[refactor]: group model-family settings under pipeline.model

pipeline.model is a tagged union: a mapping with exactly one key, which names the model family (ltx2, minimax_h3, longcat, or generic) and selects that family's options class. The former sibling sections pipeline.ltx2, pipeline.minimax_h3, and pipeline.longcat, and the DiT/VAE architecture override dicts pipeline.dit and pipeline.vae, live inside the family block. The parser implements tagged unions as a general mechanism (a dataclass with a TAGS table); the family is readable as pipeline.model.family. Resolution validates the tag against the family that the registry chose for model_path (validate_model_family raises and names the registry class and the right block) and fills the family's empty block when the input has none (fill_model_family), so readers never see None. Family-only fills and derivations (the MiniMax-H3 parallel-VAE environment variables, the LTX-2 tile-size derivation and refine preset copy, the LongCat BSA mirrors) run only for their family. Against the 3f6893a0 goldens every PipelineConfig is identical and no other field differs beyond the allowed categories; family-only settings are no longer decided on models of another family. LingBot-Video reads its refiner switch from pipeline.preset_overrides.refine.enabled and loads the refiner by default when the checkpoint has one, as at 3f6893a0. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CbSuTKA9oX4YsECdcBPAsX
[refactor]: decide every config value before the config freezes, and drop the from_pretrained keywords
2026-10-05 06:04:17 +00:00 · 2026-10-04 07:14:31 +00:00 · 2026-10-04 04:45:02 +00:00 · 2026-10-03 06:46:52 +00:00 · 2026-10-03 06:35:19 +00:00 · 2026-10-03 06:28:44 +00:00
1124 changed files with 123633 additions and 24524 deletions
@@ -80,26 +80,36 @@ def _run_fastvideo_pipeline(model_path: Path, params: dict[str, Any]) -> Any:

    generator = VideoGenerator.from_pretrained(
        str(model_path),
-        num_gpus=1,
-        use_fsdp_inference=False,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=False,
+        {
+            "engine": {
+                "num_gpus": 1,
+                "use_fsdp_inference": False,
+                "offload": {
+                    "dit": False,
+                    "vae": False,
+                    "text_encoder": False,
+                },
+            },
+        },
    )
    try:
-        return generator.generate_video(
-            prompt=params["prompt"],
-            negative_prompt=params.get("negative_prompt"),
-            output_path=f"outputs_{_MODEL_FAMILY}/pipeline_parity",
-            save_video=False,
-            height=params.get("height"),
-            width=params.get("width"),
-            num_frames=params.get("num_frames"),
-            fps=params.get("fps"),
-            num_inference_steps=params["num_inference_steps"],
-            guidance_scale=params.get("guidance_scale"),
-            seed=params["seed"],
-        )
+        return generator.generate({
+            "prompt": params["prompt"],
+            "negative_prompt": params.get("negative_prompt"),
+            "sampling": {
+                "height": params.get("height"),
+                "width": params.get("width"),
+                "num_frames": params.get("num_frames"),
+                "fps": params.get("fps"),
+                "num_inference_steps": params["num_inference_steps"],
+                "guidance_scale": params.get("guidance_scale"),
+                "seed": params["seed"],
+            },
+            "output": {
+                "output_path": f"outputs_{_MODEL_FAMILY}/pipeline_parity",
+                "save_video": False,
+            },
+        })
    finally:
        generator.shutdown()

@@ -84,7 +84,9 @@ fastvideo/configs/models/dits/__init__.py
 fastvideo/configs/models/encoders/__init__.py
 fastvideo/configs/models/vaes/__init__.py
 fastvideo/envs.py
-fastvideo/fastvideo_args.py
+fastvideo/api/schema.py
+fastvideo/api/resolution.py
+fastvideo/api/inference_resolution.py
 fastvideo/distributed/**
 fastvideo/layers/**
 fastvideo/attention/**
@@ -0,0 +1,80 @@
+---
+name: env-var-conventions
+description: Add, read, rename, or remove an environment variable in FastVideo, or change the environment-variable policy. Use before touching fastvideo/envs.py, os.environ, os.getenv, or monkeypatch.setenv in fastvideo/, and when fastvideo/tests/contract/test_env_policy.py fails.
+---
+
+# Environment Variable Conventions
+
+## Purpose
+
+FastVideo registers its environment variables as typed fields in
+`fastvideo/envs.py`. The policy that governs them is
+`docs/contributing/env_vars.md`, and the contract test
+`fastvideo/tests/contract/test_env_policy.py` enforces the policy in the unit
+CI lane. This skill routes an environment-variable change through that policy.
+The policy doc is the single source of the rules; read it instead of relying
+on a summary here.
+
+## Prerequisites
+
+- Read `docs/contributing/env_vars.md` in full.
+- Decide whether the setting belongs in an environment variable or an argument
+  (rule 5 in the policy doc). Settings that users change per deployment are
+  arguments; add them as typed config fields in `fastvideo/api/schema.py`
+  instead.
+
+## Inputs
+
+| Parameter  | Required | Description                                                    |
+| ---------- | -------- | -------------------------------------------------------------- |
+| `change`   | Yes      | Add, read, rename, or remove a variable, or change the policy. |
+| `variable` | Yes      | The variable name, with the `FASTVIDEO_` prefix.               |
+
+## Steps
+
+1. **Declare or edit the variable in `fastvideo/envs.py`.**
+   - Pick the field type and category that the policy doc lists.
+   - Write a description that states what the variable does and its units.
+   - To rename, keep the old name in `deprecated_names`. To remove, add the
+     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_*`.
+   - 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`,
+     `envs.setdefault_external`, or `envs.unset_external`.
+3. **Regenerate the table in the policy doc.**
+   - Run `python fastvideo/tests/contract/test_env_policy.py`.
+4. **Run the contract test.**
+   - Run `pytest fastvideo/tests/contract/test_env_policy.py`.
+   - When the test reports a fixed known violation, delete or lower its entry
+     in `KNOWN_VIOLATIONS`. Never add an entry to `KNOWN_VIOLATIONS`.
+5. **When the policy itself changes, update the policy doc and the contract
+   test in the same pull request.**
+   - The rules in `docs/contributing/env_vars.md`, the checks and allowlist in
+     `fastvideo/tests/contract/test_env_policy.py`, and this skill must agree.
+
+## Outputs
+
+- A registry entry in `fastvideo/envs.py` and call sites that use
+  `envs.NAME.get()`.
+- A regenerated table in `docs/contributing/env_vars.md`.
+- A passing `fastvideo/tests/contract/test_env_policy.py`.
+
+## Example Usage
+
+```
+Add a FASTVIDEO_DEBUG_MY_STAGE switch that logs MyStage inputs.
+```
+
+## References
+
+- `docs/contributing/env_vars.md`: the policy, the field types, and the
+  violation kinds that the contract test reports.
+- `fastvideo/envs.py`: the registry.
+- `fastvideo/tests/contract/test_env_policy.py`: the contract test,
+  `EXTERNAL_ALLOWLIST`, and `KNOWN_VIOLATIONS`.
@@ -57,7 +57,7 @@ Hardcoded:
 - Quality tier: **`default`**. `full_quality` is a separate, deliberate
  operation.
 - HF repo: `FastVideo/ssim-reference-videos` (override via
-  `FASTVIDEO_SSIM_REFERENCE_HF_REPO`).
+  `FASTVIDEO_TEST_SSIM_REFERENCE_HF_REPO`).
 - Device folder: `L40S_reference_videos`.

 ## Prerequisites
@@ -185,6 +185,7 @@ steps:

  - label: ":bar_chart: SSIM Tests"
    key: "ssim"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      build.env("TEST_SCOPE") == "scheduled" ||
@@ -211,6 +212,7 @@ steps:

  - label: ":test_tube: LoRA Inference Tests"
    key: "lora-inference"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -232,6 +234,7 @@ steps:

  - label: ":test_tube: LoRA Extraction Tests"
    key: "lora-extraction"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -253,6 +256,7 @@ steps:

  - label: ":test_tube: Training Tests"
    key: "training"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -276,6 +280,7 @@ steps:

  - label: ":test_tube: Distillation DMD Tests"
    key: "distillation"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -297,6 +302,7 @@ steps:

  - label: ":test_tube: Self-Forcing Tests"
    key: "self-forcing"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -318,6 +324,7 @@ steps:

  - label: ":test_tube: LoRA Training Tests"
    key: "lora-training"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -341,6 +348,7 @@ steps:

  - label: ":test_tube: Training Tests VSA"
    key: "training-vsa"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -364,6 +372,7 @@ steps:

  - label: ":test_tube: Inference Tests VMoBA"
    key: "inference-vmoba"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -385,6 +394,7 @@ steps:

  - label: ":test_tube: Performance Tests"
    key: "performance"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -406,6 +416,7 @@ steps:

  - label: ":test_tube: API Server Tests"
    key: "api-server"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -427,6 +438,7 @@ steps:

  - label: ":test_tube: Train Framework Tests"
    key: "train-framework"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -448,6 +460,7 @@ steps:

  - label: ":test_tube: Eval Metrics Tests"
    key: "eval"
+    depends_on: "golden-gate"
    if: |
      build.env("TEST_SCOPE") == "full" ||
      (build.env("TEST_SCOPE") == "merge" &&
@@ -13,7 +13,7 @@ if [ -z "$selected" ]; then
  selected=all
 fi
 if [ "$selected" = all ]; then
-  exec pytest "$golden_root" -vs
+  exec pytest "$golden_root" -xvs
 fi

 [[ $selected =~ ^test_[a-z0-9_]+\.py(,test_[a-z0-9_]+\.py)*$ ]] || {
@@ -32,4 +32,4 @@ for golden_file in "${golden_files[@]}"; do
  golden_paths+=("$golden_path")
 done

-exec pytest "${golden_paths[@]}" -vs
+exec pytest "${golden_paths[@]}" -xvs
@@ -2,4 +2,8 @@
 # Canonical Slurm CI selection for the transformer lane.
 set -euo pipefail

+# The existing block reference records an absent FASTVIDEO_FA4 (FA2). Keep
+# that reference identity; the component lane also selects FA2 explicitly.
+env -u FASTVIDEO_FA4 pytest ./fastvideo/tests/golden_gate/test_wan_t2v.py -xvs
+pytest ./fastvideo/tests/golden_gate/test_wan_causal.py -xvs
 exec pytest ./fastvideo/tests/transformers -vs
@@ -2,4 +2,5 @@
 # Canonical Slurm CI selection for the VAE lane.
 set -euo pipefail

+pytest ./fastvideo/tests/golden_gate/test_wan_vae.py -xvs
 exec pytest ./fastvideo/tests/vaes -vs
@@ -16,6 +16,7 @@ exec pytest \
  ./fastvideo/tests/worker/ \
  ./fastvideo/tests/training/test_trackers.py \
  ./fastvideo/tests/attention/test_sdpa_metadata_mask_contract.py \
+  ./fastvideo/tests/attention/test_vsa_h3_tile_grad_safety.py \
  ./fastvideo/tests/modal/test_kernel_build_cache.py \
  ./fastvideo/tests/modal/test_pr_test.py \
  ./fastvideo/tests/modal/test_ssim_test.py \
@@ -211,8 +211,8 @@ FAMILY_COVERAGE = (
        ("test_turbodiffusion_similarity.py", ),
    ),
    FamilyCoverage(
-        re.compile(r"(^|[/_.-])wan(video)?([/_.-]|$)"),
-        ("test_wan_t2v.py", ),
+        re.compile(r"(^|[/_.-])wan(video|vae)?([/_.-]|$)"),
+        ("test_wan_t2v.py", "test_wan_vae.py", "test_wan_causal.py", "test_wan_denoising.py"),
        (
            "test_causal_similarity.py",
            "test_wan_i2v_similarity.py",
@@ -291,6 +291,18 @@ def _family_coverage(path: str) -> tuple[set[str], set[str]]:
        if family.pattern.search(normalized):
            golden.update(family.golden_tests)
            ssim.update(family.ssim_tests)
+    # Select the component actually touched, including compatibility paths.
+    # Family configs/pipeline wiring can affect all four Wan gates.
+    if re.search(r"(^|[/_.-])wan(video|vae)?([/_.-]|$)", normalized):
+        if (normalized.endswith(("/wan/vae.py", "/wan/vae_config.py", "/vaes/wanvae.py"))
+                or normalized.endswith("/wan/stages/conditioning.py")):
+            golden = {"test_wan_vae.py"}
+        elif normalized.endswith(("/wan/causal_transformer.py", "/dits/causal_wanvideo.py",
+                                  "/wan/stages/causal_denoising.py")):
+            golden = {"test_wan_causal.py"}
+        elif (normalized == "fastvideo/models/dits/wanvideo.py"
+              or normalized.endswith(("/wan/transformer.py", "/wan/stages/denoising.py", "/wan/stages/dmd.py"))):
+            golden = {"test_wan_t2v.py", "test_wan_denoising.py"}
    return golden, ssim


@@ -477,7 +489,6 @@ def classify_paths(paths: list[str]) -> MergePlan:
            _select_output_coverage(plan, path)
            continue
        if path in {
-                "fastvideo/fastvideo_args.py",
                "fastvideo/forward_context.py",
                "fastvideo/image_processor.py",
                "fastvideo/registry.py",
@@ -92,6 +92,7 @@ jobs:
            fastvideo/tests/mlx/test_frame_upsample.py \
            fastvideo/tests/mlx/test_mlx_fast_spatial.py \
            fastvideo/tests/mlx/test_mlx_refine.py \
+            fastvideo/tests/mlx/test_mlx_prompt_enhance.py \
            fastvideo/tests/mlx/test_mlx_prompt_to_video_decode.py \
            fastvideo/tests/mlx/test_mlx_wan22_prompt_cache_fingerprint.py \
            fastvideo/tests/mlx/test_wan22_sample.py \
@@ -156,6 +157,7 @@ jobs:
            fastvideo/tests/mlx/test_frame_upsample.py \
            fastvideo/tests/mlx/test_mlx_fast_spatial.py \
            fastvideo/tests/mlx/test_mlx_refine.py \
+            fastvideo/tests/mlx/test_mlx_prompt_enhance.py \
            fastvideo/tests/mlx/test_mlx_prompt_to_video_decode.py \
            fastvideo/tests/mlx/test_mlx_wan22_prompt_cache_fingerprint.py \
            fastvideo/tests/mlx/test_wan22_sample.py \
@@ -63,7 +63,7 @@ jobs:
              torch-cuda-short: 'cu130'
          platform:
            # x86_64 builds the full cu126 + cu130 set. cu130 ships the
-            # data-center Blackwell sm_100a VSA and consumer sm_120a FP4
+            # data-center Blackwell sm_100a/sm_103a VSA and consumer sm_120a FP4
            # kernels.
            - os: ubuntu-22.04
              arch: x86_64
@@ -169,18 +169,18 @@ jobs:
          #     covers sm_120a; turbodiffusion covers sm_100a+sm_120a. The sm_100 FP4
          #     forward is the FA4 CuTe DSL path in the fastvideo package (PR #1221),
          #     JIT-compiled at runtime — not built into this wheel.
-          #   * x86_64 cu130 = Hopper TK + data-center Blackwell sm_100a VSA
+          #   * x86_64 cu130 = Hopper TK + data-center Blackwell sm_100a/sm_103a VSA
          #     + consumer Blackwell sm_120a FP4.
          #   * x86_64 cu126 = Hopper TK only (older drivers; CUDA < 12.8 has no FP4).
          # The per-arch split in CMakeLists pins the FP4 targets to sm_120a and builds
          # the main extension for the full arch list. CMAKE_BUILD_PARALLEL_LEVEL caps
          # Ninja so heavy CUTLASS/TK template TUs don't OOM the 16 GB runner (exit 143).
          if [ "${{ matrix.platform.arch }}" = "aarch64" ]; then
-            export TORCH_CUDA_ARCH_LIST="10.0a;12.0a"
+            export TORCH_CUDA_ARCH_LIST="10.0a;10.3a;12.0a"
            export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=OFF -DFASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER=ON"
            export CMAKE_BUILD_PARALLEL_LEVEL=1
          elif [ "${{ matrix.torch-cuda.torch-cuda-short }}" = "cu130" ]; then
-            export TORCH_CUDA_ARCH_LIST="9.0a;10.0a;12.0a"
+            export TORCH_CUDA_ARCH_LIST="9.0a;10.0a;10.3a;12.0a"
            export CMAKE_ARGS="${CMAKE_ARGS:-} -DFASTVIDEO_KERNEL_BUILD_TK=ON -DFASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER=ON -DCMAKE_CUDA_ARCHITECTURES=90a"
            # A single FP4 TU (attn_qat_infer) can use ~8-12 GB on its own, so serialize.
            export CMAKE_BUILD_PARALLEL_LEVEL=1
@@ -9,7 +9,7 @@ exclude: |
    tests/.*|
    scripts/.*|
    fastvideo/dataset/.*|
-    fastvideo/models/.*|
+    fastvideo/models/(?!wan/(config|vae_config|pipeline_config|definition|__init__)\.py$).*|
    ^apps/dreamverse/web/.*|
    examples/.*|
    \.agents/.*|
@@ -66,14 +66,18 @@ Local guidance lives next to the code. Read the in-scope file before editing:
 | `fastvideo/AGENTS.md` | Core package map, public API, registry-driven model dispatch |
 | `fastvideo/configs/AGENTS.md` | Arch + pipeline config dataclasses, `param_names_mapping` |
 | `fastvideo/models/AGENTS.md` | DiT / VAE / encoder / scheduler / loader layout (pre-commit excluded) |
+| `fastvideo/models/wan/AGENTS.md` | Wan family-local transformers, VAE, configs, and the SP sharding invariant |
 | `fastvideo/layers/AGENTS.md` | Tensor-parallel linear/attention layer rules for ports |
 | `fastvideo/attention/AGENTS.md` | Backend registry + env-var override |
 | `fastvideo/pipelines/AGENTS.md` | Stage ABC, `basic/<model>/`, `preprocess/`, presets |
+| `fastvideo/pipelines/basic/wan/AGENTS.md` | Wan sampling stages, first-frame conditioning, DMD/causal boundaries |
+| `fastvideo/pipelines/basic/magi_human/AGENTS.md` | MagiHuman umbrella repo, lazy-loaded components, packing invariants |
 | `fastvideo/training/AGENTS.md` | Legacy monolithic pipelines (frozen for existing models) |
 | `fastvideo/train/AGENTS.md` | New modular trainer (methods × models × callbacks, YAML) |
 | `fastvideo/tests/AGENTS.md` | Test taxonomy, conftest, pre-commit-excluded path |
 | `fastvideo/tests/ssim/AGENTS.md` | GPU SSIM regression authoring + reference video sync |
 | `scripts/checkpoint_conversion/AGENTS.md` | Adding a converter for a new HF/official checkpoint |
+| `apps/dreamverse/AGENTS.md` | DreamVerse app structure and conventions |

 ## Critical: Two Training Stacks Coexist

@@ -3,14 +3,16 @@
 </div>

 <p align="center">
-     | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/"><b> Quick Start</b></a> | <a href="https://github.com/hao-ai-lab/FastVideo/discussions/982"  target="_blank"><b>Weekly Dev Meeting</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://github.com/hao-ai-lab/FastVideo/discussions/1097" target="_blank"> <b> WeChat </b> </a> |
+     | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://haoailab.com/FastVideo/cookbook/"><b>Cookbook</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start/"><b> Quick Start</b></a> | <a href="https://github.com/hao-ai-lab/FastVideo/discussions/982"  target="_blank"><b>Weekly Dev Meeting</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://github.com/hao-ai-lab/FastVideo/discussions/1097" target="_blank"> <b> WeChat </b> </a> |
 </p>

 **FastVideo is a unified post-training and real-time inference framework for accelerated video generation.**

 ## NEWS
+- `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).
- `2026/08/19`: FastVideo now supports MLX on Apple Silicon with [FastMetal-QAD](https://huggingface.co/collections/FastVideo/fastmetal), a family of 1.3B, 5B, and 14B models optimized for Mac—follow the [Apple Silicon guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mps/) and read the [Blog](https://haoailab.com/blogs/fastmetal/).
+- `2026/08/19`: FastVideo now supports MLX on Apple Silicon with [FastMetal-QAD](https://huggingface.co/collections/FastVideo/fastmetal), a family of 1.3B, 5B, and 14B models optimized for Mac. Follow the [MLX install guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mlx/) and read the [Blog](https://haoailab.com/blogs/fastmetal/).
 - `2026/06/23`: Release FastWan-QAD: 5s of Video generated in 1.8s E2E. See the [FastWan-QAD models](https://huggingface.co/FastVideo/FastWan-QAD-FP8-1.3B), [Attn-QAT training guide](https://haoailab.com/FastVideo/training/attn_qat/), and [blog](https://haoailab.com/blogs/fastwan-qad/).
 - `2026/03/17`: Release demo: Into the Dreamverse: Vibe Directing in FastVideo, check out the [Blog](https://haoailab.com/blogs/dreamverse/).
 - `2026/03/13`: Release demo: Create a 5s 1080p Video in 4.5s with FastVideo on a Single GPU, check out the [Blog](https://haoailab.com/blogs/fastvideo_realtime_1080p/).
@@ -62,13 +64,12 @@ UV_TORCH_BACKEND=cu126 uv pip install fastvideo
 ```

 Use `UV_TORCH_BACKEND=cu130` on CUDA 13. Apple silicon users should follow the
-[MPS installation guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mps/).
+[MLX install guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mlx/).

-> **On an Apple Silicon Mac?** FastVideo runs FastMetal-QAD through an MLX
-> runtime. Install with `uv pip install -e '.[mlx]'`, download
-> [`FastVideo/FastMetal-1.3B-QAD`](https://huggingface.co/FastVideo/FastMetal-1.3B-QAD),
-> and follow the
-> [Apple Silicon guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mps/).
+> **On an Apple Silicon Mac?** Install with `uv pip install -e '.[mlx]'` from
+> a clone, then pick a recipe in the
+> [cookbook](https://haoailab.com/FastVideo/cookbook/). See the
+> [MLX install guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/mlx/).

 Please see our [docs](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/) for more detailed installation instructions.

@@ -86,7 +87,7 @@ Install FastVideo (https://github.com/hao-ai-lab/FastVideo) into a fresh uv virt
   https://hao-ai-lab.github.io/FastVideo/getting_started/installation/):
     - NVIDIA GPU, x86_64                         -> docs/getting_started/installation/gpu.md
     - NVIDIA DGX Spark / GB10, aarch64, CUDA 13  -> docs/getting_started/installation/spark.md
-     - Apple Silicon, macOS                       -> docs/getting_started/installation/mps.md
+     - Apple Silicon, macOS                       -> docs/getting_started/installation/mlx.md
 3. Use uv for every step. If a command fails, debug it and tell me what you changed.
 4. Verify the result:
     python -c "import fastvideo, torch; print('cuda', torch.cuda.is_available())"
@@ -134,18 +135,20 @@ def main():
    # Create a video generator with a pre-trained model
    generator = VideoGenerator.from_pretrained(
        "FastVideo/FastWan2.1-T2V-1.3B-Diffusers",
-        num_gpus=1,  # Adjust based on your hardware
+        {"engine": {"num_gpus": 1}},  # Adjust based on your hardware
    )

    # Define a prompt for your video
    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."

    # Generate the video
-    video = generator.generate_video(
-        prompt,
-        output_path="my_videos/",  # Controls where videos are saved
-        save_video=True
-    )
+    video = generator.generate({
+        "prompt": prompt,
+        "output": {
+            "output_path": "my_videos/",  # Controls where videos are saved
+            "save_video": True,
+        },
+    })

 if __name__ == '__main__':
    main()
@@ -97,13 +97,33 @@ dreamverse-server --port 8009
 dreamverse-mock-server --port 8009
 ```

+### Run Dreamverse with FastH3
+
+Select the VSA data-free FastH3 Preview profile when you start the backend:
+
+```bash
+DREAMVERSE_MODEL_ID=fast-h3 dreamverse-server --port 8009
+```
+
+The `fast-h3` profile uses four visible GPUs by default. It loads the `MiniMaxAI/MiniMax-H3` base checkpoint and the
+`vsa-datafree/adapter_model.safetensors` adapter from
+`FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA`. Each request generates a 124-frame, 768×1344 video with
+synchronized audio and five sigma-grid points. Dreamverse uses the last frame of each segment as first-frame
+conditioning for the following segment.
+
+Set `CUDA_VISIBLE_DEVICES` when you need to choose the four physical GPUs:
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 DREAMVERSE_MODEL_ID=fast-h3 dreamverse-server --port 8009
+```
+
 > **Expect a slow first boot.** With `torch.compile` and startup warmup enabled
 > (the default), the backend compiles the segment 1 and segment 2 inference
 > paths before it reports ready — this can take **tens of minutes on a cold
 > cache**, regardless of how you deploy (local, server, Docker, or Modal).
 > `/healthz` responds as soon as the process is up; `/readyz` stays `503` until
-> warmup finishes. For a faster, uncompiled startup while testing, set
-> `FASTVIDEO_ENABLE_STARTUP_WARMUP=0` before starting the backend.
+> warmup finishes. To defer compilation until the first generated request while
+> testing, set `FASTVIDEO_ENABLE_STARTUP_WARMUP=0` before starting the backend.

