[bugfix]: select architecture-specific Blackwell kernel targets

- Detect GB200 and GB300 as 10.0a and 10.3a
- Build future GB200 CI images with 10.0a
- Add regression tests for detection, cache keys, and build arguments

.agents/skills/add-model-02-parity/SKILL.md

@@ -68,7 +68,7 @@ Copy `templates/component_parity_test.py` and fill every `TODO` marker. The
 template is distilled from:
 
 - `tests/local_tests/transformers/test_ltx2.py`
-- `tests/local_tests/gen3c/test_gen3c.py`
+- `tests/local_tests/transformers/test_gamecraft_parity.py`
 - `tests/local_tests/encoders/test_ltx2_gemma_parity.py`
 - `tests/local_tests/vaes/test_oobleck_vae_parity.py`
 - `tests/local_tests/sd35/test_sd35_component_parity.py`

.agents/skills/add-model-02-parity/templates/component_parity_test.py

@@ -87,7 +87,7 @@ def _load_official_model(device: torch.device, dtype: torch.dtype) -> torch.nn.M
     # TODO: import official class/factory and load real weights strictly.
     # Examples in-tree:
     # - LTX2: SingleGPUModelBuilder(...).build(device=device, dtype=dtype)
-    # - GEN3C: torch.load(...)["state_dict"] -> official_model.load_state_dict(...)
+    # - GameCraft: torch.load(...)["module"] -> official_model.load_state_dict(...)
     # - Oobleck: create_model_from_config(config) + ckpt state_dict
     OfficialClass = _import_or_skip(OFFICIAL_MODULE, OFFICIAL_CLASS)
     model = OfficialClass()  # TODO: pass official config kwargs.

.agents/skills/add-model-03-port-dit/SKILL.md

@@ -52,8 +52,8 @@ scaling constants, dtype casts, state-dict names, and every output head.
 - Loader path: `TransformerLoader` reads `transformer/config.json`, calls
   `dit_config.update_model_arch(config)`, resolves `_class_name` through
   `ModelRegistry`, and constructs the class with `config` and `hf_config`.
-- Reference examples: `stable_audio.py`, `wanvideo.py`, `sd3.py`, and
-  `ltx2.py`.
+- Reference examples: `stable_audio.py`, `wanvideo.py`, `sd3.py`, `longcat.py`,
+  and `ltx2.py`.
 - Layer guidance: `fastvideo/layers/AGENTS.md`.
 
 ## Implementation Rules

.agents/skills/add-model-04-port-vae/SKILL.md

@@ -51,7 +51,7 @@ posterior behavior, encode/decode output objects, tiling flags, and cropping.
 - Loader path: VAE loaders resolve `_class_name` through `ModelRegistry` and
   load converted component weights from the VAE subdir.
 - Reference examples: `oobleck.py`, `autoencoder_kl.py`, `wanvae.py`,
-  `ltx2vae.py`, and `hunyuanvae.py`.
+  `ltx2vae.py`, and `gamecraftvae.py`.
 - Layer guidance: `fastvideo/layers/AGENTS.md`.
 
 ## Implementation Rules

.agents/skills/add-model-07-conversion/SKILL.md

@@ -43,12 +43,11 @@ from `../add-model/contracts/conversion_request.md`.
 - `scripts/checkpoint_conversion/stable_audio_to_diffusers.py`: monolithic
   `model.safetensors` split into transformer/VAE/conditioner, plus copied
   passthrough subfolders. Use this shape for single-checkpoint official repos.
-- `scripts/checkpoint_conversion/convert_mmaudio_to_diffusers.py`: separate
-  official sources for transformer, VAE, encoders, and vocoder, assembled under
-  a root `model_index.json`.
-- `scripts/checkpoint_conversion/convert_flux2_klein.py`: fused QKV split,
-  renamed native transformer weights, and copied passthrough text encoder,
-  tokenizer, and scheduler components.
+- `scripts/checkpoint_conversion/convert_gamecraft_full.py`: separate official
+  sources for transformer, VAE, text encoders, tokenizers, scheduler, and root
+  `model_index.json`.
+- `scripts/checkpoint_conversion/longcat_to_fastvideo.py`: fused QKV/KV split,
+  renamed native transformer weights, and copied existing Diffusers components.
 - `scripts/checkpoint_conversion/pt_to_safetensors.py`: simple `.pt` extraction
   helper for nested checkpoint dictionaries.
 

.agents/skills/add-model-09-pipeline/SKILL.md

@@ -247,7 +247,7 @@ setup gap, not a pass.
 - `fastvideo/configs/pipelines/stable_audio.py` and
   `fastvideo/pipelines/basic/stable_audio/presets.py` for config/preset shape.
 - `fastvideo/registry.py` for `register_configs(...)` and preset registration.
-- `tests/local_tests/pipelines/test_lingbot_video_pipeline_parity.py` for latent
+- `tests/local_tests/pipelines/test_gamecraft_pipeline_parity.py` for latent
   parity structure.
 - `tests/local_tests/pipelines/test_stable_audio_pipeline_parity.py` for audio
   parity structure.

.agents/skills/add-model/SKILL.md

@@ -413,7 +413,7 @@ matching `*secret*`.
 - `fastvideo/pipelines/basic/wan/` for standard T2V/I2V/DMD/Causal variants.
 - `fastvideo/pipelines/basic/ltx2/` for non-standard stages and audio/video
   patterns.
-- `tests/local_tests/pipelines/test_lingbot_video_pipeline_parity.py` for pipeline
+- `tests/local_tests/pipelines/test_gamecraft_pipeline_parity.py` for pipeline
   parity shape.
 - `tests/local_tests/transformers/test_ltx2.py`,
   `tests/local_tests/vaes/test_ltx2_vae.py`, and

.agents/skills/env-var-conventions/SKILL.md

@@ -38,10 +38,7 @@ on a summary here.
      name to `DEPRECATED_VARIABLES`. Update the uses in `examples/`,
      `scripts/`, `docs/`, `apps/`, and the tests.
 2. **Read the variable with `envs.NAME.get()` inside a function.**
-   - In tests, change the value with `envs.NAME.override(value)`, and a variable
-     outside the registry with `envs.override_external(name, value)`; the
-     `env_overrides` fixture keeps either until the end of the test.
-   - Name a variable that only tests read `FASTVIDEO_TEST_*`.
+   - In tests, change the value with `envs.NAME.override(value)`.
    - Do not call `os.environ`, `os.getenv`, or `monkeypatch.setenv` for a
      FastVideo variable.
    - To set a variable that another tool reads, call `envs.set_external`,

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

@@ -57,7 +57,7 @@ Hardcoded:
 - Quality tier: **`default`**. `full_quality` is a separate, deliberate
   operation.
 - HF repo: `FastVideo/ssim-reference-videos` (override via
-  `FASTVIDEO_TEST_SSIM_REFERENCE_HF_REPO`).
+  `FASTVIDEO_SSIM_REFERENCE_HF_REPO`).
 - Device folder: `L40S_reference_videos`.
 
 ## Prerequisites

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

@@ -105,8 +105,8 @@ Detect artefact type by inspecting the file's imports / helper call:
 - **pixel** (`.mp4`) — file imports
   `run_text_to_video_similarity_test` / `run_image_to_video_similarity_test`
   from `fastvideo.tests.ssim.inference_similarity_utils`, OR uses the
-  legacy custom-inline helper pattern (see `test_gen3c`). Default to pixel
-  when both heuristics fail.
+  legacy custom-inline helper pattern (see `test_gamecraft`,
+  `test_longcat`, etc.). Default to pixel when both heuristics fail.
 
 Record `ARTEFACT_TYPE ∈ {pixel, latent}` for use in step 4. Steps 2, 3, 5,
 and 6 are artefact-type-agnostic — `_iter_reference_files`,

.buildkite/pipeline.yml

@@ -23,100 +23,7 @@ notify:
 # dispatcher, and every test payload executes inside the Slinky Slurm tray.
 # fastvideo/tests/modal remains available only for an explicit manual rollback;
 # no active pipeline or slash-command route invokes it.
-# Buildkite hands jobs to free agents in the order they appear here. Golden-gate comes first
-# because every later merge lane waits for it; the fastcheck lanes follow from longest to
-# shortest measured runtime, so the longest lane never starts last and stretches the build.
 steps:
-  - label: ":test_tube: Golden-Gate Tests"
-    key: "golden-gate"
-    if: |
-      build.env("TEST_SCOPE") == "full" ||
-      (build.env("TEST_SCOPE") == "merge" &&
-       build.env("MERGE_TEST_PLAN") =~ /,golden-gate,/) ||
-      (build.env("TEST_SCOPE") == "direct" &&
-       (build.env("TEST_TYPE") == "golden_gate" || build.env("TEST_TYPE") == "golden_gate_ci"))
-    command: "/opt/fastvideo-ci-runner/run-ci"
-    timeout_in_minutes: 90
-    env:
-      TEST_TYPE: "golden_gate_ci"
-    retry:
-      automatic:
-        - exit_status: 128
-          limit: 3
-        - exit_status: -1
-          limit: 2
-    agents:
-      queue: "ci-runner"
-
-  - label: ":microscope: Unit Tests"
-    key: "unit"
-    if: |
-      build.env("TEST_SCOPE") == "full" ||
-      (build.env("TEST_SCOPE") == "merge" &&
-       build.env("MERGE_TEST_PLAN") =~ /,unit,/) ||
-      build.env("TEST_SCOPE") == "fastcheck" ||
-      build.env("TEST_SCOPE") == null ||
-      (build.env("TEST_SCOPE") == "direct" &&
-       (build.env("TEST_TYPE") == "unit_test" || build.env("TEST_TYPE") == "unit_test_ci"))
-    command: "/opt/fastvideo-ci-runner/run-unit"
-    timeout_in_minutes: 90
-    env:
-      TEST_TYPE: "unit_test_ci"
-    retry:
-      automatic:
-        - exit_status: 128
-          limit: 3
-        - exit_status: -1
-          limit: 2
-    agents:
-      queue: "ci-runner"
-
-  - label: ":microscope: Kernel Tests"
-    key: "kernel-tests"
-    if: |
-      build.env("TEST_SCOPE") == "full" ||
-      (build.env("TEST_SCOPE") == "merge" &&
-       build.env("MERGE_TEST_PLAN") =~ /,kernel-tests,/) ||
-      build.env("TEST_SCOPE") == "fastcheck" ||
-      build.env("TEST_SCOPE") == null ||
-      (build.env("TEST_SCOPE") == "direct" &&
-       (build.env("TEST_TYPE") == "kernel_tests" || build.env("TEST_TYPE") == "kernel_tests_ci"))
-    command: "/opt/fastvideo-ci-runner/run-ci"
-    timeout_in_minutes: 90
-    env:
-      TEST_TYPE: "kernel_tests_ci"
-    retry:
-      automatic:
-        - exit_status: 128
-          limit: 3
-        - exit_status: -1
-          limit: 2
-    agents:
-      queue: "ci-runner"
-
-  - label: ":microscope: DreamVerse App Tests"
-    key: "dreamverse"
-    if: |
-      build.env("TEST_SCOPE") == "full" ||
-      (build.env("TEST_SCOPE") == "merge" &&
-       build.env("MERGE_TEST_PLAN") =~ /,dreamverse,/) ||
-      build.env("TEST_SCOPE") == "fastcheck" ||
-      build.env("TEST_SCOPE") == null ||
-      (build.env("TEST_SCOPE") == "direct" &&
-       (build.env("TEST_TYPE") == "dreamverse_app" || build.env("TEST_TYPE") == "dreamverse_app_ci"))
-    command: "/opt/fastvideo-ci-runner/run-ci"
-    timeout_in_minutes: 90
-    env:
-      TEST_TYPE: "dreamverse_app_ci"
-    retry:
-      automatic:
-        - exit_status: 128
-          limit: 3
-        - exit_status: -1
-          limit: 2
-    agents:
-      queue: "ci-runner"
-
   - label: ":microscope: Encoder Tests"
     key: "encoder"
     if: |
@@ -186,6 +93,96 @@ steps:
     agents:
       queue: "ci-runner"
 
+  - label: ":microscope: Kernel Tests"
+    key: "kernel-tests"
+    if: |
+      build.env("TEST_SCOPE") == "full" ||
+      (build.env("TEST_SCOPE") == "merge" &&
+       build.env("MERGE_TEST_PLAN") =~ /,kernel-tests,/) ||
+      build.env("TEST_SCOPE") == "fastcheck" ||
+      build.env("TEST_SCOPE") == null ||
+      (build.env("TEST_SCOPE") == "direct" &&
+       (build.env("TEST_TYPE") == "kernel_tests" || build.env("TEST_TYPE") == "kernel_tests_ci"))
+    command: "/opt/fastvideo-ci-runner/run-ci"
+    timeout_in_minutes: 90
+    env:
+      TEST_TYPE: "kernel_tests_ci"
+    retry:
+      automatic:
+        - exit_status: 128
+          limit: 3
+        - exit_status: -1
+          limit: 2
+    agents:
+      queue: "ci-runner"
+
+  - label: ":microscope: Unit Tests"
+    key: "unit"
+    if: |
+      build.env("TEST_SCOPE") == "full" ||
+      (build.env("TEST_SCOPE") == "merge" &&
+       build.env("MERGE_TEST_PLAN") =~ /,unit,/) ||
+      build.env("TEST_SCOPE") == "fastcheck" ||
+      build.env("TEST_SCOPE") == null ||
+      (build.env("TEST_SCOPE") == "direct" &&
+       (build.env("TEST_TYPE") == "unit_test" || build.env("TEST_TYPE") == "unit_test_ci"))
+    command: "/opt/fastvideo-ci-runner/run-unit"
+    timeout_in_minutes: 90
+    env:
+      TEST_TYPE: "unit_test_ci"
+    retry:
+      automatic:
+        - exit_status: 128
+          limit: 3
+        - exit_status: -1
+          limit: 2
+    agents:
+      queue: "ci-runner"
+
+  - label: ":microscope: DreamVerse App Tests"
+    key: "dreamverse"
+    if: |
+      build.env("TEST_SCOPE") == "full" ||
+      (build.env("TEST_SCOPE") == "merge" &&
+       build.env("MERGE_TEST_PLAN") =~ /,dreamverse,/) ||
+      build.env("TEST_SCOPE") == "fastcheck" ||
+      build.env("TEST_SCOPE") == null ||
+      (build.env("TEST_SCOPE") == "direct" &&
+       (build.env("TEST_TYPE") == "dreamverse_app" || build.env("TEST_TYPE") == "dreamverse_app_ci"))
+    command: "/opt/fastvideo-ci-runner/run-ci"
+    timeout_in_minutes: 90
+    env:
+      TEST_TYPE: "dreamverse_app_ci"
+    retry:
+      automatic:
+        - exit_status: 128
+          limit: 3
+        - exit_status: -1
+          limit: 2
+    agents:
+      queue: "ci-runner"
+
+  - label: ":test_tube: Golden-Gate Tests"
+    key: "golden-gate"
+    if: |
+      build.env("TEST_SCOPE") == "full" ||
+      (build.env("TEST_SCOPE") == "merge" &&
+       build.env("MERGE_TEST_PLAN") =~ /,golden-gate,/) ||
+      (build.env("TEST_SCOPE") == "direct" &&
+       (build.env("TEST_TYPE") == "golden_gate" || build.env("TEST_TYPE") == "golden_gate_ci"))
+    command: "/opt/fastvideo-ci-runner/run-ci"
+    timeout_in_minutes: 90
+    env:
+      TEST_TYPE: "golden_gate_ci"
+    retry:
+      automatic:
+        - exit_status: 128
+          limit: 3
+        - exit_status: -1
+          limit: 2
+    agents:
+      queue: "ci-runner"
+
   - label: ":bar_chart: SSIM Tests"
     key: "ssim"
     depends_on: "golden-gate"

.buildkite/scripts/lanes/performance.sh

@@ -22,6 +22,12 @@ else
   export PERF_UPLOAD_POLICY=never
 fi
 
+# Alternate GPU backends compare against references without publishing records.
+# Their worker has read-only Hub credentials; publication is an operator task.
+if [ "${FASTVIDEO_CI_LOCAL_ONLY:-0}" = 1 ]; then
+  export PERF_UPLOAD_POLICY=never
+fi
+
 nvidia-smi \
   --query-gpu=index,timestamp,clocks.sm,clocks.max.sm,power.draw,power.limit,temperature.gpu \
   --format=csv -l 10 > "$PERF_REPORTS_DIR/gpu_telemetry.csv" 2>/dev/null &

.buildkite/scripts/unit_test.sh

@@ -15,12 +15,8 @@ exec pytest \
   ./fastvideo/tests/ops/ \
   ./fastvideo/tests/worker/ \
   ./fastvideo/tests/training/test_trackers.py \
-  ./fastvideo/tests/inference/test_basic_fasth3_omniref_pdd.py \
   ./fastvideo/tests/attention/test_sdpa_metadata_mask_contract.py \
   ./fastvideo/tests/attention/test_vsa_h3_tile_grad_safety.py \
-  ./fastvideo/tests/attention/test_vsa_h3_metadata.py \
-  ./fastvideo/tests/attention/test_vsa_h3_ref2va_regions.py \
-  ./fastvideo/tests/layers/test_pdd_linear.py \
   ./fastvideo/tests/modal/test_kernel_build_cache.py \
   ./fastvideo/tests/modal/test_pr_test.py \
   ./fastvideo/tests/modal/test_ssim_test.py \

.github/scripts/plan_merge_ci.py

@@ -12,7 +12,6 @@ from __future__ import annotations
 import argparse
 import fnmatch
 import re
-from collections.abc import Iterable
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TextIO
@@ -131,6 +130,11 @@ class FamilyCoverage:
 
 
 FAMILY_COVERAGE = (
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])dreamx(_world)?([/_.-]|$)"),
+        ("test_dreamx.py", ),
+        ("test_dreamx_world_similarity.py", ),
+    ),
     FamilyCoverage(
         re.compile(r"(^|[/_.-])flux[_-]?2([/_.-]|$)"),
         ("test_flux2_klein.py", ),
@@ -141,11 +145,21 @@ FAMILY_COVERAGE = (
         ("test_flux.py", ),
         ("test_flux_t2i_similarity.py", ),
     ),
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])(hunyuan)?gamecraft([/_.-]|$)"),
+        ("test_gamecraft.py", ),
+        ("test_gamecraft_similarity.py", ),
+    ),
     FamilyCoverage(
         re.compile(r"(^|[/_.-])gen3c([/_.-]|$)"),
         ("test_gen3c.py", ),
         ("test_gen3c_similarity.py", ),
     ),
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])glm[_-]?image([/_.-]|$)"),
+        ("test_glm_image.py", ),
+        ("test_glm_image_similarity.py", ),
+    ),
     FamilyCoverage(
         re.compile(r"(^|[/_.-])kandinsky[_-]?5([/_.-]|$)"),
         ("test_kandinsky5.py", ),
@@ -156,6 +170,11 @@ FAMILY_COVERAGE = (
         ("test_lingbot.py", ),
         ("test_lingbot_similarity.py", ),
     ),
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])longcat([/_.-]|$)"),
+        ("test_longcat.py", ),
+        ("test_longcat_similarity.py", ),
+    ),
     FamilyCoverage(
         re.compile(r"(^|[/_.-])ltx[_-]?2([/_.-]|$)"),
         ("test_ltx2.py", ),
@@ -186,6 +205,11 @@ FAMILY_COVERAGE = (
         ("test_stable_audio.py", ),
         ("test_stable_audio_similarity.py", ),
     ),
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])turbo(diffusion)?([/_.-]|$)"),
+        (),
+        ("test_turbodiffusion_similarity.py", ),
+    ),
     FamilyCoverage(
         re.compile(r"(^|[/_.-])wan(video|vae)?([/_.-]|$)"),
         ("test_wan_t2v.py", "test_wan_vae.py", "test_wan_causal.py", "test_wan_denoising.py"),
@@ -195,6 +219,11 @@ FAMILY_COVERAGE = (
             "test_wan_t2v_similarity.py",
         ),
     ),
+    FamilyCoverage(
+        re.compile(r"(^|[/_.-])z[_-]?image([/_.-]|$)"),
+        ("test_zimage.py", ),
+        ("test_zimage_similarity.py", ),
+    ),
 )
 
 
@@ -290,22 +319,16 @@ def _select_output_coverage(plan: MergePlan, path: str) -> None:
         plan.add_ssim(SSIM_SMOKE_TESTS, reason=f"shared output SSIM smoke coverage: {path}")
 
 
-def _normalize_path(raw_path: str) -> str:
-    path = raw_path.strip()
-    while path.startswith("./"):
-        path = path[2:]
-    return path
-
-
-def classify_paths(paths: list[str], removed_paths: Iterable[str] = ()) -> MergePlan:
-    """Plan merge lanes for ``paths``.
-
-    ``removed_paths`` lists changed paths that no longer exist at the PR head
-    (deleted files and rename sources); removed golden/SSIM tests are not run.
-    """
+def classify_paths(paths: list[str]) -> MergePlan:
     plan = MergePlan()
-    normalized_paths = sorted({path for path in map(_normalize_path, paths) if path})
-    removed = {path for path in map(_normalize_path, removed_paths) if path}
+    normalized_paths: list[str] = []
+    for raw_path in paths:
+        path = raw_path.strip()
+        while path.startswith("./"):
+            path = path[2:]
+        if path:
+            normalized_paths.append(path)
+    normalized_paths = sorted(set(normalized_paths))
     if not normalized_paths:
         plan.require_all("changed-file list was empty; failing closed")
         return plan
@@ -345,10 +368,7 @@ def classify_paths(paths: list[str], removed_paths: Iterable[str] = ()) -> Merge
         if path.startswith("fastvideo/tests/golden_gate/"):
             name = Path(path).name
             if name.startswith("test_") and name.endswith(".py"):
-                if path in removed:
-                    plan.reasons.append(f"removed golden test has nothing to run: {path}")
-                else:
-                    plan.add_golden((name, ), reason=f"changed golden test: {path}")
+                plan.add_golden((name, ), reason=f"changed golden test: {path}")
             elif name in {"AGENTS.md", "README.md"}:
                 plan.reasons.append(f"golden documentation only: {path}")
             else:
@@ -359,10 +379,7 @@ def classify_paths(paths: list[str], removed_paths: Iterable[str] = ()) -> Merge
         if path.startswith("fastvideo/tests/ssim/"):
             name = Path(path).name
             if name.startswith("test_") and name.endswith(".py"):
-                if path in removed:
-                    plan.reasons.append(f"removed SSIM test has nothing to run: {path}")
-                else:
-                    plan.add_ssim((name, ), reason=f"changed SSIM test: {path}")
+                plan.add_ssim((name, ), reason=f"changed SSIM test: {path}")
             elif path.endswith((".py", ".json", ".pt", ".png", ".mp4")):
                 plan.ssim_all = True
                 plan.add_lanes("ssim", reason=f"shared SSIM harness/reference: {path}")
@@ -538,7 +555,6 @@ def _write_summary(output: TextIO, plan: MergePlan) -> None:
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--paths-file", type=Path, required=True)
-    parser.add_argument("--removed-paths-file", type=Path)
     parser.add_argument("--github-output", type=Path)
     parser.add_argument("--summary-file", type=Path)
     return parser.parse_args()
@@ -547,9 +563,7 @@ def parse_args() -> argparse.Namespace:
 def main() -> int:
     args = parse_args()
     paths = args.paths_file.read_text(encoding="utf-8").splitlines()
-    removed_paths = (args.removed_paths_file.read_text(encoding="utf-8").splitlines()
-                     if args.removed_paths_file else [])
-    plan = classify_paths(paths, removed_paths)
+    plan = classify_paths(paths)
     print(f"MERGE_TEST_PLAN={plan.encoded_lanes()}")
     print(f"MERGE_GOLDEN_TESTS={plan.encoded_golden_tests()}")
     print(f"MERGE_SSIM_TESTS={plan.encoded_ssim_tests()}")

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

@@ -11,6 +11,7 @@ jobs:
     if: >-
       github.event.context == 'direct-test-completed'
       && github.event.state == 'success'
+      && (vars.CI_GPU_BACKEND == '' || vars.CI_GPU_BACKEND == 'slurm')
     runs-on: ubuntu-latest
     steps:
       - name: Check and update aggregate status

.github/workflows/ci-gpu-backend-status.yml

@@ -0,0 +1,62 @@
+name: Promote Selected GPU Backend Status
+
+on:
+  status:
+
+permissions:
+  statuses: write
+
+concurrency:
+  group: gpu-ci-status-${{ github.event.sha }}-${{ vars.CI_GPU_BACKEND }}
+  cancel-in-progress: false
+
+jobs:
+  promote:
+    if: >-
+      (vars.CI_GPU_BACKEND == 'modal' || vars.CI_GPU_BACKEND == 'vllm')
+      && (github.event.context == format('gpu-ci/{0}/fastcheck-passed', vars.CI_GPU_BACKEND)
+          || github.event.context == format('gpu-ci/{0}/full-suite-passed', vars.CI_GPU_BACKEND))
+    runs-on: ubuntu-latest
+    env:
+      SELECTED_BACKEND: ${{ vars.CI_GPU_BACKEND }}
+    steps:
+      - name: Mirror the selected backend's latest suite results
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        with:
+          script: |
+            const backend = process.env.SELECTED_BACKEND;
+            if (!['modal', 'vllm'].includes(backend)) {
+              throw new Error('Unsupported selected GPU backend');
+            }
+            const sha = context.payload.sha;
+            // Read current state after entering the serialized workflow. A
+            // delayed event must not overwrite a newer failure with success.
+            const statuses = await github.paginate(github.rest.repos.listCommitStatusesForRef, {
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              ref: sha,
+              per_page: 100,
+            });
+            for (const suffix of ['fastcheck-passed', 'full-suite-passed']) {
+              const sourceContext = `gpu-ci/${backend}/${suffix}`;
+              const matches = statuses.filter(status => status.context === sourceContext);
+              matches.sort((a, b) =>
+                Date.parse(b.updated_at) - Date.parse(a.updated_at) || b.id - a.id
+              );
+              const latest = matches[0];
+              const state = latest ? latest.state : 'pending';
+              if (!['pending', 'success', 'failure', 'error'].includes(state)) {
+                throw new Error(`Unsupported status state for ${sourceContext}`);
+              }
+              await github.rest.repos.createCommitStatus({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                sha,
+                context: suffix,
+                state,
+                description: latest
+                  ? `${backend} ${suffix}: ${state}`
+                  : `Waiting for ${backend} ${suffix}`,
+                ...(latest && latest.target_url ? {target_url: latest.target_url} : {}),
+              });
+            }

.github/workflows/ci-macos-mlx.yml

@@ -33,15 +33,15 @@ jobs:
     env:
       FASTVIDEO_ATTENTION_BACKEND: TORCH_SDPA
       TOKENIZERS_PARALLELISM: "false"
-      MASTER_ADDR: "127.0.0.1"
+      MASTER_ADDR: localhost
       MASTER_PORT: "29513"
-      GLOO_SOCKET_IFNAME: lo0
     steps:
       - uses: actions/checkout@v4
 
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
+          cache: pip
 
       - uses: astral-sh/setup-uv@v3
 
@@ -49,9 +49,9 @@ jobs:
         run: |
           uv pip install --system \
             --index-url https://download.pytorch.org/whl/cpu \
-            torch==2.12.0 torchvision torchaudio
+            torch==2.11.0 torchvision torchaudio
           uv pip install --system \
-            pytest pytest-timeout numpy scipy pillow imageio einops cloudpickle filelock \
+            pytest numpy scipy pillow imageio einops cloudpickle filelock \
             PyYAML diffusers huggingface_hub remote-pdb safetensors loguru mlx \
             "ftfy>=6.3.1" "opencv-python>=4.10.0.84" psutil "transformers>=5.0.0"
 
@@ -65,8 +65,8 @@ jobs:
           print("machine:", platform.machine())
           print("processor:", platform.processor())
           print("mlx default device:", mx.default_device())
-          device_info = mx.metal.device_info() if mx.metal.is_available() else "metal unavailable"
-          print("mlx device_info:", device_info)
+          memory_size = mx.metal.device_info().get("memory_size") if mx.metal.is_available() else "metal unavailable"
+          print("mlx memory_size:", memory_size)
           print("torch:", torch.__version__)
           print("torch mps available:", torch.backends.mps.is_available())
           PY
@@ -74,7 +74,6 @@ jobs:
       - name: Run MLX smoke tests
         run: |
           python -m pytest \
-            fastvideo/mlx_runtime/tests/ \
             fastvideo/tests/mlx/test_dmd_sampling.py \
             fastvideo/tests/mlx/test_memory_limits.py \
             fastvideo/tests/mlx/test_quant_capability.py \
@@ -102,7 +101,7 @@ jobs:
             fastvideo/tests/mlx/test_mlx_rife_interpolation.py::test_rife_backend_regression_is_not_skip_eligible \
             fastvideo/tests/platforms/test_mps_vsa_error.py \
             fastvideo/tests/platforms/test_cpu_sdpa.py \
-            -v -s --timeout=120 -o faulthandler_timeout=120
+            -v -s -o faulthandler_timeout=120
 
   # Same tests on MLX's CPU backend. Hosted macOS runners are scarce and
   # slower to schedule; this Linux job gives fast PR signal on the identical
@@ -123,6 +122,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
+          cache: pip
 
       - uses: astral-sh/setup-uv@v3
 
@@ -130,16 +130,15 @@ jobs:
         run: |
           uv pip install --system \
             --index-url https://download.pytorch.org/whl/cpu \
-            torch==2.12.0 torchvision torchaudio
+            torch==2.11.0 torchvision torchaudio
           uv pip install --system \
-            pytest pytest-timeout numpy scipy pillow imageio einops cloudpickle filelock \
+            pytest numpy scipy pillow imageio einops cloudpickle filelock \
             PyYAML diffusers huggingface_hub remote-pdb safetensors loguru "mlx[cpu]" \
             "ftfy>=6.3.1" "opencv-python>=4.10.0.84" psutil "transformers>=5.0.0"
 
       - name: Run MLX smoke tests (CPU backend)
         run: |
           python -m pytest \
-            fastvideo/mlx_runtime/tests/ \
             fastvideo/tests/mlx/test_dmd_sampling.py \
             fastvideo/tests/mlx/test_memory_limits.py \
             fastvideo/tests/mlx/test_quant_capability.py \
@@ -167,4 +166,4 @@ jobs:
             fastvideo/tests/mlx/test_mlx_rife_interpolation.py::test_rife_backend_regression_is_not_skip_eligible \
             fastvideo/tests/platforms/test_mps_vsa_error.py \
             fastvideo/tests/platforms/test_cpu_sdpa.py \
-            -v -s --timeout=120 -o faulthandler_timeout=120
+            -v -s -o faulthandler_timeout=120

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

@@ -76,8 +76,6 @@ jobs:
           set -euo pipefail
           changed_json="$RUNNER_TEMP/merge-changed-files.json"
           changed_paths="$RUNNER_TEMP/merge-changed-paths.txt"
-          removed_paths="$RUNNER_TEMP/merge-removed-paths.txt"
-          : > "$removed_paths"
           if gh api --paginate --slurp \
               "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files?per_page=100" \
               > "$changed_json"; then
@@ -85,10 +83,6 @@ jobs:
             if [ "$observed" = "$EXPECTED_CHANGED_FILES" ]; then
               jq -r '.[][] | .filename, (.previous_filename // empty)' "$changed_json" \
                 | sort -u > "$changed_paths"
-              # Paths absent from the PR head, so the planner never selects a deleted test.
-              jq -r '.[][] | if .status == "removed" then .filename
-                  elif .status == "renamed" then (.previous_filename // empty) else empty end' \
-                "$changed_json" | sort -u > "$removed_paths"
             else
               echo "::warning::Changed-file API returned $observed of $EXPECTED_CHANGED_FILES paths; selecting all merge lanes."
               echo '__FASTVIDEO_CI_PLAN_ALL__' > "$changed_paths"
@@ -104,7 +98,6 @@ jobs:
         run: |
           python3 .github/scripts/plan_merge_ci.py \
             --paths-file "$RUNNER_TEMP/merge-changed-paths.txt" \
-            --removed-paths-file "$RUNNER_TEMP/merge-removed-paths.txt" \
             --github-output "$GITHUB_OUTPUT" \
             --summary-file "$GITHUB_STEP_SUMMARY"
 

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

@@ -203,7 +203,8 @@ jobs:
           docker buildx imagetools create "${TAG_ARGS[@]}" "${IMAGE_REFS[@]}"
           docker buildx imagetools inspect "${TAGS[0]}"
 
-  # The CI runner is ARM64 like DGX Spark, but targets sm_100 rather than sm_121.
+  # The CI runner is ARM64 like DGX Spark, but targets sm_100a rather than sm_121.
+  # The architecture-specific target includes the GB200 VSA CUDA extensions.
   # Publish a single-architecture variant so the self-hosted CI runner can reuse
   # the exact prebuilt kernel instead of compiling it in every job.
   build-ci-runner-image:
@@ -219,7 +220,7 @@ jobs:
         PYTHON_VERSION=3.12
         CUDA_VERSION=13.0.0
         UV_TORCH_BACKEND=cu130
-        TORCH_CUDA_ARCH_LIST=10.0
+        TORCH_CUDA_ARCH_LIST=10.0a
         CMAKE_BUILD_PARALLEL_LEVEL=1
         FLASH_ATTN_WHEEL_TAG=cu130torch2.12
         FLASH_ATTN_WHEEL_RELEASE_ARM64=https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.22

.gitignore

@@ -36,6 +36,7 @@ env
 *.log
 weights/
 logs/
+/Z-Image/
 official_weights/
 converted_weights/
 
@@ -133,7 +134,6 @@ openspec/
 fastvideo/tests/ssim/reference_videos/**
 !fastvideo/tests/ssim/reference_videos/**/*.mp4
 !fastvideo/tests/ssim/reference_videos/**/*.png
-fastvideo/tests/ssim/.reference_videos_download.lock
 
 # Local H3 MLX kernel / exactness benches (JSON, logs, frames, videos)
 .kernel_bench/
@@ -142,5 +142,3 @@ fastvideo/tests/ssim/.reference_videos_download.lock
 *.nvimlog
 .nvimlog
 .python-version
-scripts/benchmarks/minimax_h3_pro6000/headline_results/
-fastvideo/tests/ssim/.reference_videos_download.lock

README.md

@@ -9,8 +9,6 @@
 **FastVideo is a unified post-training and real-time inference framework for accelerated video generation.**
 
 ## NEWS
-- `2026/10/06`: FastH3 V2 now runs on a single consumer machine: NVIDIA RTX 5090, RTX 4090 and RTX PRO 6000 GPUs, DGX Spark and Apple Silicon. We also release [FastH3 Trim](https://huggingface.co/FastVideo/FastVideo-FastH3-Trim-8-Step-NVFP4), an experimental pruned model that is 4.2× smaller than base H3 and runs in as little as 8 GB of GPU memory. Get the [models](https://huggingface.co/collections/FastVideo/fastvideo-fasth3) and read the [Blog](https://haoailab.com/blogs/fasth3-rtx/).
-- `2026/10/06`: FastVideo now supports [Kandinsky 6](https://x.com/kandinskylab_ai/status/2107374635218055345) from Kandinsky Lab: text- and image-to-video with synchronized audio (base and 10-step distilled pi-Flow checkpoints) plus video super-resolution up to 4x. See the [Kandinsky 6 recipes](https://haoailab.com/FastVideo/cookbook/kandinsky6/).
 - `2026/09/15`: Release [FastH3 8-Step V2](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2), an eight-forward data-free DMD2 checkpoint distilled from MiniMax-H3 with 80% Video Sparse Attention. Run it with `examples/inference/basic/basic_fasth3_8step.py` or the [FastH3 8-Step V2 recipe](https://haoailab.com/FastVideo/cookbook/minimax-h3/).
 - `2026/09/01`: FastH3 now runs locally on Apple Silicon through MLX and on NVIDIA DGX Spark through CUDA 13, including two-Spark inference. Follow the [FastH3 recipes](https://haoailab.com/FastVideo/cookbook/minimax-h3/) and read the [Blog](https://haoailab.com/blogs/fasth3-local/).
 - `2026/08/27`: [FastH3 Preview v1](https://haoailab.com/blogs/fasth3-preview/) is an open-weight 4-step sparse-distilled MiniMax-H3 model for synchronized video-and-audio generation, developed in collaboration with [Nuva Lab](https://nuvalab.ai/) and the [NVIDIA FastGen team](https://github.com/NVlabs/FastGen). Download the recommended [VSA / Data-Free weights](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree), or see the [full FastH3 collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3).

