Skip to content

Wan2.1-T2V-14B

Wan2.1-T2V-14B is the 14B dense text-to-video path in Sol-Engine. The optimized configuration runs single-GPU with kernel compilation, EasyCache, and Sol-Attn (techniques/sparse_backends/).

Performance status

Full optimization uses compile + EasyCache + 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 WanPipeline, single GPU)
python3 scripts/launch_candidate.py candidates/wan21_14b_baseline.toml --mode sbatch --confirm-submit

# optimized release stack
python3 scripts/launch_candidate.py candidates/wan21_14b_fullstack.toml --mode sbatch --confirm-submit

The optimized candidate enables kernel compilation, EasyCache, and Sol-Attn together; every technique is env-gated (all flags off = byte-identical baseline).

Techniques

  • Cache: EasyCache reuses denoising work.
  • Kernel fusion: compiled DiT glue with invariant caches.
  • Sparse attention: Sol-Attn at tau=1.0 with dense guards on the first layer and the first denoising steps.