FastVideo-FastH3-4-step-Preview-v1-VSA-Synthetic-Step1900

FastH3 4-Step Preview v1 checkpoint from FastVideo. This variant is the dmd2_sp2_fsdp64_v11_dataonly_mixed_vsa64 synthetic-data DMD2 MiniMax-H3 student at training step 1900. 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: 20260825T003627Z-v11-dataonly-sp2-fastgen-fix
• Lineage parent: 20260823T191211Z-v11-dataonly-sp2-initial
• W&B: arg07fnt
• Step: 1900
• Transformer content SHA-256: 932493fce3c32de6bb49e0d568ec493bec365a62b0e5a92262ab73a1619f0dda
• Checkpoint metadata SHA-256: d93f3974fa6a916a0989327b8861c0fb9f5c028b33f95e65c45925ea49dcdbfe
• FastVideo execution commit: 6ac300b5a7c472e3a139f1cb3ff7a03b58f3e26e
• Data commit: b92be606a439c67006220459df236d18c48fd38c3e11f4771d21fe0d078a3371

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-Synthetic-Step1900 \
  --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 and VSA-H3 at 90% sparsity with 64-token tiles. Compilation is an optional runtime optimization; first validate output parity on the pinned FastVideo commit.

Training summary

• Method: data-only DMD2 on synthetic Base-H3 video latents. The student is trained from forward-noised target latents; no data-free carried rollout is mixed into this run.
• Student grid: [999, 749, 500, 250].
• Student attention: VSA-H3 at 90% sparsity with 64-token tiles.
• Global batch size: 64.
• Student and critic learning rate: 2e-06 and 2e-06.
• Precision: fp32 training state and bf16 inference export.
• Training topology: 64 B200 GPUs, SP=2, HSDP replicate=1 / shard=64.
• Continuous score-clock shift: 2.4; warp max: 0.999.

Limitations

• Intermediate checkpoint at step 1900 of a planned 4000-step run.
• 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.

Indexed student: 688 bf16 tensors / 70,099,502,592 logical bytes.