All pull requests

Pull requests

• [Config] Qwen3.8-Flash-Next FP8 tuned configs for gfx950 (MI350X, 256 CUs, TP1)

  Config
  #5869·mavizao opened Sep 26, 2026·

  ·
  1

• [Triton] unified_attention 2D: optional block skipping (BLASST)

  #5868·amd-sonsingh opened Sep 26, 2026·

  ·
  1

• [Triton/Gluon] DSR1 GEMM tuning

  Triton/Gluon
  #5867·azaidy opened Sep 26, 2026Contributor·

  ·
  7

  

• Omuhamma/prefill kda

  #5866·omuhamma opened Sep 26, 2026Contributor·

  ·
  6

• [CI] Preserve install failure status after retries

  CI
  #5864·vgokhale opened Sep 25, 2026Contributor·

  ·
  1

• [Triton/Gluon] Expert-parallel support for moe_gemm_a16w4

  Triton/Gluon
  #5863·Rohan138 opened Sep 25, 2026Contributor·

  ·
  4

  

• [Triton/Gluon] Unify afp8wfp8 GEMM kernels

  Triton/Gluon
  #5862·azaidy opened Sep 25, 2026Contributor·

  ·
  8

  

• [Triton/Gluon] [Bugfix][MLA] Follow-up: fix stale comments + test hardening (#5648)

  Triton/Gluon
  #5860·Rohan138 opened Sep 25, 2026Contributor·2/2·

  ·
  2

  

• [Triton/Gluon] PA decode: config-driven v1/v2 dispatch, gfx950 always v2 wh…

  Triton/Gluon
  #5859·NimitPtl opened Sep 25, 2026Contributor·

  ·
  4

• [Triton/Gluon] MHA fwd: gfx950 mid_head config for 64 < head_dim <= 128

  Triton/Gluon
  #5858·NimitPtl opened Sep 25, 2026Contributor·

  ·
  3

• [Triton/Gluon] gmm: gfx950 large-K/N config with JSON-driven shape dispatch

  Triton/Gluon
  #5857·NimitPtl opened Sep 25, 2026Contributor·

  ·
  5

• [MORI] Multi-node + MORI_V2

  #5855·k50112113 opened Sep 25, 2026Contributor·

  ·
  1

  

• [Triton/Gluon] rename tuning files

  Triton/Gluon
  #5854·Boss2002n opened Sep 25, 2026Contributor·

  ·
  4

  

• [HIP] [CK] [FlyDSL] Tune DSV4.1 TP2 A4W4 fused MoE for gfx950

  CKFlyDSLHIP
  #5852·yifehuan opened Sep 25, 2026Contributor·

  ·
  1

  

• [OPUS] Tune MiMo MXFP4 MOE kernel perf

  OPUS
  #5851·sammysun0711 opened Sep 25, 2026Contributor·

  ·
  5

  

• [Triton/Gluon] [Config] Route GLM-5 qkv_a_proj decode tiers to the triton split-K kernel on gfx950

  ConfigTriton/Gluon
  #5850·xiaofei-zheng opened Sep 25, 2026Contributor·

  ·
  5

  

• [MHA] Add a delta race, resume, finalist rounds and evidence to the MHA forward tuner

  #5849·amd-bartgips opened Sep 25, 2026·

  ·

• Fix the hardcoded die and CU counts in the opus kernels

  #5848·RElbers opened Sep 25, 2026Contributor·3/3·

  ·

• [CK] blockscale split-K: pick the pipeline from the per-split loop count

  CK
  #5847·siliangchen-amd opened Sep 25, 2026·

  ·
  1

  

• [Tuning] Share mp_tuner typed statuses and add a central tuning policy

  #5846·amd-bartgips opened Sep 25, 2026·

  ·
  1

• [ASM] [HIP] [FlyDSL] Add nhead=96 and round-robin DCP to persistent MLA PS1 FP8 decode for GFX1250

  ASMFlyDSLHIPJIT
  #5844·xiangM99 opened Sep 25, 2026Member·

  ·
  1

• [tuner] Fail a task as soon as its worker process exits

  #5841·siliangchen-amd opened Sep 25, 2026·

  ·
  1

  1

  

• [Config] Add gfx942 a8w8 blockscale GEMM tunings for Qwen3/Qwen3.5/GLM/DSV4 shapes

  Config
  #5839·siliangchen-amd opened Sep 25, 2026·

  ·
  1

  

• fix: ci comm kernel ut hang

  ci:multi-gpumultigpu
  #5837·TennyWang1223 opened Sep 25, 2026Contributor·

  ·
  1

  

• [Config] Add Qwen3.8-27B gfx950 A4W4 GDN BA projection shapes

  Config
  #5836·vorapolsiloai opened Sep 25, 2026Contributor·

  ·
  1