docs/assets/cookbook-recipes.json

@@ -1,5 +1,5 @@
 {
-  "version": 15,
+  "version": 11,
   "recipes": [
     {
       "id": "fastwan21-t2v",
@@ -10,11 +10,6 @@
       "summary": "Generate a video in three denoising steps with the distilled FastWan2.1 1.3B checkpoint and video sparse attention.",
       "model": "FastVideo/FastWan2.1-T2V-1.3B-Diffusers",
       "source": "scripts/inference/inference_wan_VSA_DMD_1_3B.yaml",
-      "serving": {
-        "source": "examples/serving/openai_fastwan21_1_3b.yaml",
-        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e .",
-        "env": "FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN"
-      },
       "command": "FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN fastvideo generate --config scripts/inference/inference_wan_VSA_DMD_1_3B.yaml",
       "gpu_types": ["NVIDIA"],
       "hardware": {"gpu_count": 1, "evidence": "source-configured"},
@@ -30,13 +25,6 @@
       "summary": "The maintained high-capacity Wan2.2 text-to-video example with CPU offload settings encoded in its checked-in Python source.",
       "model": "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
       "source": "examples/inference/basic/basic_wan2_2.py",
-      "serving": {
-        "source": "examples/serving/openai_wan22_t2v_a14b.yaml",
-        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e .",
-        "limitations": [
-          "Generation at 1280 × 720 on two GPUs with DiT CPU offload is slow. The client examples stop polling after 30 minutes while the job keeps running on the server; retrieve it by ID or raise the deadline."
-        ]
-      },
       "command": "python examples/inference/basic/basic_wan2_2.py",
       "gpu_types": ["NVIDIA"],
       "hardware": {"gpu_count": 2, "evidence": "source-configured"},
@@ -66,14 +54,6 @@
       "summary": "Use one maintained 5B checkpoint for text-to-video or add an image input to switch the same recipe to image-to-video.",
       "model": "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
       "source": "examples/inference/basic/basic_wan2_2_ti2v.py",
-      "serving": {
-        "source": "examples/serving/openai_wan22_ti2v_5b.yaml",
-        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e .",
-        "task": "Text to video",
-        "limitations": [
-          "This server walkthrough sends a text prompt only. The server API also accepts an image reference (input_reference) for this checkpoint, but the checked-in clients do not upload one. For the cookbook's image-to-video workflow, use the Python command."
-        ]
-      },
       "command": "python examples/inference/basic/basic_wan2_2_ti2v.py",
       "gpu_types": ["NVIDIA"],
       "hardware": {"gpu_count": 1, "evidence": "source-configured"},
@@ -161,6 +141,53 @@
         "Native MLX FastMetal T2V on 36 GB+ unified memory. Same --fast, --fast-spatial, and --refine flags as the 1.3B script."
       ]
     },
+    {
+      "id": "turbodiffusion-wan21-1-3b-t2v",
+      "family": "turbodiffusion",
+      "stage": "inference",
+      "task": "Text to video",
+      "label": "TurboWan2.1 1.3B",
+      "summary": "A TurboDiffusion-accelerated Wan2.1 1.3B text-to-video run from its maintained single-GPU example.",
+      "model": "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers",
+      "source": "examples/inference/basic/basic_turbodiffusion.py",
+      "command": "python examples/inference/basic/basic_turbodiffusion.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "MP4 videos under video_samples_turbodiffusion/",
+      "related": ["turbodiffusion-wan21-14b-t2v", "turbowan22-i2v"]
+    },
+    {
+      "id": "turbodiffusion-wan21-14b-t2v",
+      "family": "turbodiffusion",
+      "stage": "inference",
+      "task": "Text to video",
+      "label": "TurboWan2.1 14B",
+      "summary": "TurboDiffusion acceleration applied to the 14B Wan2.1 text-to-video checkpoint; the checked-in source is configured for two GPUs.",
+      "model": "loayrashid/TurboWan2.1-T2V-14B-Diffusers",
+      "source": "examples/inference/basic/basic_turbodiffusion_14b.py",
+      "command": "python examples/inference/basic/basic_turbodiffusion_14b.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 2, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "MP4 videos under video_samples_turbodiffusion_14B/",
+      "related": ["turbodiffusion-wan21-1-3b-t2v", "turbowan22-i2v"]
+    },
+    {
+      "id": "turbowan22-i2v",
+      "family": "turbodiffusion",
+      "stage": "inference",
+      "task": "Image to video",
+      "label": "TurboWan2.2 A14B",
+      "summary": "A one-to-four-step image-to-video path using TurboDiffusion and the SLA attention backend from its maintained example.",
+      "model": "loayrashid/TurboWan2.2-I2V-A14B-Diffusers",
+      "source": "examples/inference/basic/basic_turbodiffusion_i2v.py",
+      "command": "python examples/inference/basic/basic_turbodiffusion_i2v.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 2, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "related": ["turbodiffusion-wan21-14b-t2v"]
+    },
     {
       "id": "ltx2-distilled-t2v",
       "family": "ltx2",
@@ -285,70 +312,6 @@
       "expected_artifact": "MP4 videos under video_samples_kandinsky5_i2v/",
       "related": ["kandinsky5-t2v-lite-sft"]
     },
-    {
-      "id": "kandinsky6-ti2va-base",
-      "family": "kandinsky6",
-      "stage": "inference",
-      "task": "Text or image to video with audio",
-      "label": "Kandinsky 6 Pro TI2VA",
-      "summary": "Generate a five-second, 512x768 video with synchronized audio from text, or set IMAGE_PATH in the maintained example to condition on an image.",
-      "model": "kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers",
-      "source": "examples/inference/basic/basic_kandinsky6_ti2va.py",
-      "command": "python examples/inference/basic/basic_kandinsky6_ti2va.py",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {"platform": "cuda", "gpu_count": 1, "evidence": "source-configured"},
-      "evidence": "Source-backed",
-      "expected_artifact": "MP4 video with synchronized audio under video_samples_kandinsky6_ti2va/",
-      "related": ["kandinsky6-ti2va-piflow", "kandinsky6-vsr"]
-    },
-    {
-      "id": "kandinsky6-ti2va-piflow",
-      "family": "kandinsky6",
-      "stage": "inference",
-      "task": "Distilled text or image to video with audio",
-      "label": "Kandinsky 6 Pro pi-Flow",
-      "summary": "Generate a five-second video with synchronized audio using the distilled pi-Flow checkpoint in ten inference steps and guidance 1.0.",
-      "model": "kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers",
-      "source": "examples/inference/basic/basic_kandinsky6_ti2va.py",
-      "command": "KANDINSKY6_MODEL_PATH=kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers python examples/inference/basic/basic_kandinsky6_ti2va.py",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {"platform": "cuda", "gpu_count": 1, "evidence": "source-configured"},
-      "evidence": "Source-backed",
-      "expected_artifact": "MP4 video with synchronized audio under video_samples_kandinsky6_ti2va/",
-      "related": ["kandinsky6-ti2va-base", "kandinsky6-vsr-distilled"]
-    },
-    {
-      "id": "kandinsky6-vsr",
-      "family": "kandinsky6",
-      "stage": "inference",
-      "task": "Video super-resolution",
-      "label": "Kandinsky 6 VSR",
-      "summary": "Upscale an existing clip by 2.25x with the four-step flow-matching VSR checkpoint while preserving its source audio.",
-      "model": "kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers",
-      "source": "examples/inference/basic/basic_kandinsky6_sr.py",
-      "command": ": \"${INPUT_VIDEO:?Set INPUT_VIDEO to an existing video}\"\npython examples/inference/basic/basic_kandinsky6_sr.py --video-path \"$INPUT_VIDEO\" --scale 2.25",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {"platform": "cuda", "gpu_count": 1, "evidence": "source-configured"},
-      "evidence": "Source-backed",
-      "expected_artifact": "Upscaled MP4 under outputs_video/kandinsky6_sr/",
-      "related": ["kandinsky6-vsr-distilled", "kandinsky6-ti2va-base"]
-    },
-    {
-      "id": "kandinsky6-vsr-distilled",
-      "family": "kandinsky6",
-      "stage": "inference",
-      "task": "Distilled video super-resolution",
-      "label": "Kandinsky 6 VSR distilled",
-      "summary": "Upscale an existing clip by 2.25x with the distilled two-step VSR checkpoint while preserving its source audio.",
-      "model": "kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers",
-      "source": "examples/inference/basic/basic_kandinsky6_sr.py",
-      "command": ": \"${INPUT_VIDEO:?Set INPUT_VIDEO to an existing video}\"\npython examples/inference/basic/basic_kandinsky6_sr.py --model-path kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers --video-path \"$INPUT_VIDEO\" --scale 2.25",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {"platform": "cuda", "gpu_count": 1, "evidence": "source-configured"},
-      "evidence": "Source-backed",
-      "expected_artifact": "Upscaled MP4 under outputs_video/kandinsky6_sr/",
-      "related": ["kandinsky6-vsr", "kandinsky6-ti2va-piflow"]
-    },
     {
       "id": "flux2-klein-t2i",
       "family": "flux",
@@ -397,6 +360,53 @@
       "expected_artifact": "PNG images under outputs/flux_dev/samples/",
       "limitations": ["FLUX.1 is loadable by ID but registers no model_family in fastvideo/registry.py; it is grouped under FLUX for documentation only."]
     },
+    {
+      "id": "glm-image-t2i",
+      "family": "glm_image",
+      "stage": "inference",
+      "task": "Text to image",
+      "label": "GLM-Image",
+      "summary": "GLM-Image text-to-image generation from its maintained example.",
+      "model": "zai-org/GLM-Image",
+      "source": "examples/inference/basic/basic_glm_image.py",
+      "command": "python examples/inference/basic/basic_glm_image.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "PNG image at image_output/landscape.png",
+      "related": ["glm-image-edit"]
+    },
+    {
+      "id": "glm-image-edit",
+      "family": "glm_image",
+      "stage": "inference",
+      "task": "Image editing",
+      "label": "GLM-Image editing",
+      "summary": "Edit an input image with an instruction prompt using GLM-Image, from its maintained editing example.",
+      "model": "zai-org/GLM-Image",
+      "source": "examples/inference/basic/edit_glm_image.py",
+      "command": "python examples/inference/basic/edit_glm_image.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "PNG image at image_output/edited.png (input: assets/images/couple.jpg)",
+      "related": ["glm-image-t2i"]
+    },
+    {
+      "id": "zimage-turbo-t2i",
+      "family": "zimage",
+      "stage": "inference",
+      "task": "Text to image",
+      "label": "Z-Image Turbo",
+      "summary": "Z-Image Turbo text-to-image on a single GPU from its maintained example.",
+      "model": "Tongyi-MAI/Z-Image-Turbo",
+      "source": "examples/inference/basic/basic_zimage.py",
+      "command": "python examples/inference/basic/basic_zimage.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "PNG image at outputs/zimage/zimage_turbo.png"
+    },
     {
       "id": "sd35-medium-t2i",
       "family": "sd35",
@@ -446,8 +456,7 @@
       "source": "examples/inference/basic/basic_fasth3.py",
       "serving": {
         "source": "examples/serving/openai_fasth3.yaml",
-        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\"",
-        "audio": true
+        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\""
       },
       "command": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\"\npython examples/inference/basic/basic_fasth3.py --prompt \"(S1) A presenter says <d>[English] FastVideo runs FastH3.</d>\" --profile all",
       "gpu_types": ["NVIDIA"],
@@ -472,8 +481,7 @@
       "serving": {
         "source": "examples/serving/mlx_fasth3.yaml",
         "install": "uv pip install -e \".[mlx]\"",
-        "prepare": "hf download FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2 --local-dir ./FastH3-Preview-v0.2\npython scripts/checkpoint_conversion/convert_minimax_h3_mlx.py --model-root ./FastH3-Preview-v0.2/transformer --out ./FastH3-MLX --formats \"int6\"",
-        "audio": true
+        "prepare": "hf download FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2 --local-dir ./FastH3-Preview-v0.2\npython scripts/checkpoint_conversion/convert_minimax_h3_mlx.py --model-root ./FastH3-Preview-v0.2/transformer --out ./FastH3-MLX --formats \"int6\""
       },
       "group": "fasth3-preview",
       "group_label": "FastH3 V1",
@@ -520,8 +528,7 @@
       "serving": {
         "source": "examples/serving/openai_fasth3_spark.yaml",
         "install": "UV_TORCH_BACKEND=cu130 uv pip install -e .",
-        "env": "FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3",
-        "audio": true
+        "env": "FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3"
       },
       "command": "FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 fastvideo generate --config examples/inference/basic/basic_fasth3_spark.yaml",
       "gpu_types": ["NVIDIA"],
@@ -571,77 +578,6 @@
         "Height, width, frames, and steps in the YAML are examples. Edit them or pass CLI flags. See docs/getting_started/installation/spark_pair.md."
       ]
     },
-    {
-      "id": "compacth3-rtx5090",
-      "group": "compacth3-rtx5090",
-      "group_label": "CompactH3 on RTX 5090",
-      "group_task": "4-step 42-block text to video + audio",
-      "family": "minimax_h3",
-      "stage": "inference",
-      "task": "Few-step text to video (with audio)",
-      "label": "CompactH3 NVFP4 on RTX 5090",
-      "summary": "Run the 42-block CompactH3 NVFP4 DiT with the NVFP4 Qwen3-VL encoder, Comfy int8-convrot VAE, and SageAttention3 FP4 on one 32 GB RTX 5090. Sequential load parks the encoder in pinned host RAM.",
-      "model": "./CompactH3",
-      "source": "examples/inference/basic/basic_compacth3_rtx5090.yaml",
-      "serving": {
-        "source": "examples/serving/openai_compacth3_rtx5090.yaml",
-        "install": "UV_TORCH_BACKEND=cu128 uv pip install -e \".[fasth3]\"",
-        "prepare": "hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree --local-dir ./CompactH3 --include model_index.json --include \"tokenizer/**\" --include \"processor/**\" --include \"scheduler/**\" --include \"audio_scheduler/**\" --include \"audio_vae/**\" --include \"vae/**\"\nhf download aryan5v/FastH3-20B-42block-dmd2-ckpt1400-bf16 --local-dir ./CompactH3/transformer\nhf download aryan5v/FastH3-20B-42block-dmd2-ckpt1400-nvfp4 --local-dir ./CompactH3/transformer --include nvfp4_weights.safetensors\nhf download KyleNeverGivesUp/FastH3-text-encoder-nvfp4 --local-dir ./CompactH3/text_encoder\nhf download Comfy-Org/MiniMax-H3 --local-dir ./Comfy-MiniMax-H3 --include vae/minimax_h3_video_vae_int8_convrot.safetensors\ncp ./Comfy-MiniMax-H3/vae/minimax_h3_video_vae_int8_convrot.safetensors ./CompactH3/vae/",
-        "env": "FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FLASHINFER_CUDA_ARCH_LIST=12.0a"
-      },
-      "command": "FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FLASHINFER_CUDA_ARCH_LIST=12.0a FASTVIDEO_STAGE_LOGGING=1 fastvideo generate --config examples/inference/basic/basic_compacth3_rtx5090.yaml",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {
-        "platform": "cuda",
-        "gpu_count": 1,
-        "evidence": "source-configured"
-      },
-      "evidence": "Source-backed",
-      "expected_artifact": "MP4 under outputs/compacth3_rtx5090/",
-      "modes": ["T2VA", "CompactH3 NVFP4", "RTX 5090"],
-      "limitations": [
-        "Assemble ./CompactH3 before running. The DiT NVFP4 export and encoder snapshot are gated; run huggingface-cli login and accept each repo license.",
-        "32 GB cannot keep the NVFP4 encoder and DiT on the GPU together. Keep h3_sequential_load on and lazy_module_load off.",
-        "Blackwell sm_120 needs ATTN_QAT_INFER, FLASHINFER_CUDA_ARCH_LIST=12.0a, and a CUDA 12.8 PyTorch wheel.",
-        "Legal num_frames values are 17n+5, capped at 362 (15.08 s). Native 16:9 sizes include 832x480 and 1344x768; 1344x768 on 32 GB is unmeasured."
-      ]
-    },
-    {
-      "id": "compacth3-rtx-pro6000",
-      "group": "compacth3-rtx-pro6000",
-      "group_label": "CompactH3 on RTX PRO 6000",
-      "group_task": "4-step 42-block text to video + audio",
-      "family": "minimax_h3",
-      "stage": "inference",
-      "task": "Few-step text to video (with audio)",
-      "label": "CompactH3 NVFP4 on RTX PRO 6000 Blackwell",
-      "summary": "Run CompactH3 NVFP4 with the encoder, DiT, and int8-convrot VAE resident on one 96 GB RTX PRO 6000 Blackwell. The checked-in example is 1344x768 and 124 frames (5.17 s) with VAE torch.compile. Use 832x480 for clip-queue playground traffic.",
-      "model": "./CompactH3",
-      "source": "examples/inference/basic/basic_compacth3_rtx_pro6000.yaml",
-      "serving": {
-        "source": "examples/serving/openai_compacth3_rtx_pro6000.yaml",
-        "install": "UV_TORCH_BACKEND=cu128 uv pip install -e \".[fasth3]\"",
-        "prepare": "hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree --local-dir ./CompactH3 --include model_index.json --include \"tokenizer/**\" --include \"processor/**\" --include \"scheduler/**\" --include \"audio_scheduler/**\" --include \"audio_vae/**\" --include \"vae/**\"\nhf download aryan5v/FastH3-20B-42block-dmd2-ckpt1400-bf16 --local-dir ./CompactH3/transformer\nhf download aryan5v/FastH3-20B-42block-dmd2-ckpt1400-nvfp4 --local-dir ./CompactH3/transformer --include nvfp4_weights.safetensors\nhf download KyleNeverGivesUp/FastH3-text-encoder-nvfp4 --local-dir ./CompactH3/text_encoder\nhf download Comfy-Org/MiniMax-H3 --local-dir ./Comfy-MiniMax-H3 --include vae/minimax_h3_video_vae_int8_convrot.safetensors\ncp ./Comfy-MiniMax-H3/vae/minimax_h3_video_vae_int8_convrot.safetensors ./CompactH3/vae/",
-        "env": "FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FLASHINFER_CUDA_ARCH_LIST=12.0a"
-      },
-      "command": "FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FLASHINFER_CUDA_ARCH_LIST=12.0a FASTVIDEO_STAGE_LOGGING=1 fastvideo generate --config examples/inference/basic/basic_compacth3_rtx_pro6000.yaml",
-      "gpu_types": ["NVIDIA"],
-      "hardware": {
-        "platform": "cuda",
-        "gpu_count": 1,
-        "evidence": "source-configured"
-      },
-      "evidence": "Source-backed",
-      "expected_artifact": "MP4 under outputs/compacth3_rtx_pro6000/",
-      "modes": ["T2VA", "CompactH3 NVFP4", "RTX PRO 6000"],
-      "limitations": [
-        "Assemble ./CompactH3 before running. The DiT NVFP4 export and encoder snapshot are gated; run huggingface-cli login and accept each repo license.",
-        "96 GB keeps the encoder, DiT, and VAE on GPU. Do not enable h3_sequential_load or lazy_module_load on this box.",
-        "Enable compile.vae_enabled. Leave inference_torch_compile off: FlashInfer and Sage3 custom ops cannot be compiled.",
-        "Blackwell sm_120 needs ATTN_QAT_INFER, FLASHINFER_CUDA_ARCH_LIST=12.0a, and a CUDA 12.8 PyTorch wheel.",
-        "Legal num_frames values are 17n+5, capped at 362 (15.08 s). Native 16:9 sizes include 832x480 and 1344x768. Dense CompactH3 has zero VSA gates; the pipeline raises if VIDEO_SPARSE_ATTN_H3 is loaded with all-zero to_gate_compress weights."
-      ]
-    },
     {
       "id": "fasth3-8step-v2-cuda",
       "group": "fasth3-8step-v2",
@@ -656,8 +592,7 @@
       "source": "examples/inference/basic/basic_fasth3_8step.py",
       "serving": {
         "source": "examples/serving/openai_fasth3_8step.yaml",
-        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\"",
-        "audio": true
+        "install": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\""
       },
       "command": "UV_TORCH_BACKEND=cu130 uv pip install -e \".[fasth3]\"\npython examples/inference/basic/basic_fasth3_8step.py --prompt \"(S1) A presenter says <d>[English] FastVideo runs FastH3.</d>\" --profile strict --no-inference-torch-compile --no-compile-vae",
       "gpu_types": ["NVIDIA"],
@@ -686,8 +621,7 @@
       "serving": {
         "source": "examples/serving/mlx_fasth3_8step.yaml",
         "install": "uv pip install -e \".[mlx]\"",
-        "prepare": "hf download FastVideo/FastVideo-FastH3-8-Step-V2 --local-dir ./FastH3-8-Step-V2\npython scripts/checkpoint_conversion/convert_minimax_h3_mlx.py --model-root ./FastH3-8-Step-V2/transformer --out ./FastH3-8-Step-V2-MLX --formats \"int8\" --include-vsa",
-        "audio": true
+        "prepare": "hf download FastVideo/FastVideo-FastH3-8-Step-V2 --local-dir ./FastH3-8-Step-V2\npython scripts/checkpoint_conversion/convert_minimax_h3_mlx.py --model-root ./FastH3-8-Step-V2/transformer --out ./FastH3-8-Step-V2-MLX --formats \"int8\" --include-vsa"
       },
       "group": "fasth3-8step-v2",
       "group_label": "FastH3 V2",
@@ -788,6 +722,38 @@
       ],
       "limitations": ["Supply a compatible FastH3 adapter. The script infers dense or VSA attention from the adapter payload unless you override it."]
     },
+    {
+      "id": "longcat-t2v",
+      "family": "longcat",
+      "stage": "inference",
+      "task": "Text to video",
+      "label": "LongCat Video T2V",
+      "summary": "LongCat Video text-to-video at 480p (50 steps), with distilled and 720p refinement passes included in the same maintained script.",
+      "model": "FastVideo/LongCat-Video-T2V-Diffusers",
+      "source": "examples/inference/basic/basic_longcat_t2v.py",
+      "command": "python examples/inference/basic/basic_longcat_t2v.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "MP4 videos under outputs_video/longcat_t2v_basic/, longcat_t2v_distill/, and longcat_t2v_refine_720p/",
+      "related": ["longcat-i2v"]
+    },
+    {
+      "id": "longcat-i2v",
+      "family": "longcat",
+      "stage": "inference",
+      "task": "Image to video",
+      "label": "LongCat Video I2V",
+      "summary": "LongCat Video image-to-video with optional distilled and refinement passes, from its maintained example.",
+      "model": "FastVideo/LongCat-Video-I2V-Diffusers",
+      "source": "examples/inference/basic/basic_longcat_i2v.py",
+      "command": "python examples/inference/basic/basic_longcat_i2v.py",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "MP4 videos under outputs_video/longcat_i2v_basic/ and longcat_i2v_distill/",
+      "related": ["longcat-t2v"]
+    },
     {
       "id": "stable-audio-open-t2a",
       "family": "stable_audio",

docs/assets/cookbook.js

@@ -506,10 +506,6 @@
       const runtime = runtimeFor(recipe);
       const profile = servingPanel && servingProfiles[recipe.id];
       const useServer = Boolean(profile && usagePreference === "server");
-      // Only some servers expose the browser playground or return audio; both come from the serving profile.
-      const hasPlayground = Boolean(profile && profile.playground_url);
-      const hasAudio = Boolean(profile && profile.audio);
-      const playgroundAny = familyRecipes.some((item) => servingProfiles[item.id] && servingProfiles[item.id].playground_url);
       // The measured local profile and the server config have separate evidence.
       const activeRecipe = useServer ? { ...recipe, hardware: profile.hardware, evidence: "Source-backed" } : recipe;
       const knobs = knobsFor(recipe);
@@ -519,19 +515,14 @@
         usage.querySelectorAll("[data-cookbook-mode]").forEach((option) => {
           const selected = option.dataset.cookbookMode === (useServer ? "server" : "python");
           option.disabled = option.dataset.cookbookMode === "server" && !profile;
-          const hint = option.querySelector("[data-cookbook-server-hint]");
-          if (hint) hint.textContent = (profile ? hasPlayground : playgroundAny) ? "Playground, cURL, or an API client" : "cURL or an API client";
           option.classList.toggle("cookbook-option--selected", selected);
           option.setAttribute("aria-pressed", String(selected));
         });
-        const servedNames = [...new Set(familyRecipes.filter((item) => servingProfiles[item.id]).map((item) => item.group_label || item.label))];
         servingAvailability.textContent = profile
-          ? `${hasPlayground ? "The playground" : "cURL"} and the OpenAI Python client share one server process. Both workflows can run on your own machine.`
+          ? "The playground and the OpenAI Python client share one server process. Both workflows can run on your own machine."
           : servingLoadFailed
-            ? "Server examples could not be loaded. Use Python directly."
-            : servedNames.length
-              ? `This recipe uses Python directly. ${new Intl.ListFormat("en").format(servedNames)} can also run a local server for ${playgroundAny ? "the playground and " : "cURL and "}the OpenAI Python client.`
-              : "This recipe uses Python directly.";
+            ? "Server examples could not be loaded. Open the H3 server guide below, or use Python directly."
+            : "This recipe uses Python directly. FastH3 V1 and FastH3 V2 can also run a local server for the playground and the OpenAI Python client.";
         servingPanel.hidden = !useServer;
         commandBlock.hidden = useServer;
         root.querySelector("[data-cookbook-python-note]").hidden = useServer;
@@ -540,12 +531,11 @@
       if (useServer) {
         const isMLX = profile.runtime === "mlx";
         const isSpark = runtime.id === "spark";
-        const where = hasPlayground ? "in the playground or your app" : "in your app";
         servingPanel.querySelector("[data-cookbook-server-lifetime]").textContent = isMLX
-          ? `Start once, then change prompts ${where}. MLX reuses its pipeline and prompt cache, but loads and releases model components between phases to limit unified-memory use. It does not keep all weights resident.`
+          ? "Start once, then change prompts in the playground or your app. MLX reuses its pipeline and prompt cache, but loads and releases model components between phases to limit unified-memory use. It does not keep all weights resident."
           : isSpark
-            ? `Start once, then change prompts ${where}. On a DGX Spark, lazy module load still reloads Qwen3-VL and the DiT between phases of each request, so later prompts are not a free hot cache.`
-            : `Start once, then change prompts ${where}. CUDA requests reuse the loaded model. The Python SDK can also reuse a generator within one process.`;
+            ? "Start once, then change prompts in the playground or your app. On a DGX Spark, lazy module load still reloads Qwen3-VL and the DiT between phases of each request, so later prompts are not a free hot cache."
+            : "Start once, then change prompts in the playground or your app. CUDA requests reuse the loaded model. The Python SDK can also reuse a generator within one process.";
         servingPanel.querySelector("[data-cookbook-install-guide]").href = isMLX
           ? "../../getting_started/installation/mlx/"
           : isSpark
@@ -556,10 +546,7 @@
         servingPanel.querySelector("[data-cookbook-server-install]").textContent = profile.install;
         servingPanel.querySelector("[data-cookbook-server-command]").textContent = profile.command;
         servingPanel.querySelector("[data-cookbook-health-command]").textContent = profile.health_command;
-        servingPanel.querySelectorAll("[data-cookbook-playground-only]").forEach((element) => {
-          element.hidden = !hasPlayground;
-        });
-        if (hasPlayground) servingPanel.querySelector("[data-cookbook-playground]").href = profile.playground_url;
+        servingPanel.querySelector("[data-cookbook-playground]").href = profile.playground_url;
         const client = profile.clients[selectedClient];
         const filename = client.source.split("/").pop();
         servingPanel.querySelector("[data-cookbook-client-install]").textContent = client.install;
@@ -587,17 +574,13 @@
       });
 
       description.textContent = useServer
-        ? `${recipe.group_label || recipe.label} generates video${hasAudio ? " with audio" : ""}. Start the local server, then use ${hasPlayground ? "the playground or " : "cURL or "}the OpenAI Python client. This profile uses the checked-in ${runtime.label} configuration.`
+        ? `${recipe.group_label || recipe.label} generates video with audio. Start the local server, then use the playground or the OpenAI Python client. This profile uses the checked-in ${runtime.label} configuration.`
         : recipe.summary;
       label.textContent = useServer ? `${recipe.group_label || recipe.label} · Server` : recipe.label;
       model.textContent = recipe.model;
-      task.textContent = (useServer && recipe.serving && recipe.serving.task) || recipe.task;
+      task.textContent = recipe.task;
       hardwareValue.textContent = runtimeSummary(activeRecipe);
-      if (artifact) {
-        artifact.textContent = useServer
-          ? (hasAudio ? "MP4 with audio" : "MP4 video")
-          : recipe.expected_artifact || "Not yet documented for this recipe.";
-      }
+      if (artifact) artifact.textContent = useServer ? "MP4 with audio" : recipe.expected_artifact || "Not yet documented for this recipe.";
       if (evidenceCell) {
         evidenceCell.textContent = activeRecipe.evidence || "Source-backed";
         evidenceCell.classList.toggle("cookbook-badge--verified", activeRecipe.evidence === "Verified");
@@ -619,12 +602,10 @@
           ? "This MLX server config has no recorded hardware run. Measurements from the Python recipe are not server memory requirements. Only text-to-video/audio is wired; reference inputs and fast modes are not exposed here."
           : runtime.id === "spark"
             ? "This Spark server config has no recorded serving benchmark. Lazy module load reloads Qwen3-VL and the DiT between phases of each request. Compilation of the DiT is disabled."
-            : `This server config has no recorded serving benchmark.${profile.compile_enabled === false ? " Compilation is disabled, unlike the measured Python performance profile." : ""}`,
+            : "This server config has no recorded serving benchmark. Compilation is disabled, unlike the measured Python performance profile.",
         `${profile.sampling.width} × ${profile.sampling.height} · ${profile.sampling.num_frames} frames · ${profile.sampling.fps} fps. The server supplies these defaults; the client sends the model and prompt.`,
         "Generation is serialized. Job metadata is held in memory and is lost when the server restarts.",
-      ] : recipe.limitations || []),
-      ...(useServer && recipe.serving && Array.isArray(recipe.serving.limitations) ? recipe.serving.limitations : []),
-      ...knobCaveats];
+      ] : recipe.limitations || []), ...knobCaveats];
       notes.replaceChildren();
       notes.hidden = limitations.length === 0;
       if (limitations.length) {

docs/assets/custom.css

@@ -1515,8 +1515,7 @@ img {
 
 .cookbook-serving[hidden],
 .cookbook-command[hidden],
-[data-cookbook-python-note][hidden],
-[data-cookbook-playground-only][hidden] {
+[data-cookbook-python-note][hidden] {
     display: none;
 }
 

docs/assets/logos/SOURCES.md

@@ -9,15 +9,18 @@ plain typographic tile instead — that tile is a UI placeholder, not a logo.
 | --- | --- | --- |
 | `wan-ai.webp` | [Official Wan-AI Hugging Face organization avatar](https://huggingface.co/Wan-AI) | Wan family card and page header |
 | `ltx.webp` | [Official Lightricks Hugging Face organization avatar](https://huggingface.co/Lightricks) | LTX family card and page header |
-| `tencent-hunyuan.webp` | [Official Tencent Hunyuan Hugging Face organization avatar](https://huggingface.co/Tencent-Hunyuan) | Hunyuan family card and page header |
+| `tencent-hunyuan.webp` | [Official Tencent Hunyuan Hugging Face organization avatar](https://huggingface.co/Tencent-Hunyuan) | Hunyuan and GameCraft cards, Hunyuan page header |
 | `nvidia.webp` | [Official NVIDIA Hugging Face organization avatar](https://huggingface.co/nvidia) | Cosmos and GEN3C cards, Cosmos page header |
-| `kandinsky.webp` | [Official Kandinsky Lab Hugging Face organization avatar](https://huggingface.co/kandinskylab) | Kandinsky 5 and Kandinsky 6 cards and page headers |
+| `kandinsky.webp` | [Official Kandinsky Lab Hugging Face organization avatar](https://huggingface.co/kandinskylab) | Kandinsky 5 family card and page header |
 | `black-forest-labs.webp` | [Official Black Forest Labs Hugging Face organization avatar](https://huggingface.co/black-forest-labs) | FLUX family card and page header |
 | `minimax.webp` | [Official MiniMax Hugging Face organization avatar](https://huggingface.co/MiniMaxAI) | MiniMax H3 family card and page header |
+| `tongyi.webp` | [Official Tongyi MAI Hugging Face organization avatar](https://huggingface.co/Tongyi-MAI) | Z-Image family card and page header |
+| `zai.webp` | [Official Z.ai Hugging Face organization avatar](https://huggingface.co/zai-org) | GLM-Image family card and page header |
 | `stabilityai.webp` | [Official Stability AI Hugging Face organization avatar](https://huggingface.co/stabilityai) | Stable Diffusion and Stable Audio cards and page headers |
-| `fastvideo.webp` | [Official FastVideo Hugging Face organization avatar](https://huggingface.co/FastVideo) | Matrix Game and MMAudio cards (converted weights published by this org) |
+| `meituan-longcat.webp` | [Official Meituan LongCat Hugging Face organization avatar](https://huggingface.co/meituan-longcat) | LongCat family card and page header |
+| `fastvideo.webp` | [Official FastVideo Hugging Face organization avatar](https://huggingface.co/FastVideo) | Matrix Game and MMAudio cards (converted weights published by this org), DreamX card |
 
-Typographic tiles (no vendored image): HY-World ("HY"), LingBot ("LB"),
-MMAudio page header ("MMA"). These publishers have no
+Typographic tiles (no vendored image): TurboDiffusion ("Turbo"), HY-World
+("HY"), LingBot ("LB"), MMAudio page header ("MMA"). These publishers have no
 single official mark appropriate for reuse in the catalog; add a licensed
 asset here if one becomes available.

docs/contributing/ci_architecture.md

@@ -4,6 +4,11 @@ This is the canonical reference for FastVideo's CI/CD system. Contributor-facing
 PR steps live in [Pull Requests](pull_requests.md), and test-authoring guidance
 lives in [Testing](testing.md).
 
+The existing Slurm route below remains the default. Operators can also install
+the [selectable GPU dispatcher](gpu_ci_backends.md) to run the same lane scripts
+on Modal or Kubernetes in the `vllm` namespace. That opt-in path enforces two
+active PRs, four GPUs per PR, and eight total, with separate backend statuses.
+
 ## Overview
 
 FastVideo splits validation across GitHub Actions, Buildkite, Slinky Slurm,
@@ -435,8 +440,11 @@ before later jobs consume the updated image pin.
 The same workflow publishes a single-architecture ARM64, CUDA 13, SM100 image
 for the self-hosted CI runner under the
 `py3.12-cuda13.0.0-sm100-{latest,sha-*}` tags. It carries the matching prebuilt
-kernel wheel so runner jobs can validate and install the exact source and ABI
-match instead of recompiling it in every lane.
+kernel wheel compiled with `TORCH_CUDA_ARCH_LIST=10.0a` to include the GB200
+VSA CUDA extensions. Runtime kernel detection uses the same target, so runner
+jobs can validate and install the exact source and ABI match instead of
+recompiling it in every lane. Older artifacts built for `10.0` have a different
+cache key and trigger a local rebuild when the worker detects `10.0a`.
 
 The optional Dreamverse matrix builds backend and UI images for CUDA 12.6 and
 CUDA 13 on `amd64`. Dreamverse remains `amd64`-only because its FA4 dependency

docs/contributing/env_vars.md

@@ -13,15 +13,12 @@ test in the same pull request.
    them directly with `os.environ.get("NAME")`, and the name must be in the external-variable allowlist
    (`EXTERNAL_ALLOWLIST` in the contract test). When FastVideo sets such a variable for the other tool, it calls
    `envs.set_external`, `envs.setdefault_external`, or `envs.unset_external`, and the name must be in
-   `EXTERNAL_WRITE_ALLOWLIST`. Variables that FastVideo's CI and CI tooling define (for example `TEST_SCOPE` and
-   `PERF_RUN_SOURCE`) keep their names, and test code under `fastvideo/tests/` reads them directly; they are listed
-   in `CI_ONLY_VARIABLES` in the contract test, together with the file that sets each one.
+   `EXTERNAL_WRITE_ALLOWLIST`.
 2. **Read with `envs.NAME.get()`, write with `envs.NAME.set()`, and change a value in tests with
    `envs.NAME.override()`.** Each type has one parsing rule. A value that the rule rejects raises
    `fastvideo.envs.EnvVarError` instead of falling back to the default.
 3. **Name FastVideo variables with the `FASTVIDEO_` prefix.** The second word states the purpose where one applies:
-   `ENABLE_`, `DISABLE_`, `USE_`, `FORCE_`, `DEBUG_`, `TEST_`. Variables that only tests read use
-   `FASTVIDEO_TEST_`, for example `FASTVIDEO_TEST_SD35_MODEL_DIR`, and the category `test`.
+   `ENABLE_`, `DISABLE_`, `USE_`, `FORCE_`, `DEBUG_`, `TEST_`.
 4. **Keep a renamed variable as a deprecated alias until the next minor release.** Setting the old name logs a
    warning. Delete a variable that no code reads, and list it in `DEPRECATED_VARIABLES` so that setting it logs a
    warning.
@@ -32,8 +29,7 @@ test in the same pull request.
    effect without re-importing a module. Module level, class bodies, decorators, and default argument values run at
    import time.
 7. **Do not write the environment to pass values between parts of FastVideo.** Pass an argument instead. Tests use
-   `envs.NAME.override()`, and `envs.override_external()` for variables outside the registry; both restore the
-   previous value.
+   `envs.NAME.override()`.
 
