 JSONs. It uploads one new `success=true` record per accepted source JSON; it
@@ -25,13 +22,11 @@ does not blindly replicate one measurement into 3 or 5 records. The effective
 reseed size is therefore dynamic and equals the number of provided, validated,
 internally consistent source JSONs.
 
-For baseline shifts with existing history, if the operator provides fewer than
-3 records, call out that the last-5 rolling median may not move immediately. If
-the operator provides 3 consistent shifted records, the rolling median usually
-moves immediately. If the operator provides 5 consistent shifted records, the
-last-5 window is effectively reset to the new runtime profile. For the first
-approved v2 baseline of a new exact identity, one reviewed calibration seed is
-enough for the next comparable run to leave `CALIBRATION_NEEDED`.
+If the operator provides fewer than 3 records, call out that the last-5 rolling
+median may not move immediately. If the operator provides 3 consistent shifted
+records, the rolling median usually moves immediately. If the operator provides
+5 consistent shifted records, the last-5 window is effectively reset to the new
+runtime profile.
 
 These records are intentional operator-approved baseline resets, not ordinary
 independent main-branch persistence. Mark them clearly with provenance fields
@@ -71,8 +66,8 @@ approval, then upload reviewed accepted baseline records.
 
 | Parameter | Required | Description |
 |-----------|----------|-------------|
-| `model_id` | Legacy required; v2 inferred | Benchmark id, e.g. `wan-t2v-1.3b-2gpu`. This maps to the HF subdirectory after `sanitize(model_id)`. For v2 records, use the `model_id` from each source artifact only as the upload directory; comparison is by exact identity. |
-| `gpu_type` | Legacy required; v2 inferred | Exact GPU device string from the performance record, e.g. the L40S device name emitted by CI. V2 hardware matching uses `hardware_profile_id`; preserve `gpu_type` as display metadata. |
+| `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: `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. |
@@ -83,14 +78,10 @@ Hardcoded defaults:
   supported by the code, but use the default unless the user explicitly asks).
 - Local sync root: `/tmp/perf-tracking` (`PERFORMANCE_TRACKING_ROOT` override
   is supported).
-- Prepared-record staging root: `/tmp/performance_reseed_prepared`
-  (`PERFORMANCE_RESEED_STAGING_ROOT` override is supported). Keep it separate
-  and non-nested from the sync root.
 - Backup root: `/tmp/performance_reseed_backup`.
 - Download scratch root for source artifact URLs: `/tmp/performance_reseed_source`.
-- Baseline window: last 5 `success=true`, `baseline_eligible=true` records
-  for the same legacy `(model_id, gpu_type)` target or the same v2 exact
-  comparable identity.
+- Baseline window: last 5 `success=true` records for the same
+  `(model_id, gpu_type)`.
 - Reseed count: dynamic. Upload exactly one accepted seed record per validated
   source JSON.
 
@@ -124,24 +115,12 @@ with open(source_result, encoding="utf-8") as f:
     record = json.load(f)
 ```
 
-Classify the source batch before continuing:
-
-- **Legacy source records** have no v2 exact identity fields. Stop if any
-  normalized record's `model_id` or `gpu_type` does not match the requested
-  `model_id` and `gpu_type`.
-- **V2 source records** have exact identity fields. Stop unless every source
-  record has all six comparable identity fields and they are identical across
-  the batch: `workload_id`, `variant_id`, `benchmark_version`,
-  `hardware_profile_id`, `software_profile_id`, and `recipe_fingerprint`.
-  Do not fall back to legacy `(model_id, gpu_type)` matching for v2 records.
+Stop if any normalized record's `model_id` or `gpu_type` does not match the
+requested `model_id` and `gpu_type`.
 
 The source records may have `success: false` when they came from failed
 rolling baseline comparisons. That is expected; only the reviewed reseed
 records become new `success: true` baseline records after explicit approval.
-For a first v2 baseline seed, the source records must instead be successful
-scheduled-main full-suite `CALIBRATION_NEEDED` normalized artifacts. Reject PR,
-local, direct-run, non-main-branch, or non-full-suite calibration artifacts as
-seed sources.
 
 Sort validated source records by their original `timestamp` ascending before
 preparing the seed records. If a source timestamp is missing or unparsable,
@@ -215,7 +194,7 @@ export HF_REPO_ID="${HF_REPO_ID:-FastVideo/performance-tracking}"
 python -c 'from fastvideo.performance.hf_store import sync_from_hf; import os; sync_from_hf(os.environ["PERFORMANCE_TRACKING_ROOT"], strict=True)'
 ```
 
-For legacy records, back up the sanitized model directory under `/tmp`:
+Then back up only the sanitized model directory under `/tmp`:
 
 ```bash
 SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
@@ -230,16 +209,6 @@ mkdir -p "$BACKUP_DIR"
 cp -R "${PERFORMANCE_TRACKING_ROOT}/${MODEL_SAFE}" "$BACKUP_DIR/" 2>/dev/null || true
 ```
 
-For v2 records, back up the full local tracking root after sync. Exact identity
-lookup scans across model directories, so a benchmark rename may have relevant
-history outside the current source artifact's `model_id` directory:
-
-```bash
-BACKUP_DIR="/tmp/performance_reseed_backup/${TIMESTAMP}_${SHORT_COMMIT}_v2_exact_identity"
-mkdir -p "$BACKUP_DIR"
-cp -R "${PERFORMANCE_TRACKING_ROOT}" "$BACKUP_DIR/tracking-root"
-```
-
 Write provenance next to the backup:
 
 ```bash
@@ -262,9 +231,7 @@ first baseline seed. Continue, but report that baseline history was empty.
 
 ### 3. Compute old baseline and candidate shift
 
-Load the last 5 successful baseline records for the target.
-
-For legacy targets:
+Load the last 5 successful records for the target:
 
 ```python
 from fastvideo.performance.hf_store import load_records_for_model
@@ -275,28 +242,6 @@ records = load_records_for_model(
     "<gpu_type>",
     last_n=5,
     successful_only=True,
-    baseline_eligible_only=True,
-)
-```
-
-For v2 exact-identity targets:
-
-```python
-from fastvideo.performance.hf_store import load_records_for_identity
-
-records = load_records_for_identity(
-    "/tmp/perf-tracking",
-    {
-        "workload_id": "<workload_id>",
-        "variant_id": "<variant_id>",
-        "benchmark_version": "<benchmark_version>",
-        "hardware_profile_id": "<hardware_profile_id>",
-        "software_profile_id": "<software_profile_id>",
-        "recipe_fingerprint": "<recipe_fingerprint>",
-    },
-    last_n=5,
-    successful_only=True,
-    baseline_eligible_only=True,
 )
 ```
 
@@ -312,8 +257,7 @@ medians after appending the proposed seed records, and source batch spread for:
 
 Also print how many successful old records exist. Make clear:
 
-- 1 seed record usually does not move an existing last-5 median by itself, but
-  it is enough to establish the first v2 baseline for a new exact identity.
+- 1 seed record usually does not move a last-5 median by itself.
 - 3 consistent seed records usually move the last-5 median immediately.
 - 5 consistent seed records effectively reset the last-5 window.
 - The records are intentional approved baseline resets and must be labeled
@@ -323,10 +267,10 @@ Also print how many successful old records exist. Make clear:
 
 Require an explicit confirmation phrase before preparing the upload:
 
-> About to RE-SEED performance baseline for `<target description>`.
+> About to RE-SEED performance baseline for `<model_id>` on `<gpu_type>`.
 > This will upload `<N>` new `success=true` records to
-> `FastVideo/performance-tracking/<sanitize(model_id)>/` or the source
-> artifact's v2 model directory, one per accepted source JSON.
+> `FastVideo/performance-tracking/<sanitize(model_id)>/`, one per accepted
+> source JSON.
 >
 > Reason: `<intent_rationale>`
 > Source results: `<source_results>`
@@ -344,63 +288,25 @@ Do not continue unless the user types exactly `confirm performance reseed`.
 
 ### 5. Create the accepted seed records
 
-Create one seed record from each normalized source result.
-
-For first v2 baseline seeds, use the scoped utility. It validates exact
-identity, requires successful scheduled-main full-suite `CALIBRATION_NEEDED`
-source artifacts, preserves the normalized v2 identity and metadata fields,
-and writes seed records with `success=true`, `baseline_eligible=true`, and
-`comparison_status=PASS`:
-
-```bash
-python fastvideo/tests/performance/seed_baseline.py \
-  --source-result <normalized_perf_1.json> \
-  --source-result <normalized_perf_2.json> \
-  --intent-rationale "<intent_rationale>" \
-  --max-intra-batch-regression 0.05 \
-  --tracking-root "${PERFORMANCE_TRACKING_ROOT}" \
-  --staging-root "${PERFORMANCE_RESEED_STAGING_ROOT:-/tmp/performance_reseed_prepared}"
-```
-
-The utility is prepare-only and intentionally has no upload option. Upload the
-scoped records only after the separate confirmation in step 6.
-
-The utility validates against an isolated fresh HF snapshot and leaves
-`PERFORMANCE_TRACKING_ROOT` untouched; that argument only proves the staging
-root is separate from the operator's tracking mirror. Before writing, it stops
-if the exact identity already has a successful baseline-eligible record or if
-the workload/variant/version already trusts another recipe. It atomically
-reserves the exact identity and writes a digest-protected upload manifest bound
-to the current HF endpoint, repository id, and repository type. Keep the
-prepared records, manifest, source files, and reservation unchanged until the
-operation is uploaded or explicitly cleaned up.
-
-If the prepared seed records look correct, upload only those scoped records in
-step 7. Do not rerun the utility with a different source list after approval.
-
-For legacy reseeds or accepted v2 baseline shifts from regression artifacts,
-create one seed record from each normalized source result. Do not copy the
+Create one seed record from each normalized source result. Do not copy the
 source JSON wholesale.
 
 Infer the baseline field allowlist from all existing HF records for the target
-after syncing, including both `success=true` and `success=false` records. For
-legacy targets the target is `(model_id, gpu_type)`. For v2 baseline-shift
-reseeds the target is the exact comparable identity. Use the union of
-non-provenance keys present in those target records, preserving only fields
-that also exist in the normalized source record or are explicitly set by the
-reseed workflow. Always include `model_id`, `timestamp`, `success`,
-`baseline_eligible`, and `comparison_status` because the upload path and
-baseline loader depend on them. Always set `timestamp` to a fresh reseed
-timestamp, `success` to `true`, `baseline_eligible` to `true`, and
-`comparison_status` to `PASS`. Do not include unrelated source-only fields
-that are absent from existing HF records.
+`(model_id, gpu_type)` after syncing, including both `success=true` and
+`success=false` records. Use the union of non-provenance keys present in those
+target records, preserving only fields that also exist in the normalized
+source record or are explicitly set by the reseed workflow. Always include
+`model_id`, `timestamp`, and `success` because the upload path and baseline
+loader depend on them. Always set `timestamp` to a fresh reseed timestamp and
+`success` to `true`. Do not include unrelated source-only fields that are
+absent from existing HF records.
 
 Exclude existing provenance or operator metadata from the inferred baseline
 field allowlist. At minimum, exclude keys prefixed with `baseline_reseed` and
 any fields known to be local-only audit metadata.
 
-If there are no previous HF records for the target, fall back to this default
-baseline field list:
+If there are no previous HF records for the target model/GPU, fall back to this
+default baseline field list:
 
 - `model_id`
 - `timestamp`
@@ -413,22 +319,6 @@ baseline field list:
 - `dit_time_s`
 - `vae_decode_time_s`
 - `success`
-- `baseline_eligible`
-- `comparison_status`
-
-For v2 baseline-shift reseeds with no previous HF records for the exact
-identity, also preserve:
-
-- `workload_id`
-- `variant_id`
-- `benchmark_version`
-- `recipe_fingerprint`
-- `hardware_profile_id`
-- `software_profile_id`
-- `recipe`
-- `hardware_profile`
-- `software_profile`
-- `software_comparison_profile`
 
 Do not upload extra fields from the source artifact.
 
@@ -444,22 +334,6 @@ Optional provenance fields are allowed and useful:
 - `baseline_reseed_operator`
 - `baseline_reseed_max_intra_batch_regression`
 
-The v2 calibration seed utility writes analogous first-seed provenance:
-
-- `baseline_seed: true`
-- `baseline_seed_reason`
-- `baseline_seed_source_result`
-- `baseline_seed_source_status`
-- `baseline_seed_source_timestamp`
-- `baseline_seed_source_success`
-- `baseline_seed_source_run_source`
-- `baseline_seed_source_branch`
-- `baseline_seed_source_test_scope`
-- `baseline_seed_source_pr_number`
-- `baseline_seed_batch_size`
-- `baseline_seed_batch_index`
-- `baseline_seed_operator`
-
 Use a fresh reseed timestamp for each seed record, not the original source
 result timestamp. This is required because
 `load_records_for_model(..., last_n=5)` keeps the last records after loading
@@ -482,8 +356,7 @@ Prefer uploading new accepted seed records so failed history remains visible.
 Print:
 
 - Backup directory path under `/tmp`.
-- Prepared local record paths under `PERFORMANCE_RESEED_STAGING_ROOT`.
-- Prepared upload-manifest path under the identity reservation.
+- Prepared local record paths under `PERFORMANCE_TRACKING_ROOT`.
 - HF paths that will receive the new records.
 - Old rolling medians.
 - Source batch medians, source batch spread, reseed count, and candidate
@@ -495,36 +368,22 @@ prepared records plus backup on disk.
 
 ### 7. Upload only the scoped records
 
-For a first v2 calibration seed, use the manifest uploader after the user
-replies exactly `upload`:
+Use the shared storage helper so the path and repo type match CI:
 
-```bash
-python -c 'from fastvideo.tests.performance.seed_baseline import upload_prepared_seed_manifest; print(upload_prepared_seed_manifest("<prepared_manifest>"))'
+```python
+from fastvideo.performance.hf_store import upload_record
+
+upload_record("<local_record_path>", record, strict=True)
 ```
 
-The uploader verifies the source and prepared-record digests, pins and scans
-the current HF revision, rechecks exact-identity and recipe-cohort conflicts,
-and writes the entire batch in one commit whose `parent_commit` must still be
-current. A concurrent Hub update makes the commit fail. Do not retry
-automatically: preserve staging, refresh/review remote state, and request a new
-explicit `upload` after the conflict is understood. Each record goes to:
+Run it once per prepared record. Each upload goes to:
 
 ```text
 FastVideo/performance-tracking/<sanitize(model_id)>/<record_filename>.json
 ```
 
-Never call `upload_record()` once per first-seed record: that can partially
-land the batch and has no compare-and-swap guard.
-
-For a legacy reseed or an accepted v2 baseline shift, the first-seed manifest
-validator does not apply because an eligible baseline already exists. Upload
-only the individually reviewed records prepared in step 5 with the shared
-`upload_record(local_path, record, strict=True)` helper. Stop on the first
-failure and report exactly which records reached HF; do not silently rerun or
-replicate the remainder.
-
-Never bulk upload the tracking or staging root, and never modify another
-model's directory in the same operation.
+Never bulk upload the whole tracking root. Never modify another model's
+directory in the same operation.
 
 ### 8. Report outcome and offer cleanup
 
@@ -546,14 +405,9 @@ distinguish an accepted baseline shift from a hidden regression.
 After the upload is verified, ask whether the user wants to clear temporary
 local state. Explain what each directory is for:
 
-- `PERFORMANCE_TRACKING_ROOT`, usually `/tmp/perf-tracking`: read-only local
-  synced mirror used for operator review and reporting. First-v2 preparation
-  independently proves remote state from a fresh temporary HF snapshot.
-- `PERFORMANCE_RESEED_STAGING_ROOT`, usually
-  `/tmp/performance_reseed_prepared`: prepared local seed records used for the
-  scoped upload, plus the identity reservation and digest manifest. Keeping
-  this separate prevents aborted preparations from appearing in later
-  baseline reads.
+- `PERFORMANCE_TRACKING_ROOT`, usually `/tmp/perf-tracking`: local synced
+  mirror of `FastVideo/performance-tracking` plus the prepared local seed
+  records used for scoped upload.
 - `/tmp/performance_reseed_backup/<...>`: local backup of the target model's
   pre-reseed HF history plus `PROVENANCE.txt`, kept so a bad reseed can be
   audited or corrected.
@@ -563,18 +417,14 @@ local state. Explain what each directory is for:
 Ask:
 
 > Reseed succeeded. Do you want me to delete the local temp tracking mirror,
-> this reseed's prepared staging records, source downloads, and reseed backup
-> under `/tmp`? These files are local safety/audit artifacts only; HF already
-> has the uploaded records.
+> source downloads, and reseed backup under `/tmp`? These files are local
+> safety/audit artifacts only; HF already has the uploaded records.
 >
 > Reply `cleanup reseed temp` to delete them, anything else to keep them.
 
 Do not delete anything unless the user replies exactly
 `cleanup reseed temp`. If cleanup is requested, remove only the specific
-directories and prepared record paths created for this reseed. Do not remove
-the shared staging root when it contains other records. Remove this operation's
-identity reservation only with its prepared records and manifest, and never
-remove unrelated `/tmp` contents.
+directories created for this reseed. Never remove unrelated `/tmp` contents.
 
 ## Failure modes and handling
 
@@ -586,34 +436,19 @@ remove unrelated `/tmp` contents.
   against the source batch median by more than `max_intra_batch_regression`.
   Ask for cleaner sources or a reviewed explanation before continuing.
 - **Too few source records to move the median.** Continue only after making
-  clear that one or two records may not immediately move an existing last-5
-  median. This warning does not block a first v2 calibration seed for an exact
-  identity with no eligible baseline yet.
+  clear that one or two records may not immediately move the last-5 median.
 - **The source results are noisy or suspicious.** Stop. Reseeding amplifies
   those measurements into the baseline, so they must be reviewed first.
 - **HF sync fails.** Stop for destructive reseeds. A stale or empty sync can
   make the old baseline look missing.
-- **The exact v2 identity already has an eligible baseline.** Stop. The
-  `CALIBRATION_NEEDED` artifact is stale; use the reviewed baseline-shift path
-  instead of the first-seed utility.
-- **The workload/variant/version trusts another recipe.** Stop. The source is
-  stale relative to the current recipe cohort and must not bypass
-  `RECIPE_MISMATCH` by creating a second trusted recipe.
-- **The staging root already has a prepared seed for the exact identity.**
-  Stop and reuse, upload, or explicitly clean that preparation. Do not prepare
-  another copy of the same measurement.
-- **The conditional Hub commit loses its parent race.** Stop without retrying.
-  Keep the preparation, refresh and review the new remote state, then request
-  a new explicit `upload` only if the seed is still valid.
 - **Candidate still violates fixed thresholds.** Report that this skill only
   handles the rolling HF baseline; update benchmark JSON thresholds in code
   review if maintainers accept the new absolute limit.
 - **The user aborts at either confirmation.** Leave the backup and prepared
   records on disk. Nothing should be uploaded.
-- **The user declines cleanup.** Keep `/tmp/perf-tracking`, the prepared seed
-  records under `/tmp/performance_reseed_prepared`, the source download
-  directory if any, and `/tmp/performance_reseed_backup/<...>` in place for
-  audit/debugging.
+- **The user declines cleanup.** Keep `/tmp/perf-tracking`, the source
+  download directory if any, and `/tmp/performance_reseed_backup/<...>` in
+  place for audit/debugging.
 - **A bad seed was uploaded.** Use the backup and HF history to identify the
   uploaded file, then remove or supersede it with an explicitly reviewed
   corrective record. Do not silently rewrite unrelated history.
@@ -624,9 +459,8 @@ remove unrelated `/tmp` contents.
   intentional baseline replacement.
 - `fastvideo/tests/performance/compare_baseline.py` — normalization, rolling
   median comparison, and persistence rules.
-- `fastvideo/performance/hf_store.py` — HF sync and record loading helpers.
-- `fastvideo/tests/performance/seed_baseline.py` — first-seed preparation,
-  staging reservation, manifest validation, and conditional batch upload.
+- `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/*.json` — fixed absolute benchmark
@@ -639,4 +473,3 @@ remove unrelated `/tmp` contents.
 | 2026-05-03 | Initial version. Sister workflow to `reseed-ssim-references`, scoped to one performance `(model_id, gpu_type)` baseline seed with backup, confirmation, provenance, and `success=true` upload. |
 | 2026-05-03 | Previous policy: replicate one approved shifted source result into 3 success records by default, or 5 only when explicitly requested. Add provenance marker for replicated-source reseeds. Superseded by the 2026-05-08 dynamic multi-source policy. |
 | 2026-05-08 | Replace fixed 3/5 replication with dynamic multi-source reseeding: upload one seed record per reviewed source JSON, validate intra-batch consistency, move backup/source scratch under `/tmp`, and ask whether to clean temp state after successful upload. |
-| 2026-07-13 | Keep first-v2-seed preparation outside the canonical mirror, reserve staging identities atomically, reject stale or replayed calibration seeds, and upload reviewed manifests with a single parent-guarded Hub commit. |

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

@@ -3,7 +3,7 @@
   "config_schema_version": 2,
   "workload_id": "wan-t2v",
   "variant_id": "1.3b-sp2",
-  "benchmark_version": 3,
+  "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

@@ -436,7 +436,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile"
               config:
-                command: "timeout 25m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
                 label: ":test_tube: LoRA Training Tests"
                 env:
                   - TEST_TYPE=training_lora
@@ -481,7 +481,6 @@ steps:
                 - "fastvideo/layers/**"
                 - "fastvideo/worker/**"
                 - "fastvideo/entrypoints/**"
-                - "fastvideo/performance/**"
                 - "fastvideo/tests/performance/**"
                 - ".buildkite/performance-benchmarks/**"
                 - "pyproject.toml"

.buildkite/scripts/pr_test.sh

@@ -76,27 +76,10 @@ EFFECTIVE_PR=${BUILDKITE_PULL_REQUEST:-false}
 if [ "$EFFECTIVE_PR" = "false" ] && [ -n "${PR_NUMBER:-}" ]; then
     EFFECTIVE_PR=$PR_NUMBER
 fi
-MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR BUILDKITE_BRANCH=${BUILDKITE_BRANCH:-} BUILDKITE_SOURCE=${BUILDKITE_SOURCE:-} TEST_SCOPE=${TEST_SCOPE:-} BUILDKITE_BUILD_URL=${BUILDKITE_BUILD_URL:-} BUILDKITE_BUILD_ID=${BUILDKITE_BUILD_ID:-} BUILDKITE_JOB_ID=${BUILDKITE_JOB_ID:-} IMAGE_VERSION=$IMAGE_VERSION"
+MODAL_ENV="BUILDKITE_REPO=$BUILDKITE_REPO BUILDKITE_COMMIT=$BUILDKITE_COMMIT BUILDKITE_PULL_REQUEST=$EFFECTIVE_PR BUILDKITE_BRANCH=${BUILDKITE_BRANCH:-} TEST_SCOPE=${TEST_SCOPE:-} BUILDKITE_BUILD_URL=${BUILDKITE_BUILD_URL:-} BUILDKITE_BUILD_ID=${BUILDKITE_BUILD_ID:-} BUILDKITE_JOB_ID=${BUILDKITE_JOB_ID:-} IMAGE_VERSION=$IMAGE_VERSION"
 
 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=${BUILDKITE_COMMIT:0:7}
     LOCAL_DIR="downloaded_reports"
@@ -189,12 +172,7 @@ case "$TEST_TYPE" in
         ;;
     "ssim")
         log "Running 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"
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_SSIM_TEST_FILE::run_ssim_tests"
         ;;
     "training")
         log "Running training tests..."

