Repository navigation

Optionally pad QK/V for better kernel selection - #3339

Open

janEbert wants to merge 31 commits into

NVIDIA:mainfrom

janEbert:optional-mla-pad

Open

Optionally pad QK/V for better kernel selection#3339
janEbert wants to merge 31 commits into
NVIDIA:mainfrom
janEbert:optional-mla-pad

Commits on Oct 12, 2026
- Automate MLA padding-related kernel selection
  janEbert
  committed
  15a34b4
- Add padding-related tests
  janEbert
  committed
  6e5bf36
- Fix tests
  janEbert
  committed
  748957e
- Add memoization
  janEbert
  committed
  64c347f
- Add test to CI
  janEbert
  committed
  6bdac9b
- Skip padding when using fused+packed FP8 path
  janEbert
  committed
  cf0e401
- Refactor optional pad if-guard
  janEbert
  committed
  0546cdf
- Merge output trim code paths
  janEbert
  committed
  1416805
- Reduce verbosity
  janEbert
  committed
  78c1348
- Explicitly mention padding target and limits
  janEbert
  committed
  69f7044
- Merge input pad code paths
  janEbert
  committed
  60a69a2
- Remove standalone MLA pad tests
  janEbert
  committed
  7559cad
- Assert MLA pad selects a faster backend
  janEbert
  committed
  ca7fd18
- Remove removed test from test runner script
  janEbert
  committed
  e576da3
- Reduce verbosity
  janEbert
  committed
  3702220
- Update Attn fwd type hints
  janEbert
  committed
  aae27cc
- Improve readability
  janEbert
  committed
  8f06b41
- Improve docstring
  janEbert
  committed
  f15da4f
- Fix padding selection
  janEbert
  committed
  2da4177
- Handle tuple Attn backend outputs
  janEbert
  committed
  def8ad1
- Fold FP8 DSA+packed check into main padding check
  janEbert
  committed
  4dcde58
Commits on Oct 13, 2026
- Fix backend selection cache invalidation
  janEbert
  committed
  3f00799
- Fix tests
  janEbert
  committed
  0dc265d
- To-be-reviewed: minor changes
  cyanguwa
  committed
  220c370
Commits on Oct 14, 2026
- Fix QKV layout
  janEbert
  committed
  9739216
- Use varlength tuples with type hints
  janEbert
  committed
  4029149
- Revert "Use varlength tuples with type hints"
  janEbert
  committed
  f9a2e77
- fix FA2 pad logic and should_pad overall logic
  cyanguwa
  committed
  0620e6d
- Merge branch 'main' into optional-mla-pad
  cyanguwa
  authored
  cc780bd
- Update tests/pytorch/attention/test_attention.py
  
  cyanguwa
  and
  greptile-apps[bot]
  authored
  d69ff67
Commits on Oct 21, 2026
- Improve cache clear usability
  janEbert
  committed
  f8455f1

Commits

Commits on Oct 12, 2026

Commits on Oct 13, 2026

Commits on Oct 14, 2026

Commits on Oct 21, 2026