 ## Frontend Setup

@@ -219,6 +239,7 @@ selection, and mock-server behavior:
 pytest apps/dreamverse/dreamverse/tests/test_config.py \
  apps/dreamverse/dreamverse/tests/test_entrypoints.py \
  apps/dreamverse/dreamverse/tests/test_gpu_pool.py \
+  apps/dreamverse/dreamverse/tests/test_minimax_h3_generation.py \
  apps/dreamverse/dreamverse/tests/test_mock_server.py -q
 ```

@@ -42,7 +42,7 @@ Near-term OSS note:
 - `apps/dreamverse/dreamverse/main.py`: websocket endpoint, request handling,
  session state machine, rewrite orchestration, REST routes, and stream relay
 - `apps/dreamverse/dreamverse/gpu_pool.py`: GPU worker processes, warmup, model
-  loading, and `generate_video()` calls through FastVideo
+  loading, and `generate()` calls through FastVideo
 - `apps/dreamverse/dreamverse/prompt_enhancer.py`: prompt enhancement, rollout
  rewrite execution, provider selection, and timeout/fallback behavior
 - `apps/dreamverse/dreamverse/rewrite_prompt_payload.py`: canonical rewrite request payload
@@ -139,7 +139,18 @@ session.
 - startup warmup
 - user join/leave commands
 - `USER_STEP` execution for each segment
- continuation state between segments
+- generation-command routing and stream-result delivery
+
+Model generation has a separate ownership boundary inside each GPU process:
+
+- `apps/dreamverse/dreamverse/generation_worker.py` selects the backend that the active model profile declares and owns
+  the backend lifecycle.
+- `apps/dreamverse/dreamverse/ltx2_generation.py` owns LTX-2 generator configuration, video and audio continuation, and
+  runtime LoRA application.
+- `apps/dreamverse/dreamverse/minimax_h3_generation.py` owns the VSA data-free FastH3 adapter, FastH3 generator and
+  request configuration, and last-frame continuation through MiniMax H3 first-frame conditioning.
+- `apps/dreamverse/dreamverse/generation_contracts.py` defines the decoded media and stream-trimming result that both
+  model backends return to `apps/dreamverse/dreamverse/gpu_pool.py`.

 `apps/dreamverse/dreamverse/prompt_enhancer.py` manages:

@@ -1,6 +1,6 @@
 """Benchmark the LTX-2 generation pipeline driven by the dreamverse Python SDK path.

-Mirrors how ``apps/dreamverse/dreamverse/video_generation.py`` constructs
+Mirrors how ``apps/dreamverse/dreamverse/ltx2_generation.py`` constructs
 ``GeneratorConfig`` and calls ``VideoGenerator.generate()``, then
 captures per-stage timings via the ``FASTVIDEO_STAGE_LOGGING=1`` log
 hooks (same mechanism as ``FastVideo-internal/examples/inference/basic/
@@ -52,7 +52,8 @@ import torch  # noqa: E402

 from fastvideo import VideoGenerator  # noqa: E402
 from fastvideo.api import (  # noqa: E402
-    ComponentConfig, CompileConfig, EngineConfig, GeneratorConfig, OffloadConfig, PipelineSelection, QuantizationConfig,
+    ComponentConfig, CompileConfig, EngineConfig, GenerationResult, GeneratorConfig, OffloadConfig, PipelineSelection,
+    QuantizationConfig,
 )

 DEFAULT_PROMPT = ("A cinematic drone shot over coastal cliffs at sunrise, golden "
@@ -128,9 +129,9 @@ def _build_generator_config(model_path: str, enable_compile: bool, num_gpus: int
    )


-def _extract_stage_times(result: dict) -> OrderedDict[str, float]:
+def _extract_stage_times(result: GenerationResult) -> OrderedDict[str, float]:
    out: OrderedDict[str, float] = OrderedDict()
-    info = result.get("logging_info") if isinstance(result, dict) else None
+    info = result.logging_info if isinstance(result, GenerationResult) else None
    if info is None:
        return out
    stages = getattr(info, "stages", None)
@@ -162,19 +163,25 @@ def _do_one_run(generator: VideoGenerator, prompt: str, *, height: int, width: i
    _reset_peak_gpu()
    t0 = time.perf_counter()
    try:
-        result = generator.generate_video(
-            prompt=prompt,
-            negative_prompt="",
-            save_video=False,
-            height=height,
-            width=width,
-            num_frames=num_frames,
-            fps=24,
-            num_inference_steps=num_inference_steps,
-            guidance_scale=1.0,
-            seed=seed,
-            ltx2_image_crf=0.0,
-        )
+        result = generator.generate({
+            "prompt": prompt,
+            "negative_prompt": "",
+            "sampling": {
+                "height": height,
+                "width": width,
+                "num_frames": num_frames,
+                "fps": 24,
+                "num_inference_steps": num_inference_steps,
+                "guidance_scale": 1.0,
+                "seed": seed,
+            },
+            "output": {
+                "save_video": False
+            },
+            "extensions": {
+                "ltx2_image_crf": 0.0
+            },
+        })
        if torch.cuda.is_available():
            torch.cuda.synchronize()
    except Exception as exc:
@@ -1,5 +1,6 @@
 import os
 from pathlib import Path
+from typing import cast

 _REPO_ROOT = Path(__file__).resolve().parents[1]
 _SERVER_ROOT = Path(__file__).resolve().parent
@@ -55,16 +56,34 @@ FRONTEND_STATIC_DIR_CANDIDATES = _resolve_frontend_static_dir_candidates()
 MODEL_REGISTRY = {
    "fast-ltx2": {
        "name": "FastLTX2",
+        "generation_backend": "ltx2",
+        "default_sp_size": 1,
        "model_path": "FastVideo/LTX2-Distilled-Diffusers",
        "config_model_path": "FastVideo/LTX2-Distilled-Diffusers",
        "lora_repo": "FastVideo/LTX2-OmniNFT-LoRA",
    },
    "fast-ltx23": {
        "name": "FastLTX23",
+        "generation_backend": "ltx2",
+        "default_sp_size": 1,
        "model_path": "FastVideo/LTX-2.3-Distilled-Diffusers",
        "config_model_path": "FastVideo/LTX-2.3-Distilled-Diffusers",
        "lora_repo": "FastVideo/LTX-2.3-OmniNFT-LoRA",
    },
+    "fast-h3": {
+        "name": "FastH3",
+        "generation_backend": "minimax_h3",
+        "default_sp_size": 4,
+        "model_path": "MiniMaxAI/MiniMax-H3",
+        "adapter_repo": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
+        "adapter_filename": "vsa-datafree/adapter_model.safetensors",
+        "attention_backend": "VIDEO_SPARSE_ATTN_H3",
+        "height": 768,
+        "width": 1344,
+        "num_frames": 124,
+        "num_inference_steps": 5,
+        "seed": 1000,
+    },
 }

 DEFAULT_MODEL_ID = "fast-ltx2"
@@ -171,7 +190,7 @@ def _optional_env(*names: str) -> str | None:
 DEVTOOLS_ENABLED = _env_bool("FASTVIDEO_ENABLE_DEVTOOLS", False)
 PROMPT_SAFETY_ENABLED = _env_bool("FASTVIDEO_ENABLE_PROMPT_SAFETY", False)
 DREAMVERSE_MAX_AUTOTUNE = _env_bool("DREAMVERSE_MAX_AUTOTUNE", True)
-DREAMVERSE_SP_SIZE = max(1, _env_int("DREAMVERSE_SP_SIZE", 1))
+DREAMVERSE_SP_SIZE = max(1, _env_int("DREAMVERSE_SP_SIZE", cast(int, MODEL_CONFIG["default_sp_size"])))

 DREAMVERSE_MODEL_PATH = (os.getenv("DREAMVERSE_MODEL_PATH", "").strip() or None)
 if DREAMVERSE_MODEL_PATH:
@@ -213,7 +232,7 @@ def _resolve_lora_spec(spec: str) -> str | None:
    if not spec:
        return None
    if spec.lower() == "omninft":
-        return MODEL_CONFIG.get("lora_repo")
+        return cast(str | None, MODEL_CONFIG.get("lora_repo"))
    if spec.lower() in AVAILABLE_LORAS:
        return AVAILABLE_LORAS[spec.lower()]["repo"]
    return spec
@@ -0,0 +1,46 @@
+"""Shared contract between DreamVerse generation backends and GPU workers."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+
+@dataclass
+class StepResult:
+    """Decoded media and stream-trimming metadata for one DreamVerse segment."""
+
+    frames: list
+    audio: Any
+    audio_sample_rate: int | None
+    timings: dict[str, float]
+    head_trim_frames: int
+    head_trim_audio_frames: int
+
+
+class GenerationBackend(Protocol):
+    """Model-owned generation operations used by one GPU worker process."""
+
+    def initialize(self, model_config: dict | None = None) -> None:
+        ...
+
+    def shutdown(self) -> None:
+        ...
+
+    def clear_conditioning(self) -> None:
+        ...
+
+    def generate_step(
+        self,
+        prompt: str,
+        segment_idx: int,
+        image_path: str | None,
+        reset_conditioning: bool,
+    ) -> StepResult:
+        ...
+
+    def warmup(self, prompt: str) -> dict[str, float]:
+        ...
+
+    def apply_lora_stack(self, stack: list[tuple[str, float]]) -> tuple[str | None, str | None]:
+        ...
@@ -0,0 +1,96 @@
+"""Select and own one model-specific generation backend per GPU process."""
+
+from __future__ import annotations
+
+from dreamverse.config import MODEL_CONFIG
+from dreamverse.generation_contracts import GenerationBackend, StepResult
+
+
+def _create_generation_backend(backend_name: str, gpu_id: int) -> GenerationBackend:
+    """Construct the backend that owns the selected model family's behavior."""
+    if backend_name == "ltx2":
+        from dreamverse.ltx2_generation import LTX2GenerationBackend
+
+        return LTX2GenerationBackend(gpu_id)
+    if backend_name == "minimax_h3":
+        from dreamverse.minimax_h3_generation import MiniMaxH3GenerationBackend
+
+        return MiniMaxH3GenerationBackend(gpu_id)
+    raise ValueError(f"Unsupported DreamVerse generation backend: {backend_name!r}")
+
+
+class VideoGenerationWorker:
+    """Delegate GPU lifecycle and generation calls to the active model backend."""
+
+    def __init__(self, gpu_id: int):
+        self.gpu_id = gpu_id
+        self.model_config: dict = dict(MODEL_CONFIG)
+        self.backend_name: str | None = None
+        self.backend: GenerationBackend | None = None
+
+    def initialize(self, model_config: dict | None = None) -> None:
+        """Load the requested model through its generation backend.
+
+        Model selection belongs here so the GPU process and streaming layers
+        use one stable media contract without importing model-specific code.
+        """
+        requested_model_config = dict(model_config) if model_config is not None else dict(self.model_config)
+        backend_name = requested_model_config.get("generation_backend")
+        if not isinstance(backend_name, str) or not backend_name:
+            raise ValueError("DreamVerse model configuration requires `generation_backend`.")
+
+        candidate_backend = self.backend
+        if candidate_backend is None or self.backend_name != backend_name:
+            if candidate_backend is not None:
+                candidate_backend.shutdown()
+            candidate_backend = _create_generation_backend(backend_name, self.gpu_id)
+
+        try:
+            candidate_backend.initialize(requested_model_config)
+        except Exception:
+            try:
+                candidate_backend.shutdown()
+            except Exception as shutdown_error:
+                print(f"[GPU {self.gpu_id}] Backend cleanup after initialization failure: {shutdown_error}")
+            self.backend = None
+            self.backend_name = None
+            raise
+
+        self.model_config = requested_model_config
+        self.backend = candidate_backend
+        self.backend_name = backend_name
+
+    def _require_backend(self) -> GenerationBackend:
+        """Return the initialized backend or fail before processing a command."""
+        if self.backend is None:
+            raise RuntimeError("Generation backend is not initialized.")
+        return self.backend
+
+    def shutdown(self) -> None:
+        """Release model resources owned by the selected backend."""
+        if self.backend is not None:
+            self.backend.shutdown()
+
+    def clear_conditioning(self) -> None:
+        self._require_backend().clear_conditioning()
+
+    def generate_step(
+        self,
+        prompt: str,
+        segment_idx: int,
+        image_path: str | None,
+        reset_conditioning: bool,
+    ) -> StepResult:
+        """Generate one segment through the selected model backend."""
+        return self._require_backend().generate_step(
+            prompt,
+            segment_idx,
+            image_path,
+            reset_conditioning,
+        )
+
+    def warmup(self, prompt: str) -> dict[str, float]:
+        return self._require_backend().warmup(prompt)
+
+    def apply_lora_stack(self, stack: list[tuple[str, float]]) -> tuple[str | None, str | None]:
+        return self._require_backend().apply_lora_stack(stack)
@@ -12,7 +12,7 @@ from enum import Enum
 from multiprocessing import Process, Queue

 from dreamverse.config import (
-    DEFAULT_MODEL_ID,
+    ACTIVE_MODEL_ID,
    DREAMVERSE_SP_SIZE,
    MODEL_REGISTRY,
    STARTUP_WARMUP_ENABLED,
@@ -54,7 +54,7 @@ from dreamverse.worker_ipc import (
 def _parse_requested_gpu_limit() -> int | None:
    raw_value = os.getenv("FASTVIDEO_GPU_COUNT", "").strip().lower()
    if not raw_value:
-        return 1
+        return DREAMVERSE_SP_SIZE
    if raw_value == "all":
        return None
    try:
@@ -164,12 +164,12 @@ def gpu_worker_process(
    os.environ["CUDA_VISIBLE_DEVICES"] = cuda_device
    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN"

-    from dreamverse.video_generation import VideoGenerationWorker
+    from dreamverse.generation_worker import VideoGenerationWorker

    worker = VideoGenerationWorker(gpu_id)

    def event_loop(first_cmd: Command = None):
-        """Blocking event loop for LTX2; dispatches user commands."""
+        """Block on generation commands after the model is initialized."""
        print(f"[GPU {gpu_id}] Entering event loop")

        def handle_command(cmd: Command):
@@ -435,7 +435,7 @@ class GPUSlot:
        self._response_reader_task: asyncio.Task | None = None
        self._active: bool = False
        self._reader_lock: asyncio.Lock | None = None
-        self.current_model_id: str = DEFAULT_MODEL_ID
+        self.current_model_id: str | None = ACTIVE_MODEL_ID
        self.shared_stream_buffer = None
        self.shared_stream_buffer_size = SHARED_STREAM_BUFFER_BYTES

@@ -690,7 +690,7 @@ class GPUSlot:
    async def join_user(self, user_id: str, model_id: str = None) -> JoinAck:
        """Add a user to this GPU."""
        if model_id is None:
-            model_id = DEFAULT_MODEL_ID
+            model_id = ACTIVE_MODEL_ID

        # Reload model if a different one is requested
        if model_id != self.current_model_id and model_id in MODEL_REGISTRY:
@@ -705,16 +705,23 @@ class GPUSlot:
            self.connected_users.clear()

            model_config = MODEL_REGISTRY[model_id]
-            reload_response = await self._send_command(Command(CommandType.RELOAD_MODEL,
-                                                               payload=ReloadModelPayload(model_config=model_config),
-                                                               user_id="__reload__"),
-                                                       timeout=600.0)
+            try:
+                reload_response = await self._send_command(Command(
+                    CommandType.RELOAD_MODEL,
+                    payload=ReloadModelPayload(model_config=model_config),
+                    user_id="__reload__"),
+                                                           timeout=600.0)
+            except Exception:
+                self.current_model_id = None
+                raise
            match reload_response:
                case ReloadAck():
                    pass
                case WorkerError(message=msg):
+                    self.current_model_id = None
                    raise RuntimeError(f"Model reload failed: {msg}")
                case _:
+                    self.current_model_id = None
                    raise RuntimeError(f"Unexpected reload response: "
                                       f"{type(reload_response).__name__}")

@@ -1,9 +1,9 @@
-"""LTX2 model lifecycle and continuation conditioning.
+"""LTX-2 model lifecycle and continuation conditioning.

 Runs inside a GPU worker subprocess.  Owns the model, the audio
 encoder, and the per-session continuation state carried across
 segments.  Callers must set ``os.environ["CUDA_VISIBLE_DEVICES"]``
-before constructing ``VideoGenerationWorker`` — all ``fastvideo.*``
+before constructing ``LTX2GenerationBackend`` — all ``fastvideo.*``
 imports are deferred to method bodies so nothing touches CUDA at
 module import time.
 """
@@ -14,9 +14,7 @@ import gc
 import os
 import re
 import time
-from dataclasses import dataclass
-from typing import Any
-
+from typing import TYPE_CHECKING
 import numpy as np
 import torch

@@ -35,6 +33,10 @@ from dreamverse.config import (
    DREAMVERSE_LORA_STACK,
    _resolve_lora_spec,
 )
+from dreamverse.generation_contracts import StepResult
+
+if TYPE_CHECKING:
+    from fastvideo.api import GenerationResult

 # Multi-frame decoded continuation defaults from
 # examples/inference/basic/basic_ltx2_distilled_video_continuation.py.
@@ -80,22 +82,6 @@ def _reset_lora_registry(worker) -> dict:
    return {"status": "lora_registry_reset"}


-@dataclass
-class StepResult:
-    """Output of one generation step.
-
-    ``head_trim_frames`` / ``head_trim_audio_frames`` are derived here
-    so downstream AV streaming never needs to import conditioning
-    constants.
-    """
-    frames: list
-    audio: Any
-    audio_sample_rate: int | None
-    timings: dict
-    head_trim_frames: int
-    head_trim_audio_frames: int
-
-
 class ContinuationState:
    """Per-session video + audio conditioning carried across segments."""

@@ -113,8 +99,8 @@ class ContinuationState:
        self.video_images = None
        self.audio_latents = None

-    def apply_video(self, request_kwargs: dict, segment_idx: int) -> None:
-        """Seed next-segment kwargs with the cached tail frames."""
+    def apply_video(self, request: dict, segment_idx: int) -> None:
+        """Seed the next-segment request with the cached tail frames."""
        if segment_idx <= 1 or not self.video_images:
            return
        from PIL import Image
@@ -128,21 +114,21 @@ class ContinuationState:
                arr = np.clip(arr, 0, 255).astype(np.uint8)
                noisy.append(Image.fromarray(arr))
            cond_images = noisy
-        request_kwargs["ltx2_video_conditions"] = [(
+        request["extensions"]["ltx2_video_conditions"] = [(
            cond_images,
            LTX2_VIDEO_CONDITIONING_FRAME_IDX,
            LTX2_VIDEO_CONDITIONING_STRENGTH,
        )]
-        request_kwargs["ltx2_images"] = None
-        request_kwargs["image_path"] = None
+        request["extensions"]["ltx2_images"] = None
+        request["inputs"]["image_path"] = None

    def apply_audio(
        self,
-        request_kwargs: dict,
+        request: dict,
        segment_idx: int,
        audio_lps: float,
    ) -> None:
-        """Seed next-segment kwargs with clean audio latents + denoise mask.
+        """Seed the next-segment request with clean audio latents + denoise mask.

        When audio conditioning is longer than video, extend audio
        generation and shift video RoPE forward so the audio prefix
@@ -159,9 +145,9 @@ class ContinuationState:
        audio_extra = max(0, AUDIO_CONDITIONING_NUM_FRAMES - LTX2_VIDEO_CONDITIONING_NUM_FRAMES)
        if audio_extra > 0:
            audio_num_frames = NUM_FRAMES + audio_extra
-            request_kwargs["audio_num_frames"] = (audio_num_frames)
+            request["extensions"]["audio_num_frames"] = (audio_num_frames)
            prefix_sec = float(audio_extra) / 24.0
-            request_kwargs["video_position_offset_sec"] = prefix_sec
+            request["extensions"]["video_position_offset_sec"] = prefix_sec

        new_duration = float(NUM_FRAMES + audio_extra) / 24.0
        total_T = max(
@@ -179,8 +165,8 @@ class ContinuationState:
        mask = torch.ones((B, 1, total_T, 1), dtype=torch.float32)
        mask[:, :, :audio_cond_T, :] = (1.0 - AUDIO_CONDITIONING_STRENGTH)

-        request_kwargs["ltx2_audio_clean_latent"] = clean
-        request_kwargs["ltx2_audio_denoise_mask"] = mask
+        request["extensions"]["ltx2_audio_clean_latent"] = clean
+        request["extensions"]["ltx2_audio_denoise_mask"] = mask

    def save_video(self, frames: list) -> None:
        """Snapshot trailing N frames as PIL images for next-segment conditioning."""
@@ -202,7 +188,7 @@ class ContinuationState:
        self.audio_latents = latents.detach().clone().cpu()


-class VideoGenerationWorker:
+class LTX2GenerationBackend:
    """Single-GPU LTX2 generator with continuation state.

    Caller must set ``os.environ["CUDA_VISIBLE_DEVICES"]`` before
@@ -324,7 +310,7 @@ class VideoGenerationWorker:
            ),
        )

-        self.generator = VideoGenerator.from_pretrained(config=generator_config)
+        self.generator = VideoGenerator.from_config(generator_config)
        print(f"[GPU {self.gpu_id}] After model load: {self._gpu_mem()}")

        lora_stack = DREAMVERSE_LORA_STACK or ([(DREAMVERSE_LORA_PATH,
@@ -421,7 +407,7 @@ class VideoGenerationWorker:
            return

        loader = ComponentLoader.for_module_type("audio_encoder", "diffusers")
-        enc = loader.load(audio_vae_path, self.generator.fastvideo_args)
+        enc = loader.load(audio_vae_path, self.generator.resolved_config)
        target = getattr(enc, "model", enc)

        proc = AudioProcessor(
@@ -478,51 +464,59 @@ class VideoGenerationWorker:

        prompt = self._inject_style_trigger(prompt)

-        request_kwargs = dict(
-            prompt=prompt,
-            negative_prompt="",
-            save_video=False,
-            height=FRAME_HEIGHT,
-            width=FRAME_WIDTH,
-            num_frames=NUM_FRAMES,
-            fps=24,
-            num_inference_steps=NUM_INFERENCE_STEPS,
-            guidance_scale=1.0,
-            seed=10,
-            ltx2_image_crf=0.0,
-            image_path=image_path if segment_idx == 1 else None,
-            return_continuation_state=False,
-        )
+        request = {
+            "prompt": prompt,
+            "negative_prompt": "",
+            "inputs": {
+                "image_path": image_path if segment_idx == 1 else None
+            },
+            "sampling": {
+                "height": FRAME_HEIGHT,
+                "width": FRAME_WIDTH,
+                "num_frames": NUM_FRAMES,
+                "fps": 24,
+                "num_inference_steps": NUM_INFERENCE_STEPS,
+                "guidance_scale": 1.0,
+                "seed": 10,
+            },
+            "output": {
+                "save_video": False
+            },
+            "extensions": {
+                "ltx2_image_crf": 0.0,
+                "return_continuation_state": False,
+            },
+        }

        if reset_conditioning:
            self.continuation.clear()

        audio_lps = (DEFAULT_LTX2_AUDIO_SAMPLE_RATE / DEFAULT_LTX2_AUDIO_HOP_LENGTH / DEFAULT_LTX2_AUDIO_DOWNSAMPLE)

-        # Phase 1: seed kwargs with prior-segment conditioning.
-        self.continuation.apply_video(request_kwargs, segment_idx)
-        self.continuation.apply_audio(request_kwargs, segment_idx, audio_lps)
+        # Phase 1: seed the request with prior-segment conditioning.
+        self.continuation.apply_video(request, segment_idx)
+        self.continuation.apply_audio(request, segment_idx, audio_lps)

        # Phase 2: generate.
        t0 = time.perf_counter()
-        result = self.generator.generate_video(**request_kwargs)
+        result = self.generator.generate(request)
        torch.cuda.synchronize()
        timings["generation_ms"] = (time.perf_counter() - t0) * 1000

-        if not isinstance(result, dict):
-            raise RuntimeError("Expected dictionary output from generate_video.")
-        frames = result.get("frames")
+        if isinstance(result, list):
+            raise RuntimeError("Expected a single GenerationResult from generate.")
+        frames = result.frames
        if not isinstance(frames, list) or len(frames) == 0:
            raise RuntimeError("Generation did not return frames.")
-        audio = result.get("audio")
-        audio_sample_rate = result.get("audio_sample_rate")
+        audio = result.audio
+        audio_sample_rate = result.audio_sample_rate
        if audio is not None and audio_sample_rate is None:
            # LTX2 audio decoding stage uses 24kHz output by default.
            audio_sample_rate = 24000
            print(f"[GPU {self.gpu_id}] audio_sample_rate missing from result; "
                  f"defaulting to {audio_sample_rate}Hz")

-        timings["generation_time_ms"] = result.get("generation_time", 0.0) * 1000
+        timings["generation_time_ms"] = (result.generation_time or 0.0) * 1000

        # Phase 3: snapshot continuation state for the next segment.
        t_save_start = time.perf_counter()
@@ -560,7 +554,7 @@ class VideoGenerationWorker:
        self,
        audio: object,
        audio_sample_rate: int | None,
-        result: dict,
+        result: "GenerationResult",
        segment_idx: int,
    ) -> torch.Tensor | None:
        """Pick which tensor to cache for next-segment audio conditioning."""
@@ -575,7 +569,7 @@ class VideoGenerationWorker:
                      f"for segment {segment_idx + 1}")
                return re_encoded
            return None
-        audio_latents = result.get("ltx2_audio_latents")
+        audio_latents = result.extra.get("ltx2_audio_latents")
        if audio_latents is not None:
            print(f"[GPU {self.gpu_id}] Cached audio latents "
                  f"shape={tuple(audio_latents.shape)} "
@@ -0,0 +1,294 @@
+"""FastH3 model lifecycle and first-frame continuation for DreamVerse."""
+
+from __future__ import annotations
+
+import gc
+import os
+import time
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+import torch
+
+from dreamverse.config import DREAMVERSE_SP_SIZE
+from dreamverse.generation_contracts import StepResult
+
+if TYPE_CHECKING:
+    from PIL.Image import Image
+
+
+def _required_config_str(model_config: dict, field_name: str) -> str:
+    """Read one required non-empty string from a DreamVerse model profile."""
+    value = model_config.get(field_name)
+    if not isinstance(value, str) or not value.strip():
+        raise ValueError(f"FastH3 model configuration requires `{field_name}`.")
+    return value.strip()
+
+
+class MiniMaxH3GenerationBackend:
+    """Run the VSA data-free FastH3 adapter and retain one continuation frame."""
+
+    def __init__(self, gpu_id: int):
+        self.gpu_id = gpu_id
+        self.generator: Any | None = None
+        self.model_config: dict = {}
+        self.continuation_image: Image | None = None
+
+    def _gpu_mem(self) -> str:
+        allocated_gib = torch.cuda.memory_allocated() / 1024**3
+        reserved_gib = torch.cuda.memory_reserved() / 1024**3
+        return f"alloc={allocated_gib:.2f}GiB, reserved={reserved_gib:.2f}GiB"
+
+    @staticmethod
+    def _configure_environment(attention_backend: str) -> None:
+        """Apply the fixed boot-time switches from the FastH3 reference recipe."""
+        os.environ.update({
+            "FASTVIDEO_ATTENTION_BACKEND": attention_backend,
+            "FASTVIDEO_FA4": "1",
+            "FASTVIDEO_MINIMAX_H3_FUSIONS": "all",
+            "FASTVIDEO_VSA_SM100A": "0",
+        })
+        os.environ.pop("FASTVIDEO_INFERENCE_TORCH_COMPILE", None)
+
+    def initialize(self, model_config: dict | None = None) -> None:
+        """Download the fixed Preview adapter and load the FastH3 generator.
+
+        The model profile owns the base checkpoint, adapter file, attention
+        backend, and generation geometry. The backend translates that profile
+        into FastVideo's typed generator configuration.
+        """
+        if model_config is not None:
+            self.model_config = dict(model_config)
+        if not self.model_config:
+            raise ValueError("FastH3 initialization requires a model configuration.")
+
+        if self.generator is not None:
+            self.generator.shutdown()
+            self.generator = None
+            gc.collect()
+            torch.cuda.empty_cache()
+
+        self.clear_conditioning()
+        model_path = _required_config_str(self.model_config, "model_path")
+        adapter_repo = _required_config_str(self.model_config, "adapter_repo")
+        adapter_filename = _required_config_str(self.model_config, "adapter_filename")
+        attention_backend = _required_config_str(self.model_config, "attention_backend")
+        self._configure_environment(attention_backend)
+
+        from huggingface_hub import hf_hub_download
+
+        from fastvideo import VideoGenerator
+        from fastvideo.api import (
+            AttentionConfig,
+            CompileConfig,
+            ComponentConfig,
+            EngineConfig,
+            GeneratorConfig,
+            MiniMaxH3Options,
+            OffloadConfig,
+            ParallelismConfig,
+            PipelineSelection,
+        )
+
+        adapter_path = hf_hub_download(repo_id=adapter_repo, filename=adapter_filename)
+        use_vsa = attention_backend == "VIDEO_SPARSE_ATTN_H3"
+        generator_config = GeneratorConfig(
+            model_path=model_path,
+            pipeline=PipelineSelection(
+                components=ComponentConfig(lora_path=adapter_path, lora_strength=1.0),
+                model=MiniMaxH3Options(vae_parallel_decode=True, vae_parallel_decode_strategy="gather"),
+            ),
+            engine=EngineConfig(
+                num_gpus=DREAMVERSE_SP_SIZE,
+                parallelism=ParallelismConfig(tp_size=1, sp_size=DREAMVERSE_SP_SIZE),
+                offload=OffloadConfig(
+                    dit=False,
+                    dit_layerwise=False,
+                    text_encoder=True,
+                    image_encoder=True,
+                    vae=True,
+                    pin_cpu_memory=True,
+                ),
+                compile=CompileConfig(enabled=False, vae_enabled=True, regional=attention_backend == "FLASH_ATTN"),
+                attention=AttentionConfig(
+                    backend=attention_backend,
+                    vsa_sparsity=0.9 if use_vsa else None,
+                    vsa_tile_size=64 if use_vsa else None,
+                ),
+                use_fsdp_inference=False,
+            ),
+        )
+
+        print(f"[GPU {self.gpu_id}] Loading FastH3 model: {model_path}")
+        print(f"[GPU {self.gpu_id}] FastH3 adapter: {adapter_repo}/{adapter_filename}")
+        print(f"[GPU {self.gpu_id}] Before model load: {self._gpu_mem()}")
+        self.generator = VideoGenerator.from_config(generator_config)
+        print(f"[GPU {self.gpu_id}] FastH3 loaded: {self._gpu_mem()} (warmup pending)")
+
+    def shutdown(self) -> None:
+        """Release the FastVideo generator and cached continuation image."""
+        self.clear_conditioning()
+        if self.generator is not None:
+            self.generator.shutdown()
+            self.generator = None
+
+    def clear_conditioning(self) -> None:
+        """Release the first-frame image retained for the next segment."""
+        if self.continuation_image is not None:
+            self.continuation_image.close()
+            self.continuation_image = None
+
+    @staticmethod
+    def _load_rgb_image(image_path: str) -> Image:
+        """Load an image into an independent RGB buffer with no open file handle."""
+        from PIL import Image
+
+        with Image.open(image_path) as image:
+            return image.convert("RGB").copy()
+
+    def _select_conditioning_image(
+        self,
+        segment_idx: int,
+        image_path: str | None,
+        reset_conditioning: bool,
+    ) -> tuple[Image | None, bool]:
+        """Select the initial upload or retained last frame for one segment."""
+        if reset_conditioning:
+            self.clear_conditioning()
+        if segment_idx > 1 and self.continuation_image is not None:
+            return self.continuation_image.copy(), True
+        if segment_idx > 1 and not reset_conditioning:
+            raise RuntimeError(f"FastH3 segment {segment_idx} requires a retained continuation frame.")
+        if segment_idx == 1 and image_path:
+            return self._load_rgb_image(image_path), False
+        return None, False
+
+    def _build_request(self, prompt: str, conditioning_image: Image | None):
+        """Build the typed FastVideo request owned by the FastH3 profile."""
+        from fastvideo.api import GenerationRequest, InputConfig, OutputConfig, SamplingConfig
+
+        return GenerationRequest(
+            prompt=prompt,
+            negative_prompt="",
+            inputs=InputConfig(pil_image=conditioning_image),
+            sampling=SamplingConfig(
+                height=int(self.model_config["height"]),
+                width=int(self.model_config["width"]),
+                num_frames=int(self.model_config["num_frames"]),
+                fps=24,
+                num_inference_steps=int(self.model_config["num_inference_steps"]),
+                guidance_scale=1.0,
+                batch_cfg=False,
+                seed=int(self.model_config["seed"]),
+            ),
+            output=OutputConfig(save_video=False, return_frames=True),
+        )
+
+    def _save_continuation_frame(self, frames: list) -> None:
+        """Retain the last decoded frame as first-frame conditioning."""
+        from PIL import Image
+
+        self.clear_conditioning()
+        self.continuation_image = Image.fromarray(np.ascontiguousarray(frames[-1])).convert("RGB")
+
+    def generate_step(
+        self,
+        prompt: str,
+        segment_idx: int,
+        image_path: str | None,
+        reset_conditioning: bool,
+    ) -> StepResult:
+        """Generate one synchronized FastH3 segment and retain its last frame.
+
+        Later segments use MiniMax H3's first-frame-to-video path. The first
+        conditioned frame and its matching audio duration are trimmed before
+        streaming so adjacent segments do not duplicate media.
+        """
+        if self.generator is None:
+            raise RuntimeError("FastH3 generator is not initialized.")
+        conditioning_image, uses_continuation = self._select_conditioning_image(
+            segment_idx,
+            image_path,
+            reset_conditioning,
+        )
+        request = self._build_request(prompt, conditioning_image)
+        started = time.perf_counter()
+        try:
+            result = self.generator.generate(request)
+        finally:
+            if conditioning_image is not None:
+                conditioning_image.close()
+        torch.cuda.synchronize()
+        generation_ms = (time.perf_counter() - started) * 1000.0
+
+        if isinstance(result, list):
+            raise RuntimeError("FastH3 returned multiple results for one DreamVerse segment.")
+        frames = result.frames
+        if not isinstance(frames, list) or not frames:
+            raise RuntimeError("FastH3 generation did not return decoded frames.")
+        audio = result.audio
+        audio_sample_rate = result.audio_sample_rate
+        if audio is not None and audio_sample_rate is None:
+            raise RuntimeError("FastH3 returned audio without an audio sample rate.")
+
+        save_started = time.perf_counter()
+        self._save_continuation_frame(frames)
+        save_conditioning_ms = (time.perf_counter() - save_started) * 1000.0
+        timings = {
+            "generation_ms": generation_ms,
+            "generation_time_ms": float(result.generation_time or 0.0) * 1000.0,
+            "save_conditioning_ms": save_conditioning_ms,
+            "e2e_latency_ms": (time.perf_counter() - started) * 1000.0,
+        }
+        trim_frames = 1 if uses_continuation else 0
+        print(f"[GPU {self.gpu_id}] FastH3 segment {segment_idx}: "
+              f"{len(frames)} frames, gen={generation_ms:.0f}ms, "
+              f"save_conditioning={save_conditioning_ms:.0f}ms, "
+              f"e2e={timings['e2e_latency_ms']:.0f}ms")
+        return StepResult(
+            frames=frames,
+            audio=audio,
+            audio_sample_rate=audio_sample_rate,
+            timings=timings,
+            head_trim_frames=trim_frames,
+            head_trim_audio_frames=trim_frames,
+        )
+
+    def warmup(self, prompt: str) -> dict[str, float]:
+        """Compile the FastH3 text and first-frame paths before readiness."""
+        warmup_prompt = (prompt or "").strip()
+        if not warmup_prompt:
+            raise RuntimeError("Startup warmup prompt must be non-empty.")
+        print(f"[GPU {self.gpu_id}] FastH3 startup warmup starting "
+              "(synthetic segments: text-to-video, first-frame-to-video)")
+        started = time.perf_counter()
+        text_result = self.generate_step(
+            warmup_prompt,
+            segment_idx=1,
+            image_path=None,
+            reset_conditioning=True,
+        )
+        first_frame_result = self.generate_step(
+            warmup_prompt,
+            segment_idx=2,
+            image_path=None,
+            reset_conditioning=False,
+        )
+        total_ms = (time.perf_counter() - started) * 1000.0
+        self.clear_conditioning()
+        text_ms = float(text_result.timings.get("e2e_latency_ms", 0.0))
+        first_frame_ms = float(first_frame_result.timings.get("e2e_latency_ms", 0.0))
+        print(f"[GPU {self.gpu_id}] FastH3 startup warmup complete: "
+              f"text_to_video={text_ms:.0f}ms, "
+              f"first_frame_to_video={first_frame_ms:.0f}ms, "
+              f"total={total_ms:.0f}ms")
+        return {
+            "warmup_text_to_video_ms": text_ms,
+            "warmup_first_frame_to_video_ms": first_frame_ms,
+            "warmup_total_ms": total_ms,
+        }
+
+    def apply_lora_stack(self, stack: list[tuple[str, float]]) -> tuple[str | None, str | None]:
+        """Reject runtime LoRA mutation because FastH3 uses one startup adapter."""
+        del stack
+        raise RuntimeError("FastH3 uses its fixed startup adapter and does not support runtime LoRA changes.")
@@ -30,7 +30,7 @@ from dreamverse.session_init_image import cleanup_session_init_image, persist_se
 from dreamverse.worker_ipc import MediaChunk, MediaComplete, MediaInit

 from dreamverse.config import (
-    DEFAULT_MODEL_ID,
+    ACTIVE_MODEL_ID,
    GENERATION_SEGMENT_CAP,
    PROMPT_AUTO_SLEEP_MS,
    PROMPT_AUTO_TIMEOUT_MS,
@@ -264,7 +264,7 @@ class SessionController:
            timeout_task = asyncio.create_task(session_timeout())

            # Join the engine on this GPU.
-            await slot.join_user(client_id, model_id=DEFAULT_MODEL_ID)
+            await slot.join_user(client_id, model_id=ACTIVE_MODEL_ID)

            # Notify client they're connected to a GPU.
            await ws_send_json({
@@ -2,13 +2,14 @@ from __future__ import annotations

 import importlib.util
 from pathlib import Path
+from types import ModuleType

 import pytest

 SERVER_DIR = Path(__file__).resolve().parents[1]


-def _load_config_module():
+def _load_config_module() -> ModuleType:
    spec = importlib.util.spec_from_file_location(
        "server_config_test_module",
        SERVER_DIR / "config.py",
@@ -20,7 +21,7 @@ def _load_config_module():
    return module


-def _set_required_prompt_keys(monkeypatch):
+def _set_required_prompt_keys(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setenv("CEREBRAS_API_KEY", "cerebras-key")
    monkeypatch.setenv("GROQ_API_KEY", "groq-key")

@@ -150,3 +151,38 @@ def test_config_rejects_invalid_prompt_provider(monkeypatch):

    with pytest.raises(RuntimeError, match="Invalid FASTVIDEO_PROMPT_PROVIDER"):
        _load_config_module()
+
+
+def test_config_registers_vsa_datafree_fasth3_profile(monkeypatch):
+    """The FastH3 registry entry owns the complete fixed Preview recipe."""
+    _set_required_prompt_keys(monkeypatch)
+
+    module = _load_config_module()
+
+    assert module.MODEL_REGISTRY["fast-h3"] == {
+        "name": "FastH3",
+        "generation_backend": "minimax_h3",
+        "default_sp_size": 4,
+        "model_path": "MiniMaxAI/MiniMax-H3",
+        "adapter_repo": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
+        "adapter_filename": "vsa-datafree/adapter_model.safetensors",
+        "attention_backend": "VIDEO_SPARSE_ATTN_H3",
+        "height": 768,
+        "width": 1344,
+        "num_frames": 124,
+        "num_inference_steps": 5,
+        "seed": 1000,
+    }
+
+
+def test_config_uses_fasth3_sequence_parallel_default(monkeypatch):
+    """Selecting FastH3 defaults DreamVerse to its four-GPU topology."""
+    _set_required_prompt_keys(monkeypatch)
+    monkeypatch.setenv("DREAMVERSE_MODEL_ID", "fast-h3")
+    monkeypatch.delenv("DREAMVERSE_SP_SIZE", raising=False)
+
+    module = _load_config_module()
+
+    assert module.ACTIVE_MODEL_ID == "fast-h3"
+    assert module.MODEL_CONFIG["generation_backend"] == "minimax_h3"
+    assert module.DREAMVERSE_SP_SIZE == 4
@@ -0,0 +1,9 @@
+from dreamverse.generation_worker import _create_generation_backend
+from dreamverse.ltx2_generation import LTX2GenerationBackend
+
+
+def test_create_generation_backend_ltx2_module_import():
+    backend = _create_generation_backend("ltx2", gpu_id=3)
+
+    assert isinstance(backend, LTX2GenerationBackend)
+    assert backend.gpu_id == 3
@@ -63,6 +63,14 @@ def test_get_available_gpus_defaults_to_first_visible_device(monkeypatch):
    assert gpu_pool.get_available_gpus() == [3]


+def test_get_available_gpus_defaults_to_active_model_sequence_parallel_size(monkeypatch):
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "0,1,2,3,4")
+    monkeypatch.delenv("FASTVIDEO_GPU_COUNT", raising=False)
+    monkeypatch.setattr(gpu_pool, "DREAMVERSE_SP_SIZE", 4)
+
+    assert gpu_pool.get_available_gpus() == [0, 1, 2, 3]
+
+
 def test_get_available_gpus_rejects_invalid_gpu_count(monkeypatch):
    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
    monkeypatch.setenv("FASTVIDEO_GPU_COUNT", "zero")
@@ -71,6 +79,23 @@ def test_get_available_gpus_rejects_invalid_gpu_count(monkeypatch):
        gpu_pool.get_available_gpus()


+def test_join_user_failed_reload_marks_model_uninitialized(monkeypatch):
+    """A failed model reload forces the next join to reload a model."""
+    slot = gpu_pool.GPUSlot(gpu_id=0, cuda_device="0")
+    slot.current_model_id = "fast-ltx2"
+
+    async def fake_send_command(command, timeout):
+        del command, timeout
+        return gpu_pool.WorkerError(user_id="__reload__", message="load failed")
+
+    monkeypatch.setattr(slot, "_send_command", fake_send_command)
+
+    with pytest.raises(RuntimeError, match="Model reload failed"):
+        asyncio.run(slot.join_user("client-id", model_id="fast-h3"))
+
+    assert slot.current_model_id is None
+
+
 def test_send_command_raises_on_worker_death():
    """A worker that consumes a command and exits without replying must
    surface as RuntimeError via sentinel detection, not after the long
@@ -92,9 +117,9 @@ def test_send_command_raises_on_worker_death():
    ready = resp_q.get(timeout=30.0)
    assert ready == "READY"

-    async def runner():
+    async def runner() -> None:
        slot = gpu_pool.GPUSlot(gpu_id=0, cuda_device="0")
-        slot.process = proc
+        slot.process = proc  # type: ignore[assignment]
        slot.command_queue = cmd_q
        slot.response_queue = resp_q

@@ -13,15 +13,14 @@ FORBIDDEN_PREFIXES = (
    "fastvideo.models",
    "fastvideo.layers",
    "fastvideo.worker",
-    "fastvideo.fastvideo_args",
 )
 ALLOWED_INTERNAL_IMPORTS = {
    (
-        "video_generation.py",
+        "ltx2_generation.py",
        "fastvideo.models.audio.ltx2_audio_processing",
    ),
    (
-        "video_generation.py",
+        "ltx2_generation.py",
        "fastvideo.models.loader.component_loader",
    ),
 }
@@ -0,0 +1,246 @@
+from __future__ import annotations
+
+import os
+from types import SimpleNamespace
+from typing import Any
+
+import numpy as np
+import pytest
+
+import dreamverse.generation_worker as generation_worker
+from dreamverse.minimax_h3_generation import MiniMaxH3GenerationBackend
+
+
+FASTH3_MODEL_CONFIG = {
+    "name": "FastH3",
+    "generation_backend": "minimax_h3",
+    "default_sp_size": 4,
+    "model_path": "MiniMaxAI/MiniMax-H3",
+    "adapter_repo": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
+    "adapter_filename": "vsa-datafree/adapter_model.safetensors",
+    "attention_backend": "VIDEO_SPARSE_ATTN_H3",
+    "height": 768,
+    "width": 1344,
+    "num_frames": 124,
+    "num_inference_steps": 5,
+    "seed": 1000,
+}
+
+
+class _RecordingGenerator:
+    """Record typed requests and return small synchronized media fixtures."""
+
+    def __init__(self) -> None:
+        self.requests: list[Any] = []
+        self.conditioning_pixels: list[np.ndarray | None] = []
+
+    def generate(self, request):
+        """Capture the request and return two tiny video frames with audio."""
+        self.requests.append(request)
+        conditioning_image = request.inputs.pil_image
+        self.conditioning_pixels.append(
+            None if conditioning_image is None else np.asarray(conditioning_image).copy())
+        frames = [
+            np.full((2, 3, 3), 10, dtype=np.uint8),
+            np.full((2, 3, 3), 20, dtype=np.uint8),
+        ]
+        return SimpleNamespace(
+            frames=frames,
+            audio=np.zeros((2, 16), dtype=np.float32),
+            audio_sample_rate=44100,
+            generation_time=0.25,
+        )
+
+
+def test_initialize_builds_vsa_datafree_fasth3_generator(monkeypatch):
+    """Initialization translates the DreamVerse profile into typed FastVideo config."""
+    from fastvideo import VideoGenerator
+
+    captured = {}
+    fake_generator = SimpleNamespace(shutdown=lambda: None)
+
+    def fake_from_config(config):
+        captured["config"] = config
+        return fake_generator
+
+    def fake_download(**kwargs):
+        captured["download"] = kwargs
+        return f"/models/{kwargs['filename']}"
+
+    monkeypatch.setattr("huggingface_hub.hf_hub_download", fake_download)
+    monkeypatch.setattr(VideoGenerator, "from_config", fake_from_config)
+    monkeypatch.setattr("dreamverse.minimax_h3_generation.DREAMVERSE_SP_SIZE", 4)
+    monkeypatch.setenv("FASTVIDEO_ATTENTION_BACKEND", "test-attention")
+    monkeypatch.setenv("FASTVIDEO_FA4", "0")
+    monkeypatch.setenv("FASTVIDEO_MINIMAX_H3_FUSIONS", "0")
+    monkeypatch.setenv("FASTVIDEO_VSA_SM100A", "1")
+    monkeypatch.setenv("FASTVIDEO_INFERENCE_TORCH_COMPILE", "1")
+
+    backend = MiniMaxH3GenerationBackend(gpu_id=0)
+    monkeypatch.setattr(backend, "_gpu_mem", lambda: "alloc=0.00GiB, reserved=0.00GiB")
+    backend.initialize(FASTH3_MODEL_CONFIG)
+
+    config = captured["config"]
+    assert captured["download"] == {
+        "repo_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
+        "filename": "vsa-datafree/adapter_model.safetensors",
+    }
+    assert config.model_path == "MiniMaxAI/MiniMax-H3"
+    assert config.pipeline.components.lora_path.endswith("vsa-datafree/adapter_model.safetensors")
+    assert config.pipeline.components.lora_strength == 1.0
+    assert config.pipeline.experimental == {}
+    assert config.engine.attention.backend == "VIDEO_SPARSE_ATTN_H3"
+    assert config.engine.attention.vsa_sparsity == 0.9
+    assert config.engine.attention.vsa_tile_size == 64
+    assert config.engine.compile.regional is False
+    assert config.pipeline.model.vae_parallel_decode is True
+    assert config.pipeline.model.vae_parallel_decode_strategy == "gather"
+    assert config.engine.num_gpus == 4
+    assert config.engine.parallelism.tp_size == 1
+    assert config.engine.parallelism.sp_size == 4
+    assert config.engine.offload.dit is False
+    assert config.engine.offload.dit_layerwise is False
+    assert config.engine.offload.text_encoder is True
+    assert config.engine.offload.vae is True
+    assert config.engine.compile.vae_enabled is True
+    assert config.engine.use_fsdp_inference is False
+    assert os.environ["FASTVIDEO_ATTENTION_BACKEND"] == "VIDEO_SPARSE_ATTN_H3"
+    assert os.environ["FASTVIDEO_FA4"] == "1"
+    assert os.environ["FASTVIDEO_MINIMAX_H3_FUSIONS"] == "all"
+    assert os.environ["FASTVIDEO_VSA_SM100A"] == "0"
+    assert "FASTVIDEO_INFERENCE_TORCH_COMPILE" not in os.environ
+
+
+def test_initialize_selects_declared_generation_backend(monkeypatch):
+    """The GPU worker constructs the backend that the active model profile declares."""
+    from unittest.mock import Mock
+
+    selected_backend = Mock()
+    monkeypatch.setattr(
+        generation_worker,
+        "_create_generation_backend",
+        lambda backend_name, gpu_id: selected_backend,
+    )
+    worker = generation_worker.VideoGenerationWorker(gpu_id=3)
+
+    worker.initialize(FASTH3_MODEL_CONFIG)
+
+    assert worker.backend_name == "minimax_h3"
+    assert worker.backend is selected_backend
+    selected_backend.initialize.assert_called_once_with(FASTH3_MODEL_CONFIG)
+
+
+def test_initialize_failure_clears_backend_ownership(monkeypatch):
+    """A failed family change leaves the GPU worker explicitly uninitialized."""
+    ltx_backend = SimpleNamespace(initialize=lambda config: None, shutdown=lambda: None)
+
+    def fail_initialize(config):
+        del config
+        raise RuntimeError("load failed")
+
+    fasth3_backend = SimpleNamespace(
+        initialize=fail_initialize,
+        shutdown=lambda: None,
+    )
+    backends = {
+        "ltx2": ltx_backend,
+        "minimax_h3": fasth3_backend,
+    }
+    monkeypatch.setattr(
+        generation_worker,
+        "_create_generation_backend",
+        lambda backend_name, gpu_id: backends[backend_name],
+    )
+    worker = generation_worker.VideoGenerationWorker(gpu_id=3)
+    worker.initialize({"generation_backend": "ltx2"})
+
+    with pytest.raises(RuntimeError, match="load failed"):
+        worker.initialize(FASTH3_MODEL_CONFIG)
+
+    assert worker.backend is None
+    assert worker.backend_name is None
+    assert worker.model_config == {"generation_backend": "ltx2"}
+
+
+def test_generate_step_uses_last_frame_for_continuation(monkeypatch):
+    """A later segment receives the prior segment's last decoded frame."""
+    backend = MiniMaxH3GenerationBackend(gpu_id=0)
+    backend.model_config = dict(FASTH3_MODEL_CONFIG)
+    backend.generator = _RecordingGenerator()
+    monkeypatch.setattr("dreamverse.minimax_h3_generation.torch.cuda.synchronize", lambda: None)
+
+    first_result = backend.generate_step(
+        "first prompt",
+        segment_idx=1,
+        image_path=None,
+        reset_conditioning=True,
+    )
+    second_result = backend.generate_step(
+        "second prompt",
+        segment_idx=2,
+        image_path=None,
+        reset_conditioning=False,
+    )
+
+    first_request = backend.generator.requests[0]
+    assert first_request.inputs.pil_image is None
+    assert first_request.negative_prompt == ""
+    assert first_request.sampling.height == 768
+    assert first_request.sampling.width == 1344
+    assert first_request.sampling.num_frames == 124
+    assert first_request.sampling.num_inference_steps == 5
+    assert first_request.sampling.fps == 24
+    assert first_request.sampling.guidance_scale == 1.0
+    assert first_request.sampling.batch_cfg is False
+    assert first_request.sampling.seed == 1000
+    assert first_request.output.save_video is False
+    assert first_request.output.return_frames is True
+    assert backend.generator.conditioning_pixels[1].tolist() == np.full((2, 3, 3), 20).tolist()
+    assert first_result.head_trim_frames == 0
+    assert first_result.head_trim_audio_frames == 0
+    assert second_result.head_trim_frames == 1
+    assert second_result.head_trim_audio_frames == 1
+    assert second_result.audio_sample_rate == 44100
+
+
+def test_generate_step_reset_uses_text_to_video_path(monkeypatch):
+    """Resetting continuation produces an unconditioned text-to-video request."""
+    backend = MiniMaxH3GenerationBackend(gpu_id=0)
+    backend.model_config = dict(FASTH3_MODEL_CONFIG)
+    backend.generator = _RecordingGenerator()
+    monkeypatch.setattr("dreamverse.minimax_h3_generation.torch.cuda.synchronize", lambda: None)
+
+    backend.generate_step("first prompt", 1, None, True)
+    reset_result = backend.generate_step("reset prompt", 2, None, True)
+
+    assert backend.generator.requests[-1].inputs.pil_image is None
+    assert reset_result.head_trim_frames == 0
+    assert reset_result.head_trim_audio_frames == 0
+
+
+def test_generate_step_missing_continuation_frame(monkeypatch):
+    """A later segment fails when no reset or retained frame defines its input."""
+    backend = MiniMaxH3GenerationBackend(gpu_id=0)
+    backend.model_config = dict(FASTH3_MODEL_CONFIG)
+    backend.generator = _RecordingGenerator()
+
+    with pytest.raises(RuntimeError, match="requires a retained continuation frame"):
+        backend.generate_step("later prompt", 2, None, False)
+
+    assert backend.generator.requests == []
+
+
+def test_warmup_exercises_text_and_first_frame_paths(monkeypatch):
+    """Warmup covers both request shapes used by a DreamVerse session."""
+    backend = MiniMaxH3GenerationBackend(gpu_id=0)
+    backend.model_config = dict(FASTH3_MODEL_CONFIG)
+    backend.generator = _RecordingGenerator()
+    monkeypatch.setattr("dreamverse.minimax_h3_generation.torch.cuda.synchronize", lambda: None)
+
+    timings = backend.warmup("warmup prompt")
+
+    assert backend.generator.conditioning_pixels[0] is None
+    assert backend.generator.conditioning_pixels[1] is not None
+    assert backend.continuation_image is None
+    assert "warmup_text_to_video_ms" in timings
+    assert "warmup_first_frame_to_video_ms" in timings
@@ -331,11 +331,11 @@ def test_rewrite_prompt_sequence_accepts_numbered_prose_output():
    ]


-def test_enhance_prompt_prefers_cerebras_before_groq_fallback():
+def test_enhance_prompt_uses_groq_when_it_returns_first():
    enhancer = _build_staged_enhancer(
        cerebras_payload=_chat_payload_with_content('{"prompt":"Cerebras prompt"}'),
        groq_payload=_chat_payload_with_content('{"prompt":"Groq prompt"}'),
-        cerebras_delay_s=0.01,
+        cerebras_delay_s=0.08,
        groq_delay_s=0.01,
    )

@@ -346,12 +346,12 @@ def test_enhance_prompt_prefers_cerebras_before_groq_fallback():

    assert result.fallback_used is False
    assert result.error is None
-    assert result.provider == "cerebras"
+    assert result.provider == "groq"
    assert result.model == "gpt-test"
-    assert result.prompt == "Cerebras prompt"
+    assert result.prompt == "Groq prompt"
    assert enhancer.get_provider_success_counts() == {
-        "cerebras": 1,
-        "groq": 0,
+        "cerebras": 0,
+        "groq": 1,
    }


@@ -70,7 +70,7 @@
                <mxCell id="dispatcher" value="command dispatcher&#xa;&#xa;gpu_worker_process() branches on&#xa;CommandType; asserts payload type&#xa;&#xa;INIT / WARMUP / RELOAD_MODEL&#xa;USER_JOIN / USER_STEP / USER_LEAVE&#xa;SHUTDOWN" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffe6cc;strokeColor=#d79b00;fontSize=11;align=left;spacingLeft=10;spacingTop=8;fontStyle=1;" parent="1" vertex="1">
                    <mxGeometry x="120" y="1120" width="240" height="120" as="geometry"/>
                </mxCell>
-                <mxCell id="do_step" value="VideoGenerationWorker.generate_step()&#xa;video_generation.py:380&#xa;&#xa;reads + updates ContinuationState,&#xa;calls generator" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;align=left;spacingLeft=10;spacingTop=8;fontStyle=1;" parent="1" vertex="1">
+                <mxCell id="do_step" value="VideoGenerationWorker.generate_step()&#xa;ltx2_generation.py:380&#xa;&#xa;reads + updates ContinuationState,&#xa;calls generator" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;align=left;spacingLeft=10;spacingTop=8;fontStyle=1;" parent="1" vertex="1">
                    <mxGeometry x="460" y="1120" width="240" height="120" as="geometry"/>
                </mxCell>
                <mxCell id="stream_av" value="stream_fmp4()&#xa;av_streaming.py:121&#xa;&#xa;trims overlap, pipes to ffmpeg,&#xa;publishes StreamInit / StreamChunk /&#xa;StreamComplete via callback" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#b1d8d7;strokeColor=#23445d;fontSize=11;align=left;spacingLeft=10;spacingTop=8;fontStyle=1;" parent="1" vertex="1">
@@ -79,13 +79,13 @@
                <mxCell id="Ot8BU52QTIb4EhyRSe7I-2" value="" style="edgeStyle=none;html=1;" parent="1" source="generator" target="Ot8BU52QTIb4EhyRSe7I-1" edge="1">
                    <mxGeometry relative="1" as="geometry"/>
                </mxCell>
-                <mxCell id="generator" value="VideoGenerator  (fastvideo)&#xa;&#xa;LTX2 DiT + refine upsampler&#xa;FP4 quant, torch.compile&#xa;&#xa;owned by VideoGenerationWorker&#xa;video_generation.py:211" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;" parent="1" vertex="1">
+                <mxCell id="generator" value="VideoGenerator  (fastvideo)&#xa;&#xa;LTX2 DiT + refine upsampler&#xa;FP4 quant, torch.compile&#xa;&#xa;owned by VideoGenerationWorker&#xa;ltx2_generation.py:211" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;" parent="1" vertex="1">
                    <mxGeometry x="460" y="1300" width="240" height="100" as="geometry"/>
                </mxCell>
                <mxCell id="ffmpeg" value="ffmpeg subprocess&#xa;&#xa;libx264 / *_nvenc&#xa;fragmented mp4" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffffff;strokeColor=#d79b00;fontSize=11;" parent="1" vertex="1">
                    <mxGeometry x="800" y="1300" width="260" height="100" as="geometry"/>
                </mxCell>
-                <mxCell id="caches" value="ContinuationState&#xa;video_generation.py:89&#xa;&#xa;• video_images: list[PIL.Image]&#xa;• audio_latents: torch.Tensor (CPU)&#xa;&#xa;carried across segments" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;align=left;spacingLeft=10;spacingTop=8;" parent="1" vertex="1">
+                <mxCell id="caches" value="ContinuationState&#xa;ltx2_generation.py:89&#xa;&#xa;• video_images: list[PIL.Image]&#xa;• audio_latents: torch.Tensor (CPU)&#xa;&#xa;carried across segments" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#e1d5e7;strokeColor=#9673a6;fontSize=11;align=left;spacingLeft=10;spacingTop=8;" parent="1" vertex="1">
                    <mxGeometry x="120" y="1300" width="240" height="100" as="geometry"/>
                </mxCell>
                <mxCell id="e_cp" value="acquire" style="edgeStyle=orthogonalEdgeStyle;rounded=0;html=1;strokeColor=#6c8ebf;endArrow=classic;fontSize=11;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" parent="1" source="client" target="pool" edge="1">
@@ -207,7 +207,7 @@
                        </Array>
                    </mxGeometry>
                </mxCell>
-                <mxCell id="e_dsg" value="generator.generate_video()" style="edgeStyle=orthogonalEdgeStyle;rounded=0;html=1;strokeColor=#9673a6;endArrow=classic;fontSize=10;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" parent="1" source="do_step" target="generator" edge="1">
+                <mxCell id="e_dsg" value="generator.generate()" style="edgeStyle=orthogonalEdgeStyle;rounded=0;html=1;strokeColor=#9673a6;endArrow=classic;fontSize=10;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" parent="1" source="do_step" target="generator" edge="1">
                    <mxGeometry relative="1" as="geometry"/>
                </mxCell>
                <mxCell id="e_dscache" value="read / write" style="edgeStyle=orthogonalEdgeStyle;rounded=0;html=1;strokeColor=#d6b656;endArrow=classic;startArrow=classic;fontSize=10;exitX=0;exitY=0.8;exitDx=0;exitDy=0;entryX=1;entryY=0.2;entryDx=0;entryDy=0;" parent="1" source="do_step" target="caches" edge="1">
@@ -250,7 +250,7 @@
                        <mxPoint x="690" y="880"/>
                    </Array>
                </mxCell>
-                <mxCell id="legend" value="Legend&#xa;&#xa;■ blue  client / external&#xa;■ green  main-process pool/slot&#xa;    (methods — italic label)&#xa;■ yellow  containers (routing state)&#xa;■ red  IPC primitives (mp.Queue, mp.RawArray)&#xa;&#xa;Worker subprocess modules:&#xa;■ orange  gpu_pool.py (dispatcher)&#xa;■ lavender  video_generation.py&#xa;■ teal  av_streaming.py&#xa;■ gray  worker_ipc.py (shared types)&#xa;&#xa;Flow:&#xa;  client → pool → slot&#xa;  → _send_command(_tagged) → command_queue&#xa;  → dispatcher → generate_step()&#xa;  → stream_fmp4() → ffmpeg&#xa;  → shared_buf + response_queue&#xa;  → _response_reader → futures / stream_queues&#xa;  → client awaits (via main.py AV loop)" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#999999;fontSize=11;align=left;spacingLeft=10;spacingTop=8;" parent="1" vertex="1">
+                <mxCell id="legend" value="Legend&#xa;&#xa;■ blue  client / external&#xa;■ green  main-process pool/slot&#xa;    (methods — italic label)&#xa;■ yellow  containers (routing state)&#xa;■ red  IPC primitives (mp.Queue, mp.RawArray)&#xa;&#xa;Worker subprocess modules:&#xa;■ orange  gpu_pool.py (dispatcher)&#xa;■ lavender  ltx2_generation.py&#xa;■ teal  av_streaming.py&#xa;■ gray  worker_ipc.py (shared types)&#xa;&#xa;Flow:&#xa;  client → pool → slot&#xa;  → _send_command(_tagged) → command_queue&#xa;  → dispatcher → generate_step()&#xa;  → stream_fmp4() → ffmpeg&#xa;  → shared_buf + response_queue&#xa;  → _response_reader → futures / stream_queues&#xa;  → client awaits (via main.py AV loop)" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#999999;fontSize=11;align=left;spacingLeft=10;spacingTop=8;" parent="1" vertex="1">
                    <mxGeometry x="39" y="-200" width="270" height="380" as="geometry"/>
                </mxCell>
                <mxCell id="Ot8BU52QTIb4EhyRSe7I-1" value="FastVideo video_generator" style="whiteSpace=wrap;html=1;fontSize=11;fillColor=#e1d5e7;strokeColor=#9673a6;rounded=1;" parent="1" vertex="1">
@@ -389,10 +389,10 @@
                <mxCell id="cw2" value="from fastvideo.entrypoints.video_generator import VideoGenerator&#xa;from fastvideo.models.dits.ltx2 import DEFAULT_LTX2_AUDIO_*&#xa;&#xa;** Dreamverse reaches into fastvideo internals here **" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffe0b2;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
                    <mxGeometry x="675" y="695" width="550" height="60" as="geometry"/>
                </mxCell>
-                <mxCell id="cw3" value="on Command(INIT):&#xa;   VideoGenerationWorker.initialize()    (video_generation.py:247)&#xa;      maybe_download_model(model_id)&#xa;      VideoGenerator.from_pretrained(path, FP4Config, PipelineConfig)&#xa;      load audio VAE, resolve refine upsampler&#xa;      resp_q.put(InitAck(success=True))" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffffff;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
+                <mxCell id="cw3" value="on Command(INIT):&#xa;   VideoGenerationWorker.initialize()    (ltx2_generation.py:247)&#xa;      maybe_download_model(model_id)&#xa;      VideoGenerator.from_pretrained(path, FP4Config, PipelineConfig)&#xa;      load audio VAE, resolve refine upsampler&#xa;      resp_q.put(InitAck(success=True))" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffffff;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
                    <mxGeometry x="675" y="765" width="550" height="95" as="geometry"/>
                </mxCell>
-                <mxCell id="cw4" value="on Command(WARMUP) with WarmupPayload:&#xa;   VideoGenerationWorker.warmup(payload.prompt)    (video_generation.py:518)&#xa;      two synthetic segments prime caches + torch.compile&#xa;   resp_q.put(WarmupComplete(timings=...))" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffffff;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
+                <mxCell id="cw4" value="on Command(WARMUP) with WarmupPayload:&#xa;   VideoGenerationWorker.warmup(payload.prompt)    (ltx2_generation.py:518)&#xa;      two synthetic segments prime caches + torch.compile&#xa;   resp_q.put(WarmupComplete(timings=...))" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffffff;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
                    <mxGeometry x="675" y="870" width="550" height="55" as="geometry"/>
                </mxCell>
                <mxCell id="cw5" value="enter main worker loop → waits for JOIN_USER / USER_STEP / LEAVE" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#c8e6c9;strokeColor=#388e3c;fontSize=11;fontStyle=1;fontFamily=monospace;" parent="1" vertex="1">
@@ -534,7 +534,7 @@
                        <mxPoint x="1040" y="1610" as="targetPoint"/>
                    </mxGeometry>
                </mxCell>
-                <mxCell id="dm11a" value="10a. worker runs:&#xa;VideoGenerationWorker.generate_step()&#xa;    (video_generation.py:380)&#xa;  → generator.generate_video()&#xa;  → updates ContinuationState&#xa;then stream_fmp4() (av_streaming.py:121)&#xa;  → ffmpeg (rawvideo+wav → fmp4)" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffe0b2;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
+                <mxCell id="dm11a" value="10a. worker runs:&#xa;VideoGenerationWorker.generate_step()&#xa;    (ltx2_generation.py:380)&#xa;  → generator.generate()&#xa;  → updates ContinuationState&#xa;then stream_fmp4() (av_streaming.py:121)&#xa;  → ffmpeg (rawvideo+wav → fmp4)" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#ffe0b2;strokeColor=#d79b00;fontSize=10;align=left;spacingLeft=8;fontFamily=monospace;" parent="1" vertex="1">
                    <mxGeometry x="955" y="1640" width="180" height="70" as="geometry"/>