.github/scripts/gate_full_suite.sh

@@ -1,133 +0,0 @@
-#!/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

@@ -1,122 +0,0 @@
-#!/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

@@ -47,6 +47,3 @@ 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

@@ -52,7 +52,6 @@ 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'
@@ -61,7 +60,6 @@ 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: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -73,7 +71,6 @@ jobs:
               --arg commit "$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,
@@ -83,12 +80,11 @@ jobs:
                 pull_request_id: $pr_id,
                 pull_request_base_branch: "main",
                 env: {
-                    TEST_SCOPE: "full",
-                    FULL_SUITE: "true",
-                    PR_NUMBER: ($pr_id | tostring),
-                    PR_TITLE: $pr_title
-                  }
-                }')"
+                  TEST_SCOPE: "full",
+                  FULL_SUITE: "true",
+                  PR_NUMBER: ($pr_id | tostring)
+                }
+              }')"
 
   parse-command:
     if: >-
@@ -245,7 +241,6 @@ 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: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -262,7 +257,6 @@ jobs:
               --arg full_suite "$FULL_SUITE" \
               --arg test_type "$TEST_TYPE" \
               --arg pr_number "$PR_NUMBER" \
-              --arg pr_title "$PR_TITLE" \
               '{
                 commit: $commit,
                 branch: $branch,
@@ -272,9 +266,8 @@ jobs:
                 pull_request_base_branch: "main",
                 env: {
                   TEST_SCOPE: $test_scope,
-                    FULL_SUITE: $full_suite,
-                    TEST_TYPE: $test_type,
-                    PR_NUMBER: $pr_number,
-                    PR_TITLE: $pr_title
-                  }
-                }')"
+                  FULL_SUITE: $full_suite,
+                  TEST_TYPE: $test_type,
+                  PR_NUMBER: $pr_number
+                }
+              }')"

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

@@ -7,7 +7,6 @@ on:
 permissions:
   contents: read
   pull-requests: read
-  actions: read
 
 concurrency:
   group: full-suite-${{ github.event.pull_request.number }}
@@ -19,8 +18,6 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'ready')
       || github.event.action == 'synchronize'
     runs-on: ubuntu-latest
-    # Gate below may wait for cheap checks (up to MAX_WAIT_SECS = 25 min).
-    timeout-minutes: 35
     steps:
       - name: Check ready label
         id: check
@@ -52,20 +49,6 @@ jobs:
               "https://api.buildkite.com/v2/organizations/${{ 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:
@@ -73,7 +56,6 @@ 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: ${{ vars.BUILDKITE_PIPELINE_SLUG }}
         run: |
@@ -85,7 +67,6 @@ jobs:
               --arg commit "$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: $commit,
@@ -97,7 +78,6 @@ jobs:
                 env: {
                   TEST_SCOPE: "full",
                   FULL_SUITE: "true",
-                  PR_NUMBER: ($pr_id | tostring),
-                  PR_TITLE: $pr_title
+                  PR_NUMBER: ($pr_id | tostring)
                 }
               }')"

.gitignore

@@ -72,7 +72,8 @@ docs/distillation/examples/
 # Python pickle files
 *.pkl
 
-# Reference videos (negations must come after the catch-all on line below)
+# Reference videos
+!fastvideo/tests/ssim/reference_videos/**/*.mp4
 
 # Static images
 !docs/assets/images/**/*.png
@@ -126,8 +127,6 @@ 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/dreamverse/config.py

@@ -67,7 +67,7 @@ MODEL_REGISTRY = {
     },
 }
 
-DEFAULT_MODEL_ID = "fast-ltx23"
+DEFAULT_MODEL_ID = "fast-ltx2"
 
 ACTIVE_MODEL_ID = (os.getenv("DREAMVERSE_MODEL_ID", "").strip() or DEFAULT_MODEL_ID)
 if ACTIVE_MODEL_ID not in MODEL_REGISTRY:
@@ -76,11 +76,14 @@ if ACTIVE_MODEL_ID not in MODEL_REGISTRY:
 # Active model configuration
 MODEL_CONFIG = MODEL_REGISTRY[ACTIVE_MODEL_ID]
 
+# Generation limits
+SESSION_TIMEOUT_SECONDS = 300
+
 # Frame settings
 NUM_FRAMES = 121
 FRAME_HEIGHT = 1088
 FRAME_WIDTH = 1920
-NUM_INFERENCE_STEPS = 6
+NUM_INFERENCE_STEPS = 5
 JPEG_QUALITY = 100
 BATCH_SIZE = 3
 
@@ -165,9 +168,6 @@ def _optional_env(*names: str) -> str | None:
     return None
 
 
-# Generation limits
-SESSION_TIMEOUT_SECONDS = _env_int("DREAMVERSE_SESSION_TIMEOUT_SECONDS", 300)
-
 DEVTOOLS_ENABLED = _env_bool("FASTVIDEO_ENABLE_DEVTOOLS", False)
 PROMPT_SAFETY_ENABLED = _env_bool("FASTVIDEO_ENABLE_PROMPT_SAFETY", False)
 DREAMVERSE_MAX_AUTOTUNE = _env_bool("DREAMVERSE_MAX_AUTOTUNE", True)

apps/dreamverse/dreamverse/gpu_pool.py

@@ -1023,11 +1023,7 @@ def get_available_gpus() -> list[int]:
     """Get list of available GPU IDs from environment or auto-detect."""
     cuda_visible = os.environ.get("CUDA_VISIBLE_DEVICES", "")
     if cuda_visible:
-        try:
-            visible_gpu_ids = [int(x.strip()) for x in cuda_visible.split(",") if x.strip()]
-        except ValueError as exc:
-            raise RuntimeError("CUDA_VISIBLE_DEVICES must be a comma-separated list of integer GPU "
-                               f"indices (got {cuda_visible!r}); GPU UUIDs are not supported.") from exc
+        visible_gpu_ids = [int(x.strip()) for x in cuda_visible.split(",") if x.strip()]
         return _limit_gpu_ids(visible_gpu_ids)
 
     # Auto-detect available GPUs

apps/dreamverse/dreamverse/main.py

@@ -206,9 +206,7 @@ def cli() -> None:
     args = parser.parse_args()
 
     _install_heartbeat_log_filter()
-    # A 15MB init image (session_init_image.MAX_SESSION_INIT_IMAGE_BYTES) is ~20MB
-    # as a base64 ws message, above uvicorn's default 16MiB frame cap.
-    uvicorn.run(app, host=args.host, port=args.port, ws_max_size=32 * 1024 * 1024)
+    uvicorn.run(app, host=args.host, port=args.port)
 
 
 if __name__ == "__main__":

apps/dreamverse/dreamverse/mock_server.py

@@ -30,11 +30,11 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 
 from dreamverse._deps import require_dreamverse_runtime_deps
-from dreamverse.config import FRONTEND_STATIC_DIR_CANDIDATES, GENERATION_SEGMENT_CAP, SESSION_TIMEOUT_SECONDS
+from dreamverse.config import FRONTEND_STATIC_DIR_CANDIDATES, GENERATION_SEGMENT_CAP
 from dreamverse.session_init_image import cleanup_session_init_image, persist_session_init_image
-from dreamverse.utils import _resolve_generation_segment_cap
 
 LATENCY_MS = 200
+SESSION_TIMEOUT_SECONDS = 300
 MOCK_FRAME_WIDTH = 640
 MOCK_FRAME_HEIGHT = 352
 MOCK_FPS = 24
@@ -334,7 +334,6 @@ async def websocket_endpoint(websocket: WebSocket):
         auto_extension_enabled = bool(init_data.get("auto_extension_enabled", False))
         loop_generation_enabled = bool(init_data.get("loop_generation_enabled", False))
         single_clip_mode = bool(init_data.get("single_clip_mode", False))
-        manual_continuation_mode = bool(init_data.get("manual_continuation_mode", False))
         generation_paused = False
 
         if init_type == "session_init_v2":
@@ -345,8 +344,7 @@ async def websocket_endpoint(websocket: WebSocket):
             incoming_prompts = []
 
         curated_prompts = [prompt.strip() for prompt in incoming_prompts if isinstance(prompt, str) and prompt.strip()]
-        generation_paused = bool(not manual_continuation_mode and initial_rollout_prompt and not single_clip_mode
-                                 and len(curated_prompts) == 0)
+        generation_paused = bool(initial_rollout_prompt and not single_clip_mode and len(curated_prompts) == 0)
 
         try:
             session_init_image = persist_session_init_image(init_data.get("initial_image"))
@@ -375,10 +373,7 @@ async def websocket_endpoint(websocket: WebSocket):
         prompt_sources_blocked = False
         pending_seed_reset = False
         pending_seed_reset_reason = ""
-        pending_simple_submission: PromptSubmission | None = (PromptSubmission(
-            prompt_id=str(init_data.get("initial_rollout_prompt_id") or uuid.uuid4()),
-            raw_prompt=initial_rollout_prompt,
-        ) if manual_continuation_mode and initial_rollout_prompt else None)
+        pending_simple_submission: PromptSubmission | None = None
         single_clip_waiting_for_request = False
         rollout_waiting_for_rewrite = False
         initial_rollout_waiting_for_rewrite = generation_paused
@@ -398,26 +393,14 @@ async def websocket_endpoint(websocket: WebSocket):
 
         async def send_stream_start(seed_reason: str) -> None:
             await ws_send_json({
-                "type":
-                "ltx2_stream_start",
-                "total_segments":
-                len(curated_prompts),
-                "preset_id":
-                preset_id,
-                "stream_mode":
-                "av_fmp4",
-                "live_mode":
-                True,
-                "loop_generation_enabled":
-                loop_generation_enabled,
-                "loop_iteration":
-                loop_iteration,
-                "generation_segment_cap":
-                _resolve_generation_segment_cap(
-                    single_clip_mode=single_clip_mode,
-                    cap=GENERATION_SEGMENT_CAP,
-                    manual_continuation_mode=manual_continuation_mode,
-                ),
+                "type": "ltx2_stream_start",
+                "total_segments": len(curated_prompts),
+                "preset_id": preset_id,
+                "stream_mode": "av_fmp4",
+                "live_mode": True,
+                "loop_generation_enabled": loop_generation_enabled,
+                "loop_iteration": loop_iteration,
+                "generation_segment_cap": 0,
             })
             if seed_reason == "init":
                 await ws_send_json({
@@ -510,7 +493,6 @@ async def websocket_endpoint(websocket: WebSocket):
             nonlocal auto_extension_enabled
             nonlocal loop_generation_enabled
             nonlocal single_clip_mode
-            nonlocal manual_continuation_mode
             nonlocal generation_paused
             nonlocal seed_prompt_memory
             nonlocal curated_prompts
@@ -555,7 +537,6 @@ async def websocket_endpoint(websocket: WebSocket):
             auto_extension_enabled = bool(payload.get("auto_extension_enabled", False))
             loop_generation_enabled = bool(payload.get("loop_generation_enabled", False))
             single_clip_mode = bool(payload.get("single_clip_mode", False))
-            manual_continuation_mode = bool(payload.get("manual_continuation_mode", False))
 
             seed_prompt_memory = list(next_curated_prompts)
             curated_prompts = list(seed_prompt_memory)
@@ -563,14 +544,10 @@ async def websocket_endpoint(websocket: WebSocket):
             segment_idx = 0
             pending_seed_reset = False
             pending_seed_reset_reason = ""
-            pending_simple_submission = (PromptSubmission(
-                prompt_id=str(payload.get("initial_rollout_prompt_id") or uuid.uuid4()),
-                raw_prompt=initial_rollout_prompt,
-            ) if manual_continuation_mode and initial_rollout_prompt else None)
+            pending_simple_submission = None
             single_clip_waiting_for_request = False
             rollout_waiting_for_rewrite = False
-            generation_paused = bool(not manual_continuation_mode and initial_rollout_prompt and not single_clip_mode
-                                     and len(curated_prompts) == 0)
+            generation_paused = bool(initial_rollout_prompt and not single_clip_mode and len(curated_prompts) == 0)
             initial_rollout_waiting_for_rewrite = generation_paused
             rewrite_restart_pending = False
             loop_iteration = 0
@@ -991,11 +968,10 @@ async def websocket_endpoint(websocket: WebSocket):
                     project_stream_started = True
                     await send_stream_start(pending_seed_reset_reason)
                     pending_seed_reset_reason = ""
-
-                if pending_simple_submission is not None:
-                    submission = pending_simple_submission
-                    pending_simple_submission = None
-                    await promote_submission_to_ready(submission)
+                    if pending_simple_submission is not None:
+                        submission = pending_simple_submission
+                        pending_simple_submission = None
+                        await promote_submission_to_ready(submission)
 
                 if generation_paused:
                     await asyncio.sleep(0.05)
@@ -1005,8 +981,8 @@ async def websocket_endpoint(websocket: WebSocket):
                     await asyncio.sleep(0.05)
                     continue
 
-                if (not single_clip_mode and not manual_continuation_mode and not rollout_waiting_for_rewrite
-                        and GENERATION_SEGMENT_CAP > 0 and segment_idx >= GENERATION_SEGMENT_CAP):
+                if (not single_clip_mode and not rollout_waiting_for_rewrite and GENERATION_SEGMENT_CAP > 0
+                        and segment_idx >= GENERATION_SEGMENT_CAP):
                     rollout_waiting_for_rewrite = True
                     loop_generation_enabled = False
                     project_stream_started = False
@@ -1241,9 +1217,7 @@ def cli() -> None:
     print(f"Starting mock server with {LATENCY_MS}ms latency on port {args.port}")
 
     _install_heartbeat_log_filter()
-    # A 15MB init image (session_init_image.MAX_SESSION_INIT_IMAGE_BYTES) is ~20MB
-    # as a base64 ws message, above uvicorn's default 16MiB frame cap.
-    uvicorn.run(app, host="0.0.0.0", port=args.port, ws_max_size=32 * 1024 * 1024)
+    uvicorn.run(app, host="0.0.0.0", port=args.port)
 
 
 if __name__ == "__main__":

apps/dreamverse/dreamverse/prompt_enhancer.py

@@ -313,33 +313,6 @@ def _extract_content_or_empty(response_json: dict[str, Any]) -> str:
         return ""
 
 
-def _find_balanced_object_end(text: str, start: int) -> int:
-    """Return the index just past the brace-balanced span opening at
-    ``text[start] == '{'``, honoring JSON string literals and escapes, or -1
-    if the braces never balance (i.e. the object was truncated)."""
-    depth = 0
-    in_string = False
-    escaped = False
-    for i in range(start, len(text)):
-        ch = text[i]
-        if in_string:
-            if escaped:
-                escaped = False
-            elif ch == "\\":
-                escaped = True
-            elif ch == '"':
-                in_string = False
-        elif ch == '"':
-            in_string = True
-        elif ch == "{":
-            depth += 1
-        elif ch == "}":
-            depth -= 1
-            if depth == 0:
-                return i + 1
-    return -1
-
-
 def _parse_json_response(content: str) -> dict[str, Any]:
     text = content.strip()
     if not text:
@@ -357,7 +330,6 @@ def _parse_json_response(content: str) -> dict[str, Any]:
         r"```(?:json)?\s*([\s\S]*?)```",
         flags=re.IGNORECASE,
     )
-    last_fenced: dict[str, Any] | None = None
     for match in fence_pattern.finditer(text):
         block = match.group(1).strip()
         if not block:
@@ -365,37 +337,21 @@ def _parse_json_response(content: str) -> dict[str, Any]:
         try:
             parsed = json.loads(block)
             if isinstance(parsed, dict):
-                last_fenced = parsed
+                return parsed
         except json.JSONDecodeError:
             continue
-    if last_fenced is not None:
-        return last_fenced
 
-    # Scan for all decodable JSON objects and return the last — chain-of-thought
-    # models emit draft JSON mid-reasoning; the final answer is always last.
+    # Fall back to scanning for the first decodable JSON object in free-form text.
     decoder = json.JSONDecoder()
-    last_parsed: dict[str, Any] | None = None
-    pos = 0
-    while (idx := text.find("{", pos)) != -1:
+    for idx, char in enumerate(text):
+        if char != "{":
+            continue
         try:
-            parsed, consumed = decoder.raw_decode(text[idx:])
+            parsed, _ = decoder.raw_decode(text[idx:])
         except json.JSONDecodeError:
-            # Skip the whole failed object rather than rescanning inside it:
-            # fragments nested in a malformed or truncated (finish_reason=
-            # length) object must not override an earlier complete object.
-            span_end = _find_balanced_object_end(text, idx)
-            if span_end == -1:
-                break
-            pos = span_end
             continue
-        # Skip past the consumed span so nested braces inside a decoded
-        # object are not re-parsed as standalone objects.
-        pos = idx + consumed
         if isinstance(parsed, dict):
-            last_parsed = parsed
-
-    if last_parsed is not None:
-        return last_parsed
+            return parsed
 
     raise ValueError("No JSON object found in assistant response.")
 
