Model card: match the 4-step Preview card layout; add acknowledgements

Documentation only. README.md rewritten to the FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree card structure (logo, summary, run, scope, acknowledgements). No weights or provenance files changed.

README.md

@@ -12,9 +12,7 @@ tags:
 - text-to-audio-video
 - distillation
 - dmd2
-- vsa
 - few-step
-- 8-step
 - minimax-h3
 - fastvideo
 - fasth3
@@ -26,44 +24,25 @@ tags:
 
 # FastVideo-FastH3-8-Step-V2
 
-FastH3 8-Step V2 is a text-to-audio-video checkpoint from
-[FastVideo](https://github.com/hao-ai-lab/FastVideo), distilled from
-[MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3). It generates video
-and stereo audio in eight transformer forwards using data-free DMD2 and
-Video Sparse Attention (VSA).
+The FastH3 8-Step V2 checkpoint from
+[FastVideo](https://github.com/hao-ai-lab/FastVideo). It generates synchronized
+video and audio from text with eight transformer forwards. This step-1300 model
+was trained with data-free DMD2 and VSA-H3 at 80% sparsity.
 
-[FastVideo](https://github.com/hao-ai-lab/FastVideo) ·
-[FastH3 blog](https://haoailab.com/blogs/fasth3-preview/) ·
+[Blog](https://haoailab.com/blogs/fasth3-preview/) ·
 [FastH3 collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3)
 
-Powered by MiniMax H3.
-
-
-## Model
-
-| Setting | Value |
-|---|---|
-| Task | Text to video with stereo audio (T2AV) |
-| Method | Data-free [DMD2](https://arxiv.org/abs/2405.14867), carried backward simulation |
-| Student attention | [VSA-H3](https://arxiv.org/abs/2505.13389), 80% sparsity, 64-token tiles |
-| Denoising | 8 transformer forwards |
-| Video / audio shifts | 10 / 3 |
-| Guidance scale | 1.0 |
-| Checkpoint | Training step 1300, bf16 full weights |
-
-Training used prompt conditioning and student-generated latents, not target
-video latents. The corpus used mixed resolutions and aspect ratios. This is
-an eight-step DMD2 checkpoint, not a PDD model.
-
-**This checkpoint requires FastVideo's VSA-H3 backend and Video Sparse
-Attention kernel.** The four-step models' default 90% sparsity and 12/3 shifts
-are not this model's sampling recipe. No matching eight-step LoRA is included.
+> This checkpoint requires FastVideo's VSA-H3 attention backend. Its video
+> scheduler shift is 10, not the base model's 12; use the example below, which
+> reads the trained schedule from the checkpoint.
 
 ## Run with FastVideo
 
-Start with the [installation guide](https://haoailab.com/FastVideo/getting_started/installation/)
-or [agent-guided installation](https://github.com/hao-ai-lab/FastVideo#install-with-an-ai-coding-agent).
-For NVIDIA B200 with CUDA 13, the public FastVideo environment is installed with:
+Install [uv](https://docs.astral.sh/uv/getting-started/installation/), then use
+the CUDA 13 / Blackwell path below. It selects FastVideo's published CUDA
+kernel wheel instead of compiling the kernel locally. See the
+[installation guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/)
+for other platforms.
 
 ```bash
 git clone https://github.com/hao-ai-lab/FastVideo.git
@@ -75,58 +54,24 @@ UV_TORCH_BACKEND=cu130 uv pip install \
   -e ".[fasth3]"
 ```
 
-Inference support for this checkpoint (video shift 10 and the trained
-eight-rung ladder) is in FastVideo
-[PR #1852](https://github.com/hao-ai-lab/FastVideo/pull/1852). Until it is
-merged, check out that branch; afterwards, `main` works unchanged.
-
 ```bash
 python examples/inference/basic/basic_fasth3_8step.py \
-  --prompt 'A slow cinematic drone shot glides over a coastal town at golden hour; gulls call over the harbor as a church bell rings twice.' \
-  --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
+  --prompt "your prompt" \
+  --no-warmup \
+  --repeats 1
 ```
 
-`basic_fasth3_8step.py` pins this checkpoint and its trained recipe (nine
-sigma-grid points = eight forwards, VSA 0.8, 64-token tiles) and reads the
-shifts and rung ladder from `scheduler/`, `audio_scheduler/` and
-[fastvideo_inference.json](fastvideo_inference.json). It shares the FastH3
-preview example's CLI, so every other flag works unchanged; pass
-`--model-path` to use a local snapshot.
-
-To download ahead of time:
+The tested defaults use four B200 GPUs and the trained eight-forward schedule.
+On other multi-GPU CUDA systems, follow the installation guide and add
+`--no-replicated-dit --vsa-kernel triton --no-fa4`. The GPU count must divide
+H3's 56 attention heads.
 
-```bash
-hf download FastVideo/FastVideo-FastH3-8-Step-V2 \
-  --local-dir ./FastH3-8-Step-V2
-```
+## Scope
 
-The complete package is approximately 148 GB. It includes the distilled
-transformer, text encoder, tokenizer, processor, video VAE and audio VAE.
-It is stored in a modular Diffusers layout, but the supported runtime is
-FastVideo with VSA-H3, not an unmodified dense Diffusers pipeline.
-
-## Scope and limitations
-
-- T2AV only. FL2VA and Ref2VA were not distilled or validated for this release.
-- Difficult motion, small details and some audio can remain below Base H3.
-- Sampling outside the trained schedule or attention policy is an ablation.
-- The four-step blog's latency and evaluation numbers do not establish this
-  eight-step model's performance. On 4x GB200 (SP4, eager strict profile,
-  VSA 0.8 tile 64, 832x480, 124 frames + 32 kHz audio) one measured request
-  took 6.3 s end to end including saving (FastVideo PR #1852 smoke). Quality
-  has not been formally compared against Base H3 or the four-step preview.
-
-This model inherits the [MiniMax H3 Community License](LICENSE), including
-its territorial, use and commercial restrictions. Review it before use or
-redistribution. The bundled Qwen3-VL encoder is covered by
-[Apache 2.0](LICENSE-Qwen3-VL). See [NOTICE](NOTICE) for attribution and modifications.
-
-Training and checkpoint provenance are preserved in
-[provenance.json](provenance.json), [checkpoint_metadata.json](checkpoint_metadata.json)
-and [checkpoint_content.json](checkpoint_content.json).
+This checkpoint supports text-to-audio-video generation. FL2VA and Ref2VA were
+not distilled. Difficult motion, fine detail, and some audio may remain below
+the base MiniMax H3 model. This checkpoint inherits the
+[MiniMax H3 Community License](LICENSE).
 
 ## Acknowledgements