                </mxCell>
                <mxCell id="dm11" value="10b. resp_q.put(MediaInit / MediaChunk / MediaComplete / StepComplete)" style="endArrow=classic;html=1;strokeColor=#b85450;fontSize=10;labelBackgroundColor=#ffffff;" parent="1" edge="1">
@@ -64,8 +64,8 @@ generator:
    # internal: pipeline_config.dit_config.quant_config = FP4Config()
    # set in gpu_pool.py:280 (via the legacy in-place mutation). The
    # public typed surface resolves "NVFP4" to NVFP4Config() and pins
-    # it on dit_config in FastVideoArgs.__post_init__. Comment this
-    # block out on hosts without flashinfer / NVFP4 hardware.
+    # it on dit_config when resolution materializes the PipelineConfig.
+    # Comment this block out on hosts without flashinfer / NVFP4 hardware.
    quantization:
      transformer_quant: NVFP4

@@ -67,7 +67,7 @@ test.describe('preset prompt generation', () => {
    // on a B200 plus encode/transfer time. The "Continuation flipped
    // to Generating + Leave button rendered" pair above is the proof
    // the integration works: FE → /readyz → /curated-presets → WS
-    // /ws → BE → GPU pool → VideoGenerator.generate_video, all green.
+    // /ws → BE → GPU pool → VideoGenerator.generate, all green.
    const video = page.locator('video').first();
    await expect(video).toHaveCount(1);
  });
@@ -0,0 +1,63 @@
+import { expect, test } from '@playwright/test';
+
+import { skipWithoutMock } from './helpers';
+
+test.describe('create job interactions', () => {
+  skipWithoutMock();
+
+  for (const jobType of ['inference', 'finetuning', 'distillation']) {
+    test(`${jobType} remains interactive after repeated dialog dismissals`, async ({ page }) => {
+      await page.goto(`/${jobType}`);
+      const trigger = page.getByRole('button', { name: 'Create Job', exact: true });
+      const dialog = page.getByRole('dialog');
+
+      // Exercise both dismissal paths and reopen without reloading the page.
+      for (const closeWithEscape of [false, true]) {
+        await trigger.click();
+        await page.getByRole('menuitem').first().click();
+        await expect(dialog).toBeVisible();
+        if (closeWithEscape) {
+          await page.keyboard.press('Escape');
+        } else {
+          await dialog.getByRole('button', { name: 'Close', exact: true }).click();
+        }
+        await expect(dialog).toBeHidden();
+        await expect(page.locator('body')).toHaveCSS('pointer-events', 'auto');
+        await expect(trigger).toBeFocused();
+      }
+
+      await page.getByRole('link', { name: 'Datasets', exact: true }).click();
+      await expect(page).toHaveURL(/\/datasets$/);
+    });
+  }
+
+  test('preserves keyboard menu dismissal and dialog focus trapping', async ({ page }) => {
+    await page.goto('/inference');
+    const trigger = page.getByRole('button', { name: 'Create Job', exact: true });
+    await trigger.focus();
+    await page.keyboard.press('Enter');
+    const firstItem = page.getByRole('menuitem').first();
+    await expect(firstItem).toBeFocused();
+    await page.keyboard.press('Escape');
+    await expect(page.getByRole('menu')).toBeHidden();
+    await expect(trigger).toBeFocused();
+    await expect(page.locator('body')).toHaveCSS('pointer-events', 'auto');
+
+    await page.keyboard.press('Enter');
+    await expect(firstItem).toBeFocused();
+    await page.keyboard.press('Enter');
+    const dialog = page.getByRole('dialog');
+    await expect(dialog).toBeVisible();
+    await expect(dialog.getByLabel('Name (optional)')).toBeFocused();
+
+    // Shift+Tab from the first field wraps to Close, then Tab wraps back.
+    await page.keyboard.press('Shift+Tab');
+    await expect(dialog.getByRole('button', { name: 'Close', exact: true })).toBeFocused();
+    await page.keyboard.press('Tab');
+    await expect(dialog.getByLabel('Name (optional)')).toBeFocused();
+    await page.keyboard.press('Escape');
+    await expect(dialog).toBeHidden();
+    await expect(trigger).toBeFocused();
+    await expect(page.locator('body')).toHaveCSS('pointer-events', 'auto');
+  });
+});
@@ -1,6 +1,6 @@
 import { expect, test } from '@playwright/test';

-import { skipWithoutMock } from './helpers';
+import { API_BASE, skipWithoutMock } from './helpers';

 /**
 * Create-job flow: open the Create Job modal on /inference, fill the prompt
@@ -10,7 +10,8 @@ import { skipWithoutMock } from './helpers';
 test.describe('create inference job', () => {
  skipWithoutMock();

-  test('creates a T2V job and shows it in the queue', async ({ page }) => {
+  test('creates a T2V job and starts it without refreshing', async ({ page, request }) => {
+    await request.put(`${API_BASE}/settings`, { data: { autoStartJob: false } });
    await page.goto('/inference');

    // The trigger opens a real menu on click, so this path works for touch,
@@ -38,5 +39,20 @@ test.describe('create inference job', () => {
    // Modal closes and the queue refreshes with the newly created job.
    await expect(dialog).toBeHidden();
    await expect(page.getByText(prompt)).toBeVisible();
+    await expect(page.locator('body')).toHaveCSS('pointer-events', 'auto');
+
+    const card = page.getByRole('article').filter({ hasText: prompt });
+    await expect(card.getByText('pending', { exact: true })).toBeVisible();
+    const started = page.waitForResponse((response) =>
+      response.url().startsWith(`${API_BASE}/jobs/`) &&
+      response.url().endsWith('/start') &&
+      response.request().method() === 'POST',
+    );
+    await card.getByRole('button', { name: 'Start', exact: true }).click();
+    expect((await started).ok()).toBe(true);
+    await expect(card.getByText('running', { exact: true })).toBeVisible();
+
+    await page.getByRole('link', { name: 'Datasets', exact: true }).click();
+    await expect(page).toHaveURL(/\/datasets$/);
  });
 });
@@ -862,23 +862,40 @@ class JobRunner:
            sp_size,
        )

-        gen = VideoGenerator.from_pretrained(
-            model_id,
-            workload_type=workload_type,
-            num_gpus=num_gpus,
-            dit_layerwise_offload=dit_layerwise_offload,
-            **({
-                "override_pipeline_cls_name": override_pipeline_cls_name
-            } if override_pipeline_cls_name else {}),
-            dit_cpu_offload=dit_cpu_offload,
-            text_encoder_cpu_offload=text_encoder_cpu_offload,
-            vae_cpu_offload=vae_cpu_offload,
-            image_encoder_cpu_offload=image_encoder_cpu_offload,
-            use_fsdp_inference=use_fsdp_inference,
-            enable_torch_compile=enable_torch_compile,
-            VSA_sparsity=vsa_sparsity,
-            tp_size=tp_size,
-            sp_size=sp_size,
+        gen = VideoGenerator.from_config(
+            {
+                "model_path": model_id,
+                "engine": {
+                    "num_gpus": num_gpus,
+                    "parallelism": {
+                        "tp_size": tp_size,
+                        "sp_size": sp_size,
+                    },
+                    "offload": {
+                        "dit": dit_cpu_offload,
+                        "dit_layerwise": dit_layerwise_offload,
+                        "text_encoder": text_encoder_cpu_offload,
+                        "image_encoder": image_encoder_cpu_offload,
+                        "vae": vae_cpu_offload,
+                    },
+                    "compile": {
+                        "enabled": enable_torch_compile
+                    },
+                    "attention": {
+                        "vsa_sparsity": vsa_sparsity
+                    },
+                    "use_fsdp_inference": use_fsdp_inference,
+                },
+                "pipeline": {
+                    "workload_type":
+                    workload_type,
+                    **({
+                        "components": {
+                            "override_pipeline_cls_name": override_pipeline_cls_name
+                        }
+                    } if override_pipeline_cls_name else {}),
+                },
+            },
            log_queue=log_queue,
        )

@@ -1103,30 +1120,33 @@ class JobRunner:
            # Without a name FastVideo derives the filename from the prompt.
            safe_name = re.sub(r'[\\/:*?"<>|]+', "", job.name).strip().strip(".")
            output_target = (os.path.join(job_output_dir, f"{safe_name[:80]}.mp4") if safe_name else job_output_dir)
-            gen_kwargs: dict[str, Any] = {
+            request: dict[str, Any] = {
                "prompt": job.prompt,
-                "output_path": output_target,
-                "save_video": True,
-                "num_inference_steps": job.num_inference_steps,
-                "num_frames": job.num_frames,
-                "height": job.height,
-                "width": job.width,
-                "guidance_scale": job.guidance_scale,
-                "guidance_rescale": job.guidance_rescale,
-                "fps": job.fps,
-                "seed": job.seed,
                "negative_prompt": job.negative_prompt or "",
-                "log_queue": log_queue,
+                "sampling": {
+                    "num_inference_steps": job.num_inference_steps,
+                    "num_frames": job.num_frames,
+                    "height": job.height,
+                    "width": job.width,
+                    "guidance_scale": job.guidance_scale,
+                    "guidance_rescale": job.guidance_rescale,
+                    "fps": job.fps,
+                    "seed": job.seed,
+                },
+                "output": {
+                    "output_path": output_target,
+                    "save_video": True,
+                },
            }
            if job.image_path:
-                gen_kwargs["image_path"] = job.image_path
+                request.setdefault("inputs", {})["image_path"] = job.image_path
            if job.references:
-                gen_kwargs["references"] = _build_h3_references(job.references)
+                request.setdefault("inputs", {})["references"] = _build_h3_references(job.references)
            if job.last_image_path:
                # _prepare_fl2va requires a PIL image, not a path.
                from PIL import Image as _PILImage
-                gen_kwargs["last_image"] = _PILImage.open(job.last_image_path)
-            generator.generate_video(**gen_kwargs)
+                request.setdefault("inputs", {})["last_image"] = _PILImage.open(job.last_image_path)
+            generator.generate(request, log_queue=log_queue)

            buf.phase = "saving"
            logger.info("Generation completed, searching for output file...")
@@ -17,20 +17,11 @@
 		"start:all": "concurrently --kill-others-on-fail \"npm:start:api\" \"npm:start:web\""
 	},
 	"dependencies": {
-		"@radix-ui/react-dialog": "^1.1.0",
-		"@radix-ui/react-dropdown-menu": "^2.1.24",
-		"@radix-ui/react-label": "^2.1.8",
-		"@radix-ui/react-scroll-area": "^1.2.10",
-		"@radix-ui/react-select": "^2.2.6",
-		"@radix-ui/react-separator": "^1.1.8",
-		"@radix-ui/react-slider": "^1.2.0",
-		"@radix-ui/react-slot": "^1.2.4",
-		"@radix-ui/react-switch": "^1.1.0",
-		"@radix-ui/react-tabs": "^1.1.0",
 		"class-variance-authority": "^0.7.1",
 		"clsx": "^2.1.1",
 		"lucide-react": "^0.577.0",
 		"next": "15.5.18",
+		"radix-ui": "^1.6.7",
 		"react": "^19.1.0",
 		"react-dom": "^19.1.0",
 		"sonner": "^2.0.7",
@@ -1,24 +1,26 @@
-import { render, screen } from '@testing-library/react';
+import { render, screen, waitFor, within } from '@testing-library/react';
 import userEvent from '@testing-library/user-event';
-import { describe, expect, it, vi } from 'vitest';
+import { beforeEach, describe, expect, it, vi } from 'vitest';

 import CreateJobButton from './CreateJobButton';
+import { getDatasets, getModels } from '@/lib/api';

-vi.mock('./CreateJobModal', () => ({
-  default: ({
-    isOpen,
-    workloadType,
-  }: {
-    isOpen: boolean;
-    workloadType: string;
-  }) =>
-    isOpen ? (
-      <div role="dialog" data-workload-type={workloadType}>
-        Create job form
-      </div>
-    ) : null,
+vi.mock('@/lib/api', () => ({
+  createJob: vi.fn(),
+  getModels: vi.fn(),
+  getDatasets: vi.fn(),
+  uploadImage: vi.fn(),
+  getSettings: vi.fn(),
+  updateSettings: vi.fn(),
 }));

+beforeEach(() => {
+  vi.mocked(getModels).mockResolvedValue([
+    { id: 'wan/t2v-1.3b', label: 'Wan T2V' },
+  ]);
+  vi.mocked(getDatasets).mockResolvedValue([]);
+});
+
 describe('CreateJobButton', () => {
  it('opens the workload menu on click and selects an item', async () => {
    const user = userEvent.setup();
@@ -27,10 +29,9 @@ describe('CreateJobButton', () => {
    await user.click(screen.getByRole('button', { name: 'Create Job' }));
    await user.click(screen.getByRole('menuitem', { name: /I2V/i }));

-    expect(screen.getByRole('dialog')).toHaveAttribute(
-      'data-workload-type',
-      'i2v',
-    );
+    expect(
+      screen.getByRole('dialog', { name: 'New Inference Job (I2V)' }),
+    ).toBeInTheDocument();
  });

  it('opens and operates the workload menu from the keyboard', async () => {
@@ -45,9 +46,42 @@ describe('CreateJobButton', () => {
    expect(firstItem).toHaveFocus();
    await user.keyboard('{Enter}');

-    expect(screen.getByRole('dialog')).toHaveAttribute(
-      'data-workload-type',
-      't2v',
+    expect(
+      screen.getByRole('dialog', { name: 'New Inference Job (T2V)' }),
+    ).toBeInTheDocument();
+    await user.keyboard('{Escape}');
+    await waitFor(() =>
+      expect(screen.queryByRole('dialog')).not.toBeInTheDocument(),
    );
+    await waitFor(() =>
+      expect(document.body.style.pointerEvents).not.toBe('none'),
+    );
+    expect(trigger).toHaveFocus();
  });
+
+  it.each(['inference', 'finetuning', 'distillation'] as const)(
+    'restores page interaction after closing the real %s dialog',
+    async (jobType) => {
+      const user = userEvent.setup();
+      render(<CreateJobButton jobType={jobType} />);
+      const trigger = screen.getByRole('button', { name: 'Create Job' });
+
+      // Keep the real Dialog mounted: mocking it hides conflicting Radix layers.
+      for (let attempt = 0; attempt < 2; attempt++) {
+        await user.click(trigger);
+        await user.click(screen.getAllByRole('menuitem')[0]);
+        const dialog = screen.getByRole('dialog');
+        await user.click(
+          within(dialog).getByRole('button', { name: 'Close' }),
+        );
+        await waitFor(() =>
+          expect(screen.queryByRole('dialog')).not.toBeInTheDocument(),
+        );
+        await waitFor(() =>
+          expect(document.body.style.pointerEvents).not.toBe('none'),
+        );
+        expect(trigger).toHaveFocus();
+      }
+    },
+  );
 });
@@ -2,7 +2,7 @@

 import * as React from 'react';
 import { ChevronDown } from 'lucide-react';
-import * as DropdownMenu from '@radix-ui/react-dropdown-menu';
+import { DropdownMenu } from 'radix-ui';

 import CreateJobModal from '@/components/jobs/CreateJobModal';
 import { Button } from '@/components/ui/button';
@@ -16,6 +16,7 @@ interface CreateJobButtonProps {

 export default function CreateJobButton({ jobType }: CreateJobButtonProps) {
  const options = WORKLOAD_OPTIONS[jobType] ?? [];
+  const triggerRef = React.useRef<HTMLButtonElement>(null);

  const [modalOpen, setModalOpen] = React.useState(false);
  const [workloadType, setWorkloadType] = React.useState(
@@ -36,7 +37,7 @@ export default function CreateJobButton({ jobType }: CreateJobButtonProps) {
    <>
      <DropdownMenu.Root>
        <DropdownMenu.Trigger asChild>
-          <Button type="button" className="gap-1.5">
+          <Button ref={triggerRef} type="button" className="gap-1.5">
            Create Job
            <ChevronDown className="size-3.5 opacity-85" aria-hidden />
          </Button>
@@ -66,6 +67,11 @@ export default function CreateJobButton({ jobType }: CreateJobButtonProps) {
      <CreateJobModal
        isOpen={modalOpen}
        onClose={() => setModalOpen(false)}
+        onCloseAutoFocus={(event) => {
+          // This dialog opens from a menu item, so it has no DialogTrigger.
+          event.preventDefault();
+          triggerRef.current?.focus();
+        }}
        onSuccess={handleSuccess}
        jobType={jobType}
        workloadType={workloadType}
@@ -55,6 +55,9 @@ import { jobToFormFields, type JobLike } from '@/lib/jobToFields';
 export interface CreateJobModalProps {
  isOpen: boolean;
  onClose: () => void;
+  onCloseAutoFocus?: React.ComponentProps<
+    typeof DialogContent
+  >['onCloseAutoFocus'];
  onSuccess: () => void;
  jobType: JobType;
  workloadType: string;
@@ -67,6 +70,7 @@ export interface CreateJobModalProps {
 export default function CreateJobModal({
  isOpen,
  onClose,
+  onCloseAutoFocus,
  onSuccess,
  jobType,
  workloadType,
@@ -127,8 +131,9 @@ export default function CreateJobModal({
  const editingJobId = editingJob?.id ?? null;
  const editingJobModelId = editingJob?.model_id ?? null;

-  // Layerwise offload and FSDP compete for the DiT weights and FastVideoArgs
-  // silently picks a winner (fastvideo_args.py:859); resolve it visibly here.
+  // Layerwise offload and FSDP compete for the DiT weights and the device offload
+  // policy (resolve_device_offload_conflicts in fastvideo/api/device_policy.py)
+  // silently picks a winner; resolve it visibly here.
  // dit_cpu_offload is deliberately not interlocked -- it is a modifier, not a
  // competing strategy.
  const handleDitLayerwiseOffloadChange = React.useCallback((next: boolean) => {
@@ -644,6 +649,7 @@ export default function CreateJobModal({
    >
      <DialogContent
        className="max-h-[90vh] w-[90vw] max-w-[850px] overflow-y-auto"
+        onCloseAutoFocus={onCloseAutoFocus}
        onEscapeKeyDown={(e) => {
          if (isSubmitting) e.preventDefault();
        }}
@@ -1,5 +1,7 @@
+import * as React from 'react';
 import { render, screen } from '@testing-library/react';
-import { describe, expect, it } from 'vitest';
+import userEvent from '@testing-library/user-event';
+import { describe, expect, it, vi } from 'vitest';

 import { Button } from './button';
 import { Input } from './input';
@@ -8,6 +10,24 @@ import { Slider } from './slider';
 import { Switch } from './switch';

 describe('shared control accessibility', () => {
+  it('forwards refs and click handlers to the asChild button', async () => {
+    const user = userEvent.setup();
+    const ref = React.createRef<HTMLButtonElement>();
+    const onClick = vi.fn();
+    render(
+      <Button asChild ref={ref} onClick={onClick}>
+        <button type="button">Slotted action</button>
+      </Button>,
+    );
+
+    const button = screen.getByRole('button', { name: 'Slotted action' });
+    expect(screen.getAllByRole('button')).toHaveLength(1);
+    expect(ref.current).toBe(button);
+    await user.click(button);
+    expect(onClick).toHaveBeenCalledTimes(1);
+    expect(button).toHaveFocus();
+  });
+
  it('keeps button, input, and select targets at least 44px tall', () => {
    render(
      <>
@@ -1,7 +1,7 @@
 "use client";

 import * as React from "react";
-import { Slot } from "@radix-ui/react-slot";
+import { Slot } from "radix-ui";
 import { cva, type VariantProps } from "class-variance-authority";

 import { cn } from "@/lib/utils";
@@ -37,7 +37,7 @@ export interface ButtonProps extends React.ButtonHTMLAttributes<HTMLButtonElemen
 }

 const Button = React.forwardRef<HTMLButtonElement, ButtonProps>(({ className, variant, size, asChild = false, ...props }, ref) => {
-	const Comp = asChild ? Slot : "button";
+	const Comp = asChild ? Slot.Root : "button";
 	return <Comp className={cn(buttonVariants({ variant, size, className }))} ref={ref} {...props} />;
 });
 Button.displayName = "Button";
@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as DialogPrimitive from '@radix-ui/react-dialog';
+import { Dialog as DialogPrimitive } from 'radix-ui';
 import { X } from 'lucide-react';

 import { cn } from '@/lib/utils';
@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as LabelPrimitive from '@radix-ui/react-label';
+import { Label as LabelPrimitive } from 'radix-ui';
 import { cva, type VariantProps } from 'class-variance-authority';

 import { cn } from '@/lib/utils';
@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as ScrollAreaPrimitive from '@radix-ui/react-scroll-area';
+import { ScrollArea as ScrollAreaPrimitive } from 'radix-ui';

 import { cn } from '@/lib/utils';

@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as SelectPrimitive from '@radix-ui/react-select';
+import { Select as SelectPrimitive } from 'radix-ui';
 import { Check, ChevronDown, ChevronUp } from 'lucide-react';

 import { cn } from '@/lib/utils';
@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as SeparatorPrimitive from '@radix-ui/react-separator';
+import { Separator as SeparatorPrimitive } from 'radix-ui';

 import { cn } from '@/lib/utils';

@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as SliderPrimitive from '@radix-ui/react-slider';
+import { Slider as SliderPrimitive } from 'radix-ui';

 import { cn } from '@/lib/utils';

@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as SwitchPrimitives from '@radix-ui/react-switch';
+import { Switch as SwitchPrimitives } from 'radix-ui';

 import { cn } from '@/lib/utils';

@@ -1,7 +1,7 @@
 'use client';

 import * as React from 'react';
-import * as TabsPrimitive from '@radix-ui/react-tabs';
+import { Tabs as TabsPrimitive } from 'radix-ui';

 import { cn } from '@/lib/utils';

@@ -15,6 +15,9 @@ from fastvideo import VideoGenerator as FastVideoGenerator

 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))))

+# InferenceArgs keys that are SamplingConfig fields of a GenerationRequest.
+_SAMPLING_INFERENCE_ARGS = ("height", "width", "num_frames", "num_inference_steps", "guidance_scale", "seed", "fps")
+

 # Custom exception for interruption
 class GenerationInterruptedException(Exception):
@@ -154,7 +157,17 @@ class VideoGenerator:
        """Thread function to run the generation"""
        try:
            if self.generator is not None:
-                self.generator.generate_video(prompt=prompt, output_path=output_path, **inference_args)
+                # Place each InferenceArgs value in the GenerationRequest section that owns it.
+                request: dict[str, Any] = {"prompt": prompt, "output": {"output_path": output_path}}
+                for key, value in inference_args.items():
+                    if key == "image_path":
+                        section = "inputs"
+                    elif key in _SAMPLING_INFERENCE_ARGS:
+                        section = "sampling"
+                    else:
+                        section = "extensions"
+                    request.setdefault(section, {})[key] = value
+                self.generator.generate(request)
                self._generation_result = os.path.join(output_path, f"{prompt[:100]}.mp4")
            else:
                raise RuntimeError("Generator is not initialized")
@@ -253,9 +266,24 @@ class VideoGenerator:
        if self.generator is None:
            print('generation_args', generation_args)
            print('pipeline_config', pipeline_config)
-            self.generator = FastVideoGenerator.from_pretrained(model_path=model_path,
-                                                                **generation_args,
-                                                                pipeline_config=pipeline_config)
+            # Place each generation argument at its GeneratorConfig engine path.
+            engine_config: dict[str, Any] = {}
+            if "num_gpus" in generation_args:
+                engine_config["num_gpus"] = generation_args["num_gpus"]
+            for parallelism_key in ("tp_size", "sp_size"):
+                if parallelism_key in generation_args:
+                    engine_config.setdefault("parallelism", {})[parallelism_key] = generation_args[parallelism_key]
+            if "dit_cpu_offload" in generation_args:
+                engine_config["offload"] = {"dit": generation_args["dit_cpu_offload"]}
+            self.generator = FastVideoGenerator.from_config({
+                "model_path": model_path,
+                "engine": engine_config,
+                "pipeline": {
+                    "experimental": {
+                        "pipeline_config": pipeline_config
+                    }
+                },
+            })

        print('inference_args', inference_args)

@@ -1,5 +1,5 @@
 {
-  "version": 6,
+  "version": 11,
  "recipes": [
    {
      "id": "fastwan21-t2v",
@@ -437,18 +437,21 @@
      "evidence": "Source-backed",
      "expected_artifact": "MP4 with synchronized audio at outputs/minimax_h3_t2v/minimax_h3_t2v.mp4",
      "modes": ["T2VA"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4}
+      ],
      "limitations": ["The checked-in example defaults to four-way sequence parallelism. It does not record a GPU model or memory requirement."]
    },
    {
      "id": "fasth3-preview-cuda",
      "group": "fasth3-preview",
-      "group_label": "FastH3 Preview",
+      "group_label": "FastH3 V1",
      "group_task": "4-step text to video + audio",
      "family": "minimax_h3",
      "stage": "inference",
      "task": "Few-step text to video (with audio)",
-      "label": "FastH3 Preview on CUDA",
-      "summary": "Run the DMD2-distilled FastH3 Preview with four DiT forwards, trained H3 sparse attention, compiled decode, and synchronized audio.",
+      "label": "FastH3 V1 on CUDA",
+      "summary": "Run FastH3 V1 with four DiT forwards, trained H3 sparse attention, compiled decode, and synchronized audio.",
      "model": "FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2",
      "source": "examples/inference/basic/basic_fasth3.py",
      "serving": {
@@ -467,6 +470,10 @@
      "evidence": "Verified",
      "expected_artifact": "Warmup and measured MP4 files under outputs/fasth3/",
      "modes": ["T2VA", "4-step FastH3"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4},
+        {"key": "video_decode_backend", "label": "VAE decode", "hint": "Fidelity vs. speed", "flag": "--video-decode-backend", "options": [{"value": "h3-vae", "label": "Full H3 VAE"}, {"value": "taeh3", "label": "TAEH3 preview"}], "default": "h3-vae"}
+      ],
      "limitations": ["The default all profile is the measured GB200 performance route and can change floating-point operation order. Use --profile strict --no-inference-torch-compile for the eager strict route."]
    },
    {
@@ -477,13 +484,13 @@
        "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 Preview",
+      "group_label": "FastH3 V1",
      "group_task": "4-step text to video + audio",
      "family": "minimax_h3",
      "stage": "inference",
      "task": "Few-step text to video (with audio)",
-      "label": "FastH3 Preview on MLX",
-      "summary": "Run FastH3 Preview on Apple Silicon with a locally converted INT6 DiT, streamed Qwen3-VL conditioning, and native MLX video and audio VAEs.",
+      "label": "FastH3 V1 on MLX",
+      "summary": "Run FastH3 V1 on Apple Silicon with a locally converted INT6 DiT, streamed Qwen3-VL conditioning, and native MLX video and audio VAEs.",
      "model": "FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2",
      "source": "examples/inference/basic/mlx_fasth3.py",
      "command": "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\"\npython examples/inference/basic/mlx_fasth3.py --model-root ./FastH3-Preview-v0.2 --mlx-checkpoint ./FastH3-MLX/int6 --prompt \"(S1) A presenter says <d>[English] FastVideo runs FastH3.</d>\" --height 480 --width 832 --num-frames 124 --seed 2026 --output-path ./outputs/fasth3_int6.mp4",
@@ -499,10 +506,155 @@
      "evidence": "Verified",
      "expected_artifact": "MP4 with H.264 video and stereo AAC audio at outputs/fasth3_int6.mp4",
      "modes": ["T2VA", "temporal --fast", "spatial --fast-spatial", "opt-in VSA"],
+      "knobs": [
+        {"key": "video_decode_backend", "label": "VAE decode", "hint": "Fidelity vs. speed", "flag": "--video-decode-backend", "options": [{"value": "h3-vae", "label": "Full H3 VAE"}, {"value": "taeh3", "label": "TAEH3 preview"}], "default": "h3-vae"}
+      ],
      "limitations": [
        "The MLX path supports T2VA, optional temporal --fast, optional spatial --fast-spatial, and opt-in VSA on --include-vsa checkpoints. FL2VA, Ref2VA, and two-pass refinement are not wired."
      ]
    },
+    {
+      "id": "fasth3-preview-spark",
+      "group": "fasth3-preview",
+      "group_label": "FastH3 V1",
+      "group_task": "4-step text to video + audio",
+      "family": "minimax_h3",
+      "stage": "inference",
+      "task": "Few-step text to video (with audio)",
+      "label": "FastH3 V1 on one DGX Spark",
+      "summary": "Run FastH3 V1 on one GB10 with Triton VSA, FA4 off, and lazy module load. Height, width, frames, and steps in the YAML are examples.",
+      "model": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
+      "source": "examples/inference/basic/basic_fasth3_spark.yaml",
+      "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"
+      },
+      "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"],
+      "hardware": {
+        "platform": "cuda",
+        "device": "spark",
+        "gpu_count": 1,
+        "evidence": "source-configured"
+      },
+      "evidence": "Source-backed",
+      "expected_artifact": "MP4 under outputs/fasth3_spark/",
+      "modes": ["T2VA", "1-Spark"],
+      "limitations": [
+        "Install from the DGX Spark guide, not the generic CUDA extra. GB10 has no FA4 / sm_100a VSA kernel; keep FASTVIDEO_FA4=0 and FASTVIDEO_VSA_SM100A=0.",
+        "Legal num_frames values are 17n+5, capped at 345 (15 s). Native 16:9 sizes include 832x480 and 1344x768.",
+        "Lazy module load reloads Qwen3-VL and the DiT between phases of each request. Do not set engine.offload.lazy_module_load to false on this box.",
+        "A 345-frame request on one Spark can OOM. Prefer 124 or 243 frames, TAEH3 decode, or two Sparks over QSFP."
+      ]
+    },
+    {
+      "id": "fasth3-spark-pair",
+      "group": "fasth3-preview",
+      "group_label": "FastH3 V1",
+      "group_task": "4-step text to video + audio",
+      "family": "minimax_h3",
+      "stage": "inference",
+      "task": "Few-step text to video (with audio)",
+      "label": "FastH3 V1 on two DGX Sparks",
+      "summary": "Run one FastH3 clip across two GB10s with Ray sequence parallel over QSFP RoCE. Sequential load and lazy module load stay on because SP replicates the DiT on each node.",
+      "model": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
+      "source": "examples/inference/basic/basic_fasth3_spark_pair.yaml",
+      "command": "source examples/inference/optimizations/spark_pair_env.sh && FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_VAE_PARALLEL_DECODE=1 fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair.yaml",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {
+        "platform": "cuda",
+        "device": "spark",
+        "gpu_count": 2,
+        "accelerator": "NVIDIA GB10 (DGX Spark pair)",
+        "evidence": "validated",
+        "evidence_url": "https://github.com/hao-ai-lab/FastVideo/pull/1803"
+      },
+      "evidence": "Verified",
+      "expected_artifact": "MP4 under outputs/fasth3_spark_pair/",
+      "modes": ["T2VA", "2-Spark SP"],
+      "limitations": [
+        "Requires a two-node Ray cluster on the QSFP interconnect. There is no cookbook server for this path; use Python / generate.",
+        "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": "fasth3-8step-v2-cuda",
+      "group": "fasth3-8step-v2",
+      "group_label": "FastH3 V2",
+      "group_task": "8-step text to video + audio",
+      "family": "minimax_h3",
+      "stage": "inference",
+      "task": "Few-step text to video (with audio)",
+      "label": "FastH3 V2 on CUDA",
+      "summary": "Run FastH3 V2, the eight-forward checkpoint (video/audio shifts 10/3, VSA 0.8, 64-token tiles) with the trained DMD ladder loaded from the checkpoint's fastvideo_inference.json.",
+      "model": "FastVideo/FastVideo-FastH3-8-Step-V2",
+      "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]\""
+      },
+      "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"],
+      "hardware": {
+        "platform": "cuda",
+        "gpu_count": 4,
+        "accelerator": "NVIDIA GB200",
+        "evidence": "validated",
+        "evidence_url": "https://github.com/hao-ai-lab/FastVideo/pull/1852"
+      },
+      "evidence": "Verified",
+      "expected_artifact": "Warmup and measured MP4 files under outputs/fasth3_8step/",
+      "modes": ["T2VA", "8-step FastH3"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4},
+        {"key": "video_decode_backend", "label": "VAE decode", "hint": "Fidelity vs. speed", "flag": "--video-decode-backend", "options": [{"value": "h3-vae", "label": "Full H3 VAE"}, {"value": "taeh3", "label": "TAEH3 preview"}], "default": "h3-vae"}
+      ],
+      "limitations": [
+        "Nine sigma-grid points (eight transformer forwards) are fixed by the checkpoint's trained ladder; the example rejects any other --steps.",
+        "Validated with the eager strict route (--profile strict --no-inference-torch-compile --no-compile-vae). The compiled all profile has not been measured for this checkpoint.",
+        "T2AV only; no FL2VA/Ref2VA distillation and no matching LoRA. V2 uses eight forwards rather than V1's four."
+      ]
+    },
+    {
+      "id": "fasth3-8step-v2-mlx",
+      "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"
+      },
+      "group": "fasth3-8step-v2",
+      "group_label": "FastH3 V2",
+      "group_task": "8-step text to video + audio",
+      "family": "minimax_h3",
+      "stage": "inference",
+      "task": "Few-step text to video (with audio)",
+      "label": "FastH3 V2 on MLX",
+      "summary": "Run FastH3 V2 on Apple Silicon with a locally converted INT8 DiT, the checkpoint's DMD contract ladder, and trained VSA (sparsity 0.8, 64-token tiles).",
+      "model": "FastVideo/FastVideo-FastH3-8-Step-V2",
+      "source": "examples/inference/basic/mlx_fasth3_8step.py",
+      "command": "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\npython examples/inference/basic/mlx_fasth3_8step.py --model-root ./FastH3-8-Step-V2 --mlx-checkpoint ./FastH3-8-Step-V2-MLX/int8 --prompt \"(S1) A presenter says <d>[English] FastVideo runs FastH3.</d>\" --height 480 --width 832 --num-frames 124 --seed 2026 --output-path ./outputs/fasth3_8step_int8.mp4",
+      "gpu_types": ["Apple Silicon"],
+      "hardware": {
+        "platform": "mlx",
+        "accelerator": "Apple M4 Max",
+        "system_memory": "36 GB unified memory",
+        "peak_memory": "27.39 GiB peak MLX memory during denoising",
+        "evidence": "validated",
+        "evidence_url": "https://github.com/hao-ai-lab/FastVideo/pull/1863"
+      },
+      "evidence": "Verified",
+      "expected_artifact": "MP4 with H.264 video and stereo AAC audio at outputs/fasth3_8step_int8.mp4",
+      "modes": ["T2VA", "8-step FastH3", "trained VSA"],
+      "knobs": [
+        {"key": "video_decode_backend", "label": "VAE decode", "hint": "Fidelity vs. speed", "flag": "--video-decode-backend", "options": [{"value": "h3-vae", "label": "Full H3 VAE"}, {"value": "taeh3", "label": "TAEH3 preview"}], "default": "h3-vae"}
+      ],
+      "limitations": [
+        "Convert with --include-vsa. mlx_fasth3_8step.py turns VSA on (sparsity 0.8, tile 64). A dense export fails at configure_vsa.",
+        "--steps 8 or 9 both run the eight trained forwards. Reuse the preview VAE, audio VAE, text encoder, and tokenizer if those directories already exist.",
+        "The MLX path supports T2VA only. FL2VA, Ref2VA, and two-pass refinement are not wired."
+      ]
+    },
    {
      "id": "minimax-h3-fl2va",
      "family": "minimax_h3",
@@ -518,6 +670,9 @@
      "evidence": "Source-backed",
      "expected_artifact": "MP4 with synchronized audio at outputs/minimax_h3_fl2va/minimax_h3_fl2va.mp4",
      "modes": ["FL2VA"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4}
+      ],
      "limitations": ["Pass --last-image to constrain the final frame. The checked-in source defaults to four GPUs."]
    },
    {
@@ -535,6 +690,9 @@
      "evidence": "Source-backed",
      "expected_artifact": "MP4 with synchronized audio at outputs/minimax_h3_ref2va/minimax_h3_ref2va.mp4",
      "modes": ["Ref2VA"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4}
+      ],
      "limitations": ["Pass --reference-audio for an additional audio reference. The checked-in source defaults to four GPUs."]
    },
    {
@@ -558,6 +716,10 @@
      "evidence": "Verified",
      "expected_artifact": "Warmup and measured MP4 files under outputs/fasth3_lora_preview/",
      "modes": ["T2VA", "FastH3 LoRA"],
+      "knobs": [
+        {"key": "num_gpus", "label": "GPUs", "hint": "Sequence-parallel degree", "flag": "--num-gpus", "options": [1, 2, 4, 8], "default": 4},
+        {"key": "video_decode_backend", "label": "VAE decode", "hint": "Fidelity vs. speed", "flag": "--video-decode-backend", "options": [{"value": "h3-vae", "label": "Full H3 VAE"}, {"value": "taeh3", "label": "TAEH3 preview"}], "default": "h3-vae"}
+      ],
      "limitations": ["Supply a compatible FastH3 adapter. The script infers dense or VSA attention from the adapter payload unless you override it."]
    },
    {
@@ -89,11 +89,38 @@

  const groupIdFor = (recipe) => recipe.group || recipe.id;

-  const gpuCountLabel = (hardware) => {
+  const gpuCountLabel = (hardware, runtimeId) => {
    const count = hardware?.gpu_count;
+    if (count == null) return "";
+    if (runtimeId === "spark") return `${count} Spark${count === 1 ? "" : "s"}`;
    return `${count} GPU${count === 1 ? "" : "s"}`;
  };

+  const knobsFor = (recipe) => recipe.knobs || [];
+
+  const knobOptions = (knob) =>
+    knob.options.map((option) =>
+      option !== null && typeof option === "object" ? option : { value: option, label: String(option) },
+    );
+
+  const knobDefaultLabel = (knob) => {
+    const match = knobOptions(knob).find((option) => option.value === knob.default);
+    return match ? match.label : String(knob.default);
+  };
+
+  // Knob flags are always shown explicitly in the displayed command, even at
+  // their default value, so the command stays copy-pasteable and precise
+  // about what it runs -- not just "trust the script's own default".
+  const appendKnobFlags = (commandText, knobs, knobValues) => {
+    const flags = knobs
+      .filter((knob) => knobValues[knob.key] !== undefined)
+      .map((knob) => `${knob.flag} ${knobValues[knob.key]}`);
+    if (!flags.length) return commandText;
+    const lines = commandText.split("\n");
+    lines[lines.length - 1] = `${lines[lines.length - 1]} ${flags.join(" ")}`;
+    return lines.join("\n");
+  };
+
  const runtimeFor = (recipe) => {
    const platform = recipe.hardware?.platform || "cuda";
    if (platform === "mlx") {
@@ -109,17 +136,24 @@
    if (platform === "mps") {
      return {
        id: "mps",
-        label: "Apple Silicon · MPS",
+        label: "Apple Silicon · PyTorch MPS",
        hint: recipe.hardware?.minimum_memory || recipe.hardware?.system_memory || "Memory not recorded",
      };
    }
    const hardware = recipe.hardware || {};
+    if (hardware.device === "spark") {
+      return {
+        id: "spark",
+        label: "NVIDIA DGX Spark",
+        hint: "GB10 · 128 GB unified memory",
+      };
+    }
    return {
      id: "cuda",
      label: "NVIDIA CUDA",
      hint: hardware.accelerator
-        ? `${hardware.accelerator} · ${gpuCountLabel(hardware)}`
-        : `${gpuCountLabel(hardware)} configured · GPU model not recorded`,
+        ? `${hardware.accelerator} · ${gpuCountLabel(hardware, "cuda")}`
+        : `${gpuCountLabel(hardware, "cuda")} configured · GPU model not recorded`,
    };
  };

@@ -136,8 +170,11 @@
        .filter(Boolean)
        .join(" · ");
    }
-    if (hardware.accelerator) return [hardware.accelerator, gpuCountLabel(hardware)].join(" · ");
-    return `NVIDIA CUDA · ${gpuCountLabel(hardware)} configured · GPU model and VRAM not recorded`;
+    if (runtime.id === "spark") {
+      return [hardware.accelerator || "NVIDIA GB10", gpuCountLabel(hardware, "spark")].filter(Boolean).join(" · ");
+    }
+    if (hardware.accelerator) return [hardware.accelerator, gpuCountLabel(hardware, runtime.id)].join(" · ");
+    return `NVIDIA CUDA · ${gpuCountLabel(hardware, "cuda")} configured · GPU model and VRAM not recorded`;
  };

  const renderHardwareEvidence = (container, badge, recipe) => {
@@ -157,15 +194,17 @@
    if (isValidated) {
      const recorded = [
        hardware.accelerator,
-        runtime.id === "cuda" ? gpuCountLabel(hardware) : hardware.system_memory,
+        runtime.id === "mlx" || runtime.id === "mps" ? hardware.system_memory : gpuCountLabel(hardware, runtime.id),
      ].filter(Boolean);
      const statements = [`${recorded.join(" · ")}.`];
      if (hardware.minimum_memory) statements.push(`Documented minimum: ${hardware.minimum_memory}.`);
      if (hardware.peak_memory) statements.push(`Measured: ${hardware.peak_memory}.`);
      if (!hardware.minimum_memory) statements.push("This recorded device is not a minimum requirement.");
      details.textContent = ` ${statements.join(" ")}`;
+    } else if (runtime.id === "spark") {
+      details.textContent = ` NVIDIA DGX Spark · ${gpuCountLabel(hardware, "spark")}. GB10 has no FA4 / sm_100a VSA kernel; keep Triton VSA and FA4 off.`;
    } else if (runtime.id === "cuda") {
-      details.textContent = ` NVIDIA CUDA · ${gpuCountLabel(hardware)}. The source does not record the GPU model or VRAM.`;
+      details.textContent = ` NVIDIA CUDA · ${gpuCountLabel(hardware, "cuda")}. The source does not record the GPU model or VRAM.`;
    } else {
      details.textContent = ` ${runtime.label}. The source does not record a device or memory requirement.`;
    }
@@ -223,6 +262,10 @@
    const servingPanel = root.querySelector("[data-cookbook-serving]");
    const usage = root.querySelector("[data-cookbook-usage]");
    const servingAvailability = root.querySelector("[data-cookbook-serving-availability]");
+    const knobsContainer = root.querySelector("[data-cookbook-knobs]");
+    const deviceRow = root.querySelector("[data-cookbook-device-row]");
+    const deviceOptions = root.querySelector("[data-cookbook-device-options]");
+    const deviceCaption = root.querySelector("[data-cookbook-device-caption]");

    modelOptions.setAttribute("aria-label", "Recipe");
    hardwareOptions.setAttribute("aria-label", "Runtime");
@@ -289,16 +332,96 @@
    const clientDetails = servingPanel?.querySelector(".cookbook-serving__code");
    if (clientDetails && query.has("client")) clientDetails.open = true;

+    const knobDefs = new Map();
+    familyRecipes.forEach((recipe) => knobsFor(recipe).forEach((knob) => {
+      if (!knobDefs.has(knob.key)) knobDefs.set(knob.key, knob);
+    }));
+    const knobValues = {};
+    knobDefs.forEach((knob, key) => {
+      const fromQuery = query.get(key);
+      const validValues = knobOptions(knob).map((option) => String(option.value));
+      const useQueryValue = fromQuery !== null && validValues.includes(fromQuery);
+      const raw = useQueryValue ? fromQuery : knob.default;
+      knobValues[key] = typeof knob.default === "number" ? Number(raw) : raw;
+    });
+
+    const renderKnobs = (recipe, hidden) => {
+      if (!knobsContainer) return;
+      const knobs = knobsFor(recipe);
+      const renderedKeys = [...knobsContainer.querySelectorAll("[data-knob-row]")].map((row) => row.dataset.knobRow);
+      if (renderedKeys.join(",") !== knobs.map((knob) => knob.key).join(",")) {
+        knobsContainer.replaceChildren();
+        knobs.forEach((knob) => {
+          const row = document.createElement("div");
+          row.className = "cookbook-selection-row";
+          row.dataset.knobRow = knob.key;
+          const labelWrap = document.createElement("div");
+          labelWrap.className = "cookbook-selection-row__label";
+          const strongLabel = document.createElement("strong");
+          strongLabel.textContent = knob.label;
+          const hintLabel = document.createElement("span");
+          hintLabel.textContent = knob.hint || "";
+          labelWrap.append(strongLabel, hintLabel);
+          const grid = document.createElement("div");
+          grid.className = "cookbook-option-grid cookbook-option-grid--hardware";
+          grid.setAttribute("role", "group");
+          grid.setAttribute("aria-label", knob.label);
+          knobOptions(knob).forEach((option) => {
+            const optionButton = document.createElement("button");
+            optionButton.type = "button";
+            optionButton.dataset.knobKey = knob.key;
+            optionButton.dataset.knobValue = String(option.value);
+            optionButton.setAttribute("aria-pressed", "false");
+            const optionLabel = document.createElement("strong");
+            optionLabel.textContent = option.label;
+            optionButton.append(optionLabel);
+            grid.append(optionButton);
+          });
+          row.append(labelWrap, grid);
+          knobsContainer.append(row);
+        });
+      }
+      knobsContainer.hidden = hidden || knobs.length === 0;
+      knobsContainer.querySelectorAll("button[data-knob-key]").forEach((optionButton) => {
+        const selected = String(knobValues[optionButton.dataset.knobKey]) === optionButton.dataset.knobValue;
+        optionButton.classList.toggle("cookbook-option--selected", selected);
+        optionButton.setAttribute("aria-pressed", String(selected));
+      });
+    };
+
+    const recipesForRuntime = (runtimeId, groupRecipes) =>
+      groupRecipes.filter((item) => runtimeFor(item).id === runtimeId);
+
+    const uniqueRuntimeIds = (groupRecipes) => {
+      const ids = [];
+      groupRecipes.forEach((item) => {
+        const id = runtimeFor(item).id;
+        if (!ids.includes(id)) ids.push(id);
+      });
+      return ids;
+    };
+
+    const pickRecipeForRuntime = (runtimeId, preferredCount, groupRecipes) => {
+      const siblings = recipesForRuntime(runtimeId, groupRecipes);
+      if (!siblings.length) return null;
+      if (preferredCount != null) {
+        const match = siblings.find((item) => item.hardware?.gpu_count === preferredCount);
+        if (match) return match;
+      }
+      return siblings.find((item) => item.hardware?.gpu_count === 1) || siblings[0];
+    };
+
    const renderRuntimeOptions = () => {
      const groupRecipes = groups.get(selectedGroupId) || [];
-      const renderedIds = [...hardwareOptions.querySelectorAll("[data-recipe-id]")].map((option) => option.dataset.recipeId);
-      if (renderedIds.join(",") === groupRecipes.map((recipe) => recipe.id).join(",")) return;
+      const runtimeIds = uniqueRuntimeIds(groupRecipes);
+      const renderedIds = [...hardwareOptions.querySelectorAll("[data-runtime-id]")].map((option) => option.dataset.runtimeId);
+      if (renderedIds.join(",") === runtimeIds.join(",")) return;
      hardwareOptions.replaceChildren();
-      groupRecipes.forEach((recipe) => {
-        const runtime = runtimeFor(recipe);
+      runtimeIds.forEach((runtimeId) => {
+        const representative = pickRecipeForRuntime(runtimeId, 1, groupRecipes);
+        const runtime = runtimeFor(representative);
        const option = document.createElement("button");
        option.type = "button";
-        option.dataset.recipeId = recipe.id;
        option.dataset.runtimeId = runtime.id;
        option.setAttribute("aria-pressed", "false");
        const optionLabel = document.createElement("strong");
@@ -310,6 +433,49 @@
      });
    };

+    const renderDeviceOptions = (recipe) => {
+      if (!deviceRow || !deviceOptions) return;
+      const groupRecipes = groups.get(selectedGroupId) || [];
+      const runtime = runtimeFor(recipe);
+      const siblings = recipesForRuntime(runtime.id, groupRecipes)
+        .slice()
+        .sort((left, right) => (left.hardware?.gpu_count || 0) - (right.hardware?.gpu_count || 0));
+      const show = siblings.length > 1;
+      deviceRow.hidden = !show;
+      if (deviceCaption) {
+        deviceCaption.textContent = runtime.id === "spark" ? "1 Spark or a QSFP pair" : "GPU count for this runtime";
+      }
+      if (!show) {
+        deviceOptions.replaceChildren();
+        return;
+      }
+      const renderedIds = [...deviceOptions.querySelectorAll("[data-recipe-id]")].map((option) => option.dataset.recipeId);
+      if (renderedIds.join(",") !== siblings.map((item) => item.id).join(",")) {
+        deviceOptions.replaceChildren();
+        siblings.forEach((candidate) => {
+          const option = document.createElement("button");
+          option.type = "button";
+          option.dataset.recipeId = candidate.id;
+          option.setAttribute("aria-pressed", "false");
+          const optionLabel = document.createElement("strong");
+          optionLabel.textContent = gpuCountLabel(candidate.hardware, runtime.id);
+          const optionHint = document.createElement("span");
+          optionHint.textContent = runtime.id === "spark" && candidate.hardware?.gpu_count === 2
+            ? "Ray sequence parallel over QSFP"
+            : runtime.id === "spark"
+              ? "One GB10, local process"
+              : `${gpuCountLabel(candidate.hardware, runtime.id)} configured`;
+          option.append(optionLabel, optionHint);
+          deviceOptions.append(option);
+        });
+      }
+      deviceOptions.querySelectorAll("button").forEach((option) => {
+        const selected = option.dataset.recipeId === recipe.id;
+        option.classList.toggle("cookbook-option--selected", selected);
+        option.setAttribute("aria-pressed", String(selected));
+      });
+    };
+
    let notes = root.querySelector("[data-cookbook-notes]");
    if (!notes) {
      notes = document.createElement("aside");
@@ -325,8 +491,9 @@
      let recipe = byId.get(selectedRecipeId);
      if (groupChanged || groupIdFor(recipe) !== selectedGroupId) {
        const currentRuntime = runtimeFor(recipe).id;
+        const currentCount = recipe.hardware?.gpu_count;
        const groupRecipes = groups.get(selectedGroupId) || [];
-        recipe = groupRecipes.find((candidate) => runtimeFor(candidate).id === currentRuntime) || groupRecipes[0];
+        recipe = pickRecipeForRuntime(currentRuntime, currentCount, groupRecipes) || groupRecipes[0];
        selectedRecipeId = recipe.id;
      }

@@ -335,11 +502,14 @@
        renderRuntimeOptions();
        renderedRuntimeGroup = selectedGroupId;
      }
+      renderDeviceOptions(recipe);
      const runtime = runtimeFor(recipe);
      const profile = servingPanel && servingProfiles[recipe.id];
      const useServer = Boolean(profile && usagePreference === "server");
      // 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);
+      renderKnobs(recipe, useServer);

      if (usage) {
        usage.querySelectorAll("[data-cookbook-mode]").forEach((option) => {
@@ -349,10 +519,10 @@
          option.setAttribute("aria-pressed", String(selected));
        });
        servingAvailability.textContent = profile
-          ? "The playground and API clients 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. Open the H3 server guide below, or use Python directly."
-            : "This recipe uses Python directly. For the playground and API clients, choose FastH3 Preview with CUDA or MLX.";
+            : "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;
@@ -360,11 +530,17 @@

      if (useServer) {
        const isMLX = profile.runtime === "mlx";
+        const isSpark = runtime.id === "spark";
        servingPanel.querySelector("[data-cookbook-server-lifetime]").textContent = isMLX
          ? "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."
-          : "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.";
+          : isSpark
+            ? "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/mps/#run-fasth3-preview" : "../../getting_started/installation/gpu/";
+          ? "../../getting_started/installation/mlx/"
+          : isSpark
+            ? "../../getting_started/installation/spark/"
+            : "../../getting_started/installation/gpu/";
        servingPanel.querySelector("[data-cookbook-prepare]").hidden = !profile.prepare;
        servingPanel.querySelector("[data-cookbook-server-prepare]").textContent = profile.prepare;
        servingPanel.querySelector("[data-cookbook-server-install]").textContent = profile.install;
@@ -392,18 +568,13 @@
        option.setAttribute("aria-pressed", String(selected));
      });
      hardwareOptions.querySelectorAll("button").forEach((option) => {
-        const selected = option.dataset.recipeId === recipe.id;
+        const selected = option.dataset.runtimeId === runtime.id;
        option.classList.toggle("cookbook-option--selected", selected);
        option.setAttribute("aria-pressed", String(selected));
-        const candidate = byId.get(option.dataset.recipeId);
-        const candidateProfile = servingProfiles[candidate.id];
-        const candidateRuntime = runtimeFor(candidateProfile && usagePreference === "server"
-          ? { ...candidate, hardware: candidateProfile.hardware } : candidate);
-        option.querySelector("span").textContent = candidateRuntime.hint;
      });

      description.textContent = useServer
-        ? `FastH3 Preview generates video with audio. This server profile uses the checked-in ${profile.runtime.toUpperCase()} configuration.`
+        ? `${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;
@@ -418,17 +589,23 @@
      source.href = `https://github.com/hao-ai-lab/FastVideo/blob/main/${useServer ? profile.source : recipe.source}`;
      source.textContent = useServer ? "View server configuration" : "Open example source";
      modelLink.href = `https://huggingface.co/${recipe.model}`;
-      command.textContent = recipe.command;
+      command.textContent = useServer ? recipe.command : appendKnobFlags(recipe.command, knobs, knobValues);

      renderHardwareEvidence(hardwareState, hardwareBadge, activeRecipe);

-      const limitations = useServer ? [
+      const knobCaveats = useServer ? [] : knobs
+        .filter((knob) => String(knobValues[knob.key]) !== String(knob.default))
+        .map((knob) => `${knob.label} is set away from its recorded default (${knobDefaultLabel(knob)}). ` +
+          "The script accepts this value, but it has not been benchmarked here.");
+      const limitations = [...(useServer ? [
        profile.runtime === "mlx"
          ? "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."
-          : "This server config has no recorded serving benchmark. Compilation is disabled, unlike the measured Python performance profile.",
+          : 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. 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 || [];
+      ] : recipe.limitations || []), ...knobCaveats];
      notes.replaceChildren();
      notes.hidden = limitations.length === 0;
      if (limitations.length) {
@@ -446,7 +623,16 @@
      const nextQuery = new URLSearchParams(window.location.search);
      nextQuery.set("recipe", recipe.id);
      nextQuery.set("runtime", runtime.id);
-      nextQuery.delete("gpus");
+      const deviceSiblings = recipesForRuntime(runtime.id, groups.get(selectedGroupId) || []);
+      if (deviceSiblings.length > 1 && recipe.hardware?.gpu_count != null) {
+        nextQuery.set("gpus", String(recipe.hardware.gpu_count));
+      } else {
+        nextQuery.delete("gpus");
+      }
+      knobDefs.forEach((knob, key) => {
+        if (knobs.some((activeKnob) => activeKnob.key === key)) nextQuery.set(key, String(knobValues[key]));
+        else nextQuery.delete(key);
+      });
      if (usage) {
        nextQuery.set("use", useServer ? "server" : "python");
        if (useServer && clientDetails?.open) nextQuery.set("client", selectedClient);
@@ -466,6 +652,17 @@
      render({ groupChanged: true, historyMode: "push" });
    });
    hardwareOptions.addEventListener("click", (event) => {
+      const option = event.target.closest("button[data-runtime-id]");
+      if (!option) return;
+      const groupRecipes = groups.get(selectedGroupId) || [];
+      const currentCount = byId.get(selectedRecipeId)?.hardware?.gpu_count;
+      const nextRecipe = pickRecipeForRuntime(option.dataset.runtimeId, currentCount, groupRecipes);
+      if (!nextRecipe) return;
+      selectedRecipeId = nextRecipe.id;
+      selectedGroupId = groupIdFor(nextRecipe);
+      render({ historyMode: "push" });
+    });
+    deviceOptions?.addEventListener("click", (event) => {
      const option = event.target.closest("button[data-recipe-id]");
      if (!option) return;
      selectedRecipeId = option.dataset.recipeId;
@@ -484,6 +681,14 @@
      selectedClient = option.dataset.cookbookClient;
      render({ historyMode: "push" });
    });
+    knobsContainer?.addEventListener("click", (event) => {
+      const option = event.target.closest("button[data-knob-key]");
+      if (!option) return;
+      const knob = knobDefs.get(option.dataset.knobKey);
+      knobValues[option.dataset.knobKey] = typeof knob.default === "number"
+        ? Number(option.dataset.knobValue) : option.dataset.knobValue;
+      render({ historyMode: "push" });
+    });

    render();

@@ -496,6 +701,13 @@
      usagePreference = workflow(nextQuery.get("use"));
      selectedClient = ["python", "javascript", "curl"].includes(nextQuery.get("client")) ? nextQuery.get("client") : "curl";
      if (clientDetails) clientDetails.open = nextQuery.has("client");
+      knobDefs.forEach((knob, key) => {
+        const fromQuery = nextQuery.get(key);
+        const validValues = knobOptions(knob).map((option) => String(option.value));
+        if (fromQuery !== null && validValues.includes(fromQuery)) {
+          knobValues[key] = typeof knob.default === "number" ? Number(fromQuery) : fromQuery;
+        }
+      });
      render({ historyMode: "none" });
    };
    bindFamilyPopstate();
@@ -146,8 +146,7 @@ img {
 }

 .cookbook-section,
-.cookbook-builder,
-.cookbook-roadmap {
+.cookbook-builder {
    scroll-margin-top: 4.5rem;
 }

@@ -180,8 +179,7 @@ img {
 }

 .md-typeset .cookbook-section__heading h2,
-.md-typeset .cookbook-builder__intro h2,
-.md-typeset .cookbook-roadmap h2 {
+.md-typeset .cookbook-builder__intro h2 {
    margin: 0;
    color: var(--md-default-fg-color);
    font-size: clamp(1.4rem, 2.6vw, 1.85rem);
@@ -493,8 +491,7 @@ img {
    max-width: 45rem;
 }

-.cookbook-builder__intro > p,
-.cookbook-roadmap > p {
+.cookbook-builder__intro > p {
    color: var(--md-default-fg-color--light);
    font-size: 0.75rem;
    line-height: 1.6;
@@ -567,7 +564,7 @@ img {
 }

 .cookbook-option-grid--hardware {
-    grid-template-columns: repeat(2, minmax(0, 1fr));
+    grid-template-columns: repeat(auto-fit, minmax(6.4rem, 1fr));
 }

 .cookbook-option-grid button {
@@ -769,12 +766,6 @@ img {
    font-weight: 700;
 }

-.cookbook-roadmap {
-    padding: 2.5rem 0;
-    margin: 3.5rem 0 1rem;
-    border-top: 1px solid var(--cookbook-border);
-}
-
 /* Lifecycle roadmap on family pages: inference is live; later stages are
   explicitly labelled planned and must not look runnable. */
 .cookbook-lifecycle {
@@ -1058,6 +1049,55 @@ img {
    overflow-wrap: anywhere;
 }

+.cookbook-jumpnav {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 1.1rem;
+    max-width: 48rem;
+    margin: 0 auto 1.75rem;
+    padding: 0.55rem 0;
+    border-top: 1px solid var(--cookbook-border);
+    border-bottom: 1px solid var(--cookbook-border);
+}
+
+.md-typeset .cookbook-jumpnav a {
+    font-family: var(--md-code-font-family);
+    font-size: 0.7rem;
+    letter-spacing: 0.02em;
+    color: var(--md-default-fg-color--light);
+    text-decoration: none;
+}
+
+.md-typeset .cookbook-jumpnav a:hover {
+    color: var(--cookbook-accent-strong);
+}
+
+.cookbook-family-page details.cookbook-collapsible {
+    max-width: 48rem;
+    margin: 0 auto 0.75rem;
+    border-color: var(--cookbook-border);
+    box-shadow: none;
+}
+
+.md-typeset details.cookbook-collapsible > summary {
+    background: var(--cookbook-surface-raised);
+    font-size: 0.85rem;
+}
+
+.md-typeset .cookbook-collapsible__body {
+    font-size: 0.8rem;
+    line-height: 1.65;
+    color: var(--md-default-fg-color--light);
+}
+
+.md-typeset .cookbook-collapsible__body .cookbook-eyebrow {
+    margin-top: 0.9rem;
+}
+
+.md-typeset .cookbook-collapsible__body pre {
+    margin: 0.5rem 0 1rem;
+}
+
 .cookbook-family-page {
    max-width: 62rem;
 }
@@ -1184,6 +1224,10 @@ img {
    margin-top: 0.55rem;
 }

+[data-cookbook-knobs]:not(:empty) {
+    margin-top: 0.55rem;
+}
+
 .cookbook-selection-row__label {
    padding-inline-start: 0.15rem;
 }
@@ -1207,7 +1251,7 @@ img {
 }

 .cookbook-option-grid--hardware {
-    grid-template-columns: repeat(2, minmax(0, 1fr));
+    grid-template-columns: repeat(auto-fit, minmax(6.4rem, 1fr));
 }

 .cookbook-option-grid button {
@@ -18,8 +18,8 @@ the CUDA `fastvideo-kernel` package:
 - **Dense-only checkpoints** (the default converter) drop the 50 gate
  matrices and keep fused SDPA. They remain valid for dense inference.
 - **VSA-capable checkpoints** retain those gates, quantize them on the same
-  affine grid, and record `vsa.capable` in `mlx_h3_dit.json`. Runtime VSA is
-  still off until you pass `--vsa`.
+  affine grid, and record `vsa.capable` in `mlx_h3_dit.json`. Preview leaves
+  runtime VSA off until you pass `--vsa`. `mlx_fasth3_8step.py` turns it on.
 - **Tile sizes** 64 `(4, 4, 4)` and 256 `(4, 8, 8)`. Prefix keys can be
  `exempt` or `compete`. `--vsa-dense-first-n-steps` and `--vsa-dense-layers`
  force dense SDPA on the selected steps or blocks.
@@ -32,9 +32,12 @@ the CUDA `fastvideo-kernel` package:
  but does not yet match reference video. `--vsa-impl reference` is the same
  as `auto`.

-See the [Apple Silicon guide](../../getting_started/installation/mps.md) for
-conversion and `mlx_fasth3.py` flags. Do not enable VSA on a dense-only
-checkpoint; reconvert with `--include-vsa` first.
+See the [MLX install guide](../../getting_started/installation/mlx.md)
+and the [MiniMax H3 cookbook](../../cookbook/minimax-h3.md) for conversion
+and `mlx_fasth3.py` / `mlx_fasth3_8step.py` flags. Do not enable
+VSA on a dense-only checkpoint; reconvert with `--include-vsa` first. V1
+VSA is opt-in. FastH3 V2 converts with `--include-vsa` and turns VSA on
+by default.

 H3 uses fused MLX RMSNorm by default, including dense inference. This can
 change BF16 rounding relative to the older explicit normalization path.
@@ -30,7 +30,7 @@ FastVideo and the reference model first produce different numbers?"
 | General logging | `init_logger(__name__)` |
 | Per-stage timing | `FASTVIDEO_STAGE_LOGGING` |
 | Profiling kernel timings | `FASTVIDEO_TORCH_PROFILER_DIR` (see [Profiling](profiling.md)) |
-| Function-call tracing | `FASTVIDEO_TRACE_FUNCTION` (heavy) |
+| Function-call tracing | `fastvideo.logger.enable_trace_function_call()` (heavy) |

 ## Quickstart

@@ -141,6 +141,24 @@ it), while a GitHub outage or a >25 min wait lets it run anyway (fail open).
 The complete static graph remains available through `/test full`; path
 selection never deletes or dynamically invents a Buildkite step.

+Integration steps depend on `golden-gate`. A failed selected golden prevents
+the expensive downstream jobs from starting; `/test full` still selects all
+twenty lanes. A condition-skipped golden satisfies the dependency, so direct
+lane reruns, scheduled SSIM, and training-only merge plans keep their existing
+meaning. This follows Buildkite's
+[conditional dependency rules](https://buildkite.com/docs/pipelines/configure/depends-on).
+No dependency allows failures. The trusted uploader accepts either the old
+complete graph or the complete golden-first graph during rollout, and rejects
+partial or arbitrary dependency changes.
+
+The six automatic Fastcheck lanes are unchanged. Within VAE and transformer
+lanes, small Wan goldens run before independent component parity. Wan paths
+select matching VAE, dense/trajectory, or causal-cache goldens; shared Wan
+config and pipeline wiring select all four. Shared runtime changes continue
+to select broader coverage. The existing block references retain their exact
+environment identity; new tensor gates distinguish the effective FA2/FA4
+switch and VAE gates do not depend on an unused attention backend.
+
 | Lane | Public `TEST_TYPE` | GPUs | Typical merge trigger |
 |---|---|---:|---|
 | Encoder | `encoder` | 1 | Universal Fastcheck |
@@ -43,6 +43,8 @@ FastVideo maps a Diffusers-style repo into a pipeline like:
 - `fastvideo/models/*`: model implementations (DiT, VAE, encoders, upsamplers).
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
  weight name translation.
+- `fastvideo/models/wan/`: Wan's dense transformer, VAE, and component configs
+  live together. The old Wan modules remain compatibility re-exports.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
 - `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipeline logic built from stages.
@@ -55,19 +57,15 @@ Minimal usage example (based on `examples/inference/basic/basic.py`):

 ```python
 from fastvideo import VideoGenerator
-from fastvideo.api.sampling_param import SamplingParam

 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
-generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
+generator = VideoGenerator.from_pretrained(model_id, {"engine": {"num_gpus": 1}})

-sampling = SamplingParam.from_pretrained(model_id)
-sampling.num_frames = 45
-video = generator.generate_video(
-    "A vibrant city street at sunset.",
-    sampling_param=sampling,
-    output_path="video_samples",
-    save_video=True,
-)
+video = generator.generate({
+    "prompt": "A vibrant city street at sunset.",
+    "sampling": {"num_frames": 45},
+    "output": {"output_path": "video_samples", "save_video": True},
+})
 ```

 ## Some questions to ask yourself before starting
@@ -209,7 +207,7 @@ class OfficialWanTransformer(torch.nn.Module):
    def forward(self, x):
        return self.patch_embedding(x)

-# FastVideo model (simplified) in fastvideo/models/dits/wanvideo.py
+# FastVideo model (simplified) in fastvideo/models/wan/transformer.py
 class PatchEmbed(torch.nn.Module):
    def __init__(self):
        super().__init__()
@@ -227,7 +225,7 @@ class WanTransformer3DModel(torch.nn.Module):
        return self.patch_embedding(x)

 # Mapping defined in a config (simplified; see the real mapping in
-# fastvideo/configs/models/dits/wanvideo.py)
+# fastvideo/models/wan/config.py)
 param_names_mapping = {
    r"^patch_embedding\.(.*)$": r"patch_embedding.proj.\1",
    r"^blocks\.(\d+)\.attn1\.to_q\.(.*)$": r"blocks.\1.to_q.\2",
@@ -275,7 +273,7 @@ Mapping steps:
  - Instantiate the FastVideo DiT (`WanTransformer3DModel`) and compare
    its `state_dict().keys()` to the official keys.
  - Update `param_names_mapping` in
-    fastvideo/configs/models/dits/wanvideo.py to resolve missing/unexpected keys.
+    fastvideo/models/wan/config.py to resolve missing/unexpected keys.
  - Use `load_state_dict(strict=False)` during iteration to surface mismatches.
 ```

@@ -466,19 +464,24 @@ The Wan2.1 T2V 1.3B Diffusers pipeline is a good “standard” example for
 FastVideo integration.