@@ -425,7 +381,7 @@ def _format_locked_segments(locked_segments: list[str]) -> str:
 
 class PromptEnhancer:
 
-    def __init__(self) -> None:
+    def __init__(self):
         self.provider = PROMPT_PROVIDER
         self.provider_label = _resolve_provider_label(PROMPT_PROVIDER)
         self.api_key = PROMPT_API_KEY
@@ -1461,12 +1417,15 @@ class PromptEnhancer:
             locked_text = _format_locked_segments(locked_segments_clean)
             request_system_prompt = self.enhance_system_prompt
             user_payload = {
-                "request": ("<locked_segments>\n"
-                            f"{locked_text}\n"
-                            "</locked_segments>\n\n"
-                            f"<conditioning_prompt>{cleaned}</conditioning_prompt>\n\n"
-                            f"Write exactly one new segment ({next_segment_key}) "
-                            "continuing from the locked segments."),
+                "request": (
+                    "<locked_segments>\n"
+                    f"{locked_text}\n"
+                    "</locked_segments>\n\n"
+                    f"<conditioning_prompt>{cleaned}</conditioning_prompt>\n\n"
+                    f"Write exactly one new segment ({next_segment_key}) "
+                    "continuing from the locked segments. "
+                    'Respond with valid JSON only as {"next_prompt": "..."}.'  # noqa: E501
+                ),
             }
 
         t0 = time.perf_counter()
@@ -1498,7 +1457,6 @@ class PromptEnhancer:
                             body=request_body,
                             timeout_seconds=timeout_seconds,
                         )
-                        _enhance_print("INFO", f"raw_response: {response_content}")
                         if is_single_clip_mode:
                             prompt = self._extract_single_clip_prompt(response_content)
                         else:
@@ -1581,7 +1539,7 @@ class PromptEnhancer:
                 f"Write exactly one new segment ({next_segment_key}) "
                 "that continues linearly from the locked segments. "
                 "Infer the next narrative beat from this history. "
-                'Respond with valid JSON only: {"next_prompt": "<your segment description here>"}.'  # noqa: E501
+                'Respond with valid JSON only as {"next_prompt": "..."}.'  # noqa: E501
             ),
         }
 

apps/dreamverse/dreamverse/session/controller.py

@@ -20,7 +20,6 @@ from __future__ import annotations
 # mypy: ignore-errors
 
 import asyncio
-import os
 import time
 import uuid
 from typing import TYPE_CHECKING
@@ -53,14 +52,6 @@ if TYPE_CHECKING:
     from dreamverse.prompt_enhancer import PromptEnhancer
     from dreamverse.prompt_safety import PromptSafetyFilter
 
-# Optional append-only log of every generated segment prompt; unset disables it.
-SEGMENT_PROMPT_LOG_PATH = os.environ.get("DREAMVERSE_SEGMENT_PROMPT_LOG", "")
-
-
-def _append_segment_prompt_log(path: str, text: str) -> None:
-    with open(path, "a") as f:
-        f.write(text)
-
 
 class SessionController:
     """Runs one WebSocket session from accept() through disconnect."""
@@ -207,7 +198,6 @@ class SessionController:
             auto_extension_enabled = bool(init_data.get("auto_extension_enabled", False))
             loop_generation_enabled = bool(init_data.get("loop_generation_enabled", False))
             single_clip_mode = bool(init_data.get("single_clip_mode", False))
-            manual_continuation_mode = bool(init_data.get("manual_continuation_mode", False))
             rewrite_model = self.prompt_enhancer.resolve_rewrite_model(init_data.get("rewrite_model"))
             rewrite_system_prompt_override = str(init_data.get("rewrite_window_system_prompt") or "").strip()
             rewrite_user_system_prompt_override = str(init_data.get("rewrite_user_system_prompt") or "").strip()
@@ -292,9 +282,6 @@ class SessionController:
             # Session queues and mutable state.
             raw_prompt_queue: asyncio.Queue[PromptSubmission] = asyncio.Queue()
             ready_prompt_queue: asyncio.Queue[ReadyPrompt] = asyncio.Queue()
-            # Submissions dequeued by prompt_worker_loop but not yet resolved; while
-            # non-zero the prompt sources are busy, not drained.
-            prompt_enhancement_inflight = 0
 
             curated_idx = 0
             segment_idx = 0
@@ -304,20 +291,13 @@ class SessionController:
             generation_cap_blocked = False
             auto_extension_blocked_segment_idx: int | None = None
             prompt_sources_drained_logged = False
-            generation_paused = bool(not manual_continuation_mode and initial_rollout_prompt and not single_clip_mode
-                                     and len(curated_prompts) == 0)
+            generation_paused = bool(initial_rollout_prompt and not single_clip_mode and len(curated_prompts) == 0)
             pending_seed_reset = False
             pending_seed_reset_reason = ""
             pending_reset_conditioning = False
             loop_iteration = 0 if generation_paused else 1
             force_curated_restart_segment = False
-            # The frontend records the opening scene under this id; reuse it so
-            # prompt lifecycle events for the opening prompt reach that record.
-            pending_simple_prompt_submission: PromptSubmission | None = (PromptSubmission(
-                prompt_id=str(init_data.get("initial_rollout_prompt_id") or uuid.uuid4()),
-                raw_prompt=initial_rollout_prompt,
-                created_at_s=time.time(),
-            ) if manual_continuation_mode and initial_rollout_prompt else None)
+            pending_simple_prompt_submission: PromptSubmission | None = None
             single_clip_waiting_for_request = False
             rollout_waiting_for_rewrite = False
             initial_rollout_waiting_for_rewrite = generation_paused
@@ -326,7 +306,6 @@ class SessionController:
             project_active = True
             project_stream_started = False
             pending_project_end = False
-            segment_prompt_log_warned = False
 
             def replace_session_init_image(initial_image_payload: object) -> None:
                 nonlocal session_init_image
@@ -445,7 +424,6 @@ class SessionController:
                 nonlocal auto_extension_enabled
                 nonlocal loop_generation_enabled
                 nonlocal single_clip_mode
-                nonlocal manual_continuation_mode
                 nonlocal generation_paused
                 nonlocal curated_prompts
                 nonlocal seed_prompt_memory
@@ -480,7 +458,6 @@ class SessionController:
                 next_auto_extension_enabled = bool(payload.get("auto_extension_enabled", False))
                 next_loop_generation_enabled = bool(payload.get("loop_generation_enabled", False))
                 next_single_clip_mode = bool(payload.get("single_clip_mode", False))
-                next_manual_continuation_mode = bool(payload.get("manual_continuation_mode", False))
 
                 next_preset_id = str(payload.get("preset_id") or "").strip()
                 if next_preset_id:
@@ -533,7 +510,6 @@ class SessionController:
                 auto_extension_enabled = next_auto_extension_enabled
                 loop_generation_enabled = next_loop_generation_enabled
                 single_clip_mode = next_single_clip_mode
-                manual_continuation_mode = next_manual_continuation_mode
                 rewrite_model = next_rewrite_model
                 rewrite_system_prompt_override = (next_rewrite_system_prompt_override)
                 rewrite_user_system_prompt_override = (next_rewrite_user_system_prompt_override)
@@ -554,15 +530,10 @@ class SessionController:
                 generation_cap_blocked = False
                 auto_extension_blocked_segment_idx = None
                 prompt_sources_drained_logged = False
-                pending_simple_prompt_submission = (PromptSubmission(
-                    prompt_id=str(payload.get("initial_rollout_prompt_id") or uuid.uuid4()),
-                    raw_prompt=initial_rollout_prompt,
-                    created_at_s=time.time(),
-                ) if manual_continuation_mode and initial_rollout_prompt else None)
+                pending_simple_prompt_submission = None
                 single_clip_waiting_for_request = False
                 rollout_waiting_for_rewrite = False
-                generation_paused = bool(not manual_continuation_mode and initial_rollout_prompt
-                                         and not single_clip_mode and len(curated_prompts) == 0)
+                generation_paused = bool(initial_rollout_prompt and not single_clip_mode and len(curated_prompts) == 0)
                 initial_rollout_waiting_for_rewrite = generation_paused
                 rewrite_restart_pending = False
                 loop_iteration = 0
@@ -971,7 +942,6 @@ class SessionController:
                             _resolve_generation_segment_cap(
                                 single_clip_mode=single_clip_mode,
                                 cap=GENERATION_SEGMENT_CAP,
-                                manual_continuation_mode=manual_continuation_mode,
                             ),
                         })
                         continue
@@ -1035,9 +1005,11 @@ class SessionController:
                         continue
 
             async def prompt_worker_loop():
-                nonlocal prompt_enhancement_inflight
-
-                async def process_submission(submission: PromptSubmission) -> None:
+                while not stop_event.is_set():
+                    try:
+                        submission = await asyncio.wait_for(raw_prompt_queue.get(), timeout=0.1)
+                    except asyncio.TimeoutError:
+                        continue
                     _main_print("INFO", f"Received user prompt for enhancement: {submission.raw_prompt}")
                     prompt_id = submission.prompt_id
                     raw_prompt = submission.raw_prompt
@@ -1060,9 +1032,8 @@ class SessionController:
                         await ws_send_json({
                             "type": "error",
                             "message": blocked_raw_prompt_error,
-                            "prompt_id": prompt_id,
                         })
-                        return
+                        continue
                     await log_event(
                         "enhance_request",
                         {
@@ -1122,9 +1093,8 @@ class SessionController:
                             await ws_send_json({
                                 "type": "error",
                                 "message": blocked_final_prompt_error,
-                                "prompt_id": prompt_id,
                             })
-                            return
+                            continue
                         if result.fallback_used or not final_prompt:
                             source = "user_enhancement_failed"
                             _main_print(
@@ -1143,7 +1113,7 @@ class SessionController:
                             })
                             # Enhancement is strict JSON-only; do not enqueue raw
                             # prompt when enhancement fails.
-                            return
+                            continue
                         else:
                             source = "user_enhanced"
                             await ws_send_json({
@@ -1160,7 +1130,6 @@ class SessionController:
                                 source=source,
                                 fallback_used=result.fallback_used,
                                 loop_iteration=loop_iteration,
-                                raw_prompt=raw_prompt,
                             ))
                     else:
                         await ready_prompt_queue.put(
@@ -1170,7 +1139,6 @@ class SessionController:
                                 source="user_raw",
                                 fallback_used=False,
                                 loop_iteration=loop_iteration,
-                                raw_prompt=raw_prompt,
                             ))
                         await ws_send_json({
                             "type": "prompt_ready",
@@ -1180,21 +1148,6 @@ class SessionController:
                             "latency_ms": 0.0,
                         })
 
-                while not stop_event.is_set():
-                    try:
-                        submission = raw_prompt_queue.get_nowait()
-                    except asyncio.QueueEmpty:
-                        await asyncio.sleep(0.1)
-                        continue
-                    # Dequeue and increment without an await in between so the
-                    # generation loop never sees an empty queue with zero in flight
-                    # while this submission is still being enhanced.
-                    prompt_enhancement_inflight += 1
-                    try:
-                        await process_submission(submission)
-                    finally:
-                        prompt_enhancement_inflight -= 1
-
             def queue_snapshot() -> dict[str, object]:
                 return {
                     "user_ready": ready_prompt_queue.qsize(),
@@ -1347,7 +1300,6 @@ class SessionController:
                     _resolve_generation_segment_cap(
                         single_clip_mode=single_clip_mode,
                         cap=GENERATION_SEGMENT_CAP,
-                        manual_continuation_mode=manual_continuation_mode,
                     ),
                 })
                 await ws_send_json({
@@ -1407,7 +1359,6 @@ class SessionController:
                         _resolve_generation_segment_cap(
                             single_clip_mode=single_clip_mode,
                             cap=GENERATION_SEGMENT_CAP,
-                            manual_continuation_mode=manual_continuation_mode,
                         ),
                     })
                     if nonlocal_reason == "loop_restart":
@@ -1437,9 +1388,8 @@ class SessionController:
                     await raw_prompt_queue.put(pending_simple_prompt_submission)
                     pending_simple_prompt_submission = None
 
-                if (not single_clip_mode and not manual_continuation_mode and not generation_cap_blocked
-                        and not rollout_waiting_for_rewrite and GENERATION_SEGMENT_CAP > 0
-                        and generated_segment_count >= GENERATION_SEGMENT_CAP):
+                if (not single_clip_mode and not generation_cap_blocked and not rollout_waiting_for_rewrite
+                        and GENERATION_SEGMENT_CAP > 0 and generated_segment_count >= GENERATION_SEGMENT_CAP):
                     loop_generation_enabled = False
                     rollout_waiting_for_rewrite = True
                     _main_print(
@@ -1592,10 +1542,7 @@ class SessionController:
                     if single_clip_mode:
                         await asyncio.sleep(PROMPT_AUTO_SLEEP_MS / 1000.0)
                         continue
-                    # A raw submission still queued or being enhanced will produce a
-                    # ready prompt shortly; that is not a drained/blocked state.
-                    enhancement_pending = (raw_prompt_queue.qsize() > 0 or prompt_enhancement_inflight > 0)
-                    if not prompt_sources_drained_logged and not enhancement_pending:
+                    if not prompt_sources_drained_logged:
                         snapshot = queue_snapshot()
                         _main_print(
                             "WARN",
@@ -1634,24 +1581,6 @@ class SessionController:
                 total_segments_hint = max(segment_idx, len(curated_prompts))
                 prompt = selected.prompt
                 locked_segment_prompts.append(prompt)
-                if SEGMENT_PROMPT_LOG_PATH:
-                    _ts = time.strftime("%Y-%m-%d %H:%M:%S")
-                    _lines = [
-                        f"\n=== Segment {segment_idx} [{_ts}] source={selected.source} client={client_id[:8]} ===",
-                    ]
-                    if selected.raw_prompt and selected.raw_prompt != prompt:
-                        _lines.append(f"User:     {selected.raw_prompt}")
-                    _lines.append(f"Rewritten: {prompt}")
-                    try:
-                        await asyncio.to_thread(_append_segment_prompt_log, SEGMENT_PROMPT_LOG_PATH,
-                                                "\n".join(_lines) + "\n")
-                    except Exception as exc:
-                        if not segment_prompt_log_warned:
-                            segment_prompt_log_warned = True
-                            _main_print(
-                                "WARN",
-                                f"Failed to write segment prompt log {SEGMENT_PROMPT_LOG_PATH}: {exc}",
-                            )
                 if (auto_extension_blocked_segment_idx is not None
                         and auto_extension_blocked_segment_idx <= segment_idx):
                     auto_extension_blocked_segment_idx = None

apps/dreamverse/dreamverse/session/messages.py

@@ -19,4 +19,3 @@ class ReadyPrompt:
     fallback_used: bool = False
     seed_prompt_index: int | None = None
     loop_iteration: int | None = None
-    raw_prompt: str | None = None

apps/dreamverse/dreamverse/tests/test_config.py

@@ -150,22 +150,12 @@ def test_config_enables_prompt_safety_when_requested(monkeypatch):
 
 def test_config_uses_five_minute_session_timeout(monkeypatch):
     _set_required_prompt_keys(monkeypatch)
-    monkeypatch.delenv("DREAMVERSE_SESSION_TIMEOUT_SECONDS", raising=False)
 
     module = _load_config_module()
 
     assert module.SESSION_TIMEOUT_SECONDS == 300
 
 
-def test_config_session_timeout_env_override(monkeypatch):
-    _set_required_prompt_keys(monkeypatch)
-    monkeypatch.setenv("DREAMVERSE_SESSION_TIMEOUT_SECONDS", "1800")
-
-    module = _load_config_module()
-
-    assert module.SESSION_TIMEOUT_SECONDS == 1800
-
-
 def test_config_rejects_invalid_prompt_provider(monkeypatch):
     monkeypatch.setenv("FASTVIDEO_PROMPT_PROVIDER", "unsupported")
     _set_required_prompt_keys(monkeypatch)

apps/dreamverse/dreamverse/tests/test_entrypoints.py

@@ -75,13 +75,12 @@ def _run_cli(module, monkeypatch, argv: list[str]) -> list[dict[str, object]]:
     calls: list[dict[str, object]] = []
     uvicorn_stub = types.ModuleType("uvicorn")
 
-    def run(app, host: str, port: int, **kwargs) -> None:
+    def run(app, host: str, port: int) -> None:
         calls.append(
             {
                 "app": app,
                 "host": host,
                 "port": port,
-                **kwargs,
             }
         )
 
@@ -105,7 +104,6 @@ def test_server_cli_defaults_to_local_web_port(monkeypatch):
             "app": server_main.app,
             "host": "0.0.0.0",
             "port": 8009,
-            "ws_max_size": 32 * 1024 * 1024,
         }
     ]
 
@@ -123,7 +121,6 @@ def test_server_cli_allows_explicit_host_and_port(monkeypatch):
             "app": server_main.app,
             "host": "127.0.0.1",
             "port": 8123,
-            "ws_max_size": 32 * 1024 * 1024,
         }
     ]
 
@@ -150,7 +147,6 @@ def test_mock_server_cli_defaults_to_local_web_port(monkeypatch):
             "app": mock_server.app,
             "host": "0.0.0.0",
             "port": 8009,
-            "ws_max_size": 32 * 1024 * 1024,
         }
     ]
 
@@ -170,7 +166,6 @@ def test_mock_server_cli_updates_latency(monkeypatch):
                 "app": mock_server.app,
                 "host": "0.0.0.0",
                 "port": 8111,
-                "ws_max_size": 32 * 1024 * 1024,
             }
         ]
         assert mock_server.LATENCY_MS == 321

apps/dreamverse/dreamverse/tests/test_mock_server.py

@@ -290,152 +290,6 @@ def test_mock_server_supports_initial_custom_rollout_prompt():
         mock_server.LATENCY_MS = old_latency_ms
 
 
-def test_mock_server_manual_mode_streams_initial_prompt_without_rewrite_or_cap():
-    old_segment_bytes = mock_server.MOCK_SEGMENT_BYTES
-    old_latency_ms = mock_server.LATENCY_MS
-    old_generation_segment_cap = mock_server.GENERATION_SEGMENT_CAP
-    try:
-        mock_server.MOCK_SEGMENT_BYTES = b"mock-fmp4-bytes"
-        mock_server.LATENCY_MS = 1
-        mock_server.GENERATION_SEGMENT_CAP = 1
-
-        ws = _FakeWebSocket(
-            [
-                (
-                    0.0,
-                    {
-                        "type": "session_init_v2",
-                        "preset_id": "custom_editable",
-                        "preset_label": "Custom rollout",
-                        "curated_prompts": [],
-                        "initial_rollout_prompt": "A drone skims a neon canyon",
-                        "initial_rollout_prompt_id": "steer-prompt-1",
-                        "manual_continuation_mode": True,
-                        "enhancement_enabled": True,
-                        "auto_extension_enabled": False,
-                        "loop_generation_enabled": False,
-                    },
-                ),
-                (
-                    0.08,
-                    {
-                        "type": "append_prompt",
-                        "prompt": "The drone dives toward the river",
-                        "prompt_id": "steer-prompt-2",
-                    },
-                ),
-                (0.30, {"type": "leave"}),
-            ]
-        )
-
-        asyncio.run(mock_server.websocket_endpoint(ws))
-
-        message_types = [payload["type"] for payload in ws.sent_json]
-        assert "rewrite_seed_prompts_started" not in message_types
-        assert "rewrite_seed_prompts_complete" not in message_types
-        assert "ltx2_stream_start" in message_types
-        # cap=1 must not stop a manual-mode session after the first segment
-        assert "ltx2_stream_complete" not in message_types
-
-        prompt_ready_events = [
-            payload for payload in ws.sent_json if payload["type"] == "prompt_ready"
-        ]
-        assert [payload["prompt_id"] for payload in prompt_ready_events] == [
-            "steer-prompt-1",
-            "steer-prompt-2",
-        ]
-
-        segment_start_events = [
-            payload
-            for payload in ws.sent_json
-            if payload["type"] == "ltx2_segment_start"
-        ]
-        assert [payload["prompt"] for payload in segment_start_events] == [
-            "A drone skims a neon canyon",
-            "The drone dives toward the river",
-        ]
-        segment_source_events = [
-            payload
-            for payload in ws.sent_json
-            if payload["type"] == "segment_prompt_source"
-        ]
-        assert [payload["prompt_id"] for payload in segment_source_events] == [
-            "steer-prompt-1",
-            "steer-prompt-2",
-        ]
-    finally:
-        mock_server.MOCK_SEGMENT_BYTES = old_segment_bytes
-        mock_server.LATENCY_MS = old_latency_ms
-        mock_server.GENERATION_SEGMENT_CAP = old_generation_segment_cap
-
-
-def test_mock_server_project_init_manual_mode_streams_initial_prompt():
-    old_segment_bytes = mock_server.MOCK_SEGMENT_BYTES
-    old_latency_ms = mock_server.LATENCY_MS
-    try:
-        mock_server.MOCK_SEGMENT_BYTES = b"mock-fmp4-bytes"
-        mock_server.LATENCY_MS = 1
-
-        ws = _FakeWebSocket(
-            [
-                (
-                    0.0,
-                    {
-                        "type": "session_init_v2",
-                        "preset_id": "test_preset",
-                        "curated_prompts": ["segment one"],
-                        "enhancement_enabled": True,
-                        "auto_extension_enabled": False,
-                        "loop_generation_enabled": False,
-                    },
-                ),
-                (0.05, {"type": "end_project_keep_session"}),
-                (
-                    0.15,
-                    {
-                        "type": "project_init_v1",
-                        "preset_id": "custom_editable",
-                        "preset_label": "Custom rollout",
-                        "curated_prompts": [],
-                        "initial_rollout_prompt": "A drone skims a neon canyon",
-                        "initial_rollout_prompt_id": "steer-prompt-1",
-                        "manual_continuation_mode": True,
-                        "enhancement_enabled": True,
-                        "auto_extension_enabled": False,
-                        "loop_generation_enabled": False,
-                    },
-                ),
-                (0.45, {"type": "leave"}),
-            ]
-        )
-
-        asyncio.run(mock_server.websocket_endpoint(ws))
-
-        message_types = [payload["type"] for payload in ws.sent_json]
-        assert "project_idle" in message_types
-        project_idle_index = message_types.index("project_idle")
-        # manual-mode restart must not run the rewrite rollout
-        assert "rewrite_seed_prompts_started" not in message_types[project_idle_index:]
-        assert "ltx2_stream_start" in message_types[project_idle_index:]
-
-        prompt_ready_events = [
-            payload for payload in ws.sent_json if payload["type"] == "prompt_ready"
-        ]
-        assert [payload["prompt_id"] for payload in prompt_ready_events] == [
-            "steer-prompt-1",
-        ]
-
-        segment_start_events = [
-            payload
-            for payload in ws.sent_json
-            if payload["type"] == "ltx2_segment_start"
-        ]
-        assert segment_start_events[-1]["prompt"] == "A drone skims a neon canyon"
-    finally:
-        mock_server.MOCK_SEGMENT_BYTES = old_segment_bytes
-        mock_server.LATENCY_MS = old_latency_ms
-
-
 def test_mock_server_can_start_new_project_without_reconnecting():
     old_segment_bytes = mock_server.MOCK_SEGMENT_BYTES
     old_latency_ms = mock_server.LATENCY_MS

apps/dreamverse/dreamverse/tests/test_prompt_enhancer.py

@@ -6,7 +6,6 @@ import os
 import re
 import time
 
-import pytest
 
 os.environ.setdefault("CEREBRAS_API_KEY", "dummy")
 os.environ.setdefault("GROQ_API_KEY", "dummy")
@@ -206,59 +205,6 @@ def test_parse_json_response_extracts_first_embedded_object():
     assert parsed == {"segment_prompts": ["A", "B"]}
 
