FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree

FastH3 4-Step Preview v1 checkpoint from FastVideo. This variant is the v10.5 data-free DMD2 MiniMax-H3 student at training step 1300. It generates synchronized video and audio with four transformer forwards and uses VSA-H3 at 90% sparsity.

See the complete FastH3 4-Step Preview v1 collection.

VSA requirement: This checkpoint must be run with FastVideo's Video Sparse Attention (VSA-H3) backend and kernel enabled at 90% sparsity with 64-token tiles. Dense attention is not a supported substitute.

Checkpoint identity

• Training run: 20260823T132231Z-v10p5-datafree-sp4
• W&B: a1etree2
• Step: 1300
• Transformer content SHA-256: b36987515e4c75fa4c7aaa632a7842c829ea141b235358a54d782b51230497b3
• Checkpoint metadata SHA-256: dcad0fbee2a7c7e75e53435f4fd98fccf3138844883874edf057962ab48fa428
• FastVideo execution commit: 48a047c05ff4138f20cfa33351499c6ec5945f5d

checkpoint_content.json records every indexed transformer shard and digest. checkpoint_metadata.json preserves the source training configuration.

Contents

This is a self-contained Diffusers modular-pipeline package. The distilled bf16 student is under transformer/; the text encoder, tokenizer, processor, video VAE, audio VAE, and schedulers are unmodified MiniMax-H3 components.

The transformer_ref component is not bundled because this student targets T2AV. Its modular-index entry remains on MiniMaxAI/MiniMax-H3.

Exact FastVideo inference contract

The student was trained on the explicit ladder [999, 749, 500, 250]. These are four transformer forwards; do not replace the ladder with an arbitrary four-point scheduler grid.

export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250

python examples/inference/basic/basic_fasth3.py \
  --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \
  --prompt '<MiniMax-H3 T2AV prompt>' \
  --steps 5 \
  --num-gpus 4 \
  --vsa-sparsity 0.9 \
  --vsa-tile-size 64 \
  --vsa-kernel sm100a

The five scheduler grid points execute the four trained jumps. Use guidance scale 1.0, VSA-H3 at 90% sparsity with 64-token tiles, and SP=4/FSDP across one four-GPU B200 tray for the validated route. The exact v10.5 execution commit runs the sparse DiT eagerly; regional compile is disabled for this snapshot.

Training summary

• Method: data-free DMD2 with carried backward-simulation rollouts.
• Student grid: [999, 749, 500, 250].
• Student attention: VSA-H3, 64-token tiles, 90% sparsity.
• Prompt conditioning: mixed native resolution and duration T2AV prompts; no target video latents are consumed by the data-free objective.
• Global batch size: 64.
• Student and critic learning rate: 2e-6.
• Precision: fp32 training state and bf16 inference export.
• Training topology: 32 B200 GPUs, SP=4, full-world FSDP sharding.

Limitations

• Intermediate checkpoint at step 1300 of a planned 4000-step run.
• This historical checkpoint predates the later continuous score-clock and FastGen-alignment corrections. Use it to reproduce v10.5 results, not as the reference implementation of the corrected training recipe.
• T2AV only; the reference-conditioned transformer was not distilled here.
• Quality remains below base MiniMax-H3 on difficult motion, fine structure, and some audio details.
• Sampling outside the trained ladder or attention geometry is off-distribution.
• Inherits the MiniMax H3 Community License and base-model limitations.