 1. Verify model config + mapping.
-   - DiT mapping: `fastvideo/configs/models/dits/wanvideo.py`
-   - VAE: `fastvideo/models/vaes/wanvae.py`
+   - DiT: `fastvideo/models/wan/transformer.py`
+   - DiT mapping: `fastvideo/models/wan/config.py`
+   - VAE: `fastvideo/models/wan/vae.py`
+   - VAE config: `fastvideo/models/wan/vae_config.py`
   - Text encoder: `fastvideo/models/encoders/t5.py`

 2. Parity test the core components.
+   - Start with `bash scripts/validate_wan.sh all`: contracts and tiny goldens.
   - Example tests: `fastvideo/tests/transformers/test_wanvideo.py`,
     `fastvideo/tests/vaes/test_wan_vae.py`,
     `fastvideo/tests/encoders/test_t5_encoder.py`

 3. Pipeline wiring.
   - Pipeline: `fastvideo/pipelines/basic/wan/wan_pipeline.py`
-   - Pipeline config: `fastvideo/configs/pipelines/wan.py`
-   - Sampling defaults: `fastvideo/pipelines/basic/wan/profiles.py`
+   - Denoising and first-frame preparation: `fastvideo/pipelines/basic/wan/stages/`
+   - Variant definitions: `fastvideo/models/wan/definition.py`
+   - Pipeline config: `fastvideo/models/wan/pipeline_config.py`
+   - Sampling defaults: `fastvideo/pipelines/basic/wan/presets.py`

 4. Minimal example.
   - Script: `examples/inference/basic/basic.py`
@@ -0,0 +1,294 @@
+# Environment Variables
+
+FastVideo reads environment variables for expert switches, debugging, profiling, and the settings that launchers such
+as `torchrun` provide. This page is the policy for those variables. The contract test
+`fastvideo/tests/contract/test_env_policy.py` enforces the policy in the unit CI lane, and the coding-agent skill
+`.agents/skills/env-var-conventions/SKILL.md` points here. When the policy changes, update this page and the contract
+test in the same pull request.
+
+## Rules
+
+1. **Register every FastVideo variable in `fastvideo/envs.py`.** Each entry declares a type, a default, a category,
+   and a description. Variables that other tools own (CUDA, NCCL, PyTorch, launchers) are not registered; code reads
+   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.
+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`.
+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.
+5. **Give each setting one source: an argument or an environment variable.** Settings that users change per
+   deployment are arguments (CLI or YAML). Expert switches, emergency off switches, and debugging and test switches
+   are environment variables.
+6. **Read variables inside functions.** `envs.NAME.get()` runs when the function runs, so a changed value takes
+   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.
+
+## Field types
+
+| Class        | Value type       | Parsing rule                                                                        |
+| ------------ | ---------------- | ----------------------------------------------------------------------------------- |
+| `EnvBool`    | `bool`           | `1`, `true`, `yes`, `on` are true; `0`, `false`, `no`, `off`, and `""` are false.   |
+|              |                  | Case-insensitive; surrounding whitespace is ignored.                                |
+| `EnvInt`     | `int`            | `int(value)`                                                                        |
+| `EnvFloat`   | `float`          | `float(value)`                                                                      |
+| `EnvStr`     | `str` or `None`  | The raw string. A `None` default means that the variable has no default.            |
+| `EnvPath`    | `str` or `None`  | The raw string with a leading `~` expanded.                                         |
+| `EnvChoice`  | `str`            | Stripped and lower-cased, then checked against the declared `choices`.              |
+
+A default can be a zero-argument function; `get()` calls it on each read while the variable is unset. The path roots
+use this to follow `XDG_CONFIG_HOME` and `XDG_CACHE_HOME`.
+
+Using a field without a method, as in `if envs.FASTVIDEO_FA4:`, raises `TypeError`.
+
+## Add a variable
+
+1. Declare the variable in the matching section of `fastvideo/envs.py`:
+
+    ```python
+    FASTVIDEO_DEBUG_MY_STAGE = EnvBool(False, category="debug", doc="Log the inputs of MyStage.")
+    ```
+
+    The category is one of the values in `envs.CATEGORIES`.
+
+2. Read the variable inside a function:
+
+    ```python
+    import fastvideo.envs as envs
+
+    def forward(self, batch):
+        if envs.FASTVIDEO_DEBUG_MY_STAGE.get():
+            logger.info("MyStage inputs: %s", batch.keys())
+    ```
+
+3. Regenerate the table at the end of this page:
+
+    ```bash
+    python fastvideo/tests/contract/test_env_policy.py
+    ```
+
+4. Run the contract test:
+
+    ```bash
+    pytest fastvideo/tests/contract/test_env_policy.py
+    ```
+
+In a test, change the value with `override`, which restores the previous value on exit:
+
+```python
+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`:
+
+```python
+FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS = EnvBool(True,
+                                              category="sampling",
+                                              doc="...",
+                                              deprecated_names=("LTX2_USE_DISTILLED_SIGMAS", ))
+```
+
+`get()` reads an old name only when the new name is unset, and logs a warning once. Update the uses of the old name
+in `examples/`, `scripts/`, `docs/`, `apps/`, and the tests in the same pull request. Delete the old name in the next
+minor release.
+
+To remove a variable that no code reads, delete its entry and add the name to `DEPRECATED_VARIABLES` in
+`fastvideo/envs.py` with a reason. The config resolution step `warn_deprecated_environment_variables` in
+`fastvideo/api/inference_resolution.py` calls `envs.warn_deprecated_variables()`, which logs a warning for each listed
+variable that is set. Delete the entry in the next minor release.
+
+## What the contract test checks
+
+The test parses every Python file under `fastvideo/`, including `fastvideo/tests/`, with Python's `ast` module. It
+skips `fastvideo/third_party/`, which is copied from upstream projects, and the registry `fastvideo/envs.py`. It does
+not check `apps/`, `examples/`, `scripts/`, `fastvideo-kernel/`, or `docs/`.
+
+It reports each violation as `<path>: <kind> <name>`:
+
+| Kind               | Code that triggers it                                        | Fix                                          |
+| ------------------ | ------------------------------------------------------------ | -------------------------------------------- |
+| `read`             | `os.getenv`, `os.environ.get`, `os.environ[...]`, or         | Register the variable and call               |
+|                    | `"NAME" in os.environ` with a name outside the allowlist, or | `envs.NAME.get()`. For a variable that       |
+|                    | 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 |
+|                    |                                                              | `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.                                       |
+| `bare-field`       | A registry field used without calling one of its methods,    | Call `envs.NAME.get()`.                      |
+|                    | as in `envs.NAME == "auto"` or `getter = envs.NAME.get`      |                                              |
+| `import-time-read` | `envs.NAME.get()` outside a function                         | Move the read into the function that uses    |
+|                    |                                                              | the value.                                   |
+| `prefix`           | A registry entry without the `FASTVIDEO_` prefix             | Rename the variable and keep the old name in |
+|                    |                                                              | `deprecated_names`.                          |
+| `unread`           | A registry entry that no code reads with `get()` or          | Delete the variable and add it to            |
+|                    | `is_set()`                                                   | `DEPRECATED_VARIABLES`.                      |
+
+The test recognizes `os` imported under another name, `from os import environ, getenv`, and a name held in a
+module-level string constant. Code that reaches the environment through `importlib` or `getattr(os, "environ")` is
+left to code review.
+
+The test also checks that every registry entry has a category from `envs.CATEGORIES` and a description, and that the
+table at the end of this page matches the registry.
+
+**Known violations.** `KNOWN_VIOLATIONS` in the contract test lists the violations that existed when the policy was
+introduced. The list only shrinks. A violation that is not in the list fails the test. A listed violation that no
+longer exists also fails the test, so the fixing pull request deletes its entry.
+
+## Registered variables
+
+<!-- BEGIN GENERATED ENV TABLE: python fastvideo/tests/contract/test_env_policy.py -->
+| Variable                                                   | Type                           | Default                                               | Category    | Description                                                                                                                                                                                                                                              |
+| ---------------------------------------------------------- | ------------------------------ | ----------------------------------------------------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `FASTVIDEO_CONFIG_ROOT`                                    | path                           | computed                                              | path        | Root directory for FastVideo configuration files, at runtime and at installation. Defaults to ~/.config/fastvideo, or $XDG_CONFIG_HOME/fastvideo when XDG_CONFIG_HOME is set.                                                                            |
+| `FASTVIDEO_CACHE_ROOT`                                     | path                           | computed                                              | path        | Root directory for FastVideo cache files. Defaults to ~/.cache/fastvideo, or $XDG_CACHE_HOME/fastvideo when XDG_CACHE_HOME is set.                                                                                                                       |
+| `FASTVIDEO_REASON1_WEIGHTS_PATH`                           | str                            | unset                                                 | path        | Local path or Hugging Face id of Reason1 weights to load instead of the checkpoint's own.                                                                                                                                                                |
+| `FASTVIDEO_HOST_IP`                                        | str                            | `""`                                                  | distributed | IP address of this node when the node has several network interfaces. Set it on each node for multi-node inference.                                                                                                                                      |
+| `FASTVIDEO_LOOPBACK_IP`                                    | str                            | `""`                                                  | distributed | Loopback IP address to use instead of the detected one.                                                                                                                                                                                                  |
+| `FASTVIDEO_RAY_PER_WORKER_GPUS`                            | float                          | `1.0`                                                 | distributed | GPUs per Ray worker. A fraction lets Ray schedule several actors on one GPU, so other actors can share the GPUs with FastVideo.                                                                                                                          |
+| `FASTVIDEO_NCCL_SO_PATH`                                   | str                            | unset                                                 | distributed | Path to the NCCL library file. Needed because the nccl&gt;=2.19 that PyTorch ships has a bug (https://github.com/NVIDIA/nccl/issues/1234).                                                                                                               |
+| `FASTVIDEO_HCCL_SO_PATH`                                   | str                            | unset                                                 | distributed | Path to the HCCL library file on Ascend NPUs. Deprecated names: `HCCL_SO_PATH`.                                                                                                                                                                          |
+| `FASTVIDEO_WORKER_MULTIPROC_METHOD`                        | one of spawn, fork, forkserver | `spawn`                                               | distributed | Multiprocessing start method for worker processes.                                                                                                                                                                                                       |
+| `FASTVIDEO_ULYSSES_A2A`                                    | one of off, auto               | `off`                                                 | distributed | Sequence-parallel all-to-all backend. off uses the NCCL path in DistributedAutograd.AllToAll4D. auto uses the fused NVLink kernel when the group is a load-store accessible mesh of 2, 4, 6, or 8 ranks in eager execution, and the NCCL path otherwise. |
+| `FASTVIDEO_CONFIGURE_LOGGING`                              | bool                           | `1`                                                   | logging     | Configure logging at import. When true, FastVideo uses its default logging configuration or the file in FASTVIDEO_LOGGING_CONFIG_PATH.                                                                                                                   |
+| `FASTVIDEO_LOGGING_CONFIG_PATH`                            | str                            | unset                                                 | logging     | Path to a JSON logging configuration file.                                                                                                                                                                                                               |
+| `FASTVIDEO_LOGGING_LEVEL`                                  | str                            | `INFO`                                                | logging     | Default logging level.                                                                                                                                                                                                                                   |
+| `FASTVIDEO_LOGGING_PREFIX`                                 | str                            | `""`                                                  | logging     | Prefix prepended to every log message.                                                                                                                                                                                                                   |
+| `FASTVIDEO_STAGE_LOGGING`                                  | bool                           | `0`                                                   | logging     | Log the time that each pipeline stage takes.                                                                                                                                                                                                             |
+| `FASTVIDEO_ATTENTION_BACKEND`                              | str                            | unset                                                 | attention   | Attention backend, as an AttentionBackendEnum name such as TORCH_SDPA, FLASH_ATTN, VIDEO_SPARSE_ATTN, SAGE_ATTN, or SAGE_ATTN_THREE. Config resolution uses it when engine.attention.backend is unset. An unsupported name raises an error.              |
+| `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_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.                                                                                                                                                                                    |
+| `FASTVIDEO_MLX_WINDOW_SINK`                                | int                            | `0`                                                   | attention   | Number of sink tokens that MLX windowed attention always attends to.                                                                                                                                                                                     |
+| `FASTVIDEO_INFERENCE_TORCH_COMPILE`                        | bool                           | `0`                                                   | performance | Compile each DiT transformer block with fullgraph torch.compile at inference. Same as engine.compile.regional=True.                                                                                                                                      |
+| `FASTVIDEO_VAE_PARALLEL_DECODE`                            | bool                           | `0`                                                   | performance | MiniMax-H3 VAE decode splits its temporal chunks across the sequence-parallel ranks instead of running serially on the output rank. Same as pipeline.model.minimax_h3.vae_parallel_decode=True.                                                          |
+| `FASTVIDEO_VAE_PARALLEL_ENCODE`                            | bool                           | `0`                                                   | performance | MiniMax-H3 reference-video VAE encode splits its temporal chunks across the sequence-parallel ranks. Same as pipeline.model.minimax_h3.vae_parallel_encode=True.                                                                                         |
+| `FASTVIDEO_VAE_PARALLEL_DECODE_STRATEGY`                   | str                            | unset                                                 | performance | Collective that moves chunks in parallel VAE decode: gather (used when unset) or all_gather.                                                                                                                                                             |
+| `FASTVIDEO_MINIMAX_H3_FUSIONS`                             | str                            | `""`                                                  | performance | MiniMax-H3 inference-only Triton fusions: all, 1, or a comma-separated subset of modulate,qknorm_rope,swiglu. Empty, 0, or none keeps the eager implementation.                                                                                          |
+| `FASTVIDEO_FSDP2_AUTOWRAP`                                 | bool                           | `0`                                                   | performance | FSDP2 shards modules by parameter count instead of the model's shard conditions. Not supported by self-forcing distillation.                                                                                                                             |
+| `FASTVIDEO_FSDP2_MIN_PARAMS`                               | int                            | `10000000`                                            | performance | Minimum parameter count of a module that FASTVIDEO_FSDP2_AUTOWRAP shards.                                                                                                                                                                                |
+| `FASTVIDEO_MLX_COMPILE`                                    | bool                           | `0`                                                   | performance | Compile the MLX DiT forward with mx.compile.                                                                                                                                                                                                             |
+| `FASTVIDEO_MLX_FAST_NORM`                                  | bool                           | `0`                                                   | performance | Use MLX fast normalization kernels.                                                                                                                                                                                                                      |
+| `FASTVIDEO_MLX_DQ_GEMM`                                    | str                            | `1`                                                   | performance | MLX dequantized GEMM for affine-quantized weights: 0 turns it off, 1 uses the measured minimum row count, and an integer sets the minimum row count.                                                                                                     |
+| `FASTVIDEO_LTX2_VAE_CHANNELS_LAST_3D`                      | bool                           | `1`                                                   | performance | LTX-2 VAE uses the channels_last_3d memory format.                                                                                                                                                                                                       |
+| `FASTVIDEO_LTX2_DISABLE_AUDIO_AUTOCAST`                    | bool                           | `1`                                                   | performance | LTX-2 audio decoding runs without CUDA autocast. Deprecated names: `LTX2_DISABLE_AUDIO_AUTOCAST`.                                                                                                                                                        |
+| `FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION` | bool                           | `0`                                                   | performance | Flux denoising disables reduced-precision reductions in bf16 matmuls, which tightens accumulation for the 4-step Klein model.                                                                                                                            |
+| `FASTVIDEO_FFMPEG_BIN`                                     | str                            | `ffmpeg`                                              | output      | ffmpeg executable used to save video with audio.                                                                                                                                                                                                         |
+| `FASTVIDEO_VIDEO_CODEC`                                    | str                            | `libx264`                                             | output      | ffmpeg video codec for saved videos.                                                                                                                                                                                                                     |
+| `FASTVIDEO_NVENC_PRESET`                                   | str                            | `p1`                                                  | output      | NVENC preset when the codec is an \*_nvenc codec.                                                                                                                                                                                                        |
+| `FASTVIDEO_NVENC_TUNE`                                     | str                            | `ull`                                                 | output      | NVENC tune option.                                                                                                                                                                                                                                       |
+| `FASTVIDEO_NVENC_RC`                                       | str                            | `constqp`                                             | output      | NVENC rate-control mode.                                                                                                                                                                                                                                 |
+| `FASTVIDEO_NVENC_QP`                                       | str                            | `28`                                                  | output      | NVENC quantization parameter.                                                                                                                                                                                                                            |
+| `FASTVIDEO_NVENC_BF`                                       | str                            | `0`                                                   | output      | NVENC number of B-frames.                                                                                                                                                                                                                                |
+| `FASTVIDEO_X264_PRESET`                                    | str                            | `ultrafast`                                           | output      | x264 preset for non-NVENC codecs.                                                                                                                                                                                                                        |
+| `FASTVIDEO_OUTPUT_PIX_FMT`                                 | str                            | `yuv420p`                                             | output      | ffmpeg pixel format for saved videos.                                                                                                                                                                                                                    |
+| `FASTVIDEO_NVTX_PROFILE`                                   | bool                           | `0`                                                   | profiling   | Emit NVTX ranges for external profilers such as Nsight Systems.                                                                                                                                                                                          |
+| `FASTVIDEO_TORCH_PROFILER_DIR`                             | path                           | unset                                                 | profiling   | Enables the torch profiler and sets the directory for its traces. Must be an absolute path.                                                                                                                                                              |
+| `FASTVIDEO_TORCH_PROFILER_RECORD_SHAPES`                   | bool                           | `0`                                                   | profiling   | Torch profiler records shapes.                                                                                                                                                                                                                           |
+| `FASTVIDEO_TORCH_PROFILER_WITH_PROFILE_MEMORY`             | bool                           | `0`                                                   | profiling   | Torch profiler profiles memory.                                                                                                                                                                                                                          |
+| `FASTVIDEO_TORCH_PROFILER_WITH_STACK`                      | bool                           | `0`                                                   | profiling   | Torch profiler captures stacks. Costs about 1.5x runtime and 1.4x trace size.                                                                                                                                                                            |
+| `FASTVIDEO_TORCH_PROFILER_WITH_FLOPS`                      | bool                           | `0`                                                   | profiling   | Torch profiler profiles FLOPs.                                                                                                                                                                                                                           |
+| `FASTVIDEO_TORCH_PROFILE_REGIONS`                          | str                            | `""`                                                  | profiling   | Comma-separated profiler regions to record. The torch profiler requires at least one region.                                                                                                                                                             |
+| `FASTVIDEO_TRACE_ACTIVATIONS`                              | bool                           | `0`                                                   | debug       | Enable activation trace hooks.                                                                                                                                                                                                                           |
+| `FASTVIDEO_TRACE_LAYERS`                                   | str                            | `""`                                                  | debug       | Regex filter for traced module names. Empty means all.                                                                                                                                                                                                   |
+| `FASTVIDEO_TRACE_STATS`                                    | str                            | `abs_mean,sum`                                        | debug       | Comma-separated activation statistics dumped for each output tensor.                                                                                                                                                                                     |
+| `FASTVIDEO_TRACE_OUTPUT`                                   | str                            | `/tmp/fv_trace_<pid>.jsonl`                           | debug       | JSONL path for activation traces. The literal &lt;pid&gt; is replaced at runtime.                                                                                                                                                                        |
+| `FASTVIDEO_TRACE_STEPS`                                    | str                            | `""`                                                  | debug       | Comma-separated denoising step indices. Empty means all.                                                                                                                                                                                                 |
+| `FASTVIDEO_H3_VSA_PROBE`                                   | str                            | unset                                                 | debug       | Output directory for the VSA-H3 attention-mass probe, which writes one .pt file per step, layer, and rank. Keeps the model out of regional compile.                                                                                                      |
+| `FASTVIDEO_LTX2_GEMMA_LOG`                                 | str                            | `""`                                                  | debug       | Log file for LTX-2 Gemma text-encoder hidden states, used by parity tests. Deprecated names: `LTX2_FASTVIDEO_GEMMA_LOG`.                                                                                                                                 |
+| `FASTVIDEO_COSMOS25_LOG_KNOBS`                             | bool                           | `0`                                                   | debug       | Log the Cosmos 2.5 latent-preparation conditioning inputs.                                                                                                                                                                                               |
+| `FASTVIDEO_CFG_GATE_STEP`                                  | float                          | `1.0`                                                 | sampling    | CFG gating fraction in [0, 1]. Steps before len(timesteps) \* X run the conditional and unconditional forwards; later steps reuse the cached difference. 1.0 disables gating.                                                                            |
+| `FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS`                      | bool                           | `1`                                                   | sampling    | LTX-2 uses the distilled sigma schedule when pipeline.model.ltx2.use_distilled_sigmas is also true. Deprecated names: `LTX2_USE_DISTILLED_SIGMAS`.                                                                                                       |
+| `FASTVIDEO_EVAL_CACHE`                                     | path                           | computed                                              | eval        | Cache directory for evaluation models and datasets. Defaults to $FASTVIDEO_CACHE_ROOT/eval.                                                                                                                                                              |
+| `FASTVIDEO_PHYSICS_IQ_BUCKET_URL`                          | str                            | `https://storage.googleapis.com/physics-iq-benchmark` | eval        | Base URL of the Physics-IQ benchmark bucket.                                                                                                                                                                                                             |
+| `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_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`.                                                                                                                                                   |
+| `FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR`                   | str                            | `data/ltx2_overfit_preprocessed`                      | test        | Output directory for preprocess_ltx2_overfit.py. Deprecated names: `LTX2_OVERFIT_OUTPUT_DIR`.                                                                                                                                                            |
+| `FASTVIDEO_TEST_LTX2_OVERFIT_MODEL`                        | str                            | `FastVideo/LTX2-Distilled-Diffusers`                  | test        | Model repository whose encoders preprocess_ltx2_overfit.py uses. Deprecated names: `LTX2_OVERFIT_MODEL`.                                                                                                                                                 |
+| `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_DREAMX_WORLD_SSIM_MODEL_PATH`              | str                            | `FastVideo/DreamX-World-5B-Cam-Diffusers`             | test        | Model for the DreamX-World camera SSIM test. Deprecated names: `DREAMX_WORLD_SSIM_MODEL_PATH`.                                                                                                                                                           |
+| `FASTVIDEO_TEST_DREAMX_WORLD_AR_SSIM_MODEL_PATH`           | str                            | `FastVideo/DreamX-World-5B-Diffusers`                 | test        | Model for the DreamX-World autoregressive SSIM test. Deprecated names: `DREAMX_WORLD_AR_SSIM_MODEL_PATH`.                                                                                                                                                |
+| `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_GAMECRAFT_MODEL_PATH`                      | str                            | `FastVideo/HunyuanGameCraft-Diffusers`                | test        | Model for the HunyuanGameCraft SSIM test. Deprecated names: `GAMECRAFT_MODEL_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_GLM_IMAGE_LOCAL_WEIGHTS_DIR`               | str                            | unset                                                 | test        | Local official GLM-Image weights for the GLM-Image SSIM test. Deprecated names: `GLM_IMAGE_LOCAL_WEIGHTS_DIR`.                                                                                                                                           |
+| `FASTVIDEO_TEST_GLM_IMAGE_MODEL_DIR`                       | str                            | unset                                                 | test        | Model for the GLM-Image SSIM test. Deprecated names: `GLM_IMAGE_MODEL_DIR`.                                                                                                                                                                              |
+| `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_LONGCAT_MODEL_ROOT`                        | str                            | unset                                                 | test        | Local LongCat-Video checkpoint for the golden-gate test. Deprecated names: `LONGCAT_MODEL_ROOT`.                                                                                                                                                         |
+| `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`.                                                                                                                                                          |
+| `FASTVIDEO_TEST_ZIMAGE_MODEL_DIR`                          | str                            | `Tongyi-MAI/Z-Image-Turbo`                            | test        | Model for the Z-Image SSIM test. Deprecated names: `ZIMAGE_MODEL_DIR`.                                                                                                                                                                                   |
+| `FASTVIDEO_TEST_ZIMAGE_MODEL_REVISION`                     | str                            | `f332072aa78be7aecdf3ee76d5c247082da564a6`            | test        | Hugging Face revision of the Z-Image model for its SSIM test. Deprecated names: `ZIMAGE_MODEL_REVISION`.                                                                                                                                                 |
+
+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 |
+<!-- END GENERATED ENV TABLE -->
@@ -112,7 +112,7 @@ per-metric policy with direction, percent threshold, absolute threshold, and a

 `test_inference_performance.py` temporarily sets `FASTVIDEO_STAGE_LOGGING=1`
 while it runs so pipeline stage execution times are available in
-`generate_video(...).logging_info`. Stage logs use pipeline-unique keys such as
+`generate(...).logging_info`. Stage logs use pipeline-unique keys such as
 `prompt_encoding_stage` so duplicate stage classes do not collide. For
 `PipelineStage` entries, shared component stage bases emit a stable
 `component_metric`: text encoding stages map to `text_encoder_time_s`,
@@ -10,6 +10,7 @@ slash-command mappings, and workflow ownership live in
 |---|---|---|
 | Unit tests | `fastvideo/tests/api`, `fastvideo/tests/dataset`, `fastvideo/tests/entrypoints`, `fastvideo/tests/workflow`, CPU-safe `fastvideo/tests/train` subsets | Validate individual functions, APIs, contracts, and lightweight workflows. |
 | Component tests | `fastvideo/tests/encoders`, `fastvideo/tests/transformers`, `fastvideo/tests/vaes` | Validate loading and basic behavior for model components. |
+| Golden gates | `fastvideo/tests/golden_gate` | Compare small, deterministic component outputs exactly against device/runtime-matched reference tensors. |
 | Train framework tests | `fastvideo/tests/train/models`, `fastvideo/tests/train/methods` | Exercise the new `fastvideo/train/` framework on real checkpoints and tiny synthetic batches. |
 | SSIM tests | `fastvideo/tests/ssim` | Compare generated videos against references to catch visual regressions. |
 | Training tests | `fastvideo/tests/training` | Validate legacy training loops, LoRA, distillation, self-forcing, and VSA behavior. |
@@ -20,14 +21,74 @@ slash-command mappings, and workflow ownership live in

 ## Running Tests Locally

-Run the narrowest useful suite while iterating:
+Start with the cheapest checks that cover the changed behavior, and stop on a
+failure before starting heavier dependent checks:
+
+1. Run pre-commit on changed files and focused import, config, and contract tests.
+2. Run the smallest matching component golden gate. Prefer one GPU, tiny fixed
+   inputs, cached component weights, and direct tensor comparisons over renders.
+3. Run focused component parity or default SSIM when the golden does not cover
+   the changed behavior, such as VAE normalization or pipeline wiring.
+4. Run full-quality renders or broad suites when required by the change, an
+   explicit request, or CI policy, rather than on every edit.
+
+A golden must cover the component being changed. Wan has four small gates:
+
+| Gate | Boundary |
+|---|---|
+| `test_wan_t2v.py` | Dense transformer block 0 |
+| `test_wan_vae.py` | FP32 encode, BF16 decode, streaming, and cache reset |
+| `test_wan_causal.py` | Real block weights, cache append/rewrite, and sink eviction |
+| `test_wan_denoising.py` | Three real 1.3B DiT/UniPC steps with fixed prompt embeddings |
+
+These use an immutable Wan checkpoint revision and only download the required
+component. The trajectory gate needs no tokenizer, text encoder, VAE, or video
+reference. Weight-free stage tests also exercise every step of a 50-step UniPC
+loop, CFG caching, expert switching, conditioning layouts, and DMD RNG order.
+These checks do not replace independent Diffusers parity or end-to-end SSIM.
+
+Use the golden's matching GPU, dtype, backend, and runtime. A missing reference
+or environment mismatch is not a pass. Do not replace a reference with the
+candidate output just to clear a failure. For a relocation, the unchanged
+parent is the baseline; two imports of the same class are not numerical proof.
+
+The VAE and transformer CI lanes run their Wan component goldens first.
+Selected integration lanes wait for the golden lane in merge/full builds.
+See [CI/CD Architecture](ci_architecture.md) for direct-rerun and skip semantics.
+
+For Wan, one command enforces the local ordering and stops on failure:
+
+The initial contracts include `tests/api/test_wan_definitions.py`: all registered
+Wan aliases, local-manifest detector precedence, sampling/precision defaults,
+config isolation, and legacy serialized-config compatibility. These require no
+weights or Hub access; the current package import still needs its prepared
+runtime environment.

 ```bash
-pytest tests/
-pytest fastvideo/tests/ -v
-pytest fastvideo/tests/encoders -vs
-pytest fastvideo/tests/transformers -vs
-pytest fastvideo/tests/vaes -vs
+bash scripts/validate_wan.sh vae           # contracts, then the VAE golden
+bash scripts/validate_wan.sh dense parity  # contracts, goldens, Diffusers parity
+bash scripts/validate_wan.sh all default   # then focused T2V/I2V/causal SSIM
+```
+
+The second argument is an upper validation tier, not a reference override.
+Supply the GPU/backend/runtime and SSIM model/tier settings that match the
+reference. The script never updates references. Causal component coverage is
+a block-cache fingerprint, not independent full causal-pipeline parity.
+
+New named-tensor gates write a missing output to `*.candidate.pt` and fail;
+running candidate code again cannot turn it into an approved reference. Seed
+from unchanged, pushed source, verify two independent processes bit-for-bit,
+then review and publish only the new reference files. Preserve the baseline
+source SHA, test recipe SHA, checkpoint revision, runtime, and comparison
+receipt with the run artifacts. Never overwrite an existing video reference
+as a side effect of adding a tensor gate.
+
+Examples of focused checks:
+
+```bash
+pytest fastvideo/tests/loader/test_wan_family_imports.py -q
+pytest fastvideo/tests/golden_gate/test_wan_t2v.py -q
+pytest fastvideo/tests/vaes/test_wan_vae.py -q
 ```

 GPU-heavy suites need the right hardware, credentials, local caches, and
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>World-generation prompts work best describing a scene and camera motion; the built-in prompt in the example is a known-good starting point.</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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- World-generation prompts work best describing a scene and camera motion; the built-in prompt in the example is a known-good starting point.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,24 +112,30 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>FLUX.1 dev defaults to a local <code>official_weights/FLUX.1-dev</code> directory in the example; the cookbook command passes the Hugging Face ID explicitly instead.</li>
+        <li>Image outputs land under <code>outputs/</code>; adjust <code>--output</code> if that path is not writable.</li>
+        <li>Gated checkpoints (FLUX.1 dev): run <code>huggingface-cli login</code> and accept the license on Hugging Face first.</li>
+      </ul>
+  </div>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- FLUX.1 dev defaults to a local `official_weights/FLUX.1-dev` directory in the example; the cookbook command passes the Hugging Face ID explicitly instead.
- Image outputs land under `outputs/`; adjust `--output` if that path is not writable.
- Gated checkpoints (FLUX.1 dev): run `huggingface-cli login` and accept the license on Hugging Face first.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -5,7 +5,7 @@ hide:

 # GLM-Image recipes

-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="glm_image" data-recipes="../../assets/cookbook-recipes.json?v=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<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>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- The editing example reads `assets/images/couple.jpg` from the repository root, so run it from a repo checkout rather than an arbitrary working directory.
- Output paths default under `image_output/`; pass `--output` to change them.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,24 +112,30 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Out of memory: <code>basic_hy15.py</code> already enables dit/VAE/text-encoder CPU offload; further tradeoffs are described in <a href="../../inference/optimizations/">Optimizations</a>.</li>
+        <li><code>pin_cpu_memory</code> errors on low-RAM machines are documented inline in the example; set it to false as the source comment suggests.</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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Out of memory: `basic_hy15.py` already enables dit/VAE/text-encoder CPU offload; further tradeoffs are described in [Optimizations](../inference/optimizations.md).
- `pin_cpu_memory` errors on low-RAM machines are documented inline in the example; set it to false as the source comment suggests.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -41,13 +41,14 @@ hide:
        <span class="cookbook-family-tile__footer">
          <span class="cookbook-family-tile__footer-top">
            <span><strong>MiniMax H3</strong><small>Video + stereo audio</small></span>
-            <span class="cookbook-count">6 recipes</span>
+            <span class="cookbook-count">9 recipes</span>
          </span>
          <ul class="cookbook-mode-row">
            <li>T2VA</li>
            <li>FL2VA</li>
            <li>Ref2VA</li>
            <li>MLX T2VA</li>
+            <li>DGX Spark</li>
          </ul>
        </span>
      </a>
@@ -454,16 +455,6 @@ hide:
      </article>
    </div>
  </section>
-
-  <section class="cookbook-roadmap" aria-labelledby="roadmap-heading">
-    <h2 id="roadmap-heading">Inference first, then the full workflow</h2>
-    <p>
-      Inference is the first complete stage. Distillation, fine-tuning,
-      training, evaluation, optimization, and deployment will reuse the same
-      family-first structure as their recipes land. Each family page shows
-      which stages are available and which are planned.
-    </p>
-  </section>
 </div>

 <small class="cookbook-logo-credit">
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Alternative Lite/Pro checkpoints are listed inline in the maintained examples; swap the model string only after checking its Hugging Face card.</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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Alternative Lite/Pro checkpoints are listed inline in the maintained examples; swap the model string only after checking its Hugging Face card.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-Both recipes map to checked-in FastVideo examples and recorded single-GPU B200 runs. The image-to-video run also records 10,365.89 MB peak GPU memory. These measurements describe the recorded runs; they are not minimum hardware requirements.
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>Both recipes map to checked-in FastVideo examples and recorded single-GPU B200 runs. The image-to-video run also records 10,365.89 MB peak GPU memory. These measurements describe the recorded runs; they are not minimum hardware requirements.</p>
+  </div>
+</details>
@@ -5,7 +5,7 @@ hide:

 # LongCat recipes

-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="longcat" data-recipes="../../assets/cookbook-recipes.json?v=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<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>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Each LongCat script runs multiple passes (basic, distilled, refine); total runtime scales accordingly, and every pass prints its own output directory.
- Out of memory: the sources already enable VAE and text-encoder CPU offload; further options are covered in [Offloading](../inference/offloading.md).
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,24 +112,30 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>The distilled recipe is source-configured for four GPUs; running it on fewer GPUs is unverified and may fail during distributed setup.</li>
+        <li>Audio-less output usually means the base checkpoint resolved instead of the distilled LTX-2 checkpoint with audio; check the loaded model ID in the logs.</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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- The distilled recipe is source-configured for four GPUs; running it on fewer GPUs is unverified and may fail during distributed setup.
- Audio-less output usually means the base checkpoint resolved instead of the distilled LTX-2 checkpoint with audio; check the loaded model ID in the logs.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Matrix Game 3.0 downloads its reference input image from GitHub; offline machines should pre-download it and edit the <code>IMAGE_URL</code> constant locally.</li>
+        <li>Streaming variants of Matrix Game 2.0 exist under <code>examples/inference/basic/basic_matrixgame2_streaming.py</code> but are not included as cookbook recipes.</li>
+      </ul>
+  </div>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Matrix Game 3.0 downloads its reference input image from GitHub; offline machines should pre-download it and edit the `IMAGE_URL` constant locally.
- Streaming variants of Matrix Game 2.0 exist under `examples/inference/basic/basic_matrixgame2_streaming.py` but are not included as cookbook recipes.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -5,7 +5,13 @@ hide:

 # MiniMax H3 recipes

-<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=6">
+FastH3 is two distilled MiniMax-H3 checkpoints. **V1** is the four-step
+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).
+
+<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">
@@ -15,8 +21,8 @@ hide:
      <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 or Apple Silicon MLX to iterate on prompts, or call the pipeline directly from Python.</p>
-        <span class="cookbook-count" data-cookbook-count>6 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">
@@ -29,14 +35,21 @@ hide:
      <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>

  <details class="cookbook-modes">
    <summary>Compare H3 modes and options</summary>
    <h2 id="h3-modes-heading">Supported modes</h2>
    <p>
      CUDA covers T2VA, FL2VA, and Ref2VA on the full checkpoint, plus FastH3
-      Preview and FastH3 LoRA. MLX is T2VA only. Temporal <code>--fast</code>,
-      spatial <code>--fast-spatial</code>, and opt-in VSA are flags on the same
+      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.
    </p>
    <div class="cookbook-modes__table-wrap">
@@ -51,8 +64,8 @@ hide:
        <tbody>
          <tr>
            <td>T2VA</td>
-            <td>Full H3, FastH3 Preview, FastH3 LoRA</td>
-            <td>FastH3 Preview after a local DiT conversion</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>
            <td>FL2VA</td>
@@ -84,6 +97,11 @@ hide:
            <td>No cookbook recipe</td>
            <td>Not wired</td>
          </tr>
+          <tr>
+            <td>DGX Spark</td>
+            <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>
        </tbody>
      </table>
    </div>
@@ -92,7 +110,7 @@ hide:
  <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 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>

@@ -118,6 +136,17 @@ hide:
          </div>
        </div>

+        <div class="cookbook-selection-row" data-cookbook-device-row hidden>
+          <div class="cookbook-selection-row__label">
+            <strong>Devices</strong>
+            <span data-cookbook-device-caption>1 Spark or a QSFP pair</span>
+          </div>
+          <div class="cookbook-option-grid cookbook-option-grid--hardware" data-cookbook-device-options role="group" aria-label="Devices">
+          </div>
+        </div>
+
+        <div data-cookbook-knobs></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">
@@ -217,36 +246,44 @@ hide:
  </section>
 </div>

-## Before you run
+<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>
+      <p class="cookbook-eyebrow">CUDA</p>
+      <pre><code>UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"</code></pre>
+      <p class="cookbook-eyebrow">Apple Silicon</p>
+      <pre><code>uv pip install -e ".[mlx]"</code></pre>
+      <p>Follow the <a href="../../getting_started/installation/mlx/">MLX install guide</a> for the extra, <code>ffmpeg</code>, and a clone. Then pick FastH3 V1 or V2 in the builder above.</p>
+      <p class="cookbook-eyebrow">NVIDIA DGX Spark</p>
+      <pre><code>UV_TORCH_BACKEND=cu130 uv pip install -e .</code></pre>
+      <p>Follow the <a href="../../getting_started/installation/spark/">DGX Spark install guide</a> for ARM64 CUDA 13. One Spark is a local process. Two Sparks need Ray on the QSFP link:</p>
+      <pre><code>uv pip install ray</code></pre>
+      <p>Bring up the cluster from <a href="../../getting_started/installation/spark_pair/">pairing two Sparks</a> before selecting 2 Sparks in the builder.</p>
+  </div>
+</details>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>The full CUDA H3 examples request four GPUs by default. Their sources do not claim a GPU model or memory minimum.</li>
+        <li>The FastH3 CUDA performance profile was measured on four GB200 GPUs. Use its strict profile when exact operation order matters more than the measured performance configuration.</li>
+        <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>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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-CUDA FastH3 uses the pinned performance dependencies:
-
-    UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"
-
-Apple Silicon uses the native MLX extra and a locally converted H3 DiT:
-
-    uv pip install -e ".[mlx]"
-
-Follow the [Apple Silicon guide](../getting_started/installation/mps.md#run-fasth3-preview)
-for the download, conversion, and storage requirements.
-
-## Troubleshooting
-
- The full CUDA H3 examples request four GPUs by default. Their sources do not claim a GPU model or memory minimum.
- The FastH3 CUDA performance profile was measured on four GB200 GPUs. Use its strict profile when exact operation order matters more than the measured performance configuration.
- The MLX source runtime supports T2VA, optional temporal `--fast`, optional spatial `--fast-spatial`, and opt-in VSA on `--include-vsa` checkpoints. FL2VA, Ref2VA, and two-pass refinement are not wired.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-Every command, model ID, and flag on this page maps to a checked-in FastVideo source. Recipes marked **Verified** also have a recorded hardware path in linked FastVideo evidence. The full H3 CUDA examples remain **Source-backed** where the source records a GPU count but no GPU model or memory requirement. Unlisted hardware is unknown, not unsupported.
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>Every command, model ID, and flag on this page maps to a checked-in FastVideo source. Recipes marked <strong>Verified</strong> also have a recorded hardware path in linked FastVideo evidence. The full H3 CUDA examples remain <strong>Source-backed</strong> where the source records a GPU count but no GPU model or memory requirement. Unlisted hardware is unknown, not unsupported.</p>
+  </div>
+</details>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>If the example reports a missing local <code>converted_weights/mmaudio/large_44k_v2</code> path, the <code>MMAUDIO_MODEL_PATH</code> env var from the cookbook command was not applied; export it in the same shell.</li>
+        <li>Alternatively convert upstream weights yourself with <code>scripts/checkpoint_conversion/convert_mmaudio_to_diffusers.py</code> and point the env var at the result.</li>
+      </ul>
+  </div>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- If the example reports a missing local `converted_weights/mmaudio/large_44k_v2` path, the `MMAUDIO_MODEL_PATH` env var from the cookbook command was not applied; export it in the same shell.
- Alternatively convert upstream weights yourself with `scripts/checkpoint_conversion/convert_mmaudio_to_diffusers.py` and point the env var at the result.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -1,6 +1,6 @@
 # Run H3 with a server and playground

-Start FastH3 on CUDA or Apple Silicon MLX, then iterate on prompts in a browser,
+Start FastH3 on CUDA, one DGX Spark, or Apple Silicon MLX, then iterate on prompts in a browser,
 with cURL, or from your app. The server and clients can run on the same machine.

 The server supports the OpenAI-compatible video-job API. The Python and
@@ -8,8 +8,9 @@ JavaScript client examples use that interface, but requests go to FastVideo.
 You do not need an OpenAI account or cloud key.

 The [H3 recipe selector](minimax-h3.md) provides the same workflow with runtime
-selection. This guide covers FastH3 Preview text-to-video/audio. Other H3
-recipes keep their direct Python commands.
+selection. This guide covers FastH3 V1 (four forwards) and FastH3 V2
+(eight forwards). Start a server, then iterate in the playground or with the
+OpenAI Python client. Other H3 recipes keep their direct Python commands.

