HunyuanVideo-13B¶
HunyuanVideo-13B is a 13B text-to-video path in Sol-Engine. The optimized
configuration runs single-GPU with torch.compile, TeaCache, and
Sol-Attn
(techniques/sparse_backends/).
Performance status¶
Full optimization uses compile + TeaCache + the released Sol-Attn kernel. Historical timing used the retired sparse backend and is not quoted for this configuration; a fresh same-config benchmark is pending.
Launch¶
# baseline (vanilla diffusers HunyuanVideoPipeline, single GPU)
python3 scripts/launch_candidate.py candidates/hunyuan_video_baseline.toml --mode sbatch --confirm-submit
# optimized release stack
python3 scripts/launch_candidate.py candidates/hunyuan_video_full.toml --mode sbatch --confirm-submit
The optimized candidate enables torch.compile, TeaCache, and Sol-Attn together; every technique is env-gated (all flags off = byte-identical baseline).
Techniques¶
- Cache: TeaCache reuses denoising steps (threshold 0.15, start step 6, max 2 consecutive hits).
- Kernel fusion: torch.compile over the transformer.
- Sparse attention: Sol-Attn at
tau=1.0; valid text K/V is an exact sink and valid text-query rows stay dense.