 
-def test_parse_json_response_returns_outer_object_not_nested_value():
-    parsed = _parse_json_response(
-        "Final answer: {\"next_prompt\": \"a scene\", \"style\": {\"mood\": \"noir\"}} done."
-    )
-    assert parsed == {"next_prompt": "a scene", "style": {"mood": "noir"}}
-
-
-def test_parse_json_response_returns_last_of_multiple_objects():
-    parsed = _parse_json_response(
-        "Draft: {\"next_prompt\": \"draft\"}\nRefined: {\"next_prompt\": \"final\"}"
-    )
-    assert parsed == {"next_prompt": "final"}
-
-
-def test_parse_json_response_ignores_fragments_of_truncated_trailing_object():
-    # finish_reason=length cut the refined object short; the complete draft must
-    # win over a nested fragment of the truncated object.
-    parsed = _parse_json_response(
-        '{"next_prompt": "draft"} refined: {"next_prompt": "final", "style": {"mood": "noir"}'
-    )
-    assert parsed == {"next_prompt": "draft"}
-
-
-def test_parse_json_response_ignores_fragments_of_mid_string_truncated_object():
-    # Unterminated-string truncation reports the error at the opening quote,
-    # not end-of-text; nested fragments still must not win over the draft.
-    parsed = _parse_json_response(
-        '{"next_prompt": "draft"} refined: {"style": {"mood": "noir"}, "next_prompt": "cut off'
-    )
-    assert parsed == {"next_prompt": "draft"}
-
-
-def test_parse_json_response_ignores_fragments_of_malformed_object_with_trailing_prose():
-    parsed = _parse_json_response(
-        '{"next_prompt": "draft"} {"final": {"mood": "noir"}, "x": 1 and then some prose'
-    )
-    assert parsed == {"next_prompt": "draft"}
-
-
-def test_parse_json_response_raises_when_only_object_is_truncated():
-    with pytest.raises(ValueError):
-        _parse_json_response('{"style": {"mood": "noir"}, "next_prompt": "cut off')
-
-
-def test_parse_json_response_returns_outer_rollout_dict():
-    parsed = _parse_json_response(
-        "{\"rollout\": {\"segment_prompts\": [{\"prompt\": \"a\"}, {\"prompt\": \"b\"}]}}"
-    )
-    assert parsed == {
-        "rollout": {"segment_prompts": [{"prompt": "a"}, {"prompt": "b"}]}
-    }
-
-
 def test_load_prompt_required_falls_back_to_default_path(tmp_path):
     fallback_path = tmp_path / "next_segment_system_prompt.md"
     fallback_path.write_text("fallback prompt\n", encoding="utf-8")

apps/dreamverse/dreamverse/utils.py

@@ -15,6 +15,5 @@ def _utc_now_iso() -> str:
 PROMPT_EXTENSION_FAILURE_USER_MESSAGE = ("Prompt extension failed for this request.")
 
 
-def _resolve_generation_segment_cap(*, single_clip_mode: bool, cap: int, manual_continuation_mode: bool = False) -> int:
-    # Steering (manual continuation) lets the user keep going indefinitely, like single-clip mode.
-    return 0 if (single_clip_mode or manual_continuation_mode) else cap
+def _resolve_generation_segment_cap(*, single_clip_mode: bool, cap: int) -> int:
+    return 0 if single_clip_mode else cap

apps/dreamverse/dreamverse/video_generation.py

@@ -489,7 +489,7 @@ class VideoGenerationWorker:
             num_inference_steps=NUM_INFERENCE_STEPS,
             guidance_scale=1.0,
             seed=10,
-            ltx2_image_crf=(33.0 if image_path and segment_idx == 1 else 0.0),
+            ltx2_image_crf=0.0,
             image_path=image_path if segment_idx == 1 else None,
             return_continuation_state=False,
         )

apps/dreamverse/pyproject.toml

@@ -29,9 +29,6 @@ test = [
 dreamverse-server = "dreamverse.server_entry:cli"
 dreamverse-mock-server = "dreamverse.mock_server:cli"
 
-[tool.setuptools.packages.find]
-include = ["dreamverse*"]
-
 [tool.uv]
 package = false
 

apps/dreamverse/scripts/launch-dreamverse.sh

@@ -1,67 +0,0 @@
-#!/usr/bin/env bash
-# launch-dreamverse.sh — launch dreamverse-server on a compute node.
-#
-# Usage (from repo root):
-#   bash apps/dreamverse/scripts/launch-dreamverse.sh                          # GPUs 0-3, SP_SIZE=4
-#   CUDA_VISIBLE_DEVICES=0 bash apps/dreamverse/scripts/launch-dreamverse.sh   # single GPU, SP_SIZE=1
-
-set -euo pipefail
-
-CONDA_PREFIX="$HOME/miniconda3/envs/dreamverse"
-CUDA_RT_DIR="$CONDA_PREFIX/lib/python3.11/site-packages/nvidia/cuda_runtime/lib"
-GXX="$CONDA_PREFIX/bin/aarch64-conda-linux-gnu-g++"
-
-
-export CUDA_HOME="$CONDA_PREFIX"
-export FASTVIDEO_ENABLE_STARTUP_WARMUP=true
-export FASTVIDEO_ENABLE_PROMPT_SAFETY=false
-export DREAMVERSE_MAX_AUTOTUNE=true
-export LTX2_USE_DISTILLED_SIGMAS=0
-export LTX2_VIDEO_CONDITIONING_NUM_FRAMES=1
-export AUDIO_CONDITIONING_NUM_FRAMES=41
-export DREAMVERSE_SESSION_TIMEOUT_SECONDS="${DREAMVERSE_SESSION_TIMEOUT_SECONDS:-1800}"
-# GB200 max-autotune warmup compiles can run for hours; keep the watchdog generous here
-export FASTVIDEO_STARTUP_WARMUP_TIMEOUT_SECONDS="${FASTVIDEO_STARTUP_WARMUP_TIMEOUT_SECONDS:-24000}"
-export CEREBRAS_API_KEY="${CEREBRAS_API_KEY:-}"  # set this in your env or ~/.env
-export FASTVIDEO_PROMPT_CEREBRAS_MODEL="gpt-oss-120b"
-export TORCHINDUCTOR_CACHE_DIR="$HOME/.cache/torchinductor"
-export TRITON_CACHE_DIR="$HOME/.triton/cache"
-export TORCH_CUDA_ARCH_LIST="10.0a"
-
-# Compiler env (needed for flashinfer JIT compilation at server startup)
-export CXX="$CONDA_PREFIX/compiler_compat/g++"
-export CC="$CONDA_PREFIX/compiler_compat/gcc"
-export CUDAHOSTCXX="$GXX"
-export NVCC_PREPEND_FLAGS="-ccbin $GXX -allow-unsupported-compiler"
-
-# Link against libcudart.so.12 at JIT compile time; stubs for libcuda.so
-# cuda-compat has libcudart.so -> libcudart.so.12 (linker needs unversioned name)
-export LIBRARY_PATH="$CONDA_PREFIX/lib/cuda-compat:$CONDA_PREFIX/lib/stubs"
-
-# Only libcudart.so.12 at runtime — prevents cuDNN from seeing .so.13
-export LD_LIBRARY_PATH="$CUDA_RT_DIR"
-
-CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3}"
-export FASTVIDEO_GPU_COUNT="${FASTVIDEO_GPU_COUNT:-all}"
-# Default SP size to the usable GPU count so single-GPU invocations work.
-# A numeric FASTVIDEO_GPU_COUNT caps the pool below the visible count, and an
-# SP size above the pool size fails GPUPool startup with "Not enough GPUs".
-IFS=',' read -ra _VISIBLE_GPUS <<< "$CUDA_VISIBLE_DEVICES"
-# Count only non-empty tokens, matching gpu_pool.get_available_gpus (e.g. ",0,1" is 2 GPUs).
-_USABLE_GPU_COUNT=0
-for _gpu in "${_VISIBLE_GPUS[@]}"; do
-    [[ -n "${_gpu//[[:space:]]/}" ]] && _USABLE_GPU_COUNT=$((_USABLE_GPU_COUNT + 1))
-done
-if [[ "$FASTVIDEO_GPU_COUNT" =~ ^[0-9]+$ ]] && (( FASTVIDEO_GPU_COUNT < _USABLE_GPU_COUNT )); then
-    _USABLE_GPU_COUNT="$FASTVIDEO_GPU_COUNT"
-fi
-export DREAMVERSE_SP_SIZE="${DREAMVERSE_SP_SIZE:-$_USABLE_GPU_COUNT}"
-PORT="${DREAMVERSE_PORT:-8009}"
-
-FFMPEG_ENV="$(dirname "$0")/ffmpeg-env.sh"
-# shellcheck source=ffmpeg-env.sh
-[[ -f "$FFMPEG_ENV" ]] && source "$FFMPEG_ENV"
-
-echo "==> Launching dreamverse-server on GPU $CUDA_VISIBLE_DEVICES port $PORT"
-CUDA_VISIBLE_DEVICES="$CUDA_VISIBLE_DEVICES" \
-    "$CONDA_PREFIX/bin/dreamverse-server" --host 0.0.0.0 --port "$PORT"

apps/dreamverse/scripts/launch-frontend.sh

@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-# launch-frontend.sh — start the Dreamverse Next.js dev server and ngrok tunnel.
-#
-# Usage (from repo root):
-#   bash apps/dreamverse/scripts/launch-frontend.sh
-#
-# Override backend or ngrok URL via env:
-#   BACKEND_HOST=1.2.3.4 bash apps/dreamverse/scripts/launch-frontend.sh
-
-set -euo pipefail
-
-CONDA_PREFIX="$HOME/miniconda3/envs/dreamverse"
-BACKEND_HOST="${BACKEND_HOST:-10.244.18.228}"
-BACKEND_PORT="${BACKEND_PORT:-8009}"
-NGROK_URL="${NGROK_URL:-ltx23.ngrok.app}"
-WEB_DIR="$(git rev-parse --show-toplevel)/apps/dreamverse/web"
-
-cleanup() {
-    echo "==> Shutting down..."
-    kill "$FRONTEND_PID" 2>/dev/null || true
-}
-trap cleanup EXIT
-
-echo "==> Starting frontend (backend: $BACKEND_HOST:$BACKEND_PORT)"
-BACKEND_HOST="$BACKEND_HOST" BACKEND_PORT="$BACKEND_PORT" \
-    npm run --prefix "$WEB_DIR" dev &
-FRONTEND_PID=$!
-
-echo "==> Starting ngrok tunnel -> $NGROK_URL"
-"$CONDA_PREFIX/bin/ngrok" http --url="$NGROK_URL" 5299

apps/dreamverse/scripts/setup-dreamverse-env.sh

@@ -1,105 +0,0 @@
-#!/usr/bin/env bash
-# setup-dreamverse-env.sh — create and configure the dreamverse conda env
-# from scratch on this aarch64 NFS Slurm cluster.
-#
-# Run from the login node (from the repo root):
-#   bash apps/dreamverse/scripts/setup-dreamverse-env.sh
-#
-# After this script completes, use launch-dreamverse.sh on a compute node.
-
-set -euo pipefail
-
-REPO_ROOT="$(git rev-parse --show-toplevel)"
-ENV_NAME="dreamverse"
-LOCAL_DIR="/mnt/local/hal-kevin"       # cache/pkgs — keep on local disk
-CONDA_PREFIX="$HOME/miniconda3/envs/$ENV_NAME"  # env — on shared NFS so it survives node changes
-
-echo "==> Removing existing env if present"
-conda env remove -p "$CONDA_PREFIX" -y 2>/dev/null || true
-rm -rf "$CONDA_PREFIX" 2>/dev/null || true
-
-echo "==> Creating conda env at $CONDA_PREFIX"
-CONDA_PKGS_DIRS="$LOCAL_DIR/conda/pkgs" conda create -p "$CONDA_PREFIX" python=3.11 -y
-
-GXX="$CONDA_PREFIX/bin/aarch64-conda-linux-gnu-g++"
-GCC="$CONDA_PREFIX/bin/aarch64-conda-linux-gnu-gcc"
-
-echo "==> Installing compiler"
-CONDA_PKGS_DIRS="$LOCAL_DIR/conda/pkgs" conda install -p "$CONDA_PREFIX" gxx_linux-aarch64 -y
-
-echo "==> Installing CUDA toolkit (nvcc + headers)"
-CONDA_PKGS_DIRS="$LOCAL_DIR/conda/pkgs" conda install -p "$CONDA_PREFIX" -c nvidia cuda-toolkit -y
-
-echo "==> Hiding conflicting libcudart.so.13 immediately"
-mkdir -p "$CONDA_PREFIX/lib/hidden"
-mv "$CONDA_PREFIX"/lib/libcudart.so* "$CONDA_PREFIX/lib/hidden/" 2>/dev/null || true
-
-echo "==> Fixing compiler_compat symlinks"
-mkdir -p "$CONDA_PREFIX/compiler_compat"
-ln -sf "$GXX" "$CONDA_PREFIX/compiler_compat/g++"
-ln -sf "$GCC" "$CONDA_PREFIX/compiler_compat/gcc"
-
-echo "==> Symlinking CUDA headers to standard location"
-for f in "$CONDA_PREFIX/targets/sbsa-linux/include/"*; do
-    ln -sf "$f" "$CONDA_PREFIX/include/$(basename "$f")" 2>/dev/null || true
-done
-
-echo "==> Installing ffmpeg (native build with x264 + NVENC)"
-CUDA_PREFIX="$CONDA_PREFIX" bash "$REPO_ROOT/apps/dreamverse/scripts/install_native_ffmpeg.sh"
-
-echo "==> Installing pip and uv"
-CONDA_PKGS_DIRS="$LOCAL_DIR/conda/pkgs" conda install -p "$CONDA_PREFIX" pip -y
-"$CONDA_PREFIX/bin/pip" install uv
-
-echo "==> Setting compiler env vars"
-export UV_CACHE_DIR="$LOCAL_DIR/cache"
-export UV_LINK_MODE=copy
-export CXX="$CONDA_PREFIX/compiler_compat/g++"
-export CC="$CONDA_PREFIX/compiler_compat/gcc"
-export CUDAHOSTCXX="$GXX"
-export NVCC_PREPEND_FLAGS="-ccbin $GXX -allow-unsupported-compiler"
-export CUDA_HOME="$CONDA_PREFIX"
-# Only build for GB200 (sm_100a); CUDA 13 dropped support for older archs
-export TORCH_CUDA_ARCH_LIST="10.0a"
-
-echo "==> Installing torch with CUDA 12.8"
-UV_CACHE_DIR="$LOCAL_DIR/cache" "$CONDA_PREFIX/bin/uv" pip install torch==2.11.0 torchvision \
-    --index-url https://download.pytorch.org/whl/cu128
-
-echo "==> Hiding any newly introduced libcudart.so.13"
-mv "$CONDA_PREFIX"/lib/libcudart.so* "$CONDA_PREFIX/lib/hidden/" 2>/dev/null || true
-
-# Set paths now that torch (and its nvidia packages) are installed
-CUDA_RT_DIR="$CONDA_PREFIX/lib/python3.11/site-packages/nvidia/cuda_runtime/lib"
-CUDA_RT_SO="$(ls "$CUDA_RT_DIR"/libcudart.so.* 2>/dev/null | head -1)"
-
-# The pip nvidia package only has libcudart.so.12 (versioned), not libcudart.so.
-# The linker needs the unversioned name to satisfy -lcudart. Create a compat dir.
-mkdir -p "$CONDA_PREFIX/lib/cuda-compat"
-ln -sf "$CUDA_RT_SO" "$CONDA_PREFIX/lib/cuda-compat/libcudart.so"
-
-export LIBRARY_PATH="$CONDA_PREFIX/lib/cuda-compat:$CONDA_PREFIX/lib/stubs"
-export CMAKE_ARGS="-DCUDA_CUDART_LIBRARY=$CUDA_RT_SO -DCUDA_INCLUDE_DIRS=$CONDA_PREFIX/targets/sbsa-linux/include"
-
-echo "==> Installing build tools"
-"$CONDA_PREFIX/bin/pip" install scikit-build-core cmake ninja
-
-echo "==> Initializing git submodules"
-cd "$REPO_ROOT"
-git submodule update --init fastvideo-kernel/include/cutlass fastvideo-kernel/include/tk
-
-echo "==> Building fastvideo-kernel from local source"
-UV_CACHE_DIR="$LOCAL_DIR/cache" "$CONDA_PREFIX/bin/uv" pip install \
-    -e "./fastvideo-kernel" --no-build-isolation
-
-echo "==> Installing fastvideo + dreamverse extras"
-UV_CACHE_DIR="$LOCAL_DIR/cache" "$CONDA_PREFIX/bin/uv" pip install \
-    -e ".[dreamverse]" --no-build-isolation
-
-echo "==> Installing flashinfer-python (pinned, must be last)"
-UV_CACHE_DIR="$LOCAL_DIR/cache" "$CONDA_PREFIX/bin/uv" pip install \
-    https://github.com/flashinfer-ai/flashinfer/releases/download/v0.6.11.post3/flashinfer_python-0.6.11.post3-py3-none-any.whl
-
-echo ""
-echo "Done. On a compute node run (GPUs 0-3 by default; set CUDA_VISIBLE_DEVICES to restrict):"
-echo "  bash apps/dreamverse/scripts/launch-dreamverse.sh"

apps/dreamverse/serve_configs/streaming_demo.yaml

@@ -110,7 +110,7 @@ default_request:
     fps: 24                  # internal: gpu_pool.py:85 TARGET_FPS
 
 streaming:
-  # internal: config.py SESSION_TIMEOUT_SECONDS (env DREAMVERSE_SESSION_TIMEOUT_SECONDS, default 300)
+  # internal: config.py:33 SESSION_TIMEOUT_SECONDS = 300
   session_timeout_seconds: 300
   # internal: config.py:282-284 GENERATION_SEGMENT_CAP default 6
   generation_segment_cap: 6

apps/dreamverse/web/src/app/page.tsx

@@ -9,7 +9,7 @@ import SessionTimeoutModal from "@/components/SessionTimeoutModal";
 import Sidebar from "@/components/Sidebar";
 import Header from "@/components/Header";
 import VideoPlayer from "@/components/VideoPlayer";
-import Workspace, { SceneHistoryList } from "@/components/Workspace";
+import Workspace from "@/components/Workspace";
 import { saveProject, saveProjectMetadata, listProjects, loadProjectClips, deleteProject, pruneOldProjects, type StoredProject, type StoredClip } from "@/lib/projectStorage";
 import { isInfrastructureError } from "@/lib/ws/reducer";
 import { useStore } from "@/hooks/useStore";
@@ -31,7 +31,7 @@ import { applyNormalizedSocketEvent } from "@/lib/ws/reducer";
 import { createPromptWindowStore } from "@/stores/promptWindow";
 import { createRewriteStore } from "@/stores/rewrite";
 import { createSessionStore } from "@/stores/session";
-import { createStreamStore, USER_PROMPT_SOURCES } from "@/stores/stream";
+import { createStreamStore } from "@/stores/stream";
 import { createUiStore } from "@/stores/ui";
 import { Button } from "@/components/ui/button";
 