 CUDA requests reuse one loaded `VideoGenerator`. The Python SDK can do the same
 when you reuse the generator across `generate()` calls. MLX keeps one
@@ -36,6 +37,40 @@ advertises it as `fasth3`. It configures four CUDA GPUs but does not record
 a GPU model or VRAM requirement. This is a source-backed server profile, not
 the measured GB200 Python performance profile. Compilation is disabled.

+For FastH3 V2, use the same install and the 8-step config (nine sigma
+points, eight DiT forwards, VSA sparsity 0.8):
+
+```bash
+UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"
+fastvideo serve --config examples/serving/openai_fasth3_8step.yaml --server.host 127.0.0.1
+```
+
+Keep the server running. In another terminal, check readiness:
+
+```bash
+curl --fail-with-body http://127.0.0.1:8000/health
+```
+
+After model loading completes, the response is `{"status":"ok"}`.
+
+### NVIDIA DGX Spark
+
+Complete the [DGX Spark installation](../getting_started/installation/spark.md)
+before running these commands. GB10 has no FA4 / sm_100a VSA kernel:
+
+```bash
+UV_TORCH_BACKEND=cu130 uv pip install -e .
+FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 \
+  fastvideo serve --config examples/serving/openai_fasth3_spark.yaml --server.host 127.0.0.1
+```
+
+The configuration loads `FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree`
+on one GB10 and advertises it as `fasth3`. Lazy module load still reloads
+Qwen3-VL and the DiT between phases of each request. Legal `num_frames` values
+are `17n+5`, capped at 362 (15.08 s); a 345-frame request on one Spark can OOM.
+There is no cookbook server for two Sparks; use the generate YAML after
+[pairing two Sparks](../getting_started/installation/spark_pair.md).
+
 Keep the server running. In another terminal, check readiness:

 ```bash
@@ -46,7 +81,7 @@ After model loading completes, the response is `{"status":"ok"}`.

 ### Apple Silicon MLX

-Complete the [Apple Silicon installation](../getting_started/installation/mps.md#run-fasth3-preview),
+Complete the [MLX install](../getting_started/installation/mlx.md),
 including `ffmpeg`. From your FastVideo clone, install the MLX extra:

 ```bash
@@ -77,14 +112,30 @@ LoRA selection, and alternate decoders are not exposed by this server adapter.
 Unsupported request options return HTTP 400 before a job starts.

 The server binds to `127.0.0.1:8000` and advertises `fasth3`, so the playground
-and clients below work unchanged. Do not run CUDA and MLX servers on the same
-port. MLX readiness means the pipeline is initialized; components load during
+and clients below work unchanged. Do not run CUDA, Spark, and MLX servers on the
+same port. MLX readiness means the pipeline is initialized; components load during
 generation. The first request can take longer than a repeated cached prompt.

 The MLX server has no recorded device or unified-memory requirement. The
 direct Python recipe's M4 Max measurements are not a server benchmark or a
 minimum-memory claim.

+For FastH3 V2, convert that checkpoint's transformer with `--include-vsa`
+and start the 8-step config. Do not overwrite a V1 export. The
+[MiniMax H3 cookbook](minimax-h3.md) has the same download and convert
+commands as the Python recipe.
+
+```bash
+hf download FastVideo/FastVideo-FastH3-8-Step-V2 --local-dir ./FastH3-8-Step-V2
+python 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
+python -m fastvideo.entrypoints.openai.mlx_server --config examples/serving/mlx_fasth3_8step.yaml
+```
+
+That adapter passes `num_steps=8` when the HTTP field is `num_inference_steps=9`,
+and it enables the trained VSA recipe (sparsity 0.8, 64-token tiles). Reuse the
+preview VAE, audio VAE, text encoder, and tokenizer if those directories already
+exist; edit the YAML paths if your files live elsewhere.
+
 ## Open the playground

 Open [the local H3 playground](http://127.0.0.1:8000/playground/) after startup.
@@ -110,9 +161,11 @@ or manage a GPU server for you.
 ## Generate with cURL or an SDK

 These examples use the server's resolution, frame count, and sampling defaults.
-Do not copy Sora-specific durations or resolutions onto H3. Both server configs
-use 124 frames, 24 fps, and the five-point distilled sigma schedule with four
-DiT forwards. CUDA uses 1344 × 768; MLX uses 832 × 480.
+Do not copy Sora-specific durations or resolutions onto H3. V1 configs use
+124 frames, 24 fps, and the five-point distilled sigma schedule with four DiT
+forwards. V2 configs use nine sigma points and eight DiT forwards. CUDA
+and one Spark use 1344 × 768; MLX uses 832 × 480. The OpenAI Python client is
+the same for every FastH3 server that advertises `fasth3`.

 Each client submits a job, checks for completion or failure, and downloads an
 MP4 named after the job ID. Polling stops after 30 minutes; a timeout does not
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Loader errors about monolithic checkpoints mean an upstream <code>stabilityai/stable-audio-open-*</code> ID was used; use the FastVideo converted repos from the recipes.</li>
+        <li>Duration and step knobs (<code>audio_end_in_s</code>, <code>num_inference_steps</code>) are documented inline in the example source.</li>
+      </ul>
+  </div>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Loader errors about monolithic checkpoints mean an upstream `stabilityai/stable-audio-open-*` ID was used; use the FastVideo converted repos from the recipes.
- Duration and step knobs (`audio_end_in_s`, `num_inference_steps`) are documented inline in the example source.
-
-## Evidence status
-
-The Stable Audio Open 1.0 recipe maps to a checked-in example and a recorded single-GPU B200 run. The Stable Audio Open Small recipe remains **Source-backed** because the implementation PR did not record a full run for that gated checkpoint. Neither recipe claims a minimum VRAM requirement.
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>The Stable Audio Open 1.0 recipe maps to a checked-in example and a recorded single-GPU B200 run. The Stable Audio Open Small recipe remains <strong>Source-backed</strong> because the implementation PR did not record a full run for that gated checkpoint. Neither recipe claims a minimum VRAM requirement.</p>
+  </div>
+</details>
@@ -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=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>The example writes several PNGs under <code>outputs/sd35/samples/</code>; make sure the output directory is writable.</li>
+        <li>Gated checkpoints: Stability AI models require accepting the license and <code>huggingface-cli login</code>.</li>
+      </ul>
+  </div>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- The example writes several PNGs under `outputs/sd35/samples/`; make sure the output directory is writable.
- Gated checkpoints: Stability AI models require accepting the license and `huggingface-cli login`.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -5,7 +5,7 @@ hide:

 # TurboDiffusion recipes

-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="turbodiffusion" data-recipes="../../assets/cookbook-recipes.json?v=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<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>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- TurboDiffusion paths load community-published `loayrashid/TurboWan*` checkpoints; availability is governed by those repos.
- The SLA attention backend used by the I2V recipe is selected inside the example source; do not combine it with another `FASTVIDEO_ATTENTION_BACKEND` override in the same shell.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -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=6">
+<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">
@@ -29,8 +29,15 @@ hide:
      <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-modes" aria-labelledby="wan-modes-heading">
+  <details class="cookbook-modes">
+    <summary>Compare Wan modes and options</summary>
    <h2 id="wan-modes-heading">Supported modes</h2>
    <p>
      FastMetal MLX is T2V in the checked-in examples. Image-to-video and
@@ -84,7 +91,7 @@ hide:
        </tbody>
      </table>
    </div>
-  </section>
+  </details>

  <section class="cookbook-builder" id="recipe-builder" aria-labelledby="builder-heading">
    <div class="cookbook-builder__intro">
@@ -163,26 +170,32 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <li>Out of memory on the A14B recipes: the checked-in sources already enable CPU offload; see <a href="../../inference/configuration/">Configuration</a> for the offload surface before reducing resolution or frames.</li>
+        <li>The FastWan2.1 recipe requires <code>VIDEO_SPARSE_ATTN</code>; confirm the environment variable in the command was set in the same shell.</li>
+        <li>FastMetal MLX: install with the <a href="../../getting_started/installation/mlx/">MLX install guide</a>, then pick a FastMetal recipe in the builder. CUDA FastWan-QAD checkpoints are refused on the MLX runtime.</li>
+        <li>FastMetal 5B uses <code>mlx_wan22_generate.py</code>. 1.3B and 14B use <code>mlx_wan_prompt_to_video.py</code>.</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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Out of memory on the A14B recipes: the checked-in sources already enable CPU offload; see [Configuration](../inference/configuration.md) for the offload surface before reducing resolution or frames.
- The FastWan2.1 recipe requires `VIDEO_SPARSE_ATTN`; confirm the environment variable in the command was set in the same shell.
- FastMetal MLX: install with `uv pip install -e ".[mlx]"`, then follow the [Apple Silicon guide](../getting_started/installation/mps.md). CUDA FastWan-QAD checkpoints are refused on the MLX runtime.
- FastMetal 5B uses `mlx_wan22_generate.py`. 1.3B and 14B use `mlx_wan_prompt_to_video.py`.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-Every recipe on this page maps to a checked-in FastVideo source. The FastMetal MLX releases include the recorded M4 Max system memory, documented unified-memory floor, and measured peak MLX memory. CUDA entries remain **Source-backed** where the examples record a GPU count but no exact GPU model or VRAM. Unlisted hardware is unknown, not unsupported.
+<details class="cookbook-collapsible" id="cookbook-evidence">
+  <summary>Evidence status</summary>
+  <div class="cookbook-collapsible__body">
+      <p>Every recipe on this page maps to a checked-in FastVideo source. The FastMetal MLX releases include the recorded M4 Max system memory, documented unified-memory floor, and measured peak MLX memory. CUDA entries remain <strong>Source-backed</strong> where the examples record a GPU count but no exact GPU model or VRAM. Unlisted hardware is unknown, not unsupported.</p>
+  </div>
+</details>
@@ -5,7 +5,7 @@ hide:

 # Z-Image recipes

-<div class="cookbook-shell cookbook-family-page" data-cookbook data-family="zimage" data-recipes="../../assets/cookbook-recipes.json?v=6">
+<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">
@@ -28,6 +28,12 @@ hide:
      <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">
@@ -106,23 +112,29 @@ hide:
  </section>
 </div>

-## Before you run
+<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>

-The generated commands expect a local clone:
+<details class="cookbook-collapsible" id="cookbook-troubleshooting">
+  <summary>Troubleshooting</summary>
+  <div class="cookbook-collapsible__body">
+      <ul>
+        <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>
+</details>

-    git clone https://github.com/hao-ai-lab/FastVideo.git
-    cd FastVideo
-
-Use [Configuration](../inference/configuration.md) for supported Python and
-CLI settings, [Optimizations](../inference/optimizations.md) for attention and
-memory tradeoffs, and the [support matrix](../inference/support_matrix.md) for
-the supported model and optimization surface.
-
-## Troubleshooting
-
- Output defaults to `outputs/zimage/zimage_turbo.png`; pass `--output` to redirect.
- Gated or missing checkpoints: run `huggingface-cli login` and confirm you accepted the model's license on Hugging Face.
-
-## Evidence status
-
-All recipes on this page are **Source-backed**: 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 **Unknown** and deliberately not claimed. Runtime buttons show only the GPU counts configured in checked-in sources.
+<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>
@@ -2,132 +2,158 @@ status_definitions:
  kept: "Public field remains on a public adapter surface with the same meaning."
  moved: "Public field remains supported but normalizes into a different nested path."
  preset_owned: "Public field remains supported only through a model/preset-specific surface."
-  compatibility_only: "Legacy public field remains adapter-only during migration and is not part of the canonical typed schema."
+  compatibility_only: "Public field that an adapter or an open mapping accepts outside the typed fields of the canonical schema."
  private_only: "Field should only be handled by private adapters and is not a public FastVideo compatibility promise."
-  internal_only: "Field is runtime/config plumbing and should not be part of the new public typed inference API."
+  internal_only: "Field is runtime/config plumbing that model code, config resolution, or the runtime fills; it is not a public input."
+  unsupported: "Typed config field that no runtime code reads; resolution rejects a value."

 surfaces:
-  fastvideo_args:
+  generator_config:
+    kept:
+      - model_path
+      - mode
+      - revision
+      - trust_remote_code
+      - engine.num_gpus
+      - engine.execution_backend
+      - engine.parallelism.tp_size
+      - engine.parallelism.sp_size
+      - engine.parallelism.hsdp_replicate_dim
+      - engine.parallelism.hsdp_shard_dim
+      - engine.parallelism.dist_timeout
+      - engine.offload.dit
+      - engine.offload.dit_layerwise
+      - engine.offload.text_encoder
+      - engine.offload.image_encoder
+      - engine.offload.vae
+      - engine.offload.pin_cpu_memory
+      - engine.compile.enabled
+      - engine.compile.backend
+      - engine.compile.fullgraph
+      - engine.compile.mode
+      - engine.compile.dynamic
+      - engine.compile.extras
+      - engine.enable_stage_verification
+      - engine.use_fsdp_inference
+      - engine.disable_autocast
+      - engine.attention.nvfp4_fa4
+      - pipeline.components.lora_path
+      - pipeline.components.lora_strength
+      - pipeline.output_type
+      - engine.parallelism.master_port
+      - engine.offload.lazy_module_load
+      - engine.compile.text_encoder_enabled
+      - engine.compile.vae_enabled
+      - engine.compile.audio_vae_enabled
+      - engine.compile.regional
+      - engine.compile.dit_kwargs
+      - engine.compile.text_encoder_kwargs
+      - engine.compile.vae_kwargs
+      - engine.compile.audio_vae_kwargs
+      - engine.attention.backend
+      - engine.attention.vsa_sparsity
+      - engine.attention.vsa_tile_size
+      - engine.attention.moba_config_path
+      - engine.precision.dit
+      - engine.precision.vae
+      - engine.precision.vae_decode
+      - engine.precision.image_encoder
+      - engine.precision.text_encoders
+      - engine.quantization.text_encoder_quant
+      - engine.quantization.transformer_quant
+      - pipeline.workload_type
+      - pipeline.components.config_root
+      - pipeline.components.pipeline_config_path
+      - pipeline.components.text_encoder_weights
+      - pipeline.components.transformer_weights
+      - pipeline.components.transformer_2_weights
+      - pipeline.components.upsampler_weights
+      - pipeline.components.lora_nickname
+      - pipeline.components.lora_target_modules
+      - pipeline.components.override_pipeline_cls_name
+      - pipeline.components.override_transformer_cls_name
+      - pipeline.vae_tiling
+      - pipeline.vae_sp
+      - pipeline.flow_shift
+      - pipeline.embedded_cfg_scale
+      - pipeline.dmd_denoising_steps
+      - pipeline.boundary_ratio
+      - pipeline.model.generic.dit
+      - pipeline.model.generic.vae
+      - pipeline.model.ltx2.dit
+      - pipeline.model.ltx2.vae
+      - pipeline.model.minimax_h3.dit
+      - pipeline.model.minimax_h3.vae
+      - pipeline.model.longcat.dit
+      - pipeline.model.longcat.vae
+    unsupported:
+      - pipeline.preset
+      - pipeline.preset_version
+      - pipeline.components.vae_weights
    moved:
-      model_path: generator.model_path
-      workload_type: generator.pipeline.workload_type
-      distributed_executor_backend: generator.engine.execution_backend
-      trust_remote_code: generator.trust_remote_code
-      revision: generator.revision
-      num_gpus: generator.engine.num_gpus
-      tp_size: generator.engine.parallelism.tp_size
-      sp_size: generator.engine.parallelism.sp_size
-      hsdp_replicate_dim: generator.engine.parallelism.hsdp_replicate_dim
-      hsdp_shard_dim: generator.engine.parallelism.hsdp_shard_dim
-      dist_timeout: generator.engine.parallelism.dist_timeout
-      lora_path: generator.pipeline.components.lora_path
-      lora_nickname: generator.pipeline.components.lora_nickname
-      lora_strength: generator.pipeline.components.lora_strength
-      dit_cpu_offload: generator.engine.offload.dit
-      use_fsdp_inference: generator.engine.use_fsdp_inference
-      dit_layerwise_offload: generator.engine.offload.dit_layerwise
-      text_encoder_cpu_offload: generator.engine.offload.text_encoder
-      image_encoder_cpu_offload: generator.engine.offload.image_encoder
-      vae_cpu_offload: generator.engine.offload.vae
-      pin_cpu_memory: generator.engine.offload.pin_cpu_memory
-      enable_torch_compile: generator.engine.compile.enabled
-      enable_torch_compile_text_encoder: generator.engine.compile.text_encoder_enabled
-      enable_torch_compile_vae: generator.engine.compile.vae_enabled
-      enable_torch_compile_audio_vae: generator.engine.compile.audio_vae_enabled
-      torch_compile_kwargs: generator.engine.compile.backend,fullgraph,mode,dynamic,extras
-      torch_compile_kwargs_dit: generator.engine.compile.dit_kwargs
-      torch_compile_kwargs_text_encoder: generator.engine.compile.text_encoder_kwargs
-      torch_compile_kwargs_vae: generator.engine.compile.vae_kwargs
-      torch_compile_kwargs_audio_vae: generator.engine.compile.audio_vae_kwargs
-      transformer_quant: generator.engine.quantization.transformer_quant
-      disable_autocast: generator.engine.disable_autocast
-      enable_stage_verification: generator.engine.enable_stage_verification
-      prompt_txt: request.inputs.prompt_path
-      override_text_encoder_safetensors: generator.pipeline.components.text_encoder_weights
-      override_text_encoder_quant: generator.engine.quantization.text_encoder_quant
-      transformer_quant: generator.engine.quantization.transformer_quant
-      override_transformer_cls_name: generator.pipeline.components.override_transformer_cls_name
-      init_weights_from_safetensors: generator.pipeline.components.transformer_weights
-      init_weights_from_safetensors_2: generator.pipeline.components.transformer_2_weights
-      override_pipeline_cls_name: generator.pipeline.components.override_pipeline_cls_name
-      boundary_ratio: request.sampling.boundary_ratio
-      ltx2_vae_tiling: generator.pipeline.vae_tiling
-      refine_enabled: generator.pipeline.preset_overrides.refine.enabled
-      refine_upsampler_path: generator.pipeline.components.upsampler_weights
-      refine_lora_path: generator.pipeline.components.lora_path
-      refine_num_inference_steps: request.stage_overrides.refine.num_inference_steps
-      refine_guidance_scale: request.stage_overrides.refine.guidance_scale
-      refine_add_noise: generator.pipeline.preset_overrides.refine.add_noise
-      ltx2_refine_enabled: generator.pipeline.preset_overrides.refine.enabled
-      ltx2_refine_upsampler_path: generator.pipeline.components.upsampler_weights
-      ltx2_refine_lora_path: generator.pipeline.components.lora_path
-      ltx2_refine_num_inference_steps: request.stage_overrides.refine.num_inference_steps
-      ltx2_refine_guidance_scale: request.stage_overrides.refine.guidance_scale
-      ltx2_refine_add_noise: generator.pipeline.preset_overrides.refine.add_noise
+      pipeline.preset_overrides:
+        target: generator.pipeline.model.ltx2.refine
+        note: "Only the refine mapping applies: resolution copies pipeline.preset_overrides.refine into the pipeline.model.ltx2.refine fields of the same names when the model is LTX-2. Other keys have no effect."
    preset_owned:
-      ltx2_vae_spatial_tile_size_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_size_in_pixels
-      ltx2_vae_spatial_tile_overlap_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_overlap_in_pixels
-      ltx2_vae_temporal_tile_size_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_size_in_frames
-      ltx2_vae_temporal_tile_overlap_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_overlap_in_frames
-      ltx2_initial_latent_path: request.extensions.ltx2.initial_latent_path
-      ltx2_audio_latent_path: request.extensions.ltx2.audio_latent_path
+      - pipeline.model.ltx2.vae_spatial_tile_size_in_pixels
+      - pipeline.model.ltx2.vae_spatial_tile_overlap_in_pixels
+      - pipeline.model.ltx2.vae_temporal_tile_size_in_frames
+      - pipeline.model.ltx2.vae_temporal_tile_overlap_in_frames
+      - pipeline.model.ltx2.initial_latent_path
+      - pipeline.model.ltx2.audio_latent_path
+      - pipeline.model.ltx2.legacy_native_noise_order
+      - pipeline.model.ltx2.use_distilled_sigmas
+      - pipeline.model.ltx2.refine.enabled
+      - pipeline.model.ltx2.refine.num_inference_steps
+      - pipeline.model.ltx2.refine.guidance_scale
+      - pipeline.model.ltx2.refine.add_noise
+      - pipeline.model.ltx2.refine.image_crf
+      - pipeline.model.ltx2.refine.video_position_offset_sec
+      - pipeline.model.ltx2.refine.transformer_path
+      - pipeline.model.ltx2.refine.lora_path
+      - pipeline.model.ltx2.refine.noise_path
+      - pipeline.model.ltx2.refine.audio_noise_path
+      - pipeline.model.minimax_h3.sequential_load
+      - pipeline.model.minimax_h3.video_decode_backend
+      - pipeline.model.minimax_h3.taeh3_checkpoint
+      - pipeline.model.minimax_h3.taeh3_chunk_size
+      - pipeline.model.minimax_h3.vae_parallel_decode
+      - pipeline.model.minimax_h3.vae_parallel_encode
+      - pipeline.model.minimax_h3.vae_parallel_decode_strategy
+      - pipeline.model.longcat.enable_bsa
+      - pipeline.model.longcat.bsa_sparsity
+      - pipeline.model.longcat.bsa_cdf_threshold
+      - pipeline.model.longcat.bsa_chunk_q
+      - pipeline.model.longcat.bsa_chunk_k
    compatibility_only:
-      mode: "Legacy multi-mode FastVideoArgs switch; typed inference config should not expose execution mode."
-      inference_mode: "Legacy boolean mirror of mode; kept only through adapters while FastVideoArgs remains."
-      lora_target_modules: "Legacy LoRA configuration surface pending dedicated component API."
-      output_type: "Legacy output formatting surface pending GenerationResult cleanup."
-      VSA_sparsity: "Model-specific inference optimization not yet represented in the typed public schema."
-      VSA_tile_size: "VSA-H3 tile geometry request; model-specific optimization not yet represented in the typed public schema."
-      inference_torch_compile: "Regional inference compile opt-in currently carried through PipelineSelection.experimental rather than CompileConfig."
-      vae_parallel_decode: "MiniMax-H3 sequence-parallel VAE decode opt-in; model-specific optimization not yet represented in the typed public schema."
-      h3_sequential_load: "MiniMax-H3 sequential text-encoder then DiT/VAE load; model-specific optimization not yet represented in the typed public schema."
-      vae_parallel_encode: "MiniMax-H3 sequence-parallel reference VAE encode opt-in; model-specific optimization not yet represented in the typed public schema."
-      vae_parallel_decode_strategy: "Chunk-transport collective for vae_parallel_decode; model-specific optimization not yet represented in the typed public schema."
-      attention_backend: "Process-wide default attention-backend request applied per component at load time; kernel-selection knob not yet represented in the typed public schema."
-      moba_config_path: "Model-specific MoBA optimization surface not yet represented in the typed public schema."
-      master_port: "Executor/bootstrap compatibility field; not part of the canonical inference schema."
-      refine_transformer_path: "Generic stage-2 refine transformer override; no typed equivalent yet."
-      refine_noise_path: "Generic stage-2 refine noise override; no typed equivalent yet."
-      refine_audio_noise_path: "Generic stage-2 refine audio noise override; no typed equivalent yet."
-      ltx2_refine_transformer_path: "LTX-2 refine transformer carrier; no typed equivalent yet."
-      ltx2_refine_noise_path: "LTX-2 refine noise carrier; no typed equivalent yet."
-      ltx2_refine_audio_noise_path: "LTX-2 refine audio noise carrier; no typed equivalent yet."
-      ltx2_legacy_native_noise_order: "LTX-2 SSIM compatibility knob preserving legacy native latent noise ordering."
-      ltx2_use_distilled_sigmas: "LTX-2 compatibility knob gating use of distilled sigma schedule."
-    private_only:
-      ray_placement_group: "Ray deployment-only field."
-      ray_runtime_env: "Ray deployment-only field."
+      pipeline.experimental: "Open mapping for settings without a typed path: the pipeline_config source (a JSON path, a mapping, or a PipelineConfig), keys that runtime code reads by name (for example ray_runtime_env), and PipelineConfig attribute overrides for model-only fields (for example flow_shift_sr). Resolution rejects a key whose PipelineConfig attribute has a typed path."
    internal_only:
-      pipeline_config: "Legacy internal carrier object."
-      preprocess_config: "Legacy preprocess carrier object."
-      moba_config: "Derived runtime config loaded from moba_config_path."
-      model_paths: "Runtime bookkeeping."
-      model_loaded: "Runtime bookkeeping."
+      engine.attention.moba_config: "V-MoBA attention settings that resolution loads from engine.attention.moba_config_path."

  pipeline_config_base:
    moved:
+      model_path: generator.model_path
      pipeline_config_path: generator.pipeline.components.pipeline_config_path
+      embedded_cfg_scale: generator.pipeline.embedded_cfg_scale
+      flow_shift: generator.pipeline.flow_shift
+      disable_autocast: generator.engine.disable_autocast
+      vae_tiling: generator.pipeline.vae_tiling
+      vae_sp: generator.pipeline.vae_sp
+      dmd_denoising_steps: generator.pipeline.dmd_denoising_steps
+      boundary_ratio: generator.pipeline.boundary_ratio
+      dit_precision: generator.engine.precision.dit
+      vae_precision: generator.engine.precision.vae
+      vae_decode_precision: generator.engine.precision.vae_decode
+      image_encoder_precision: generator.engine.precision.image_encoder
+      text_encoder_precisions: generator.engine.precision.text_encoders
    preset_owned:
-      embedded_cfg_scale: generator.pipeline.preset_overrides.embedded_cfg_scale
-      flow_shift: generator.pipeline.preset_overrides.flow_shift
-      flow_shift_sr: generator.pipeline.preset_overrides.flow_shift_sr
-      is_causal: generator.pipeline.preset_overrides.is_causal
-      vae_tiling: generator.pipeline.preset_overrides.vae_tiling
-      vae_sp: generator.pipeline.preset_overrides.vae_sp
-      dmd_denoising_steps: generator.pipeline.preset_overrides.dmd_denoising_steps
-      ti2v_task: generator.pipeline.preset_overrides.ti2v_task
-      lucy_edit_task: generator.pipeline.preset_overrides.lucy_edit_task
-      boundary_ratio: generator.pipeline.preset_overrides.boundary_ratio
+      flow_shift_sr: generator.pipeline.experimental.flow_shift_sr
+      is_causal: generator.pipeline.experimental.is_causal
+      ti2v_task: generator.pipeline.experimental.ti2v_task
+      lucy_edit_task: generator.pipeline.experimental.lucy_edit_task
    compatibility_only:
-      model_path: "Redundant with generator.model_path."
-      disable_autocast: "Duplicated by generator.engine.disable_autocast during migration."
-      dit_precision: "Precision override pending dedicated typed component precision design."
      upsampler_precision: "Precision override pending dedicated typed component precision design."
-      vae_precision: "Precision override pending dedicated typed component precision design."
-      vae_decode_precision: "Decode-only precision override pending dedicated typed component precision design."
-      image_encoder_precision: "Precision override pending dedicated typed component precision design."
      image_encoder_precisions: "Precision overrides pending dedicated typed component precision design."
-      text_encoder_precisions: "Precision override pending dedicated typed component precision design."
    internal_only:
      dit_config: "Legacy internal component config object."
      upsampler_config: "Legacy internal component config object."
@@ -140,6 +166,22 @@ surfaces:
      scheduler_step_in_fp32: "Runtime scheduler precision toggle; not part of the public typed inference API."

  pipeline_config_extensions:
+    moved:
+      enable_bsa:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+        target: generator.pipeline.model.longcat.enable_bsa
+      bsa_sparsity:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+        target: generator.pipeline.model.longcat.bsa_sparsity
+      bsa_cdf_threshold:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+        target: generator.pipeline.model.longcat.bsa_cdf_threshold
+      bsa_chunk_q:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+        target: generator.pipeline.model.longcat.bsa_chunk_q
+      bsa_chunk_k:
+        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
+        target: generator.pipeline.model.longcat.bsa_chunk_k
    preset_owned:
      flux2_text_encoder_type:
        sources:
@@ -320,18 +362,8 @@ 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:
@@ -516,6 +548,7 @@ surfaces:
      cfg_truncation: request.sampling.cfg_truncation
      guidance_rescale: request.sampling.guidance_rescale
      use_embedded_guidance: request.sampling.use_embedded_guidance
+      embedded_cfg_scale: request.sampling.embedded_cfg_scale
      true_cfg_scale: request.sampling.true_cfg_scale
      boundary_ratio: request.sampling.boundary_ratio
      sigmas: request.sampling.sigmas
@@ -11,6 +11,8 @@ FastVideo maps a Diffusers-style repo into a pipeline like this:
 - `fastvideo/models/*`: model implementations (DiT, VAE, encoders, upsamplers).
 - `fastvideo/configs/models/*`: arch configs and `param_names_mapping` for
  weight name translation.
+- `fastvideo/models/wan/`: Wan's transformers, VAE, configs, and variant definitions
+  together; the old Wan component/config modules remain compatibility imports.
 - `fastvideo/configs/pipelines/*`: pipeline wiring (component classes + names).
 - `fastvideo/api/sampling_param.py`: runtime sampling parameters.
 - `fastvideo/pipelines/basic/*`: end-to-end pipelines.
@@ -26,19 +28,15 @@ Minimal usage (from `examples/inference/basic/basic.py`):

 ```python
 from fastvideo import VideoGenerator
-from fastvideo.api.sampling_param import SamplingParam

 model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # or official_weights/<model_name>/
-generator = VideoGenerator.from_pretrained(model_id, num_gpus=1)
+generator = VideoGenerator.from_pretrained(model_id, {"engine": {"num_gpus": 1}})

-sampling = SamplingParam.from_pretrained(model_id)
-sampling.num_frames = 45
-video = generator.generate_video(
-    "A vibrant city street at sunset.",
-    sampling_param=sampling,
-    output_path="video_samples",
-    save_video=True,
-)
+video = generator.generate({
+    "prompt": "A vibrant city street at sunset.",
+    "sampling": {"num_frames": 45},
+    "output": {"output_path": "video_samples", "save_video": True},
+})
 ```

 ## Configuration system
@@ -48,16 +46,35 @@ runtime parameters consistent:

 - `fastvideo/configs/models/`: architecture definitions, layer shapes, and
  `param_names_mapping` rules for key renaming.
+- `fastvideo/models/wan/config.py`: Wan's transformer architecture and mapping rules,
+  co-located with `transformer.py`.
+- `fastvideo/models/wan/vae_config.py`: Wan's VAE architecture and runtime
+  settings, co-located with `vae.py`.
+- `fastvideo/models/wan/pipeline_config.py`: Wan component and pipeline defaults;
+  `configs/pipelines/wan.py` remains a compatibility import.
+- `fastvideo/models/wan/definition.py`: data-only Wan variants linking HF aliases,
+  config classes, presets, workload metadata, and default sampling algorithms.
 - `fastvideo/configs/pipelines/`: pipeline wiring and required components.
 - `fastvideo/api/sampling_param.py`: sampling parameters (steps, frames,
-  guidance scale, resolution, fps). Defaults come from profiles in
-  `fastvideo/pipelines/basic/<family>/profiles.py`.
+  guidance scale, resolution, fps). Defaults come from presets in
+  `fastvideo/pipelines/basic/<family>/presets.py`.
 - `fastvideo/registry.py`: unified registry for pipeline config + sampling
-  defaults and model metadata resolution, defined via explicit
-  `register_configs(...)` blocks (no separate dict registries).
+  defaults and model metadata resolution. Wan registrations consume its
+  family-local definitions; other families use `register_configs(...)` blocks.

-`FastVideoArgs` (in `fastvideo/fastvideo_args.py`) provides runtime settings and
-is passed into pipeline construction and stages.
+Wan definitions reference existing defaults rather than copying them. Dense
+UniPC, dense DMD, and causal DMD remain separate sampling algorithms. Dense
+DMD uses a full training-noise scheduler with shift 8.0, separate from the
+configurable scheduler mutated during timestep preparation. The catalog does
+not override checkpoint manifests, user pipeline overrides, or component
+precision settings. HF IDs, local checkpoints, and old config imports retain
+their existing resolution behavior, including first-match detector ordering.
+
+`ResolvedGeneratorConfig` (in `fastvideo/api/resolution.py`) provides runtime
+settings and is passed into pipeline construction and stages as `resolved_config`.
+`resolve_inference_config` (in `fastvideo/api/inference_resolution.py`) builds it
+from the typed config in `fastvideo/api/schema.py` and attaches the model's
+frozen `PipelineConfig` as `resolved_config.pipeline_config`.

 ## Weights and Diffusers format

@@ -96,7 +113,8 @@ Note on tensor names:

 Official checkpoints often use different `state_dict` names than FastVideo's
 module layout. We translate tensor names via the DiT arch config mapping
-(`param_names_mapping` under `fastvideo/configs/models/dits/`). This is similar
+(`param_names_mapping` under `fastvideo/configs/models/dits/`, or
+`fastvideo/models/wan/config.py` for Wan). This is similar
 in spirit to name-translation layers used in systems like vLLM and SGLang.

 Example HF repo (Wan 2.1 T2V 1.3B Diffusers):
@@ -137,13 +155,16 @@ Example `model_index.json` from that repo:
 How this maps to FastVideo:

 - `WanPipeline` -> `fastvideo/pipelines/basic/wan/wan_pipeline.py`
- `WanTransformer3DModel` -> `fastvideo/models/dits/wanvideo.py`
- `AutoencoderKLWan` -> `fastvideo/models/vaes/wanvae.py`
+- `WanTransformer3DModel` -> `fastvideo/models/wan/transformer.py`
+- `WanVideoConfig` -> `fastvideo/models/wan/config.py`
+- `AutoencoderKLWan` -> `fastvideo/models/wan/vae.py`
+- `WanVAEConfig` -> `fastvideo/models/wan/vae_config.py`
 - `UMT5EncoderModel` -> `fastvideo/models/encoders/t5.py`
 - `T5TokenizerFast` -> loaded via HF in `fastvideo/models/loader/`
 - `UniPCMultistepScheduler` -> loaded via Diffusers scheduler utilities
- Pipeline defaults -> `fastvideo/configs/pipelines/wan.py`
- Sampling defaults -> `fastvideo/pipelines/basic/wan/profiles.py`
+- Variant definitions -> `fastvideo/models/wan/definition.py`
+- Pipeline defaults -> `fastvideo/models/wan/pipeline_config.py`
+- Sampling defaults -> `fastvideo/pipelines/basic/wan/presets.py`

 ## Pipeline system

@@ -157,8 +178,8 @@ How this maps to FastVideo:

 ## Model components

- DiT models: `fastvideo/models/dits/`
- VAEs: `fastvideo/models/vaes/`
+- DiT models: `fastvideo/models/dits/`; dense Wan: `fastvideo/models/wan/`
+- VAEs: `fastvideo/models/vaes/`; Wan: `fastvideo/models/wan/vae.py`
 - Text/image encoders: `fastvideo/models/encoders/`
 - Schedulers: `fastvideo/models/schedulers/`
 - Upsamplers: `fastvideo/models/upsamplers/`
@@ -32,7 +32,7 @@ from fastvideo.api import (

 | Surface | Availability | Notes |
 | --- | --- | --- |
-| `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` | Today | `typed_kwargs` is a stable subset from `GeneratorConfig` — no flat legacy LTX-2 kwargs (guaranteed after PR 6) |
+| `VideoGenerator.from_pretrained(model_path, config)` | Today | `config` is a nested `GeneratorConfig` mapping without `model_path`; no flat keywords |
 | `VideoGenerator.generate(request: GenerationRequest) -> GenerationResult` | Today | Aggregated; Dynamo wraps in `asyncio.to_thread` under `asyncio.Lock` |
 | `VideoGenerator.generate_async(request) -> AsyncGenerator[VideoEvent, None]` | **PR 7.10** | Canonical execution substrate; sync wrapper reroutes through this |
 | `VideoGenerator.default_health_check_request() -> GenerationRequest` | **PR 7.10** | 256x256 / 8 frames / 1 step; lets Dynamo build its health payload without knowing any FastVideo internals |
@@ -225,7 +225,7 @@ async def init_video_generation(runtime, config, shutdown_endpoints):
    from fastvideo.api import config_to_dict

    server_args, dynamo_args = config.server_args, config.dynamo_args
-    generator = VideoGenerator.from_pretrained(**config.fastvideo_kwargs())
+    generator = VideoGenerator.from_config(build_generator_config(server_args))

    dump_config(dynamo_args.dump_config_to, config)