 ## Field types
 
@@ -91,21 +87,6 @@ with envs.FASTVIDEO_DEBUG_MY_STAGE.override(True):
     run_stage()
 ```
 
-For a variable outside the registry, such as `MASTER_PORT` or `TEST_SCOPE`, use `envs.override_external`. To keep
-an override until the end of a test, enter it through the `env_overrides` fixture from `fastvideo/tests/conftest.py`,
-which restores every value at teardown:
-
-```python
-def test_my_stage(env_overrides):
-    env_overrides.enter_context(envs.FASTVIDEO_DEBUG_MY_STAGE.override(True))
-    env_overrides.enter_context(envs.override_external("MASTER_PORT", "29512"))
-    run_stage()
-```
-
-In test code under `fastvideo/tests/`, `override_external` accepts any name that code may read directly
-(`EXTERNAL_ALLOWLIST`, `CI_ONLY_VARIABLES`) or that is in `EXTERNAL_WRITE_ALLOWLIST`. Library code may write only
-the names in `EXTERNAL_WRITE_ALLOWLIST`.
-
 ## Rename or remove a variable
 
 To rename a variable, declare it under the new name and list the old name in `deprecated_names`:
@@ -140,11 +121,9 @@ It reports each violation as `<path>: <kind> <name>`:
 |                    | with a name built at runtime (`<dynamic>`)                   | another tool owns, add it to                 |
 |                    |                                                              | `EXTERNAL_ALLOWLIST` with a reason.          |
 | `write`            | `os.environ[...] = ...`, `setdefault`, `pop`, `del`,         | Pass an argument instead. In tests, use      |
-|                    | `os.putenv`, `os.unsetenv`, `monkeypatch.setenv`/`delenv`,   | `envs.NAME.override()`, or                   |
-|                    | or an `envs.*_external` call with a name that the            | `envs.override_external()` for a variable    |
-|                    | helper does not accept                                       | outside the registry. For a variable that    |
-|                    |                                                              | another tool reads, call an                  |
-|                    |                                                              | `envs.*_external` helper and add the name to |
+|                    | `os.putenv`, `os.unsetenv`, `monkeypatch.setenv`/`delenv`,   | `envs.NAME.override()`. For a variable that  |
+|                    | or an `envs.*_external` call with a name outside             | another tool reads, call an                  |
+|                    | `EXTERNAL_WRITE_ALLOWLIST`                                   | `envs.*_external` helper and add the name to |
 |                    |                                                              | `EXTERNAL_WRITE_ALLOWLIST` with a reason.    |
 | `whole-environ`    | `os.environ.copy()`, `dict(os.environ)`, iteration,          | Read the specific variables that the code    |
 |                    | `mock.patch.dict(os.environ, ...)`, `os.environ.update`      | needs.                                       |
@@ -192,7 +171,6 @@ longer exists also fails the test, so the fixing pull request deletes its entry.
 | `FASTVIDEO_FA4`                                            | bool                           | `0`                                                   | attention   | The FLASH_ATTN backend uses FlashAttention-4 (flash_attn.cute) instead of FA3 or FA2.                                                                                                                                                                    |
 | `FASTVIDEO_MINIMAX_H3_FA4_PACKED_VARLEN`                   | bool                           | `0`                                                   | attention   | MiniMax-H3 dense DiT self-attention uses the FlashAttention-4 packed-varlen entry point. This changes the floating-point reduction order, so it is an inference-only opt-in.                                                                             |
 | `FASTVIDEO_VSA_SM100A`                                     | bool                           | `0`                                                   | attention   | VIDEO_SPARSE_ATTN_H3 sends no-grad tile-64 forwards to the data-center Blackwell (sm_100a) kernel. fastvideo-kernel reads the same variable with the same rule.                                                                                          |
-| `FASTVIDEO_VSA_TRITON`                                     | bool                           | `0`                                                   | attention   | Force the Triton MiniMax-H3 sparse attention kernel. fastvideo-kernel reads the same variable.                                                                                                                                                           |
 | `FASTVIDEO_NVFP4_FA4`                                      | bool                           | `0`                                                   | attention   | FlashAttention-4 quantizes Q and K to NVFP4. An explicit nvfp4_fa4 attention implementation argument takes precedence.                                                                                                                                   |
 | `FASTVIDEO_DISABLE_ATTENTION_COMPILE`                      | bool                           | `1`                                                   | attention   | Keep attention forward out of torch.compile graphs (torch.compiler.disable). Set it to 0 to let attention constructed under that setting be traced. Setting it explicitly to true also blocks regional compile.                                          |
 | `FASTVIDEO_MLX_WINDOW`                                     | int                            | `0`                                                   | attention   | MLX FastWan windowed attention size in tokens. 0 uses full attention.                                                                                                                                                                                    |
@@ -241,35 +219,6 @@ longer exists also fails the test, so the fixing pull request deletes its entry.
 | `FASTVIDEO_VBENCH_FULL_INFO_JSON`                          | str                            | unset                                                 | eval        | Path to VBench_full_info.json, used instead of the vendored copy. Deprecated names: `VBENCH_FULL_INFO_JSON`.                                                                                                                                             |
 | `FASTVIDEO_FVD_REF_FEATURES`                               | str                            | unset                                                 | eval        | Cached reference-feature file for the FVD metric.                                                                                                                                                                                                        |
 | `FASTVIDEO_FAD_REF_FEATURES`                               | str                            | unset                                                 | eval        | Cached reference-feature file for the audio Frechet distance metric.                                                                                                                                                                                     |
-| `FASTVIDEO_H3_VSA_FP4`                                     | bool                           | `0`                                                   | attention   | Run MiniMax-H3 VSA attention on the block-sparse SageAttention3 FP4 kernel (sm_120, no-grad, single sequence-parallel rank).                                                                                                                             |
-| `FASTVIDEO_H3_VSA_TILE_FIRST`                              | bool                           | `0`                                                   | attention   | Single-rank MiniMax-H3 VSA with one tile gather of the block input instead of separate Q/K/V/gate scatters.                                                                                                                                              |
-| `FASTVIDEO_H3_VSA_SM89_KERNEL`                             | one of original, bf16, int8    | `original`                                            | attention   | Fine-attention kernel for MiniMax-H3 VSA on sm_89: original, bf16, or int8 (INT8 QK, BF16 PV).                                                                                                                                                           |
-| `FASTVIDEO_H3_SIM_SP_FP8`                                  | bool                           | `0`                                                   | debug       | Simulate the FP8 sequence-parallel exchange of the MiniMax-H3 FP4 VSA path on one rank.                                                                                                                                                                  |
-| `FASTVIDEO_H3_FFN_CHUNK_TOKENS`                            | int                            | `0`                                                   | performance | Inference-only MiniMax-H3 FFN token chunk size; 0 runs the FFN unchunked.                                                                                                                                                                                |
-| `FASTVIDEO_H3_FP8_ATTENTION`                               | bool                           | `0`                                                   | performance | With NVFP4 layer_profile h3_dit_ffn, run MiniMax-H3 attention projections in FP8.                                                                                                                                                                        |
-| `FASTVIDEO_H3_FP8_GRANULARITY`                             | one of tensor, channel         | `tensor`                                              | performance | FP8 scaling granularity for FASTVIDEO_H3_FP8_ATTENTION.                                                                                                                                                                                                  |
-| `FASTVIDEO_NVFP4_MM_BACKEND`                               | str                            | `auto`                                                | performance | FlashInfer mm_fp4 backend for NVFP4 linears, e.g. auto or cutlass.                                                                                                                                                                                       |
-| `FASTVIDEO_NVFP4_ACT_AMAX`                                 | path                           | unset                                                 | performance | JSON of calibrated NVFP4 input amax per linear, keyed b&lt;block&gt;.&lt;sub&gt; or full prefix; sets a static activation scale.                                                                                                                         |
-| `FASTVIDEO_NVFP4_DYNAMIC_ACT`                              | str                            | `""`                                                  | performance | NVFP4 linears that derive the activation scale per call: all, or comma-separated layer-name suffixes such as ff.fc_out.                                                                                                                                  |
-| `FASTVIDEO_H3_ADALN_CACHE`                                 | bool                           | `0`                                                   | performance | Cache MiniMax-H3 AdaLN modulation per timestep instead of keeping the projection weights resident.                                                                                                                                                       |
-| `FASTVIDEO_H3_ADALN_TABLE`                                 | path                           | unset                                                 | performance | Precomputed MiniMax-H3 AdaLN modulation table; enables the cache and skips loading the AdaLN projection weights.                                                                                                                                         |
-| `FASTVIDEO_H3_ADALN_DUMP`                                  | path                           | unset                                                 | debug       | Write the MiniMax-H3 AdaLN modulation table to this path while sampling.                                                                                                                                                                                 |
-| `FASTVIDEO_H3_SPLICE_TRANSFORMER`                          | path                           | unset                                                 | eval        | Second MiniMax-H3 transformer that runs the late DMD steps (checkpoint step-splice evaluation).                                                                                                                                                          |
-| `FASTVIDEO_H3_SPLICE_FROM_STEP`                            | int                            | `4`                                                   | eval        | First denoising step run by FASTVIDEO_H3_SPLICE_TRANSFORMER.                                                                                                                                                                                             |
-| `FASTVIDEO_H3_ENCODER_LAYERWISE`                           | bool                           | `0`                                                   | performance | Stream MiniMax-H3 text-encoder language layers through exact-size pinned host memory (text-only prompts).                                                                                                                                                |
-| `FASTVIDEO_H3_ENCODER_FUSED_DEQUANT`                       | bool                           | `0`                                                   | performance | Expand the serialized NVFP4 MiniMax-H3 text encoder with one fused Triton pass on GPUs without FP4 GEMM.                                                                                                                                                 |
-| `FASTVIDEO_H3_VAE_TILE_BATCH`                              | int                            | `1`                                                   | performance | Spatial tiles per MiniMax-H3 video VAE decoder call; 1 decodes per tile.                                                                                                                                                                                 |
-| `FASTVIDEO_H3_VAE_INT8_SHARED_QKV`                         | bool                           | `0`                                                   | performance | Share the INT8 activation rotation and quantization across the MiniMax-H3 VAE Q/K/V projections.                                                                                                                                                         |
-| `FASTVIDEO_H3_VAE_INT8_TRANSPOSE_VIEW`                     | bool                           | `0`                                                   | performance | Use transposed weight views in the MiniMax-H3 VAE INT8 projections.                                                                                                                                                                                      |
-| `FASTVIDEO_H3_VAE_INT8_FUSED_DEQUANT`                      | bool                           | `0`                                                   | performance | Fused dequantization epilogue for the MiniMax-H3 VAE INT8 projections.                                                                                                                                                                                   |
-| `FASTVIDEO_H3_PINNED_SWAP`                                 | bool                           | `1`                                                   | performance | Swap offloaded MiniMax-H3 modules through exact-size pinned host arenas.                                                                                                                                                                                 |
-| `FASTVIDEO_H3_PARK_MODULES`                                | str                            | unset                                                 | performance | Comma-separated MiniMax-H3 denoise modules parked on the host while the text encoder runs, e.g. vae,audio_vae.                                                                                                                                           |
-| `FASTVIDEO_LAYERWISE_OFFLOAD_BUFFERS`                      | bool                           | `0`                                                   | performance | Layerwise offload also streams large buffers such as packed FP4/FP8 weights.                                                                                                                                                                             |
-| `FASTVIDEO_LAYERWISE_RESIDENT_BLOCKS`                      | int                            | `0`                                                   | performance | Keep the first N layerwise-offloaded blocks resident on the GPU.                                                                                                                                                                                         |
-| `FASTVIDEO_H3_SP_PROFILE`                                  | bool                           | `0`                                                   | profiling   | CUDA-event spans per stage over one MiniMax-H3 FP4 VSA DiT forward.                                                                                                                                                                                      |
-| `FASTVIDEO_H3_CAPTURE_QKV`                                 | path                           | unset                                                 | debug       | Directory for captured real MiniMax-H3 Q/K/V attention inputs.                                                                                                                                                                                           |
-| `FASTVIDEO_CUDA_MEMORY_CAP_GIB`                            | float                          | `0.0`                                                 | debug       | Cap this process's CUDA allocator at this many GiB to emulate a smaller GPU; 0 leaves it uncapped.                                                                                                                                                       |
-| `FASTVIDEO_MEMORY_REPORT`                                  | bool                           | `0`                                                   | debug       | Log bytes held per pipeline component by device and dtype after loading.                                                                                                                                                                                 |
 | `FASTVIDEO_TEST_LTX2_OVERFIT_DATA_DIR`                     | str                            | `data/cats`                                           | test        | Raw data directory for preprocess_ltx2_overfit.py. Deprecated names: `LTX2_OVERFIT_DATA_DIR`.                                                                                                                                                            |
 | `FASTVIDEO_TEST_LTX2_OVERFIT_CAPTION_JSON`                 | str                            | `videos2caption_1_sample.json`                        | test        | Caption file, relative to the raw data directory. Deprecated names: `LTX2_OVERFIT_CAPTION_JSON`.                                                                                                                                                         |
 | `FASTVIDEO_TEST_LTX2_OVERFIT_VIDEO_SUBDIR`                 | str                            | `video`                                               | test        | Video subdirectory, relative to the raw data directory. Deprecated names: `LTX2_OVERFIT_VIDEO_SUBDIR`.                                                                                                                                                   |
@@ -278,54 +227,16 @@ longer exists also fails the test, so the fixing pull request deletes its entry.
 | `FASTVIDEO_TEST_LTX2_OVERFIT_NUM_COPIES`                   | int                            | `4`                                                   | test        | Number of copies of the overfit sample in the parquet file. Deprecated names: `LTX2_OVERFIT_NUM_COPIES`.                                                                                                                                                 |
 | `FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR`               | str                            | `data/kandinsky5_overfit`                             | test        | Raw data directory for preprocess_kandinsky5_overfit.py. Deprecated names: `KANDINSKY5_OVERFIT_DATA_DIR`.                                                                                                                                                |
 | `FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR`             | str                            | `data/kandinsky5_overfit_preprocessed`                | test        | Output directory for preprocess_kandinsky5_overfit.py. Deprecated names: `KANDINSKY5_OVERFIT_OUTPUT_DIR`.                                                                                                                                                |
-| `FASTVIDEO_TEST_SSIM_REFERENCE_HF_REPO`                    | str                            | `FastVideo/ssim-reference-videos`                     | test        | Hugging Face repository that holds the SSIM reference videos. Deprecated names: `FASTVIDEO_SSIM_REFERENCE_HF_REPO`.                                                                                                                                      |
-| `FASTVIDEO_TEST_SSIM_REFERENCE_HF_REPO_TYPE`               | str                            | `dataset`                                             | test        | Repository type of FASTVIDEO_TEST_SSIM_REFERENCE_HF_REPO. Deprecated names: `FASTVIDEO_SSIM_REFERENCE_HF_REPO_TYPE`.                                                                                                                                     |
-| `FASTVIDEO_TEST_SSIM_SKIP_REFERENCE_DOWNLOAD`              | bool                           | `0`                                                   | test        | SSIM tests use local reference videos without downloading. Deprecated names: `FASTVIDEO_SSIM_SKIP_REFERENCE_DOWNLOAD`.                                                                                                                                   |
-| `FASTVIDEO_TEST_SSIM_FULL_QUALITY`                         | bool                           | `0`                                                   | test        | SSIM tests use the full-quality sampling configurations. Deprecated names: `FASTVIDEO_SSIM_FULL_QUALITY`.                                                                                                                                                |
-| `FASTVIDEO_TEST_NIGHTLY`                                   | bool                           | `0`                                                   | test        | Run the nightly end-to-end overfit tests. Deprecated names: `FASTVIDEO_NIGHTLY`.                                                                                                                                                                         |
-| `FASTVIDEO_TEST_ULYSSES_FAULT_RANK`                        | str                            | unset                                                 | test        | Rank that fails in the Ulysses fault-injection test. The test sets it for its worker processes. Deprecated names: `FASTVIDEO_ULYSSES_FAULT_RANK`.                                                                                                        |
-| `FASTVIDEO_TEST_ULYSSES_FAULT_STAGE`                       | str                            | unset                                                 | test        | Stage that fails in the Ulysses fault-injection test. The test sets it for its worker processes. Deprecated names: `FASTVIDEO_ULYSSES_FAULT_STAGE`.                                                                                                      |
-| `FASTVIDEO_TEST_GOLDEN_GATE_DIR`                           | str                            | unset                                                 | test        | Local directory of golden-gate reference tensors. Deprecated names: `FASTVIDEO_GOLDEN_GATE_DIR`.                                                                                                                                                         |
-| `FASTVIDEO_TEST_WAN22_5B_ALLOW_LOW_MEMORY`                 | bool                           | `0`                                                   | test        | Run the MLX Wan2.2 5B real-weights parity test on hosts with little memory. Deprecated names: `FASTVIDEO_WAN22_5B_ALLOW_LOW_MEMORY`.                                                                                                                     |
-| `FASTVIDEO_TEST_WAN22_5B_ROOT`                             | str                            | unset                                                 | test        | Local Wan2.2 5B checkpoint for the MLX real-weights parity test. Deprecated names: `FASTVIDEO_WAN22_5B_ROOT`.                                                                                                                                            |
-| `FASTVIDEO_TEST_GRADNORM_UPDATE`                           | bool                           | `0`                                                   | test        | Gradient-norm regression tests update their references. Deprecated names: `FASTVIDEO_GRADNORM_UPDATE`.                                                                                                                                                   |
-| `FASTVIDEO_TEST_FLUX_T2I_MODEL_DIR`                        | str                            | `black-forest-labs/FLUX.1-dev`                        | test        | Model for the Flux text-to-image SSIM test. Deprecated names: `FLUX_T2I_MODEL_DIR`.                                                                                                                                                                      |
-| `FASTVIDEO_TEST_FLUX_TRANSFORMER_PATH`                     | str                            | unset                                                 | test        | Local Flux transformer for the Flux transformer test. Deprecated names: `FLUX_TRANSFORMER_PATH`.                                                                                                                                                         |
-| `FASTVIDEO_TEST_GEN3C_MODEL_PATH`                          | str                            | `FastVideo/GEN3C-Cosmos-7B-Diffusers`                 | test        | Model for the GEN3C SSIM test. Deprecated names: `GEN3C_MODEL_PATH`.                                                                                                                                                                                     |
-| `FASTVIDEO_TEST_GEN3C_IMAGE_PATH`                          | str                            | unset                                                 | test        | Input image for the GEN3C SSIM test. Deprecated names: `GEN3C_TEST_IMAGE_PATH`.                                                                                                                                                                          |
-| `FASTVIDEO_TEST_KANDINSKY5_E2E_NUM_GPUS`                   | int                            | `1`                                                   | test        | GPUs for the Kandinsky5 nightly end-to-end overfit test. Deprecated names: `KANDINSKY5_E2E_NUM_GPUS`.                                                                                                                                                    |
-| `FASTVIDEO_TEST_KANDINSKY5_E2E_WRITE_REFERENCE`            | bool                           | `0`                                                   | test        | The Kandinsky5 nightly end-to-end test writes a missing reference video. Deprecated names: `KANDINSKY5_E2E_WRITE_REFERENCE`.                                                                                                                             |
-| `FASTVIDEO_TEST_MINIMAX_H3_GATE_GOLDEN_DIR`                | str                            | unset                                                 | test        | Local directory of MiniMax-H3 golden-gate tensors. Deprecated names: `MINIMAX_H3_GATE_GOLDEN_DIR`.                                                                                                                                                       |
-| `FASTVIDEO_TEST_MINIMAX_H3_GATE_LAYER`                     | int                            | `0`                                                   | test        | Transformer layer that the MiniMax-H3 golden-gate test checks. Deprecated names: `MINIMAX_H3_GATE_LAYER`.                                                                                                                                                |
-| `FASTVIDEO_TEST_MINIMAX_H3_MODEL_ROOT`                     | str                            | unset                                                 | test        | Local MiniMax-H3 checkpoint for the golden-gate test. Deprecated names: `MINIMAX_H3_MODEL_ROOT`.                                                                                                                                                         |
-| `FASTVIDEO_TEST_SD35_MODEL_DIR`                            | str                            | `stabilityai/stable-diffusion-3.5-medium`             | test        | Model for the Stable Diffusion 3.5 SSIM test. Deprecated names: `SD35_MODEL_DIR`.                                                                                                                                                                        |
-| `FASTVIDEO_TEST_TAEH3_REFERENCE_DIR`                       | str                            | unset                                                 | test        | Upstream taehv checkout for the MLX TAEH3 parity test. Deprecated names: `TAEH3_REFERENCE_DIR`.                                                                                                                                                          |
 
 Variables that FastVideo no longer reads; setting one logs a warning:
 
-| Deprecated variable                              | Reason                       |
-| ------------------------------------------------ | ---------------------------- |
-| `FASTVIDEO_TARGET_DEVICE`                        | no code reads it             |
-| `FASTVIDEO_USE_PRECOMPILED`                      | no code reads it             |
-| `FASTVIDEO_RINGBUFFER_WARNING_INTERVAL`          | no code reads it             |
-| `FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S`           | no code reads it             |
-| `FASTVIDEO_SERVER_DEV_MODE`                      | no code reads it             |
-| `FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE`        | no code reads it             |
-| `FASTVIDEO_TRACE_FUNCTION`                       | no code reads it             |
-| `FASTVIDEO_TEST_DREAMX_WORLD_SSIM_MODEL_PATH`    | DreamX World was removed     |
-| `DREAMX_WORLD_SSIM_MODEL_PATH`                   | DreamX World was removed     |
-| `FASTVIDEO_TEST_DREAMX_WORLD_AR_SSIM_MODEL_PATH` | DreamX World was removed     |
-| `DREAMX_WORLD_AR_SSIM_MODEL_PATH`                | DreamX World was removed     |
-| `FASTVIDEO_TEST_GAMECRAFT_MODEL_PATH`            | HunyuanGameCraft was removed |
-| `GAMECRAFT_MODEL_PATH`                           | HunyuanGameCraft was removed |
-| `FASTVIDEO_TEST_GLM_IMAGE_LOCAL_WEIGHTS_DIR`     | GLM-Image was removed        |
-| `GLM_IMAGE_LOCAL_WEIGHTS_DIR`                    | GLM-Image was removed        |
-| `FASTVIDEO_TEST_GLM_IMAGE_MODEL_DIR`             | GLM-Image was removed        |
-| `GLM_IMAGE_MODEL_DIR`                            | GLM-Image was removed        |
-| `FASTVIDEO_TEST_LONGCAT_MODEL_ROOT`              | LongCat was removed          |
-| `LONGCAT_MODEL_ROOT`                             | LongCat was removed          |
-| `FASTVIDEO_TEST_ZIMAGE_MODEL_DIR`                | Z-Image was removed          |
-| `ZIMAGE_MODEL_DIR`                               | Z-Image was removed          |
-| `FASTVIDEO_TEST_ZIMAGE_MODEL_REVISION`           | Z-Image was removed          |
-| `ZIMAGE_MODEL_REVISION`                          | Z-Image was removed          |
+| Deprecated variable                       | Reason           |
+| ----------------------------------------- | ---------------- |
+| `FASTVIDEO_TARGET_DEVICE`                 | no code reads it |
+| `FASTVIDEO_USE_PRECOMPILED`               | no code reads it |
+| `FASTVIDEO_RINGBUFFER_WARNING_INTERVAL`   | no code reads it |
+| `FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S`    | no code reads it |
+| `FASTVIDEO_SERVER_DEV_MODE`               | no code reads it |
+| `FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE` | no code reads it |
+| `FASTVIDEO_TRACE_FUNCTION`                | no code reads it |
 <!-- END GENERATED ENV TABLE -->

docs/contributing/gpu_ci_backends.md

@@ -0,0 +1,252 @@
+# Selectable GPU CI Backends
+
+GPU CI can use the existing Slurm dispatcher, Modal, or GB200 Kubernetes Jobs
+in the `vllm` namespace. GitHub and Buildkite remain the trigger and reporting
+systems. `vllm` names the cluster namespace here; tests run FastVideo's existing
+lane scripts rather than a vLLM inference server.
+
+The default remains Slurm. The existing `.buildkite/pipeline.yml`, private
+Slurm uploader, and dormant Modal launchers are preserved. Deploying the new
+trusted uploader is an operator step; merging these files does not change the
+live Buildkite pipelines or create cluster resources. See
+[CI/CD Architecture](ci_architecture.md) for the existing installation.
+
+## Execution and Limits
+
+```text
+GitHub PR / slash command / schedule
+  -> trusted Buildkite bootstrap
+     -> slurm: existing validated graph and private dispatcher
+     -> modal or vllm: one trusted suite coordinator
+        -> shared PR admission and GPU reservations
+        -> isolated GPU worker per existing lane
+        -> lane results, logs, and backend-specific suite status
+```
+
+All new Modal and `vllm` coordinators share one admission database. It enforces:
+
+- At most two distinct PRs with admitted GPU work.
+- At most four reserved GPUs per PR across its concurrent builds and lanes.
+- At most eight reserved GPUs across both backends combined.
+- Additional PRs wait without creating GPU workers.
+
+A PR is identified by repository and PR number, so Fastcheck, merge builds,
+manual reruns, and separate attempts share its allowance. Its slot remains
+occupied between lanes until all admitted builds for that PR finish. A
+non-PR build, including a schedule on `main`, consumes its own slot and the
+same GPU budget. The preserved Slurm dispatcher has its existing independent
+limits; this new admission database does not control legacy Slurm jobs.
+
+Lanes retain their existing one-, two-, or four-GPU requirements. Four-GPU
+SSIM and training lanes wait for that PR's other lanes to release capacity.
+Selected integration lanes wait for the golden gate and are skipped if it
+fails. CPU-only GitHub checks are unaffected. Pending admission is bounded by
+`queue_timeout_seconds`, initially six hours. The coordinator's Buildkite
+command timeout is eight hours, including admission and lane waits after the
+command starts; time waiting for a Buildkite agent is outside that timeout.
+
+Reservations persist before a worker is created. Cancellation releases them
+only after worker termination is confirmed. A lost coordinator, uncertain
+create response, or unavailable backend retains its reservations until
+recovery confirms cleanup. There is no heartbeat expiry that silently frees
+GPUs while an old worker could still be running.
+
+## Operator Installation
+
+Use an operator-reviewed immutable checkout under
+`/opt/fastvideo-gpu-ci/source`. Install the supplied `scripts/gpu_ci/run` and
+`scripts/gpu_ci/upload` wrappers as `/opt/fastvideo-gpu-ci/run` and
+`/opt/fastvideo-gpu-ci/upload`. They invoke the reviewed Python entrypoint
+with isolated import mode. Install the supplied files instead of generating
+shell wrappers from build metadata:
+
+```bash
+install -m 755 /opt/fastvideo-gpu-ci/source/scripts/gpu_ci/run /opt/fastvideo-gpu-ci/run
+install -m 755 /opt/fastvideo-gpu-ci/source/scripts/gpu_ci/upload /opt/fastvideo-gpu-ci/upload
+```
+
+The controller needs Python 3.10 or newer, PyYAML,
+the Buildkite agent, and `kubectl`; Modal additionally needs the reviewed
+Modal SDK and a controller-side Modal credential.
+
+Copy [the configuration example](../../scripts/gpu_ci/config.example.json)
+to `/etc/fastvideo-gpu-ci.json`. Keep the installation and configuration
+operator-owned and unwritable by workers. Configure the actual legacy
+`slurm_uploader` command, replace each enabled backend's image placeholder
+with a reviewed registry digest, and leave `default_backend` as `slurm`
+during canaries. The checked-in placeholders deliberately cannot start jobs.
+
+The permanent dispatcher must reach the Kubernetes API independently of a
+developer laptop, SSH tunnel, or Tailscale session. The existing CPU development
+pod is useful for investigating access, but is not a production CI service.
+Use a dedicated controller identity and a separate `gpu-ci-dispatch` Buildkite
+queue. The trusted bootstrap also needs access to the existing Slurm uploader
+while that backend remains enabled.
+
+Run all coordinators on **one host**, with `state_path` and `artifacts_dir` on
+its persistent local disk. The implementation uses SQLite transactions and
+local file locks. Do not place the database or locks on Lustre/NFS, run
+independent database copies, or scale controller replicas across nodes. Those
+configurations would invalidate the global limits. Multiple Buildkite agent
+processes on the same host can share the installation and ledger; enough
+agents are needed for concurrent suite coordinators and queued work.
+
+Configure agent-owned hooks to skip repository checkout and reject commands
+other than the trusted uploader and coordinator. Disable repository hooks
+and plugins on this queue. Do not use a PR checkout to load `scripts/gpu_ci`,
+its configuration, or its worker entrypoint. Build metadata is validated by
+the dispatcher; arbitrary environment variables are not forwarded into GPU
+workers. Buildkite, Kubernetes, and Modal control-plane credentials stay on
+the dispatcher.
+
+For Kubernetes, provision namespace-scoped permissions to create/get/delete
+Jobs and read Pods and their logs. Worker Pods disable service-account token
+mounting and request the exact GPU count on ARM64 GB200 nodes. The image must
+include the expected `/opt/venv` runtime, CUDA/SM100 support, and the reviewed
+FA4 dependencies. Use a separate AMD64 image digest for Modal. The Modal
+profile preserves H100 for encoder, custom-kernel, and VSA lanes, and L40S
+for the remaining lanes. Its reviewed image must support both SM89 and
+SM90a. Attention settings are selected per backend and lane to preserve the
+existing Modal and GB200 FA4 profiles; do not replace them with one global
+attention override. Validate the images against their respective hardware
+before enabling merge gates.
+
+Optional Kubernetes configuration includes `context`, `hf_secret`,
+`hf_secret_key`, `cache_pvc`, `cache_subpath`, `artifacts_pvc`, and
+`artifacts_subpath`. Only provide a read-only Hugging Face credential when
+private/gated downloads require it. PR workers are untrusted and can access
+any credential supplied to them, so never provide a reference-publication
+or other write-capable token. Prefer a pre-populated read-only cache.
+
+Use dedicated CI PVCs and relative CI subpaths. The personal
+`lustre-pvc-vllm` and `nfs-pvc-vllm` claims are rejected. The cache is mounted
+read-only; mutable references and locks stay inside the worker. The artifact
+init container prepares each Job's dedicated artifact subdirectory before
+the worker mounts it. An
+operator quota or admission policy scoped to CI can provide a second GPU
+ceiling; do not apply an eight-GPU quota to the shared `vllm` namespace if it
+would also cap other users' work.
+
+## Wire Every Buildkite Entry Pipeline
+
+Set each pipeline's operator-owned bootstrap command to
+`/opt/fastvideo-gpu-ci/upload`. Apply this to all three entry pipelines:
+
+| Pipeline | Trigger and scope |
+|---|---|
+| `pr-fastcheck` | Automatic PR webhook; `TEST_SCOPE=fastcheck` or unset. |
+| `ci` | Existing API triggers for Fastcheck, full, merge, direct, and scheduled SSIM. Keep its incoming PR webhook disabled to avoid duplicate builds. |
+| `fastvideo-performance-lane` | Existing schedule; `TEST_SCOPE=direct`, `TEST_TYPE=performance`. |
+
+The wrapper resolves `CI_GPU_BACKEND` from the build environment, falling
+back to `default_backend` in the trusted configuration. Allowed values are
+exactly `slurm`, `modal`, and `vllm`; unknown values fail. For Slurm, upload
+delegates to the unchanged private uploader. For Modal or `vllm`, it uploads
+one fixed `/opt/fastvideo-gpu-ci/run` command on the dedicated queue, with
+the selected backend and scope pinned in step environment.
+
+Set `CI_GPU_BACKEND=vllm` in the Buildkite build environment for a rack canary,
+or `modal` for a Modal canary. The existing API build payload can carry the
+same string in its `env` object. Automatic PR webhook builds inherit the
+operator's configured default unless pipeline/build configuration explicitly
+overrides it. Do not put backend selection inside a PR-controlled command.
+
+Keep the existing exact `BUILDKITE_COMMIT`, repository, PR identity, and
+`TEST_SCOPE` metadata. Direct runs also need an allowlisted `TEST_TYPE`.
+Merge runs need the trusted base-branch planner's `MERGE_TEST_PLAN`,
+`MERGE_GOLDEN_TESTS`, and `MERGE_SSIM_TESTS`. Full/direct/scheduled quality
+runs keep their complete matrices. The worker fetches and verifies the
+immutable commit before installing dependencies or invoking a lane script.
+
+The new Modal adapter uses bounded one-to-four-GPU sandboxes and the same
+lane scripts as Kubernetes. It does not reactivate `pr_test.sh` or the old
+Modal SSIM fan-out, which cannot enforce this shared four-GPU-per-PR budget.
+The legacy files remain available for their existing manual workflows.
+
+## Statuses and Default-Backend Cutover
+
+The selected adapter reports separate suite contexts:
+
+| Scope | GitHub context |
+|---|---|
+| Fastcheck | `gpu-ci/<backend>/fastcheck-passed` |
+| Merge or explicit full suite | `gpu-ci/<backend>/full-suite-passed` |
+| Direct lane | `gpu-ci/<backend>/direct-test-completed` |
+| Scheduled SSIM | `gpu-ci/<backend>/scheduled-ssim-passed` |
+
+The repository variable `CI_GPU_BACKEND` controls which new backend can
+publish the existing required `fastcheck-passed` and `full-suite-passed`
+contexts. Empty or `slurm` preserves the existing status behavior. `modal`
+or `vllm` enables `ci-gpu-backend-status.yml`, which reads the latest statuses
+for that one backend and mirrors both required contexts. A missing result
+becomes pending; failure and error remain failures. It never combines one
+backend's Fastcheck with another backend's full-suite result. Late status
+events re-read current state instead of replaying stale event payloads.
+
+Direct tests are diagnostic and do not promote a whole suite on the new
+backends. Rerun the matching Fastcheck or merge/full suite to clear its gate.
+Per-build tests on the other new backend remain separate diagnostics. The
+legacy direct-test aggregation workflow is disabled while a new backend is
+selected. These workflows retain the existing trust assumption that only
+authorized status-writing integrations can publish CI status contexts.
+
+For production cutover, drain existing Slurm builds: its preserved pipeline
+still emits canonical status contexts. The new uploader rejects a Slurm
+override when `default_backend` is `modal` or `vllm`, preventing later legacy
+builds from overwriting the promoted backend's checks. Ensure no old bootstrap
+bypasses the new uploader. Synchronize the operator `default_backend` and
+the GitHub repository variable, then clear or rerun required checks for all
+open PRs. Old green contexts do not become new-backend validation merely
+because a setting changed. Run both Fastcheck and the merge/full suite on
+the promoted backend before allowing merge. Apply the same drain and rerun
+procedure when rolling back to Slurm.
+
+## Validation and Recovery
+
+Inspect the rendered opt-in pipeline without submitting a build:
+
+```bash
+CI_GPU_BACKEND=vllm TEST_SCOPE=fastcheck \
+  /opt/fastvideo-gpu-ci/venv/bin/python -I \
+  /opt/fastvideo-gpu-ci/source/scripts/gpu_ci/entrypoint.py render \
+  --config /etc/fastvideo-gpu-ci.json
+```
+
+Validate a one-GPU lane, then a two-/four-GPU lane and the complete Fastcheck
+suite. Exercise two PRs plus a third waiter, concurrent builds of the same
+PR, cancellation, retries, and coordinator restart. Confirm observed GPU
+reservations never exceed two PRs, four per PR, or eight in total. A GPU
+reservation includes a pending worker, so unavailable nodes cannot cause
+the controller to submit more work than its allowance.
+
+Run SSIM, training, and performance canaries separately. References must
+match the effective GPU/runtime/attention backend; do not silently reuse
+L40S performance results as GB200 baselines or reseed references as part of
+routine CI. Workers keep W&B offline and disable reference publication.
+
+Inspect the persistent ledger and recover an abandoned attempt with:
+
+```bash
+/opt/fastvideo-gpu-ci/venv/bin/python -I \
+  /opt/fastvideo-gpu-ci/source/scripts/gpu_ci/entrypoint.py status \
+  --config /etc/fastvideo-gpu-ci.json
+
+/opt/fastvideo-gpu-ci/venv/bin/python -I \
+  /opt/fastvideo-gpu-ci/source/scripts/gpu_ci/entrypoint.py recover \
+  --config /etc/fastvideo-gpu-ci.json --build-id BUILD_ID.JOB_ID
+```
+
+Use the exact ledger ID from `status`; each Buildkite retry has a distinct
+job ID. Recovery refuses a live coordinator, persists cancellation, stops
+owned resources, and releases reservations only after confirming termination.
+If creation or cleanup is ambiguous, investigate the recorded handle on its
+backend and retain the reservation until the outcome is known. Do not delete
+the database or manually zero counters to unblock the queue.
+
+The dispatcher uploads controller logs, per-lane numeric results, request
+metadata, and the suite summary to Buildkite. These are the sources for its
+exit status. Generated videos and JUnit files stay in the worker unless a
+dedicated Kubernetes artifact PVC is configured; automatic publication of
+those worker files and Modal worker artifacts is not implemented. This is
+a deployment limitation to account for before replacing existing artifact
+review workflows.

docs/cookbook/cosmos.md

@@ -5,7 +5,7 @@ hide:
 
 # Cosmos recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="cosmos" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="cosmos" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/flux.md

@@ -5,7 +5,7 @@ hide:
 
 # FLUX recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="flux" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="flux" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/glm-image.md

@@ -0,0 +1,140 @@
+---
+hide:
+- toc
+---
+
+# GLM-Image recipes
+
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="glm_image" data-recipes="../../assets/cookbook-recipes.json?v=11">
+  <header class="cookbook-family-header">
+    <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
+    <div class="cookbook-family-header__body">
+      <span class="cookbook-family-header__logo">
+        <img class="off-glb" src="../../assets/logos/zai.webp" alt="Z.ai" width="112" height="112">
+      </span>
+      <div>
+        <p class="cookbook-eyebrow">Maintained family · Inference</p>
+        <h2>GLM-Image inference recipes</h2>
+        <p>GLM-Image from Z.ai supports both text-to-image generation and instruction-based image editing, each with a maintained example.</p>
+      </div>
+    </div>
+    <div class="cookbook-lifecycle" aria-label="Lifecycle stages">
+      <span class="cookbook-lifecycle__stage cookbook-lifecycle__stage--active">Inference <small>live</small></span>
+      <span class="cookbook-lifecycle__stage">Distillation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Fine-tuning <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Training <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Evaluation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Optimization <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Deployment <small>planned</small></span>
+    </div>
+  </header>
+  <nav class="cookbook-jumpnav" aria-label="Recipe page sections">
+    <a href="#recipe-builder">Builder</a>
+    <a href="#cookbook-setup">Setup</a>
+    <a href="#cookbook-troubleshooting">Troubleshooting</a>
+    <a href="#cookbook-evidence">Evidence</a>
+  </nav>
+
+  <section class="cookbook-builder" id="recipe-builder" aria-labelledby="builder-heading">
+    <div class="cookbook-builder__intro">
+      <h2 id="builder-heading">Pick a recipe and runtime</h2>
+      <p>Start with the result you want, then choose one of the runtimes FastVideo actually maintains for it.</p>
+    </div>
+
+    <div class="cookbook-builder__layout">
+      <div class="cookbook-controls">
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Recipe</strong>
+            <span>Task and checkpoint</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--models" data-cookbook-model-options role="group" aria-label="Recipe">
+            <button type="button" disabled>Loading GLM-Image recipes...</button>
+          </div>
+        </div>
+
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Runtime</strong>
+            <span>Maintained paths only</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--hardware" data-cookbook-hardware-options role="group" aria-label="Runtime">
+            <button type="button" disabled>Loading runtimes...</button>
+          </div>
+        </div>
+
+        <p class="cookbook-selection-description" data-cookbook-description>Loading recipe details...</p>
+        <p class="cookbook-hardware-note">Exact device and memory details appear only when a recorded run supports them.</p>
+
+        <div class="cookbook-hardware-state" data-cookbook-hardware-state role="status" aria-live="polite">
+          Reading recipe evidence...
+        </div>
+      </div>
+
+      <article class="cookbook-result">
+        <div class="cookbook-result__header">
+          <h3 data-cookbook-label>Loading...</h3>
+          <div class="cookbook-result__badges">
+            <span class="cookbook-badge">Maintained</span>
+            <span class="cookbook-badge" data-cookbook-evidence>Source-backed</span>
+            <span class="cookbook-badge cookbook-badge--neutral" data-cookbook-hardware-badge>Source config</span>
+          </div>
+        </div>
+
+        <dl class="cookbook-result__facts">
+          <div><dt>Model</dt><dd data-cookbook-model>Loading...</dd></div>
+          <div><dt>Workload</dt><dd data-cookbook-task>Loading...</dd></div>
+          <div><dt>Source configuration</dt><dd data-cookbook-gpus>Loading...</dd></div>
+          <div><dt>Expected output</dt><dd data-cookbook-artifact>Loading...</dd></div>
+        </dl>
+
+        <div class="cookbook-command">
+          <div class="cookbook-command__bar">
+            <span>Terminal</span>
+          </div>
+          <pre><code class="language-bash" data-cookbook-command>Loading...</code></pre>
+        </div>
+
+        <div class="cookbook-result__footer">
+          <a data-cookbook-source href="../../inference/examples/basic/">Open example source</a>
+          <a data-cookbook-model-link href="https://huggingface.co/zai-org">View model card</a>
+        </div>
+        <p class="cookbook-picker__status" role="status" aria-live="polite" data-cookbook-status></p>
+      </article>
+    </div>
+
+    <noscript>
+      <div class="cookbook-noscript">
+        JavaScript is needed for the guided selector. You can still browse the
+        <a href="../../inference/examples/examples_inference_index/">maintained inference examples</a>.
+      </div>
+    </noscript>
+  </section>
+</div>
+
+<details class="cookbook-collapsible" id="cookbook-setup">
+  <summary>Setup</summary>
+  <div class="cookbook-collapsible__body">
+      <p>The generated commands expect a local clone:</p>
+      <pre><code>git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo</code></pre>
+      <p>Use <a href="../../inference/configuration/">Configuration</a> for supported Python and CLI settings, <a href="../../inference/optimizations/">Optimizations</a> for attention and memory tradeoffs, and the <a href="../../inference/support_matrix/">support matrix</a> for the supported model and optimization surface.</p>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>The editing example reads <code>assets/images/couple.jpg</code> from the repository root, so run it from a repo checkout rather than an arbitrary working directory.</li>
+        <li>Output paths default under <code>image_output/</code>; pass <code>--output</code> to change them.</li>
+      </ul>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>All recipes on this page are <strong>Source-backed</strong>: their commands, model IDs, and flags were validated against the checked-in FastVideo sources listed above (static validation). No runtime GPU validation is recorded for these recipes, so GPU model fit, memory use, throughput, and runtime duration are <strong>Unknown</strong> and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.</p>
+  </div>
+</details>

docs/cookbook/hunyuan.md

@@ -5,7 +5,7 @@ hide:
 
 # Hunyuan recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="hunyuan" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="hunyuan" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/index.md

@@ -141,25 +141,46 @@ hide:
         </span>
       </a>
 
-      <a class="cookbook-family-tile cookbook-family-tile--ready" href="./kandinsky6/" aria-label="Open Kandinsky 6 recipes">
+      <a class="cookbook-family-tile cookbook-family-tile--ready" href="./longcat/" aria-label="Open LongCat recipes">
         <span class="cookbook-family-tile__visual" data-evervault>
           <span class="cookbook-evervault" aria-hidden="true">
             <span class="cookbook-evervault__gradient"></span>
             <span class="cookbook-evervault__noise" data-cookbook-pattern></span>
           </span>
           <span class="cookbook-family-tile__logo-wrap">
-            <img class="off-glb" src="../assets/logos/kandinsky.webp" alt="" width="132" height="132" loading="lazy">
+            <img class="off-glb" src="../assets/logos/meituan-longcat.webp" alt="" width="132" height="132" loading="lazy">
           </span>
         </span>
         <span class="cookbook-family-tile__footer">
           <span class="cookbook-family-tile__footer-top">
-            <span><strong>Kandinsky 6</strong><small>Video with audio, and video SR</small></span>
-            <span class="cookbook-count">4 recipes</span>
+            <span><strong>LongCat</strong><small>T2V, I2V, optional refine</small></span>
+            <span class="cookbook-count">2 recipes</span>
           </span>
           <ul class="cookbook-mode-row">
-            <li>T2VA</li>
-            <li>I2VA</li>
-            <li>VSR</li>
+            <li>T2V</li>
+            <li>I2V</li>
+          </ul>
+        </span>
+      </a>
+
+      <a class="cookbook-family-tile cookbook-family-tile--ready" href="./turbodiffusion/" aria-label="Open TurboDiffusion recipes">
+        <span class="cookbook-family-tile__visual" data-evervault>
+          <span class="cookbook-evervault" aria-hidden="true">
+            <span class="cookbook-evervault__gradient"></span>
+            <span class="cookbook-evervault__noise" data-cookbook-pattern></span>
+          </span>
+          <span class="cookbook-family-tile__logo-wrap">
+            <span class="cookbook-family-tile__monogram" aria-hidden="true">Turbo</span>
+          </span>
+        </span>
+        <span class="cookbook-family-tile__footer">
+          <span class="cookbook-family-tile__footer-top">
+            <span><strong>TurboDiffusion</strong><small>Accelerated Wan profiles</small></span>
+            <span class="cookbook-count">3 recipes</span>
+          </span>
+          <ul class="cookbook-mode-row">
+            <li>T2V</li>
+            <li>I2V</li>
           </ul>
         </span>
       </a>
@@ -192,6 +213,49 @@ hide:
         </span>
       </a>
 
+      <a class="cookbook-family-tile cookbook-family-tile--ready" href="./glm-image/" aria-label="Open GLM-Image recipes">
+        <span class="cookbook-family-tile__visual" data-evervault>
+          <span class="cookbook-evervault" aria-hidden="true">
+            <span class="cookbook-evervault__gradient"></span>
+            <span class="cookbook-evervault__noise" data-cookbook-pattern></span>
+          </span>
+          <span class="cookbook-family-tile__logo-wrap">
+            <img class="off-glb" src="../assets/logos/zai.webp" alt="" width="132" height="132" loading="lazy">
+          </span>
+        </span>
+        <span class="cookbook-family-tile__footer">
+          <span class="cookbook-family-tile__footer-top">
+            <span><strong>GLM-Image</strong><small>Generate and edit</small></span>
+            <span class="cookbook-count">2 recipes</span>
+          </span>
+          <ul class="cookbook-mode-row">
+            <li>T2I</li>
+            <li>Edit</li>
+          </ul>
+        </span>
+      </a>
+
+      <a class="cookbook-family-tile cookbook-family-tile--ready" href="./z-image/" aria-label="Open Z-Image recipes">
+        <span class="cookbook-family-tile__visual" data-evervault>
+          <span class="cookbook-evervault" aria-hidden="true">
+            <span class="cookbook-evervault__gradient"></span>
+            <span class="cookbook-evervault__noise" data-cookbook-pattern></span>
+          </span>
+          <span class="cookbook-family-tile__logo-wrap">
+            <img class="off-glb" src="../assets/logos/tongyi.webp" alt="" width="132" height="132" loading="lazy">
+          </span>
+        </span>
+        <span class="cookbook-family-tile__footer">
+          <span class="cookbook-family-tile__footer-top">
+            <span><strong>Z-Image</strong><small>Turbo text to image</small></span>
+            <span class="cookbook-count">1 recipe</span>
+          </span>
+          <ul class="cookbook-mode-row">
+            <li>T2I</li>
+          </ul>
+        </span>
+      </a>
+
       <a class="cookbook-family-tile cookbook-family-tile--ready" href="./stable-diffusion/" aria-label="Open Stable Diffusion recipes">
         <span class="cookbook-family-tile__visual" data-evervault>
           <span class="cookbook-evervault" aria-hidden="true">
@@ -320,6 +384,20 @@ hide:
       <p>These families already have runnable examples. The cookbook page is not ready, so the cards are not links.</p>
     </div>
     <div class="cookbook-family-grid">
+      <article class="cookbook-family-tile cookbook-family-tile--coming" aria-label="GameCraft cookbook page planned; runnable examples exist">
+        <span class="cookbook-family-tile__visual">
+          <span class="cookbook-family-tile__logo-wrap">
+            <img class="off-glb" src="../assets/logos/tencent-hunyuan.webp" alt="" width="132" height="132" loading="lazy">
+          </span>
+        </span>
+        <span class="cookbook-family-tile__footer">
+          <span class="cookbook-family-tile__footer-top">
+            <span><strong>GameCraft</strong><small>Game world generation</small></span>
+            <span class="cookbook-count">Page planned</span>
+          </span>
+        </span>
+      </article>
+
       <article class="cookbook-family-tile cookbook-family-tile--coming" aria-label="GEN3C cookbook page planned; runnable examples exist">
         <span class="cookbook-family-tile__visual">
           <span class="cookbook-family-tile__logo-wrap">
@@ -348,6 +426,20 @@ hide:
         </span>
       </article>
 
+      <article class="cookbook-family-tile cookbook-family-tile--coming" aria-label="DreamX cookbook page planned; runnable examples exist">
+        <span class="cookbook-family-tile__visual">
+          <span class="cookbook-family-tile__logo-wrap">
+            <img class="off-glb" src="../assets/logos/fastvideo.webp" alt="" width="132" height="132" loading="lazy">
+          </span>
+        </span>
+        <span class="cookbook-family-tile__footer">
+          <span class="cookbook-family-tile__footer-top">
+            <span><strong>DreamX</strong><small>World generation</small></span>
+            <span class="cookbook-count">Page planned</span>
+          </span>
+        </span>
+      </article>
+
       <article class="cookbook-family-tile cookbook-family-tile--coming" aria-label="LingBot cookbook page planned; runnable examples exist">
         <span class="cookbook-family-tile__visual">
           <span class="cookbook-family-tile__logo-wrap">

docs/cookbook/kandinsky5.md

@@ -5,7 +5,7 @@ hide:
 
 # Kandinsky 5 recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="kandinsky5" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="kandinsky5" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/longcat.md

@@ -0,0 +1,140 @@
+---
+hide:
+- toc
+---
+
+# LongCat recipes
+
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="longcat" data-recipes="../../assets/cookbook-recipes.json?v=11">
+  <header class="cookbook-family-header">
+    <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
+    <div class="cookbook-family-header__body">
+      <span class="cookbook-family-header__logo">
+        <img class="off-glb" src="../../assets/logos/meituan-longcat.webp" alt="Meituan LongCat" width="112" height="112">
+      </span>
+      <div>
+        <p class="cookbook-eyebrow">Maintained family · Inference</p>
+        <h2>LongCat inference recipes</h2>
+        <p>LongCat Video from Meituan covers text-to-video and image-to-video, and its maintained examples chain optional distilled and 720p refinement passes.</p>
+      </div>
+    </div>
+    <div class="cookbook-lifecycle" aria-label="Lifecycle stages">
+      <span class="cookbook-lifecycle__stage cookbook-lifecycle__stage--active">Inference <small>live</small></span>
+      <span class="cookbook-lifecycle__stage">Distillation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Fine-tuning <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Training <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Evaluation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Optimization <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Deployment <small>planned</small></span>
+    </div>
+  </header>
+  <nav class="cookbook-jumpnav" aria-label="Recipe page sections">
+    <a href="#recipe-builder">Builder</a>
+    <a href="#cookbook-setup">Setup</a>
+    <a href="#cookbook-troubleshooting">Troubleshooting</a>
+    <a href="#cookbook-evidence">Evidence</a>
+  </nav>
+
+  <section class="cookbook-builder" id="recipe-builder" aria-labelledby="builder-heading">
+    <div class="cookbook-builder__intro">
+      <h2 id="builder-heading">Pick a recipe and runtime</h2>
+      <p>Start with the result you want, then choose one of the runtimes FastVideo actually maintains for it.</p>
+    </div>
+
+    <div class="cookbook-builder__layout">
+      <div class="cookbook-controls">
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Recipe</strong>
+            <span>Task and checkpoint</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--models" data-cookbook-model-options role="group" aria-label="Recipe">
+            <button type="button" disabled>Loading LongCat recipes...</button>
+          </div>
+        </div>
+
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Runtime</strong>
+            <span>Maintained paths only</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--hardware" data-cookbook-hardware-options role="group" aria-label="Runtime">
+            <button type="button" disabled>Loading runtimes...</button>
+          </div>
+        </div>
+
+        <p class="cookbook-selection-description" data-cookbook-description>Loading recipe details...</p>
+        <p class="cookbook-hardware-note">Exact device and memory details appear only when a recorded run supports them.</p>
+
+        <div class="cookbook-hardware-state" data-cookbook-hardware-state role="status" aria-live="polite">
+          Reading recipe evidence...
+        </div>
+      </div>
+
+      <article class="cookbook-result">
+        <div class="cookbook-result__header">
+          <h3 data-cookbook-label>Loading...</h3>
+          <div class="cookbook-result__badges">
+            <span class="cookbook-badge">Maintained</span>
+            <span class="cookbook-badge" data-cookbook-evidence>Source-backed</span>
+            <span class="cookbook-badge cookbook-badge--neutral" data-cookbook-hardware-badge>Source config</span>
+          </div>
+        </div>
+
+        <dl class="cookbook-result__facts">
+          <div><dt>Model</dt><dd data-cookbook-model>Loading...</dd></div>
+          <div><dt>Workload</dt><dd data-cookbook-task>Loading...</dd></div>
+          <div><dt>Source configuration</dt><dd data-cookbook-gpus>Loading...</dd></div>
+          <div><dt>Expected output</dt><dd data-cookbook-artifact>Loading...</dd></div>
+        </dl>
+
+        <div class="cookbook-command">
+          <div class="cookbook-command__bar">
+            <span>Terminal</span>
+          </div>
+          <pre><code class="language-bash" data-cookbook-command>Loading...</code></pre>
+        </div>
+
+        <div class="cookbook-result__footer">
+          <a data-cookbook-source href="../../inference/examples/basic/">Open example source</a>
+          <a data-cookbook-model-link href="https://huggingface.co/meituan-longcat">View model card</a>
+        </div>
+        <p class="cookbook-picker__status" role="status" aria-live="polite" data-cookbook-status></p>
+      </article>
+    </div>
+
+    <noscript>
+      <div class="cookbook-noscript">
+        JavaScript is needed for the guided selector. You can still browse the
+        <a href="../../inference/examples/examples_inference_index/">maintained inference examples</a>.
+      </div>
+    </noscript>
+  </section>
+</div>
+
+<details class="cookbook-collapsible" id="cookbook-setup">
+  <summary>Setup</summary>
+  <div class="cookbook-collapsible__body">
+      <p>The generated commands expect a local clone:</p>
+      <pre><code>git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo</code></pre>
+      <p>Use <a href="../../inference/configuration/">Configuration</a> for supported Python and CLI settings, <a href="../../inference/optimizations/">Optimizations</a> for attention and memory tradeoffs, and the <a href="../../inference/support_matrix/">support matrix</a> for the supported model and optimization surface.</p>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Each LongCat script runs multiple passes (basic, distilled, refine); total runtime scales accordingly, and every pass prints its own output directory.</li>
+        <li>Out of memory: the sources already enable VAE and text-encoder CPU offload; further options are covered in <a href="../../inference/offloading/">Offloading</a>.</li>
+      </ul>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>All recipes on this page are <strong>Source-backed</strong>: their commands, model IDs, and flags were validated against the checked-in FastVideo sources listed above (static validation). No runtime GPU validation is recorded for these recipes, so GPU model fit, memory use, throughput, and runtime duration are <strong>Unknown</strong> and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.</p>
+  </div>
+</details>

docs/cookbook/ltx.md

@@ -5,7 +5,7 @@ hide:
 
 # LTX recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="ltx2" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="ltx2" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/matrix-game.md

@@ -5,7 +5,7 @@ hide:
 
 # Matrix Game recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="matrixgame" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="matrixgame" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/minimax-h3.md

@@ -10,15 +10,8 @@ launch. Some Hub repo names still say Preview. That name is historical. V1 is
 a full model, not a demo. **V2** is the eight-step checkpoint. More forwards
 is why V2 is the higher-quality FastH3. The V2 schedule contract is in
 [FastH3 distilled checkpoint schedules](../inference/fasth3-distilled.md).
-**CompactH3** is the 42-block 20B NVFP4 H3 checkpoint for one Blackwell GPU
-(RTX 5090 or RTX PRO 6000). The FastH3 V1 and V2 recipes are unchanged.
 
-The 42-block pruned checkpoint has an [MLX INT8/INT6 conversion and
-eight-forward T2VA command](../getting_started/installation/mlx.md#pruned-eight-forward-checkpoint).
-It reads `fastvideo_inference.json` for the trained schedule. The command
-uses native 832x480 resolution and all requested frames.
-
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="minimax_h3" data-default-recipe="fasth3-preview-cuda" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="minimax_h3" data-default-recipe="fasth3-preview-cuda" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">
@@ -28,8 +21,8 @@ uses native 832x480 resolution and all requested frames.
       <div>
         <p class="cookbook-eyebrow">Primary focus · Inference</p>
         <h2>MiniMax H3 recipes</h2>
-        <p>Generate video and audio with H3. Run a server on CUDA, one Blackwell GPU, one DGX Spark, or Apple Silicon MLX to iterate on prompts, or call the pipeline directly from Python.</p>
-        <span class="cookbook-count" data-cookbook-count>11 maintained recipes</span>
+        <p>Generate video and audio with H3. Run a server on CUDA, one DGX Spark, or Apple Silicon MLX to iterate on prompts, or call the pipeline directly from Python.</p>
+        <span class="cookbook-count" data-cookbook-count>9 maintained recipes</span>
       </div>
     </div>
     <div class="cookbook-lifecycle" aria-label="Lifecycle stages">
@@ -54,7 +47,7 @@ uses native 832x480 resolution and all requested frames.
     <h2 id="h3-modes-heading">Supported modes</h2>
     <p>
       CUDA covers T2VA, FL2VA, and Ref2VA on the full checkpoint, plus FastH3
-      V1 and FastH3 V2, plus CompactH3 NVFP4 on one Blackwell GPU. FastH3 V1 also has a DGX Spark runtime with
+      V1 and FastH3 V2. FastH3 V1 also has a DGX Spark runtime with
       a 1-Spark or 2-Spark device row. MLX is T2VA only: V1 and V2.
       Temporal <code>--fast</code>, spatial <code>--fast-spatial</code>, and opt-in VSA are flags on the same
       MLX script, not extra recipes.
@@ -71,7 +64,7 @@ uses native 832x480 resolution and all requested frames.
         <tbody>
           <tr>
             <td>T2VA</td>
-            <td>Full H3, FastH3 V1, FastH3 LoRA, FastH3 V2, CompactH3 NVFP4</td>
+            <td>Full H3, FastH3 V1, FastH3 LoRA, FastH3 V2</td>
             <td>FastH3 V1 or FastH3 V2 after a local DiT conversion</td>
           </tr>
           <tr>
@@ -109,11 +102,6 @@ uses native 832x480 resolution and all requested frames.
             <td>FastH3 V1 on one GB10, or two Sparks with Ray sequence parallel (<code>sp_size=2</code>) over QSFP RoCE. Select NVIDIA DGX Spark, then 1 Spark or 2 Sparks.</td>
             <td>Not wired</td>
           </tr>
-          <tr>
-            <td>CompactH3 NVFP4</td>
-            <td>42-block 20B checkpoint on one RTX 5090 (32 GB, sequential encoder offload) or one RTX PRO 6000 Blackwell (96 GB, encoder+DiT+VAE resident). SageAttention3 FP4, packed NVFP4 DiT, Comfy int8-convrot VAE.</td>
-            <td>Not wired</td>
-          </tr>
         </tbody>
       </table>
     </div>
@@ -122,7 +110,7 @@ uses native 832x480 resolution and all requested frames.
   <section class="cookbook-builder" id="recipe-builder" aria-labelledby="builder-heading">
     <div class="cookbook-builder__intro">
       <h2 id="builder-heading">Pick an H3 recipe and runtime</h2>
-      <p>Choose the result you want, then use a maintained CUDA, Blackwell, DGX Spark, or MLX path.
+      <p>Choose the result you want, then use a maintained CUDA, DGX Spark, or MLX path.
       Device claims stay tied to checked-in sources and recorded runs.</p>
     </div>
 
@@ -166,7 +154,7 @@ uses native 832x480 resolution and all requested frames.
             <span>Both can run locally</span>
           </div>
           <div class="cookbook-option-grid cookbook-option-grid--hardware" role="group" aria-label="How to run this recipe">
-            <button type="button" data-cookbook-mode="server" aria-pressed="false"><strong>Run a server</strong><span data-cookbook-server-hint>Playground, cURL, or an API client</span></button>
+            <button type="button" data-cookbook-mode="server" aria-pressed="false"><strong>Run a server</strong><span>Playground, cURL, or an API client</span></button>
             <button type="button" data-cookbook-mode="python" aria-pressed="false"><strong>Use Python directly</strong><span>Call the model in your own process</span></button>
           </div>
         </div>
@@ -213,17 +201,17 @@ uses native 832x480 resolution and all requested frames.
           </section>
           <section class="cookbook-serving__step" aria-labelledby="serving-start-heading">
             <h4 id="serving-start-heading"><span aria-hidden="true">2</span> Start the server</h4>
-            <p>Keep this terminal running while you use <span data-cookbook-playground-only>the playground or </span>API clients.</p>
+            <p>Keep this terminal running while you use the playground or API clients.</p>
             <div class="cookbook-command"><div class="cookbook-command__bar"><span>GPU machine · Terminal</span></div><pre id="cookbook-server-command"><code class="language-bash" data-cookbook-server-command></code></pre></div>
             <details class="cookbook-serving__check"><summary>Check that the server is ready</summary><p>In another terminal, this returns <code>{"status":"ok"}</code> after startup.</p><div class="cookbook-command"><pre id="cookbook-health-command"><code class="language-bash" data-cookbook-health-command></code></pre></div></details>
           </section>
           <section class="cookbook-serving__step" aria-labelledby="serving-client-heading">
             <h4 id="serving-client-heading"><span aria-hidden="true">3</span> Generate and download a video</h4>
-            <div class="cookbook-serving__playground" data-cookbook-playground-only>
+            <div class="cookbook-serving__playground">
               <div><strong>Try prompts in your browser</strong><p>Edit a prompt, generate, and watch the result. The playground uses the same server as cURL and your app.</p></div>
               <a class="cookbook-serving__launch" data-cookbook-playground href="http://127.0.0.1:8000/playground/" target="_blank" rel="noopener">Open playground <span aria-hidden="true">↗</span></a>
             </div>
-            <p class="cookbook-serving__local-hint"><span data-cookbook-playground-only>Open after the server is ready. </span>On a remote GPU machine, <a href="../openai-api/#connect-your-app">forward port 8000</a> to your computer first.<span data-cookbook-playground-only> This opens a local page, not a hosted demo.</span></p>
+            <p class="cookbook-serving__local-hint">Open after the server is ready. On a remote GPU machine, <a href="../openai-api/#connect-your-app">forward port 8000</a> to your computer first. This opens a local page, not a hosted demo.</p>
             <details class="cookbook-serving__code"><summary>Use cURL or an SDK</summary>
             <p>Each example submits a job, checks its status, and saves the MP4. The Python and JavaScript examples use OpenAI-compatible clients; no OpenAI account is needed.</p>
             <div class="cookbook-serving__clients" role="group" aria-label="API client language">
@@ -287,8 +275,7 @@ cd FastVideo</code></pre>
         <li>The MLX source runtime supports T2VA, optional temporal <code>--fast</code>, optional spatial <code>--fast-spatial</code>, and opt-in VSA on <code>--include-vsa</code> checkpoints. FastH3 V2 MLX converts with <code>--include-vsa</code> and runs eight forwards. FL2VA, Ref2VA, and two-pass refinement are not wired.</li>
         <li>GPU count and VAE decode backend are configurable in the builder above for FastH3 CUDA recipes. Only the value shown by default has a recorded run; other supported values are unmeasured here.</li>
         <li>DGX Spark is a runtime on FastH3 V1, not a separate family card. Select NVIDIA DGX Spark, then 1 Spark or 2 Sparks. The CUDA GPU-count knob does not apply to Spark.</li>