@@ -80,7 +80,7 @@ function yieldToEventLoop(): Promise<void> {
 
 const HERO_WAVE_LIGHT = ["#2A4A98", "#4878E5", "#6FA0F2", "#B0BCC8", "#E8D99E", "#D8C844", "#C2A620"];
 const HERO_WAVE_DARK = ["#143468", "#1E58B8", "#3892F0", "#80B8E8", "#B8D0EA", "#E2D498", "#DABB50"];
-const HERO_TEXT = "Direct scenes in seconds with";
+const HERO_TEXT = "Direct scenes in seconds";
 
 function HeroTagline() {
 	const ref = useRef<HTMLHeadingElement>(null);
@@ -166,8 +166,6 @@ function HeroTagline() {
 					</span>
 				</Fragment>
 			))}
-			<span data-char className="transition-[color,filter] duration-150">{" "}</span>
-			<img src="/logo.svg" alt="FastVideo" className="inline-block h-[1.1em] w-auto align-middle" />
 		</h1>
 	);
 }
@@ -239,7 +237,6 @@ export default function Page() {
 		enhancementEnabled,
 		promptExtensionError,
 		autoExtensionEnabled,
-		manualContinuationMode,
 		autoExtensionTimeoutHint,
 		loopGenerationEnabled,
 		generationPaused,
@@ -252,8 +249,6 @@ export default function Page() {
 		livePromptRewriteMode,
 		sessionExpired,
 		projectResetPending,
-		waitingForSegmentPrompt,
-		generatingNextScene,
 	} = sessionState;
 
 	const {
@@ -328,11 +323,6 @@ export default function Page() {
 	const [ttffValueMs, setTtffValueMs] = useState<number | null>(null);
 	const ttffIntervalRef = useRef<ReturnType<typeof setInterval> | null>(null);
 	const pendingInitialPromptRef = useRef("");
-	// Prompt id the opening scene is recorded under; sent as initial_rollout_prompt_id
-	// so the backend's pre-seeded opening PromptSubmission emits status updates
-	// (prompt_enhancing/prompt_ready/prompt_fallback_used) against the same id.
-	const pendingInitialPromptIdRef = useRef("");
-	const [initialImageDataUrl, setInitialImageDataUrl] = useState("");
 	const lastArchivedReplayKeyRef = useRef("");
 	const [sidebarOpen, setSidebarOpen] = useState(false);
 	const [currentThumbnail, setCurrentThumbnail] = useState<string | null>(null);
@@ -432,11 +422,6 @@ export default function Page() {
 			if (String(e?.source || "") === "user_rewrite" && typeof e?.text === "string" && e.text.trim()) {
 				return e.text.trim();
 			}
-			// Steering opening: the backend overwrites text/source with the enhanced
-			// prompt once ready, so fall back to the stable rawText record.
-			if (e?.steeringUserPrompt && typeof e?.rawText === "string" && e.rawText.trim()) {
-				return e.rawText.trim();
-			}
 		}
 		return "Untitled project";
 	}, [selectedPreset, promptEvents]);
@@ -447,39 +432,11 @@ export default function Page() {
 	const canDownloadVideo = useMemo(() => {
 		const currentActiveClip = activeClip as Record<string, any> | null;
 		if (currentActiveClip?.blob instanceof Blob) return true;
-		if ((completedClips as Record<string, any>[]).some((clip) => clip?.blob instanceof Blob)) return true;
-		// Steering only: once playback has started the live AV pipeline holds playable segments, so
-		// the user can download the in-progress video at any time (handleDownloadVideo remuxes live
-		// segments). Auto mode keeps its original blob-gated behavior.
-		return Boolean(manualContinuationMode) && Boolean(avPlaybackStarted);
-	}, [activeClip, completedClips, avPlaybackStarted, manualContinuationMode]);
-
-	// Steering mode scene list (oldest first). Primary source is the user's own words, captured
-	// stably at submit time as `rawText` (the backend later overwrites text/source with the
-	// enhanced prompt, so we never read those). A segment with no user prompt — e.g. a preset's
-	// opening scene — falls back to promptHistory (the actual prompt that drove that segment).
-	const steeringScenes = useMemo(() => {
-		if (!manualContinuationMode) return [] as Record<string, any>[];
-		const userScenes = (promptEvents as Record<string, any>[])
-			.filter((e) => e?.steeringUserPrompt && !e?.steeringFailed && typeof e?.rawText === "string" && e.rawText.trim())
-			.slice()
-			.reverse() // oldest -> newest
-			.map((e) => ({ id: e.promptId, prompt: e.rawText as string }));
-		const scenes: Record<string, any>[] = [];
-		// Preset opening segments: curated seeds with no user prompt of their own.
-		const curatedHists = (promptHistory as Record<string, any>[])
-			.slice()
-			.reverse() // oldest first
-			.filter((h) => !USER_PROMPT_SOURCES.has(String(h?.source || "")) && typeof h?.prompt === "string" && (h.prompt as string).trim());
-		scenes.push(...curatedHists.map((h) => ({ id: h.id || "scene_open", prompt: h.prompt })));
-		scenes.push(...userScenes);
-		return scenes;
-	}, [manualContinuationMode, promptEvents, promptHistory]);
+		return (completedClips as Record<string, any>[]).some((clip) => clip?.blob instanceof Blob);
+	}, [activeClip, completedClips]);
 
 	const hasEdits = useMemo(
-		() => Boolean(sessionStarted) && (
-			(promptEvents as Record<string, any>[]).some((e) => typeof e?.text === "string" && e.text.trim() && String(e?.source || "").trim() === "user_rewrite")
-		),
+		() => Boolean(sessionStarted) && (promptEvents as Record<string, any>[]).some((e) => typeof e?.text === "string" && e.text.trim() && String(e?.source || "").trim() === "user_rewrite"),
 		[sessionStarted, promptEvents],
 	);
 
@@ -1464,8 +1421,7 @@ export default function Page() {
 		if (!prompt) return;
 		lastSubmitTimeRef.current = now;
 		const rCAR = !uiStore.get().devtoolsMode && !uiStore.get().demoMode;
-		const inManualMode = sessionStore.get().manualContinuationMode;
-		if (!inManualMode && (rCAR || (sessionStore.get().livePromptRewriteMode && !uiStore.get().demoMode))) {
+		if (rCAR || (sessionStore.get().livePromptRewriteMode && !uiStore.get().demoMode)) {
 			if (rewriteStore.get().rewritingSeedPrompts) return;
 			const rewriteSourcePromptWindowPrompts = getActivePromptWindowPrompts();
 			const nextPendingClip = {
@@ -1506,11 +1462,6 @@ export default function Page() {
 			status: "submitted",
 			source: "user_raw",
 			text: prompt,
-			// Stable record of the user's own words for the steering scene list. The backend
-			// later overwrites `text`/`source` with the enhanced prompt via prompt/ready, but
-			// these two fields are never touched by trackPromptEvent.
-			steeringUserPrompt: true,
-			rawText: prompt,
 		});
 		ws.send(
 			JSON.stringify({
@@ -1524,14 +1475,7 @@ export default function Page() {
 			activeClipId: shouldUseArchivedPlaybackFallback() ? streamStore.get().activeClipId : "",
 			activePlaybackStartTime: shouldUseArchivedPlaybackFallback() ? streamStore.get().activePlaybackStartTime : 0,
 		});
-		sessionStore.patch({
-			livePromptDraft: "",
-			waitingForSegmentPrompt: false,
-			sessionNotice: "",
-			// Light the "Generating next scene" overlay immediately on a real submit;
-			// stream/media_init (or a fallback/error) clears it.
-			...(inManualMode ? { generatingNextScene: true } : {}),
-		});
+		sessionStore.patch({ livePromptDraft: "" });
 	}
 
 	function setLivePromptRewriteMode(enabled: boolean) {
@@ -1606,14 +1550,6 @@ export default function Page() {
 		);
 	}
 
-	// Steering (manual continuation) vs the automatic 6-segment rollout — a pre-session
-	// preference honored when the session starts.
-	function handleManualContinuationToggle(event: any) {
-		sessionStore.patch({
-			manualContinuationMode: Boolean(event.currentTarget.checked),
-		});
-	}
-
 	function handleLoopGenerationToggle(event: any) {
 		sessionStore.patch({
 			loopGenerationEnabled: Boolean(event.currentTarget.checked),
@@ -1789,9 +1725,6 @@ export default function Page() {
 			sessionNotice: preserveSessionNotice ? sessionStore.get().sessionNotice : "",
 			sessionExpired: preserveSessionNotice ? sessionStore.get().sessionExpired : false,
 			projectResetPending: false,
-			manualContinuationMode: true,
-			waitingForSegmentPrompt: false,
-			generatingNextScene: false,
 		});
 		rewriteStore.resetSessionState();
 		streamStore.resetSessionState();
@@ -1804,7 +1737,6 @@ export default function Page() {
 	function resetToProjectLobbyState() {
 		setVideoMuted(true);
 		pendingInitialPromptRef.current = "";
-		pendingInitialPromptIdRef.current = "";
 		sessionStore.patch({
 			sessionStarted: false,
 			livePromptDraft: "",
@@ -1818,9 +1750,6 @@ export default function Page() {
 			sessionNotice: "",
 			sessionExpired: false,
 			projectResetPending: false,
-			manualContinuationMode: true,
-			waitingForSegmentPrompt: false,
-			generatingNextScene: false,
 		});
 		rewriteStore.resetSessionState();
 		streamStore.resetSessionState();
@@ -1831,14 +1760,7 @@ export default function Page() {
 	}
 
 	function buildProjectInitPayload(type: "session_init_v2" | "project_init_v1") {
-		const manualMode = Boolean(sessionStore.get().manualContinuationMode);
-		let segmentPrompts = getSessionInitPrompts();
-		// Steering mode: seed the first 2 segments from the preset so there's no
-		// gap between segment 1 and 2; the user drives every subsequent segment.
-		// Force auto/loop off so the backend waits after the seeded prompts run out.
-		if (manualMode) {
-			segmentPrompts = segmentPrompts.slice(0, 2);
-		}
+		const segmentPrompts = getSessionInitPrompts();
 		setSeedPrompts(segmentPrompts);
 		return {
 			type,
@@ -1846,17 +1768,11 @@ export default function Page() {
 			preset_label: getInitialPresetLabel(),
 			curated_prompts: segmentPrompts,
 			initial_rollout_prompt: normalizeInitialPrompt(pendingInitialPromptRef.current),
-			// Ties the backend's pre-seeded opening PromptSubmission to the prompt event
-			// recorded in beginProjectLocally so its status updates land on that record.
-			initial_rollout_prompt_id: pendingInitialPromptIdRef.current,
-			initial_image: initialImageDataUrl
-				? { data_url: initialImageDataUrl, mime_type: initialImageDataUrl.split(";")[0].split(":")[1] || "image/png", name: "upload.png" }
-				: null,
+			initial_image: null,
 			single_clip_mode: false,
 			enhancement_enabled: sessionStore.get().enhancementEnabled,
-			auto_extension_enabled: manualMode ? false : sessionStore.get().autoExtensionEnabled,
-			loop_generation_enabled: manualMode ? false : sessionStore.get().loopGenerationEnabled,
-			manual_continuation_mode: manualMode,
+			auto_extension_enabled: sessionStore.get().autoExtensionEnabled,
+			loop_generation_enabled: sessionStore.get().loopGenerationEnabled,
 		};
 	}
 
@@ -2035,11 +1951,7 @@ export default function Page() {
 		setCurrentThumbnail(null);
 		const initialPrompt = normalizeInitialPrompt(sessionStore.get().livePromptDraft as string);
 		pendingInitialPromptRef.current = initialPrompt;
-		setInitialImageDataUrl("");
 		const rCAR = !uiStore.get().devtoolsMode && !uiStore.get().demoMode;
-		// The "Steering mode" toggle is authoritative: checked = manual per-segment steering,
-		// unchecked = automatic 6-segment rollout (default).
-		const nextManualContinuationMode = Boolean(sessionStore.get().manualContinuationMode);
 		sessionStore.patch({
 			sessionNotice: "",
 			sessionExpired: false,
@@ -2054,9 +1966,6 @@ export default function Page() {
 			autoExtensionTimeoutHint: "",
 			generationPaused: false,
 			projectResetPending: false,
-			manualContinuationMode: nextManualContinuationMode,
-			waitingForSegmentPrompt: false,
-			generatingNextScene: false,
 		});
 		resetPlaybackState();
 		streamStore.patch({
@@ -2070,15 +1979,12 @@ export default function Page() {
 			selectedHistoryId: "",
 		});
 		rewriteStore.resetSessionState();
-		pendingInitialPromptIdRef.current = initialPrompt ? makePromptId() : "";
 		if (initialPrompt) {
 			addPromptEvent({
-				promptId: pendingInitialPromptIdRef.current,
+				promptId: makePromptId(),
 				status: "rewrite_requested",
 				source: "user_rewrite",
 				text: initialPrompt,
-				// In steering mode the typed opening is the user's Scene 1 — record it stably.
-				...(nextManualContinuationMode ? { steeringUserPrompt: true, rawText: initialPrompt } : {}),
 			});
 		}
 		setSeedPrompts(getSessionInitPrompts());
@@ -2594,7 +2500,6 @@ export default function Page() {
 				selectedPresetId={selectedPresetId as string}
 				enhancementEnabled={enhancementEnabled as boolean}
 				autoExtensionEnabled={autoExtensionEnabled as boolean}
-				manualContinuationEnabled={manualContinuationMode as boolean}
 				loopGenerationEnabled={loopGenerationEnabled as boolean}
 				canJoinSession={canJoinSession as boolean}
 				canSubmitContinuation={canSubmitContinuation}
@@ -2607,7 +2512,6 @@ export default function Page() {
 				onEnhancementToggle={handleEnhancementToggle}
 				onCuratedPromptLimitChange={handleCuratedPromptLimitChange}
 				onAutoExtensionToggle={handleAutoExtensionToggle}
-				onManualContinuationToggle={handleManualContinuationToggle}
 				onLoopToggle={handleLoopGenerationToggle}
 				onJoin={joinSession}
 				onLeave={leaveSession}
@@ -2737,7 +2641,6 @@ export default function Page() {
 			/>
 			<Header timeLeft={headerTimeLeft} formatTime={formatTime} onToggleSidebar={() => setSidebarOpen((prev) => !prev)} />
 
-			<div className={cn("flex flex-1 min-h-0 flex-col", sessionStarted && "pb-16 sm:pb-28")}>
 			<div className="relative flex flex-1 min-h-0 flex-col justify-center px-4 pb-2 sm:px-6 sm:pb-12">
 				{isViewingMode && (
 					<>
@@ -2821,8 +2724,6 @@ export default function Page() {
 									showLivePlayback={showLivePlayback}
 									defaultMuted={videoMuted}
 									canDownload={canDownloadVideo}
-									waitingForSegmentPrompt={waitingForSegmentPrompt as boolean}
-									generatingNextScene={generatingNextScene as boolean}
 									onPlaying={markFirstFrameRendered}
 									onDownload={handleDownloadVideo}
 								/>
@@ -2833,7 +2734,6 @@ export default function Page() {
 					<section className={cn("mx-auto w-full max-w-2xl", hasEdits && "flex-1 min-h-0 overflow-y-auto")}>
 						<Workspace
 							promptEvents={promptEvents as any[]}
-							manualMode={manualContinuationMode as boolean}
 							currentThumbnail={currentThumbnail}
 							originalLabel={pendingInitialPromptRef.current || (selectedPreset as Record<string, any>)?.label || ""}
 							sessionStarted={sessionStarted as boolean}
@@ -2880,17 +2780,11 @@ export default function Page() {
 							isGenerating={loadingAnimation as boolean}
 							storyPresets={storyPresets as any[]}
 							continuationDraft={livePromptDraft as string}
-							manualContinuationEnabled={manualContinuationMode as boolean}
-							onModeChange={(manual) => sessionStore.patch({ manualContinuationMode: manual })}
-							initialImageDataUrl={initialImageDataUrl}
-							onImageUpload={(dataUrl) => setInitialImageDataUrl(dataUrl)}
-							onImageClear={() => setInitialImageDataUrl("")}
 							canJoinSession={canStartSession}
 							canSubmitContinuation={canSubmitContinuation}
 							sessionExpired={sessionExpired as boolean}
 							sessionNotice={sessionNotice as string}
 							projectResetPending={projectResetPending as boolean}
-							waitingForSegmentPrompt={waitingForSegmentPrompt as boolean}
 							onPresetGenerate={handlePresetGenerate}
 							onContinuationInput={handleLivePromptInput}
 							onContinuationKeydown={handleLivePromptKeydown}
@@ -2904,12 +2798,6 @@ export default function Page() {
 					</motion.div>
 				</div>
 			</div>
-			{manualContinuationMode && (
-				<div className="px-4 sm:px-6">
-					<SceneHistoryList sceneHistory={steeringScenes as any[]} />
-				</div>
-			)}
-			</div>
 		</main>
 	);
 }

apps/dreamverse/web/src/components/ChatBar.tsx

@@ -2,16 +2,13 @@
 
 import React, { useRef, useState, useCallback, useEffect } from "react";
 import Image from "next/image";
-import { Film, ArrowUp, X, Loader2, ArrowLeft, ImagePlus } from "lucide-react";
+import { Film, ArrowUp, X, Loader2, ArrowLeft } from "lucide-react";
 import { Button } from "@/components/ui/button";
 import LeaveSessionModal, { shouldShowLeaveWarning } from "@/components/LeaveSessionModal";
 import SpeechToTextButton from "@/components/SpeechToTextButton";
 import { cn } from "@/lib/utils";
 
 const PROMPT_MAX_LENGTH = 500;
-// Must match backend session_init_image.py: MAX_SESSION_INIT_IMAGE_BYTES / SUPPORTED_SESSION_INIT_IMAGE_MIME_TYPES.
-const IMAGE_MAX_BYTES = 15 * 1024 * 1024;
-const IMAGE_ALLOWED_TYPES = ["image/png", "image/jpeg", "image/webp"];
 
 interface Props {
 	sessionStarted?: boolean;
@@ -25,12 +22,6 @@ interface Props {
 	sessionNotice?: string;
 	projectResetPending?: boolean;
 	viewingReadOnly?: boolean;
-	waitingForSegmentPrompt?: boolean;
-	manualContinuationEnabled?: boolean;
-	onModeChange?: (manual: boolean) => void;
-	initialImageDataUrl?: string;
-	onImageUpload?: (dataUrl: string, mimeType: string, name: string) => void;
-	onImageClear?: () => void;
 	onPresetGenerate?: (presetId: string) => void;
 	onContinuationInput?: (e: React.ChangeEvent<HTMLTextAreaElement>) => void;
 	onContinuationKeydown?: (e: React.KeyboardEvent<HTMLTextAreaElement>) => void;
@@ -55,12 +46,6 @@ export default function ChatBar({
 	sessionNotice = "",
 	projectResetPending = false,
 	viewingReadOnly = false,
-	waitingForSegmentPrompt = false,
-	manualContinuationEnabled = false,
-	onModeChange = () => {},
-	initialImageDataUrl = "",
-	onImageUpload = () => {},
-	onImageClear = () => {},
 	onPresetGenerate = () => {},
 	onContinuationInput = () => {},
 	onContinuationKeydown = () => {},
@@ -74,51 +59,15 @@ export default function ChatBar({
 }: Props) {
 	const [sttBusy, setSttBusy] = useState(false);
 	const [leaveModalOpen, setLeaveModalOpen] = useState(false);
-	const [imageError, setImageError] = useState("");
-	const fileInputRef = useRef<HTMLInputElement>(null);
-
-	const processImageFile = useCallback((file: File) => {
-		if (!IMAGE_ALLOWED_TYPES.includes(file.type)) {
-			setImageError("Unsupported image type. Use a PNG, JPEG, or WebP image.");
-			return;
-		}
-		if (file.size > IMAGE_MAX_BYTES) {
-			setImageError("Image is too large. The maximum size is 15MB.");
-			return;
-		}
-		const reader = new FileReader();
-		reader.onload = (e) => {
-			const dataUrl = e.target?.result as string;
-			if (dataUrl) {
-				setImageError("");
-				onImageUpload(dataUrl, file.type, file.name);
-			}
-		};
-		reader.onerror = () => {
-			setImageError("Could not read the image file. Please try again.");
-		};
-		reader.readAsDataURL(file);
-	}, [onImageUpload]);
-
-	const handleImagePaste = useCallback((e: React.ClipboardEvent) => {
-		if (sessionStarted) return;
-		const items = Array.from(e.clipboardData?.items ?? []);
-		const imageItem = items.find((item) => item.type.startsWith("image/"));
-		if (!imageItem) return;
-		const file = imageItem.getAsFile();
-		if (file) processImageFile(file);
-	}, [sessionStarted, processImageFile]);
 	const showSpinner = isGenerating || rewritingSeedPrompts;
 	const isBusy = isGenerating || rewritingSeedPrompts || projectResetPending;
 	const messagePlaceholder = projectResetPending
 		? "Starting new project\u2026"
 		: isBusy
 			? "Generating video\u2026"
-			: waitingForSegmentPrompt
-				? "Describe the next scene\u2026"
-				: !sessionStarted
-					? "What video are you imagining?"
-					: "What do you want to edit?";
+			: !sessionStarted
+				? "What video are you imagining?"
+				: "What do you want to edit?";
 	const actionLabel = !sessionStarted ? "Generate" : "Rewrite rollout";
 
 	const inputRef = useRef<HTMLTextAreaElement>(null);
@@ -318,9 +267,9 @@ export default function ChatBar({
 						<Button onClick={onStartNewProject} size="sm" className="rounded-full px-5">
 							New Project
 						</Button>
-						<a href="https://haoailab.com/blogs/dreamverse/" target="_blank" rel="noopener noreferrer">
+						<a href="https://docs.google.com/forms/d/e/1FAIpQLSe5zpO1iD8Ds-Ih-fOLm64qd7YZVvuvAyHuJaAfw1hkRHTe_A/viewform?usp=publish-editor" target="_blank" rel="noopener noreferrer">
 							<Button variant="outline" size="sm" className="rounded-full px-5">
-								Blog
+								Join Waitlist
 							</Button>
 						</a>
 					</div>
@@ -397,31 +346,6 @@ export default function ChatBar({
 				</div>
 			)}
 
