All pull requests

Pull requests

• [Config] [MoE] [gfx950] Qwen3-235B FP8 EP: drop the token=1,2 tuned fmoe rows that lose to the CK 2-stage default

  Config
  #6681·afriedri opened Oct 19, 2026Contributor·

  ·
  1

• [Triton/Gluon] [Config] gfx950 a16w16 N=2880 K=4096: retune M_LEQ_4 to avoid Triton 3.8 VGPR spill (gpt-oss-20b o_proj decode)

  ConfigTriton/Gluon
  #6680·afriedri opened Oct 19, 2026Contributor·

  ·
  1

• [Triton/Gluon] [Bugfix] gfx1250 Gluon a8w8 blockscale: NUM_BUFFERS config key, clamped to the split

  Triton/Gluon
  #6679·Boss2002n opened Oct 19, 2026Contributor·

  ·
  9

  

• [Triton] Add fused_reduce_add_rmsnorm

  #6678·eky-amd opened Oct 19, 2026Contributor·1 task done·

  ·
  1

• [HIP] [Bugfix] Bound per-tensor quantization vector tails

  HIP
  #6677·yuzhouo7 opened Oct 19, 2026·2 tasks done·

  ·
  1

  2

• [Triton] gemm_a16w16: round to the output type before the activation (GELU 19-34% faster on gfx1151)

  #6676·klin2024 opened Oct 19, 2026·

  ·
  1

• [ASM] [HIP] [OPUS] Add gfx950 MLA v4.1 sparse prefill and move the opus MLA sparse prefill to ctypes

  ci:atom
  #6675·kaiyang-1 opened Oct 19, 2026Contributor·

  ·
  3

• [Triton] Add fused_gemm_a16w16_mul_add for the DiT gated residual

  #6674·klin2024 opened Oct 19, 2026·

  ·
  1

• [Triton/Gluon] [Config] [gfx1250] Tune DeepSeek-R1 high-batch attention GEMM tiles

  ConfigTriton/Gluon
  #6673·karverma-amd opened Oct 19, 2026Contributor·

  ·
  4

• [Triton/Gluon] Support BF16 state storage in gfx950 KDA

  Triton/Gluon
  #6672·YukioZzz opened Oct 19, 2026·

  ·
  4

  

• [Triton/Gluon] [Config] Add gfx1151 gemm_a16w16 DEFAULT.json

  #6671·klin2024 opened Oct 19, 2026·

  ·
  1

• [FlyDSL] Add fused QSA sparse decode for gfx950

  FlyDSL
  #6670·amd-nprotaso opened Oct 19, 2026·

  ·
  1

• [FlyDSL] Add gfx11 (RDNA3 / RDNA3.5) flash attention forward

  #6669·klin2024 opened Oct 19, 2026·

  ·
  1

• [FlyDSL] [Perf] MiniMax-M3 MoE on gfx1250: fused route psum and tuned grouped A8W4 tiles

  #6668·lirui927 opened Oct 19, 2026Contributor·

  ·
  3

• [ASM] update code objects

  ASM
  #6666·slippedJim opened Oct 19, 2026Contributor·1 task·

  ·
  2

• [JIT] [Bugfix] Round CK-Tile a8w8 GEMM bf16 outputs to nearest even on gfx942

  JIT
  #6665·siliangchen-amd opened Oct 19, 2026Contributor·4 tasks done·

  ·
  2

• [FlyDSL] [gfx1250] Fix the dropped K tail and the bias column offset in the grouped MoE TDM GEMM

  FlyDSL
  #6663·lalala-sh opened Oct 19, 2026Contributor·

  ·
  1

• [FlyDSL] Add gfx1250 implicit-GEMM conv3d

  #6662·huizzhan opened Oct 19, 2026Contributor·3 of 5 tasks·

  ·
  11

• [Triton/Gluon] Fix unified_attention gfx950 wrong output when packed KV cache is 2–4 GiB

  Triton/Gluon
  #6661·zhentaocc opened Oct 19, 2026Contributor·

  ·
  3

• feat(attention): per-token fp8 KV and split-K verify in gfx1250 unified attention

  Triton/Gluon
  #6660·lirui927 opened Oct 19, 2026Contributor·

  ·
  5

  

• [MoE] Cap the default fused_moe split-K at 15 (KBatch >= 16 gives wrong results)

  #6658·chuyeh opened Oct 19, 2026·1 task done·

  ·
  1

• [HIP] [Perf] scaled_silu_and_mul: split short batches across workgroups

  HIP
  #6656·AndreasKaratzas opened Oct 18, 2026Contributor·

  ·
  1

  

• [Triton/Gluon] [Config] Tune gfx1250 Gluon MXFP4 GEMMs for Llama3 405B

  ConfigTriton/Gluon
  #6654·Boss2002n opened Oct 18, 2026Contributor·

  ·
  2

  

• [Triton/Gluon] [Config] Tune gfx1250 Gluon MXFP4 GEMMs for Llama3 70B

  ConfigTriton/Gluon
  #6653·Boss2002n opened Oct 18, 2026Contributor·

  ·
  3

  

• [Triton/Gluon] [Config] Tune gfx1250 Gluon MXFP4 GEMMs for Llama3 8B

  ConfigTriton/Gluon
  #6652·Boss2002n opened Oct 17, 2026Contributor·

  ·
  2