| --- |
| license: other |
| base_model: FastVideo/FastVideo-FastH3-8-Step-V2 |
| tags: [video-generation, text-to-video, audio, fastvideo, nvfp4, qad] |
| --- |
| # FastH3-QAD-v2, 8-step, NVFP4 for one GPU |
|
|
| FastH3-QAD-v2 is an 8-step video-and-audio model with NVFP4 weights, trained with **quantization-aware distillation |
| (QAD)**. It runs on one Blackwell GPU (RTX 5090, RTX PRO 6000) or a DGX Spark. |
|
|
| - **Transformer:** 50 blocks, NVFP4 W4A4 on all attention q/k/v/out and MLP linears, plus the sparse-attention |
| (VSA) gates. The q/k/v/out and MLP weights were trained with QAD. The VSA gates and all norms stay in BF16. |
| - **Attention:** block-sparse VSA with 64-token tiles and sparsity 0.8. The checkpoint was trained with BF16 VSA. |
| FastVideo also exposes an FP4 VSA inference lane on Blackwell. |
| - **Text encoder:** NVFP4 Qwen3-VL trimmed to the 50 layers H3 reads. |
| - **VAE:** LynnReal lightweight video VAE with Kijai's INT8 weights; H3 audio VAE. |
| - **Sampling:** `fastvideo_inference.json` holds the 8-step DMD schedule, which FastVideo reads automatically. |
|
|
| ## Quantization-aware distillation |
|
|
| QAD trains the NVFP4 student against a frozen BF16 teacher along the teacher's 8-step trajectory. The student runs |
| the deployed 4-bit weights and activations, so it learns weights that work with the quantized forward pass. Training |
| uses text prompts only: 300 steps, learning rate 5e-6, with 25% of rows on the student's own trajectory. Only the |
| quantized linears train. The VSA gates and norms are frozen because changing them changes which video tiles the |
| model attends to, not just the arithmetic of attention. |
|
|
| Error against the BF16 teacher on 30 held-out prompts (relative L2 of the x0 prediction, lower is better): |
|
|
| | Metric | FastH3-QAD-v2 | |
| |---|---:| |
| | Mean (video + audio) | 0.351 | |
| | Video | 0.382 | |
| | Audio | 0.319 | |
|
|
| ## Machines with less than 64 GB of system RAM |
|
|
| The 50 AdaLN timestep projections are 26 GB in BF16. `transformer/adaln_tables.pt` (155 MB) holds their precomputed |
| outputs for the 8-step schedule, so FastVideo can skip loading them: |
|
|
| ```python |
| import os |
| from huggingface_hub import hf_hub_download |
| |
| os.environ["FASTVIDEO_H3_ADALN_TABLE"] = hf_hub_download( |
| "FastVideo/FastH3-QAD-v2", "transformer/adaln_tables.pt") |
| ``` |
|
|
| The table covers only the shipped schedule (`fastvideo_inference.json`); FastVideo raises an error for other timesteps. |
|
|