-        <li>CompactH3 NVFP4 is one Blackwell GPU. RTX 5090 (32 GB) parks the encoder in pinned host RAM. RTX PRO 6000 Blackwell (96 GB) keeps encoder, DiT, and VAE resident. Keep <code>FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER</code>, <code>FASTVIDEO_FA4=0</code>, <code>FASTVIDEO_VSA_SM100A=0</code>, and <code>FLASHINFER_CUDA_ARCH_LIST=12.0a</code>. On PRO 6000 enable VAE compile and leave DiT <code>inference_torch_compile</code> off. CompactH3 is a dense prune; do not enable <code>VIDEO_SPARSE_ATTN_H3</code> until a VSA-trained student exists.</li>
-        <li>GB10 has no FA4 / sm_100a VSA kernel. Keep <code>FASTVIDEO_FA4=0</code> and <code>FASTVIDEO_VSA_SM100A=0</code>. Legal <code>num_frames</code> values are <code>17n+5</code>, capped at 362 (15.08 s). A 345-frame request on one Spark can OOM. Native 16:9 sizes include 832×480 and 1344×768.</li>
+        <li>GB10 has no FA4 / sm_100a VSA kernel. Keep <code>FASTVIDEO_FA4=0</code> and <code>FASTVIDEO_VSA_SM100A=0</code>. Legal <code>num_frames</code> values are <code>17n+5</code>, capped at 362 (15.08 s). A 345-frame request on one Spark can OOM.</li>
         <li>Gated or missing checkpoints: run <code>huggingface-cli login</code> and confirm you accepted the model's license on Hugging Face.</li>
       </ul>
   </div>

docs/cookbook/mmaudio.md

@@ -5,7 +5,7 @@ hide:
 
 # MMAudio recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="mmaudio" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="mmaudio" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/stable-audio.md

@@ -5,7 +5,7 @@ hide:
 
 # Stable Audio recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="stable_audio" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="stable_audio" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/stable-diffusion.md

@@ -5,7 +5,7 @@ hide:
 
 # Stable Diffusion recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="sd35" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="sd35" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">

docs/cookbook/turbodiffusion.md

@@ -0,0 +1,140 @@
+---
+hide:
+- toc
+---
+
+# TurboDiffusion recipes
+
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="turbodiffusion" data-recipes="../../assets/cookbook-recipes.json?v=11">
+  <header class="cookbook-family-header">
+    <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
+    <div class="cookbook-family-header__body">
+      <span class="cookbook-family-header__logo">
+        <span class="cookbook-family-tile__monogram" aria-hidden="true">Turbo</span>
+      </span>
+      <div>
+        <p class="cookbook-eyebrow">Maintained family · Inference</p>
+        <h2>TurboDiffusion inference recipes</h2>
+        <p>TurboDiffusion profiles accelerate Wan checkpoints with step-distilled sampling and the SLA attention backend. These recipes follow the registry's `turbodiffusion` model family.</p>
+      </div>
+    </div>
+    <div class="cookbook-lifecycle" aria-label="Lifecycle stages">
+      <span class="cookbook-lifecycle__stage cookbook-lifecycle__stage--active">Inference <small>live</small></span>
+      <span class="cookbook-lifecycle__stage">Distillation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Fine-tuning <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Training <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Evaluation <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Optimization <small>planned</small></span>
+      <span class="cookbook-lifecycle__stage">Deployment <small>planned</small></span>
+    </div>
+  </header>
+  <nav class="cookbook-jumpnav" aria-label="Recipe page sections">
+    <a href="#recipe-builder">Builder</a>
+    <a href="#cookbook-setup">Setup</a>
+    <a href="#cookbook-troubleshooting">Troubleshooting</a>
+    <a href="#cookbook-evidence">Evidence</a>
+  </nav>
+
+  <section class="cookbook-builder" id="recipe-builder" aria-labelledby="builder-heading">
+    <div class="cookbook-builder__intro">
+      <h2 id="builder-heading">Pick a recipe and runtime</h2>
+      <p>Start with the result you want, then choose one of the runtimes FastVideo actually maintains for it.</p>
+    </div>
+
+    <div class="cookbook-builder__layout">
+      <div class="cookbook-controls">
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Recipe</strong>
+            <span>Task and checkpoint</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--models" data-cookbook-model-options role="group" aria-label="Recipe">
+            <button type="button" disabled>Loading TurboDiffusion recipes...</button>
+          </div>
+        </div>
+
+        <div class="cookbook-selection-row">
+          <div class="cookbook-selection-row__label">
+            <strong>Runtime</strong>
+            <span>Maintained paths only</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--hardware" data-cookbook-hardware-options role="group" aria-label="Runtime">
+            <button type="button" disabled>Loading runtimes...</button>
+          </div>
+        </div>
+
+        <p class="cookbook-selection-description" data-cookbook-description>Loading recipe details...</p>
+        <p class="cookbook-hardware-note">Exact device and memory details appear only when a recorded run supports them.</p>
+
+        <div class="cookbook-hardware-state" data-cookbook-hardware-state role="status" aria-live="polite">
+          Reading recipe evidence...
+        </div>
+      </div>
+
+      <article class="cookbook-result">
+        <div class="cookbook-result__header">
+          <h3 data-cookbook-label>Loading...</h3>
+          <div class="cookbook-result__badges">
+            <span class="cookbook-badge">Maintained</span>
+            <span class="cookbook-badge" data-cookbook-evidence>Source-backed</span>
+            <span class="cookbook-badge cookbook-badge--neutral" data-cookbook-hardware-badge>Source config</span>
+          </div>
+        </div>
+
+        <dl class="cookbook-result__facts">
+          <div><dt>Model</dt><dd data-cookbook-model>Loading...</dd></div>
+          <div><dt>Workload</dt><dd data-cookbook-task>Loading...</dd></div>
+          <div><dt>Source configuration</dt><dd data-cookbook-gpus>Loading...</dd></div>
+          <div><dt>Expected output</dt><dd data-cookbook-artifact>Loading...</dd></div>
+        </dl>
+
+        <div class="cookbook-command">
+          <div class="cookbook-command__bar">
+            <span>Terminal</span>
+          </div>
+          <pre><code class="language-bash" data-cookbook-command>Loading...</code></pre>
+        </div>
+
+        <div class="cookbook-result__footer">
+          <a data-cookbook-source href="../../inference/examples/basic/">Open example source</a>
+          <a data-cookbook-model-link href="https://huggingface.co/loayrashid">View model card</a>
+        </div>
+        <p class="cookbook-picker__status" role="status" aria-live="polite" data-cookbook-status></p>
+      </article>
+    </div>
+
+    <noscript>
+      <div class="cookbook-noscript">
+        JavaScript is needed for the guided selector. You can still browse the
+        <a href="../../inference/examples/examples_inference_index/">maintained inference examples</a>.
+      </div>
+    </noscript>
+  </section>
+</div>
+
+<details class="cookbook-collapsible" id="cookbook-setup">
+  <summary>Setup</summary>
+  <div class="cookbook-collapsible__body">
+      <p>The generated commands expect a local clone:</p>
+      <pre><code>git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo</code></pre>
+      <p>Use <a href="../../inference/configuration/">Configuration</a> for supported Python and CLI settings, <a href="../../inference/optimizations/">Optimizations</a> for attention and memory tradeoffs, and the <a href="../../inference/support_matrix/">support matrix</a> for the supported model and optimization surface.</p>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>TurboDiffusion paths load community-published <code>loayrashid/TurboWan*</code> checkpoints; availability is governed by those repos.</li>
+        <li>The SLA attention backend used by the I2V recipe is selected inside the example source; do not combine it with another <code>FASTVIDEO_ATTENTION_BACKEND</code> override in the same shell.</li>
+      </ul>
+  </div>
+</details>
+
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>All recipes on this page are <strong>Source-backed</strong>: their commands, model IDs, and flags were validated against the checked-in FastVideo sources listed above (static validation). No runtime GPU validation is recorded for these recipes, so GPU model fit, memory use, throughput, and runtime duration are <strong>Unknown</strong> and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.</p>
+  </div>
+</details>

docs/cookbook/wan.md

@@ -5,7 +5,7 @@ hide:
 
 # Wan recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="wan" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="wan" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">
@@ -122,17 +122,6 @@ hide:
         </div>
 
         <p class="cookbook-selection-description" data-cookbook-description>Loading recipe details...</p>
-        <div class="cookbook-selection-row" data-cookbook-usage>
-          <div class="cookbook-selection-row__label">
-            <strong>Workflow</strong>
-            <span>Both can run locally</span>
-          </div>
-          <div class="cookbook-option-grid cookbook-option-grid--hardware" role="group" aria-label="How to run this recipe">
-            <button type="button" data-cookbook-mode="server" aria-pressed="false"><strong>Run a server</strong><span data-cookbook-server-hint>Playground, cURL, or an API client</span></button>
-            <button type="button" data-cookbook-mode="python" aria-pressed="false"><strong>Use Python directly</strong><span>Call the model in your own process</span></button>
-          </div>
-        </div>
-        <p class="cookbook-hardware-note" data-cookbook-serving-availability></p>
         <p class="cookbook-hardware-note">Exact device and memory details appear only when a recorded run supports them.</p>
 
         <div class="cookbook-hardware-state" data-cookbook-hardware-state role="status" aria-live="polite">
@@ -161,45 +150,7 @@ hide:
           <div class="cookbook-command__bar">
             <span>Terminal</span>
           </div>
-          <pre id="cookbook-local-command"><code class="language-bash" data-cookbook-command>Loading...</code></pre>
-        </div>
-        <p class="cookbook-hardware-note" data-cookbook-python-note>Running this script again starts a new process and reloads the model. To iterate in Python, create the generator once and reuse it for multiple prompts.</p>
-
-        <div class="cookbook-serving" data-cookbook-serving hidden>
-          <p class="cookbook-serving__intro" data-cookbook-server-lifetime>Start once, then change prompts in the playground or your app. You can run the server and clients on the same machine.</p>
-          <section class="cookbook-serving__step" aria-labelledby="serving-install-heading">
-            <h4 id="serving-install-heading"><span aria-hidden="true">1</span> Prepare the machine</h4>
-            <p>Run from your FastVideo clone in an activated Python environment. See <a data-cookbook-install-guide href="../../getting_started/installation/gpu/">installation requirements</a>.</p>
-            <div class="cookbook-command"><div class="cookbook-command__bar"><span>GPU machine · Terminal</span></div><pre id="cookbook-server-install"><code class="language-bash" data-cookbook-server-install></code></pre></div>
-            <details class="cookbook-serving__prepare" data-cookbook-prepare hidden><summary>Download and convert MLX weights once</summary><p>Skip this if the weights are already prepared. Edit the paths in the serving config to use your existing files.</p><div class="cookbook-command"><pre id="cookbook-server-prepare"><code class="language-bash" data-cookbook-server-prepare></code></pre></div></details>
-          </section>
-          <section class="cookbook-serving__step" aria-labelledby="serving-start-heading">
-            <h4 id="serving-start-heading"><span aria-hidden="true">2</span> Start the server</h4>
-            <p>Keep this terminal running while you use <span data-cookbook-playground-only>the playground or </span>API clients.</p>
-            <div class="cookbook-command"><div class="cookbook-command__bar"><span>GPU machine · Terminal</span></div><pre id="cookbook-server-command"><code class="language-bash" data-cookbook-server-command></code></pre></div>
-            <details class="cookbook-serving__check"><summary>Check that the server is ready</summary><p>In another terminal, this returns <code>{"status":"ok"}</code> after startup.</p><div class="cookbook-command"><pre id="cookbook-health-command"><code class="language-bash" data-cookbook-health-command></code></pre></div></details>
-          </section>
-          <section class="cookbook-serving__step" aria-labelledby="serving-client-heading">
-            <h4 id="serving-client-heading"><span aria-hidden="true">3</span> Generate and download a video</h4>
-            <div class="cookbook-serving__playground" data-cookbook-playground-only>
-              <div><strong>Try prompts in your browser</strong><p>Edit a prompt, generate, and watch the result. The playground uses the same server as cURL and your app.</p></div>
-              <a class="cookbook-serving__launch" data-cookbook-playground href="http://127.0.0.1:8000/playground/" target="_blank" rel="noopener">Open playground <span aria-hidden="true">↗</span></a>
-            </div>
-            <p class="cookbook-serving__local-hint"><span data-cookbook-playground-only>Open after the server is ready. </span>On a remote GPU machine, <a href="../openai-api/#connect-your-app">forward port 8000</a> to your computer first.<span data-cookbook-playground-only> This opens a local page, not a hosted demo.</span></p>
-            <details class="cookbook-serving__code"><summary>Use cURL or an SDK</summary>
-            <p>Each example submits a job, checks its status, and saves the MP4. The Python and JavaScript examples use OpenAI-compatible clients; no OpenAI account is needed.</p>
-            <div class="cookbook-serving__clients" role="group" aria-label="API client language">
-              <button type="button" data-cookbook-client="curl" aria-pressed="false">cURL</button>
-              <button type="button" data-cookbook-client="python" aria-pressed="true">Python</button>
-              <button type="button" data-cookbook-client="javascript" aria-pressed="false">JavaScript</button>
-            </div>
-            <div class="cookbook-command"><div class="cookbook-command__bar"><span>Client dependencies</span></div><pre id="cookbook-client-install"><code class="language-bash" data-cookbook-client-install></code></pre></div>
-            <div class="cookbook-command cookbook-command--client"><div class="cookbook-command__bar"><span data-cookbook-client-filename>video.py</span><a data-cookbook-client-source href="https://github.com/hao-ai-lab/FastVideo/tree/main/examples/serving/clients">View source</a></div><pre id="cookbook-client-code"><code data-cookbook-client-code></code></pre></div>
-            <p data-cookbook-client-run></p>
-            </details>
-          </section>
-          <p class="cookbook-serving__boundary">This is a local development server without built-in API-key authentication. The client key <code>local</code> is a placeholder. Keep the server on loopback; use an authenticated TLS proxy before exposing it publicly. Run the JavaScript client in your webapp's backend, not in a browser with a private key.</p>
-          <a href="../openai-api/">Server guide and API compatibility →</a>
+          <pre><code class="language-bash" data-cookbook-command>Loading...</code></pre>
         </div>
 
         <div class="cookbook-result__footer">

docs/cookbook/z-image.md

@@ -3,19 +3,19 @@ hide:
 - toc
 ---
 
-# Kandinsky 6 recipes
+# Z-Image recipes
 
-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="kandinsky6" data-recipes="../../assets/cookbook-recipes.json?v=15">
+<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="zimage" data-recipes="../../assets/cookbook-recipes.json?v=11">
   <header class="cookbook-family-header">
     <a class="cookbook-back-link" href="../"><span aria-hidden="true">←</span> All model families</a>
     <div class="cookbook-family-header__body">
       <span class="cookbook-family-header__logo">
-        <img class="off-glb" src="../../assets/logos/kandinsky.webp" alt="Kandinsky Lab" width="112" height="112">
+        <img class="off-glb" src="../../assets/logos/tongyi.webp" alt="Tongyi MAI" width="112" height="112">
       </span>
       <div>
         <p class="cookbook-eyebrow">Maintained family · Inference</p>
-        <h2>Kandinsky 6 inference recipes</h2>
-        <p>Kandinsky 6 from the Kandinsky Lab generates five-second video with synchronized audio from text or an image, with base and distilled pi-Flow checkpoints, and upscales existing clips with base or distilled video super-resolution.</p>
+        <h2>Z-Image inference recipes</h2>
+        <p>Z-Image Turbo from Tongyi MAI is a fast text-to-image model. The maintained example runs it on a single GPU.</p>
       </div>
     </div>
     <div class="cookbook-lifecycle" aria-label="Lifecycle stages">
@@ -49,7 +49,7 @@ hide:
             <span>Task and checkpoint</span>
           </div>
           <div class="cookbook-option-grid cookbook-option-grid--models" data-cookbook-model-options role="group" aria-label="Recipe">
-            <button type="button" disabled>Loading Kandinsky 6 recipes...</button>
+            <button type="button" disabled>Loading Z-Image recipes...</button>
           </div>
         </div>
 
@@ -97,7 +97,7 @@ hide:
 
         <div class="cookbook-result__footer">
           <a data-cookbook-source href="../../inference/examples/basic/">Open example source</a>
-          <a data-cookbook-model-link href="https://huggingface.co/kandinskylab">View model card</a>
+          <a data-cookbook-model-link href="https://huggingface.co/Tongyi-MAI">View model card</a>
         </div>
         <p class="cookbook-picker__status" role="status" aria-live="polite" data-cookbook-status></p>
       </article>
@@ -119,7 +119,6 @@ hide:
       <pre><code>git clone https://github.com/hao-ai-lab/FastVideo.git
 cd FastVideo</code></pre>
       <p>Use <a href="../../inference/configuration/">Configuration</a> for supported Python and CLI settings, <a href="../../inference/optimizations/">Optimizations</a> for attention and memory tradeoffs, and the <a href="../../inference/support_matrix/">support matrix</a> for the supported model and optimization surface.</p>
-      <p>For model-specific behavior, see <a href="../../inference/kandinsky6/">Kandinsky 6 video with audio</a> and <a href="../../inference/kandinsky6_sr/">Kandinsky 6 video super-resolution</a>.</p>
   </div>
 </details>
 
@@ -127,10 +126,7 @@ cd FastVideo</code></pre>
   <summary>Troubleshooting</summary>
   <div class="cookbook-collapsible__body">
       <ul>
-        <li>Text and image conditioning use the same TI2VA pipeline and example. Set <code>IMAGE_PATH</code> in the script to condition on an image; both base and distilled outputs include synchronized audio.</li>
-        <li>For pi-Flow, set <code>KANDINSKY6_MODEL_PATH</code> to the distilled model ID when running <code>basic_kandinsky6_ti2va.py</code>. Its registered preset uses 10 inference steps, guidance 1.0, <code>eps=1e-6</code>, <code>final_step_size_scale=0.5</code>, and <code>num_policy_substeps=128</code>.</li>
-        <li>For VSR, set <code>INPUT_VIDEO</code> in the generated command. The pipeline accepts x2, x2.25 and x4 scales, processes up to 121 frames at 24 fps, and preserves source audio.</li>
-        <li>Checkpoint key-layout errors indicate that the checkpoint and FastVideo checkout target different Kandinsky 6 Diffusers revisions.</li>
+        <li>Output defaults to <code>outputs/zimage/zimage_turbo.png</code>; pass <code>--output</code> to redirect.</li>
         <li>Gated or missing checkpoints: run <code>huggingface-cli login</code> and confirm you accepted the model's license on Hugging Face.</li>
       </ul>
   </div>

docs/design/inference_schema_parity_inventory.yaml

@@ -205,6 +205,9 @@ surfaces:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig
           - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      color_correction_strength:
+        sources:
+          - fastvideo.configs.pipelines.dreamx_world.DreamXWorld5BARPipelineConfig
       default_camera_rotation:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig
@@ -279,6 +282,11 @@ surfaces:
         sources:
           - fastvideo.configs.pipelines.lingbotworld.LingBotWorldI2V480PConfig
           - fastvideo.configs.pipelines.lingbotworld.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_1_3B_Config
           - fastvideo.configs.pipelines.wan.FastWan2_1_T2V_480P_Config
           - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
           - fastvideo.configs.pipelines.matrixgame2.MatrixGame2BaseI2V480PConfig
@@ -297,6 +305,11 @@ surfaces:
         sources:
           - fastvideo.configs.pipelines.lingbotworld.LingBotWorldI2V480PConfig
           - fastvideo.configs.pipelines.lingbotworld.Wan2_2_I2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionI2V_A14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2VConfig
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_14B_Config
+          - fastvideo.configs.pipelines.turbodiffusion.TurboDiffusionT2V_1_3B_Config
           - fastvideo.configs.pipelines.wan.FastWan2_1_T2V_480P_Config
           - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
           - fastvideo.configs.pipelines.matrixgame2.MatrixGame2BaseI2V480PConfig
@@ -311,21 +324,48 @@ surfaces:
           - fastvideo.configs.pipelines.wan.WanI2V720PConfig
           - fastvideo.configs.pipelines.wan.WanT2V480PConfig
           - fastvideo.configs.pipelines.wan.WanT2V720PConfig
+      bsa_cdf_threshold:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_chunk_k:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_chunk_q:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_params:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      bsa_sparsity:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enable_bsa:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enable_kv_cache:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      enhance_hf:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      offload_kv_cache:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      t_thresh:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+      use_distill:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
       scheduler_arch:
         sources:
           - fastvideo.configs.pipelines.sd35.SD35Config
+          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
       text_encoder_archs:
         sources:
           - fastvideo.configs.pipelines.sd35.SD35Config
+          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
       tokenizer_archs:
         sources:
           - fastvideo.configs.pipelines.sd35.SD35Config
+          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
       transformer_arch:
         sources:
           - fastvideo.configs.pipelines.sd35.SD35Config
+          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
       vae_arch:
         sources:
           - fastvideo.configs.pipelines.sd35.SD35Config
+          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
       expand_timesteps:
         sources:
           - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
@@ -418,14 +458,6 @@ surfaces:
         sources: [fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanBaseConfig]
       video_txt_guidance_scale:
         sources: [fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanBaseConfig]
-      piflow_eps:
-        sources: [fastvideo.configs.pipelines.kandinsky6.Kandinsky6TI2VAConfig]
-      piflow_final_step_size_scale:
-        sources: [fastvideo.configs.pipelines.kandinsky6.Kandinsky6TI2VAConfig]
-      piflow_num_policy_substeps:
-        sources: [fastvideo.configs.pipelines.kandinsky6.Kandinsky6TI2VAConfig]
-      sample_fps:
-        sources: [fastvideo.configs.pipelines.kandinsky6.Kandinsky6TI2VAConfig]
     compatibility_only:
       batch_size: "Gen3C inference-only tuning field pending typed batching design."
       gradient_checkpointing: "Gen3C inference-only compatibility field pending typed batching design."
@@ -438,17 +470,15 @@ surfaces:
       coords_style: "MagiHuman internal data-proxy coordinate convention."
       frame_receptive_field: "MagiHuman internal data-proxy receptive-field setting."
       image_conditioning: "MagiHuman preset variant marker for reference-image conditioning."
-      pdd_step_indices: "MiniMax-H3 PDD fine-grid partition; resolve_checkpoint_settings sets it from fastvideo_inference.json."
       ref_audio_offset: "MagiHuman internal data-proxy audio alignment offset."
+      scheduler_sigma_min: "Z-Image scheduler parity invariant; not part of the public typed inference API."
+      scheduler_use_reference_discrete_timesteps: "Z-Image scheduler parity invariant; not part of the public typed inference API."
       sr_local_attn_layers: "MagiHuman SR internal sparse-attention layer selection."
       text_offset: "MagiHuman internal data-proxy text alignment offset."
       vae_stride: "MagiHuman internal VAE/data-proxy stride setting."
-      vsa_ref_keep_rate: "MiniMax-H3 PDD reference-video VSA keep rate; defaults to the checkpoint's trained value."
       z_dim: "MagiHuman internal VAE latent channel setting."
       vocoder_config: "Legacy internal component config object."
       vocoder_precision: "Precision override pending dedicated component precision design."
-      audio_sample_rate: "Kandinsky6 audio-latent length constant (audio VAE sample rate); tied to the checkpoint, not a request knob."
-      audio_downsample_factor: "Kandinsky6 audio-latent length constant (audio VAE downsample factor); tied to the checkpoint, not a request knob."
 
   sampling_param_base:
     moved:
@@ -465,6 +495,8 @@ surfaces:
       pose: request.inputs.pose
       c2ws_plucker_emb: request.inputs.c2ws_plucker_emb
       action_path: request.inputs.action_path
+      refine_from: request.inputs.refine_from
+      stage1_video: request.inputs.stage1_video
       prompt: request.prompt
       negative_prompt: request.negative_prompt
       prompt_path: request.inputs.prompt_path
@@ -484,6 +516,8 @@ surfaces:
       guidance_scale: request.sampling.guidance_scale
       batch_cfg: request.sampling.batch_cfg
       guidance_scale_2: request.sampling.guidance_scale_2
+      cfg_normalization: request.sampling.cfg_normalization
+      cfg_truncation: request.sampling.cfg_truncation
       guidance_rescale: request.sampling.guidance_rescale
       use_embedded_guidance: request.sampling.use_embedded_guidance
       true_cfg_scale: request.sampling.true_cfg_scale
@@ -498,12 +532,19 @@ surfaces:
       return_continuation_state: request.output.return_state
     preset_owned:
       t_thresh: request.stage_overrides.refine.t_thresh
+      spatial_refine_only: request.stage_overrides.refine.spatial_refine_only
       num_cond_frames: request.stage_overrides.refine.num_cond_frames
       trajectory_type: request.extensions.gen3c.trajectory_type
       movement_distance: request.extensions.gen3c.movement_distance
       camera_rotation: request.extensions.gen3c.camera_rotation
       prompt_attention_mask: request.extensions.hyworld.prompt_attention_mask
       negative_attention_mask: request.extensions.hyworld.negative_attention_mask
+      camera_states: request.extensions.hunyuangamecraft.camera_states
+      camera_trajectory: request.extensions.hunyuangamecraft.camera_trajectory
+      action_list: request.extensions.hunyuangamecraft.action_list
+      action_speed_list: request.extensions.hunyuangamecraft.action_speed_list
+      gt_latents: request.extensions.hunyuangamecraft.gt_latents
+      conditioning_mask: request.extensions.hunyuangamecraft.conditioning_mask
       ltx2_cfg_scale_video: request.extensions.ltx2.cfg_scale_video
       ltx2_cfg_scale_audio: request.extensions.ltx2.cfg_scale_audio
       ltx2_modality_scale_video: request.extensions.ltx2.modality_scale_video

docs/generate_examples.py

@@ -38,14 +38,17 @@ COOKBOOK_SOURCE_ROOTS = (
 # (e.g. black-forest-labs/FLUX.1-dev) and are grouped for documentation only.
 COOKBOOK_FAMILIES = {
     "wan",
+    "turbodiffusion",
     "ltx2",
     "hunyuan",
     "cosmos",
     "kandinsky5",
-    "kandinsky6",
     "flux",
+    "glm_image",
+    "zimage",
     "sd35",
     "minimax_h3",
+    "longcat",
     "stable_audio",
     "mmaudio",
     "matrixgame",
@@ -290,16 +293,12 @@ def cookbook_serving_profile(recipe: dict) -> dict:
     for language, (filename, install) in COOKBOOK_CLIENTS.items():
         path = ROOT_DIR / "examples/serving/clients" / filename
         text = path.read_text(encoding="utf-8")
-        # Keep displayed snippets identical to the executable sources, with
-        # only endpoint and alias substitutions.
+        # Keep displayed snippets identical to executable sources for the
+        # checked-in H3 profile, with only endpoint and alias substitutions.
         text = text.replace("http://127.0.0.1:8000/v1", base_url)
         text = text.replace('"fasth3"', json.dumps(model)).replace("${FASTVIDEO_MODEL:-fasth3}",
                                                                    "${FASTVIDEO_MODEL:-" + model + "}")
         clients[language] = {"source": path.relative_to(ROOT_DIR).as_posix(), "code": text, "install": install}
-    # The playground router only serves H3 servers (see require_h3 in
-    # fastvideo/entrypoints/openai/playground.py), so other families must not link to it.
-    has_playground = recipe["family"] == "minimax_h3"
-    compile_enabled = ((generator.get("engine") or {}).get("compile") or {}).get("enabled")
     return {
         "source": serving["source"],
         "install": serving["install"],
@@ -308,10 +307,7 @@ def cookbook_serving_profile(recipe: dict) -> dict:
         "prepare": serving.get("prepare", ""),
         "model": model,
         "base_url": base_url,
-        "playground_url": f"http://127.0.0.1:{port}/playground/" if has_playground else None,
-        "audio": bool(serving.get("audio")),
-        # None when the config leaves compilation at its default.
-        "compile_enabled": compile_enabled,
+        "playground_url": f"http://127.0.0.1:{port}/playground/",
         "health_command": f"curl --fail-with-body http://127.0.0.1:{port}/health",
         "hardware": hardware,
         "sampling": config["default_request"]["sampling"],

docs/getting_started/installation/mlx.md

@@ -46,91 +46,6 @@ is the higher-quality FastH3.
 Recorded shapes and evidence live in the
 [support matrix](../../inference/support_matrix.md#apple-silicon-native-runtime).
 
-## FastH3 V2 and Trim with INT6
-
-The released sources are `FastVideo/FastVideo-FastH3-8-Step-V2` and
-`FastVideo/FastVideo-FastH3-Trim-8-Step`. Trim has 42 transformer blocks and
-rank-16 AdaLN. Both use eight denoising forwards, video/audio shifts of 10/3,
-VSA sparsity 0.8, a native NVFP4 text encoder, and the 26-layer light video VAE.
-Keep `fastvideo_inference.json` beside `transformer/`; conversion reads its
-schedule to build the AdaLN cache.
-
-Convert the BF16 transformer to affine INT6 with its VSA gates:
-
-```bash
-hf download FastVideo/FastVideo-FastH3-Trim-8-Step \
-  --local-dir ./FastH3-Trim
-
-python scripts/checkpoint_conversion/convert_minimax_h3_mlx.py \
-  --model-root ./FastH3-Trim/transformer \
-  --out ./FastH3-Trim-MLX \
-  --formats "int6" --include-vsa
-```
-
-For V2, use `FastVideo/FastVideo-FastH3-8-Step-V2` and separate source/output
-directories. Preconverted release snapshots use the same names with the
-`-MLX-INT6` suffix. Each snapshot includes the encoder in MLX layout, both VAEs,
-and the trained schedule, so it does not require a second encoder download.
-
-The 36 GiB M4 Max release recipe uses phased placement. It loads the encoder,
-DiT, and decoders in turn. Use reference attention and native output geometry:
-
-```python
-from pathlib import Path
-from fastvideo.mlx_runtime.minimax_h3_pipeline import MiniMaxH3MLXPipeline
-
-root = Path("./FastH3-Trim")
-pipeline = MiniMaxH3MLXPipeline(
-    model_root=root,
-    mlx_dit_checkpoint="./FastH3-Trim-MLX/int6",
-    conditioner_mode="nvfp4",
-    resident=False,
-    vae_dtype="fp16",
-    metal_wired_limit_gib=27,
-)
-try:
-    pipeline.generate(
-        "A corgi news anchor sits behind a desk and gives a cheerful bark.",
-        output_path="./outputs/trim-int6-corgi.mp4",
-        width=832, height=480, num_frames=124, seed=1234,
-        num_steps=8, vsa=True, vsa_sparsity=0.8, vsa_tile_size=64,
-        vsa_impl="reference", vae_tile_height=256, vae_tile_width=256,
-    )
-finally:
-    pipeline.close()
-```
-
-Set `FASTVIDEO_MLX_DQ_GEMM=1` before running this Python command. It selects the
-validated affine dequantization followed by dense matrix multiplication.
-124 frames at 24 fps is roughly five seconds. The recipe preserves all frames
-and the requested resolution.
-
-### Native NVFP4 encoder and cache
-
-The MLX conditioner reads the released packed NVFP4 weights without
-requantization. It retains the layers H3 reads and can cache the packed weights
-in MLX layout. The cache is written in a staging directory and published by a
-single rename. A cache hit changes storage layout, not encoder arithmetic.
-
-MLX uses BF16 embeddings and FP32 activations; CUDA uses quantized activations.
-Generated video and audio must be reviewed before claiming cross-runtime
-quality parity. MLX 0.32.2 supports the required operator on Apple Silicon.
-
-### Metal wired memory
-
-MLX's allocation limit and wired-memory limit are separate. The optional
-`metal_wired_limit_gib` calls `mx.set_wired_limit` to keep selected Metal
-allocations in physical memory. It does not add RAM. An explicit request fails
-if the installed MLX build cannot apply it. `close()` restores the previous
-wired limit. The phased pipeline also sets a 30 GiB maximum allocator guideline
-and restores its previous value on close. Resident placement keeps the existing
-allocator limit so larger Macs can hold all components.
-
-The tested 36 GiB M4 Max recipe uses 27 GiB and phased placement. Resident
-placement also requires capacity for all components and peak activations;
-wiring cannot make an oversized stack fit. Inspect `mx.device_info()` before
-choosing a limit on another Mac.
-
 ## Hardware
 
 - FastMetal 1.3B and 5B: 16 GB unified memory and up

docs/getting_started/installation/spark.md

@@ -144,9 +144,6 @@ for which models are practical on the GB10, what makes them faster, and what
 won't help on this hardware (and why) — so you don't spend a night tuning knobs
 that can't move here.
 
-For the eight-forward FastH3 V2 NVFP4 stack with a trimmed encoder and light
-VAE, use the [one-Spark resident recipe](spark_performance.md#fasth3-v2-nvfp4-on-one-spark).
-
 Two Sparks with QSFP cables: [Pair two NVIDIA DGX Sparks](spark_pair.md) for
 one FastH3 clip across both GPUs (`sp_size=2` over Ray). Copy-paste commands
 for one or two Sparks also live on the

docs/getting_started/installation/spark_pair.md

@@ -149,29 +149,6 @@ fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair.yam
 
 Stop the cluster when you are done: `ray stop` on both nodes.
 
-## Released V2 and Trim NVFP4 stacks
-
-The public eight-forward V2 and Trim stacks include the NVFP4 encoder and
-lightweight video VAE. Use the environment above plus the release kernel
-settings, then run one of these configs from the head:
-
-```bash
-export FASTVIDEO_MINIMAX_H3_FUSIONS=all FASTVIDEO_NVFP4_MM_BACKEND=cutlass
-export FASTVIDEO_H3_VAE_TILE_BATCH=1 FASTVIDEO_VSA_TRITON=1
-export FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0
-export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1
-fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair_v2_nvfp4.yaml
-# Or basic_fasth3_spark_pair_pruned_nvfp4.yaml for FastH3 Trim.
-```
-
-Both configs use their released Hugging Face model paths, `h3_dit_vsa`,
-832x480, 124 frames, seed 1234, VSA 0.8 and 64-token tiles. All components
-stay resident; lazy/sequential loading and compilation are disabled for
-these compact stacks. The older BF16/Preview memory guidance below applies
-to those larger stacks. Set 1344x768 for native 768p with the same 124 frames.
-See [the resident release recipe](spark_performance.md#fasth3-v2-nvfp4-on-one-spark)
-for checkpoint contents and the timing protocol.
-
 ## FastH3 frame counts
 
 H3 is 24 fps. Legal `num_frames` values are `17n+5`. The pipeline rejects

docs/getting_started/installation/spark_performance.md

@@ -161,16 +161,14 @@ is power-cycled. To avoid it:
   on: "CPU" offload uses the same unified RAM. Multi-GPU FSDP sharding remains
   available because it partitions weights without parking them in a separate
   host pool.
-- **Older MiniMax H3 / FastH3 bf16 weights** need deferred loading on one GB10.
-  The full Qwen3-VL conditioner is tens of gigabytes of BF16. If the DiT and
-  VAEs load while that encoder is still resident, the process can be killed by
-  `earlyoom`. On unified memory, `lazy_module_load` auto-enables and owns that
+- **MiniMax H3 / FastH3** still needs deferred loading on one GB10. The Qwen3-VL
+  conditioner is tens of gigabytes of BF16. If the DiT and VAEs load while that
+  encoder is still resident, the process is a typical `earlyoom` kill (Python is
+  preferred). On unified memory, `lazy_module_load` auto-enables and owns that
   split (encoder, then DiT, then VAE; DiT can drop before decode). Sequential
-  load is the H3-only fallback when lazy is off. Keep deferred loading for
-  those older checkpoints. The trimmed NVFP4 encoder and light VAE in the
-  [V2 resident recipe](#fasth3-v2-nvfp4-on-one-spark) are a different memory
-  profile. Geometry scalars come from checkpoint `config.json`, not live
-  weights. See [Offloading](../../inference/offloading.md).
+  load is the H3-only fallback when lazy is off; do not pass
+  `--no-lazy-module-load` here. Geometry scalars come from checkpoint
+  `config.json`, not live weights. See [Offloading](../../inference/offloading.md).
 - **FastH3 TAEH3** (`--video-decode-backend taeh3`) is an opt-in preview decoder.
   T2VA never materializes the 9.7 GiB video VAE (DiT still loads after Qwen via
   sequential start). On this box, alpine 768×1344×124 decoded in **2.4 s** versus
@@ -209,121 +207,6 @@ A few things that surprise people on this box (beyond the memory notes above):
   `Released MiniMax-H3 text encoder after conditioning` before
   `Loading MiniMax-H3 denoise modules`).
 
-## FastH3 V2 NVFP4 on one Spark
-
-This recipe uses the full V2 eight-forward transformer, the 50-layer NVFP4
-Qwen3-VL encoder, and the light H3 video VAE. Its configuration keeps all
-three resident on one GB10. This stack fits in the Spark's unified memory;
-benchmark your installed runtime and review the clips before publishing a
-speed claim. The earlier bf16 H3 memory guidance above concerns a larger checkpoint.
-
-Install FastVideo following [the Spark install guide](spark.md). The released
-repositories are complete inference stacks:
-
-| Model | Repository | Packed DiT profile |
-|---|---|---|
-| FastH3 V2 | `FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer` | `h3_dit_vsa` |
-| FastH3 Trim | `FastVideo/FastVideo-FastH3-Trim-8-Step-NVFP4` | `h3_dit_vsa` |
-
-Each ships its own trained schedule, NVFP4 transformer and 50-layer NVFP4
-text encoder, lightweight 26-layer video VAE with the INT8-weight overlay,
-and audio VAE. Download the complete repository; the runtime selects these
-components from its model index. The released Trim transformer also packs
-attention and VSA gates, unlike the earlier FFN-only pruned export. Neither
-release requires local checkpoint conversion or a separate encoder download.
-
-```bash
-hf download FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer
-hf download FastVideo/FastVideo-FastH3-Trim-8-Step-NVFP4
-```
-
-Keep `fastvideo_inference.json` with the transformer if you stage the stack
-in a local directory. It declares the trained eight-forward ladder and
-video/audio shifts of 10/3. The recipes use `num_inference_steps: 9` for
-nine sigma points and eight DiT forwards.
-
-On GB10 with FlashInfer 0.6.18, FastVideo fences activation quantization before
-releasing its padded input. Without this completion fence, identical H3
-requests produced different DiT latents and occasionally corrupt video.
-The fence applies to `sm_121`; other architectures retain asynchronous execution.
-
-Run `examples/inference/basic/basic_fasth3_spark_v2_nvfp4.yaml` from the
-repository root. It uses 832x480, 124 frames and seed 1234, VSA sparsity 0.8 with
-64-token tiles, and the light H3 VAE through the `h3-vae` decode backend.
-It does not use frame dropping or spatial upscaling.
-
-```bash
-FASTVIDEO_MINIMAX_H3_FUSIONS=all \
-FASTVIDEO_NVFP4_MM_BACKEND=cutlass \
-FASTVIDEO_H3_VAE_TILE_BATCH=1 \
-FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 \
-FASTVIDEO_STAGE_LOGGING=1 \
-nice -n 19 fastvideo generate \
-  --config examples/inference/basic/basic_fasth3_spark_v2_nvfp4.yaml
-```
-
-The defaults produce a roughly five-second clip at 24 fps. For native 768p,
-set `--request.sampling.width 1344` and `--request.sampling.height 768`,
-keeping 124 frames. For the separate ten-second setting, use 243 frames.
-H3 permits frame counts of `17n+5`; 124 is the nearest legal count above
-five seconds and 243 is the nearest above ten seconds.
-
-For Trim, use `basic_fasth3_spark_pruned_nvfp4.yaml` with the same environment.
-Both recipes keep the encoder, DiT and VAEs resident, disable compilation,
-and use `h3_dit_vsa`. On two Sparks, use the corresponding
-`basic_fasth3_spark_pair_{pruned,v2}_nvfp4.yaml` after following the
-[pair setup guide](spark_pair.md). The pair uses SP2/TP1 and parallel VAE
-gathering, with tile batch 1 on each worker.
-
-For release timing, create one generator per model/resolution. Run one
-untimed ceramics warmup, then two timed ceramics calls and two timed harbor
-calls in that same process, using the exact release prompt strings, seed
-1234 and the settings above. Measure each `generate()` call through finished
-MP4 output and report the median of the two timed calls per prompt. Keep the
-warmup excluded. Record the model revision, code commit, command, environment
-and peak-memory scope with the results. Review every clip's video and audio
-before publishing a quality or speed claim.
-
-The benchmark helper requires a local stack so it can validate the schedule
-before loading. For example:
-
-```bash
-hf download FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer \
-  --local-dir ./FastH3-V2-Consumer
-python examples/inference/basic/benchmark_fasth3_spark_nvfp4.py \
-  --config examples/inference/basic/basic_fasth3_spark_v2_nvfp4.yaml \
-  --model-path ./FastH3-V2-Consumer --frames 124 \
-  --prompts /path/to/benchmark_prompts.json \
-  --output-dir outputs/fasth3_spark_v2_nvfp4/benchmark-124
-```
-
-Set the environment from the generation command above before benchmarking.
-The prompt JSON must contain `latency-ceramics-005` and
-`latency-harbor-005`. Pass `--width 1344 --height 768` for the native 768p
-protocol. Use the Trim repository and config for its corresponding run.
-
-### Released model measurements
-
-The released Trim stack at revision `cae9ceb6feefe77d34a56640782cda3909363f19`
-completed the native 832x480, 124-frame protocol on one Spark, seed 1234.
-The tested code is `6ccdbc761e6b854003f472e39b826c06cb54de60`, using the
-resident recipe above. Medians exclude warmups and cover finished MP4 output.
-
-| Released model | Resolution | One Spark, ceramics / harbor | Two Sparks |
-|---|---|---:|---|
-| FastH3 Trim NVFP4 | 832x480 | 124.090 / 127.519 s | Pending |
-| FastH3 Trim NVFP4 | 1344x768 | Review pending | Pending |
-| FastH3 V2 NVFP4-Consumer | 832x480 | Repeatability review pending | Pending |
-| FastH3 V2 NVFP4-Consumer | 1344x768 | Review pending | Pending |
-
-The completed Trim 480p batch used one extra untimed harbor warmup in the
-same process, six calls total. Both warmups are excluded from the medians.
-Each prompt's three decoded videos and audio streams match exactly, and
-sampled frames are coherent. These checks do not establish BF16 parity,
-speech accuracy or lip sync. Pending cells are not measured substitutes
-from older checkpoints.
-
 ## Reproduce these numbers
 
 Two scripts under `examples/inference/optimizations/` reproduce the claims on

docs/inference/architecture.md

@@ -232,7 +232,8 @@ Dataclass carrying all pipeline state between stages. Key field groups:
 - **Scheduler**: `timesteps`, `num_inference_steps`, `guidance_scale`,
   `sigmas`.
 - **Task-specific**: `mouse_cond`/`keyboard_cond` (Matrix-Game 2.0), `pose`
-  (HYWorld), `c2ws_plucker_emb` (LingBotWorld).
+  (HYWorld), `camera_states` (GameCraft), `c2ws_plucker_emb`
+  (LingBotWorld).
 - **Output**: `output: Tensor | None`.
 - **Logging**: `logging_info: PipelineLoggingInfo`.
 
@@ -255,6 +256,7 @@ Standard stages (typical execution order):
 | `DecodingStage` | `stages/decoding.py` | Decodes latents to video via VAE |
 
 Specialized variants: `CausalDenoisingStage`, `LTX2DenoisingStage`,
+`LongCatDenoisingStage`, `GameCraftDenoisingStage`,
 `HYWorldDenoisingStage`, `MatrixGame2CausalDenoisingStage`,
 `SRDenoisingStage`, `LTX2AudioDecodingStage`, `SD35ConditioningStage`,
 `LTX2TextEncodingStage`, `LTX2LatentPreparationStage`.

docs/inference/fasth3-distilled.md

@@ -74,130 +74,6 @@ This documents execution support for the published checkpoint. It is not a
 quality claim: compare video/audio output against base MiniMax-H3 on your own
 prompts before adopting it.
 
-## Ref2VA PDD students
-
-A Parallel Decoding Distillation (PDD) student widens the transformer's two
-output projections to `pdd_steps` heads, one per interval of a fixed fine
-time grid on `[0, 0.999]`. Each transformer forward fuses a block of
-consecutive heads into their integration-weighted mean, and an ordinary Euler
-step over the block's two node sigmas applies it. The FastH3 OmniRef PDD-8
-student is a Ref2VA (`transformer_ref`) student with 32 heads, sampled in
-eight blocks of four.
-
-Its export carries only what distillation changed:
-
-| Path | Contents |
-| --- | --- |
-| `transformer_ref/` | Widened `proj_out` and `audio_proj_out`, trained VSA compression gates; `config.json` records `pdd_steps` |
-| `scheduler/`, `audio_scheduler/` | Video and audio shifts (12 and 3) |
-| `fastvideo_inference.json` | The sampling contract below |
-| `modular_model_index.json` | The Diffusers manifest |
-
-The text encoder, tokenizer, processor, and both VAEs are base MiniMax-H3's.
-`basic_fasth3_omniref_pdd.py` composes the two into one local directory of
-symlinks and runs the recipe the contract records. `--model-path` is the
-export, as a local directory or a Hugging Face repo id. The base comes from
-the revision the contract pins in `base_model_revision`, or from
-`--base-model-path`:
-
-```bash
-python examples/inference/basic/basic_fasth3_omniref_pdd.py \
-  --model-path <local export directory or Hugging Face repo id> \
-  --video reference.mp4 --image character.png \
-  --prompt 'The dancer from the video performs the routine in the pictured outfit.' \
-  --height 480 --width 832 --num-frames 124 \
-  --output outputs/fasth3-omniref-pdd
-```
-
-References are ordered: pass `--image`, `--video`, and `--audio` in the order
-the prompt refers to them. At least one image or video is required.
-
-A PDD export uses `fasth3-inference-contract-v1` with PDD fields in place of
-the DMD ladder:
-
-```json
-{
-  "schema_version": "fasth3-inference-contract-v1",
-  "model_type": "ref2va",
-  "transformer_component": "transformer_ref",
-  "pdd_steps": 32,
-  "pdd_step_indices": [0, 4, 8, 12, 16, 20, 24, 28, 32],
-  "num_inference_steps": 8,
-  "transformer_forwards": 8,
-  "grid_max_t": 0.999,
-  "video_scheduler_shift": 12.0,
-  "audio_scheduler_shift": 3.0,
-  "guidance_scale": 1.0,
-  "attention_backend": "VIDEO_SPARSE_ATTN_H3",
-  "vsa_sparsity": 0.9,
-  "vsa_tile_size": 128,
-  "vsa_ref_policy": "p2_multi_region",
-  "vsa_ref_keep_rate": 0.1
-}
-```
-
-The export also records `schema`, `conditioning`, and `base_model_revision`,
-the base snapshot it was distilled against, as `hf://<repo id>@<revision>`
-(for example `hf://MiniMaxAI/MiniMax-H3@<commit>`); any other form is an
-error. FastVideo reads the file once, when the run's arguments are built and
-before any weights load. The file must carry exactly these fields: a missing
-or unknown field is an error. Fields that repeat a value stored elsewhere must
-equal it: `pdd_steps` equals `transformer_ref/config.json`, the shifts equal
-the scheduler configs, and `num_inference_steps` and `transformer_forwards`
-equal the block count of `pdd_step_indices`, a strictly increasing partition
-of the grid from 0 to `pdd_steps`. Only `MiniMaxH3Ref2VAModularPipeline` runs
-the export. A `transformer_ref` whose `config.json` sets `pdd_steps` without
-this file is an error.
-
-Each sampling setting has one source, the file:
-
-| Setting                         | Run leaves it unset    | Run sets a different value                     |
-| ------------------------------- | ---------------------- | ---------------------------------------------- |
-| `pdd_step_indices`              | Taken from the file    | Error                                          |
-| `num_inference_steps` (request) | Set to the block count | Error                                          |
-| `attention_backend`             | Taken from the file    | Error, including `FASTVIDEO_ATTENTION_BACKEND` |
-| `VSA_tile_size`                 | Taken from the file    | Error                                          |
-| `VSA_sparsity`                  | Taken from the file    | The run's value, with a warning                |
-| `vsa_ref_keep_rate`             | Taken from the file    | The run's value, with a warning                |
-
-A request leaves `num_inference_steps` unset only when it is parsed from a
-mapping or a config file. A `GenerationRequest` built in Python counts every
-field as set, so it must pass the block count. The trained compression gates
-of `transformer_ref` load only under `VIDEO_SPARSE_ATTN_H3`, so the attention
-backend cannot change. `dmd_denoising_steps` must stay unset.
-
-A MiniMax-H3 checkpoint without PDD fields, such as a DMD export, whose
-transformer carries VSA compression gates (`to_gate_compress` weights) also
-runs only with `VIDEO_SPARSE_ATTN_H3`. The pipeline reads the shard index (or
-the safetensors headers) and rejects any other backend, including automatic
-selection, before it loads any component.
-
-### Reference-video sparsity
-
-Every PDD contract sets `"vsa_ref_policy": "p2_multi_region"`, the only value FastVideo accepts, which means:
-`VIDEO_SPARSE_ATTN_H3` tiles every reference video as its own sparse region,
-in place in the packed sequence. Each video query keeps `vsa_ref_keep_rate` of
-every reference video's tiles and `1 - VSA_sparsity` of the target video's
-tiles. Text, audio, and image references stay dense. With `VSA_sparsity` 0,
-every region is dense and the reference keep rate has no effect. Checkpoints
-other than PDD students keep every conditioning row dense.
-
-### Hardware
-
-The contract's 128-token tiles, `(4, 4, 8)`, run only on the sm_100a/sm_103a
-CUDA block-sparse kernel (B200, B300, GB200, GB300) of a fastvideo-kernel
-build with the Blackwell VSA extension. There is no Triton fallback for tile
-128: the backend raises instead. Tile 128 runs eagerly. Regional compile
-requires 64-token tiles. With `--num-gpus` above 1 the example shards the DiT
-across the GPUs (FSDP) and splits the sequence across them.
-
-On GB200, a 480x832, 124-frame request fits on one GPU: the eight forwards
-take about 54 s, and device memory in use peaks near 103 GiB. A 768x1344,
-345-frame request with `--num-gpus 4` takes about 36 s for the eight
-forwards. With the DiT sharded, peak allocated memory is about 66 GiB per GPU
-(about 95 GiB in use), against 83 GiB (103 GiB) with a full DiT copy on each
-GPU; the output is identical. Model loading and decoding come on top of this.
-
 ## Apple Silicon
 
 `mlx_fasth3.py` stays on FastH3 V1 and its uniform AdaLN cache.

docs/inference/fasth3_rtx_pro_6000.md

@@ -1,260 +0,0 @@
-# FastH3 NVFP4 on RTX PRO 6000 (sm_120)
-
-This page covers serving the FastH3 NVFP4 checkpoints on one RTX PRO 6000
-Blackwell (sm_120, 96 GB) with every component resident: the NVFP4 text
-encoder, the NVFP4 denoiser and an INT8 light VAE. It also records what was
-measured and tried along the way. Every switch is opt-in and defaults to the
-existing behavior.
-
-## Results
-
-All results are for one RTX PRO 6000 (Modal), a 10.1 s clip at 1344x768
-(243 frames, 73.6k packed tokens), the light INT8 VAE and warm runs.
-
-| Checkpoint | Denoise | Video decode | End to end | Peak memory |
-| --- | ---: | ---: | ---: | ---: |
-| [`FastH3-4-step-Preview-v1-VSA-DataFree-NVFP4`](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree-NVFP4) (4 forwards, VSA 0.9) | 30.1–31.0 s | 9.0 s | **41.2 / 42.4 s** | 79.6 GB |
-| [`FastH3-8-Step-V2-NVFP4`](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4) (8 forwards, VSA 0.8) | 75.6 s | 9.0 s | **86.5 s** | — |
-
-These end-to-end runs predate the warp-skip kernel change below, which cuts
-sparse attention by a further 1.5x, so they are upper bounds. Conditioning
-takes about 0.12 s, frame post-processing plus MP4 writing about 0.9 s, and
-the first clip at a new shape about 200 s (VAE compile).
-
-At 480p (124 frames, 15.1k tokens) V2 8-step denoises in 12.5 s on a cold
-run, down from 16.8 s warm before this work.
-
-## Usage
-
-### 1. Convert the checkpoint
-
-The published NVFP4 checkpoints use ModelOpt's unified Hugging Face layout.
-`convert_minimax_h3_modelopt_nvfp4_dit.py` repacks it into FastVideo's packed
-export, `transformer/nvfp4_weights.safetensors`. The calibrated FFN weights
-and scales are carried over bit for bit and only the scale bytes are
-swizzled. Optionally it also quantizes the BF16 attention projections and the
-VSA compression gates:
-
-```bash
-python scripts/checkpoint_conversion/convert_minimax_h3_modelopt_nvfp4_dit.py \
-  --src /models/FastH3-8-Step-V2-NVFP4/transformer \
-  --dst /models/fasth3-v2-fv/transformer \
-  --quantize-attention --quantize-gate
-```
-
-Every exported linear is probed through the loader's `mm_fp4` path against a
-BF16 matmul with the dequantized weight. Conversion refuses to write when any
-error exceeds `--max-probe-error` (default 0.3; the converted V2 and 4-step
-checkpoints probe at 0.14). The rest of the model folder (text encoder,
-VAEs, schedulers, `fastvideo_inference.json`) is used as-is; the
-NVFP4 text encoder comes from
-`convert_minimax_h3_text_encoder_nvfp4.py`.
-
-The flags select the `layer_profile` to load the result with:
-
-| Flags | Linears in NVFP4 | `layer_profile` |
-| --- | --- | --- |
-| *(none)* | FFN `fc_in`/`fc_out` | `h3_dit_ffn` |
-| `--quantize-attention` | + attention `to_{q,k,v,out}` | `h3_dit` |
-| `--quantize-attention --quantize-gate` | + VSA `to_gate_compress` | `h3_dit_vsa` |
-
-### 2. Generate
-
-```python
-import os
-os.environ.update({
-    "FASTVIDEO_H3_VSA_FP4": "1",              # sparse FP4 attention
-    "FASTVIDEO_MINIMAX_H3_FUSIONS": "all",    # Triton norm/modulate/RoPE/SwiGLU fusions
-    "FASTVIDEO_NVFP4_MM_BACKEND": "cutlass",  # see "FP4 GEMM backend" below
-    "FASTVIDEO_H3_VAE_TILE_BATCH": "28",      # one decoder call per 1344x768 tile grid
-})
-from fastvideo import VideoGenerator
-
-generator = VideoGenerator.from_config({
-    "model_path": "/models/fasth3-v2-fv",
-    "engine": {
-        "num_gpus": 1,
-        "quantization": {"transformer_quant": "NVFP4", "layer_profile": "h3_dit_vsa"},
-        "compile": {"enabled": False, "vae_enabled": True},
-    },
-    "pipeline": {"experimental": {"attention_backend": "VIDEO_SPARSE_ATTN_H3",
-                                  "VSA_sparsity": 0.8, "VSA_tile_size": 64}},
-})
-generator.generate({"prompt": "...", "sampling": {"height": 768, "width": 1344, "num_frames": 243,
-                                                   "num_inference_steps": 9}})
-```
-
-Use `VSA_sparsity` 0.9 and `num_inference_steps` 5 for the 4-step checkpoint
-(see its `fastvideo_inference.json`). Frame counts must be `17n + 5`: 243
-frames is the closest to 10 s.
-
-## What changed
-
-### Block-sparse FP4 attention for VSA tiles (`fastvideo-kernel`)
-
-SageAttention3's sm_120 FP4 kernel (`attn_qat_infer`) gains a block-sparse
-forward, `fwd_sparse`, exposed as `sageattn_blackwell_sparse` (head-major
-inputs) and `sageattn_blackwell_sparse_bshd` (sequence-major inputs, quantized
-in place without a transpose). `vsa_tile_mask_to_fp4_blocks` turns a VSA tile
-mask into the kernel's lists:
-
-- **Block lists.** Query block `m` visits only the 128-token KV blocks in
-  `q2k_idx[b, h, m, :q2k_num[b, h, m]]`.
-- **Quadrant masks for 64-token tiles.** The kernel computes on 128x128
-  blocks, but V2 and the 4-step preview use 64-token VSA tiles. Each listed
-  block carries a 4-bit `q2k_quad` (one bit per 64x64 quadrant). The kernel
-  masks unselected quadrants to `-inf`, so the result is exactly VSA's tile-64
-  semantics.
-- **Valid counts per 64-column half.** `kv_valid` gives the valid tokens in
-  each 64-column half, so partially filled tiles can pad mid-block.
-- **Warp-level skipping.** Each MMA warp owns 16 query rows and so sits
-  inside one 64-row half. A warp skips a listed block that its half did not
-  select, and the P·V chunk of a key half it did not select. Masked scores
-  contribute exactly zero, and the warp sharing its tensor-core partition runs
-  faster meanwhile. This recovers most of the work that pairing 64-token tiles
-  into 128-token blocks adds.
-- **First-visited block.** Lists run in descending block order because the
-  kernel visits them last entry first. Block 0 (the first prefix tile, which
-  VSA-H3's exempt mode gives every query) is therefore visited first, so every
-  row starts from a finite running max. `validate=True` checks this. The model
-  integration uses only exempt mode.
-
-The dense and sparse entry points also stop allocating `delta_s`. With Q
-smoothing off (the default), each call used to allocate and zero a
-`[B, H, L/128, L]` fp32 tensor: 9.5 GB at 73k tokens. Its int32 batch stride
-also overflowed the TMA descriptor ("Failed to initialize the TMA descriptor
-1", then an illegal instruction), so FP4 attention could not run 10 s 1344x768
-clips at all. A cached `[B, H, 1, L]` zero row read with `per_block_mean=False`
-replaces it, and the outputs are bit-identical.
-
-Correctness (`fastvideo-kernel/tests/test_attn_qat_infer_sparse.py`, RTX PRO
-6000):
-
-| Layout | Error vs token-masked fp32 | Dense FP4 floor |
-| --- | ---: | ---: |
-| 64-token tiles, odd count, partial tiles | 0.190 | 0.191 |
-| 64-token tiles, even count | 0.191 | 0.193 |
-| 256-token tiles, partial tile | 0.195 | 0.196 |
-
-Errors are relative L2. The sparse kernel sits exactly at the FP4 noise
-floor; random Gaussian inputs make that floor large. Full block lists
-reproduce the dense kernel bit for bit.
-
-### Model integration (`FASTVIDEO_H3_VSA_FP4=1`)
-
-`fastvideo/models/dits/minimax_h3_vsa_fp4.py` replaces only the attention core
-of `MiniMaxH3Attention`. VSA-H3's tile pooling, top-k mask, exempt prefix and
-gated compression branch are unchanged. Per block, it:
-
-1. Gathers the attention input into tile order once (one `hidden_size`-wide
-   pass). Pad rows stay zero, so the q/k/v pad rows are exactly zero through
-   the bias-free projections, RMSNorm and RoPE.
-2. Quantizes that input once and shares it between `to_q`, `to_k` and `to_v`.
-   NVFP4 activations use a unit global scale, so this is exact.
-3. Applies QK-norm and RoPE with tile-ordered `cos`/`sin`, computed once per
-   step.
-4. Runs the sparse FP4 kernel on sequence-major tensors and gathers the output
-   back to packed order before `to_out`.
-
-This replaces the generic path's concat, four tile scatters and three
-transposes. The route applies only to no-grad, non-compiled, single
-sequence-parallel-rank calls in exempt mode; everything else keeps the
-existing path.
-
-Two smaller pieces ship alongside it:
-
-- **Packed gate check.** With `--quantize-gate`, `to_gate_compress` loses its
-  BF16 weight, so the gate-activity check reads the packed E2M1 bytes instead.
-- **Compiled residual.** With `FASTVIDEO_MINIMAX_H3_FUSIONS` enabled, each
-  block's final `hidden + gate[indices] * ffn_out` runs as one compiled op
-  instead of materializing the gathered gate.
-
-### FP4 GEMM backend (`FASTVIDEO_NVFP4_MM_BACKEND`)
-
-FlashInfer's `mm_fp4(backend="auto")` picks a kernel about 2x slower than
-`cutlass` or `cudnn` on sm_120 once activations reach tens of thousands of
-rows. At 15k rows all three match.
-
-| Linear | 73.6k rows: `auto` | 73.6k rows: `cutlass` | 73.6k rows: `cudnn` | 15.1k rows: `auto` |
-| --- | ---: | ---: | ---: | ---: |
-| `to_q` (5376→7168) | 7.96 ms | 3.96 ms | 4.24 ms | 0.88 ms |
-| `to_out` (7168→5376) | 9.07 ms | 4.09 ms | 4.27 ms | 0.92 ms |
-| `fc_in` (5376→28672) | 21.54 ms | 16.59 ms | 16.17 ms | 3.01 ms |
-| `fc_out` (14336→5376) | 18.06 ms | 8.09 ms | 8.43 ms | 1.65 ms |
-
-### Batched VAE tile decode (`FASTVIDEO_H3_VAE_TILE_BATCH`)
-
-The H3 video VAE decodes 256-pixel spatial tiles one at a time. A 1344x768
-clip is a 4x7 grid per temporal chunk, so a 10 s clip is roughly 400 small
-decoder calls. The ViT decoder treats batch entries independently, so
-`FASTVIDEO_H3_VAE_TILE_BATCH=N` decodes up to `N` equal-shaped tiles per call;
-28 covers a full 1344x768 grid. The decoded tiles are the same as per-tile
-decoding.
-
-## Per-block measurements
-
-One H3 transformer block (hidden 5376, 56 heads, FFN 14336), RTX PRO 6000:
-
-| Component | 480p, 124 f (15.1k tokens) | 768p, 243 f (73.6k tokens) |
-| --- | ---: | ---: |
-| VSA Triton BF16 attention (kernel + pooling/mask) | 14.8 ms | 180.1 ms |
-| Tile scatter of q/k/v/gate + gather (generic path) | 2.7 ms | 12.7 ms |
-| Dense FP4 attention (SageAttention3) | 10.5 ms | 211.9 ms |
-| Sparse FP4, quadrant masks | 7.9 ms | 123.4 ms |
-| Sparse FP4, quadrant masks + warp skip | — | **83.0 ms** (VSA 0.8) / **49.3 ms** (VSA 0.9) |
-| Dense BF16 SDPA | 18.1 ms | — |
-| Modulation: eager / fused / compiled | 4.07 / 1.42 / 0.58 ms | 20.3 / 7.0 / 2.8 ms |
-| SwiGLU: eager / fused | 1.53 / 0.88 ms | 7.37 / 4.24 ms |
-| QK-norm + RoPE: eager / fused | 4.64 / 1.12 ms | 22.4 / 5.2 ms |
-
-Before this work, a 480p block cost about 40 ms: 17.7 ms of attention, 12 ms
-of linears and 10 ms of eager elementwise ops. Over 50 blocks that is 2.0 s
-per step, which matches the measured 2.1 s.
-
-Block density each kernel granularity computes at 768p (fraction of the
-dense attention). "Selected" is what VSA needs; the other columns are what
-each block shape computes:
-
-| VSA sparsity | Selected (64x64) | 128x128 blocks | 64-row x 128 | 128 x 64-col |
-| --- | ---: | ---: | ---: | ---: |
-| 0.8 | 0.222 | 0.434 | 0.317 | 0.313 |
-| 0.9 | 0.125 | 0.254 | 0.180 | 0.177 |
-
-## What was tried and not shipped
-
-- **Dense FP4 attention for VSA-trained students.** `ATTN_QAT_INFER` does not
-  build `to_gate_compress`. A VSA-distilled checkpoint such as V2 carries
-  trained gates, so the strict loader refuses it ("Parameter
-  ...to_gate_compress.weight not found"). Dense FP4 attention is also slower
-  than sparse FP4 at 768p (212 vs 83 ms per block).
-- **Multi-GPU (Ulysses) FP8 exchange.** On 8x RTX PRO 6000 (PCIe only), NCCL
-  all-to-all moves about 21 GB/s per GPU, with NCCL P2P on or off. A BF16
-  q/k/v/gate exchange at 73.6k tokens therefore costs 24 ms per block, and the
-  attention output another 6.6 ms; with FP4 payloads q/k/v/gate drop to 7.2 ms.
-  The branch `h3-sm120-experimental` keeps a sequence-parallel path that:
-  - sends q/k/v as FP8 with one scale per token and head;
-  - never sends the VSA gate, applying it on each rank after a small
-    all-gather of the per-tile compression output;
-  - returns the attention output as FP8.
-
-  It is estimated at about 18–20 s per 10 s clip for V2 8-step on 8 GPUs. It
-  has not executed yet (8-GPU capacity was unavailable), so it is not part of
-  this change. The same branch holds the Modal drivers behind every number on
-  this page.
-- **64-row query blocks.** The kernel traits allow `kBlockM = 64`, which would
-  remove the query-side pairing waste. Warp-level skipping recovers most of
-  that waste without a second kernel instantiation, so it was not built.
-
-## Known limitations
-
-- **End-to-end quality.** The kernel matches a masked reference at the FP4
-  noise floor. Generated videos have not yet been A/B-compared against the
-  BF16 Triton VSA path on the H3 audio/video metrics.
-- **Activation scales.** The packed export drops ModelOpt's calibrated static
-  `input_scale` and quantizes activations with a unit global scale and dynamic
-  per-16 block scales, as FastVideo's NVFP4 linears do elsewhere.
-- **Decode cost.** Video decode (9 s at 10 s/768p with the light INT8 VAE) is
-  the next largest cost after denoising.
-- **Hardware.** Everything here targets sm_120. The GeForce RTX 5090 shares
-  the architecture but has 32 GB, which needs a reduced-AdaLN checkpoint and a
-  non-resident text encoder at this resolution.

docs/inference/kandinsky6.md

@@ -1,107 +0,0 @@
-# Kandinsky 6 Text/Image to Video with Audio (T2IVA)
-
-`Kandinsky6TI2VAPipeline` generates a video and a synchronized audio track from a text prompt, optionally conditioned
-on an image: one pipeline serves both, so pass `image_path` to condition on an image and leave it out for text only.
-The audio is decoded by the checkpoint's audio VAE (mel decoder plus vocoder in one component) and muxed into the saved
-mp4 automatically. To upscale a generated clip, see [Kandinsky 6 Video SR](kandinsky6_sr.md).
-
-## Models
-
-Both variants are official Diffusers repos, loaded directly through their `model_index.json`:
-
-| Variant | Hub repo | Scheduler | Steps | Guidance | Example |
-|---|---|---|---|---|---|
-| T2IVA | `kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers` | `FlowMatchEulerDiscreteScheduler` (shift 5.0) | 50 | 5.0 | [`basic_kandinsky6_ti2va.py`](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky6_ti2va.py) |
-| T2IVA distilled | `kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers` | `PiflowScheduler` (`n_grid` 10, shift 5.0) | 10 | 1.0 | [`basic_kandinsky6_ti2va.py`](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky6_ti2va.py) |
-
-The steps and guidance columns are the defaults of the preset the registry selects for each repo id. Everything else is
-shared: 512x768, 121 frames (5 s at 24 fps) and the Diffusers default negative prompt (only used when
-`guidance_scale > 1`). The distilled preset is named `kandinsky6_ti2va_distilled`.
-
-## Usage
-
-```bash
-python examples/inference/basic/basic_kandinsky6_ti2va.py
-KANDINSKY6_MODEL_PATH=kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers \
-  python examples/inference/basic/basic_kandinsky6_ti2va.py
-```
-
-Set `IMAGE_PATH` in the script to condition on an image.
-
-```python
-from fastvideo import VideoGenerator
-
-generator = VideoGenerator.from_pretrained(
-    "kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers",
-    num_gpus=1,
-    dit_cpu_offload=False,
-    text_encoder_cpu_offload=True,
-)
-generator.generate_video(
-    "cinematic shot: a giant stone samurai on a stormy cliff above a neon city opens glowing golden eyes and "
-    "raises a katana. Blue lightning strikes the blade, creating a massive shockwave through the clouds. The "
-    "camera rapidly pulls back from a low angle. Photorealistic, epic scale, dark blue and gold lighting, rain, "
-    "sparks, volumetric lightning, blockbuster quality. Audio: heavy rain, deep thunder, metallic sword hum, "
-    "rising brass and choir, electrical crackle, perfectly synchronized lightning impact, sub-bass shockwave. "
-    "No dialogue, text, or logos.",
-    image_path=None,  # or the path of a conditioning image
-    output_path="video_samples_kandinsky6_ti2va",
-    height=512,
-    width=768,
-    num_frames=121,
-)
-```
-
-A local copy of a repo works the same way. A local directory is treated as the distilled variant only when its name is
-a Kandinsky-6 name containing `distill` (for example `Kandinsky-6.0-Pro-distill-5s-Diffusers`); any other directory name
-selects the base preset (see below).
-
-## Distilled (pi-Flow) checkpoint
-
-- The distilled repo replaces the flow-matching scheduler with `PiflowScheduler` (`n_grid` 10, `eps` 1e-6,
-  `final_step_size_scale` 0.5, `num_policy_substeps` 128). Its DiT emits `n_grid` predictions per latent channel
-  (`out_visual_dim` 160 = 16 x 10, `out_audio_dim` 400 = 40 x 10).
-- pi-Flow runs without classifier-free guidance: `guidance_scale` must be exactly `1.0`. The official Diffusers
-  pipeline rejects any other guidance for a `PiflowScheduler` too; FastVideo raises a `ValueError` naming the
-  required value. `num_inference_steps` is not constrained by the checkpoint -- `PiflowScheduler.set_timesteps`
-  accepts any step count and ignores the scheduler's `nfe`; `10` is only the value the distilled checkpoint was
-  trained for and the `kandinsky6_ti2va_distilled` preset's default.
-- The `kandinsky6_ti2va_distilled` preset (10 steps, guidance 1.0) is the default for the distilled repo id and for
-  local directories named like it. Set `KANDINSKY6_MODEL_PATH` to either one when running the shared
-  `basic_kandinsky6_ti2va.py` example. A distilled copy under another directory name selects the base preset, so
-  retain `Kandinsky-6.0-Pro-distill-5s-Diffusers` as the final directory name.
-- The policy values can be overridden on `Kandinsky6TI2VAConfig` (`piflow_eps`, `piflow_final_step_size_scale`,
-  `piflow_num_policy_substeps`); `None` keeps the values from `scheduler_config.json`.
-
-## Differences from the Diffusers pipeline
-
-A few Diffusers pipeline options are not ported, and are surfaced here instead of as a knob that would silently do
-nothing:
-
-- `sample_audio=False` (video-only, no audio stream) is not exposed; FastVideo's DiT raises `NotImplementedError` for
-  a partial-modality call instead of denoising video alone.
-- `expand_prompts` (the Qwen prompt-beautifier pass) is not ported; FastVideo's `PromptEnhancerConfig` is a separate,
-  external (Cerebras/Groq streaming) feature, not this pipeline's built-in expansion.
-- Of the Diffusers reference's `visual_cond_scheme` values, only `tail_cond_first_frame` (append one clean reference
-  frame to the end of the sequence) is implemented; `pretrain` and plain `i2v` are not.
-- Precomputed `prompt_embeds`/`negative_prompt_embeds` are not accepted; every call encodes its own prompt text.
-- MagCache is not ported: a `magcache` block in a checkpoint's `transformer/config.json` is parsed and dropped
-  by `update_model_arch`, not read automatically or exposed as an opt-in cache config.
-- RNG differs: FastVideo draws video then audio noise from one per-request CPU generator seeded by `seed` (default
-  1024); Diffusers seeds a device generator from a value drawn out of `generator` (audio uses `seed+1`). The same
-  numeric seed produces different noise on the two stacks -- pass `latents`/`audio_latents` directly for bit-level
-  comparisons.
-- Qwen prompt tokens are unpadded and carry no attention mask (Diffusers pads to a fixed length and masks the
-  padding in text self/cross-attention). Mathematically equivalent for a single prompt (measured 2e-7 relative
-  difference), but FastVideo has no attention-mask plumbing, so a hand-built batch of unequal-length prompts is not
-  supported.
-- Attention is dense (`LocalAttention`, flash/SDPA); NABLA sparse attention is wired but unverified against a real
-  NABLA-flagged checkpoint. This matches the Diffusers pipeline itself, which never enables NABLA either.
-- VAE tiling is on by default (`vae_tiling=True`); Diffusers decodes untiled.
-
-## Memory
-
-The Pro DiT has 30.1B parameters, about 60 GB in bf16 (`dit_precision` defaults to `bf16`), and the Qwen2.5-VL text
-encoder adds 16.6 GB. FastVideo enables `dit_cpu_offload` by default; the examples turn it off (`dit_cpu_offload=False`)
-to keep the DiT resident on the GPU and offload the text encoder instead (`text_encoder_cpu_offload=True`). See
-[Offloading](offloading.md) for the memory knobs.

docs/inference/kandinsky6_sr.md

@@ -1,83 +0,0 @@
-# Kandinsky 6 Video Super-Resolution
-
-`Kandinsky6SRPipeline` upscales a low-resolution video by x2, x2.25 or x4 (video-to-video, no text prompt). The clip is
-encoded once with the SR VAE (KVAE); its latent is cut into overlapping tiles, each tile is enlarged by the *latent
-upscaler*, refined by a text-free SR DiT in a few denoising steps and decoded, and the tiles are blended back together.
-The source audio is kept. The clip can come from anywhere, for example from the [Kandinsky 6 T2IVA](kandinsky6.md)
-pipeline.
-
-## Models
-
-| Variant | Hub repo | Scheduler | Default steps per tile |
-|---|---|---|---|
-| Flow matching | `kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers` | `FlowMatchEulerDiscreteScheduler` (shift 5.0) | 4 |
-| Distilled | `kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers` | `PiflowScheduler` (shift 3.5, `n_grid` 10) | 2 |
-
-Both repos have the same layout and share `vae/` and `latent_upscaler/`:
-
-```text
-model_index.json      _class_name = Kandinsky6SRPipeline
-transformer/          Kandinsky6SRTransformer3DModel (sr_params: trained base resolution, RoPE scale, noise level)
-vae/                  Kandinsky6SRVAE
-latent_upscaler/      Kandinsky6SRLatentUpscalerBank (x2 and x4 models)
-scheduler/            FlowMatchEulerDiscreteScheduler or PiflowScheduler
-```
-
-The scheduler component drives the denoising loop, and each repo resolves to its own preset (4 or 2 steps). The
-distilled transformer's head holds `n_grid` predictions per latent channel, which `PiflowScheduler` integrates.
-
-## Usage
-
-```bash
-python examples/inference/basic/basic_kandinsky6_sr.py --video-path input.mp4 --scale 2.25
-python examples/inference/basic/basic_kandinsky6_sr.py --video-path input.mp4 \
-    --model-path kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers
-```
-
-```python
-from fastvideo import VideoGenerator
-
-generator = VideoGenerator.from_pretrained("kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers", num_gpus=1)
-result = generator.generate({
-    "inputs": {"video_path": "input.mp4"},
-    "output": {"output_path": "outputs_video/sr", "return_frames": False},
-    "extensions": {"sr_resolution_scale": 2.25, "sr_target_resolution": "fullhd"},
-})
-```
-
-The output geometry and frame rate follow the input clip; the request's `height` / `width` / `num_frames` are not used.
-Clips are resampled to 24 fps by a fixed stride and only the first 121 frames (5 s, `1 + 8k`-aligned) are processed.
-The source audio (mono, 44.1 kHz) is trimmed to the processed span and muxed into the output.
-
-## Request parameters
-
-The `sr_*` options belong only to Kandinsky6 SR. Pass them in `request.extensions` (as above), or under
-`request.stage_overrides.sr`. The SR pipeline reads them from `ForwardBatch.extra`; they are not fields of the
-shared `SamplingParam` or `ForwardBatch`. Other model families reject these options. Existing
-`generate_video(..., sr_resolution_scale=...)` calls remain supported for Kandinsky6 SR; replace direct
-`SamplingParam(sr_...=...)` construction with request extensions or these keyword arguments.
-For the config-based CLI, use dotted overrides such as `--request.extensions.sr_resolution_scale 4`, rather than
-shared `--sr-*` flags. The model-specific example above keeps its `--scale` and `--tiles-batch-size` flags.
-
-| Field | Default | Meaning |
-|---|---|---|
-| `num_inference_steps` | 4 / 2 | Denoising steps (DiT calls) per tile. The upstream Diffusers pipeline counts grid points instead (its 5 is 4 steps here). The distilled model was trained for 2; other values run with a warning. |
-| `sr_resolution_scale` | `2.25` | Total upscale: `2`, `4` or `2.25` (x1.125 pixel pre-upscale, then x2). |
-| `sr_tiles_batch_size` | `1` | Tiles denoised per DiT call (raise it only if memory allows). |
-| `sr_tile_min_overlap` | `0.20` | Minimum overlap between neighbouring tiles, as a fraction of the tile. |
-| `sr_target_resolution` | `None` | Downscale the result to `hd`, `fullhd`, `2k` or `WxH`. |
-| `sr_target_resize_mode` | `fit` | `fit` keeps the aspect ratio, `exact` uses the bucket dimensions. |
-| `seed` | `42` | Tile group *k* (of `sr_tiles_batch_size` tiles) is seeded with `seed + index of its first tile`. |
-
-Two Python-only inputs take raw tensors and are passed as keyword arguments of `generate_video()`:
-
-- `sr_lr_latent`: an unscaled KVAE latent `[T, C, H, W]` of the source, instead of `video_path` (skips decoding and
-  encoding the video; scale 2 or 4 only, since 2.25 needs the pixel pre-upscale).
-- `sr_audio` / `sr_audio_sample_rate`: a mono waveform in `[-1, 1]` to mux instead of the source's own audio.
-
-## Limitations
-
-- NABLA block-sparse attention, which both repos request for 512-pixel tiles, is not wired; the DiT runs dense attention
-  and logs a warning.
-- No sequence or tensor parallelism: the DiT runs on one GPU. Tiles are processed one group after another.
-- One clip per request.

docs/inference/mlx_taeh3.md

@@ -70,7 +70,7 @@ Run the numerical tests against a local TAEHV checkout containing the released
 weights:
 
