All pull requests

Pull requests

• Flydsl bwd c dropout

  cla signed
  #585·amd-weisun opened Oct 27, 2026Collaborator·

  ·

• Flydsl bwd b deterministic window value dim

  cla signed
  #584·amd-weisun opened Oct 27, 2026Collaborator·

  ·

• Flydsl bwd a dispatch gfx950

  cla signed
  #583·amd-weisun opened Oct 27, 2026Collaborator·

  ·
  1

• [Perf][ROCm][backward attention]Launch FlyDSL gfx950 backward CTAs in n_tile-major order

  cla signed
  #582·amd-weisun opened Oct 26, 2026Collaborator·

  ·
  1

• Fix paged length bounds, CUDA graph split-merge workspaces and NaNs, and add the paged test matrix

  cla signed
  #581·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Consolidate the paged decode routes and replace the FlyDSL paged tuning knobs with constants

  cla signed
  #580·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Trade paged splits for chain length and fix the paged decode routes

  cla signed
  #579·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Make FlyDSL window_left follow the flash-attention convention

  cla signed
  #578·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Add the single-launch pa_decode_swa paged decode kernel and its routing

  cla signed
  #577·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Fuse the paged split-K combine and narrow paged decode Q tiles

  cla signed
  #576·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Speed up paged KV loading and split-K occupancy in the generic FlyDSL kernel

  cla signed
  #575·RichardChamberlain1 opened Oct 23, 2026Contributor·

  ·
  1

• Serve f4f4bf16 from aiter on ROCm where it is faster

  cla signed
  #574·aryaman-gupta opened Oct 23, 2026Contributor·

  ·
  1

• Add shared infrastructure for serving MSLK ops from aiter per shape

  cla signed
  #573·aryaman-gupta opened Oct 23, 2026Contributor·

  ·
  1

• Fix f4f4bf16 on ROCm when K is not a multiple of the K tile

  cla signed
  #572·aryaman-gupta opened Oct 23, 2026Contributor·

  ·
  1

• Add aiter as an optional, pinned ROCm dependency

  cla signed
  #570·aryaman-gupta opened Oct 23, 2026Contributor·

  ·
  1

• Speed up FlyDSL paged attention decode

  cla signed
  #569·RichardChamberlain1 opened Oct 22, 2026Contributor·

  ·
  1

• Blackwell FMHA: fix NaNs, hang and bad masking for empty rows

  cla signedmeta-exported
  #553·bottler opened Oct 15, 2026Contributor·

  ·
  1

• Remove dead int32-KV branch from both split-K decode ops

  cla signed
  #545·aryaman-gupta opened Oct 9, 2026Contributor·3 of 4 tasks·

  ·
  2

• Memoise paged launch metadata on the attention bias

  cla signed
  #535·RichardChamberlain1 opened Sep 30, 2026Contributor·

  ·
  3

• Suppress type errors for Pyre upgrade - mslk

  cla signedmeta-exported
  #532·maggiemoss opened Sep 29, 2026Contributor·

  ·
  1

• FlyDSL paged decode: maintain MFMA occupancy on short-query workload

  cla signed
  #531·RichardChamberlain1 opened Sep 29, 2026Contributor·

  ·
  1

• Skip dispatch reporting during torch.compile (#530)

  cla signedmeta-exported
  #530·amolagrawalmeta opened Sep 28, 2026·

  ·
  1

• Stop the CUDA conda-build CI job from OOMing the worker

  cla signedmeta-exported
  #511·sryap opened Sep 4, 2026Contributor·

  ·
  1

• Autotune async copy for rowwise FP8 GEMM (#494)

  cla signedmeta-exported
  #494·warrendeng opened Aug 21, 2026·

  ·

• Add Triton FP8×INT4 rowwise GEMM for ROCm gfx942/gfx950

  cla signedmodule: rocm
  #479·apicciau opened Aug 11, 2026Contributor·

  ·