All pull requests

Pull requests

• [Triton/Gluon] [Config] Tune gfx1201 unified attention 2d for D=512 for full attention and D=256 for Sliding Window (Gemma4 shapes) and fix D=1024 launches

  ConfigTriton/Gluon
  #5790·GodRishUniverse opened Sep 23, 2026·

  ·
  2

• Topk index score minimax m3

  #5788·demonsan opened Sep 23, 2026·

  ·
  23

• [HIP] Enable CK fmha LLC head grouping on RDNA (batch mode)

  HIP
  #5787·qiangpan2 opened Sep 23, 2026·

  ·
  1

• [JIT] Make CK fmha kernel generation follow GPU_ARCHS

  JIT
  #5786·qiangpan2 opened Sep 23, 2026·

  ·
  1

• [CI] review-pr: fix agent-timeout envelope so large-diff reviews do not spuriously time out

  CI
  #5785·zufayu opened Sep 23, 2026·

  ·
  

• [CI] Enable DeepSeek-V4-Pro 2P1D TP8 ATOM DI smoke cases

  CI
  #5784·avininjamay8 opened Sep 23, 2026·

  ·
  1

• [Config] Tune a8w8 bpreshuffle for Kimi-K3's KDA gate at tp4

  Config
  #5783·XiaobingSuper opened Sep 23, 2026·

  ·
  2

• Fp8 mqa logits gfx950 dsa tune

  #5782·EricKing626 opened Sep 23, 2026·

  ·
  1

• [HIP] [JIT] Remove unintended CK dependency from plain top-k

  HIPJIT
  #5779·UD-mmcminn opened Sep 23, 2026·

  ·
  1

• [OPUS] Exp/mx32 plain

  OPUS
  #5778·yzhou103 opened Sep 23, 2026·

  ·
  17

• MLA v4 nm: faster cross-split merge for small decode grids

  #5776·AMD-yanfeiwang opened Sep 23, 2026·

  ·
  1

• fix(custom_all_reduce): send the real IPC offset instead of 0

  #5775·zhou9402 opened Sep 23, 2026·

  ·
  2

• [Triton/Gluon] Add rotating-buffer GEMM A16W16 benchmark

  Triton/Gluon
  #5774·dyokelso opened Sep 23, 2026·

  ·
  6

• [SMI] Add opt-in summary plots and raw-sample timelines

  #5773·arakowsk-amd opened Sep 22, 2026·

  ·
  2

  

• [Triton/Gluon] Assert SPLIT_UNMASKED_LOOP is not used with sliding window

  Triton/Gluon
  #5771·amd-xavierwang opened Sep 22, 2026·

  ·
  5

  

• [CI] chore: add security scanning workflows

  CI
  #5770·haribabug opened Sep 22, 2026·

  ·
  2

  

• [HIP] Mask partial vectors in sampling kernels

  HIP
  #5767·UD-mmcminn opened Sep 22, 2026·

  ·
  1

  

• WORKAROUND - fix a1_scale out-of-bounds read in fmoe_fp8_blockscale_g1u1

  #5764·afriedri opened Sep 22, 2026·

  ·
  1

  

• Add gfx950 MXFP4 SiTUv2 FLAT 16x192/16x32 kernels and KK3 fake tuner …

  #5763·JohnNikolay84 opened Sep 22, 2026·

  ·
  1

• [MHA] Add tuned-config dispatch and delta-race tuner for packed varlen

  #5762·amd-bartgips opened Sep 22, 2026·

  ·

• [HIP] [OPUS] [JIT] Unify gfx950+gfx1250 MXFP4 MQA-logits

  HIPJITOPUS
  #5761·shay-li77 opened Sep 22, 2026·

  ·
  16

  

• [Bugfix][MLA] Size PS prefill partial buffers from the real token budget

  #5759·simondanielsson opened Sep 22, 2026·

  ·
  1

  1

  

• [Config] [Perf] [gfx950] Add MiniMax-M3-MXFP4 BF16 tuned GEMM configs

  Config
  #5758·amd-bishwoadhikari opened Sep 22, 2026·

  ·
  11

  

• Qwe38 flash mxfp4 asm

  #5757·JohnNikolay84 opened Sep 22, 2026·

  ·
  1

• [Config] Restore accurate Kimi-K3 A4W4 FMoE decode tiles

  Config
  #5756·amd-mghanimi opened Sep 22, 2026·

  ·
  2