 ```bash
-FASTVIDEO_TEST_TAEH3_REFERENCE_DIR=/path/to/taehv \
+TAEH3_REFERENCE_DIR=/path/to/taehv \
   python -m pytest fastvideo/tests/mlx/test_mlx_taeh3.py -q
 ```
 

docs/inference/offloading.md

@@ -168,8 +168,8 @@ and logs a warning if the flag is set.
 
 Deferral is opt-in per pipeline. Releasing a component and loading it again is
 only safe when nothing outside the loader has changed it, and two common habits
-break that without raising: mutating a component after load, such as enabling
-a sparse-attention mode in `initialize_pipeline`, and reading a component's attributes
+break that without raising: mutating a component after load, as LongCat does
+when it enables block-sparse attention, and reading a component's attributes
 while stages are built, as the shared denoising stage does to pick an attention
 backend. A pipeline therefore lists the components it has checked in
 `_lazy_module_names`, which is empty in the base class. MiniMax-H3 opts in. On

docs/inference/optimizations.md

@@ -198,11 +198,6 @@ The `attn_qat_infer` kernel hard-gates on **sm_120 (consumer Blackwell / RTX
 5090)**; on other GPUs the backend logs a notice and falls back to Flash
 Attention. See the [Attn-QAT paper](https://arxiv.org/abs/2603.00040).
 
-For VSA-distilled MiniMax-H3 students, the same kernel has a block-sparse
-forward that runs VSA's 64-token tile selection in FP4
-(`FASTVIDEO_H3_VSA_FP4=1`). See
-[FastH3 NVFP4 on RTX PRO 6000](fasth3_rtx_pro_6000.md).
-
 Enable both halves — attention via the env var, linear via `transformer_quant`:
 
