feat(cake_softmax): add optimized Blackwell softmax - #4282

Open

yyihuang wants to merge 14 commits into

flashinfer-ai:mainfrom

yyihuang:perf/blackwell-softmax-loom

Commits on Jul 30, 2026

Optimize Blackwell softmax with Loom kernels
averyhNV
committed
317fa85
Use CUDA types in generated Blackwell header
averyhNV
committed
6e025b1
Keep Blackwell softmax dispatch separate
averyhNV
committed
210ec16
Format Blackwell softmax sources
averyhNV
committed
6a8e024
Integrate final Blackwell softmax dispatcher
averyhNV
committed
2093dcb
Preserve generated Softmax payload formatting
averyhNV
committed
b82c54b
Sync final Loom Softmax dispatcher
averyhNV
committed
657d533
Use standard integer types in exported kernels
averyhNV
committed
bc6f334
style: apply pre-commit formatting
averyhNV
committed
754aee2

Commits on Aug 5, 2026

Commits on Aug 6, 2026

fix(logits-processor): honor softmax architecture dispatch
averyhNV
committed
f0edac6