Repository navigation

[FP8 3/5] Integrate FP8 attention and caching into TabICLv2 - #879

Draft

aw471 wants to merge 11 commits into

fp8-stack/kernelsfrom

fp8-stack/01-tabiclv2

Draft

[FP8 3/5] Integrate FP8 attention and caching into TabICLv2#879
aw471 wants to merge 11 commits into
fp8-stack/kernelsfrom
fp8-stack/01-tabiclv2

Commits on Oct 24, 2026
- Enable FP8 attention in shared attention and TabICLv2
  aw471
  committed
  b401074
- Enable FP8 attention with a scoped optimization policy
  aw471
  committed
  ce8a9fe
- Test model cache reuse across optimization scopes
  aw471
  committed
  1062545
- Make FP8 fallback and numerical regressions explicit
  aw471
  committed
  691c908
- Cover FP8 automatic chunking and estimator cache offloading
  aw471
  committed
  9b54f89
- Use centralized FP8 dispatch with existing SDPA fallback
  aw471
  committed
  3bf9ada
- Explain why final query-only attention stays on SDPA
  aw471
  committed
  458cc32
- Preserve causal attention behavior when FP8 is enabled
  aw471
  committed
  d1cb352
- Simplify FP8 eligibility and keep focused integration coverage
  aw471
  committed
  62fe0a2
- Keep the attention forward path focused on projection and dispatch
  aw471
  committed
  10bb081
Commits on Oct 26, 2026
- Enable FP8 attention on CUDA architectures with native FP8 support
  aw471
  committed
  6c4171b