 ```python

docs/inference/support_matrix.md

@@ -31,10 +31,14 @@ column links a runnable script in `examples/inference/basic/` where one exists.
 | cosmos | `nvidia/Cosmos-Predict2-2B-Video2World` | T2V | — |
 | cosmos25 | `KyleShao/Cosmos-Predict2.5-2B-Diffusers` | T2V | [basic_cosmos2_5_t2w.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_cosmos2_5_t2w.py) |
 | cosmos25 | `nvidia/Cosmos-Predict2.5-14B` | T2V | [basic_cosmos2_5_t2w.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_cosmos2_5_t2w.py) |
+| dreamx_world | `FastVideo/DreamX-World-5B-Cam-Diffusers` | I2V | [basic_dreamx_world.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dreamx_world.py) |
+| dreamx_world | `FastVideo/DreamX-World-5B-Diffusers` | I2V | [basic_dreamx_world.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dreamx_world.py) |
 | flux | `black-forest-labs/FLUX.1-dev` | T2I | [basic_flux_dev.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux_dev.py) |
 | flux2 | `black-forest-labs/FLUX.2-klein-4B`<br>`black-forest-labs/FLUX.2-klein-9B` | T2I | [basic_flux2_klein.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux2_klein.py) |
 | flux2 | `black-forest-labs/FLUX.2-dev` | T2I | [basic_flux2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux2.py) |
+| gamecraft | `FastVideo/HunyuanGameCraft-Diffusers` | I2V | [basic_gamecraft.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_gamecraft.py) |
 | gen3c | `FastVideo/GEN3C-Cosmos-7B-Diffusers` | T2V | [basic_gen3c.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_gen3c.py) |
+| glm_image | `zai-org/GLM-Image` | T2I | [basic_glm_image.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_glm_image.py) |
 | hunyuan | `hunyuanvideo-community/HunyuanVideo` | T2V | — |
 | hunyuan | `FastVideo/FastHunyuan-diffusers` | T2V | — |
 | hunyuan15 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | T2V | [basic_hy15.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15.py) |
@@ -50,13 +54,13 @@ column links a runnable script in `examples/inference/basic/` where one exists.
 | kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
 | kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
 | kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
-| kandinsky6 | `kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers`<br>`kandinskylab/Kandinsky-6.0-Pro-sft-5s-Diffusers` | T2V, I2V | [basic_kandinsky6_ti2va.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky6_ti2va.py) |
-| kandinsky6 | `kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers` | T2V, I2V | [basic_kandinsky6_ti2va.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky6_ti2va.py) |
-| kandinsky6_sr | `kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers`<br>`kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers` | — | [basic_kandinsky6_sr.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky6_sr.py) |
 | lingbot_video | `FastVideo/LingBot-Video-MoE-30B-A3B-Diffusers` | T2V | [basic_lingbot_video.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbot_video.py) |
 | lingbot_video | `FastVideo/LingBot-Video-Dense-1.3B-Diffusers` | T2V | [basic_lingbot_video.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbot_video.py) |
 | lingbotworld | `FastVideo/LingBot-World-Base-Cam-Diffusers` | I2V | [basic_lingbotworld_base_cam.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbotworld_base_cam.py) |
 | lingbotworld2 | `robbyant/lingbot-world-v2-14b-causal-fast` | I2V | [basic_lingbotworld2_causal_fast.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbotworld2_causal_fast.py) |
+| longcat | `FastVideo/LongCat-Video-T2V-Diffusers` | T2V | [basic_longcat_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_t2v.py) |
+| longcat | `FastVideo/LongCat-Video-I2V-Diffusers` | I2V | [basic_longcat_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_i2v.py) |
+| longcat | `FastVideo/LongCat-Video-VC-Diffusers` | — | [basic_longcat_vc.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_vc.py) |
 | ltx2 | `FastVideo/LTX2-Distilled-Diffusers`<br>`FastVideo/LTX2.3-Distilled-Diffusers`<br>`FastVideo/LTX-2.3-Distilled-Diffusers` | T2V | [basic_ltx2_distilled.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2_distilled.py) |
 | ltx2 | `Lightricks/LTX-2.3`<br>`FastVideo/LTX2.3-base`<br>`FastVideo/LTX2.3-Diffusers` | T2V | [basic_ltx2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2.py) |
 | ltx2 | `Lightricks/LTX-2`<br>`FastVideo/LTX2-base`<br>`FastVideo/LTX2-Diffusers` | T2V | [basic_ltx2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2.py) |
@@ -67,6 +71,9 @@ column links a runnable script in `examples/inference/basic/` where one exists.
 | sd35 | `stabilityai/stable-diffusion-3.5-medium` | T2I | [basic_sd35_t2i.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_sd35_t2i.py) |
 | stable_audio | `FastVideo/stable-audio-open-1.0-Diffusers` | T2V | [basic_stable_audio.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_stable_audio.py) |
 | stable_audio | `FastVideo/stable-audio-open-small-Diffusers` | T2V | [basic_stable_audio_small.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_stable_audio_small.py) |
+| turbodiffusion | `loayrashid/TurboWan2.1-T2V-1.3B-Diffusers` | T2V | [basic_turbodiffusion.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion.py) |
+| turbodiffusion | `loayrashid/TurboWan2.1-T2V-14B-Diffusers` | T2V | [basic_turbodiffusion_14b.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion_14b.py) |
+| turbodiffusion | `loayrashid/TurboWan2.2-I2V-A14B-Diffusers` | I2V | [basic_turbodiffusion_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion_i2v.py) |
 | wan | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | T2V | [basic.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic.py) |
 | wan | `Wan-AI/Wan2.1-T2V-14B-Diffusers`<br>`FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers` | T2V | — |
 | wan | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | I2V | — |
@@ -82,6 +89,7 @@ column links a runnable script in `examples/inference/basic/` where one exists.
 | wan | `wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers` | T2V | [basic_self_forcing_causal.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal.py) |
 | wan | `rand0nmr/SFWan2.2-T2V-A14B-Diffusers` | T2V | [basic_self_forcing_causal_wan2_2_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py) |
 | wan | `FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers` | I2V | [basic_self_forcing_causal_wan2_2_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py) |
+| zimage | `Tongyi-MAI/Z-Image-Turbo` | T2I | [basic_zimage.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_zimage.py) |
 
 **Note (stable_audio)**: the Stable Audio Open pipelines generate audio
 (`StableAudioT2AConfig` / `StableAudioOpenSmallConfig`); they are registered
@@ -93,15 +101,7 @@ to convert the official weights locally and set `MMAUDIO_MODEL_PATH`.
 
 **Note (MiniMax H3)**: T2VA, FL2VA, and Ref2VA all generate video with stereo
 audio. Use the Ref2VA example when passing ordered image, video, or audio
-references. Distilled Ref2VA PDD students (eight transformer forwards, with
-reference videos as sparse VSA regions) run through
-[basic_fasth3_omniref_pdd.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_pdd.py);
-see [FastH3 distilled checkpoint schedules](fasth3-distilled.md#ref2va-pdd-students).
-
-**Note (Kandinsky 6)**: the two `kandinsky6` IDs generate video with audio from
-text, optionally plus an image (see the [T2IVA guide](kandinsky6.md)); the
-`kandinsky6_sr` IDs upscale an existing video (see the
-[Video SR guide](kandinsky6_sr.md)).
+references.
 
 **Note (Wan-VACE)**: not currently supported — no VACE pipeline or registered
 model ID exists on `main`
@@ -152,25 +152,31 @@ optimizations: absence means **untested**, not incompatible.
   }
 </style>
 
-| Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn (Legacy Branch) | Sage Attn | VSA |
-|------------|---------------------|-------------|----------|-------------------|-----------|-----|
-| FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ |
-| FastWan2.2 TI2V 5B Full Attn | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ |
-| Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ |
-| Lucy Edit Dev 5B*** | `decart-ai/Lucy-Edit-Dev` | 480P | ⭕ | ⭕ | ⭕ | ⭕ |
-| Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
-| Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ |
-| HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
-| FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ |
-| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ |
-| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅ | ✅ | ⭕ |
-| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ |
-| Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ |
-| Matrix Game 2.0 Base Distilled | `FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ |
-| Matrix Game 2.0 GTA Distilled | `FastVideo/Matrix-Game-2.0-GTA-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ |
-| Matrix Game 2.0 TempleRun Distilled | `FastVideo/Matrix-Game-2.0-TempleRun-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ |
-| Matrix Game 3.0 Base Distilled | `FastVideo/Matrix-Game-3.0-Base-Distilled-Diffusers` | 720x1280 | ⭕ | ⭕ | ⭕ | ⭕ |
-| GEN3C Cosmos 7B | `FastVideo/GEN3C-Cosmos-7B-Diffusers` | 704px1280p | ❌ | ❌ | ❌ | ⭕ |
+| Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn (Legacy Branch) | Sage Attn | VSA | BSA |
+|------------|---------------------|-------------|----------|-------------------|-----------|-----|-----|
+| FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
+| FastWan2.2 TI2V 5B Full Attn | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
+| Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ | ⭕ |
+| DreamX-World 5B Cam | `FastVideo/DreamX-World-5B-Cam-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| DreamX-World 5B AR | `FastVideo/DreamX-World-5B-Diffusers` | 704px1280p | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Lucy Edit Dev 5B*** | `decart-ai/Lucy-Edit-Dev` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
+| Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
+| HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
+| FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| TurboWan2.1 T2V 1.3B | `loayrashid/TurboWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| TurboWan2.1 T2V 14B | `loayrashid/TurboWan2.1-T2V-14B-Diffusers` | 480P, 720P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| TurboWan2.2 I2V A14B | `loayrashid/TurboWan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| LongCat T2V 13.6B | `FastVideo/LongCat-Video-T2V-Diffusers` | 480P<br>720P | ❌ | ❌ | ❌ | ⭕ | ✅ |
+| Matrix Game 2.0 Base Distilled | `FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Matrix Game 2.0 GTA Distilled | `FastVideo/Matrix-Game-2.0-GTA-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Matrix Game 2.0 TempleRun Distilled | `FastVideo/Matrix-Game-2.0-TempleRun-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| Matrix Game 3.0 Base Distilled | `FastVideo/Matrix-Game-3.0-Base-Distilled-Diffusers` | 720x1280 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| GEN3C Cosmos 7B | `FastVideo/GEN3C-Cosmos-7B-Diffusers` | 704px1280p | ❌ | ❌ | ❌ | ⭕ | ⭕ |
 
 ## Apple Silicon native runtime
 
@@ -234,6 +240,11 @@ listed under [Special requirements](#special-requirements).
   https://github.com/hao-ai-lab/FastVideo/tree/sta_do_not_delete
 - STA currently requires Hopper GPUs (H100s).
 
+### TurboWan2.1 (TurboDiffusion)
+- Uses TurboDiffusionPipeline with RCM scheduler for 1-4 step generation
+- Requires SLA attention backend: `export FASTVIDEO_ATTENTION_BACKEND=SLA_ATTN`
+- Uses `guidance_scale=1.0` (no classifier-free guidance)
+
 ### Matrix Game 2.0
 - Image-to-video game world models with keyboard/mouse control input
 - Three variants available: Base (universal), GTA, and TempleRun

examples/inference/basic/README.md

@@ -167,24 +167,6 @@ launchers still use sparse attention at strength `0` and require FastVideo's
 tile-64 VSA kernel; the dense launcher selects FA4. Each launcher writes to its
 own variant directory by default so comparison outputs do not collide.
 
-### FastH3 OmniRef PDD (Ref2VA)
-
-[basic_fasth3_omniref_pdd.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_pdd.py)
-runs a Parallel Decoding Distillation Ref2VA student in eight transformer
-forwards. The export carries only its `transformer_ref`, scheduler configs,
-and `fastvideo_inference.json`; the script links them with the base
-MiniMax-H3 components into one local model directory. Its 128-token VSA tiles
-need the sm_100a/sm_103a kernel:
-
-```bash
-python examples/inference/basic/basic_fasth3_omniref_pdd.py \
-  --model-path <local export directory or Hugging Face repo id> \
-  --video reference.mp4 --image character.png --prompt "your prompt"
-```
-
-See [FastH3 distilled checkpoint schedules](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3-distilled.md#ref2va-pdd-students)
-for the contract and the reference-video sparsity policy.
-
 ## Basic Walkthrough
 
 All you need to generate videos using multi-gpus from state-of-the-art diffusion pipelines is the following few lines!

examples/inference/basic/basic_compacth3_rtx5090.yaml

@@ -1,49 +0,0 @@
-#   FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-#     FLASHINFER_CUDA_ARCH_LIST=12.0a FASTVIDEO_STAGE_LOGGING=1 \
-#     fastvideo generate --config examples/inference/basic/basic_compacth3_rtx5090.yaml
-generator:
-  model_path: ./CompactH3
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: true
-      vae: false
-      pin_cpu_memory: true
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    experimental:
-      attention_backend: ATTN_QAT_INFER
-      h3_sequential_load: true
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-request:
-  prompt: >-
-    A wide cinematic shot of an alpine meadow at sunrise, pale pink mountain
-    peaks above a blue valley filled with thin morning mist.
-  negative_prompt: ""
-  sampling:
-    seed: 2026
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 5
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/compacth3_rtx5090/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_compacth3_rtx_pro6000.yaml

@@ -1,49 +0,0 @@
-#   FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-#     FLASHINFER_CUDA_ARCH_LIST=12.0a FASTVIDEO_STAGE_LOGGING=1 \
-#     fastvideo generate --config examples/inference/basic/basic_compacth3_rtx_pro6000.yaml
-generator:
-  model_path: ./CompactH3
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      vae: false
-      pin_cpu_memory: true
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: true
-  pipeline:
-    experimental:
-      attention_backend: ATTN_QAT_INFER
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-request:
-  prompt: >-
-    A wide cinematic shot of an alpine meadow at sunrise, pale pink mountain
-    peaks above a blue valley filled with thin morning mist.
-  negative_prompt: ""
-  sampling:
-    seed: 2026
-    height: 768
-    width: 1344
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 5
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/compacth3_rtx_pro6000/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_dreamx_world.py

@@ -0,0 +1,61 @@
+import os
+
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = os.getenv("DREAMX_WORLD_OUTPUT_PATH", "video_samples_dreamx_world")
+
+
+def _env_int(name: str, default: int) -> int:
+    return int(os.getenv(name, str(default)))
+
+
+def _env_float(name: str, default: float) -> float:
+    return float(os.getenv(name, str(default)))
+
+
+def main():
+    model_name = os.getenv("DREAMX_WORLD_MODEL_DIR", "FastVideo/DreamX-World-5B-Cam-Diffusers")
+    generator = VideoGenerator.from_pretrained(
+        model_name,
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        override_pipeline_cls_name="DreamXWorldPipeline",
+    )
+
+    prompt = os.getenv(
+        "DREAMX_WORLD_PROMPT",
+        "A cinematic first-person drive through a futuristic coastal city at "
+        "sunrise, reflective glass towers, clean streets, soft volumetric light.",
+    )
+    image_path = os.getenv(
+        "DREAMX_WORLD_IMAGE_PATH",
+        "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG",
+    )
+
+    kwargs = {
+        "output_path": OUTPUT_PATH,
+        "save_video": os.getenv("DREAMX_WORLD_SAVE_VIDEO", "1") != "0",
+        "height": _env_int("DREAMX_WORLD_HEIGHT", 480),
+        "width": _env_int("DREAMX_WORLD_WIDTH", 832),
+        "num_frames": _env_int("DREAMX_WORLD_NUM_FRAMES", 161),
+        "num_inference_steps": _env_int("DREAMX_WORLD_STEPS", 30),
+        "guidance_scale": _env_float("DREAMX_WORLD_GUIDANCE", 5.0),
+        "action_list": os.getenv("DREAMX_WORLD_ACTIONS", "w,d,w").split(","),
+        "action_speed_list":
+        [float(value) for value in os.getenv("DREAMX_WORLD_ACTION_SPEEDS", "4.0,2.0,4.0").split(",")],
+    }
+    if image_path:
+        kwargs["image_path"] = image_path
+
+    try:
+        generator.generate_video(prompt, **kwargs)
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_fasth3_omniref_pdd.py

@@ -1,268 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Eight-forward Ref2VA video+audio generation with a FastH3 OmniRef PDD student.
-
-A FastH3 OmniRef PDD export is a Parallel Decoding Distillation (PDD) student of
-MiniMax-H3's reference-conditioned ``transformer_ref`` partition. Its directory
-(local, or a Hugging Face repo) carries only what distillation changed:
-``transformer_ref/`` (output heads widened to the fine grid, trained VSA
-compression gates), the two scheduler configs, and ``fastvideo_inference.json``.
-The text encoder, tokenizer, processor, and both VAEs are base MiniMax-H3's.
-
-This script composes the two into one local model directory of symlinks and
-runs the recipe the contract records: fused blocks of the fine grid (one
-transformer forward each), the video/audio shifts, and VIDEO_SPARSE_ATTN_H3
-with its sparsity and tile size, where every reference video is its own
-sparse region. The contract's 128-token tiles run only on the sm_100a/sm_103a
-CUDA kernel (B200/B300/GB200/GB300) of a fastvideo-kernel build with the
-Blackwell VSA extension; the VSA-H3 backend raises at the first attention call
-when either is missing.
-With ``--num-gpus`` above 1 the DiT is sharded across the GPUs (FSDP) and the
-sequence is split across them (sequence parallelism).
-
-References are ordered. Pass them in order with --image / --video / --audio,
-for example ``--video dance.mp4 --image outfit.png --audio voice.wav``.
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-from collections.abc import Sequence
-from pathlib import Path
-from typing import Any
-
-from fastvideo import VideoGenerator
-from fastvideo.api import (
-    ComponentConfig,
-    EngineConfig,
-    GenerationRequest,
-    GeneratorConfig,
-    InputConfig,
-    OffloadConfig,
-    OutputConfig,
-    ParallelismConfig,
-    PipelineSelection,
-    SamplingConfig,
-)
-from fastvideo.configs.pipelines.minimax_h3 import parse_base_model_revision
-from fastvideo.pipelines.basic.minimax_h3 import MiniMaxH3Reference
-
-CONTRACT = "fastvideo_inference.json"
-# Components the distilled export replaces, and the ones it shares with base MiniMax-H3.
-EXPORT_COMPONENTS = ("transformer_ref", "scheduler", "audio_scheduler")
-BASE_COMPONENTS = ("text_encoder", "tokenizer", "processor", "vae", "audio_vae")
-MANIFESTS = ("modular_model_index.json", "model_index.json")
-# Components the composed directory's manifest must declare, besides its transformer_ref.
-MANIFEST_COMPONENTS = ("scheduler", "audio_scheduler", *BASE_COMPONENTS)
-
-
-class _AppendReference(argparse.Action):
-    """Collect --image/--video/--audio into one list that keeps command-line order."""
-
-    def __call__(self, parser, namespace, value, option_string=None):
-        references = list(getattr(namespace, self.dest, None) or [])
-        references.append((self.const, value))
-        setattr(namespace, self.dest, references)
-
-
-def build_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
-    parser.add_argument("--model-path",
-                        required=True,
-                        help="FastH3 OmniRef PDD export: a local directory or a Hugging Face repo id")
-    parser.add_argument("--revision", default=None, help="revision of --model-path when it is a Hugging Face repo")
-    parser.add_argument("--base-model-path",
-                        default=None,
-                        help="base MiniMax-H3 snapshot (local directory or repo id). Default: the repo and revision "
-                        "pinned by the export's base_model_revision")
-    parser.add_argument("--base-revision",
-                        default=None,
-                        help="revision of --base-model-path when it is a repo id (the export's pinned revision "
-                        "applies only to its own base repo); without --base-model-path, overrides that pin")
-    parser.add_argument("--composed-dir",
-                        default=None,
-                        help="where to write the composed model directory of symlinks (default: "
-                        "<output>/fasth3_omniref_model)")
-    for flag in ("image", "video", "audio"):
-        parser.add_argument(f"--{flag}",
-                            dest="references",
-                            action=_AppendReference,
-                            const=flag,
-                            metavar="PATH",
-                            help=f"an ordered {flag} reference (repeatable)")
-    parser.add_argument("--prompt", required=True)
-    parser.add_argument("--output", default="outputs/fasth3_omniref_pdd")
-    parser.add_argument("--height", type=int, default=480)
-    parser.add_argument("--width", type=int, default=832)
-    parser.add_argument("--num-frames", type=int, default=124)
-    parser.add_argument("--seed", type=int, default=0)
-    parser.add_argument("--num-gpus", type=int, default=1, help="sequence-parallel GPUs (must divide 56 heads)")
-    return parser
-
-
-def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
-    parser = build_parser()
-    args = parser.parse_args(argv)
-    if not args.references:
-        parser.error("pass at least one ordered reference with --image, --video, or --audio")
-    if not any(kind != "audio" for kind, _ in args.references):
-        parser.error("Ref2VA needs at least one image or video reference")
-    return args
-
-
-def _snapshot(path_or_repo: str, revision: str | None, allow_patterns: list[str]) -> Path:
-    local = Path(path_or_repo).expanduser()
-    if local.is_dir():
-        return local
-    from huggingface_hub import snapshot_download
-
-    return Path(snapshot_download(repo_id=path_or_repo, revision=revision, allow_patterns=allow_patterns))
-
-
-def load_contract(export_dir: Path) -> dict[str, Any]:
-    path = export_dir / CONTRACT
-    if not path.is_file():
-        raise FileNotFoundError(f"{export_dir} has no {CONTRACT}; it is not a FastH3 distilled export.")
-    contract = json.loads(path.read_text(encoding="utf-8"))
-    if "pdd_steps" not in contract or contract.get("model_type") != "ref2va":
-        raise ValueError(f"{path} is not a Ref2VA PDD contract; use the example matching the checkpoint.")
-    return contract
-
-
-def base_model_source(args: argparse.Namespace, contract: dict[str, Any]) -> tuple[str, str | None]:
-    """The base snapshot and revision: --base-model-path as given, else the export's pin."""
-    base_repo, base_revision = base_model_from_contract(contract)
-    if args.base_model_path:
-        return args.base_model_path, args.base_revision
-    return base_repo, args.base_revision or base_revision
-
-
-def base_model_from_contract(contract: dict[str, Any]) -> tuple[str, str]:
-    """The base repo id and revision that the export pins as ``hf://<repo id>@<revision>``."""
-    return parse_base_model_revision(contract.get("base_model_revision"))
-
-
-def _link(destination: Path, source: Path) -> None:
-    if not source.exists():
-        raise FileNotFoundError(f"Missing checkpoint component: {source}")
-    source = source.resolve()
-    if destination.is_symlink():
-        if destination.resolve() == source:
-            return
-        destination.unlink()
-    elif destination.exists():
-        raise FileExistsError(f"{destination} exists and is not a symlink; choose another --composed-dir.")
-    destination.symlink_to(source, target_is_directory=source.is_dir())
-
-
-def _declares_components(manifest: Path) -> bool:
-    """Whether a Diffusers manifest declares every component of the composed directory."""
-    declared = {
-        name
-        for name, spec in json.loads(manifest.read_text(encoding="utf-8")).items()
-        if isinstance(spec, list) and spec and spec[0] is not None
-    }
-    return set(MANIFEST_COMPONENTS) <= declared and bool({"transformer", "transformer_ref"} & declared)
-
-
-def select_manifest(export_dir: Path, base_dir: Path) -> Path:
-    """The export's manifest when it declares every composed component, else the base's."""
-    for root in (export_dir, base_dir):
-        for name in MANIFESTS:
-            if (root / name).is_file() and _declares_components(root / name):
-                return root / name
-    raise FileNotFoundError("Neither the export nor the base snapshot has a Diffusers model manifest that declares "
-                            f"{', '.join(MANIFEST_COMPONENTS)} and transformer_ref.")
-
-
-def compose_model_dir(export_dir: Path, base_dir: Path, composed_dir: Path) -> Path:
-    """One MiniMax-H3 model directory: distilled components from the export, the rest from the base."""
-    composed_dir.mkdir(parents=True, exist_ok=True)
-    for name in (*EXPORT_COMPONENTS, CONTRACT):
-        _link(composed_dir / name, export_dir / name)
-    manifest_source = select_manifest(export_dir, base_dir)
-    for name in MANIFESTS:
-        stale = composed_dir / name
-        if name != manifest_source.name and stale.is_symlink():
-            stale.unlink()
-    _link(composed_dir / manifest_source.name, manifest_source)
-    for name in BASE_COMPONENTS:
-        _link(composed_dir / name, base_dir / name)
-    return composed_dir
-
-
-def resolve_model(args: argparse.Namespace) -> tuple[Path, dict[str, Any]]:
-    export_dir = _snapshot(args.model_path, args.revision,
-                           [CONTRACT, *MANIFESTS, *(f"{name}/**" for name in EXPORT_COMPONENTS)])
-    contract = load_contract(export_dir)
-    base_source, base_revision = base_model_source(args, contract)
-    base_dir = _snapshot(base_source, base_revision, [*MANIFESTS, *(f"{name}/**" for name in BASE_COMPONENTS)])
-    composed_dir = Path(args.composed_dir) if args.composed_dir else Path(args.output) / "fasth3_omniref_model"
-    return compose_model_dir(export_dir, base_dir, composed_dir), contract
-
-
-def build_generator_config(model_dir: Path, num_gpus: int) -> GeneratorConfig:
-    return GeneratorConfig(
-        model_path=str(model_dir),
-        engine=EngineConfig(
-            num_gpus=num_gpus,
-            # Shard the DiT across the GPUs rather than holding a full copy on each.
-            use_fsdp_inference=num_gpus > 1,
-            parallelism=ParallelismConfig(tp_size=1, sp_size=num_gpus),
-            offload=OffloadConfig(dit=False, dit_layerwise=False, text_encoder=True, vae=True, pin_cpu_memory=False),
-        ),
-        pipeline=PipelineSelection(
-            workload_type="i2v",
-            # FastVideo reads the trained fused-block partition and attention
-            # settings from the composed directory's fastvideo_inference.json.
-            components=ComponentConfig(override_pipeline_cls_name="MiniMaxH3Ref2VAModularPipeline"),
-        ),
-    )
-
-
-def main() -> None:
-    args = parse_args()
-    output_dir = Path(args.output)
-    output_dir.mkdir(parents=True, exist_ok=True)
-    model_dir, contract = resolve_model(args)
-    print(f"Composed model directory: {model_dir}")
-    print(f"Contract: {contract['num_inference_steps']} fused blocks of a {contract['pdd_steps']}-interval grid, "
-          f"{contract['attention_backend']} with sparsity {contract['vsa_sparsity']}, "
-          f"{contract['vsa_tile_size']}-token tiles, reference keep rate {contract['vsa_ref_keep_rate']}")
-    references = [MiniMaxH3Reference(source=path, media_type=kind) for kind, path in args.references]
-
-    generator = VideoGenerator.from_config(build_generator_config(model_dir, args.num_gpus))
-    try:
-        result = generator.generate(
-            GenerationRequest(
-                prompt=args.prompt,
-                negative_prompt="",
-                inputs=InputConfig(references=references),
-                sampling=SamplingConfig(
-                    height=args.height,
-                    width=args.width,
-                    num_frames=args.num_frames,
-                    fps=24,
-                    # PDD counts fused blocks: one transformer forward each.
-                    num_inference_steps=contract["num_inference_steps"],
-                    guidance_scale=1.0,
-                    batch_cfg=False,
-                    seed=args.seed,
-                ),
-                output=OutputConfig(
-                    output_path=str(output_dir / "fasth3_omniref_pdd.mp4"),
-                    save_video=True,
-                    return_frames=False,
-                ),
-            ))
-        print(f"Output written to: {result.video_path}")
-        if result.generation_time is not None:
-            print(f"Generation time: {result.generation_time:.1f} s")
-        if result.peak_memory_mb is not None:
-            print(f"Peak GPU memory allocated (rank 0): {result.peak_memory_mb / 1024:.1f} GiB")
-    finally:
-        generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_fasth3_spark_pair_pruned_nvfp4.yaml

@@ -1,64 +0,0 @@
-# Start Ray on both Sparks and source spark_pair_env.sh first; see spark_pair.md.
-# FastH3 Trim eight-forward video+audio on two DGX Sparks over QSFP RoCE.
-# Download the complete checkpoint, including its trained schedule, as described in
-# docs/getting_started/installation/spark_performance.md.
-#
-# GB10 uses Triton VSA. The release packs attention, FFN and VSA gate weights.
-# The NVFP4 encoder and lightweight VAE ship in the same repository.
-#
-# FASTVIDEO_MINIMAX_H3_FUSIONS=all FASTVIDEO_NVFP4_MM_BACKEND=cutlass \
-# FASTVIDEO_H3_VAE_TILE_BATCH=1 \
-# FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 \
-# fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair_pruned_nvfp4.yaml
-generator:
-  model_path: FastVideo/FastVideo-FastH3-Trim-8-Step-NVFP4
-  engine:
-    num_gpus: 2
-    execution_backend: ray
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit_vsa
-    parallelism:
-      tp_size: 1
-      sp_size: 2
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      image_encoder: false
-      vae: false
-      pin_cpu_memory: false
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    workload_type: t2v
-    vae_tiling: true
-    experimental:
-      attention_backend: VIDEO_SPARSE_ATTN_H3
-      VSA_sparsity: 0.8
-      VSA_tile_size: 64
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: true
-      vae_parallel_decode_strategy: gather
-      video_decode_backend: h3-vae
-request:
-  prompt: A quiet pottery studio with a potter finishing a bowl at the wheel.
-  negative_prompt: ""
-  sampling:
-    seed: 1234
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 9  # nine sigma points, eight DiT forwards
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/fasth3_spark_pair_pruned_nvfp4/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_fasth3_spark_pair_v2_nvfp4.yaml

@@ -1,64 +0,0 @@
-# Start Ray on both Sparks and source spark_pair_env.sh first; see spark_pair.md.
-# FastH3 V2 eight-forward video+audio on two DGX Sparks over QSFP RoCE.
-# Download the complete release checkpoint as described in
-# docs/getting_started/installation/spark_performance.md.
-#
-# GB10 uses Triton VSA. The release packs attention, FFN and VSA gate weights.
-# The NVFP4 encoder and lightweight VAE ship in the same repository.
-#
-# FASTVIDEO_MINIMAX_H3_FUSIONS=all FASTVIDEO_NVFP4_MM_BACKEND=cutlass \
-# FASTVIDEO_H3_VAE_TILE_BATCH=1 \
-# FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 \
-# fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair_v2_nvfp4.yaml
-generator:
-  model_path: FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer
-  engine:
-    num_gpus: 2
-    execution_backend: ray
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit_vsa
-    parallelism:
-      tp_size: 1
-      sp_size: 2
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      image_encoder: false
-      vae: false
-      pin_cpu_memory: false
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    workload_type: t2v
-    vae_tiling: true
-    experimental:
-      attention_backend: VIDEO_SPARSE_ATTN_H3
-      VSA_sparsity: 0.8
-      VSA_tile_size: 64
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: true
-      vae_parallel_decode_strategy: gather
-      video_decode_backend: h3-vae
-request:
-  prompt: A quiet pottery studio with a potter finishing a bowl at the wheel.
-  negative_prompt: ""
-  sampling:
-    seed: 1234
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 9  # nine sigma points, eight DiT forwards
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/fasth3_spark_pair_v2_nvfp4/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_fasth3_spark_pruned_nvfp4.yaml

@@ -1,61 +0,0 @@
-# FastH3 Trim eight-forward video+audio on one DGX Spark.
-# Download the complete checkpoint, including its trained schedule, as described in
-# docs/getting_started/installation/spark_performance.md.
-#
-# GB10 uses Triton VSA. The release packs attention, FFN and VSA gate weights.
-# The NVFP4 encoder and lightweight VAE ship in the same repository.
-#
-# FASTVIDEO_MINIMAX_H3_FUSIONS=all FASTVIDEO_NVFP4_MM_BACKEND=cutlass \
-# FASTVIDEO_H3_VAE_TILE_BATCH=1 \
-# FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 \
-# fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pruned_nvfp4.yaml
-generator:
-  model_path: FastVideo/FastVideo-FastH3-Trim-8-Step-NVFP4
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit_vsa
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      image_encoder: false
-      vae: false
-      pin_cpu_memory: false
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    workload_type: t2v
-    vae_tiling: true
-    experimental:
-      attention_backend: VIDEO_SPARSE_ATTN_H3
-      VSA_sparsity: 0.8
-      VSA_tile_size: 64
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-request:
-  prompt: A quiet pottery studio with a potter finishing a bowl at the wheel.
-  negative_prompt: ""
-  sampling:
-    seed: 1234
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 9  # nine sigma points, eight DiT forwards
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/fasth3_spark_pruned_nvfp4/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_fasth3_spark_v2_nvfp4.yaml

@@ -1,61 +0,0 @@
-# FastH3 V2 eight-forward video+audio on one DGX Spark.
-# Download the complete release checkpoint as described in
-# docs/getting_started/installation/spark_performance.md.
-#
-# GB10 uses Triton VSA. The release packs attention, FFN and VSA gate weights.
-# The NVFP4 encoder and lightweight VAE ship in the same repository.
-#
-# FASTVIDEO_MINIMAX_H3_FUSIONS=all FASTVIDEO_NVFP4_MM_BACKEND=cutlass \
-# FASTVIDEO_H3_VAE_TILE_BATCH=1 \
-# FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 \
-# fastvideo generate --config examples/inference/basic/basic_fasth3_spark_v2_nvfp4.yaml
-generator:
-  model_path: FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit_vsa
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      image_encoder: false
-      vae: false
-      pin_cpu_memory: false
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    workload_type: t2v
-    vae_tiling: true
-    experimental:
-      attention_backend: VIDEO_SPARSE_ATTN_H3
-      VSA_sparsity: 0.8
-      VSA_tile_size: 64
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-request:
-  prompt: A quiet pottery studio with a potter finishing a bowl at the wheel.
-  negative_prompt: ""
-  sampling:
-    seed: 1234
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 9  # nine sigma points, eight DiT forwards
-    guidance_scale: 1.0
-    batch_cfg: false
-  output:
-    output_path: outputs/fasth3_spark_v2_nvfp4/
-    save_video: true
-    return_frames: false

examples/inference/basic/basic_gamecraft.py

@@ -0,0 +1,118 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Basic inference script for HunyuanGameCraft video generation.
+
+HunyuanGameCraft generates game-like videos with camera/action control.
+It takes an optional image input and generates video with camera motion
+based on simple action commands (forward, left, right, backward, rotations).
+
+Available actions:
+    - forward (w): Move camera forward
+    - backward (s): Move camera backward  
+    - left (a): Move camera left (strafe)
+    - right (d): Move camera right (strafe)
+    - left_rot: Rotate camera left (pan)
+    - right_rot: Rotate camera right (pan)
+    - up_rot: Rotate camera up (tilt)
+    - down_rot: Rotate camera down (tilt)
+
+T2V vs I2V:
+    - Default: I2V (uses a default reference image). Set GAMECRAFT_I2V_IMAGE to a
+      URL or path to use a different image.
+    - T2V only (no reference image): run with GAMECRAFT_I2V_IMAGE= (empty).
+"""
+import os
+
+import torch
+
+from fastvideo import VideoGenerator
+from fastvideo.models.camera import create_camera_trajectory
+
+# Model configuration (use GAMECRAFT_MODEL_PATH for local weights)
+MODEL_PATH = os.environ.get("GAMECRAFT_MODEL_PATH", "FastVideo/HunyuanGameCraft-Diffusers")
+
+# Default prompts for demo
+DEFAULT_PROMPTS = {
+    "village":
+    "A charming medieval village with cobblestone streets, thatched-roof houses, and vibrant flower gardens under a bright blue sky.",
+    "temple":
+    "A majestic ancient temple stands under a clear blue sky, its grandeur highlighted by towering Doric columns and intricate architectural details.",
+    "forest":
+    "A lush green forest with tall trees, dappled sunlight filtering through the leaves, and a winding dirt path.",
+    "beach": "A tropical beach with crystal clear turquoise water, white sand, and palm trees swaying in the breeze.",
+}
+
+# I2V: default reference image (URL). Can override with a local path.
+DEFAULT_I2V_IMAGE_URL = ("https://huggingface.co/datasets/huggingface/documentation-images/"
+                         "resolve/main/diffusers/astronaut.jpg")
+DEFAULT_I2V_PROMPT = ("An astronaut hatching from an egg, on the surface of the moon, "
+                      "the darkness and depth of space realised in the background.")
+
+OUTPUT_PATH = "video_samples_gamecraft"
+
+
+def main():
+    # Initialize generator
+    # FastVideo will automatically download weights from HuggingFace
+    generator = VideoGenerator.from_pretrained(
+        MODEL_PATH,
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+    )
+
+    # Video parameters
+    height = 704
+    width = 1280
+    num_frames = 33
+    action = "forward"
+    action_speed = 0.2
+
+    # Create camera trajectory (Plücker coordinates)
+    camera_states = create_camera_trajectory(
+        action=action,
+        height=height,
+        width=width,
+        num_frames=num_frames,
+        action_speed=action_speed,
+        dtype=torch.bfloat16,
+    )
+    print(f"Camera states shape: {camera_states.shape}")
+
+    # I2V vs T2V: unset GAMECRAFT_I2V_IMAGE -> I2V (default image). Set to "" -> T2V.
+    env_image = os.environ.get("GAMECRAFT_I2V_IMAGE")
+    if env_image is None:
+        image_path = DEFAULT_I2V_IMAGE_URL  # default: I2V
+    elif env_image.strip() == "":
+        image_path = None  # T2V
+    else:
+        image_path = env_image.strip()  # I2V with given URL/path
+
+    is_i2v = image_path is not None
+    prompt = DEFAULT_I2V_PROMPT if is_i2v else DEFAULT_PROMPTS["temple"]
+    print(f"Mode: {'I2V' if is_i2v else 'T2V'}, prompt: {prompt[:60]}...")
+
+    gen_kw = dict(
+        prompt=prompt,
+        negative_prompt="",
+        camera_states=camera_states,
+        height=height,
+        width=width,
+        num_frames=num_frames,
+        num_inference_steps=50,
+        guidance_scale=6.0,
+        seed=42,
+        fps=24,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+    )
+    if is_i2v:
+        gen_kw["image_path"] = image_path
+    generator.generate_video(**gen_kw)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_glm_image.py

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

examples/inference/basic/basic_kandinsky6_sr.py

@@ -1,53 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Kandinsky6 video super-resolution: upscale a low-resolution video x2, x2.25 or x4.
-
-The request needs no prompt and no height / width / num_frames: the geometry and frame rate follow the input video
-(the first 121 frames, 5 s at 24 fps, are processed and the source audio is kept).  Use
-``kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers`` for the 2-step distilled model.
-See docs/inference/kandinsky6_sr.md.
-"""
-import argparse
-
-from fastvideo import VideoGenerator
-
-
-def main():
-    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
-    parser.add_argument("--model-path", default="kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers")
-    parser.add_argument("--video-path", required=True, help="Low-resolution source video")
-    parser.add_argument("--output-path", default="outputs_video/kandinsky6_sr",
-                        help="Output directory (the file is named after the input) or .mp4 path")
-    parser.add_argument("--scale", type=float, default=2.25, choices=[2.0, 4.0, 2.25], help="Total upscale factor")
-    parser.add_argument("--num-steps", type=int, default=None,
-                        help="Denoising steps per tile (default: the checkpoint's preset, 4 or 2)")
-    parser.add_argument("--tiles-batch-size", type=int, default=1, help="Tiles denoised per DiT call")
-    parser.add_argument("--target-resolution", default=None, help="Optional final size: hd, fullhd, 2k or WxH")
-    parser.add_argument("--seed", type=int, default=42)
-    args = parser.parse_args()
-
-    generator = VideoGenerator.from_pretrained(args.model_path, num_gpus=1)
-    sampling = {"seed": args.seed}
-    if args.num_steps is not None:
-        sampling["num_inference_steps"] = args.num_steps
-    result = generator.generate({
-        "inputs": {
-            "video_path": args.video_path
-        },
-        "sampling": sampling,
-        "output": {
-            "output_path": args.output_path,
-            "save_video": True,
-            "return_frames": False,
-        },
-        "extensions": {
-            "sr_resolution_scale": args.scale,
-            "sr_tiles_batch_size": args.tiles_batch_size,
-            "sr_target_resolution": args.target_resolution,
-        },
-    })
-    print(f"Wrote {result.video_path}")
-    generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_kandinsky6_ti2va.py

