nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[docs] Update FastH3 README (#1749)

Browse Source
This commit is contained in:
William Lin
2026-08-23 05:05:58 -07:00
committed by GitHub
parent f9e3680f11
commit a837fe841a
3 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/inference/basic/README.md
+4 -2
View File
@@ -33,7 +33,9 @@ For the typed config/request path added during the inference API refactor:
python examples/inference/basic/basic_dmd_new_api.py
```
For the few-step (4-forward, DMD2-distilled) MiniMax-H3 preview, generating synchronized video and audio with its trained block-sparse VSA attention:
### FastH3 Preview
The verified [basic FastH3 example](basic_fasth3.py) runs the few-step (4-forward, DMD2-distilled) MiniMax-H3 preview, generating synchronized video and audio with its trained block-sparse VSA attention:
```bash
UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"
@@ -45,7 +47,7 @@ the measured GB200 profile. Then run:
```
python examples/inference/basic/basic_fasth3.py --prompt "your prompt"
```
The default checkpoint `FastVideo/FastVideo-Minimax-FastH3-Preview-v0.1` is private on the Hub while its license review completes; until it flips public, pass `--model-path` with a local snapshot of the release.
The default checkpoint, [FastH3 Preview v0.2](https://huggingface.co/FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2), is public on the Hub under the MiniMax H3 Community License. Review its model card and license before use or redistribution.
The default `all` profile is the fastest measured four-GPU Preview recipe on GB200. It selects VSA sparsity 0.9 with 64-token tiles and the sm_100a sparse kernel, enables FA4 for eligible non-VSA paths, regionally compiles and replicates the sparse DiT, compiles and temporally parallelizes the video VAE with the `gather` strategy, and pins CPU-offloaded component memory. It also pins the benchmark protocol: five sigma-grid points (exactly four DiT forwards), one excluded seed-999 warmup, then three timed seed-1000 requests with distinct output paths.
examples/inference/basic/basic_fasth3.py
+1 -1
View File
@@ -36,7 +36,7 @@ from fastvideo.api import (
SamplingConfig,
)
DEFAULT_MODEL = "FastVideo/FastVideo-Minimax-FastH3-Preview-v0.1"
DEFAULT_MODEL = "FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2"
def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace: