FP8 Sol-Attn with Support to Dense&Sparse Wan2.1 and Minimax-H3 - #30

Merged

lzx1413 merged 14 commits into

Tele-AI:mainfrom

Uxito-Ada:fp8_quant_and_sol

Aug 20, 2026

Commits on Aug 20, 2026

Add Wan FP8 quantization with Sol-Attn
Uxito-Ada
committed
4cb350b
Fix Wan FP8 test without tf-kernel
Uxito-Ada
committed
bef7efd
Consolidate Wan optimization example
Uxito-Ada
committed
a481741
feat(wan): run Sol-Attn with FP8 QKV GEMMs
Uxito-Ada
committed
f9b5b9f
bench(wan): update tf-kernel FP8 Sol-Attn results
Uxito-Ada
committed
b62bc3e
perf(wan): keep H100 Sol-Attn on CuTe
Uxito-Ada
committed
1713240
bench(wan): refresh clean H100 FP8 Sol results
Uxito-Ada
committed
aef4f9b
docs(wan): keep benchmark artifacts local
Uxito-Ada
committed
6ed8551
feat: add native FP8 Sol-Attn for Wan
Uxito-Ada
committed
79a5151
fix: preserve Wan FP8 Sol generation quality
Uxito-Ada
committed
638d2d7
feat: add FP8 dense attention ablation for Wan
Uxito-Ada
committed
5c22448
perf: accelerate Wan FP8 Sol attention
Uxito-Ada
committed
ef49ab4
feat: add MiniMax H3 FP8 Sol attention
Uxito-Ada
committed
b7df154
docs: add FP8 Sol-Attn technical article
Uxito-Ada
committed
3ea630c