@@ -1,51 +0,0 @@
-import os
-
-from fastvideo import VideoGenerator
-
-DEFAULT_PROMPT = (
-    "cinematic shot: a giant stone samurai on a stormy cliff above a neon city opens glowing golden eyes and "
-    "raises a katana. Blue lightning strikes the blade, creating a massive shockwave through the clouds. The "
-    "camera rapidly pulls back from a low angle. Photorealistic, epic scale, dark blue and gold lighting, rain, "
-    "sparks, volumetric lightning, blockbuster quality. Audio: heavy rain, deep thunder, metallic sword hum, "
-    "rising brass and choir, electrical crackle, perfectly synchronized lightning impact, sub-bass shockwave. "
-    "No dialogue, text, or logos."
-)
-
-OUTPUT_PATH = "video_samples_kandinsky6_ti2va"
-
-# The base checkpoint uses 50 steps and guidance 5.0. Set KANDINSKY6_MODEL_PATH to run the
-# distilled pi-Flow checkpoint with this same example:
-#   kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers
-# or a local directory named ``Kandinsky-6.0-Pro-distill-5s-Diffusers``.
-# That name selects the registered distilled defaults automatically: 10 steps,
-# guidance 1.0, eps 1e-6, final_step_size_scale 0.5, and
-# num_policy_substeps 128.
-MODEL_PATH = os.environ.get("KANDINSKY6_MODEL_PATH", "kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers")
-
-IMAGE_PATH = None  # e.g. "assets/girl.png" to condition on an image
-
-
-def main():
-    generator = VideoGenerator.from_pretrained(
-        MODEL_PATH,
-        num_gpus=1,
-        use_fsdp_inference=False,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=True,
-        pin_cpu_memory=True,
-    )
-
-    _ = generator.generate_video(
-        DEFAULT_PROMPT,
-        image_path=IMAGE_PATH,
-        output_path=OUTPUT_PATH,
-        save_video=True,
-        height=512,
-        width=768,
-        num_frames=121,
-    )
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_longcat_i2v.py

@@ -0,0 +1,204 @@
+"""
+LongCat Image-to-Video (I2V) Example Script
+
+This script demonstrates LongCat I2V inference using the FastVideo Python API.
+LongCat I2V takes an input image and generates a video from it.
+
+It runs both basic generation (50 steps) and distill+refine generation
+(16 steps distill + 50 steps refinement to 720p with BSA).
+
+Usage:
+    python examples/inference/basic/basic_longcat_i2v.py
+
+Note:
+    Refinement uses 768x768 dimensions where latent (48x48) is divisible by 8,
+    compatible with BSA chunks [4, 4, 8].
+"""
+
+import glob
+import os
+
+from fastvideo import VideoGenerator
+
+# Common prompts and settings matching the shell script examples
+PROMPT = ("A woman sits at a wooden table by the window in a cozy café. She reaches out "
+          "with her right hand, picks up the white coffee cup from the saucer, and gently "
+          "brings it to her lips to take a sip. After drinking, she places the cup back on "
+          "the table and looks out the window, enjoying the peaceful atmosphere.")
+
+NEGATIVE_PROMPT = ("Bright tones, overexposed, static, blurred details, subtitles, style, works, "
+                   "paintings, images, static, overall gray, worst quality, low quality, JPEG compression "
+                   "residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, "
+                   "deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, "
+                   "three legs, many people in the background, walking backwards")
+
+# Input image path
+IMAGE_PATH = "assets/girl.png"
+
+SEED = 42
+
+
+def basic_generation():
+    """
+    Run basic LongCat I2V generation (50 steps at 480p).
+    
+    This uses the full 50-step denoising process for highest quality.
+    """
+    print("=" * 60)
+    print("LongCat I2V: Basic Generation (50 steps, 480p)")
+    print("=" * 60)
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-I2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=False,  # set to True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+    )
+
+    output_path = "outputs_video/longcat_i2v_basic"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        image_path=IMAGE_PATH,
+        output_path=output_path,
+        save_video=True,
+        height=480,
+        width=480,  # Square
+        num_frames=93,
+        num_inference_steps=50,
+        fps=15,
+        guidance_scale=4.0,
+        seed=SEED,
+    )
+
+    print(f"\nBasic generation complete! Video saved to: {output_path}")
+    generator.shutdown()
+
+
+def distill_refine_generation():
+    """
+    Run LongCat I2V with distill+refine pipeline (16 steps + refinement to 768p).
+    
+    This uses the distilled LoRA for fast 480p generation (16 steps),
+    then refines to 768p using the refinement LoRA with BSA enabled.
+    """
+    print("\n" + "=" * 60)
+    print("LongCat I2V: Distill + Refine Pipeline")
+    print("=" * 60)
+
+    # Stage 1: Distilled generation (16 steps at 480p)
+    print("\n[Stage 1] Distilled generation (16 steps, 480p)")
+    print("-" * 40)
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-I2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+        lora_path="FastVideo/LongCat-Video-T2V-Distilled-LoRA",
+        lora_nickname="distilled",
+    )
+
+    distill_output_path = "outputs_video/longcat_i2v_distill"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        image_path=IMAGE_PATH,
+        output_path=distill_output_path,
+        save_video=True,
+        height=480,
+        width=480,  # Square
+        num_frames=93,
+        num_inference_steps=16,
+        fps=15,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Distilled generation complete! Video saved to: {distill_output_path}")
+    generator.shutdown()
+
+    # Stage 2: Refinement (480p -> 768p)
+    print("\n[Stage 2] Refinement (480p -> 768p with BSA)")
+    print("-" * 40)
+
+    # Find the actual saved video file from stage 1
+    video_files = glob.glob(os.path.join(distill_output_path, "*.mp4"))
+    if not video_files:
+        raise FileNotFoundError(f"No video file found in {distill_output_path}")
+    # Use the most recently created video file
+    distill_video_path = max(video_files, key=os.path.getmtime)
+    print(f"Using stage 1 video: {distill_video_path}")
+
+    # Create a new generator with refinement LoRA and BSA enabled
+    # Note: Refinement uses the T2V model (not I2V) since it's upscaling the generated video
+    # For BSA [4, 4, 8]: latent must be divisible by 8
+    # 768x768: latent 48x48, 48%8=0 ✓
+    refine_generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=True,
+        bsa_sparsity=0.875,
+        bsa_chunk_q=[4, 4, 4],
+        bsa_chunk_k=[4, 4, 4],
+        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
+        lora_nickname="refinement",
+    )
+
+    refine_output_path = "outputs_video/longcat_i2v_refine_720p"
+
+    refine_generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=refine_output_path,
+        save_video=True,
+        refine_from=distill_video_path,
+        t_thresh=0.5,
+        spatial_refine_only=False,
+        num_cond_frames=0,
+        height=720,
+        width=720,
+        num_inference_steps=50,
+        fps=30,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Refinement complete! Video saved to: {refine_output_path}")
+    refine_generator.shutdown()
+
+
+def main():
+    """Run both basic and distill+refine generation pipelines."""
+    print("\n" + "=" * 60)
+    print("LongCat Image-to-Video Example")
+    print("=" * 60 + "\n")
+
+    # Run basic generation
+    basic_generation()
+
+    # Run distill+refine pipeline
+    distill_refine_generation()
+
+    print("\n" + "=" * 60)
+    print("All generations complete!")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_longcat_t2v.py

@@ -0,0 +1,192 @@
+"""
+LongCat Text-to-Video (T2V) Example Script
+
+This script demonstrates LongCat T2V inference using the FastVideo Python API.
+It runs both basic generation (50 steps) and distill+refine generation
+(16 steps distill + 50 steps refinement to 720p).
+
+Usage:
+    python examples/inference/basic/basic_longcat_t2v.py
+"""
+
+import glob
+import os
+
+from fastvideo import VideoGenerator
+
+# Common prompts and settings matching the shell script examples
+PROMPT = ("In a realistic photography style, a white boy around seven or eight years old "
+          "sits on a park bench, wearing a light blue T-shirt, denim shorts, and white sneakers. "
+          "He holds an ice cream cone with vanilla and chocolate flavors, and beside him is a "
+          "medium-sized golden Labrador. Smiling, the boy offers the ice cream to the dog, "
+          "who eagerly licks it with its tongue. The sun is shining brightly, and the background "
+          "features a green lawn and several tall trees, creating a warm and loving scene.")
+
+NEGATIVE_PROMPT = ("Bright tones, overexposed, static, blurred details, subtitles, style, works, "
+                   "paintings, images, static, overall gray, worst quality, low quality, JPEG compression "
+                   "residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, "
+                   "deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, "
+                   "three legs, many people in the background, walking backwards")
+
+SEED = 42
+
+
+def basic_generation():
+    """
+    Run basic LongCat T2V generation (50 steps at 480p).
+    
+    This uses the full 50-step denoising process for highest quality.
+    """
+    print("=" * 60)
+    print("LongCat T2V: Basic Generation (50 steps, 480p)")
+    print("=" * 60)
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=False,  # set to True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+    )
+
+    output_path = "outputs_video/longcat_t2v_basic"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=output_path,
+        save_video=True,
+        height=480,
+        width=832,
+        num_frames=93,
+        num_inference_steps=50,
+        fps=15,
+        guidance_scale=4.0,
+        seed=SEED,
+    )
+
+    print(f"\nBasic generation complete! Video saved to: {output_path}")
+    generator.shutdown()
+
+
+def distill_refine_generation():
+    """
+    Run LongCat T2V with distill+refine pipeline (16 steps + refinement to 720p).
+    
+    This uses the distilled LoRA for fast 480p generation (16 steps),
+    then refines to 720p using the refinement LoRA with BSA enabled.
+    """
+    print("\n" + "=" * 60)
+    print("LongCat T2V: Distill + Refine Pipeline")
+    print("=" * 60)
+
+    # Stage 1: Distilled generation (16 steps at 480p)
+    print("\n[Stage 1] Distilled generation (16 steps, 480p)")
+    print("-" * 40)
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+        lora_path="FastVideo/LongCat-Video-T2V-Distilled-LoRA",
+        lora_nickname="distilled",
+    )
+
+    distill_output_path = "outputs_video/longcat_t2v_distill"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=distill_output_path,
+        save_video=True,
+        height=480,
+        width=832,
+        num_frames=93,
+        num_inference_steps=16,
+        fps=15,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Distilled generation complete! Video saved to: {distill_output_path}")
+    generator.shutdown()
+
+    # Stage 2: Refinement (480p -> 720p)
+    print("\n[Stage 2] Refinement (480p -> 720p with BSA)")
+    print("-" * 40)
+
+    # Find the actual saved video file from stage 1
+    video_files = glob.glob(os.path.join(distill_output_path, "*.mp4"))
+    if not video_files:
+        raise FileNotFoundError(f"No video file found in {distill_output_path}")
+    # Use the most recently created video file
+    distill_video_path = max(video_files, key=os.path.getmtime)
+    print(f"Using stage 1 video: {distill_video_path}")
+
+    # Create a new generator with refinement LoRA and BSA enabled
+    refine_generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=True,
+        bsa_sparsity=0.875,
+        bsa_chunk_q=[4, 4, 8],
+        bsa_chunk_k=[4, 4, 8],
+        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
+        lora_nickname="refinement",
+    )
+
+    refine_output_path = "outputs_video/longcat_t2v_refine_720p"
+
+    refine_generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=refine_output_path,
+        save_video=True,
+        refine_from=distill_video_path,
+        t_thresh=0.5,
+        spatial_refine_only=False,
+        num_cond_frames=0,
+        height=720,
+        width=1280,
+        num_inference_steps=50,
+        fps=30,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Refinement complete! Video saved to: {refine_output_path}")
+    refine_generator.shutdown()
+
+
+def main():
+    """Run both basic and distill+refine generation pipelines."""
+    print("\n" + "=" * 60)
+    print("LongCat Text-to-Video Example")
+    print("=" * 60 + "\n")
+
+    # Run basic generation
+    basic_generation()
+
+    # Run distill+refine pipeline
+    distill_refine_generation()
+
+    print("\n" + "=" * 60)
+    print("All generations complete!")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_longcat_vc.py

@@ -0,0 +1,218 @@
+"""
+LongCat Video Continuation (VC) Example Script
+
+This script demonstrates LongCat VC inference using the FastVideo Python API.
+LongCat VC takes an input video and generates a continuation of it.
+
+It runs both basic generation (50 steps) and distill+refine generation
+(16 steps distill + 50 steps refinement to 720p).
+
+Usage:
+    python examples/inference/basic/basic_longcat_vc.py
+
+Prerequisites:
+    - Ensure the input video exists at assets/motorcycle.mp4
+      (or provide your own video)
+"""
+
+import glob
+import os
+
+from fastvideo import VideoGenerator
+
+# Common prompts and settings matching the shell script examples
+PROMPT = ("A person rides a motorcycle along a long, straight road that stretches between "
+          "a body of water and a forested hillside. The rider steadily accelerates, keeping "
+          "the motorcycle centered between the guardrails, while the scenery passes by on "
+          "both sides. The video captures the journey from the rider's perspective, emphasizing "
+          "the sense of motion and adventure.")
+
+NEGATIVE_PROMPT = ("Bright tones, overexposed, static, blurred details, subtitles, style, works, "
+                   "paintings, images, static, overall gray, worst quality, low quality, JPEG compression "
+                   "residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, "
+                   "deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, "
+                   "three legs, many people in the background, walking backwards")
+
+# Input video path
+VIDEO_PATH = "assets/motorcycle.mp4"
+
+# Number of conditioning frames from the input video
+NUM_COND_FRAMES = 13
+
+SEED = 42
+
+
+def basic_generation():
+    """
+    Run basic LongCat VC generation (50 steps at 480p).
+    
+    This uses the full 50-step denoising process for highest quality.
+    """
+    print("=" * 60)
+    print("LongCat VC: Basic Generation (50 steps, 480p)")
+    print("=" * 60)
+
+    # Check if video exists
+    if not os.path.exists(VIDEO_PATH):
+        raise FileNotFoundError(f"Video not found at {VIDEO_PATH}. "
+                                "Please provide a valid video path.")
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-VC-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=False,  # set to True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+    )
+
+    output_path = "outputs_video/longcat_vc_basic"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        video_path=VIDEO_PATH,
+        num_cond_frames=NUM_COND_FRAMES,
+        output_path=output_path,
+        save_video=True,
+        height=480,
+        width=832,
+        num_frames=93,
+        num_inference_steps=50,
+        fps=15,
+        guidance_scale=4.0,
+        seed=SEED,
+    )
+
+    print(f"\nBasic generation complete! Video saved to: {output_path}")
+    generator.shutdown()
+
+
+def distill_refine_generation():
+    """
+    Run LongCat VC with distill+refine pipeline (16 steps + refinement to 720p).
+    
+    This uses the distilled LoRA for fast 480p generation (16 steps),
+    then refines to 720p using the refinement LoRA with BSA enabled.
+    """
+    print("\n" + "=" * 60)
+    print("LongCat VC: Distill + Refine Pipeline")
+    print("=" * 60)
+
+    # Check if video exists
+    if not os.path.exists(VIDEO_PATH):
+        raise FileNotFoundError(f"Video not found at {VIDEO_PATH}. "
+                                "Please provide a valid video path.")
+
+    # Stage 1: Distilled generation (16 steps at 480p)
+    print("\n[Stage 1] Distilled generation (16 steps, 480p)")
+    print("-" * 40)
+
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-VC-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=False,
+        lora_path="FastVideo/LongCat-Video-T2V-Distilled-LoRA",
+        lora_nickname="distilled",
+    )
+
+    distill_output_path = "outputs_video/longcat_vc_distill"
+
+    generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        video_path=VIDEO_PATH,
+        num_cond_frames=NUM_COND_FRAMES,
+        output_path=distill_output_path,
+        save_video=True,
+        height=480,
+        width=832,
+        num_frames=93,
+        num_inference_steps=16,
+        fps=15,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Distilled generation complete! Video saved to: {distill_output_path}")
+    generator.shutdown()
+
+    # Stage 2: Refinement (480p -> 720p)
+    print("\n[Stage 2] Refinement (480p -> 720p with BSA)")
+    print("-" * 40)
+
+    # Find the actual saved video file from stage 1
+    video_files = glob.glob(os.path.join(distill_output_path, "*.mp4"))
+    if not video_files:
+        raise FileNotFoundError(f"No video file found in {distill_output_path}")
+    # Use the most recently created video file
+    distill_video_path = max(video_files, key=os.path.getmtime)
+    print(f"Using stage 1 video: {distill_video_path}")
+
+    # Create a new generator with refinement LoRA and BSA enabled
+    # Note: Refinement uses the T2V model (not VC) since it's upscaling the generated video
+    refine_generator = VideoGenerator.from_pretrained(
+        "FastVideo/LongCat-Video-T2V-Diffusers",
+        num_gpus=1,
+        use_fsdp_inference=True,
+        dit_cpu_offload=True,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        enable_bsa=True,
+        bsa_sparsity=0.875,
+        bsa_chunk_q=[4, 4, 8],
+        bsa_chunk_k=[4, 4, 8],
+        lora_path="FastVideo/LongCat-Video-T2V-Refinement-LoRA",
+        lora_nickname="refinement",
+    )
+
+    refine_output_path = "outputs_video/longcat_vc_refine_720p"
+
+    refine_generator.generate_video(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        output_path=refine_output_path,
+        save_video=True,
+        refine_from=distill_video_path,
+        t_thresh=0.5,
+        spatial_refine_only=False,
+        num_cond_frames=0,  # For refinement, no conditioning frames
+        height=720,
+        width=1280,
+        num_inference_steps=50,
+        fps=30,
+        guidance_scale=1.0,
+        seed=SEED,
+    )
+
+    print(f"Refinement complete! Video saved to: {refine_output_path}")
+    refine_generator.shutdown()
+
+
+def main():
+    """Run both basic and distill+refine generation pipelines."""
+    print("\n" + "=" * 60)
+    print("LongCat Video Continuation Example")
+    print("=" * 60 + "\n")
+
+    # Run basic generation
+    basic_generation()
+
+    # Run distill+refine pipeline
+    distill_refine_generation()
+
+    print("\n" + "=" * 60)
+    print("All generations complete!")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_turbodiffusion.py

@@ -0,0 +1,51 @@
+import os
+
+# Set SLA attention backend BEFORE fastvideo imports
+os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
+
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_turbodiffusion"
+
+
+def main() -> None:
+    # TurboDiffusion: 1-4 step video generation using RCM scheduler + SLA attention
+    # FastVideo will automatically use TurboDiffusionPipeline when specified
+    generator = VideoGenerator.from_pretrained(
+        "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers",
+        # FastVideo will automatically handle distributed setup
+        num_gpus=1,
+        use_fsdp_inference=False,  # set to True if GPU is out of memory
+
+        # set to false if using RTX 4090
+        # pin_cpu_memory=False,
+    )
+
+    # Generate videos with the same simple API, regardless of GPU count
+    # TurboDiffusion defaults: guidance_scale=1.0 and num_inference_steps=4 (from config)
+    prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+              "wide with interest. The playful yet serene atmosphere is complemented by soft "
+              "natural light filtering through the petals. Mid-shot, warm and cheerful tones.")
+    video = generator.generate_video(
+        prompt,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        seed=42,
+    )
+
+    # Generate another video with a different prompt, without reloading the model!
+    prompt2 = ("A majestic lion strides across the golden savanna, its powerful frame "
+               "glistening under the warm afternoon sun. The tall grass ripples gently in "
+               "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+               "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+               "cinematic.")
+    video2 = generator.generate_video(
+        prompt2,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        seed=42,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_turbodiffusion_14b.py

@@ -0,0 +1,45 @@
+import os
+
+# Set SLA attention backend BEFORE fastvideo imports
+os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
+
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_turbodiffusion_14B"
+
+
+def main() -> None:
+    # TurboDiffusion 14B: 1-4 step video generation using RCM scheduler + SLA attention
+    # FastVideo will automatically use TurboDiffusionPipeline when specified
+    generator = VideoGenerator.from_pretrained(
+        "loayrashid/TurboWan2.1-T2V-14B-Diffusers",
+        # 14B model needs more GPUs
+        num_gpus=2,
+    )
+
+    prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+              "wide with interest. The playful yet serene atmosphere is complemented by soft "
+              "natural light filtering through the petals. Mid-shot, warm and cheerful tones.")
+    video = generator.generate_video(
+        prompt,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        seed=42,
+    )
+
+    # Generate another video with a different prompt, without reloading the model!
+    prompt2 = ("A majestic lion strides across the golden savanna, its powerful frame "
+               "glistening under the warm afternoon sun. The tall grass ripples gently in "
+               "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+               "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+               "cinematic.")
+    video2 = generator.generate_video(
+        prompt2,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        seed=42,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_turbodiffusion_i2v.py

@@ -0,0 +1,38 @@
+import os
+
+# Set SLA attention backend BEFORE fastvideo imports
+os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLA_ATTN"
+
+from fastvideo import VideoGenerator
+
+# Use local model path
+MODEL_PATH = "loayrashid/TurboWan2.2-I2V-A14B-Diffusers"
+OUTPUT_PATH = "video_samples_turbodiffusion_i2v"
+
+
+def main() -> None:
+    # TurboDiffusion I2V: 1-4 step image-to-video generation
+    generator = VideoGenerator.from_pretrained(
+        MODEL_PATH,
+        num_gpus=2,
+    )
+
+    # Example prompt and image for I2V
+    prompt = (
+        "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside."
+    )
+
+    # Use an example image path
+    image_path = "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG"
+
+    video = generator.generate_video(
+        prompt,
+        image_path=image_path,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        seed=42,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_zimage.py

@@ -0,0 +1,94 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Run Z-Image-Turbo text-to-image generation through FastVideo.
+
+User story:
+    "I want the official Z-Image-Turbo defaults and a deterministic PNG from
+    a local or Hugging Face checkpoint."
+"""
+
+import argparse
+from pathlib import Path
+
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    OutputConfig,
+    ParallelismConfig,
+    PipelineSelection,
+    SamplingConfig,
+)
+
+DEFAULT_PROMPT = (
+    "Young Chinese woman in red Hanfu, intricate embroidery. Impeccable makeup, red floral forehead pattern. "
+    "Elaborate high bun, golden phoenix headdress, red flowers, beads. Holds round folding fan with lady, trees, bird. "
+    "Neon lightning-bolt lamp (⚡️), bright yellow glow, above extended left palm. Soft-lit outdoor night background, "
+    "silhouetted tiered pagoda (西安大雁塔), blurred colorful distant lights.")
+DEFAULT_REVISION = "f332072aa78be7aecdf3ee76d5c247082da564a6"
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Run Z-Image-Turbo text-to-image generation.")
+    parser.add_argument("--model-path", default="Tongyi-MAI/Z-Image-Turbo")
+    parser.add_argument("--revision", default=DEFAULT_REVISION)
+    parser.add_argument("--output", default="outputs/zimage/zimage_turbo.png")
+    parser.add_argument("--prompt", default=DEFAULT_PROMPT)
+    parser.add_argument("--negative-prompt", default="")
+    parser.add_argument("--height", type=int, default=1024)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--steps", type=int, default=8)
+    parser.add_argument("--guidance-scale", type=float, default=0.0)
+    parser.add_argument("--max-sequence-length", type=int, default=512)
+    parser.add_argument("--cfg-normalization", action=argparse.BooleanOptionalAction, default=False)
+    parser.add_argument("--cfg-truncation", type=float, default=1.0)
+    parser.add_argument("--seed", type=int, default=42)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    output = Path(args.output)
+    output.parent.mkdir(parents=True, exist_ok=True)
+
+    generator = VideoGenerator.from_config(
+        GeneratorConfig(
+            model_path=args.model_path,
+            revision=args.revision,
+            engine=EngineConfig(
+                num_gpus=1,
+                parallelism=ParallelismConfig(tp_size=1, sp_size=1),
+                use_fsdp_inference=False,
+            ),
+            # The model registry selects the native zimage_turbo preset.
+            pipeline=PipelineSelection(workload_type="t2i"),
+        ))
+    try:
+        generator.generate(
+            GenerationRequest(
+                prompt=args.prompt,
+                negative_prompt=args.negative_prompt,
+                sampling=SamplingConfig(
+                    height=args.height,
+                    width=args.width,
+                    num_frames=1,
+                    fps=1,
+                    num_inference_steps=args.steps,
+                    guidance_scale=args.guidance_scale,
+                    max_sequence_length=args.max_sequence_length,
+                    cfg_normalization=args.cfg_normalization,
+                    cfg_truncation=args.cfg_truncation,
+                    seed=args.seed,
+                ),
+                output=OutputConfig(
+                    output_path=str(output),
+                    save_video=True,
+                    return_frames=False,
+                ),
+            ))
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/benchmark_fasth3_spark_nvfp4.py

@@ -1,149 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Time a resident FastH3 eight-forward Spark recipe with the release prompts.
-
-One process loads the model, runs one excluded ceramics warmup, then times
-each prompt at least twice. Every call writes a video. This script does not alter
-the V2 schedule, VSA sparsity, or video resolution.
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-import statistics
-import time
-from copy import deepcopy
-from pathlib import Path
-
-from fastvideo import VideoGenerator
-from fastvideo.api.parser import load_raw_config, parse_config
-from fastvideo.api.schema import RunConfig
-
-PROMPT_IDS = ("latency-ceramics-005", "latency-harbor-005")
-
-
-def _stage_metrics(result: object) -> dict[str, dict]:
-    logging_info = getattr(result, "logging_info", None)
-    stages = getattr(logging_info, "stages", None)
-    if isinstance(logging_info, dict):
-        stages = logging_info.get("stages", stages)
-    if not isinstance(stages, dict):
-        return {}
-    return {name: metrics for name, metrics in stages.items() if isinstance(metrics, dict)}
-
-
-def _stage_seconds(metrics: dict[str, dict]) -> dict[str, float]:
-    return {name: float(stage["execution_time"]) for name, stage in metrics.items()
-            if stage.get("execution_time") is not None}
-
-
-def _peak_mb(metrics: dict[str, dict], key: str) -> float | None:
-    values = [float(stage[key]) for stage in metrics.values() if stage.get(key) is not None]
-    return max(values) if values else None
-
-
-def _stage_total(stages: dict[str, float], fragment: str, exclude: str | None = None) -> float | None:
-    matches = [
-        seconds for name, seconds in stages.items()
-        if fragment in name.lower() and (exclude is None or exclude not in name.lower())
-    ]
-    return sum(matches) if matches else None
-
-
-def _request(base: RunConfig, prompt: str, frames: int, width: int, height: int,
-             output: Path):
-    request = deepcopy(base.request)
-    request.prompt = prompt
-    request.inputs.prompt_path = None
-    request.sampling.num_frames = frames
-    request.sampling.width = width
-    request.sampling.height = height
-    request.output.output_path = str(output)
-    request.output.save_video = True
-    request.output.return_frames = False
-    return request
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--config", type=Path, required=True)
-    parser.add_argument("--prompts", type=Path, required=True)
-    parser.add_argument("--output-dir", type=Path, required=True)
-    parser.add_argument("--model-path", type=Path)
-    parser.add_argument("--frames", type=int, default=124)
-    parser.add_argument("--width", type=int, default=832)
-    parser.add_argument("--height", type=int, default=480)
-    parser.add_argument("--repeats", type=int, default=2)
-    args = parser.parse_args()
-
-    if args.frames not in (124, 243):
-        parser.error("use 124 frames for roughly five seconds or 243 for roughly ten seconds")
-    if args.repeats < 2:
-        parser.error("the release protocol requires at least two timed calls")
-
-    config = parse_config(RunConfig, load_raw_config(args.config))
-    if args.model_path:
-        config.generator.model_path = str(args.model_path)
-    if config.request.sampling.num_inference_steps != 9:
-        parser.error("the V2 contract requires nine sigma points for eight DiT forwards")
-    if config.generator.engine.offload.lazy_module_load is not False:
-        parser.error("the resident recipe requires lazy_module_load: false")
-    contract = Path(config.generator.model_path) / "fastvideo_inference.json"
-    if not contract.is_file():
-        parser.error(f"missing trained V2 schedule: {contract}")
-    inference = json.loads(contract.read_text())
-    if inference.get("num_inference_steps") != 9 or inference.get("transformer_forwards") != 8:
-        parser.error("the checkpoint is not the trained V2 eight-forward schedule")
-
-    prompts = json.loads(args.prompts.read_text())
-    if any(prompt_id not in prompts for prompt_id in PROMPT_IDS):
-        parser.error(f"prompt JSON must contain {', '.join(PROMPT_IDS)}")
-    args.output_dir.mkdir(parents=True, exist_ok=True)
-
-    generator = VideoGenerator.from_config(config.generator)
-    try:
-        for prompt_index, prompt_id in enumerate(PROMPT_IDS):
-            times = []
-            first_index = 0 if prompt_index == 0 else 1
-            for index in range(first_index, args.repeats + 1):
-                warmup = index == 0
-                label = "warmup" if warmup else f"run-{index:02d}"
-                requested_path = args.output_dir / f"{prompt_id}-{args.width}x{args.height}-{args.frames}-{label}.mp4"
-                request = _request(config, prompts[prompt_id], args.frames, args.width, args.height,
-                                   requested_path)
-                started = time.perf_counter()
-                result = generator.generate(request)
-                wall = time.perf_counter() - started
-                output = Path(result.video_path) if result.video_path else requested_path
-                if not output.is_file():
-                    raise RuntimeError(f"generation returned without an MP4: {output}")
-                metrics = _stage_metrics(result)
-                stages = _stage_seconds(metrics)
-                row = {
-                    "prompt_id": prompt_id,
-                    "warmup": warmup,
-                    "frames": args.frames,
-                    "width": args.width,
-                    "height": args.height,
-                    "e2e_seconds": round(wall, 3),
-                    "denoise_seconds": _stage_total(stages, "denois"),
-                    "decode_seconds": _stage_total(stages, "decod", exclude="postdecode"),
-                    "postprocess_seconds": _stage_total(stages, "postdecode"),
-                    "peak_memory_mb": _peak_mb(metrics, "peak_allocated_mb"),
-                    "peak_reserved_mb": _peak_mb(metrics, "peak_reserved_mb"),
-                    "result_peak_memory_mb": result.peak_memory_mb,
-                    "stages": stages,
-                    "mp4": str(output),
-                }
-                print(json.dumps(row, sort_keys=True), flush=True)
-                if not warmup:
-                    times.append(wall)
-            print(json.dumps({"prompt_id": prompt_id, "timed_runs": len(times),
-                              "median_e2e_seconds": round(statistics.median(times), 3)},
-                             sort_keys=True), flush=True)
-    finally:
-        generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/edit_glm_image.py

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

examples/inference/basic/mlx_fasth3.py

@@ -190,37 +190,34 @@ def run(args: argparse.Namespace) -> None:
         taeh3_chunk_size=args.taeh3_chunk_size,
         prompt_cache_dir=args.prompt_cache_dir,
     )
-    try:
-        result = pipeline.generate(
-            args.prompt,
-            output_path=args.output_path,
-            height=args.height,
-            width=args.width,
-            num_frames=args.num_frames,
-            seed=args.seed,
-            num_steps=args.steps,
-            tiled_video_decode=args.tiled_video_decode,
-            vae_tile_height=args.vae_tile_height,
-            vae_tile_width=args.vae_tile_width,
-            inter_step_cooldown_s=args.inter_step_cooldown_s,
-            fast=args.fast,
-            fast_factor=args.fast_factor,
-            fast_sharpen=args.fast_sharpen,
-            rife_weights_dir=args.rife_weights_dir,
-            fast_spatial=args.fast_spatial,
-            fast_spatial_scale=args.fast_spatial_scale,
-            fast_spatial_upsample_mode=args.fast_spatial_upsample_mode,
-            fast_spatial_sharpen=args.fast_spatial_sharpen,
-            vsa=args.vsa,
-            vsa_sparsity=args.vsa_sparsity,
-            vsa_tile_size=args.vsa_tile_size,
-            vsa_prefix_mode=args.vsa_prefix_mode,
-            vsa_dense_first_n_steps=args.vsa_dense_first_n_steps,
-            vsa_dense_layers=args.vsa_dense_layers,
-            vsa_impl=args.vsa_impl,
-        )
-    finally:
-        pipeline.close()
+    result = pipeline.generate(
+        args.prompt,
+        output_path=args.output_path,
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        seed=args.seed,
+        num_steps=args.steps,
+        tiled_video_decode=args.tiled_video_decode,
+        vae_tile_height=args.vae_tile_height,
+        vae_tile_width=args.vae_tile_width,
+        inter_step_cooldown_s=args.inter_step_cooldown_s,
+        fast=args.fast,
+        fast_factor=args.fast_factor,
+        fast_sharpen=args.fast_sharpen,
+        rife_weights_dir=args.rife_weights_dir,
+        fast_spatial=args.fast_spatial,
+        fast_spatial_scale=args.fast_spatial_scale,
+        fast_spatial_upsample_mode=args.fast_spatial_upsample_mode,
+        fast_spatial_sharpen=args.fast_spatial_sharpen,
+        vsa=args.vsa,
+        vsa_sparsity=args.vsa_sparsity,
+        vsa_tile_size=args.vsa_tile_size,
+        vsa_prefix_mode=args.vsa_prefix_mode,
+        vsa_dense_first_n_steps=args.vsa_dense_first_n_steps,
+        vsa_dense_layers=args.vsa_dense_layers,
+        vsa_impl=args.vsa_impl,
+    )
     print(json.dumps({
         "video_path": result.video_path,
         "timings_s": {k: round(v, 2) for k, v in result.timings.items()},

examples/serving/openai_compacth3_rtx5090.yaml

@@ -1,53 +0,0 @@
-#   FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-#     FLASHINFER_CUDA_ARCH_LIST=12.0a \
-#     fastvideo serve --config examples/serving/openai_compacth3_rtx5090.yaml
-generator:
-  model_path: ./CompactH3
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: true
-      image_encoder: true
-      vae: false
-      pin_cpu_memory: true
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: false
-  pipeline:
-    workload_type: t2v
-    experimental:
-      attention_backend: ATTN_QAT_INFER
-      h3_sequential_load: true
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-
-server:
-  host: 127.0.0.1
-  port: 8000
-  output_dir: outputs/openai_compacth3_rtx5090
-  served_model_name: compacth3
-
-default_request:
-  negative_prompt: ""
-  sampling:
-    height: 480
-    width: 832
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 5
-    guidance_scale: 1.0
-    batch_cfg: false
-    seed: 2026
-  output:
-    return_frames: false

examples/serving/openai_compacth3_rtx_pro6000.yaml

@@ -1,53 +0,0 @@
-#   FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
-#     FLASHINFER_CUDA_ARCH_LIST=12.0a \
-#     fastvideo serve --config examples/serving/openai_compacth3_rtx_pro6000.yaml
-generator:
-  model_path: ./CompactH3
-  engine:
-    num_gpus: 1
-    use_fsdp_inference: false
-    quantization:
-      transformer_quant: NVFP4
-      layer_profile: h3_dit
-    parallelism:
-      tp_size: 1
-      sp_size: 1
-    offload:
-      dit: false
-      dit_layerwise: false
-      text_encoder: false
-      image_encoder: true
-      vae: false
-      pin_cpu_memory: true
-      lazy_module_load: false
-    compile:
-      enabled: false
-      vae_enabled: true
-  pipeline:
-    workload_type: t2v
-    experimental:
-      attention_backend: ATTN_QAT_INFER
-      h3_sequential_load: false
-      inference_torch_compile: false
-      vae_parallel_decode: false
-      video_decode_backend: h3-vae
-
-server:
-  host: 127.0.0.1
-  port: 8000
-  output_dir: outputs/openai_compacth3_rtx_pro6000
-  served_model_name: compacth3
-
-default_request:
-  negative_prompt: ""
-  sampling:
-    height: 768
-    width: 1344
-    num_frames: 124
-    fps: 24
-    num_inference_steps: 5
-    guidance_scale: 1.0
-    batch_cfg: false
-    seed: 2026
-  output:
-    return_frames: false

examples/serving/openai_wan22_t2v_a14b.yaml

@@ -25,6 +25,5 @@ default_request:
     num_frames: 81
     height: 720
     width: 1280
-    fps: 16
   output:
     return_frames: false

examples/serving/openai_wan22_ti2v_5b.yaml

@@ -1,8 +1,6 @@
 # OpenAI-compatible server for Wan2.2 TI2V 5B. Supports both text-to-video
 # (omit the image reference) and image-to-video (set input_reference or
-# image_reference) from the same model. The checked-in cookbook clients send
-# text prompts only. The sampling block pins the values the cookbook displays;
-# they match the registered preset, and a test keeps them in sync.
+# image_reference) from the same model.
 #   fastvideo serve --config examples/serving/openai_wan22_ti2v_5b.yaml
 generator:
   model_path: Wan-AI/Wan2.2-TI2V-5B-Diffusers
@@ -25,10 +23,5 @@ server:
   served_model_name: wan22-ti2v-5b
 
 default_request:
-  sampling:
-    height: 704
-    width: 1280
-    num_frames: 121
-    fps: 24
   output:
     return_frames: false

examples/train/configs/fine_tuning/kandinsky5/README.md

@@ -187,7 +187,7 @@ to a supported dense backend while the FP4 linear layers still run.
   `Kandinsky5DMDConfig` override, generating deterministically (fixed
   seed), and comparing MS-SSIM against a committed reference video. The
   test fails (does not skip) if the reference is missing; record it once on
-  a sanctioned GPU box with `FASTVIDEO_TEST_KANDINSKY5_E2E_WRITE_REFERENCE=1`, review the
+  a sanctioned GPU box with `KANDINSKY5_E2E_WRITE_REFERENCE=1`, review the
   written video, and commit it (see the test's module docstring). Nightly
   tests are not collected per-PR (`fastvideo/tests/contract/
   test_ci_test_collection.py` allowlists the directory), so run it

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

@@ -0,0 +1,71 @@
+# LongCat-Video T2V 13.6B bidirectional finetune.
+
+models:
+  student:
+    _target_: fastvideo.train.models.longcat.LongCatModel
+    init_from: FastVideo/LongCat-Video-T2V-Diffusers
+    trainable: true
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
+
+training:
+  model_path: FastVideo/LongCat-Video-T2V-Diffusers
+
+  distributed:
+    num_gpus: 8
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 8
+
+  data:
+    data_path: data/LongCat-Syn
+    dataloader_num_workers: 4
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 1000
+    num_latent_t: 20
+    num_height: 480
+    num_width: 848
+    num_frames: 77
+
+  optimizer:
+    learning_rate: 1.0e-6
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 4000
+    gradient_accumulation_steps: 1
+
+  checkpoint:
+    output_dir: outputs/longcat_finetune
+    training_state_checkpointing_steps: 1000
+    checkpoints_total_limit: 3
+
+  tracker:
+    project_name: fastvideo
+    run_name: longcat_finetune
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+  validation:
+    _target_: fastvideo.train.callbacks.validation.ValidationCallback
+    pipeline_target: fastvideo.pipelines.basic.longcat.longcat_pipeline.LongCatPipeline
+    dataset_file: data/validation_prompts.json
+    every_steps: 100
+    sampling_steps: [50]
+    guidance_scale: 5.0
+
+pipeline:
+  # Match the released LongCat scheduler config. flow_shift=0.0 collapses
+  # FlowMatch training timesteps to zero in FastVideo's scheduler.
+  flow_shift: 12.0

examples/train/configs/fine_tuning/wan/turbo_t2v_lora_sla.yaml

@@ -0,0 +1,89 @@
+# TurboWan 2.1 T2V 1.3B LoRA finetune with SLA attention.
+#
+# Notes:
+# - TurboDiffusion checkpoints expect the SLA attention backend. The training
+#   entrypoint now auto-selects `SLA_ATTN` for model IDs containing
+#   `turbodiffusion` or `turbowan`, unless you override the environment
+#   variable manually.
+# - This debug config reuses the existing crush-smol T2V parquet recipe. If you
+#   preprocess data with the official TurboWan recipe, update the temporal and
+#   spatial fields to match it.
+
+models:
+  student:
+    _target_: fastvideo.train.models.wan.WanModel
+    init_from: loayrashid/TurboWan2.1-T2V-1.3B-Diffusers
+    trainable: true
+    lora:
+      enable: true
+      rank: 16
+      alpha: 32
+      target_modules:
+        - to_q
+        - to_k
+        - to_v
+        - to_out
+
+method:
+  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
+
+training:
+  dit_precision: bf16
+
+  distributed:
+    num_gpus: 1
+    sp_size: 1
+    tp_size: 1
+    hsdp_replicate_dim: 1
+    hsdp_shard_dim: 1
+
+  data:
+    data_path: data/crush-smol_processed_t2v/combined_parquet_dataset
+    dataloader_num_workers: 2
+    train_batch_size: 1
+    training_cfg_rate: 0.0
+    seed: 42
+    num_latent_t: 20
+    num_height: 480
+    num_width: 832
+    num_frames: 77
+
+  optimizer:
+    learning_rate: 1.0e-5
+    betas: [0.9, 0.999]
+    weight_decay: 0.01
+    lr_scheduler: constant
+    lr_warmup_steps: 0
+
+  loop:
+    max_train_steps: 10
+    gradient_accumulation_steps: 5
+
+  checkpoint:
+    output_dir: outputs/turbo_wan_t2v_lora_sla
+    training_state_checkpointing_steps: 5
+    checkpoints_total_limit: 2
+    resume_from_checkpoint: null
+
+  tracker:
+    project_name: fastvideo
+    run_name: turbo_wan_t2v_lora_sla
+
+  model:
+    enable_gradient_checkpointing_type: full
+
+callbacks:
+  grad_clip:
+    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
+    max_grad_norm: 1.0
+
+  validation:
+    _target_: fastvideo.train.callbacks.validation.ValidationCallback
+    pipeline_target: fastvideo.pipelines.basic.turbodiffusion.turbodiffusion_pipeline.TurboDiffusionPipeline
+    dataset_file: examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json
+    every_steps: 5
+    sampling_steps: [4]
+    guidance_scale: 1.0
+
+pipeline:
+  flow_shift: 3

fastvideo-kernel/attn_qat_infer/api.py

@@ -106,32 +106,12 @@ def preprocess_qkv(q: torch.Tensor,
     if enable_smoothing_q:
         delta_s = torch.matmul(qm, k.transpose(-2, -1)).to(torch.float32).contiguous()
     else:  # used to disable q smoothing
-        delta_s = _zero_delta_s(q.shape[0], q.shape[1], k.shape[2], q.device)
+        B, H, L, D = q.shape
+        delta_s = torch.zeros((B, H, L // BLOCK_M, k.shape[2]), device=q.device, dtype=torch.float32)
 
     return q, k, v, delta_s
 
 
-_ZERO_DELTA_S: dict = {}
-
-
-def _zero_delta_s(batch: int, heads: int, kv_len: int, device: torch.device) -> torch.Tensor:
-    """Cached all-zero delta_s for unsmoothed Q, read with per_block_mean=False.
-
-    The kernel reads delta_s as a contiguous [B, H, rows, KL] tensor with one
-    row per query block when per_block_mean is set and a single row otherwise.
-    With Q smoothing off every row is zero, so one shared row replaces the
-    [B, H, L/128, KL] tensor that was allocated and zero-filled on every call
-    (9.5 GB at 73k tokens, whose int32 batch stride also broke the TMA
-    descriptor). The kernel only reads it.
-    """
-    key = (batch, heads, kv_len, device)
-    zeros = _ZERO_DELTA_S.get(key)
-    if zeros is None:
-        zeros = torch.zeros((batch, heads, 1, kv_len), device=device, dtype=torch.float32)
-        _ZERO_DELTA_S[key] = zeros
-    return zeros
-
-
 def scale_and_quant_fp4(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
     assert x.ndim == 4
     B, H, N, D = x.shape
@@ -174,191 +154,6 @@ def blockscaled_fp4_attn(qlist: Tuple,
                             softmax_scale, is_causal, per_block_mean, is_bf16, single_level_p_quant)
 