-
-			{!sessionStarted && imageError && (
-				<div className="rounded-xl border border-rose-500/20 bg-rose-500/10 px-4 py-2.5 text-center text-xs text-rose-700 dark:text-rose-300">
-					{imageError}
-				</div>
-			)}
-
-			{!sessionStarted && initialImageDataUrl && (
-				<div className="flex items-center gap-2 rounded-2xl border border-input bg-card/65 px-3 py-2">
-					<img src={initialImageDataUrl} alt="Initial frame" className="h-12 w-12 rounded-lg object-cover" />
-					<span className="flex-1 truncate text-xs text-muted-foreground">Starting image set</span>
-					<button type="button" onClick={() => { setImageError(""); onImageClear(); }} className="text-muted-foreground hover:text-foreground transition-colors">
-						<X className="size-4" />
-					</button>
-				</div>
-			)}
-
-			<input
-				ref={fileInputRef}
-				type="file"
-				accept={IMAGE_ALLOWED_TYPES.join(",")}
-				className="hidden"
-				onChange={(e) => { const f = e.target.files?.[0]; if (f) processImageFile(f); e.target.value = ""; }}
-			/>
-
 			<div
 				className={cn(
 					"flex min-w-0 items-center gap-1.5 rounded-4xl border py-2.5 pl-5 pr-2.5 shadow-md backdrop-blur-sm transition-all duration-200",
@@ -435,7 +359,6 @@ export default function ChatBar({
 					value={continuationDraft}
 					onChange={onContinuationInput}
 					onKeyDown={handleKeyDown}
-					onPaste={handleImagePaste}
 					placeholder={sttBusy ? "Listening\u2026" : messagePlaceholder}
 					maxLength={PROMPT_MAX_LENGTH}
 					disabled={isBusy || sttBusy}
@@ -446,19 +369,6 @@ export default function ChatBar({
 					)}
 				/>
 				{onSpeechTranscript && <SpeechToTextButton disabled={isBusy} onTranscript={onSpeechTranscript} onInterimChange={onSpeechInterimChange} onBusyChange={setSttBusy} />}
-				{!sessionStarted && (
-					<Button
-						type="button"
-						variant="ghost"
-						size="icon-sm"
-						title="Add starting image"
-						onClick={() => fileInputRef.current?.click()}
-						disabled={isBusy}
-						className="shrink-0 rounded-full text-muted-foreground hover:text-foreground"
-					>
-						<ImagePlus className="size-4" />
-					</Button>
-				)}
 				{!sessionStarted ? (
 					<Button
 						aria-label={actionLabel}

apps/dreamverse/web/src/components/Header.tsx

@@ -8,6 +8,7 @@ import { Badge } from "@/components/ui/badge";
 import { Button } from "@/components/ui/button";
 import { ThemeToggle } from "@/components/ui/theme-toggle";
 
+const FASTVIDEO_REPO_URL = "https://haoailab.com/blogs/dreamverse/";
 const FASTVIDEO_BLOG_URL = "https://haoailab.com/blogs/dreamverse/";
 
 interface Props {
@@ -28,13 +29,13 @@ export default function Header({ timeLeft = null, formatTime = (seconds) => `${s
 							<SidePanelOpenFilled size={20} />
 						</Button>
 					)}
-					<a href="/" title="FastVideo home">
+					<a href={FASTVIDEO_REPO_URL} target="_blank" rel="noopener noreferrer" title="FastVideo on GitHub">
 						<Image src="/logo.svg" alt="FastVideo" width={32} height={32} className="h-8 w-auto sm:h-9 transition-opacity hover:opacity-70" />
 					</a>
 					<div className="hidden sm:flex items-center gap-3">
-						<a href={FASTVIDEO_BLOG_URL} target="_blank" rel="noopener noreferrer">
+						<a href="https://docs.google.com/forms/d/e/1FAIpQLSe5zpO1iD8Ds-Ih-fOLm64qd7YZVvuvAyHuJaAfw1hkRHTe_A/viewform?usp=publish-editor" target="_blank" rel="noopener noreferrer">
 							<Button variant="outline" size="sm" className="gap-1.5 rounded-full px-3 text-xs">
-								Blog
+								Join Waitlist
 								<ExternalLink className="size-3 opacity-60" />
 							</Button>
 						</a>
@@ -52,9 +53,9 @@ export default function Header({ timeLeft = null, formatTime = (seconds) => `${s
 			</div>
 
 			<div className="flex sm:hidden items-center gap-2 px-4 pb-3">
-				<a href={FASTVIDEO_BLOG_URL} target="_blank" rel="noopener noreferrer">
+				<a href="https://docs.google.com/forms/d/e/1FAIpQLSe5zpO1iD8Ds-Ih-fOLm64qd7YZVvuvAyHuJaAfw1hkRHTe_A/viewform?usp=publish-editor" target="_blank" rel="noopener noreferrer">
 					<Button variant="outline" size="sm" className="gap-1.5 rounded-full px-3 text-xs">
-						Blog
+						Join Waitlist
 						<ExternalLink className="size-3 opacity-60" />
 					</Button>
 				</a>

apps/dreamverse/web/src/components/VideoPlayer.tsx

@@ -3,7 +3,7 @@
 import React, { useState, useEffect, useRef, useCallback } from "react";
 import { cn } from "@/lib/utils";
 import { PlayFilledAlt } from "@carbon/icons-react";
-import { Check, ChevronDown, Download, Loader2, Share } from "lucide-react";
+import { Download, Loader2, Share } from "lucide-react";
 import { Button } from "@/components/ui/button";
 interface VideoPlayerProps {
 	videoRef?: React.RefCallback<HTMLVideoElement>;
@@ -21,8 +21,6 @@ interface VideoPlayerProps {
 	showLivePlayback?: boolean;
 	defaultMuted?: boolean;
 	rewritePending?: boolean;
-	waitingForSegmentPrompt?: boolean;
-	generatingNextScene?: boolean;
 	onPlaying?: () => void;
 	onDownload?: () => void;
 }
@@ -59,8 +57,6 @@ export default function VideoPlayer({
 	showLivePlayback = true,
 	defaultMuted = true,
 	rewritePending = false,
-	waitingForSegmentPrompt = false,
-	generatingNextScene = false,
 	onPlaying = () => {},
 	onDownload,
 }: VideoPlayerProps) {
@@ -92,137 +88,6 @@ export default function VideoPlayer({
 		setCanShare(typeof navigator.canShare === "function" && window.matchMedia("(pointer: coarse)").matches);
 	}, []);
 
-	// Steering mode: the backend may signal "waiting for next prompt" while the current
-	// segment is still PLAYING (it generates ahead). Only surface the "Segment complete"
-	// overlay once the playhead actually reaches the end of the buffered segment.
-	const [playbackReachedEnd, setPlaybackReachedEnd] = useState(false);
-	// Steering mode: when the user submits the next scene, the segment is generated
-	// (a few seconds of latency) before frames stream. Show a "Generating next scene…"
-	// indicator across that gap so the frozen frame isn't silent. Driven by the explicit
-	// generatingNextScene state (set on scene submit / prompt selection), never inferred
-	// from waitingForSegmentPrompt edges.
-	const [generatingNext, setGeneratingNext] = useState(false);
-	// End of the buffered timeline captured the moment generation starts; the freshly
-	// generated segment extends the buffer past this, which is how we know it landed.
-	const genBoundaryRef = useRef(0);
-
-	// Steering mode: the backend may signal "waiting for next prompt" while the current
-	// segment is still PLAYING (it generates ahead). Track whether the playhead has reached
-	// the end of the buffered segment so end-overlays only show there. Keep tracking through
-	// the generating phase too, so scrubbing back and replaying to the end re-shows them.
-	useEffect(() => {
-		if (!waitingForSegmentPrompt && !generatingNext) {
-			setPlaybackReachedEnd(false);
-			return;
-		}
-		const el = liveVideoEl.current;
-		if (!el) return;
-		const check = () => {
-			try {
-				const buffered = el.buffered;
-				if (buffered.length === 0) return;
-				const end = buffered.end(buffered.length - 1);
-				// Track proximity both ways: scrubbing back off the end hides the overlay,
-				// playing forward to the end re-shows it.
-				setPlaybackReachedEnd(el.ended || end - el.currentTime <= 0.2);
-			} catch {
-				/* buffered access can throw mid-append */
-			}
-		};
-		check();
-		el.addEventListener("timeupdate", check);
-		el.addEventListener("ended", check);
-		el.addEventListener("waiting", check);
-		el.addEventListener("stalled", check);
-		el.addEventListener("pause", check);
-		el.addEventListener("seeking", check);
-		el.addEventListener("seeked", check);
-		el.addEventListener("playing", check);
-		el.addEventListener("progress", check);
-		return () => {
-			el.removeEventListener("timeupdate", check);
-			el.removeEventListener("ended", check);
-			el.removeEventListener("waiting", check);
-			el.removeEventListener("stalled", check);
-			el.removeEventListener("pause", check);
-			el.removeEventListener("seeking", check);
-			el.removeEventListener("seeked", check);
-			el.removeEventListener("playing", check);
-			el.removeEventListener("progress", check);
-		};
-	}, [waitingForSegmentPrompt, generatingNext]);
-
-	useEffect(() => {
-		if (waitingForSegmentPrompt || !sessionStarted) {
-			// Back to waiting (or session over): nothing is generating.
-			setGeneratingNext(false);
-			return;
-		}
-		if (!generatingNextScene) return;
-		// Snapshot the current end of the buffered timeline; the generated segment will
-		// extend the buffer past this boundary.
-		const el = liveVideoEl.current;
-		let boundary = el?.currentTime ?? 0;
-		try {
-			const b = el?.buffered;
-			if (b && b.length) boundary = Math.max(boundary, b.end(b.length - 1));
-		} catch {
-			/* buffered access can throw mid-append */
-		}
-		genBoundaryRef.current = boundary;
-		setGeneratingNext(true);
-	}, [generatingNextScene, waitingForSegmentPrompt, sessionStarted]);
-
-	useEffect(() => {
-		if (!generatingNext) return;
-		if (!sessionStarted) {
-			setGeneratingNext(false);
-			return;
-		}
-		const el = liveVideoEl.current;
-		if (!el) return;
-		// Clear the instant the freshly generated segment lands: the buffer grows past the
-		// boundary captured at generation start (or the playhead advances into the new
-		// frames). Deliberately NOT a bare "playing" handler — scrubbing back and replaying
-		// the EXISTING segment must keep "Generating" up until the new frames actually arrive.
-		const check = () => {
-			try {
-				const b = el.buffered;
-				const end = b.length ? b.end(b.length - 1) : 0;
-				if (end > genBoundaryRef.current + 0.25 || el.currentTime > genBoundaryRef.current + 0.1) {
-					setGeneratingNext(false);
-				}
-			} catch {
-				/* buffered access can throw mid-append */
-			}
-		};
-		check();
-		el.addEventListener("progress", check);
-		el.addEventListener("timeupdate", check);
-		el.addEventListener("durationchange", check);
-		return () => {
-			el.removeEventListener("progress", check);
-			el.removeEventListener("timeupdate", check);
-			el.removeEventListener("durationchange", check);
-		};
-	}, [generatingNext, sessionStarted]);
-
-	// Drive a ~10.5s progress bar during generation so the wait has a visible ETA.
-	const GEN_DURATION_MS = 10500;
-	const [genProgress, setGenProgress] = useState(0);
-	useEffect(() => {
-		if (!generatingNext) {
-			setGenProgress(0);
-			return;
-		}
-		const start = performance.now();
-		setGenProgress(0);
-		const id = setInterval(() => {
-			setGenProgress(Math.min((performance.now() - start) / GEN_DURATION_MS, 1));
-		}, 50);
-		return () => clearInterval(id);
-	}, [generatingNext]);
-
 	return (
 		<div className="mx-auto w-full max-w-3xl mb-2 sm:mb-6">
 			<div className="rounded-2xl border border-border bg-card/50 p-2 shadow-lg backdrop-blur-md">
@@ -239,7 +104,7 @@ export default function VideoPlayer({
 								<PlayFilledAlt className="size-10 text-white/25" />
 								<p className="text-sm text-white/50">Your video will appear here</p>
 							</div>
-						) : !avPlaybackStarted && !mediaAppendError && !inQueue && !waitingForSegmentPrompt && loadingAnimation ? (
+						) : !avPlaybackStarted && !mediaAppendError && !inQueue && loadingAnimation ? (
 							<div className="absolute inset-0 flex flex-col items-center justify-center gap-4 bg-slate-900/60 p-4 backdrop-blur-[2px]">
 								<div className="pointer-events-none absolute inset-0 overflow-hidden">
 									<div className="absolute inset-0 -translate-x-full animate-[shimmer_3s_ease-in-out_infinite] bg-gradient-to-r from-transparent via-white/[0.04] to-transparent" />
@@ -249,35 +114,6 @@ export default function VideoPlayer({
 							</div>
 						) : null}
 
-						{/* Steering mode: this segment finished — wait gracefully for the user's next scene
-							instead of spinning. The last frame stays visible behind a soft bottom gradient. */}
-						{sessionStarted && waitingForSegmentPrompt && playbackReachedEnd && !mediaAppendError && !inQueue && (
-							<div className="pointer-events-none absolute inset-0 flex flex-col items-center justify-end gap-2 bg-gradient-to-t from-slate-950/85 via-slate-950/15 to-transparent p-5 pb-6 text-center">
-								<div className="flex size-9 items-center justify-center rounded-full border border-white/25 bg-white/10 shadow-lg backdrop-blur-md">
-									<Check className="size-4 text-white/90" />
-								</div>
-								<div className="space-y-0.5">
-									<p className="text-sm font-medium text-white/95">Segment complete</p>
-									<p className="text-xs text-white/65">Describe the next scene below to keep going</p>
-								</div>
-								<ChevronDown className="size-4 animate-bounce text-white/45" />
-							</div>
-						)}
-
-						{/* Steering mode: generating the next segment — show a ~10.5s progress bar so the wait has an ETA.
-							Gated on playbackReachedEnd like "Segment complete": scrubbing back hides it, playing to the end re-shows it. */}
-						{sessionStarted && generatingNext && playbackReachedEnd && !mediaAppendError && !inQueue && (
-							<div className="pointer-events-none absolute inset-0 flex flex-col items-center justify-end gap-3 bg-gradient-to-t from-slate-950/85 via-slate-950/15 to-transparent p-5 pb-7 text-center">
-								<p className="text-sm font-medium text-white/95">Generating next scene&hellip;</p>
-								<div className="h-1.5 w-48 overflow-hidden rounded-full bg-white/15 shadow-sm">
-									<div
-										className="h-full rounded-full bg-white/85 transition-[width] duration-100 ease-linear"
-										style={{ width: `${Math.round(genProgress * 100)}%` }}
-									/>
-								</div>
-							</div>
-						)}
-
 						{rewritePending && avPlaybackStarted && (
 							<div className="absolute inset-x-0 bottom-0 z-10 flex items-center justify-center gap-2 bg-gradient-to-t from-black/60 to-transparent px-4 pb-12 pt-8 pointer-events-none">
 								<Loader2 className="size-4 animate-spin text-white/90" />

apps/dreamverse/web/src/components/Workspace.tsx

@@ -3,90 +3,11 @@ import React, { useRef, useMemo, useEffect, useCallback, useState } from "react"
 import { motion, useAnimationControls } from "framer-motion";
 import { Badge } from "@/components/ui/badge";
 import { cn } from "@/lib/utils";
-import { Check, Clapperboard, Lightbulb, Pencil } from "lucide-react";
+import { Check, Lightbulb, Pencil } from "lucide-react";
 
 export const WORKSPACE_ORIGINAL_SELECTION_KEY = "original";
 export const WORKSPACE_CURRENT_SELECTION_KEY = "current";
 
-export function SceneHistoryList({ sceneHistory = [] }: { sceneHistory?: Record<string, any>[] }) {
-	const bottomSentinelRef = useRef<HTMLDivElement>(null);
-	const topSentinelRef = useRef<HTMLDivElement>(null);
-	const [showTopFade, setShowTopFade] = useState(false);
-
-	const scenes = useMemo(
-		() => (sceneHistory || []).filter((s) => normalizeText(s?.prompt)),
-		[sceneHistory],
-	);
-
-	const scrollToBottom = useCallback(() => {
-		setTimeout(() => {
-			bottomSentinelRef.current?.scrollIntoView({ block: "end", behavior: "smooth" });
-		}, 60);
-	}, []);
-
-	useEffect(() => {
-		if (scenes.length > 0) scrollToBottom();
-	}, [scenes.length, scrollToBottom]);
-
-	useEffect(() => {
-		const sentinel = bottomSentinelRef.current;
-		if (!sentinel || typeof ResizeObserver === "undefined") return;
-		let container: HTMLElement | null = sentinel.parentElement;
-		while (container) {
-			const oy = getComputedStyle(container).overflowY;
-			if (oy === "auto" || oy === "scroll") break;
-			container = container.parentElement;
-		}
-		if (!container) return;
-		const ro = new ResizeObserver(() => {
-			const nearBottom = container!.scrollHeight - container!.scrollTop - container!.clientHeight < 96;
-			if (nearBottom) scrollToBottom();
-		});
-		ro.observe(container);
-		return () => ro.disconnect();
-	}, [scenes.length, scrollToBottom]);
-
-	useEffect(() => {
-		const el = topSentinelRef.current;
-		if (!el) return;
-		const observer = new IntersectionObserver(([entry]) => setShowTopFade(!entry.isIntersecting), { threshold: 0.1 });
-		observer.observe(el);
-		return () => observer.disconnect();
-	}, [scenes.length]);
-
-	if (scenes.length === 0) return null;
-
-	return (
-		<section className="relative z-10 flex flex-col mx-auto w-full max-w-2xl max-h-32 overflow-y-auto">
-			<div
-				className={cn(
-					"pointer-events-none sticky top-0 z-20 -mb-12 h-12 bg-linear-to-b from-background to-transparent transition-opacity duration-200",
-					showTopFade ? "opacity-100" : "opacity-0",
-				)}
-				aria-hidden="true"
-			/>
-			<div ref={topSentinelRef} className="h-0 w-0" aria-hidden="true" />
-			<div className="flex flex-col gap-2 pt-12 pb-4">
-				{scenes.map((scene, index) => (
-					<div
-						key={scene.id || index}
-						className="flex items-start gap-3 rounded-xl p-3 transition-colors duration-200 hover:bg-slate-200/50 hover:dark:bg-slate-800/30"
-					>
-						<div className="flex min-w-0 flex-1 flex-col gap-2">
-							<Badge variant="secondary" className="horizontal gap-2 items-center w-fit">
-								<Clapperboard className="size-3 opacity-70" />
-								{`Scene ${index + 1}`}
-							</Badge>
-							<p className="line-clamp-2 text-sm leading-5 text-muted-foreground">{scene.prompt}</p>
-						</div>
-					</div>
-				))}
-			</div>
-			<div ref={bottomSentinelRef} className="h-0 w-0" aria-hidden="true" />
-		</section>
-	);
-}
-
 interface WorkspaceProps {
 	promptEvents?: Record<string, any>[];
 	currentThumbnail?: string | null;
@@ -98,7 +19,6 @@ interface WorkspaceProps {
 	selectedClipId?: string;
 	selectedEntryKey?: string;
 	originalClipId?: string;
-	manualMode?: boolean;
 }
 
 function normalizeText(value: any): string {
@@ -298,7 +218,7 @@ function ChromaGradient({ sessionStarted = false }: { sessionStarted?: boolean }
 	);
 }
 
-export default function Workspace({ promptEvents = [], currentThumbnail = null, originalLabel = "", sessionStarted = false, onSelectOriginal, onSelectEvent, onSelectCurrent, selectedClipId, selectedEntryKey: selectedEntryKeyProp, originalClipId = "", manualMode = false }: WorkspaceProps) {
+export default function Workspace({ promptEvents = [], currentThumbnail = null, originalLabel = "", sessionStarted = false, onSelectOriginal, onSelectEvent, onSelectCurrent, selectedClipId, selectedEntryKey: selectedEntryKeyProp, originalClipId = "" }: WorkspaceProps) {
 	const bottomSentinelRef = useRef<HTMLDivElement>(null);
 	const topSentinelRef = useRef<HTMLDivElement>(null);
 	const [showTopFade, setShowTopFade] = useState(false);
@@ -339,14 +259,6 @@ export default function Workspace({ promptEvents = [], currentThumbnail = null,
 		return () => observer.disconnect();
 	}, [conversationEvents.length]);
 
-	if (manualMode) {
-		return (
-			<div className="mt-auto flex flex-col">
-				<ChromaGradient sessionStarted={sessionStarted} />
-			</div>
-		);
-	}
-
 	return (
 		<div className="mt-auto flex flex-col">
 			<ChromaGradient sessionStarted={sessionStarted} />

apps/dreamverse/web/src/components/devtools/DevtoolsComposer.tsx

@@ -34,7 +34,6 @@ interface DevtoolsComposerProps {
   demoMode?: boolean;
   enhancementEnabled?: boolean;
   autoExtensionEnabled?: boolean;
-  manualContinuationEnabled?: boolean;
   loopGenerationEnabled?: boolean;
   curatedPromptLimit?: number;
   maxCuratedPromptCount?: number;
@@ -50,7 +49,6 @@ interface DevtoolsComposerProps {
   onEnhancementToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onCuratedPromptLimitChange?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onAutoExtensionToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
-  onManualContinuationToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onLoopToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onLivePromptModeToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onSpeechTranscript?: (text: string) => void;
@@ -73,7 +71,6 @@ export default function DevtoolsComposer({
   demoMode = false,
   enhancementEnabled = true,
   autoExtensionEnabled = false,
-  manualContinuationEnabled = false,
   loopGenerationEnabled = false,
   curatedPromptLimit = 0,
   maxCuratedPromptCount = 0,
@@ -89,7 +86,6 @@ export default function DevtoolsComposer({
   onEnhancementToggle = () => {},
   onCuratedPromptLimitChange = () => {},
   onAutoExtensionToggle = () => {},
-  onManualContinuationToggle = () => {},
   onLoopToggle = () => {},
   onLivePromptModeToggle = () => {},
   onSpeechTranscript,
@@ -332,28 +328,6 @@ export default function DevtoolsComposer({
                   </div>
                 </div>
 