@@ -262,7 +262,8 @@ this adapter can build the config purely from the public typed schema:
 def build_generator_config(args) -> "GeneratorConfig":
    from fastvideo.api import (
        CompileConfig, ComponentConfig, EngineConfig, GeneratorConfig,
-        OffloadConfig, ParallelismConfig, PipelineSelection,
+        LTX2Options, LTX2RefineOptions, OffloadConfig, ParallelismConfig,
+        PipelineSelection,
    )
    return GeneratorConfig(
        model_path=args.model_path,
@@ -275,10 +276,8 @@ def build_generator_config(args) -> "GeneratorConfig":
        pipeline=PipelineSelection(
            workload_type=args.workload or "t2v",
            preset=args.preset,  # e.g. "ltx2_two_stage"
-            components=ComponentConfig(
-                upsampler_weights=args.refine_upsampler,
-                lora_path=args.refine_lora,
-            ),
+            components=ComponentConfig(upsampler_weights=args.refine_upsampler),
+            model=LTX2Options(refine=LTX2RefineOptions(lora_path=args.refine_lora)),
        ),
    )
 ```
@@ -310,8 +309,11 @@ re-chase FastVideo drift:
 2. `ContinuationState.payload` is JSON-serializable or references
   opaque blob ids. Dynamo can round-trip it through RPC without
   special-casing torch tensors.
-3. `VideoGenerator.from_pretrained` accepts a typed `GeneratorConfig`;
-   legacy flat kwargs are compatibility-only and deprecate in PR 13.
+3. `VideoGenerator.from_pretrained(model_path, config)` takes a typed
+   `GeneratorConfig` or its nested mapping; any flat keyword raises
+   `TypeError` that points to the nested config.
+   `VideoGenerator.from_config(...)` takes the same settings with
+   `model_path` inside.
 4. `generate_async` (PR 7.10+) emits events in order
   `Progress* → Partial* → Final`; the final event always has exactly
   one occurrence per request.
@@ -329,7 +331,6 @@ at FastVideo's CI — before the Dynamo-side integration even knows.
 * Anything under `fastvideo.pipelines.*` directly (pipelines are
  internal; presets identify them by name on
  `PipelineSelection.preset`).
-* `fastvideo.fastvideo_args.FastVideoArgs` (legacy compat type).
 * `fastvideo.api.compat.*` private helpers
  (`_validate_continuation_state` etc.) — the public boundary is
  `VideoGenerator` + `fastvideo.api`.
@@ -25,11 +25,13 @@ requests. HTTP handling and job polling remain asynchronous.
 | `GET` | `/v1/videos/{id}/content` | Download a completed MP4 |
 | `DELETE` | `/v1/videos/{id}` | Delete a job and its completed artifact |
 | `POST` | `/v1/images` | Generate an image |
+| `POST` | `/v1/images/generations` | OpenAI-compatible alias for image generation |
 | `POST` | `/v1/images/edits` | Generate an image from image references |
 | `GET` | `/v1/images/{id}/content` | Download a generated image |
 | `GET` | `/health` | Liveness probe |

-`POST /v1/videos/generations` remains an alias for older FastVideo clients.
+`POST /v1/videos/generations` remains an alias for older FastVideo clients, and
+`POST /v1/images/generations` is the OpenAI Python client's image generation path.
 The OpenAI Python and JavaScript clients can create, retrieve, list, download,
 and delete video jobs. Use the [H3 server cookbook](../../cookbook/openai-api.md)
 for pinned client versions and executable examples. Download variants other
@@ -76,6 +76,7 @@ COOKBOOK_EVIDENCE_STATES = {
 }
 COOKBOOK_HARDWARE_EVIDENCE = {"validated", "source-configured", "estimated", "unknown"}
 COOKBOOK_HARDWARE_PLATFORMS = {"cuda", "mlx", "mps"}
+COOKBOOK_HARDWARE_DEVICES = {"spark"}
 COOKBOOK_GPU_TYPES = {"NVIDIA", "Apple Silicon"}
 COOKBOOK_HARDWARE_TEXT_FIELDS = {
    "accelerator",
@@ -128,6 +129,12 @@ def validate_cookbook() -> None:
            raise ValueError(f"CUDA cookbook recipe needs an integer gpu_count >= 1: {recipe['id']}")
        if platform != "cuda" and gpu_count is not None:
            raise ValueError(f"Non-CUDA cookbook recipe must not use gpu_count: {recipe['id']}")
+        device = hardware.get("device")
+        if device is not None:
+            if device not in COOKBOOK_HARDWARE_DEVICES:
+                raise ValueError(f"Cookbook recipe has an unknown hardware device: {recipe['id']}: {device}")
+            if platform != "cuda":
+                raise ValueError(f"Cookbook hardware device requires platform cuda: {recipe['id']}: {device}")
        hardware_evidence = hardware.get("evidence")
        if hardware_evidence not in COOKBOOK_HARDWARE_EVIDENCE:
            raise ValueError(f"Cookbook recipe has unknown hardware evidence: {recipe['id']}: {hardware_evidence}\n"
@@ -163,6 +170,31 @@ def validate_cookbook() -> None:
                                  or any(not isinstance(item, str) or not item.strip() for item in modes)):
            raise ValueError(f"Cookbook recipe modes must be a non-empty list of strings: {recipe['id']}")

+        knobs = recipe.get("knobs", [])
+        if not isinstance(knobs, list):
+            raise ValueError(f"Cookbook recipe knobs must be a list: {recipe['id']}")
+        knob_keys: set[str] = set()
+        for knob in knobs:
+            if not isinstance(knob, dict):
+                raise ValueError(f"Cookbook recipe knob must be an object: {recipe['id']}")
+            required_knob = ("key", "label", "flag", "options", "default")
+            missing_knob = {key for key in required_knob if key not in knob}
+            if missing_knob:
+                raise ValueError(f"Cookbook recipe knob is missing: {recipe['id']}: {', '.join(sorted(missing_knob))}")
+            if knob["key"] in knob_keys:
+                raise ValueError(f"Duplicate cookbook recipe knob key: {recipe['id']}: {knob['key']}")
+            knob_keys.add(knob["key"])
+            if not isinstance(knob["flag"], str) or not knob["flag"].startswith("--"):
+                raise ValueError(f"Cookbook recipe knob flag must start with --: {recipe['id']}: {knob['key']}")
+            options = knob["options"]
+            if not isinstance(options, list) or not options:
+                raise ValueError(
+                    f"Cookbook recipe knob options must be a non-empty list: {recipe['id']}: {knob['key']}")
+            option_values = [option["value"] if isinstance(option, dict) else option for option in options]
+            if knob["default"] not in option_values:
+                raise ValueError(
+                    f"Cookbook recipe knob default must be one of its options: {recipe['id']}: {knob['key']}")
+
        source = (ROOT_DIR / recipe["source"]).resolve()
        if not any(source.is_relative_to(root.resolve()) for root in COOKBOOK_SOURCE_ROOTS):
            raise ValueError(f"Cookbook source is outside an approved directory: {recipe['source']}")
@@ -170,6 +202,19 @@ def validate_cookbook() -> None:
            raise ValueError(f"Cookbook source does not exist: {recipe['source']}")

        source_text = source.read_text(encoding="utf-8")
+        if knobs:
+            # A script that shares its argument parser with a sibling module
+            # (e.g. `from . import basic_fasth3`) inherits that module's
+            # flags without the flag text appearing in its own source.
+            knob_source_text = source_text
+            for match in re.finditer(r'(?:from\s+\.\s+import|^\s*import)\s+(\w+)', source_text, flags=re.MULTILINE):
+                sibling = source.parent / f"{match.group(1)}.py"
+                if sibling.is_file() and sibling != source:
+                    knob_source_text += "\n" + sibling.read_text(encoding="utf-8")
+            for knob in knobs:
+                if knob["flag"] not in knob_source_text:
+                    raise ValueError(f"Cookbook recipe knob flag is not in its source: {recipe['id']}: {knob['flag']}")
+
        # The model must be traceable to the checked-in source itself, or be
        # passed explicitly on the command line (e.g. --model-path <model> or
        # MODEL_PATH=<model>) when the source reads it from arguments/env.
@@ -223,12 +268,22 @@ def cookbook_serving_profile(recipe: dict) -> dict:
    if type(port) is not int or not 1 <= port <= 65535:
        raise ValueError(f"Serving config needs a valid port: {recipe['id']}")
    hardware = {"platform": runtime, "evidence": "source-configured"}
+    device = recipe["hardware"].get("device")
+    if device:
+        hardware["device"] = device
    if runtime == "cuda":
        count = generator["engine"]["num_gpus"]
        if type(count) is not int or count < 1:
            raise ValueError(f"Serving config needs a valid GPU count: {recipe['id']}")
        hardware["gpu_count"] = count
-        command = f"fastvideo serve --config {serving['source']} --server.host 127.0.0.1"
+        env = serving.get("env")
+        if env is not None:
+            if not isinstance(env, str) or not env.strip() or "\n" in env:
+                raise ValueError(f"Serving env must be a single-line command prefix: {recipe['id']}")
+            prefix = env.strip() + " "
+        else:
+            prefix = ""
+        command = f"{prefix}fastvideo serve --config {serving['source']} --server.host 127.0.0.1"
    else:
        if server.get("host") != "127.0.0.1":
            raise ValueError(f"MLX cookbook server must bind to loopback: {recipe['id']}")
@@ -4,9 +4,10 @@
 FastVideo supports the following hardware platforms:

 - [NVIDIA CUDA](installation/gpu.md)
- [NVIDIA DGX Spark / GB10 (ARM64 + CUDA 13)](installation/spark.md)
-  ([performance & tuning](installation/spark_performance.md))
- [Apple silicon](installation/mps.md)
+- **NVIDIA DGX Spark / GB10 (ARM64 + CUDA 13)** — [install](installation/spark.md),
+  [performance](installation/spark_performance.md),
+  [pair two Sparks](installation/spark_pair.md)
+- [Apple silicon (MLX)](installation/mlx.md)

 ## Quick Installation

@@ -14,7 +15,7 @@ FastVideo supports the following hardware platforms:

 Use uv as the default environment manager for faster and more stable installs.
 The commands below target NVIDIA CUDA 12; use `UV_TORCH_BACKEND=cu130` on
-CUDA 13. Apple silicon users should follow the [MPS guide](installation/mps.md).
+CUDA 13. Apple silicon users should follow the [MLX install guide](installation/mlx.md).

 ```bash
 # Create and activate a new uv environment
@@ -0,0 +1,65 @@
+# Install FastVideo with MLX
+
+Install FastVideo on a Mac, then generate from the cookbook. Local video uses
+the native MLX runtime, not CUDA, and not the old PyTorch MPS demo at
+`examples/inference/basic/basic_mps.py`.
+
+## Requirements
+
+- macOS 14 or newer
+- Python 3.12
+- `ffmpeg` (`brew install ffmpeg`)
+
+## Install
+
+Cookbook commands run from a clone.
+
+```bash
+git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
+uv venv --python 3.12 --seed
+source .venv/bin/activate
+brew install ffmpeg
+uv pip install -e ".[mlx]"
+```
+
+Conda is optional. After you activate a Conda env, still install with
+`uv pip` as above.
+
+`uv pip install "fastvideo[mlx]"` from PyPI installs the extra only. It does
+not ship the example scripts the cookbook copies.
+
+## Generate a video
+
+Open the cookbook. Select Apple Silicon as the runtime. Each recipe has a
+Python command. FastH3 also has a server path for the playground and the
+OpenAI Python client.
+
+- [Wan recipes](../../cookbook/wan.md) for FastMetal 1.3B, 5B, and 14B
+- [MiniMax H3 recipes](../../cookbook/minimax-h3.md) for FastH3 V1 and FastH3 V2
+- [H3 server guide](../../cookbook/openai-api.md) for the playground, cURL, and SDKs
+
+FastH3 is two distilled MiniMax-H3 checkpoints. V1 is the four-step 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.
+
+Recorded shapes and evidence live in the
+[support matrix](../../inference/support_matrix.md#apple-silicon-native-runtime).
+
+## Hardware
+
+- FastMetal 1.3B and 5B: 16 GB unified memory and up
+- FastMetal 14B: 36 GB unified memory and up
+- FastH3 V1 and V2: validated on an M4 Max with 36 GB unified memory
+
+## Troubleshooting
+
+- **`basic_mps.py` is the wrong path.** That script is PyTorch MPS. Use an
+  Apple Silicon recipe in the cookbook.
+- **Muxing fails.** Install `ffmpeg` with Homebrew.
+- **A cookbook command cannot find a script.** Run it from the FastVideo
+  clone after `uv pip install -e ".[mlx]"`.
+
+If that does not match what you see, open an issue on the
+[GitHub repository](https://github.com/hao-ai-lab/FastVideo) or ask in the
+[Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ).
@@ -1,260 +0,0 @@
-# MPS (Apple Silicon)
-
-Install FastVideo on Apple Silicon and run FastMetal-QAD or FastH3 Preview.
-
-Apple Silicon uses the MLX runtime. FastMetal-QAD ships ready-to-run MLX
-checkpoints; FastH3 Preview currently requires a local MLX DiT conversion.
-See the [FastMetal-QAD blog](https://haoailab.com/blogs/fastmetal/) and the
-[FastMetal collection](https://huggingface.co/collections/FastVideo/fastmetal).
-
-## Requirements
-
- **OS: macOS 14 or newer**
- **Python: 3.12.4**
-
-## Set up using Python
-
-### Create a new Python environment
-
-#### uv
-Recommended default: use [uv](https://docs.astral.sh/uv/) for faster and more stable environment setup.
-
-Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, create a new environment using:
-
-```console
-# (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools`.
-uv venv --python 3.12 --seed
-source .venv/bin/activate
-```
-
-#### Conda (alternative)
-
-You can also create a Python environment using [Conda](https://docs.conda.io/projects/conda/en/stable/user-guide/getting-started.html).
-
-##### 1. Install Miniconda (if not already installed)
-
-```bash
-wget https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh
-bash Miniconda3-latest-MacOSX-arm64.sh
-source ~/.zshrc
-```
-
-##### 2. Create and activate a Conda environment for FastVideo
-
-```bash
-conda create -n fastvideo python=3.12.4 -y
-conda activate fastvideo
-```
-
-### Dependencies
-
-```
-brew install ffmpeg
-```
-
-### Installation
-
-FastMetal's native Apple Silicon runtime requires the `mlx` extra.
-
-#### With uv (recommended)
-
-```bash
-uv pip install "fastvideo[mlx]"
-```
-
-#### With Conda environment (alternative)
-
-`uv` works inside an active conda env too, so prefer `uv pip` for the actual install:
-
-```bash
-uv pip install "fastvideo[mlx]"
-```
-
-### Installation from Source
-
-#### 1. Clone the FastVideo repository
-
-```bash
-git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
-```
-
-#### 2. Install FastVideo
-
-Basic installation:
-
-```bash
-uv pip install -e ".[mlx]"
-```
-
-Alternative with Conda environment:
-
-```bash
-uv pip install -e ".[mlx]"
-```
-
-## Run FastMetal-QAD
-
-Each release is self-contained. Download one checkpoint and point both
-`--model-root` and `--mlx-checkpoint` at it (the example also auto-detects
-`mlx_dit.json` under `--model-root`).
-
-| Checkpoint | Script | Mac tier |
-| --- | --- | --- |
-| [`FastVideo/FastMetal-1.3B-QAD`](https://huggingface.co/FastVideo/FastMetal-1.3B-QAD) | `mlx_wan_prompt_to_video.py` | 16 GB+ |
-| [`FastVideo/FastMetal-5B-QAD`](https://huggingface.co/FastVideo/FastMetal-5B-QAD) | `mlx_wan22_generate.py` | 16 GB+ |
-| [`FastVideo/FastMetal-14B-QAD`](https://huggingface.co/FastVideo/FastMetal-14B-QAD) | `mlx_wan_prompt_to_video.py` | 36 GB+ |
-
-```bash
-hf download FastVideo/FastMetal-1.3B-QAD --local-dir ./FastMetal-1.3B-QAD
-
-python examples/inference/basic/mlx_wan_prompt_to_video.py \
-  --model-root ./FastMetal-1.3B-QAD \
-  --mlx-checkpoint ./FastMetal-1.3B-QAD \
-  --height 480 --width 832 --num-frames 81 \
-  --prompt "A bird's-eye view of a misty forest valley at dawn."
-```
-
-14B uses the same script. Point both flags at `./FastMetal-14B-QAD`. That repo also ships an EMA variant: keep `--model-root` at the repo root and set `--mlx-checkpoint ./FastMetal-14B-QAD/ema`.
-
-Wan2.2 5B uses a different latent layout, so it has its own entrypoint:
-
-```bash
-hf download FastVideo/FastMetal-5B-QAD --local-dir ./FastMetal-5B-QAD
-
-python examples/inference/basic/mlx_wan22_generate.py \
-  --mlx-checkpoint ./FastMetal-5B-QAD \
-  --text-encoder-root ./FastMetal-5B-QAD \
-  --vae-root ./FastMetal-5B-QAD/vae \
-  --height 704 --width 1280 --num-frames 81 \
-  --prompt "A cinematic portrait with soft neon lighting and smooth camera motion."
-```
-
-CUDA FastWan-QAD (`FastVideo/FastWan-QAD-1.3B`, `FastVideo/FastWan-QAD-FP8-1.3B`) is a separate NVIDIA release. The MLX examples look for FastMetal packed weights (`mlx_dit.json`).
-
-`basic_mps.py` is a generic PyTorch MPS demo. For local video on Mac, use the FastMetal commands above.
-
-## Run FastH3 Preview
-
-FastH3 Preview uses the existing MLX runtime for text-to-video-with-audio
-(T2VA). The runtime streams the Qwen3-VL text conditioner, loads one
-heavyweight component at a time, denoises synchronized video and audio
-latents with a converted INT8, INT6, or INT4 DiT, and decodes both modalities
-with native MLX VAEs.
-
-Download the FastH3 snapshot, then convert one or more DiT formats:
-
-```bash
-hf download FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2 \
-  --local-dir ./FastH3-Preview-v0.2
-
-python scripts/checkpoint_conversion/convert_minimax_h3_mlx.py \
-  --model-root ./FastH3-Preview-v0.2/transformer \
-  --out ./FastH3-MLX \
-  --formats "int6"
-```
-
-Dense conversion drops the trained VSA gate projections. To keep them (INT6
-weight-only, same affine grid as the other linear matrices) write a **new**
-directory:
-
-```bash
-python scripts/checkpoint_conversion/convert_minimax_h3_mlx.py \
-  --model-root ./FastH3-Preview-v0.2/transformer \
-  --out ./FastH3-MLX-vsa \
-  --formats "int6" \
-  --include-vsa
-```
-
-Do not overwrite an existing dense export such as `./FastH3-MLX/int6`.
-
-Run the baseline path:
-
-```bash
-python examples/inference/basic/mlx_fasth3.py \
-  --model-root ./FastH3-Preview-v0.2 \
-  --mlx-checkpoint ./FastH3-MLX/int6 \
-  --prompt "(S1) A presenter says <d>[English] Fast H3 is amazing.</d>" \
-  --height 480 --width 832 --num-frames 124 --seed 2026 \
-  --output-path ./outputs/fasth3_int6.mp4
-```
-
-Add `--fast` for temporal fast mode. It denoises a shorter video sequence,
-uses MLX RIFE to restore the requested frame count, and keeps the audio
-sequence at full duration:
-
-```bash
-python examples/inference/basic/mlx_fasth3.py \
-  --model-root ./FastH3-Preview-v0.2 \
-  --mlx-checkpoint ./FastH3-MLX/int6 \
-  --prompt "(S1) A presenter says <d>[English] Fast H3 is even faster.</d>" \
-  --height 720 --width 1280 --num-frames 124 --seed 2027 \
-  --fast \
-  --output-path ./outputs/fasth3_int6_fast_720p.mp4
-```
-
-Add `--fast-spatial` for spatial fast mode, `--fast`'s spatial twin. It
-denoises and decodes on the smallest 32px-aligned canvas covering the
-requested size divided by `--fast-spatial-scale` (a 480x832 request runs on a
-256x416 canvas), then resamples the decoded frames up to the requested size
-in pixel space. It composes with `--fast`. This is a speed/quality trade-off
-and stays off by default: the output carries the reduced canvas's detail
-budget, so it reads softer than a native-resolution render, with the unsharp
-pass countering some but not all of the difference:
-
-```bash
-python examples/inference/basic/mlx_fasth3.py \
-  --model-root ./FastH3-Preview-v0.2 \
-  --mlx-checkpoint ./FastH3-MLX/int6 \
-  --prompt "(S1) A presenter says <d>[English] Fast H3 is fastest.</d>" \
-  --height 480 --width 832 --num-frames 124 --seed 2028 \
-  --fast --fast-spatial \
-  --output-path ./outputs/fasth3_int6_fast_spatial.mp4
-```
-
-VSA is off by default. A dense-only checkpoint (no `--include-vsa`) keeps the
-existing fused-SDPA path. After converting with `--include-vsa`, enable the
-sparse path explicitly:
-
-```bash
-python examples/inference/basic/mlx_fasth3.py \
-  --model-root ./FastH3-Preview-v0.2 \
-  --mlx-checkpoint ./FastH3-MLX-vsa/int6 \
-  --vsa --vsa-sparsity 0.9 --vsa-tile-size 64 --vsa-prefix-mode exempt \
-  --prompt "(S1) A presenter says <d>[English] Fast H3 is amazing.</d>" \
-  --height 720 --width 1280 --num-frames 124 --seed 2026 \
-  --output-path ./outputs/fasth3_int6_vsa_720p.mp4
-```
-
-`--vsa-impl auto` uses the chunked gather+SDPA **reference** path.
-`--vsa-impl simd` is an opt-in SIMD-group kernel (tile 64, head dim 128) that
-falls back to reference on unsupported shapes. It is not the default.
-`--vsa-impl reference` is the same as `auto`.
-
-!!! note "Current MLX scope"
-    This source runtime supports T2VA, temporal `--fast`, spatial
-    `--fast-spatial`, and opt-in VSA. FL2VA, Ref2VA, two-pass refinement, and
-    `VideoGenerator` registry dispatch are not wired yet. INT8/INT6/INT4
-    quantization is **weight-only**; VSA attention Q/K/V stay BF16. Old dense
-    MLX checkpoints remain valid for dense inference and raise a reconvert
-    error if `--vsa` is set. The checkpoint uses the MiniMax H3
-    Community License; review the model card before use or redistribution.
-
-## Development Environment Setup
-
-If you're planning to contribute to FastVideo please see the following page:
-[Contributor Guide](../../contributing/overview.md)
-
-## Hardware Requirements
-
- **1.3B / 5B:** 16 GB unified memory and up (M1 and later)
- **14B:** 36 GB unified memory and up
- **FastH3 Preview:** validated on an M4 Max with 36 GB unified memory; use one
-  converted DiT format at a time and leave substantial free disk space for the
-  source snapshot plus the converted checkpoint
- Fanless 13-inch MacBook Air can run 1.3B and 5B at the same resolutions
-
-## Troubleshooting
-
-If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).
-
-You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ) for additional support.
@@ -144,6 +144,12 @@ 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.

+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
+[MiniMax H3 cookbook](../../cookbook/minimax-h3.md): pick FastH3 V1,
+then NVIDIA DGX Spark, then 1 Spark or 2 Sparks.
+
 ## Development Environment Setup

 If you're planning to contribute to FastVideo please see the
@@ -0,0 +1,213 @@
+# Pair two NVIDIA DGX Sparks
+
+One GB10 is 128 GB of unified LPDDR5X. FastH3 still fits on a single Spark with
+[`lazy_module_load`](../../inference/offloading.md) (auto on GB10; sequential
+load stands down when lazy owns deferral). Two boxes connected
+by the QSFP ConnectX-7 cables can run **one clip faster** and can hold a
+**longer clip** (up to the FastH3 15 s cap).
+
+Copy-paste commands for both counts are on the
+[MiniMax H3 cookbook](../../cookbook/minimax-h3.md): FastH3 V1 → NVIDIA DGX
+Spark → 1 Spark or 2 Sparks.
+
+This is FastVideo sequence parallel (`sp_size=2`) over Ray, not a third-party
+xDiT vendor. Do not install xDiT for this path.
+
+## What two Sparks buy you
+
+| Goal | How | Use two Sparks? |
+|---|---|---|
+| Two independent videos at once | One process per box, `engine.num_gpus: 1` | Throughput only. Each clip still takes the 1-GPU time for that size. |
+| One clip, faster | Ray + `sp_size=2` + parallel VAE | **Yes.** One 768×1344×124 recipe was 292 s vs 374 s on one GB10. |
+| One clip, longer | Same, more frames | **Yes.** 345 frames (~14.4 s at 24 fps) finished in 587 s at 768×1344. |
+
+Sequence parallel **replicates** the DiT (~66 GiB per node). Lazy module load
+is still required on each box. FSDP would shard weights;
+it is untested on this fabric and is likely slower because every layer gathers
+over ~21 GB/s RoCE.
+
+## Requirements
+
+- Two DGX Sparks with FastVideo [installed](spark.md) (CUDA 13, `aarch64`).
+- The QSFP cables that ship with a dual-Spark kit, **ACTIVE** at 200 Gb/s:
+  `ibstat` should show the ConnectX-7 ports `LinkUp`.
+- The same FastH3 snapshot on **both** NVMes. Copy the Hugging Face cache over
+  QSFP; do not download 100+ GB twice over Wi-Fi.
+- Ray in the FastVideo venv (`uv pip install ray` if it is not already there).
+
+Each Spark has **one** GPU. `engine.num_gpus: 2` therefore means two nodes, which is
+why the executor must be Ray (`mp` only works inside one process tree).
+
+## 1. Put IPv4 on the QSFP NICs
+
+The RoCE links often come up with no IPv4. Wi-Fi (`192.168.1.x`) is fine for
+SSH and must stay the default route. NCCL and Ray must **not** use it.
+
+Pick a /24 that does not collide with your LAN. Example:
+
+| Node | QSFP IPv4 | Interface (typical) |
+|---|---|---|
+| Spark A | `192.168.23.1/24` | `enp1s0f1np1` |
+| Spark B (Ray head) | `192.168.23.2/24` | `enp1s0f1np1` |
+
+Confirm names with `ibdev2netdev` and `ip -br link`. Then, as root, on each
+box (NetworkManager likes to steal the NIC; unmanaged is enough for a session):
+
+```bash
+sudo nmcli device set enp1s0f1np1 managed no
+sudo ip addr replace 192.168.23.1/24 dev enp1s0f1np1   # .2 on the other box
+sudo ip link set enp1s0f1np1 mtu 9000 up
+```
+
+These addresses do **not** survive reboot. Ping across the cable before
+continuing: `ping -c 3 -I enp1s0f1np1 192.168.23.2`.
+
+A healthy fabric on this hardware looks like:
+
+- TCP iperf (jumbo 9000): ~40 Gb/s
+- NCCL allreduce 1 GiB × 10: ~21 GB/s busbw (NVIDIA's dual-Spark figure is ~21.7)
+
+## 2. Start a two-node Ray cluster on the cable
+
+On **both** nodes, from the FastVideo repo, with the venv active:
+
+```bash
+source examples/inference/optimizations/spark_pair_env.sh
+```
+
+That script pins NCCL and Gloo to the QSFP NIC/HCA, disables NVLink-style P2P
+(there is none between boxes), and turns off Ray's memory monitor. The monitor
+treats GB10 unified RSS during a 14-shard DiT load as a runaway and SIGTERMs
+the worker around shard 11/14. Override `NCCL_SOCKET_IFNAME` /
+`GLOO_SOCKET_IFNAME` if `ibdev2netdev` shows a different name.
+
+Cap Ray's object store. The default (~30% of 128 GB) leaves too little room
+for the DiT:
+
+```bash
+# Spark B — head
+export FASTVIDEO_HOST_IP=192.168.23.2
+ray start --head --node-ip-address=192.168.23.2 --port=6379 --num-gpus=1 \
+  --disable-usage-stats --object-store-memory=2147483648 --memory=4294967296
+
+# Spark A — worker
+export FASTVIDEO_HOST_IP=192.168.23.1
+ray start --address=192.168.23.2:6379 --node-ip-address=192.168.23.1 --num-gpus=1 \
+  --disable-usage-stats --object-store-memory=2147483648 --memory=4294967296
+```
+
+`FASTVIDEO_HOST_IP` **must** match `--node-ip-address`. If you omit it, Ray
+advertises the Wi-Fi address, FastVideo builds a placement group for
+`node:192.168.1.x`, and the QSFP workers never match.
+
+Check `ray status` on the head: `0.0/2.0 GPU` idle.
+
+## 3. Generate one FastH3 clip on both GPUs
+
+Run the driver on the **head**, same venv, same QSFP IP.
+
+`basic_fasth3.py` defaults target a four-GPU GB200 profile: 768×1344, `sm100a`
+VSA, FA4, four GPUs. On Sparks you must override the kernel flags. Height,
+width, frames, steps, seed, and prompt are yours. Change them. Legal
+`num_frames` values are `17n+5`, capped at 362.
+
+GB10 has no FA4 / sm_100a VSA kernel, so `--vsa-kernel triton --no-fa4` stays
+required on this box. `--execution-backend ray` is optional when `RAY_ADDRESS`
+is already set.
+
+`--warmup --repeats 3` prints a median of three `generate()` calls after an
+excluded warmup. Sequential load reloads Qwen for each later request, so that
+protocol works. For a single cold process, pass `--no-warmup --repeats 1`.
+
+The command below is one example, not a required recipe:
+
+```bash
+source examples/inference/optimizations/spark_pair_env.sh
+export RAY_ADDRESS=192.168.23.2:6379
+export FASTVIDEO_HOST_IP=192.168.23.2
+
+python examples/inference/basic/basic_fasth3.py \
+  --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \
+  --num-gpus 2 --execution-backend ray \
+  --vsa-kernel triton --no-fa4 \
+  --warmup --repeats 3 --parallel-vae \
+  --height 768 --width 1344 --num-frames 124 --steps 5 \
+  --seed 2026 \
+  --prompt "A wide cinematic shot of an alpine meadow at sunrise, pale pink mountain peaks above a blue valley filled with thin morning mist." \
+  --output outputs/fasth3_spark_pair
+```
+
+Config-first equivalent. Edit the YAML the same way, `request.sampling` is not
+locked:
+
+```bash
+FASTVIDEO_VSA_SM100A=0 FASTVIDEO_FA4=0 \
+FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 \
+FASTVIDEO_VAE_PARALLEL_DECODE=1 FASTVIDEO_STAGE_LOGGING=1 \
+fastvideo generate --config examples/inference/basic/basic_fasth3_spark_pair.yaml
+```
+
+Stop the cluster when you are done: `ray stop` on both nodes.
+
+## FastH3 frame counts
+
+H3 is 24 fps. Legal `num_frames` values are `17n+5`. The pipeline rejects
+clips longer than **15 s**. The longest legal length is **362 frames**
+(15.083 s). 360 frames aligns to 362 and is accepted.
+
+## Measured on two GB10s (2026-08-31)
+
+These rows are full H3 VAE decode, Triton VSA, sequential + lazy load (auto on
+GB10), parallel VAE. They are not a required size. Denoise times include
+deferred DiT load (~35 s on the first generate).
+
+Cold process, `--no-warmup --repeats 1`, alpine prompt, 768×1344, 5 sigma
+points (4 DiT forwards):
+
+| Run | GPUs | Frames | E2E | Denoise | VAE decode |
+|---|---:|---:|---:|---:|---:|
+| One Spark | 1 | 124 | 374–393 s | 180–188 s | 151–156 s |
+| Two Sparks, SP=2 | 2 | 124 | **292 s** | **122 s** | **102 s** |
+| Two Sparks, SP=2 | 2 | 345 | **587 s** | **351 s** | **173 s** |
+
+Warmup excluded, `--warmup --repeats 3` median, 512×896, 5 sigma points, full
+VAE, same 4-step schedule:
+
+| Run | GPUs | Frames | Median E2E | Median denoise |
+|---|---:|---:|---:|---:|
+| One Spark | 1 | 124 | **251.4 s** | 94.2 s |
+| Two Sparks, SP=2 | 2 | 124 | **215.2 s** | 72.4 s |
+
+Those medians used `--height` / `--width` / `--num-frames` as CLI flags. Swap
+them. Native 480p on this model is 480×832, 124 frames. The 15 s cap is 362
+frames.
+
+The first VAE decode still pays `torch.compile`. Later `generate()` calls in
+the same workers are cheaper. GB10 regional DiT compile stays off because the
+sm_100a VSA kernel is not on this chip, so denoise is slower than a GB200
+`sm100a` run at the same geometry.
+
+## Troubleshooting
+
+| Symptom | Fix |
+|---|---|
+| Placement group waits forever / `node:192.168.1.x` | Set `FASTVIDEO_HOST_IP` to the QSFP address on **every** `ray start` **and** on the driver. |
+| `RayDistributedExecutor` TypeError / abstract `set_log_queue` | Use a FastVideo build that implements those methods on the Ray executor (this page). |
+| Worker SIGTERM during DiT shard 11/14 | `RAY_memory_monitor_refresh_ms=0` **before** `ray start`. Do not leave Ray's default 30% object store. |
+| NCCL hangs or uses Wi-Fi | `source spark_pair_env.sh`. Confirm `NCCL_SOCKET_IFNAME` is the QSFP NIC. |
+| Gloo `connectFullMesh` / `remote=[127.0.0.1]` | Two 1-GPU nodes must not use loopback as the Gloo store. Source `spark_pair_env.sh` so `GLOO_SOCKET_IFNAME` is the QSFP NIC on **each** box. FastVideo no longer copies that NIC name from the driver onto workers. |
+| Second `generate()` crashes `NoneType.parameters` | Sequential load used to drop the text encoder without reloading it. This branch reloads Qwen for later requests so `--warmup --repeats N` works. |
+| OOM / `earlyoom` prefers Python | Lazy module load must stay on (do not pass `--no-lazy-module-load` to `basic_fasth3.py` or set `engine.offload.lazy_module_load: false`). Peak GPU during 345-frame denoise is ~90 GiB/node. |
+| `engine.num_gpus: 2` on one Spark | Each Spark has one GPU. Use Ray across two nodes, or `engine.num_gpus: 1` on one box. |
+
+## What we are not claiming
+
+- **Throughput of many clips.** Two independent 1-GPU jobs still win if you
+  want two videos, not one faster video.
+- **xDiT PipeFusion / CFG-parallel.** FastH3 is 4-step and has no CFG.
+- **FSDP or tensor parallel as a speedup** on this 21 GB/s link.
+- **Persistent networking.** The example IPs are session `ip addr replace`.
+
+More GPUs are legal while `num_attention_heads` (56 on FastH3) is divisible by
+`sp_size`. Four Sparks would need a four-node fabric that this bring-up did
+not exercise.
@@ -62,10 +62,12 @@ nothing to set. If you run a model that still defaults to an fp32 decode, set th
 decode-only override yourself:

 ```python
-from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo import VideoGenerator

-pipeline_config = PipelineConfig.from_pretrained(model_id)
-pipeline_config.vae_decode_precision = "bf16"   # decode-only; leaves encode precision alone
+generator = VideoGenerator.from_config({
+    "model_path": model_id,
+    "engine": {"precision": {"vae_decode": "bf16"}},  # decode-only; leaves encode precision alone
+})
 ```

 Decode is output-only, so lowering its precision is safe. (Encode seeds the
@@ -161,15 +163,30 @@ 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.
- **MiniMax H3 / FastH3** still needs sequential loading on one GB10. The Qwen3-VL
+- **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). `h3_sequential_load` defaults to auto and turns this split on for
-  unified-memory devices. Do not pass `--no-h3-sequential-load` here. Force
-  `--h3-sequential-load` only if auto-detect misses the device. The CUDA pipeline
-  encodes first, releases the encoder, then loads DiT and VAEs onto the
-  accelerator (`to_cpu` follows `cpu_offload`, which is off here). See
-  [Offloading](../../inference/offloading.md).
+  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; do not set
+  `engine.offload.lazy_module_load` to false here (`--no-lazy-module-load` in
+  `basic_fasth3.py` and `basic_minimax_h3_t2v.py`). 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
+  **68 s** for the full VAE, and one T2VA generation finished in **224 s**
+  end-to-end. Reconstruction is approximate, not lossless. FL2VA/Ref2VA still
+  need the full VAE to encode references.
+- **Two Sparks, one clip.** Sequence parallel (`sp_size=2`) over the QSFP RoCE
+  link ran one 768×1344×124 FastH3 recipe in **292 s** vs **374–393 s** on
+  one GB10, and a 345-frame (~14.4 s) clip in **587 s**. Other heights, widths,
+  and frame counts are valid. Weights stay replicated, so lazy module load
+  (auto on GB10) is still required on each box. Bring-up and knobs:
+  [Pair two NVIDIA DGX Sparks](spark_pair.md).
+- A worker's SIGTERM log and traceback show where it was interrupted, not why it
+  was selected; confirm the cause in the `earlyoom` service or system logs. A
+  later SIGKILL or kernel OOM kill cannot be caught and reported by Python.

 ## Gotchas specific to the GB10

@@ -187,10 +204,11 @@ A few things that surprise people on this box (beyond the memory notes above):
  build recent enough to include its `transformers`-compatibility handling before
  running it.
 - **MiniMax H3 worker init can look healthy and still die on the first generate**
-  if sequential load is off (`--no-h3-sequential-load`, or auto-off on a
-  misclassified device) and encoder, VAE, and DiT load together. Confirm the log
-  contains `Released MiniMax-H3 text encoder after conditioning` before
-  `Loading MiniMax-H3 denoise modules`.
+  if deferred loading is off (`engine.offload.lazy_module_load: false` and sequential also off)
+  and encoder, VAE, and DiT load together. On GB10 the log should show
+  `lazy_module_load owns deferral` (or, if lazy is off, sequential
+  `Released MiniMax-H3 text encoder after conditioning` before
+  `Loading MiniMax-H3 denoise modules`).

 ## Reproduce these numbers

@@ -210,7 +210,7 @@ from fastvideo.pipelines.stages import (
    InputValidationStage, CLIPTextEncodingStage, TimestepPreparationStage,
    LatentPreparationStage, DenoisingStage, DecodingStage
 )
-from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.api.resolution import ResolvedGeneratorConfig
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 import torch

@@ -226,11 +226,11 @@ class MyCustomPipeline(ComposedPipelineBase):
    def required_config_modules(self) -> List[str]:
        return self._required_config_modules
        
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+    def initialize_pipeline(self, resolved_config: ResolvedGeneratorConfig):
        """Initialize pipeline-specific components."""
        pass
        
-    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+    def create_pipeline_stages(self, resolved_config: ResolvedGeneratorConfig):
        """Set up pipeline stages with proper dependency injection."""
        self.add_stage(
            stage_name="input_validation_stage",
@@ -294,7 +294,7 @@ class MyCustomStage(PipelineStage):
        self.custom_module = custom_module
        self.other_param = other_param
        
-    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+    def forward(self, batch: ForwardBatch, resolved_config: ResolvedGeneratorConfig) -> ForwardBatch:
        # Access input data
        input_data = batch.some_attribute
        
@@ -23,7 +23,7 @@ to FastVideo model classes. Two discovery mechanisms:
   `fastvideo/models/` and parses each `.py` file's AST looking for an
   `EntryClass` variable assignment. Discovered models take priority over
   hardcoded entries. For example,
-   `fastvideo/models/dits/wanvideo.py` exports
+   `fastvideo/models/wan/transformer.py` exports
   `EntryClass = WanTransformer3DModel`.

 Both feed into a unified `_FAST_VIDEO_MODELS` dict, which populates the
@@ -86,7 +86,7 @@ from `model_index.json`.

 ```
 PipelineConfig                    (fastvideo/configs/pipelines/base.py)
-├── WanT2V480PConfig              (fastvideo/configs/pipelines/wan.py)
+├── WanT2V480PConfig              (fastvideo/models/wan/pipeline_config.py)
 │   ├── WanT2V720PConfig
 │   └── WanI2V480PConfig
 ├── HunyuanConfig                 (fastvideo/configs/pipelines/hunyuan.py)
@@ -103,10 +103,19 @@ PipelineConfig                    (fastvideo/configs/pipelines/base.py)
 - Precision settings: `dit_precision`, `vae_precision`,
  `text_encoder_precisions`.

+These generation and precision attributes hold the model defaults. Resolution
+copies them into the typed fields of the resolved config (`pipeline.flow_shift`,
+`engine.precision.dit`, ...), and runtime code reads the typed fields.
+
 Model-specific subclasses override defaults. For example,
 `WanT2V480PConfig` sets `flow_shift=3.0` and uses `WanVideoConfig` as
 its DiT config.

+Wan's `models/wan/definition.py` links each registered variant to its pipeline
+config and sampling preset. The shared registry consumes these definitions
+without changing detector precedence or checkpoint/override-based pipeline
+selection. `configs/pipelines/wan.py` remains a compatibility import.
+
 ### ModelConfig / ArchConfig (`fastvideo/configs/models/base.py`)

 `ModelConfig` wraps an `ArchConfig` using `__getattr__` proxy — attribute
@@ -122,9 +131,11 @@ Concrete hierarchy: `DiTConfig` → `DiTArchConfig`, `VAEConfig` →

 - `PipelineConfig.from_pretrained(model_path)` — resolves config class
  via `get_pipeline_config_cls_from_name()`, instantiates with defaults.
- `PipelineConfig.from_kwargs(kwargs)` — resolves class, optionally loads
-  JSON via `load_from_json()`, then applies CLI overrides via
-  `update_config_from_dict()`.
+- `PipelineConfig.from_source(model_path, source)` — resolves the registry
+  class of `model_path`, then updates it from `source`: a JSON path loaded
+  via `load_from_json()`, a mapping of field values applied via
+  `update_pipeline_config()`, or a `PipelineConfig` that replaces the
+  registry instance.
 - `dump_to_json()` / `load_from_json()` — JSON persistence. Callable
  fields and `arch_config` are excluded from dumps.

@@ -142,7 +153,7 @@ sp = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")

 ### ComponentLoader (`fastvideo/models/loader/component_loader.py`)

-Abstract base with a `load(model_path, fastvideo_args)` method.
+Abstract base with a `load(model_path, resolved_config)` method.
 `ComponentLoader.for_module_type(module_type, library)` is a factory
 that dispatches to specialized loaders via a `module_loaders` dict:

@@ -162,7 +173,7 @@ that dispatches to specialized loaders via a `module_loaders` dict:
 `TransformerLoader` reads `config.json` from the component directory,
 resolves the class via `ModelRegistry.resolve_model_cls()`, instantiates
 the model, and loads safetensors weights. CPU offload and layerwise
-offload are applied based on `FastVideoArgs`.
+offload are applied based on `resolved_config.engine.offload`.

 Unknown module types fall back to `GenericComponentLoader`.

@@ -200,14 +211,14 @@ loading by calling `ComponentLoader.for_module_type()` then `.load()`.

 Abstract base class using the Template Method pattern:

- `__call__(batch, fastvideo_args)` — orchestrates verification, timing,
+- `__call__(batch, resolved_config)` — orchestrates verification, timing,
  and error handling. Not overridden by subclasses.
- `forward(batch, fastvideo_args) -> ForwardBatch` — abstract, contains
+- `forward(batch, resolved_config) -> ForwardBatch` — abstract, contains
  the stage logic.
 - `verify_input()` / `verify_output()` — optional hooks returning
  `VerificationResult`. Default: no checks.

-When `fastvideo_args.enable_stage_verification` is `True`, `__call__`
+When `resolved_config.engine.enable_stage_verification` is `True`, `__call__`
 runs input verification before `forward()` and output verification after.
 When `envs.FASTVIDEO_STAGE_LOGGING` is set, execution time is measured
 with `torch.cuda.synchronize()` and logged.
@@ -256,6 +267,16 @@ Specialized variants: `CausalDenoisingStage`, `LTX2DenoisingStage`,
 `SRDenoisingStage`, `LTX2AudioDecodingStage`, `SD35ConditioningStage`,
 `LTX2TextEncodingStage`, `LTX2LatentPreparationStage`.

+Wan owns its sampling recipes under `basic/wan/stages/`. `WanDenoisingStage`
+specializes input packing, expert selection, timesteps, and first-frame
+restoration around the shared dense loop. `WanFirstFrameEncodingStage`
+produces normalized `ForwardBatch.first_frame_latent` before sampling; the
+sampler no longer executes a VAE. Dense DMD and the two causal samplers have
+family-local implementations and explicit scheduler ownership. Standard and
+DMD causal sampling share cache allocation, not their sampling algorithm.
+Legacy imports from `stages/` remain compatibility aliases. Sampling invariants
+also live beside the code in `fastvideo/pipelines/basic/wan/AGENTS.md`.
+
 ### Verification System (`fastvideo/pipelines/stages/validators.py`)

 `StageValidators` (aliased as `V`) provides static validators:
@@ -280,7 +301,7 @@ provides detailed error messages. Failed verification raises

 Abstract base for all inference pipelines. Lifecycle:

-1. **`__init__(model_path, fastvideo_args)`** — initializes distributed
+1. **`__init__(model_path, resolved_config)`** — initializes distributed
   environment via `maybe_init_distributed_environment_and_model_parallel
   (tp_size, sp_size)`, then calls `load_modules()` to populate
   `self.modules`.
@@ -288,7 +309,7 @@ Abstract base for all inference pipelines. Lifecycle:
   setup), `create_pipeline_stages()` (abstract — subclasses wire stages),
   optionally applies `torch.compile` to transformers, and calls
   `warmup_sequence_parallel_communication()`.
-3. **`forward(batch, fastvideo_args)`** — iterates `self.stages` calling
+3. **`forward(batch, resolved_config)`** — iterates `self.stages` calling
   each stage in order. Decorated with `@torch.no_grad()`.

 Key class attributes:
@@ -301,8 +322,9 @@ Key methods:
 - `add_stage(name, stage)` — appends to `_stages` list and
  `_stage_name_mapping` dict, also sets attribute on `self`.
 - `get_module(name, default)` — retrieves a loaded module.
- `from_pretrained(model_path, **kwargs)` — class method constructing
-  `FastVideoArgs` and calling `cls(...)` then `post_init()`.
+- `from_pretrained(model_path, *, resolved_config)` — class method that
+  builds the pipeline from a resolved config (from
+  `resolve_inference_config({...})`) by calling `cls(...)` then `post_init()`.

 ### LoRAPipeline (`fastvideo/pipelines/lora_pipeline.py`)

@@ -337,12 +359,12 @@ Key APIs: `get_tp_rank()`, `get_tp_world_size()`, `get_sp_rank()`,
 `warmup_sequence_parallel_communication()` pre-warms NCCL communicators
 to avoid slow first forward passes.

-Usage: `torchrun --nproc-per-node=N -m fastvideo.entrypoints.cli.main
-generate --model-path ... --tp-size N --sp-size M`.
+Usage: `fastvideo generate --config run.yaml
+--generator.engine.parallelism.tp_size N --generator.engine.parallelism.sp_size M`.

 ### torch.compile Integration

-When `fastvideo_args.enable_torch_compile` is `True`,
+When `resolved_config.engine.compile.enabled` is `True`,
 `_maybe_compile_pipeline_module()` checks for a `_compile_conditions`
 attribute on the module. If present, only matching submodules are
 compiled. Otherwise, the entire module is compiled. FSDP-wrapped
@@ -354,47 +376,53 @@ modules are skipped.