 
-def blockscaled_fp4_attn_sparse(qlist: Tuple,
-                                klist: Tuple,
-                                vlist: Tuple,
-                                delta_s: torch.Tensor,
-                                KL: int,
-                                q2k_idx: torch.Tensor,
-                                q2k_num: torch.Tensor,
-                                kv_valid: torch.Tensor | None = None,
-                                q2k_quad: torch.Tensor | None = None,
-                                per_block_mean: bool = True,
-                                is_bf16: bool = True,
-                                single_level_p_quant: bool = False,
-                                sm_scale: float | None = None):
-    softmax_scale = sm_scale if sm_scale is not None else (qlist[0].shape[-1] * 2)**(-0.5)
-    return fp4attn_cuda.fwd_sparse(qlist[0], klist[0], vlist[0], qlist[1], klist[1], vlist[1], delta_s, KL, None,
-                                   softmax_scale, per_block_mean, is_bf16, single_level_p_quant, q2k_idx, q2k_num,
-                                   kv_valid, q2k_quad)
-
-
-HALF_N = BLOCK_N // 2
-
-
-def vsa_tile_mask_to_fp4_blocks(
-    tile_mask: torch.Tensor,
-    tile_tokens: int,
-    tile_valid: torch.Tensor | None = None,
-    validate: bool = False,
-) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None]:
-    """Convert a VSA tile mask into the FP4 kernel's block lists.
-
-    ``tile_mask`` is a ``[B, H, T, T]`` bool mask over VSA tiles of
-    ``tile_tokens`` tokens (query tile, key tile); ``tile_tokens`` is 64 or a
-    multiple of 128. ``tile_valid`` holds each tile's valid token count (valid
-    tokens first, padding last, as VSA tiling lays them out). The kernel works
-    on 128x128 blocks, so 64-token tiles are paired: a block is listed when any
-    of its four 64x64 quadrants is selected, and ``q2k_quad`` carries which.
-
-    Lists run in descending block order because the kernel visits them last
-    entry first: block 0 (the first prefix tile, which VSA-H3's exempt mode
-    gives every query) is then visited first, so every query row starts from a
-    finite running max. ``validate=True`` checks that (host sync).
-
-    Returns ``(q2k_idx [B, H, M, N], q2k_num [B, H, M], kv_valid [2N],
-    q2k_quad [B, H, M, N] | None)``; the queries cover ``M * 128`` padded rows.
-    """
-    if tile_tokens % HALF_N or (tile_tokens > HALF_N and tile_tokens % BLOCK_N):
-        raise ValueError(f"tile_tokens={tile_tokens} must be {HALF_N} or a multiple of {BLOCK_N}")
-    batch, heads, n_tiles, _ = tile_mask.shape
-    device = tile_mask.device
-    halves_per_tile = tile_tokens // HALF_N
-    half_mask = tile_mask
-    if halves_per_tile > 1:
-        half_mask = half_mask.repeat_interleave(halves_per_tile, dim=2).repeat_interleave(halves_per_tile, dim=3)
-    n_halves = n_tiles * halves_per_tile
-    if tile_valid is None:
-        half_valid = torch.full((n_halves, ), HALF_N, device=device, dtype=torch.int32)
-    else:
-        offsets = torch.arange(halves_per_tile, device=device, dtype=torch.int32) * HALF_N
-        half_valid = (tile_valid.to(torch.int32)[:, None] - offsets[None, :]).clamp(0, HALF_N).reshape(-1)
-    if n_halves % 2:
-        # Pad to whole 128-blocks: the extra key half is empty, the extra query
-        # half (discarded output) attends block 0 only.
-        half_mask = F.pad(half_mask, (0, 1, 0, 1), value=False)
-        half_mask[:, :, -1, 0] = True
-        half_valid = F.pad(half_valid, (0, 1), value=0)
-        n_halves += 1
-    half_mask = half_mask & (half_valid > 0)[None, None, None, :]
-    n_blocks = n_halves // 2
-    quads = half_mask.view(batch, heads, n_blocks, 2, n_blocks, 2)
-    weights = torch.tensor([[1, 2], [4, 8]], device=device, dtype=torch.uint8)  # [row_half, col_half]
-    quad = (quads.to(torch.uint8) * weights[None, None, None, :, None, :]).sum(dim=(3, 5), dtype=torch.uint8)
-    block_mask = quad != 0
-    # Descending compaction without a sort: position = running count from the right.
-    rev = block_mask.flip(-1)
-    pos = rev.cumsum(-1, dtype=torch.int32) - 1
-    q2k_num = (pos[..., -1] + 1).contiguous()
-    cols = torch.arange(n_blocks - 1, -1, -1, device=device, dtype=torch.int32).expand_as(pos)
-    slot = torch.where(rev, pos, torch.full_like(pos, n_blocks)).long()
-    q2k_idx = torch.zeros((batch, heads, n_blocks, n_blocks + 1), device=device, dtype=torch.int32)
-    q2k_idx.scatter_(-1, slot, cols)
-    q2k_idx = q2k_idx[..., :n_blocks].contiguous()
-    kv_valid = half_valid.contiguous()
-    q2k_quad = None
-    if halves_per_tile == 1:
-        q2k_quad = torch.zeros((batch, heads, n_blocks, n_blocks + 1), device=device, dtype=torch.uint8)
-        q2k_quad.scatter_(-1, slot, quad.flip(-1))
-        q2k_quad = q2k_quad[..., :n_blocks].contiguous()
-    if validate:
-        if int(q2k_num.min()) < 1:
-            raise ValueError("every query block must attend to at least one non-empty KV block")
-        last = (q2k_num - 1).long().unsqueeze(-1)
-        first_block = q2k_idx.gather(-1, last.int().long()).squeeze(-1).long()
-        first_quad = (q2k_quad.gather(-1, last).squeeze(-1).int() if q2k_quad is not None else
-                      torch.full_like(first_block, 15, dtype=torch.int32))
-        v0 = (kv_valid[2 * first_block] > 0).int()
-        v1 = (kv_valid[2 * first_block + 1] > 0).int()
-        row0 = ((first_quad & 1).bool() & v0.bool()) | ((first_quad & 2).bool() & v1.bool())
-        row1 = ((first_quad & 4).bool() & v0.bool()) | ((first_quad & 8).bool() & v1.bool())
-        if not bool((row0 & row1).all()):
-            raise ValueError("the first block each query block visits must give both 64-row halves a valid key")
-    return q2k_idx, q2k_num, kv_valid, q2k_quad
-
-
-def check_sparse_block_lists(q2k_idx: torch.Tensor, q2k_num: torch.Tensor, kv_len: int) -> None:
-    """Reject block lists the sparse kernel would read out of bounds (host sync).
-
-    Each row needs ``1 <= q2k_num <= q2k_idx.size(-1)`` and every listed index
-    in ``[0, ceil(kv_len / BLOCK_N))``; a zero count or an out-of-range index
-    makes the kernel load outside its index row or the KV tensors.
-    """
-    num_kv_blocks = -(-kv_len // BLOCK_N)
-    if int(q2k_num.min()) < 1 or int(q2k_num.max()) > q2k_idx.size(-1):
-        raise ValueError(f"q2k_num must be in [1, {q2k_idx.size(-1)}]")
-    listed = torch.arange(q2k_idx.size(-1), device=q2k_idx.device) < q2k_num.unsqueeze(-1)
-    idx = q2k_idx[listed]
-    if idx.numel() and (int(idx.min()) < 0 or int(idx.max()) >= num_kv_blocks):
-        raise ValueError(f"q2k_idx entries must be in [0, {num_kv_blocks})")
-
-
-def sageattn_blackwell_sparse(q,
-                              k,
-                              v,
-                              q2k_idx: torch.Tensor,
-                              q2k_num: torch.Tensor,
-                              kv_valid: torch.Tensor | None = None,
-                              q2k_quad: torch.Tensor | None = None,
-                              per_block_mean=True,
-                              single_level_p_quant=True,
-                              sm_scale: float | None = None,
-                              validate: bool = True):
-    """Block-sparse SageAttention3 FP4 forward (non-causal).
-
-    Query block ``m`` (``BLOCK_M`` rows) of each (batch, head) attends only to
-    the ``BLOCK_N``-token KV blocks in ``q2k_idx[b, h, m, :q2k_num[b, h, m]]``,
-    restricted to the quadrants in ``q2k_quad`` when given; see
-    :func:`vsa_tile_mask_to_fp4_blocks`. Q/K/V are ``[B, H, L, D]``.
-    Block lists are checked with :func:`check_sparse_block_lists` (a host
-    sync) unless ``validate=False``; pass that only for lists built by
-    :func:`vsa_tile_mask_to_fp4_blocks`, which are in range by construction.
-    """
-    QL = q.size(2)
-    KL = k.size(2)
-    if validate:
-        check_sparse_block_lists(q2k_idx, q2k_num, KL)
-    is_bf16 = q.dtype == torch.bfloat16
-    q, k, v, delta_s = preprocess_qkv(q, k, v, per_block_mean)
-    per_block_mean = delta_s.shape[2] > 1
-    qlist_from_cuda = scale_and_quant_fp4(q)
-    klist_from_cuda = scale_and_quant_fp4_permute(k)
-    vlist_from_cuda = scale_and_quant_fp4_transpose(v)
-    o_fp4 = blockscaled_fp4_attn_sparse(qlist_from_cuda, klist_from_cuda, vlist_from_cuda, delta_s, KL, q2k_idx,
-                                        q2k_num, kv_valid, q2k_quad, per_block_mean, is_bf16, single_level_p_quant,
-                                        sm_scale)[0][:, :, :QL, :].contiguous()
-    return o_fp4
-
-
-def sageattn_blackwell_sparse_bshd(q,
-                                   k,
-                                   v,
-                                   q2k_idx: torch.Tensor,
-                                   q2k_num: torch.Tensor,
-                                   kv_valid: torch.Tensor | None = None,
-                                   q2k_quad: torch.Tensor | None = None,
-                                   single_level_p_quant=True,
-                                   sm_scale: float | None = None,
-                                   validate: bool = True) -> torch.Tensor:
-    """:func:`sageattn_blackwell_sparse` for ``[B, L, H, D]`` inputs, without copies.
-
-    The FP4 quantizers read strided input, so the sequence-major tensors a
-    linear produces are quantized in place of a transpose + pad. ``L`` must be
-    a multiple of ``BLOCK_M`` (callers allocate the padding) and Q is
-    unsmoothed. Returns ``[B, H, L, D]``.
-    """
-    batch, seq_len, heads, _ = q.shape
-    if seq_len % BLOCK_M:
-        raise ValueError(f"sequence length {seq_len} must be a multiple of {BLOCK_M}")
-    if validate:
-        check_sparse_block_lists(q2k_idx, q2k_num, seq_len)
-    qh, kh, vh = (x.transpose(1, 2) for x in (q, k, v))
-    delta_s = _zero_delta_s(batch, heads, seq_len, q.device)
-    return blockscaled_fp4_attn_sparse(scale_and_quant_fp4(qh), scale_and_quant_fp4_permute(kh),
-                                       scale_and_quant_fp4_transpose(vh), delta_s, seq_len, q2k_idx, q2k_num, kv_valid,
-                                       q2k_quad, False, q.dtype == torch.bfloat16, single_level_p_quant, sm_scale)[0]
-
-
 def sageattn_blackwell(q,
                        k,
                        v,
@@ -396,7 +191,6 @@ def sageattn_blackwell(q,
     KL = k.size(2)
     is_bf16 = q.dtype == torch.bfloat16
     q, k, v, delta_s = preprocess_qkv(q, k, v, per_block_mean)
-    per_block_mean = delta_s.shape[2] > 1
     qlist_from_cuda = scale_and_quant_fp4(q)
     klist_from_cuda = scale_and_quant_fp4_permute(k)
     vlist_from_cuda = scale_and_quant_fp4_transpose(v)

fastvideo-kernel/attn_qat_infer/blackwell/api.cu

@@ -204,17 +204,8 @@ void run_mha_fwd(Flash_fwd_params &params, cudaStream_t stream, bool force_split
     }));
 }
 
-struct SparseKvLists {
-    int const *q2k_idx = nullptr;
-    int const *q2k_num = nullptr;
-    int q2k_max = 0;
-    int num_m_blocks = 0;
-    int const *kv_valid = nullptr;
-    uint8_t const *q2k_quad = nullptr;
-};
-
-static std::vector<at::Tensor>
-mha_fwd_impl(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head_size // 2)
+std::vector<at::Tensor>
+mha_fwd(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head_size // 2)
         const at::Tensor &k,         // batch_size x seqlen_k x num_heads_k x (head_size // 2)
         const at::Tensor &v,         // batch_size x seqlen_k x num_heads_k x (head_size // 2)
         const at::Tensor &sfq,
@@ -227,8 +218,7 @@ mha_fwd_impl(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head
         bool is_causal, 
         bool per_block_mean,
         bool is_bf16,
-        bool single_level_p_quant,  // If true, use only per-row scale s_P2 (no per-block s_P1)
-        SparseKvLists const &sparse
+        bool single_level_p_quant=false  // If true, use only per-row scale s_P2 (no per-block s_P1)
     ) {
 
     auto dprops = at::cuda::getCurrentDeviceProperties();
@@ -326,12 +316,6 @@ mha_fwd_impl(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head
     // stack-local tensor whose data pointer would dangle after mha_fwd returns
     // while the async kernel may still be running.
     params.tile_count_semaphore = nullptr;
-    params.q2k_idx = sparse.q2k_idx;
-    params.q2k_num = sparse.q2k_num;
-    params.q2k_max = sparse.q2k_max;
-    params.num_m_blocks = sparse.num_m_blocks;
-    params.kv_valid = sparse.kv_valid;
-    params.q2k_quad = sparse.q2k_quad;
 
     if (seqlen_k > 0) {
         auto stream = at::cuda::getCurrentCUDAStream().stream();
@@ -357,72 +341,7 @@ mha_fwd_impl(at::Tensor &q,         // batch_size x seqlen_q x num_heads x (head
 
 
 
-std::vector<at::Tensor>
-mha_fwd(at::Tensor &q, const at::Tensor &k, const at::Tensor &v,
-        const at::Tensor &sfq, const at::Tensor &sfk, const at::Tensor &sfv,
-        const at::Tensor &delta_s, int unpadded_k, c10::optional<at::Tensor> &out_,
-        const float softmax_scale, bool is_causal, bool per_block_mean, bool is_bf16,
-        bool single_level_p_quant=false) {
-    return mha_fwd_impl(q, k, v, sfq, sfk, sfv, delta_s, unpadded_k, out_, softmax_scale,
-                        is_causal, per_block_mean, is_bf16, single_level_p_quant, SparseKvLists{});
-}
-
-// Block-sparse forward: query block m of (b, h) attends only to the KV blocks
-// listed in q2k_idx[b, h, m, :q2k_num[b, h, m]] (BLOCK_M x BLOCK_N granularity).
-// kv_valid, if given, holds the valid token count of each 64-column half of
-// every KV block ([2 * num_kv_blocks], valid tokens first within a half).
-// q2k_quad, if given (uint8, same shape as q2k_idx), restricts each listed
-// block to the 64x64 quadrants whose bit (2 * row_half + col_half) is set, so
-// 64-token VSA tiles run on 128x128 blocks. The block a list visits first
-// (its last entry) must leave every query row at least one valid key.
-// Non-causal only.
-std::vector<at::Tensor>
-mha_fwd_sparse(at::Tensor &q, const at::Tensor &k, const at::Tensor &v,
-               const at::Tensor &sfq, const at::Tensor &sfk, const at::Tensor &sfv,
-               const at::Tensor &delta_s, int unpadded_k, c10::optional<at::Tensor> &out_,
-               const float softmax_scale, bool per_block_mean, bool is_bf16,
-               bool single_level_p_quant,
-               const at::Tensor &q2k_idx, const at::Tensor &q2k_num,
-               c10::optional<at::Tensor> &kv_valid_,
-               c10::optional<at::Tensor> &q2k_quad_) {
-    const int batch_size = q.size(0);
-    const int num_heads = q.size(1);
-    const int num_m_blocks = (q.size(2) + flash::BLOCK_M - 1) / flash::BLOCK_M;
-    const int num_n_blocks = (k.size(2) + flash::BLOCK_N - 1) / flash::BLOCK_N;
-    for (auto const *t : {&q2k_idx, &q2k_num}) {
-        TORCH_CHECK(t->scalar_type() == torch::kInt32, "q2k_idx / q2k_num must be int32");
-        CHECK_DEVICE((*t)); CHECK_CONTIGUOUS((*t));
-    }
-    TORCH_CHECK(q2k_idx.dim() == 4, "q2k_idx must be [batch, heads, num_m_blocks, max_kv_blocks]");
-    TORCH_CHECK(q2k_idx.size(0) == batch_size && q2k_idx.size(1) == num_heads && q2k_idx.size(2) == num_m_blocks,
-                "q2k_idx leading dims must be [batch, heads, ceil(seqlen_q / BLOCK_M)]");
-    TORCH_CHECK(q2k_idx.size(3) >= 1 && q2k_idx.size(3) <= num_n_blocks, "q2k_idx last dim must be in [1, num_kv_blocks]");
-    CHECK_SHAPE(q2k_num, batch_size, num_heads, num_m_blocks);
-    SparseKvLists sparse;
-    sparse.q2k_idx = q2k_idx.data_ptr<int>();
-    sparse.q2k_num = q2k_num.data_ptr<int>();
-    sparse.q2k_max = q2k_idx.size(3);
-    sparse.num_m_blocks = num_m_blocks;
-    if (kv_valid_.has_value()) {
-        auto const &kv_valid = kv_valid_.value();
-        TORCH_CHECK(kv_valid.scalar_type() == torch::kInt32, "kv_valid must be int32");
-        CHECK_DEVICE(kv_valid); CHECK_CONTIGUOUS(kv_valid);
-        CHECK_SHAPE(kv_valid, 2 * num_n_blocks);
-        sparse.kv_valid = kv_valid.data_ptr<int>();
-    }
-    if (q2k_quad_.has_value()) {
-        auto const &q2k_quad = q2k_quad_.value();
-        TORCH_CHECK(q2k_quad.scalar_type() == torch::kUInt8, "q2k_quad must be uint8");
-        CHECK_DEVICE(q2k_quad); CHECK_CONTIGUOUS(q2k_quad);
-        TORCH_CHECK(q2k_quad.sizes() == q2k_idx.sizes(), "q2k_quad must match q2k_idx's shape");
-        sparse.q2k_quad = q2k_quad.data_ptr<uint8_t>();
-    }
-    return mha_fwd_impl(q, k, v, sfq, sfk, sfv, delta_s, unpadded_k, out_, softmax_scale,
-                        /*is_causal=*/false, per_block_mean, is_bf16, single_level_p_quant, sparse);
-}
-
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     m.doc() = "FlashAttention";
     m.def("fwd", &mha_fwd, "Forward pass");
-    m.def("fwd_sparse", &mha_fwd_sparse, "Block-sparse forward pass (non-causal)");
 }

fastvideo-kernel/attn_qat_infer/blackwell/kernel_ws.h

@@ -185,14 +185,14 @@ __global__ void __launch_bounds__(Ktraits::kNWarps * cutlass::NumThreadsPerWarp,
             auto block_coord = work_tile_info.get_block_coord(scheduler_params);
             auto [m_block, bidh, bidb] = block_coord;
 
-            int n_block_count = collective_mainloop.get_n_block_count(mainloop_params, m_block, bidh, bidb);
-            if (Is_causal && n_block_count <= 0) {  // We exit early and write 0 to gO and -inf to gLSE.
+            int n_block_max = collective_mainloop.get_n_block_max(mainloop_params, m_block);
+            if (Is_causal && n_block_max <= 0) {  // We exit early and write 0 to gO and -inf to gLSE.
                 collective_epilogue.store_zero(epilogue_params, threadIdx.x - NumCopyThreads, block_coord);
                 continue;
             }
 
             collective_mainloop.mma(mainloop_params, pipeline_q, pipeline_k, pipeline_v, smem_pipe_read_q, smem_pipe_read_k, smem_pipe_read_v,
-                                    tOrO, softmax_fused, n_block_count, threadIdx.x - NumCopyThreads, work_idx, m_block, bidh, bidb, shared_storage);
+                                    tOrO, softmax_fused, n_block_max, threadIdx.x - NumCopyThreads, work_idx, m_block, shared_storage);
             barrier_o.wait();
             collective_epilogue.mma_store(shared_storage, tiled_mma_pv, tOrO, threadIdx.x - NumCopyThreads); 
             barrier_o.arrive();

fastvideo-kernel/attn_qat_infer/blackwell/launch.h

@@ -62,9 +62,7 @@ void run_flash_fwd(Flash_fwd_params &params, cudaStream_t stream) {
             static_cast<float const*>(params.delta_s_ptr),
             {params.seqlen_s, params.seqlen_k, params.h_k, params.b},
             {params.ds_row_stride, _1{}, params.ds_head_stride, params.ds_batch_stride},
-            params.scale_softmax_log2,
-            params.q2k_idx, params.q2k_num, params.q2k_max,
-            params.num_m_blocks, params.h, params.kv_valid, params.q2k_quad
+            params.scale_softmax_log2
         });
     typename CollectiveEpilogue::Params epilogue_params =
         CollectiveEpilogue::to_underlying_arguments({

fastvideo-kernel/attn_qat_infer/blackwell/mainloop_tma_ws.h

@@ -174,13 +174,6 @@ struct CollectiveMainloopFwd {
         ShapeQKV const shape_ds;
         StrideQKV const stride_ds;
         float const softmax_scale_log2;
-        int const* ptr_q2k_idx{nullptr};
-        int const* ptr_q2k_num{nullptr};
-        int q2k_max{0};
-        int num_m_blocks{0};
-        int num_heads{0};
-        int const* ptr_kv_valid{nullptr};
-        uint8_t const* ptr_q2k_quad{nullptr};
     };
 
     // Device side kernel params
@@ -201,13 +194,6 @@ struct CollectiveMainloopFwd {
         TMA_SFVt tma_load_SFVt;
         TMA_DS tma_load_DS;
         float const softmax_scale_log2;
-        int const* ptr_q2k_idx;
-        int const* ptr_q2k_num;
-        int q2k_max;
-        int num_m_blocks;
-        int num_heads;
-        int const* ptr_kv_valid;
-        uint8_t const* ptr_q2k_quad;
     };
 
 
@@ -275,9 +261,7 @@ struct CollectiveMainloopFwd {
                 tma_load_K, tma_load_sfk,
                 tma_load_Vt, tma_load_sfvt,
                 tma_load_ds, 
-                args.softmax_scale_log2,
-                args.ptr_q2k_idx, args.ptr_q2k_num, args.q2k_max,
-                args.num_m_blocks, args.num_heads, args.ptr_kv_valid, args.ptr_q2k_quad};
+                args.softmax_scale_log2};
     }
 
     /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance
@@ -306,51 +290,6 @@ struct CollectiveMainloopFwd {
         return n_block_max;
     }
 
-    // Number of KV blocks query block m_block visits: all of them when dense,
-    // its index-list length when block-sparse.
-    CUTLASS_DEVICE
-    int get_n_block_count(Params const& mainloop_params, int m_block, int bidh, int bidb) {
-        if (mainloop_params.ptr_q2k_idx == nullptr) {
-            return get_n_block_max(mainloop_params, m_block);
-        }
-        return mainloop_params.ptr_q2k_num[(bidb * mainloop_params.num_heads + bidh) * mainloop_params.num_m_blocks + m_block];
-    }
-
-    // KV block visited at iteration i (iterations run from count-1 down to 0).
-    CUTLASS_DEVICE
-    int get_kv_block(Params const& mainloop_params, int m_block, int bidh, int bidb, int i) {
-        if (mainloop_params.ptr_q2k_idx == nullptr) {
-            return i;
-        }
-        int64_t const row = (int64_t(bidb) * mainloop_params.num_heads + bidh) * mainloop_params.num_m_blocks + m_block;
-        return mainloop_params.ptr_q2k_idx[row * mainloop_params.q2k_max + i];
-    }
-
-    // Valid key columns of each 64-column half of KV block n_block (valid tokens
-    // first within a half). kv_valid, when given, stores two counts per block so
-    // 64-token VSA tiles can pad mid-block; otherwise only the sequence tail is
-    // masked.
-    CUTLASS_DEVICE
-    int get_kv_valid_half(Params const& mainloop_params, int n_block, int half, int unpadded_seqlen_k) {
-        static constexpr int kBlockN = get<1>(TileShape_MNK{});
-        static constexpr int kHalfN = kBlockN / 2;
-        if (mainloop_params.ptr_kv_valid != nullptr) {
-            return mainloop_params.ptr_kv_valid[2 * n_block + half];
-        }
-        return max(0, min(kHalfN, unpadded_seqlen_k - n_block * kBlockN - half * kHalfN));
-    }
-
-    // Quadrant mask of list entry i: bit (2 * row_half + col_half) is set when the
-    // 64-row query half attends the 64-column key half. 0xF when absent.
-    CUTLASS_DEVICE
-    int get_quad(Params const& mainloop_params, int m_block, int bidh, int bidb, int i) {
-        if (mainloop_params.ptr_q2k_quad == nullptr) {
-            return 0xF;
-        }
-        int64_t const row = (int64_t(bidb) * mainloop_params.num_heads + bidh) * mainloop_params.num_m_blocks + m_block;
-        return mainloop_params.ptr_q2k_quad[row * mainloop_params.q2k_max + i];
-    }
-
     template <class SFATensor, class Atom, class TiledThr, class TiledPerm>
     CUTE_HOST_DEVICE constexpr
     auto
@@ -511,7 +450,7 @@ struct CollectiveMainloopFwd {
 
         auto [m_block, bidh, bidb] = work_tile_info.get_block_coord(scheduler_params);
 
-        int n_block_count = get_n_block_count(mainloop_params, m_block, bidh, bidb);
+        int n_block_max = get_n_block_max(mainloop_params, m_block);
 
         Tensor sQ = make_tensor(make_smem_ptr(shared_storage.smem_q.begin()), SmemLayoutQ{});
         Tensor sK = make_tensor(make_smem_ptr(shared_storage.smem_k.begin()), SmemLayoutK{});
@@ -567,8 +506,7 @@ struct CollectiveMainloopFwd {
         Tensor tDSsDS = group_modes<0, 3>(block_tma_ds.partition_D(sDS));
         uint16_t mcast_mask_kv = 0;
 
-        int n_iter = n_block_count - 1;
-        int n_block = get_kv_block(mainloop_params, m_block, bidh, bidb, n_iter);
+        int n_block = n_block_max - 1;
         int lane_predicate = cute::elect_one_sync();
         if (lane_predicate) {
         pipeline_q.producer_acquire(smem_pipe_write_q);
@@ -591,12 +529,11 @@ struct CollectiveMainloopFwd {
         ++smem_pipe_write_v;
         }
 
-        --n_iter;
+        n_block--;
         if (lane_predicate) {
             // CUTLASS_PRAGMA_NO_UNROLL
             #pragma unroll 2
-            for (; n_iter >= 0; --n_iter) {
-                n_block = get_kv_block(mainloop_params, m_block, bidh, bidb, n_iter);
+            for (; n_block >= 0; --n_block) {
                 pipeline_k.producer_acquire(smem_pipe_write_k);
                 copy(mainloop_params.tma_load_K.with(*pipeline_k.producer_get_barrier(smem_pipe_write_k), mcast_mask_kv),
                     tKgK(_, n_block), tKsK(_, smem_pipe_write_k.index()));
@@ -648,8 +585,6 @@ struct CollectiveMainloopFwd {
         int thread_idx,
         int work_idx,
         int m_block,
-        int bidh,
-        int bidb,
         SharedStorage& shared_storage
         ) {
 
@@ -732,21 +667,7 @@ struct CollectiveMainloopFwd {
         int const seqlen_q = get<0>(mainloop_params.shape_Q);
         int const seqlen_k = get<0>(mainloop_params.shape_K);
         int const unpadded_seqlen_k = get<0>(mainloop_params.unpadded_shape_K);
-        int n_iter = n_block_count - 1;
-        int n_block = get_kv_block(mainloop_params, m_block, bidh, bidb, n_iter);
-        bool const per_block_masking = mainloop_params.ptr_q2k_idx != nullptr || mainloop_params.ptr_kv_valid != nullptr;
-        static_assert(kBlockM == 128 && kBlockN == 128, "quadrant masking assumes 128x128 blocks");
-        // Each MMA warp owns 16 consecutive query rows, so a warp lies in one
-        // 64-row half. With quadrant lists a warp skips blocks its half did not
-        // select and the P.V chunk of a key half it did not select: masked
-        // scores contribute exactly zero, and the warp sharing its tensor-core
-        // partition runs faster meanwhile.
-        int const my_row_half = __shfl_sync(0xffffffff, [&] {
-            Tensor cS0 = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{}));
-            Tensor tScS0 = thread_mma_qk.partition_C(cS0);
-            return int(get<0>(tScS0(0))) >= kBlockM / 2 ? 1 : 0;
-        }(), 0);
-        bool const quad_skip = mainloop_params.ptr_q2k_quad != nullptr;
+        int n_block = n_block_count - 1;
 
         auto copy_k_block = [&](auto block_id) {
             auto tSsK_stage = tSsK(_, _, _, smem_pipe_read_k.index());
@@ -823,26 +744,6 @@ struct CollectiveMainloopFwd {
             int local = c & 31;
             return (c & ~31) | (local & 1) | ((local & 24) >> 2) | ((local & 6) << 2);
         };
-        // Sparse lists may visit a block whose 64-token halves are partially valid
-        // (VSA tile tails, short prefix chunks) or that only one query half
-        // selected (64-token tiles), at any iteration.
-        auto apply_sparse_mask = [&](auto& acc, int n_blk, int it) {
-            int const quad = get_quad(mainloop_params, m_block, bidh, bidb, it);
-            int const valid0 = get_kv_valid_half(mainloop_params, n_blk, 0, unpadded_seqlen_k);
-            int const valid1 = get_kv_valid_half(mainloop_params, n_blk, 1, unpadded_seqlen_k);
-            if (quad == 0xF && valid0 == kBlockN / 2 && valid1 == kBlockN / 2) { return; }
-            Tensor cS = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{}));
-            Tensor tScS = thread_mma_qk.partition_C(cS);
-            CUTLASS_PRAGMA_UNROLL
-            for (int i = 0; i < size(acc); ++i) {
-                int const col = actual_col(int(get<1>(tScS(i))));
-                int const col_half = col >= kBlockN / 2;
-                int const row_half = int(get<0>(tScS(i))) >= kBlockM / 2;
-                bool const keep = ((quad >> (2 * row_half + col_half)) & 1)
-                                  && (col & (kBlockN / 2 - 1)) < (col_half ? valid1 : valid0);
-                if (!keep) { acc(i) = -INFINITY; }
-            }
-        };
         {
             Tensor cS = cute::make_identity_tensor(select<0, 1>(TileShape_MNK{}));
             Tensor tScS = thread_mma_qk.partition_C(cS);
@@ -850,9 +751,7 @@ struct CollectiveMainloopFwd {
             for (int i = 0; i < size(tSrS); ++i) {
                 int col = actual_col(int(get<1>(tScS(i))));
                 if constexpr (!Is_causal) {  // Just masking based on col
-                    if (!per_block_masking) {
-                        if (col >= int(unpadded_seqlen_k - n_block * kBlockN)) { tSrS(i) = -INFINITY; }
-                    }
+                    if (col >= int(unpadded_seqlen_k - n_block * kBlockN)) { tSrS(i) = -INFINITY; }
                 } else {
                     if (col >= std::min(seqlen_k - n_block * kBlockN,
                                        col_limit_causal(int(get<0>(tScS(i))), n_block))) {
@@ -861,9 +760,6 @@ struct CollectiveMainloopFwd {
                 }
             }
         }
-        if constexpr (!Is_causal) {
-            if (per_block_masking) { apply_sparse_mask(tSrS, n_block, n_iter); }
-        }
         auto quantize = [&](auto mma_k, auto acc_conversion_view) {
             Tensor AbsMaxP_stagek = AbsMaxP(_, make_coord(_, _, mma_k));
             Tensor acc_conversion_stagek = acc_conversion_view(_, _, mma_k);
@@ -931,12 +827,11 @@ struct CollectiveMainloopFwd {
             }
         }
         
-        --n_iter;
+        n_block--;
         constexpr int n_masking_steps = !Is_causal ? 1 : cute::ceil_div(kBlockM, kBlockN) + 1;
         // // Only go through these if Is_causal, since n_masking_steps = 1 when !Is_causal
         CUTLASS_PRAGMA_UNROLL
-        for (int masking_step = 0; masking_step < n_masking_steps - 1 && n_iter >= 0; ++masking_step, --n_iter) {
-            n_block = get_kv_block(mainloop_params, m_block, bidh, bidb, n_iter);
+        for (int masking_step = 0; masking_step < n_masking_steps - 1 && n_block >= 0; ++masking_step, --n_block) {
             Tensor tSrS = partition_fragment_C(tiled_mma_qk, select<0, 1>(TileShape_MNK{}));
             Tensor tSrS_converion_view = make_tensor(tSrS.data(), flash::convert_to_conversion_layout(tSrS.layout()));
             consumer_wait(pipeline_k, smem_pipe_read_k);
@@ -983,22 +878,7 @@ struct CollectiveMainloopFwd {
         }
 
         #pragma unroll 1
-        for (; n_iter >= 0; --n_iter) {
-            n_block = get_kv_block(mainloop_params, m_block, bidh, bidb, n_iter);
-            int const row_bits = quad_skip
-                                 ? (get_quad(mainloop_params, m_block, bidh, bidb, n_iter) >> (2 * my_row_half)) & 3
-                                 : 3;
-            if (row_bits == 0) {
-                // Neither key half is selected for this warp's rows: its softmax
-                // state and output are unchanged. Keep the pipeline in step.
-                consumer_wait(pipeline_k, smem_pipe_read_k);
-                pipeline_k.consumer_release(smem_pipe_read_k);
-                ++smem_pipe_read_k;
-                consumer_wait(pipeline_v, smem_pipe_read_v);
-                pipeline_v.consumer_release(smem_pipe_read_v);
-                ++smem_pipe_read_v;
-                continue;
-            }
+        for (; n_block >= 0; --n_block) {
             Tensor tSrS = partition_fragment_C(tiled_mma_qk, select<0, 1>(TileShape_MNK{}));
             Tensor tSrS_converion_view = make_tensor(tSrS.data(), flash::convert_to_conversion_layout(tSrS.layout()));
             consumer_wait(pipeline_k, smem_pipe_read_k);
@@ -1016,21 +896,16 @@ struct CollectiveMainloopFwd {
                 }
             }
 
-            if (per_block_masking) { apply_sparse_mask(tSrS, n_block, n_iter); }
 
             softmax_fused.template online_softmax_with_quant</*Is_first=*/false>(tSrS, AbsMaxP, mainloop_params.softmax_scale_log2);
             Tensor tOrO = make_fragment_like(tOrO_store);
-            clear(tOrO);
             consumer_wait(pipeline_v, smem_pipe_read_v);
             copy_v_block(_0{});
             quantize(_0{}, tSrS_converion_view);
             CUTLASS_PRAGMA_UNROLL
             for (int v_block = 0; v_block < size<2>(tOrP); ++v_block) {
-                // v_block spans one 64-column key half (P's K mode is 2 x 64).
-                if ((row_bits >> v_block) & 1) {
-                    cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, v_block), tOrSFP(_, _, v_block)),
-                                        make_zip_tensor(tOrVt(_, _, v_block), tOrSFVt(_, _, v_block)), tOrO);
-                }
+                cute::gemm(tiled_mma_pv, make_zip_tensor(tOrP(_, _, v_block), tOrSFP(_, _, v_block)),
+                                    make_zip_tensor(tOrVt(_, _, v_block), tOrSFVt(_, _, v_block)), tOrO);
                 if (v_block < size<2>(tOrP) - 1) {
                     copy_v_block(v_block + 1);
                     quantize(v_block + 1, tSrS_converion_view);

fastvideo-kernel/attn_qat_infer/blackwell/params.h

@@ -114,23 +114,6 @@
      int * __restrict__ seqused_k;
  
      int *__restrict__ blockmask;
-
-     // Block-sparse KV iteration (non-causal only). When q2k_idx is null the
-     // kernel is dense. Otherwise query block m of (batch b, head h) visits
-     // the q2k_num[(b*h + h)*num_m_blocks + m] KV blocks listed at
-     // q2k_idx[((b*h + h)*num_m_blocks + m)*q2k_max + i], in any order.
-     int const *__restrict__ q2k_idx;
-     int const *__restrict__ q2k_num;
-     int q2k_max;
-     int num_m_blocks;
-     // Optional valid token count of each 64-column half of every KV block,
-     // kv_valid[2*n + half] (valid tokens first within a half); null means only
-     // the sequence tail beyond the unpadded key length is masked.
-     int const *__restrict__ kv_valid;
-     // Optional quadrant mask per list entry (same layout as q2k_idx): bit
-     // (2*row_half + col_half) set when that 64-row query half attends that
-     // 64-column key half. Lets 64-token VSA tiles run on 128x128 blocks.
-     uint8_t const *__restrict__ q2k_quad;
  
      // The K_new and V_new matrices.
      void * __restrict__ knew_ptr;

fastvideo-kernel/csrc/attention/block_sparse_kernel_sm100a.cuh

@@ -2,6 +2,8 @@
 // data-center Blackwell sm_100a/sm_103a. The filename is retained for API compatibility.
 // Warp-specialized: load / MMA (tcgen05) / softmax / correction / epilogue / scheduler.
 // Writes O and, when asked, the log-sum-exp the backward consumes.
+//
+// Generated (comments stripped). Do not edit by hand.
 #ifndef BLOCK_SPARSE_VSA_KERNEL_SM100A_CUH
 #define BLOCK_SPARSE_VSA_KERNEL_SM100A_CUH
 
@@ -504,11 +506,9 @@ fmha_context_bf16_gen_kernel(const __grid_constant__ CUtensorMap tmap_q,
         const uint32_t o_tmem_addr = tmem_base + (uint32_t)(2 * S_COLS + i * O_COLS);
 
         int slot = 0;
-        // Declared outside the loop: with BLK128 only p == 0 initializes it and p == 1 continues
-        // the p == 0 descriptor walk.
-        SmemDescPair desc_bv;
         #pragma unroll
         for (int p = 0; p < V_SUBTILES; ++p) {
+        SmemDescPair desc_bv;
           if (!BLK128 || p == 0) {
             slot = kv_ph.get_stage();
             mbarrier_wait_parity(smem_ptr_u32(&full_bar[slot]), kv_ph.get_phase());
@@ -518,6 +518,8 @@ fmha_context_bf16_gen_kernel(const __grid_constant__ CUtensorMap tmap_q,
               desc_bv.u64 = desc_v0;
               desc_bv.w.x += (uint32_t)(slot * (int)KV_DESC_DELTA);
             }
+          const uint64_t dbV = desc_kv0 + (uint64_t)slot * KV_DESC_DELTA
+                             + (BLK128 ? (uint64_t)p * (V_BLK_BYTES >> 4) : 0u);
           #pragma unroll
           for (int ki = 0; ki < K_ATOMS_PER_TILE; ++ki) {
             const int a = p * K_ATOMS_PER_TILE + ki;

fastvideo-kernel/csrc/attention/primitives.cuh

@@ -1,6 +1,8 @@
 // primitives.cuh -- device primitives for the sm_100a/sm_103a VSA block-sparse attention
 // forward: tcgen05 (alloc / mma / ld / st / commit / wait / fence), TMA load / store /
 // tensormap, mbarrier, cluster launch control, setmaxnreg, fast math, and the FMHA helpers.
+//
+// Generated and pruned to what the kernel reaches -- do not edit by hand.
 #pragma once
 #include <cstdint>
 #include <cstdio>
@@ -216,8 +218,8 @@ template <int CLUSTER_SHAPE_M, int CLUSTER_SHAPE_N, ClcRasterOrder ORDER>
 __device__ __forceinline__
 ClcTileInfo clc_parse_response(uint32_t resp_smem_addr) {
   uint32_t d0, d1, d2, d3;
-  clc_load_response(resp_smem_addr, d0, d1, d2, d3);
   fence_proxy_async_shared_cta();
+  clc_load_response(resp_smem_addr, d0, d1, d2, d3);
   const int  ctaid_x = static_cast<int>(d0);
   const int  ctaid_y = static_cast<int>(d1 & 0xFFFFu);
   const bool valid   = (d2 & 1u) != 0u;
@@ -249,7 +251,6 @@ ClcTileInfo clc_fetch_next_tile(
       __cvta_generic_to_shared(&clc_response[clc_cons_stage * 4]));
   ClcTileInfo t = clc_parse_response<
       CLUSTER_SHAPE_M, CLUSTER_SHAPE_N, ORDER>(resp_addr);
-  __syncwarp();
   if (do_release) {
     uint32_t empty_local = static_cast<uint32_t>(
         __cvta_generic_to_shared(&clc_empty_bar[clc_cons_stage]));