-                <div className="flex items-start gap-3">
-                  <Checkbox
-                    id="devtools-steering-mode"
-                    checked={manualContinuationEnabled}
-                    onCheckedChange={(checked) =>
-                      onManualContinuationToggle({
-                        target: { checked: Boolean(checked) },
-                        currentTarget: { checked: Boolean(checked) },
-                      } as React.ChangeEvent<HTMLInputElement>)
-                    }
-                  />
-                  <div className="space-y-1">
-                    <Label htmlFor="devtools-steering-mode">
-                      Steering mode
-                    </Label>
-                    <p className="text-sm text-muted-foreground">
-                      Drive each segment manually — type the next scene to
-                      continue (vs the automatic 6-segment rollout).
-                    </p>
-                  </div>
-                </div>
-
                 <div className="flex items-start gap-3">
                   <Checkbox
                     id="devtools-loop-generation"

apps/dreamverse/web/src/components/devtools/DevtoolsShell.tsx

@@ -21,7 +21,6 @@ interface DevtoolsShellProps {
   selectedPresetId?: string;
   enhancementEnabled?: boolean;
   autoExtensionEnabled?: boolean;
-  manualContinuationEnabled?: boolean;
   loopGenerationEnabled?: boolean;
   canJoinSession?: boolean;
   canSubmitContinuation?: boolean;
@@ -35,7 +34,6 @@ interface DevtoolsShellProps {
   onEnhancementToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onCuratedPromptLimitChange?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onAutoExtensionToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
-  onManualContinuationToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onLoopToggle?: (e: React.ChangeEvent<HTMLInputElement>) => void;
   onJoin?: () => void;
   onLeave?: () => void;
@@ -130,7 +128,6 @@ export default function DevtoolsShell({
   selectedPresetId = '',
   enhancementEnabled = true,
   autoExtensionEnabled = false,
-  manualContinuationEnabled = false,
   loopGenerationEnabled = false,
   canJoinSession = false,
   canSubmitContinuation = false,
@@ -144,7 +141,6 @@ export default function DevtoolsShell({
   onEnhancementToggle = () => {},
   onCuratedPromptLimitChange = () => {},
   onAutoExtensionToggle = () => {},
-  onManualContinuationToggle = () => {},
   onLoopToggle = () => {},
   onJoin = () => {},
   onLeave = () => {},
@@ -288,7 +284,6 @@ export default function DevtoolsShell({
           demoMode={demoMode}
           enhancementEnabled={enhancementEnabled}
           autoExtensionEnabled={autoExtensionEnabled}
-          manualContinuationEnabled={manualContinuationEnabled}
           loopGenerationEnabled={loopGenerationEnabled}
           curatedPromptLimit={curatedPromptLimit}
           maxCuratedPromptCount={maxCuratedPromptCount}
@@ -304,7 +299,6 @@ export default function DevtoolsShell({
           onEnhancementToggle={onEnhancementToggle}
           onCuratedPromptLimitChange={onCuratedPromptLimitChange}
           onAutoExtensionToggle={onAutoExtensionToggle}
-          onManualContinuationToggle={onManualContinuationToggle}
           onLoopToggle={onLoopToggle}
           onLivePromptModeToggle={onLivePromptModeToggle}
           onSpeechTranscript={onSpeechTranscript}

apps/dreamverse/web/src/lib/promptEvents.test.ts

@@ -57,32 +57,4 @@ describe('prependPromptEvent', () => {
     expect(next[0].promptId).toBe('new');
     expect(next.some((item: any) => item.promptId === 'p-23')).toBe(false);
   });
-
-  it('never drops steering scene events when capping', () => {
-    // 30 scenes interleaved with 30 other events — well past the cap.
-    let events: Record<string, any>[] = [];
-    for (let i = 0; i < 30; i += 1) {
-      events = prependPromptEvent(events, {
-        promptId: `scene-${i}`,
-        status: 'submitted',
-        steeringUserPrompt: true,
-        rawText: `scene ${i}`,
-      });
-      events = prependPromptEvent(events, {
-        promptId: `other-${i}`,
-        status: 'submitted',
-      });
-    }
-
-    const scenes = events.filter((e) => e.steeringUserPrompt);
-    expect(scenes).toHaveLength(30);
-    // Oldest-first scene order (and therefore numbering) is stable and complete.
-    expect(scenes[scenes.length - 1].promptId).toBe('scene-0');
-    expect(scenes[0].promptId).toBe('scene-29');
-    // Non-scene events are still capped, oldest dropped first.
-    const others = events.filter((e) => !e.steeringUserPrompt);
-    expect(others.length).toBeLessThanOrEqual(24);
-    expect(others.some((e) => e.promptId === 'other-0')).toBe(false);
-    expect(others[0].promptId).toBe('other-29');
-  });
 });

apps/dreamverse/web/src/lib/promptEvents.ts

@@ -16,19 +16,5 @@ export function prependPromptEvent(
   events: Record<string, any>[],
   event: Record<string, any>,
 ): Record<string, any>[] {
-  const next = [event, ...events];
-  if (next.length <= MAX_PROMPT_EVENTS) {
-    return next;
-  }
-  // Steering scene events (steeringUserPrompt) are exempt from the cap: the
-  // scene list is derived from them and must stay complete and stably numbered
-  // for long sessions. Only the oldest non-scene events are dropped.
-  let nonSceneKept = 0;
-  return next.filter((e) => {
-    if (e?.steeringUserPrompt) {
-      return true;
-    }
-    nonSceneKept += 1;
-    return nonSceneKept <= MAX_PROMPT_EVENTS;
-  });
+  return [event, ...events].slice(0, MAX_PROMPT_EVENTS);
 }

apps/dreamverse/web/src/lib/ws/reducer.test.ts

@@ -1,11 +1,6 @@
 import { describe, expect, it } from 'vitest';
 
-import { applyNormalizedSocketEvent, resolveSessionErrorMessage } from './reducer';
-import { createSessionStore } from '../../stores/session';
-import { createRewriteStore } from '../../stores/rewrite';
-import { createStreamStore } from '../../stores/stream';
-import { createUiStore } from '../../stores/ui';
-import { createPromptWindowStore } from '../../stores/promptWindow';
+import { resolveSessionErrorMessage } from './reducer';
 
 describe('resolveSessionErrorMessage', () => {
   it('returns a dedicated message for IP session limit errors', () => {
@@ -24,162 +19,3 @@ describe('resolveSessionErrorMessage', () => {
     })).toBe('Backend replica unavailable. Rejoin session.');
   });
 });
-
-function buildContext(overrides: Record<string, unknown> = {}) {
-  const sessionStore = createSessionStore();
-  const rewriteStore = createRewriteStore();
-  const streamStore = createStreamStore();
-  const uiStore = createUiStore();
-  const promptWindowStore = createPromptWindowStore();
-  const avPipeline = {
-    reset: () => {},
-    setStreamCompleted: () => {},
-    noteSegmentInit: () => {},
-    noteSegmentComplete: () => {},
-    maybeStartPlayback: () => {},
-    ensurePipeline: async () => {},
-  };
-  return {
-    sessionStore,
-    promptWindowStore,
-    rewriteStore,
-    streamStore,
-    uiStore,
-    avPipeline,
-    tick: async () => {},
-    defaultAvMime: 'video/mp4',
-    fixedRewriteModel: 'model',
-    parseLatencyMs: () => null,
-    formatPromptWindowEventText: () => '',
-    makePromptId: () => 'generated-id',
-    buildClipLabel: () => 'clip',
-    startSessionCountdown: () => {},
-    clearCountdownInterval: () => {},
-    resetTtffTimer: () => {},
-    startTtffTimer: () => {},
-    preserveArchivedPlaybackSelection: false,
-    finalizeStreamCompletion: async () => {},
-    ...overrides,
-  };
-}
-
-describe('steering generatingNextScene flow', () => {
-  it('sets generatingNextScene on prompt/sources_resumed in manual mode', async () => {
-    const context = buildContext();
-    await applyNormalizedSocketEvent(
-      { type: 'prompt/sources_resumed', payload: { segment_idx: 2 } },
-      context,
-    );
-    expect(context.sessionStore.get().generatingNextScene).toBe(true);
-    expect(context.sessionStore.get().waitingForSegmentPrompt).toBe(false);
-  });
-
-  it('does NOT set generatingNextScene on session/auto_extension_updated', async () => {
-    const context = buildContext();
-    await applyNormalizedSocketEvent(
-      { type: 'session/auto_extension_updated', payload: { enabled: true } },
-      context,
-    );
-    expect(context.sessionStore.get().generatingNextScene).toBe(false);
-    expect(context.sessionStore.get().waitingForSegmentPrompt).toBe(false);
-  });
-
-  it('clears generatingNextScene when segment media arrives', async () => {
-    const context = buildContext();
-    context.sessionStore.patch({ generatingNextScene: true });
-    await applyNormalizedSocketEvent(
-      {
-        type: 'stream/media_init',
-        payload: { segment_idx: 2, stream_id: 's', mime: 'video/mp4' },
-      },
-      context,
-    );
-    expect(context.sessionStore.get().generatingNextScene).toBe(false);
-  });
-
-  it('clears generatingNextScene and returns to waiting on prompt/sources_blocked', async () => {
-    const context = buildContext();
-    context.sessionStore.patch({ generatingNextScene: true });
-    await applyNormalizedSocketEvent(
-      { type: 'prompt/sources_blocked', payload: { segment_idx: 3 } },
-      context,
-    );
-    expect(context.sessionStore.get().generatingNextScene).toBe(false);
-    expect(context.sessionStore.get().waitingForSegmentPrompt).toBe(true);
-  });
-
-  it('clears generatingNextScene when the opening prompt falls back', async () => {
-    const context = buildContext();
-    context.sessionStore.patch({ generatingNextScene: true });
-    await applyNormalizedSocketEvent(
-      {
-        type: 'prompt/fallback_used',
-        payload: { prompt_id: 'p1', prompt: '', source: 'user_enhancement_failed' },
-      },
-      context,
-    );
-    expect(context.sessionStore.get().generatingNextScene).toBe(false);
-    expect(context.sessionStore.get().waitingForSegmentPrompt).toBe(true);
-  });
-});
-
-describe('opening prompt id tracking', () => {
-  it('routes prompt lifecycle updates to the frontend-recorded opening event', async () => {
-    // The frontend records the opening scene under its own prompt id and sends it
-    // as initial_rollout_prompt_id; the backend echoes it in status updates.
-    const context = buildContext();
-    context.rewriteStore.addPromptEvent({
-      promptId: 'opening-id',
-      status: 'rewrite_requested',
-      source: 'user_rewrite',
-      text: 'a castle at dawn',
-      steeringUserPrompt: true,
-      rawText: 'a castle at dawn',
-    });
-
-    await applyNormalizedSocketEvent(
-      { type: 'prompt/enhancing', payload: { prompt_id: 'opening-id' } },
-      context,
-    );
-    let opening = (context.rewriteStore.get().promptEvents as Record<string, any>[])
-      .find((e) => e.promptId === 'opening-id');
-    expect(opening?.status).toBe('enhancing');
-
-    await applyNormalizedSocketEvent(
-      {
-        type: 'prompt/fallback_used',
-        payload: { prompt_id: 'opening-id', prompt: '', source: 'user_enhancement_failed' },
-      },
-      context,
-    );
-    opening = (context.rewriteStore.get().promptEvents as Record<string, any>[])
-      .find((e) => e.promptId === 'opening-id');
-    expect(opening?.status).toBe('ready_fallback');
-    // A failed opening is dropped from the steering scene list instead of
-    // lingering as a ghost "Scene 1".
-    expect(opening?.steeringFailed).toBe(true);
-  });
-
-  it('marks a prompt-scoped session/error (e.g. safety block) as steeringFailed', async () => {
-    const context = buildContext();
-    context.rewriteStore.addPromptEvent({
-      promptId: 'blocked-id',
-      status: 'queued',
-      source: 'user_raw',
-      text: 'a blocked prompt',
-      steeringUserPrompt: true,
-      rawText: 'a blocked prompt',
-    });
-
-    await applyNormalizedSocketEvent(
-      {
-        type: 'session/error',
-        payload: { message: 'Prompt blocked by safety filter.', prompt_id: 'blocked-id' },
-      },
-      context,
-    );
-    const blocked = (context.rewriteStore.get().promptEvents as Record<string, any>[])
-      .find((e) => e.promptId === 'blocked-id');
-    expect(blocked?.steeringFailed).toBe(true);
-  });
-});

apps/dreamverse/web/src/lib/ws/reducer.ts

@@ -99,19 +99,7 @@ export async function applyNormalizedSocketEvent(event: any, context: any): Prom
 				status: "ready_fallback",
 				source: payload.source || "user_raw",
 				text: payload.prompt,
-				// Steering: this prompt produced no segment — drop it from the scene list.
-				steeringFailed: true,
 			});
-			// Steering recovery: enhancement failed so the backend enqueued nothing AND won't
-			// re-emit prompt_sources_blocked (its drained flag is still set). Put the user back to
-			// "describe the next scene" ourselves so the generating overlay clears and they can retry.
-			if (!uiStore.get().simpleMode && sessionStore.get().manualContinuationMode) {
-				sessionStore.patch({
-					waitingForSegmentPrompt: true,
-					generatingNextScene: false,
-					sessionNotice: "Couldn't continue from that prompt — try rephrasing the next scene.",
-				});
-			}
 			console.warn("[PromptEnhanceFallback] Prompt extension failed for this request.");
 			return;
 
@@ -255,32 +243,19 @@ export async function applyNormalizedSocketEvent(event: any, context: any): Prom
 		}
 
 		case "prompt/sources_blocked":
-			if (sessionStore.get().manualContinuationMode) {
-				sessionStore.patch({ waitingForSegmentPrompt: true, generatingNextScene: false, autoExtensionTimeoutHint: "" });
-			} else {
-				sessionStore.patch({
-					autoExtensionTimeoutHint: uiStore.get().simpleMode ? "" : "blocked on user input, increase prompt count for smoother experience",
-				});
-			}
+			sessionStore.patch({
+				autoExtensionTimeoutHint: uiStore.get().simpleMode ? "" : "blocked on user input, increase prompt count for smoother experience",
+			});
 			return;
 
 		case "prompt/sources_resumed":
-			sessionStore.patch({
-				autoExtensionTimeoutHint: "",
-				waitingForSegmentPrompt: false,
-				// A real prompt was just selected for the next segment; media arriving
-				// (stream/media_init) clears this again.
-				...(sessionStore.get().manualContinuationMode ? { generatingNextScene: true } : {}),
-			});
-			return;
-
 		case "session/auto_extension_updated":
-			// Deliberately does NOT touch generatingNextScene: toggling auto extension
-			// starts no generation.
-			sessionStore.patch({ autoExtensionTimeoutHint: "", waitingForSegmentPrompt: false });
-			console.log("[AutoExtensionUpdated]", {
-				enabled: sessionStore.get().autoExtensionEnabled,
-			});
+			sessionStore.patch({ autoExtensionTimeoutHint: "" });
+			if (event.type === "session/auto_extension_updated") {
+				console.log("[AutoExtensionUpdated]", {
+					enabled: sessionStore.get().autoExtensionEnabled,
+				});
+			}
 			return;
 
 		case "segment/step_complete":
@@ -302,7 +277,6 @@ export async function applyNormalizedSocketEvent(event: any, context: any): Prom
 				projectResetPending: false,
 				sessionExpired: true,
 				sessionNotice: "",
-				generatingNextScene: false,
 			});
 			console.log("Session timed out");
 			clearCountdownInterval();
@@ -380,8 +354,6 @@ export async function applyNormalizedSocketEvent(event: any, context: any): Prom
 			return;
 
 		case "stream/media_init":
-			// Segment media is arriving — the "Generating next scene" phase is over.
-			sessionStore.patch({ generatingNextScene: false });
 			streamStore.patch({
 				mediaAppendError: null,
 				loadingAnimation: streamStore.get().avPlaybackStarted ? streamStore.get().loadingAnimation : true,
@@ -499,30 +471,17 @@ export async function applyNormalizedSocketEvent(event: any, context: any): Prom
 			sessionStore.patch({
 				generationCapReached: false,
 				sessionNotice: "",
-				generatingNextScene: false,
 			});
 			await finalizeStreamCompletion();
 			return;
 
 		case "session/error": {
 			const errorMessage = resolveSessionErrorMessage(payload);
-			if (payload?.prompt_id) {
-				// Prompt-scoped error (e.g. safety-blocked): the prompt produced no
-				// segment, so drop it from the steering scene list.
-				rewriteStore.trackPromptEvent(payload.prompt_id, {
-					steeringFailed: true,
-				});
-			}
 			sessionStore.patch({
 				generationCapReached: false,
 				preservePlaybackOnClose: false,
 				promptExtensionError: "",
 				sessionNotice: errorMessage,
-				// Steering: a blocked/failed prompt produced no segment and the backend won't re-emit
-				// prompt_sources_blocked, so recover the "describe the next scene" state ourselves.
-				...(!uiStore.get().simpleMode && sessionStore.get().manualContinuationMode
-					? { waitingForSegmentPrompt: true, generatingNextScene: false }
-					: {}),
 			});
 			rewriteStore.patch({
 				rewritingSeedPrompts: false,

apps/dreamverse/web/src/stores/session.ts

@@ -24,9 +24,6 @@ export interface SessionState {
   livePromptRewriteMode: boolean;
   sessionExpired: boolean;
   projectResetPending: boolean;
-  manualContinuationMode: boolean;
-  waitingForSegmentPrompt: boolean;
-  generatingNextScene: boolean;
 }
 
 const DEFAULT_SESSION_STATE: SessionState = {
@@ -52,12 +49,6 @@ const DEFAULT_SESSION_STATE: SessionState = {
   livePromptRewriteMode: false,
   sessionExpired: false,
   projectResetPending: false,
-  // Steering-only product: every session drives scenes manually. There is no
-  // auto-rollout mode and no UI selector, so this stays true throughout.
-  manualContinuationMode: true,
-  waitingForSegmentPrompt: false,
-  // True from scene submit / prompt selection until the segment's media arrives.
-  generatingNextScene: false,
 };
 
 export type SessionStore = ManagedStore<SessionState> & {

apps/dreamverse/web/src/stores/stream.ts

@@ -1,18 +1,5 @@
 import { createManagedStore, type ManagedStore } from "./createManagedStore";
 
-// Prompt-history sources driven by the user's own submissions. Curated (non-user)
-// entries — e.g. a preset's opening scene — feed the steering scene list and are
-// exempt from the history cap so Scene 1 survives long sessions.
-export const USER_PROMPT_SOURCES = new Set([
-	"user_raw",
-	"user",
-	"user_enhanced",
-	"user_rewrite",
-	"user_enhancement_failed",
-]);
-
-const PROMPT_HISTORY_CAP = 120;
-
 export interface StreamState {
 	playingSeedPromptIndex: number | null;
 	generatingSeedPromptIndex: number | null;
@@ -172,14 +159,10 @@ export function createStreamStore(initialState: Partial<StreamState> = {}): Stre
 					loopIteration: typeof loopIteration === "number" ? loopIteration : null,
 				};
 
-				const nextHistory = [entry, ...state.promptHistory];
 				return {
 					...state,
 					promptHistoryCounter: nextCounter,
-					promptHistory: nextHistory.length > PROMPT_HISTORY_CAP
-						? nextHistory.filter((item, index) =>
-							index < PROMPT_HISTORY_CAP || !USER_PROMPT_SOURCES.has(String(item.source || "")))
-						: nextHistory,
+					promptHistory: [entry, ...state.promptHistory].slice(0, 120),
 					selectedHistoryId: state.selectedHistoryId || (entry.id as string),
 				};
 			});

apps/performance_dashboard/README.md

@@ -104,11 +104,6 @@ and does not override stored status.
 - `GET /api/performance/trends?days=90&run_source=scheduled_main`
 - `GET /api/performance/records?days=90&run_source=local`
 
-V2 records use the same comparison cohort as CI: `workload_id`, `variant_id`,
-`benchmark_version`, `recipe_fingerprint`, `hardware_profile_id`, and
-`software_profile_id`. `model_id` and `gpu_type` remain display/filter
-metadata, so renaming either does not split history. Legacy records still group
-by `(model_id, gpu_type)`. Dashboard baselines use the latest five previous
-successful, baseline-eligible records in each group. Summary and trend filters
-match the latest display metadata after grouping, while the raw records endpoint
-continues to filter individual records.
+The current v1 grouping key is `(model_id, gpu_type)`. Baselines are computed
+from the latest five previous successful records in each group for dashboard
+context. CI gating uses only records marked `baseline_eligible=true`.

assets/eval/worldmodel_synthetic_flow_calibration.json

@@ -1,22 +0,0 @@
-{
-  "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
-}

docs/contributing/ci_architecture.md

@@ -99,13 +99,6 @@ 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` |

docs/contributing/performance_benchmarks.md

@@ -12,8 +12,7 @@ It serves three audiences:
 * **Maintainers** — surfaces regressions in a Markdown summary on every
   performance build and a long-form Plotly dashboard.
 * **Local developers** — lets you run the same benchmark on your own machine,
-  then compare against the historical baseline for the same comparable
-  identity.
+  then compare against the historical baseline for the same model and GPU.
 