 ```python
 generator = VideoGenerator.from_pretrained(
-    model_path="Wan-AI/Wan2.1-T2V-14B-Diffusers",
-    num_gpus=1, tp_size=1, sp_size=1,
-)
-result = generator.generate_video(
-    prompt="A cat dancing",
-    height=720, width=1280, num_frames=81,
+    "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+    {"engine": {"num_gpus": 1, "parallelism": {"tp_size": 1, "sp_size": 1}}},
 )
+result = generator.generate({
+    "prompt": "A cat dancing",
+    "sampling": {"height": 720, "width": 1280, "num_frames": 81},
+})
 ```

 **CLI** (`fastvideo/entrypoints/cli/`):

 ```bash
+# run.yaml holds `generator: {model_path: Wan-AI/Wan2.1-T2V-14B-Diffusers}`.
 fastvideo generate \
-    --model-path "Wan-AI/Wan2.1-T2V-14B-Diffusers" \
-    --prompt "A cat dancing" \
-    --num-gpus 1
+    --config run.yaml \
+    --request.prompt "A cat dancing" \
+    --generator.engine.num_gpus 1
 ```

-**FastVideoArgs** (`fastvideo/fastvideo_args.py`): Central args dataclass.
-Key fields: `model_path`, `mode` (`ExecutionMode`), `workload_type`
-(`WorkloadType`), `pipeline_config` (`PipelineConfig`), `num_gpus`,
-`tp_size`, `sp_size`, `lora_path`, `dit_cpu_offload`,
-`dit_layerwise_offload`, `enable_torch_compile`,
-`enable_stage_verification`.
+**ResolvedGeneratorConfig** (`fastvideo/api/resolution.py`): The frozen
+runtime config that the executor, workers, pipelines, stages, and loaders
+read. Key paths: `model_path`, `mode` (`ExecutionMode`),
+`pipeline.workload_type` (`WorkloadType`), `engine.num_gpus`,
+`engine.parallelism.tp_size`, `engine.parallelism.sp_size`,
+`pipeline.components.lora_path`, `engine.offload.dit`,
+`engine.offload.dit_layerwise`, `engine.compile.enabled`,
+`engine.enable_stage_verification`, and `pipeline_config` (the frozen
+`PipelineConfig`).

-Constructed via `FastVideoArgs.from_kwargs(**kwargs)` which resolves the
-`PipelineConfig` from the registry, applies JSON config if provided, and
-merges CLI overrides.
+Built by `resolve_inference_config(config)`
+(`fastvideo/api/inference_resolution.py`), which runs the named resolution
+steps (environment variables, model defaults, derived values, validation) in
+order, records each decision, and then builds the `PipelineConfig` from the
+registry, applies a JSON config if provided, and freezes it.

 ## End-to-End Inference Flow

 ```
-User: VideoGenerator.from_pretrained(model_path, **kwargs)
+User: VideoGenerator.from_pretrained(model_path, config)
  │
-  ├─ FastVideoArgs.from_kwargs() → PipelineConfig resolved via registry
+  ├─ resolve_inference_config() → PipelineConfig resolved via registry
  ├─ get_model_info() → ModelInfo(pipeline_cls, sampling_param_cls, ...)
  │   ├─ model_index.json read → _class_name extracted
  │   ├─ pipeline_registry resolves pipeline_cls from _class_name
  │   └─ config_registry resolves config classes from model_path
  │
-  ├─ pipeline_cls.__init__(model_path, fastvideo_args)
+  ├─ pipeline_cls.__init__(model_path, resolved_config)
  │   ├─ maybe_init_distributed(tp_size, sp_size)
  │   └─ load_modules() → reads model_index.json, loads each component
  │       ├─ ComponentLoader.for_module_type() → specialized loader
@@ -406,10 +434,10 @@ User: VideoGenerator.from_pretrained(model_path, **kwargs)
      ├─ torch.compile (if enabled)
      └─ warmup_sequence_parallel_communication()

-User: generator.generate_video(prompt, ...)
+User: generator.generate(request)
  │
  ├─ ForwardBatch constructed from SamplingParam + user args
-  └─ pipeline.forward(batch, fastvideo_args)
+  └─ pipeline.forward(batch, resolved_config)
      ├─ InputValidationStage → validates dims
      ├─ TextEncodingStage → prompt → embeddings
      ├─ ConditioningStage → prepares conditioning
@@ -8,64 +8,75 @@ FastVideo automatically distributes the generation process when multiple GPUs ar
 # Will use 4 GPUs in parallel for faster generation
 generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    num_gpus=4,
+    {"engine": {"num_gpus": 4}},
 )
 ```

+One node uses the multiprocessing executor (`execution_backend: mp`, the
+default). Two machines — for example two DGX Sparks, one GPU each — need Ray:
+
+```yaml
+generator:
+  engine:
+    num_gpus: 2
+    execution_backend: ray
+    parallelism:
+      sp_size: 2
+```
+
+Set `RAY_ADDRESS` and `FASTVIDEO_HOST_IP` to the interconnect IPs, not Wi-Fi.
+The FastH3 example selects Ray automatically when `RAY_ADDRESS` is set. Full
+bring-up: [Pair two NVIDIA DGX Sparks](../getting_started/installation/spark_pair.md).
+
 ## Customizing Generation

- `PipelineConfig`: Initialization time parameters
- `SamplingParam`: Generation time parameters
-
-You can customize generation behavior using `PipelineConfig` and
-`SamplingParam`:
+`VideoGenerator.from_pretrained(model_path, config)` takes the startup
+settings as a nested mapping at their typed config paths, such as
+`{"engine": {"num_gpus": 2, "offload": {"dit": False}}}`; it is
+`VideoGenerator.from_config` with `model_path` added to the mapping. Pass
+generation settings to `VideoGenerator.generate` as a request:

 ```python
-from fastvideo import VideoGenerator, SamplingParam, PipelineConfig
+from fastvideo import VideoGenerator

 def main():
    model_name = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-    config = PipelineConfig.from_pretrained(model_name)
-    config.vae_precision = "fp16"

    # Create the generator
-    generator = VideoGenerator.from_pretrained(
-        model_name,
-        num_gpus=1,
-        dit_layerwise_offload=True,  # FastVideoArgs option
-        pipeline_config=config
-    )
-
-    # Create and customize sampling parameters
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
-    
-    # How many frames to generate
-    sampling_param.num_frames = 45
-    
-    # Video resolution (width, height)
-    sampling_param.width = 1024
-    sampling_param.height = 576
-    
-    # How many steps we denoise the video (higher = better quality, slower generation)
-    sampling_param.num_inference_steps = 30
-    
-    # How strongly the video conforms to the prompt (higher = more faithful to prompt)
-    sampling_param.guidance_scale = 7.5
-    
-    # Random seed for reproducibility
-    sampling_param.seed = 42  # Optional, leave unset for random results
+    generator = VideoGenerator.from_config({
+        "model_path": model_name,
+        "engine": {
+            "num_gpus": 1,
+            "offload": {"dit_layerwise": True},
+            "precision": {"vae": "fp16"},
+        },
+    })

    # Generate video with custom parameters
    prompt = "A beautiful sunset over a calm ocean, with gentle waves."
-    video = generator.generate_video(
-        prompt, 
-        sampling_param=sampling_param, 
-        output_path="my_videos/",  # Controls where videos are saved
-        save_video=True
-    )
+    video = generator.generate({
+        "prompt": prompt,
+        "sampling": {
+            # How many frames to generate
+            "num_frames": 45,
+            # Video resolution (width, height)
+            "width": 1024,
+            "height": 576,
+            # How many steps we denoise the video (higher = better quality, slower generation)
+            "num_inference_steps": 30,
+            # How strongly the video conforms to the prompt (higher = more faithful to prompt)
+            "guidance_scale": 7.5,
+            # Random seed for reproducibility
+            "seed": 42,  # Optional, leave unset for random results
+        },
+        "output": {
+            "output_path": "my_videos/",  # Controls where videos are saved
+            "save_video": True,
+        },
+    })

-    # If return_frames=True, frames are available in video["frames"]
-    print(f"Generated {len(video['frames'])} frames")
+    # If return_frames=True, frames are available in video.frames
+    print(f"Generated {len(video.frames)} frames")

 if __name__ == '__main__':
    main()
@@ -109,6 +120,26 @@ Override individual values from the CLI with dotted paths:
 fastvideo generate --config config.yaml --request.sampling.seed 42
 ```

+## Where a Value Came From
+
+FastVideo resolves the generator config once at startup and records the source of every value: the input config
+(`input`; `explicit` tells whether you wrote the value or it is the schema default), a `FASTVIDEO_*` environment
+variable, the model's defaults, or a derived value. A worker's device policy and values read from checkpoint files
+are recorded too.
+
+```python
+generator = VideoGenerator.from_config(config)
+generator.resolved_config.provenance("engine.parallelism.sp_size")
+# PathProvenance(path='engine.parallelism.sp_size', value=2, source='derive_parallel_sizes', ...)
+
+result = generator.generate(request)
+result.resolved_request.provenance("sampling.num_frames")
+# PathProvenance(..., value=81, source='fill_sampling_defaults[preset wan_t2v_1_3b]', explicit=False)
+```
+
+`resolved_config.provenance_table()` lists every path. Every value is decided before resolution ends, including the
+device offload policy and the checkpoint defaults; after that, `resolved_config` is read-only.
+
 ## Performance Optimization

 For configuring optimizations, please see our [optimizations guide](optimizations.md)
@@ -0,0 +1,90 @@
+# FastH3 distilled checkpoint schedules
+
+Base MiniMax-H3 still uses the scheduler shifts in its checkpoint (video 12,
+audio 3), BF16 text encoding, and the existing uniform schedule. The default
+`basic_fasth3.py` example still targets the four-forward preview. Selecting a
+shift-10 eight-forward checkpoint is an explicit choice of model and recipe;
+it does not change either default or enable NVFP4.
+
+## Eight-forward T2AV recipe
+
+The public checkpoint is
+[`FastVideo/FastVideo-FastH3-8-Step-V2`](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2)
+(MiniMax H3 Community License), trained with video/audio shifts 10/3, VSA
+sparsity 0.8, 64-token tiles, and the DMD rungs
+`[999, 874, 749, 624, 500, 375, 250, 125]`. `basic_fasth3_8step.py` pins that
+checkpoint and recipe as defaults; it shares the preview example's CLI, so every
+other flag works unchanged:
+
+```bash
+python examples/inference/basic/basic_fasth3_8step.py \
+  --prompt 'A slow cinematic drone shot glides over a coastal town; gulls call over the harbor.' \
+  --num-gpus 4 --vsa-kernel sm100a \
+  --profile strict --no-inference-torch-compile --no-compile-vae \
+  --height 768 --width 1344 --num-frames 124 \
+  --output outputs/fasth3-8step
+```
+
+Pass `--model-path` to use a local snapshot of the full export (not just its
+`transformer` subdirectory). `--steps` is the number of sigma-grid points,
+including the terminal zero; nine points run exactly eight transformer
+forwards, and the script rejects any other value because the checkpoint's
+ladder has eight rungs. The rungs are unshifted noise levels on the 1000-step training
+clock; each scheduler applies its own shift once, and the transformer receives
+H3 clean-time values (`1 - sigma`). A uniform nine-point grid is not a substitute
+for those rungs.
+
+Compilation and H3 fusions are disabled above to establish an eager reference;
+they can be evaluated separately. On hardware without the sm100a extension,
+use `--vsa-kernel triton`; compare outputs and performance before adopting that
+backend. This recipe is T2AV-only, not a distilled `transformer_ref` model.
+
+## Export metadata and validation
+
+The export's `fastvideo_inference.json` supplies the trained ladder. The
+schedule fields of `fasth3-inference-contract-v1` are:
+
+```json
+{
+  "schema_version": "fasth3-inference-contract-v1",
+  "dmd_denoising_steps": [999, 874, 749, 624, 500, 375, 250, 125],
+  "num_inference_steps": 9,
+  "transformer_forwards": 8,
+  "video_scheduler_shift": 10.0,
+  "audio_scheduler_shift": 3.0
+}
+```
+
+The loader keeps this file when downloading the selected H3 components from
+Hugging Face. It checks that the two declared shifts agree with
+`scheduler/scheduler_config.json` and `audio_scheduler/scheduler_config.json`.
+Missing/invalid rungs, inconsistent counts, or an explicit conflicting ladder
+are errors. The denoiser rejects a request with the wrong number of grid points.
+The metadata does not silently change request dimensions, step count, attention
+backend, sparsity, precision, or offload/compile settings: set those explicitly
+as above.
+
+For exports without this sidecar, an explicit ladder is supported via
+`MiniMaxH3PipelineConfig.dmd_denoising_steps`, or through the typed API's
+`PipelineSelection(dmd_denoising_steps=[...])`. The shifts
+still come from the checkpoint scheduler configs. Keep generic `flow_shift`
+unset: H3 has separate video and audio shifts, not one shared shift.
+
+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.
+
+## Apple Silicon
+
+`mlx_fasth3.py` stays on FastH3 V1 and its uniform AdaLN cache.
+`mlx_fasth3_8step.py` is the eight-forward MLX recipe for FastH3 V2. It
+reads the same `fastvideo_inference.json` rungs and shifts, and it expects an
+MLX DiT whose AdaLN cache was converted from that contract. Reuse the V1
+snapshot's VAE, audio VAE, text encoder, and tokenizer; only the DiT and the
+sidecar change. Rank-reduced AdaLN checkpoints are unchanged and are not
+produced by the MLX converter.
+
+Install is in the
+[MLX install guide](../getting_started/installation/mlx.md). Generation and
+serving are in the [MiniMax H3 cookbook](../cookbook/minimax-h3.md) and the
+[H3 server guide](../cookbook/openai-api.md).
@@ -5,10 +5,10 @@ This page contains step-by-step instructions to get you quickly started with vid
 ## Requirements

 - **OS**: Linux (tested on Ubuntu 22.04+), or macOS on Apple silicon via the
-  [MPS installation guide](../getting_started/installation/mps.md)
+  [MLX install guide](../getting_started/installation/mlx.md)
 - **Python**: 3.10-3.12
 - **CUDA**: 12.6 or 13.0 (NVIDIA GPUs)
- **GPU**: At least one NVIDIA GPU, or an Apple silicon chip with MPS
+- **GPU**: At least one NVIDIA GPU, or an Apple silicon chip with the MLX runtime

 ## Installation

@@ -38,18 +38,20 @@ def main():
    # Create a video generator with a pre-trained model
    generator = VideoGenerator.from_pretrained(
        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-        num_gpus=1,  # Adjust based on your hardware
+        {"engine": {"num_gpus": 1}},  # Adjust based on your hardware
    )

    # Define a prompt for your video
    prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest."

    # Generate the video
-    video = generator.generate_video(
-        prompt,
-        output_path="my_videos/",  # Controls where videos are saved
-        save_video=True
-    )
+    video = generator.generate({
+        "prompt": prompt,
+        "output": {
+            "output_path": "my_videos/",  # Controls where videos are saved
+            "save_video": True,
+        },
+    })

 if __name__ == '__main__':
    main()
@@ -75,23 +77,24 @@ Please see the [support matrix](support_matrix.md) for the list of supported mod
 You can generate a video starting from an initial image:

 ```python
-from fastvideo import VideoGenerator, SamplingParam
+from fastvideo import VideoGenerator

 def main():
    # Create the generator
    model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
-    generator = VideoGenerator.from_pretrained(model_name, num_gpus=1)
-
-    # Set up parameters with an initial image
-    sampling_param = SamplingParam.from_pretrained(model_name)
-    sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
-    sampling_param.num_frames = 107
+    generator = VideoGenerator.from_pretrained(model_name, {"engine": {"num_gpus": 1}})

    # Generate video based on the image
    prompt = "A photograph coming to life with gentle movement"
-    generator.generate_video(prompt, sampling_param=sampling_param,
-                             output_path="my_videos/",
-                             save_video=True)
+    generator.generate({
+        "prompt": prompt,
+        # Set up parameters with an initial image
+        "inputs": {
+            "image_path": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
+        },
+        "sampling": {"num_frames": 107},
+        "output": {"output_path": "my_videos/", "save_video": True},
+    })

 if __name__ == '__main__':
    main()
@@ -106,12 +109,12 @@ Common issues and their solutions:
 If you encounter CUDA out of memory errors:

 - Reduce `num_frames` or video resolution
- Enable FastVideo offloading options such as `dit_layerwise_offload=True`
-  (single GPU) or `use_fsdp_inference=True` (multi-GPU)
+- Enable FastVideo offloading options such as `engine.offload.dit_layerwise: true`
+  (single GPU) or `engine.use_fsdp_inference: true` (multi-GPU)
 - Try a smaller model or use distilled versions
- Use `num_gpus` > 1 if multiple GPUs are available
- Try enabling FSDP inference with `use_fsdp_inference=True` (may slow down generation)
- Try enabling DiT layerwise offload with `dit_layerwise_offload=True` (now only a few models support this, but may introduce less overhead than FSDP)
+- Use `engine.num_gpus` > 1 if multiple GPUs are available
+- Try enabling FSDP inference with `engine.use_fsdp_inference: true` (may slow down generation)
+- Try enabling DiT layerwise offload with `engine.offload.dit_layerwise: true` (now only a few models support this, but may introduce less overhead than FSDP)

 ### Slow Generation

@@ -10,7 +10,9 @@ that tradeoff. It is not a lossless acceleration of the full VAE.

 ## Generate a video

-Use your existing MLX FastH3 environment and converted checkpoint:
+Use your existing MLX FastH3 environment and converted checkpoint. The same
+`--video-decode-backend taeh3` flag works on `mlx_fasth3.py` (V1) and
+`mlx_fasth3_8step.py` (V2):

 ```bash
 python examples/inference/basic/mlx_fasth3.py \
@@ -68,7 +70,7 @@ Run the numerical tests against a local TAEHV checkout containing the released
 weights:

 ```bash
-TAEH3_REFERENCE_DIR=/path/to/taehv \
+FASTVIDEO_TEST_TAEH3_REFERENCE_DIR=/path/to/taehv \
  python -m pytest fastvideo/tests/mlx/test_mlx_taeh3.py -q
 ```

@@ -4,14 +4,17 @@ This page describes how to use offloading techniques for inference to reduce GPU

 ## Default Behavior

-```python
-dit_cpu_offload: bool = True
-use_fsdp_inference: bool = False
-dit_layerwise_offload: bool = True
-text_encoder_cpu_offload: bool = True
-image_encoder_cpu_offload: bool = True
-vae_cpu_offload: bool = True
-pin_cpu_memory: bool = True
+```yaml
+engine:
+  use_fsdp_inference: false
+  offload:
+    dit: true               # dit_cpu_offload
+    dit_layerwise: true     # dit_layerwise_offload
+    text_encoder: true      # text_encoder_cpu_offload
+    image_encoder: true     # image_encoder_cpu_offload
+    vae: true               # vae_cpu_offload
+    pin_cpu_memory: true
+    lazy_module_load: null  # auto
 ```

 On unified-memory accelerators such as NVIDIA GB10 and Apple silicon, FastVideo
@@ -21,23 +24,33 @@ pool there, so offload adds transfers and duplicate residency instead of freeing
 memory. CUDA FSDP sharding remains enabled when requested; MPS continues to
 disable FSDP. `pin_cpu_memory` is not an offload mode and is left unchanged.

-MiniMax H3 CUDA inference can use a second lever that does not copy weights to a
-host pool: load the Qwen3-VL text encoder, run conditioning, then release that
-encoder before loading the DiT and video/audio VAEs. `h3_sequential_load`
-defaults to auto (`None`): on for unified-memory devices such as GB10, off on
-discrete GPUs. Pass `--h3-sequential-load` to force it, or
-`--no-h3-sequential-load` to keep the encoder resident for later `generate()`
-calls on the same worker. Sequential load currently cannot re-encode a new prompt
-on that worker; start a new generator until prompt-cache reload exists. The MLX
-FastH3 runtime always uses this phase order. When host offload is off, DiT
-safetensors are read onto the accelerator instead of CPU-then-copy.
-Input-preparation geometry (spatial ratio, latent channels, audio sample rate)
-comes from the VAE arch configs until those weights load.
+MiniMax H3 CUDA inference can use two levers that do not copy weights to a host
+pool. `lazy_module_load` is the general path: each opted-in component loads on
+first use and is freed after its last stage, so a later `generate()` reloads
+from disk in-process and the DiT can drop before VAE decode. On GB10 it
+auto-enables and owns deferral. `h3_sequential_load` is the H3-only fallback
+when lazy is off: load Qwen3-VL, run conditioning, release that encoder, then
+load the DiT and VAEs. When both would arm, sequential stands down so VAE
+`torch.compile` can attach to the lazy proxy. Input preparation and unpatchify
+read geometry from checkpoint `config.json` (VAE spatial ratio / latent
+channels, DiT patch size) so those stages do not materialize weights just to
+read two integers. The MLX FastH3 runtime always uses this phase order. When
+host offload is off, DiT safetensors are read onto the accelerator instead of
+CPU-then-copy. Both flags default to auto (`None`) and turn on for
+unified-memory devices such as GB10; lazy then disables sequential. Set
+`engine.offload.lazy_module_load: false` to keep every component resident (sequential may still
+auto-arm). Two-node Spark
+jobs still need this split: sequence parallel replicates the DiT on each GB10
+(~66 GiB of weights plus activations). See
+[Pair two NVIDIA DGX Sparks](../getting_started/installation/spark_pair.md).

 ## Behavior Explanation

 !!! note
-    For CLI usage, replace underscores (`_`) with hyphens (`-`).
+    `VideoGenerator.from_pretrained` accepts the option names below as keywords, except `lazy_module_load` and
+    `h3_sequential_load`. In a YAML config or a dotted override, each option is a typed field:
+    `engine.use_fsdp_inference`, `engine.offload.<field>` as listed in the defaults above, and
+    `pipeline.model.minimax_h3.sequential_load` for `h3_sequential_load`.

 ### `use_fsdp_inference`

@@ -93,9 +106,12 @@ because the encoder has been released.

 #### Usage Recommendation

-Leave the default on Spark / DGX Spark. Force `--h3-sequential-load` only when
-you need the split on a discrete GPU. Use `--no-h3-sequential-load` when you
-need more than one prompt per worker and have enough memory to keep the encoder.
+Leave the default on Spark / DGX Spark when `lazy_module_load` is off. When
+both would arm (the GB10 auto case), lazy owns deferral and sequential stands
+down so VAE `torch.compile` can attach to the lazy proxy. Force
+`pipeline.model.minimax_h3.sequential_load: true` only when you need the split on a discrete GPU without
+lazy load. Set `pipeline.model.minimax_h3.sequential_load: false` when you need more than one prompt per
+worker and have enough memory to keep the encoder.

 ### `text_encoder_cpu_offload`

@@ -121,6 +137,52 @@ These options introduce performance overhead due to PCIe data transfer.

 We recommend enabling these options when OOM happens.

+### `lazy_module_load`
+
+Every option above moves weights between host and device. This one changes
+whether they are in memory at all.
+
+By default a pipeline loads every component before the first stage runs, so
+peak memory is the sum of all of them even though no two are needed at the same
+moment. With `lazy_module_load` enabled, each heavy component loads on first use
+and is freed once the last stage that needs it has returned, so peak memory
+becomes the largest overlapping set instead of the sum. MiniMax-H3 T2VA is
+`max(text encoder, DiT, VAE)` rather than `text encoder + DiT + VAE`, because
+the DiT is not held through VAE decode.
+
+#### Performance Impact
+
+A freed component is read from disk again on the next generation, so a
+multi-prompt run pays one reload per component per request. For a large text
+encoder that is tens of seconds. If pipeline-level `torch.compile` is enabled,
+the compile setup is reapplied after each reload; PyTorch can reuse its graph
+and kernel caches when the component structure and input shapes are unchanged.
+
+#### Usage Recommendation
+
+Enable this when a model does not fit at load time, which the CPU offload
+options above cannot help with because they act after loading. It is
+particularly relevant on unified-memory devices, where host and device draw on
+the same pool and moving weights to the host frees nothing. FastVideo
+auto-enables it there (`lazy_module_load=None`). Leave it off when the model
+already fits, or set `engine.offload.lazy_module_load: false` to keep components resident for
+later `generate()` calls.
+
+This option applies to inference only. Training keeps every component resident
+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, 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
+a pipeline that has not, the flag is a no-op: hooks are not installed and no
+warning is logged. Sequential MiniMax-H3 (`h3_sequential_load`) reloads the
+text encoder for a later `generate()` on the same worker; you do not need to
+start a new generator.
+
 ## General Recommendations

 ### Single GPU Inference
@@ -131,6 +193,12 @@ We recommend enabling `dit_layerwise_offload`. If OOM happens, also enable `imag

 We recommend enabling `use_fsdp_inference` and disabling both `dit_layerwise_offload` and `dit_cpu_offload`. If OOM happens, consider enabling `text_encoder_cpu_offload`, `image_encoder_cpu_offload`, and `vae_cpu_offload`. If OOM still happens, consider enabling `dit_cpu_offload`.

+### When the Model Does Not Fit at Load Time
+
+The offload options only help once loading has finished. If the run dies while
+components are still being placed, or if the machine has unified memory so
+there is no separate host pool to offload into, enable `lazy_module_load`.
+
 ## Examples

 ### Single GPU with Layerwise Offloading
@@ -140,19 +208,25 @@ from fastvideo import VideoGenerator

 generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    num_gpus=1,
-    # Recommended for single GPU
-    dit_layerwise_offload=True,
-    # Enable if OOM happens
-    vae_cpu_offload=True,
-    image_encoder_cpu_offload=True,
-    text_encoder_cpu_offload=True,
-    # Speeds up CPU-GPU transfer
-    pin_cpu_memory=True,
+    {
+        "engine": {
+            "num_gpus": 1,
+            "offload": {
+                # Recommended for single GPU
+                "dit_layerwise": True,
+                # Enable if OOM happens
+                "vae": True,
+                "image_encoder": True,
+                "text_encoder": True,
+                # Speeds up CPU-GPU transfer
+                "pin_cpu_memory": True,
+            },
+        },
+    },
 )

 prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers."
-video = generator.generate_video(prompt, output_path="output/", save_video=True)
+video = generator.generate({"prompt": prompt, "output": {"output_path": "output/", "save_video": True}})
 ```

 ### Multi-GPU with FSDP
@@ -162,18 +236,24 @@ from fastvideo import VideoGenerator

 generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    num_gpus=2,
-    # Recommended for multi-GPU
-    use_fsdp_inference=True,
-    dit_layerwise_offload=False,
-    dit_cpu_offload=False,
-    # Enable if OOM happens
-    vae_cpu_offload=True,
-    image_encoder_cpu_offload=True,
-    text_encoder_cpu_offload=True,
-    pin_cpu_memory=True,
+    {
+        "engine": {
+            "num_gpus": 2,
+            # Recommended for multi-GPU
+            "use_fsdp_inference": True,
+            "offload": {
+                "dit_layerwise": False,
+                "dit": False,
+                # Enable if OOM happens
+                "vae": True,
+                "image_encoder": True,
+                "text_encoder": True,
+                "pin_cpu_memory": True,
+            },
+        },
+    },
 )

 prompt = "A majestic lion strides across the golden savanna."
-video = generator.generate_video(prompt, output_path="output/", save_video=True)
+video = generator.generate({"prompt": prompt, "output": {"output_path": "output/", "save_video": True}})
 ```
@@ -7,7 +7,8 @@ This page describes the various options for speeding up generation times in Fast
    Several options on this page behave differently on the GB10's unified-memory
    hardware — some give little or nothing there. See
    [DGX Spark: Performance & Tuning](../getting_started/installation/spark_performance.md)
-    for what actually helps on that platform and why.
+    for what actually helps on that platform and why. Two Sparks, one clip:
+    [Pair two NVIDIA DGX Sparks](../getting_started/installation/spark_pair.md).

 ## Table of Contents

@@ -165,22 +166,26 @@ Enable FP4 attention via the `--nvfp4_fa4` flag:
 python examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py --nvfp4_fa4
 ```

-Or in Python via the `nvfp4_fa4` kwarg (sets env vars automatically):
+Or in Python via the `engine.attention.nvfp4_fa4` field (resolution sets the env vars):

 ```python
 from fastvideo import VideoGenerator
 gen = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    nvfp4_fa4=True,
-    num_gpus=1,
-    use_fsdp_inference=False,  # FSDP is incompatible with FP4 pointer path
+    {
+        "engine": {
+            "attention": {"nvfp4_fa4": True},
+            "num_gpus": 1,
+            "use_fsdp_inference": False,  # FSDP is incompatible with FP4 pointer path
+        },
+    },
 )
-gen.generate_video(prompt="A raccoon in sunflowers", save_video=True)
+gen.generate(request={"prompt": "A raccoon in sunflowers", "output": {"save_video": True}})
 ```

 #### Known Limitations

- `use_fsdp_inference=True` is incompatible with the FP4 path (FSDP shards invalidate tensor pointers)
+- `engine.use_fsdp_inference: true` is incompatible with the FP4 path (FSDP shards invalidate tensor pointers)
 - Per-call cosine similarity vs BF16: ~0.99 (slight quantization error accumulates over denoising steps)
 - Only supports `headdim >= 128`

@@ -204,15 +209,15 @@ import os
 os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "ATTN_QAT_INFER"

 from fastvideo import VideoGenerator
-from fastvideo.layers.quantization import get_quantization_config
-gen = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    num_gpus=1,
-    # Wan-2.1 uses the nvfp4_qat config (NVFP4 is LTX2-specific). Pass an
-    # instance — the bare string is not resolved on the from_pretrained path.
-    transformer_quant=get_quantization_config("nvfp4_qat")(),
-    use_fsdp_inference=False,     # FSDP shards invalidate the FP4 tensor pointers
-)
+gen = VideoGenerator.from_config({
+    "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    "engine": {
+        "num_gpus": 1,
+        "use_fsdp_inference": False,     # FSDP shards invalidate the FP4 tensor pointers
+        # Wan-2.1 uses the nvfp4_qat config (NVFP4 is LTX2-specific).
+        "quantization": {"transformer_quant": "nvfp4_qat"},
+    },
+})
 gen.generate(request={"prompt": "A raccoon in sunflowers", "output": {"save_video": True}})
 ```

@@ -293,21 +298,41 @@ automatically.
 ### Requirements

 - **GPU**: sm89+ (H100, L40S, RTX 4090, or newer) for hardware FP8 compute
+- **ROCm**: CDNA4 (MI350X / MI355X, gfx950) runs the FP8 `_scaled_mm` path through
+  hipBLASLt (OCP e4m3fn). MI300X (gfx942) only exposes the `fnuz` FP8 formats and
+  takes the bf16 dequant fallback like a pre-sm89 GPU.
 - No additional packages required beyond the base FastVideo install

 ### Usage

+```python
+from fastvideo import VideoGenerator
+
+gen = VideoGenerator.from_config({
+    "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    "engine": {"quantization": {"transformer_quant": "FP8"}},  # per-tensor (default)
+})
+gen.generate(request={"prompt": "A raccoon in sunflowers", "output": {"save_video": True}})
+```
+
+`engine.quantization.transformer_quant` takes a quantization registry name and builds that config with its default
+arguments. To pass constructor arguments, such as per-channel granularity, set the config instance on the DiT config
+through `pipeline.model.generic.dit` instead:
+
 ```python
 from fastvideo import VideoGenerator
 from fastvideo.layers.quantization import get_quantization_config

-gen = VideoGenerator.from_pretrained(
-    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    # Pass an instance — the bare string is not resolved on the from_pretrained path.
-    transformer_quant=get_quantization_config("FP8")(),          # per-tensor (default)
-    # transformer_quant=get_quantization_config("FP8")(granularity="channel"),  # slower, higher accuracy
-)
-gen.generate(request={"prompt": "A raccoon in sunflowers", "output": {"save_video": True}})
+gen = VideoGenerator.from_config({
+    "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    "pipeline": {
+        "model": {
+            "generic": {
+                "dit": {"quant_config": get_quantization_config("FP8")(granularity="channel")},  # slower, higher accuracy
+            },
+        },
+    },
+})
 ```

 Or run the example script:
@@ -337,7 +362,7 @@ end-to-end speedup. It is **off by default** and enabled per-run.
 ```python
 generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    enable_torch_compile=True,
+    {"engine": {"compile": {"enabled": True}}},
 )
 ```

@@ -372,12 +397,14 @@ device is unsupported. Legacy VSA, MiniMax-H3 tile-256 VSA, and the explicit
 eager with one warning instead of failing mid-denoise.

 ```python
-generator = VideoGenerator.from_pretrained(
-    "MiniMaxAI/MiniMax-H3",
-    inference_torch_compile=True,  # or FASTVIDEO_INFERENCE_TORCH_COMPILE=1
-)
+generator = VideoGenerator.from_config({
+    "model_path": "MiniMaxAI/MiniMax-H3",
+    "engine": {"compile": {"regional": True}},  # or FASTVIDEO_INFERENCE_TORCH_COMPILE=1
+})
 ```

+In a YAML config, set `generator.engine.compile.regional: true`.
+
 Do not combine it with `torch_compile_kwargs['mode']` (the loader injects
 inductor options, and torch.compile forbids mode+options); it is
 independent of `enable_torch_compile`, and when both are set the regional
@@ -386,7 +413,7 @@ compile wins for the DiT.
 ### What to expect from generic compile

 The Wan result below measures the existing generic
-`enable_torch_compile=True` path. It is useful evidence that compile can help,
+`engine.compile.enabled: true` path. It is useful evidence that compile can help,
 but it is **not** a benchmark or numerical gate for the stricter regional
 fullgraph path above.

@@ -429,12 +456,12 @@ not asserted by any standing SSIM regression here — the SSIM tests in
 run with `enable_torch_compile` disabled. If you depend on compile
 output staying close to eager (or your previous compiled run), run an
 MS-SSIM gate on *your* config, especially when combining
-`enable_torch_compile=True` with other numerics-affecting flags
+`engine.compile.enabled: true` with other numerics-affecting flags
 (quantized attention backends, FP4, layerwise offload edge cases).

 ### Known interactions

- **Layerwise CPU offload** (`dit_layerwise_offload=True`, the default):
+- **Layerwise CPU offload** (`engine.offload.dit_layerwise: true`, the default):
  the offload hook previously caused an implicit graph break once per
  transformer layer, fragmenting the compiled region. Addressed in
  hao-ai-lab/FastVideo#1365 — keep that fix to get a clean compiled
@@ -449,16 +476,17 @@ MS-SSIM gate on *your* config, especially when combining
  grad-enabled path remain outside it. Use the default inductor mode shown
  above unless your exact configuration has its own gate.

-Extra `torch.compile` options are passed through `torch_compile_kwargs`
-(a dict), accepted by `VideoGenerator.from_pretrained(...)` and by the
-CLI as a JSON string via `--torch-compile-kwargs`. Example (currently
+Extra `torch.compile` options live at `engine.compile.backend`,
+`fullgraph`, `mode`, and `dynamic`; any other `torch.compile` kwargs go in
+`engine.compile.extras`. Set them in the nested config, in a config file,
+or as a CLI dotted override (for example
+`--generator.engine.compile.mode reduce-overhead`). Example (currently
 **not** recommended — see the CUDA-graphs caveat above):

 ```python
 VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
-    enable_torch_compile=True,
-    torch_compile_kwargs={"mode": "reduce-overhead"},  # may error today
+    {"engine": {"compile": {"enabled": True, "mode": "reduce-overhead"}}},  # may error today
 )
 ```

@@ -471,7 +499,7 @@ config; **discard the first generation** (graph build):
 import time
 from fastvideo import VideoGenerator

-gen = VideoGenerator.from_pretrained("your-model-id", enable_torch_compile=True)
+gen = VideoGenerator.from_pretrained("your-model-id", {"engine": {"compile": {"enabled": True}}})
 req = {"prompt": "Your prompt", "sampling": {"seed": 1024},
       "output": {"save_video": False}}
 gen.generate(req)                                            # warmup: graph build, discard
@@ -495,10 +523,10 @@ for backend in ["TORCH_SDPA", "FLASH_ATTN", "SAGE_ATTN"]:
    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = backend
    generator = VideoGenerator.from_pretrained("your-model-id")
    start_time = time.perf_counter()
-    generator.generate_video(
-        prompt="Your prompt",
-        seed=1024,
-    )
+    generator.generate({
+        "prompt": "Your prompt",
+        "sampling": {"seed": 1024},
+    })
    elapsed = time.perf_counter() - start_time
    print(f"{backend}: {elapsed:.2f}s")
 ```
--- a/Show More
+++ b/Show More