Repository navigation

[NPU] Add opt-in MLA prefix FIA and Gemma norm fallback - #40131

Open

McZyWu wants to merge 10 commits into

sgl-project:mainfrom

McZyWu:fix/ascend-mla-prefix-fia-opt-in

Commits on Sep 18, 2026
- [NPU] Add opt-in FIA for MLA prefix prefill
  McZyWu
  committed
  d71e2ef
- style(npu): remove MLA FIA inline comments
  McZyWu
  committed
  61428f5
- fix(npu): use torch_npu quantization entry points
  McZyWu
  committed
  1344e1e
Commits on Sep 21, 2026
- Merge branch 'main' into fix/ascend-mla-prefix-fia-opt-in
  McZyWu
  authored
  3ba07dd
Commits on Sep 22, 2026
- fix(npu): honor native fallback for Gemma RMSNorm variants
  McZyWu
  committed
  ec72a1d
Commits on Sep 24, 2026
- Revert "fix(npu): use torch_npu quantization entry points"
  McZyWu
  committed
  ef8a80b
- fix(npu): gather logical MLA prefix pages from NZ cache
  McZyWu
  committed
  0c8ebbc
Commits on Sep 28, 2026
- fix(npu): revert Gemma native fallback argument change
  McZyWu
  committed
  c797eef
- refactor(npu): keep MLA prefix FIA and RingMLA paths separate
  McZyWu
  committed
  ee3dfe9
- Merge latest main into MLA prefix FIA branch
  McZyWu
  committed
  a2dd434