 ## Quick start (local)
 
@@ -80,11 +79,10 @@ fastvideo/performance/
 ```
 
 The HF dataset (`FastVideo/performance-tracking` by default) holds one
-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:
-`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.
+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.
 
 ## Planned Coverage
 
@@ -160,15 +158,13 @@ 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 comparison cohort)
+### Rolling baseline (per `(model_id, gpu_type)`)
 
 `compare_baseline.py` loads the last 5 successful, baseline-eligible records
-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 `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.
+for the same `(model_id, gpu_type)` from the HF dataset, computes the median
+for each available metric, and evaluates the current run with the metric's
+rolling regression policy. 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:
 
@@ -188,33 +184,10 @@ slowly add up. Only scheduled-main successful records are baseline eligible.
 Local and pull-request runs can upload dashboard-visible records, but they do
 not update future gating baselines.
 
-Comparator summaries and normalized artifacts include an explicit
-`comparison_status`:
-
-| Status | Meaning | CI behavior |
-|---|---|---|
-| `PASS` | Comparable baseline exists and no gated metric regressed. Legacy records with no baseline also keep the historical initialization behavior. | Passes |
-| `REGRESSION` | The record exceeds one of its static thresholds or at least one gated metric regressed against a comparable baseline. | Fails |
-| `CALIBRATION_NEEDED` | A v2 record has no exact comparable baseline. | Passes, may upload when `PERF_UPLOAD_POLICY=pass`, never seeds a baseline |
-| `RECIPE_MISMATCH` | The same workload, variant, and benchmark version has trusted successful records under another recipe fingerprint, including records from other hardware or software profiles. | Fails |
-| `INFRA_ERROR` | The comparator cannot safely classify the record, such as a v2 record missing required identity fields. | Fails |
-
-`QUALITY_BLOCKED` is reserved for a future variant-promotion workflow and is
-not emitted by normal rolling-baseline comparison.
-
-For `RECIPE_MISMATCH`, trusted records are scheduled-main records or records
-already marked `baseline_eligible`. PR and local calibration uploads remain
-visible but do not authorize future CI-gating recipe mismatch failures.
-
 When the baseline shifts for a legitimate reason (torch upgrade, kernel
 change, etc.) and CI starts failing, use the
 [`reseed-performance-baseline`](https://github.com/hao-ai-lab/FastVideo/blob/main/.agents/skills/reseed-performance-baseline/SKILL.md)
-agent skill to advance the rolling median. To approve the first baseline for
-a new v2 exact identity, follow that skill with a reviewed scheduled-main
-full-suite `CALIBRATION_NEEDED` normalized artifact. Its prepare step uses
-`fastvideo/tests/performance/seed_baseline.py`; after a separate human gate,
-the skill rechecks the current HF revision and conditionally uploads the whole
-seed batch in one commit.
+agent skill to advance the rolling median.
 
 ## Schemas
 
@@ -230,14 +203,14 @@ configs and remain loadable. New or migrated configs should use
   "config_schema_version": 2,
   "workload_id": "wan-t2v",
   "variant_id": "1.3b-sp2",
-  "benchmark_version": 3
+  "benchmark_version": 2
 }
 ```
 
-`benchmark_id` is still required because raw artifact names, generated-video
-directories, normalized record paths, and legacy storage directories depend on
-it. The v2 comparator does not use it as part of the comparison cohort. The v2
-identity fields make the measured workload explicit:
+`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 |
 |---|---|
@@ -248,26 +221,18 @@ identity fields make the measured workload explicit:
 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 `quality_metadata`
-and the v1/v2 `regression_thresholds` policy must be JSON objects when present.
-(`recipe` is emitted by the harness and is not config-declarable.)
+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.)
 
-V2 records compare only within their exact identity cohort. A record that opens
-a new cohort is marked `baseline_status: "initialized_new_cohort"` and
-`comparison_status: "CALIBRATION_NEEDED"`; it remains ineligible until a
-reviewed scheduled-main artifact is seeded explicitly. 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 are active. `QUALITY_BLOCKED` remains reserved for future variant
-promotion policy.
-
-The shipped Wan benchmark uses `benchmark_version: 3` because recipe schema 2
-changed the recipe fingerprint by removing the legacy `benchmark_id` display
-name. This intentionally opens a new comparison cohort: after deployment, a
-reviewed scheduled-main full-suite `CALIBRATION_NEEDED` artifact must be seeded
-once before rolling regression gating resumes for that exact identity. Static
-thresholds remain active during calibration.
+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 keep
+the `(model_id, gpu_type)` rolling-baseline comparison. Metric-specific
+threshold policies and promoted baselines remain separate follow-ups.
 
 ### Raw record (`results/perf_*.json`)
 
@@ -276,10 +241,10 @@ Written by `test_inference_performance.py`. One file per benchmark run.
 ```jsonc
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
-  "result_schema_version": 2,
+  "config_schema_version": 2,
   "workload_id": "wan-t2v",
   "variant_id": "1.3b-sp2",
-  "benchmark_version": 3,
+  "benchmark_version": 2,
   "model_short_name": "Wan2.1-T2V-1.3B-Diffusers",
   "device": "NVIDIA L40S",
   "num_gpus": 2,
@@ -305,25 +270,18 @@ Written by `test_inference_performance.py`. One file per benchmark run.
     }
   },
   "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,
   "recipe": {
-    "recipe_schema_version": 2,
+    "recipe_schema_version": 1,
     "benchmark": {
       "benchmark_id": "wan-t2v-1.3b-2gpu",
       "workload_id": "wan-t2v",
       "variant_id": "1.3b-sp2",
-      "benchmark_version": 3
+      "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 },
@@ -343,9 +301,6 @@ Written by `test_inference_performance.py`. One file per benchmark run.
     "python": "3.12",
     "pytorch": "2.12",
     "cuda": "13.0",
-    "attention_backend": "FLASH_ATTN",
-    "flash_attention_4_enabled": true,
-    "container_image_version": "py3.12-cuda13.0.0",
     "packages": {
       "fastvideo_kernel": "0.3.2",
       "flashinfer": "0.2.11",
@@ -354,12 +309,7 @@ Written by `test_inference_performance.py`. One file per benchmark run.
     }
   },
   "software_profile_id": "sw-<sha256-prefix>",
-  "environment_metadata": {
-    "env": {
-      "IMAGE_VERSION": "py3.12-cuda13.0.0",
-      "FASTVIDEO_CONTAINER_IMAGE_REF": "ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:py3.12-cuda13.0.0@sha256:<digest>"
-    }
-  },
+  "environment_metadata": { "env": { "IMAGE_VERSION": "py3.12-cuda13.0.0" } },
   "environment_fingerprint": "env-<sha256-prefix>"
 }
 ```
@@ -372,10 +322,9 @@ 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": 3,
+  "benchmark_version": 2,
   "timestamp": "2026-05-08T22:00:00+00:00",
   "commit_sha": "<full sha>",
   "gpu_type": "NVIDIA L40S",
@@ -396,41 +345,21 @@ result, used as the rolling-baseline source of truth.
   "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" },
-  "baseline_status": "compared",
-  "comparison_status": "PASS",
-  "comparison_status_reason": "Comparable baseline found with no gated regressions",
-  "baseline_eligible": false,
   "success": true
 }
 ```
 
 ### Compatibility with legacy records
 
-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.
+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.
 
-New v2 records compare only against the same `workload_id`, `variant_id`,
-`benchmark_version`, `recipe_fingerprint`, `hardware_profile_id`, and
-`software_profile_id` cohort, independent of the legacy `model_id` directory
-and `gpu_type` display string. Historical v1 records remain readable for
-reporting, but current legacy artifacts do not perform a `(model_id, gpu_type)`
-rolling comparison or seed new rolling baselines.
+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
@@ -448,18 +377,13 @@ FlashInfer, Cutlass DSL, SageAttention, Triton, and xFormers when installed.
 | `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` | `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_RUN_SOURCE` | inferred | `compare_baseline.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` | Performance pytest exit code. Measured static-threshold failures are attributed per record; otherwise a nonzero code reports an infrastructure error. |
+| `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` | `fastvideo/performance/hf_store.py` | Freshness window for reusing an existing HF sync when requested by dashboard consumers. |
-| `FASTVIDEO_ATTENTION_BACKEND` | `auto` | `test_inference_performance.py` | Requested attention backend included in `software_profile_id`. |
-| `FASTVIDEO_FA4` | `0` | `test_inference_performance.py` | FlashAttention-4 toggle included in `software_profile_id`. |
-| `FASTVIDEO_PERFORMANCE_PROFILE_VERSION` | unset | `test_inference_performance.py` | Optional explicit software cohort/profile version included in `software_profile_id`. |
-| `IMAGE_VERSION` | unset | `test_inference_performance.py` | CI container image/profile version included in `software_profile_id` when available. |
-| `FASTVIDEO_CONTAINER_IMAGE_REF` | unset | `pr_test.py`, `launch_l40s_job.py`, `test_inference_performance.py` | Resolved CI container image ref or digest recorded in `environment_metadata` for audit without changing `software_profile_id`. |
 | `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
@@ -470,16 +394,11 @@ 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. 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 pytest fails. Each raw
-record is checked against its own static thresholds: a measured breach reports
-`REGRESSION`, while unaffected records retain their rolling-baseline status. A
-nonzero pytest exit with no attributable static-threshold breach reports
-`INFRA_ERROR`. Failed records have `success=false` and are excluded from future
-rolling baselines. The dashboard still runs best-effort for observability.
+Each performance build runs pytest first. If that fixed-threshold phase fails,
+PR/direct runs skip `compare_baseline.py` because they only upload passing
+records. Scheduled-main runs still execute `compare_baseline.py` with
+`PERF_PYTEST_RC` set so the failed canonical attempt is visible in normalized
+JSON and dashboard history. The dashboard runs best-effort for observability.
 When the rolling-baseline phase runs, it emits:
 
 * **Markdown summary** — appended to `$GITHUB_STEP_SUMMARY` when that variable
@@ -487,7 +406,7 @@ When the rolling-baseline phase runs, it emits:
   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 comparison cohort.
+  each metric grouped by `(model_id, gpu_type)`.
 * **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)
@@ -538,13 +457,10 @@ When the rolling-baseline phase runs, it emits:
 2. The pytest test auto-discovers all configs — no test code needed. CI
    picks it up on the next `/test performance` run.
 
-3. Legacy benchmarks are gated only by their static thresholds. Their current
-   records skip rolling-baseline comparison and are never baseline eligible.
-   V2 benchmarks with no exact comparable baseline report
-   `CALIBRATION_NEEDED`; the record remains visible but does not become
-   baseline eligible until a comparable scheduled-main full-suite run is
-   reviewed and seeded through the prepare, review, and conditional-upload
-   steps in the `reseed-performance-baseline` skill.
+3. The first persisted main-branch run with no HF history initializes the
+   baseline (passes automatically). Subsequent runs compare against it. Local
+   and pull-request runs with no HF history also pass, but they do not seed the
+   shared baseline.
 
 4. If the benchmark targets a GPU not currently in `thresholds`, either add
    that GPU as a key or rely on the `default` block. Note that `default` is
@@ -564,12 +480,8 @@ When the rolling-baseline phase runs, it emits:
 
 ## Troubleshooting
 
-**`CALIBRATION_NEEDED` / "No baseline found for exact comparable identity"** —
-the v2 run passes, but its normalized record remains
-`baseline_eligible=false`. Review a successful scheduled-main full-suite
-normalized artifact, then follow the `reseed-performance-baseline` skill. The
-utility only prepares a digest-protected manifest; the separately confirmed
-upload rechecks remote state and commits the batch atomically.
+**"No baseline for ... Initializing"** — first run for this `(model_id,
+gpu_type)`. 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

docs/contributing/testing.md

@@ -180,30 +180,6 @@ 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

@@ -458,8 +458,6 @@ 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

@@ -1,116 +0,0 @@
-# 🌊 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/training/train_infra.md

@@ -161,18 +161,6 @@ 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.
 
@@ -338,40 +326,6 @@ 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_flux_dev.py

@@ -1,140 +0,0 @@
-# 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

@@ -1,107 +0,0 @@
-# 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

@@ -1,37 +0,0 @@
-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

@@ -1,37 +0,0 @@
-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

@@ -1,120 +0,0 @@
-# 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

@@ -1,69 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,38 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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

@@ -1,33 +0,0 @@
-# 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/distribution_matching/wan/anyflow_onpolicy_t2v.yaml

@@ -1,105 +0,0 @@
-# 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

@@ -1,83 +0,0 @@
-# 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

@@ -100,6 +100,9 @@ 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/scenario/worldmodel/README.md

@@ -1,119 +1,32 @@
 # World-Model: Matrix-Game 2.0 I2V
 
-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
+Three training scenarios for the Matrix-Game 2.0 I2V world model on the
+new YAML-driven trainer (`fastvideo/train/entrypoint/train.py`).
 
 | Config | Method | Student | Notes |
 |---|---|---|---|
-| `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`.
+| `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. |
 
 ## Usage
 
-### Solaris
-
 ```bash
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/solaris/finetune_i2v.yaml
+    examples/train/scenario/worldmodel/finetune_i2v.yaml
 
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/solaris/dfsft_causal_i2v.yaml
+    examples/train/scenario/worldmodel/dfsft_causal_i2v.yaml
 
 bash examples/train/run.sh \
-    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
+    examples/train/scenario/worldmodel/self_forcing_causal_i2v.yaml
 ```
 
 Override any field on the command line:
 
 ```bash
 bash examples/train/run.sh \
-    examples/train/scenario/worldmodel/solaris/dfsft_causal_i2v.yaml \
+    examples/train/scenario/worldmodel/dfsft_causal_i2v.yaml \
     --training.distributed.num_gpus 8 \
     --training.optimizer.learning_rate 1e-5
 ```

examples/train/scenario/worldmodel/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

@@ -1,94 +0,0 @@
-# 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

@@ -1,88 +0,0 @@
-# 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

@@ -1,120 +0,0 @@
-# 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

@@ -1,140 +0,0 @@
-# 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/api/flux.py

@@ -1,27 +0,0 @@
-# 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

@@ -90,10 +90,6 @@ 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
@@ -329,18 +325,6 @@ 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

@@ -150,7 +150,6 @@ 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

@@ -1,15 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import importlib.util
 import os
+
 import torch
 import torch.nn.functional as F
 from dataclasses import dataclass
 
-from fastvideo.attention.utils.flash_attn_default import (
-    fa_version,
-    flash_attn_func_compilable,
-)
-
+from fastvideo import envs
 from fastvideo.attention.backends.abstract import (
     AttentionBackend,
     AttentionImpl,
@@ -19,6 +17,119 @@ from fastvideo.attention.backends.abstract import (
 from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
+
+# FA4 (flash_attn.cute) is explicit opt-in via FASTVIDEO_FA4=1, mirroring the
+# kernel package's FASTVIDEO_VSA_CUTEDSL: 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. Below sm90 a capability
+# gate in flash_attn_cute routes to FA2 the calls FA4 cannot serve there:
+# grad-enabled (its backward asserts sm90+) and GQA (pack_gqa fails CuTeDSL
+# JIT, observed on sm_89).
+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 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"
+    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
+
+# 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` (from `flash_attn_cute`) goes through a
+    # registered torch.library custom op (with an FA4 backward on sm90+;
+    # grad-enabled and GQA calls below sm90 route to FA2), 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.")
+
 logger.info("Using FlashAttention-%s backend", fa_version)
 
 # FP4 FA4 support: quantize Q/K to NVFP4 E2M1 for block-scaled MMA on Blackwell.
@@ -198,17 +309,9 @@ 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_compilable as flash_attn_no_pad,
-                flash_attn_varlen_qk_no_pad_compilable as flash_attn_varlen_qk_no_pad,
+                flash_attn_no_pad,
+                flash_attn_varlen_qk_no_pad,
             )
 
             attn_mask = attn_metadata.attn_mask
@@ -235,11 +338,7 @@ class FlashAttentionImpl(AttentionImpl):
                 )
 
             qkv = torch.stack([query, key, value], dim=2)
-            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)
+            attn_mask_padded = F.pad(attn_mask, (qkv.shape[1] - attn_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

@@ -1,147 +0,0 @@
-# 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

@@ -2,7 +2,6 @@
 
 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
@@ -15,18 +14,12 @@ class SDPABackend(AttentionBackend):
     accept_output_buffer: bool = True
 
     @staticmethod
-    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
+    def get_supported_head_sizes() -> list[int]:
+        return [32, 64, 96, 128, 160, 192, 224, 256]
 
     @staticmethod
     def get_name() -> str:
-        return "TORCH_SDPA"
+        return "SDPA"
 
     @staticmethod
     def get_impl_cls() -> type["SDPAImpl"]:
@@ -56,51 +49,9 @@ 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__(
@@ -131,7 +82,6 @@ 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

@@ -252,7 +252,6 @@ 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:
@@ -263,10 +262,7 @@ 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,
-                                        default_backend=default_backend)
+        attn_backend = get_attn_backend(head_size, dtype, supported_attention_backends=supported_attention_backends)
         impl_cls = attn_backend.get_impl_cls()
         self.attn_impl = impl_cls(num_heads=num_heads,
                                   head_size=head_size,

fastvideo/attention/selector.py

@@ -84,9 +84,8 @@ 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, default_backend)
+    return _cached_get_attn_backend(head_size, dtype, supported_attention_backends)
 
 
 @cache
@@ -95,7 +94,6 @@ 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.
@@ -114,12 +112,6 @@ 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_default.py

@@ -1,251 +0,0 @@
-# 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

@@ -24,7 +24,7 @@ from flash_attn.bert_padding import pad_input, unpad_input
 from fastvideo import envs
 
 
-def _resolve_flash_attn_varlen_func() -> tuple[Any, str]:
+def _resolve_flash_attn_varlen_func() -> Any:
     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
@@ -39,28 +39,20 @@ def _resolve_flash_attn_varlen_func() -> tuple[Any, str]:
                                "fix the FA4 install (see the flash-attn-4 pin in pyproject.toml) "
                                "or unset FASTVIDEO_FA4.") from e
 
-        return flash_attn_varlen_func_cute, "4"
+        return flash_attn_varlen_func_cute
     try:
         from flash_attn_interface import (
             flash_attn_varlen_func as flash_attn_varlen_func_interface, )
 
-        return flash_attn_varlen_func_interface, "3"
+        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_flash, "2"
+        return flash_attn_varlen_func_flash
 
 
-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, )
+flash_attn_varlen_func_impl = _resolve_flash_attn_varlen_func()
 
 
 def flash_attn_no_pad(
@@ -199,473 +191,3 @@ 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

@@ -1,9 +1,7 @@
 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
@@ -17,7 +15,6 @@ from fastvideo.configs.models.dits.kandinsky5 import Kandinsky5VideoConfig
 
 __all__ = [
     "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "DreamXWorldConfig",
-    "DreamXWorldARConfig", "CosmosVideoConfig", "Cosmos25VideoConfig", "FluxDiTConfig", "Flux2Config",
-    "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig", "MagiHumanVideoConfig",
-    "StableAudioConfig", "GlmImageDiTConfig"
+    "DreamXWorldARConfig", "CosmosVideoConfig", "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig",
+    "HYWorldConfig", "Kandinsky5VideoConfig", "MagiHumanVideoConfig", "StableAudioConfig", "Flux2Config"
 ]

fastvideo/configs/models/dits/flux.py

@@ -1,27 +0,0 @@
-# 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

@@ -1,61 +0,0 @@
-# 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

@@ -2,24 +2,14 @@
 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: [_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,
-    )
+    _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()
+    ])
 
     # 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

@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 from dataclasses import dataclass, field
-from typing import Literal
 
 from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
 
@@ -20,11 +19,6 @@ 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",
@@ -92,14 +86,6 @@ 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

@@ -43,16 +43,11 @@ class TextEncoderArchConfig(EncoderArchConfig):
     require_processor: bool = False
 
     def __post_init__(self) -> None:
-        # 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 = {
+        self.tokenizer_kwargs = {
             "truncation": True,
             "max_length": self.text_len,
             "return_tensors": "pt",
         }
-        self.tokenizer_kwargs = defaults | self.tokenizer_kwargs
 
 
 @dataclass

fastvideo/configs/models/vaes/__init__.py

@@ -2,7 +2,6 @@ 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
@@ -22,5 +21,4 @@ __all__ = [
     "OobleckVAEArchConfig",
     "OobleckVAEConfig",
     "Flux2VAEConfig",
-    "GlmImageVAEConfig",
 ]

fastvideo/configs/models/vaes/glm_image.py

@@ -1,95 +0,0 @@
-# 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