All pull requests

Pull requests

• [FlyDSL] Add opt-in gfx950 D256 wave-local PA pipeline with per-token KV

  #6646·fsx950223 opened Oct 17, 2026Contributor·

  ·
  1

• Use HIP to query XCC count and share it with FlyDSL PA

  ci:triton-300xci:triton-355FlyDSLTriton/Gluon
  #6645·fsx950223 opened Oct 17, 2026Contributor·

  ·
  2

  

• feat(moe): support MXFP8 experts on gfx1250

  #6642·XiaobingSuper opened Oct 17, 2026Contributor·

  ·
  1

• [HIP] [Bugfix] scaled_silu_and_mul: device-side scale, fp8 saturation, row bounds

  HIP
  #6641·AndreasKaratzas opened Oct 17, 2026Contributor·1 task·

  ·
  1

• [JIT] Round custom all-reduce BF16 stores to nearest even

  JIT
  #6640·AndreasKaratzas opened Oct 17, 2026Contributor·1 task·

  ·
  1

• [ASM] [HIP] Fix wrong results in gfx942 fp8 persistent / blockscale fused-MoE dispatch

  ASMHIP
  #6639·AndreasKaratzas opened Oct 17, 2026Contributor·1 task·

  ·
  1

• [Config] Add gfx950 128-CU A4W4 GEMM tunings for Qwen3.8-27B (MI350P)

  Config
  #6638·fyun89 opened Oct 17, 2026·3 of 6 tasks·

  ·
  1

• [Triton/Gluon] : Add rot_bufs and do_bench_aiter_triton support for some attn benchmarks

  Triton/Gluon
  #6637·rahulbatra85 opened Oct 17, 2026Contributor·1 task·

  ·
  9

• [Triton/Gluon] : Use rot_bufs and do_bench_aiter_triton for bench_rmsnorm,layernorm and rope

  Triton/Gluon
  #6636·rahulbatra85 opened Oct 16, 2026Contributor·1 task·

  ·
  5

• [Triton/Gluon] : Use rot_bufs and do_bench_aiter_triton in bench_fused_*

  Triton/Gluon
  #6635·rahulbatra85 opened Oct 16, 2026Contributor·1 task·

  ·
  8

• [Triton/Gluon] satya/gluon_fused_qk_norm

  Triton/Gluon
  #6634·Boss2002n opened Oct 16, 2026Contributor·

  ·
  10

  

• [Triton/Gluon] : Remove bad mla_decode.py op

  Triton/Gluon
  #6633·rahulbatra85 opened Oct 16, 2026Contributor·1 task·

  ·
  2

• Draft 1016

  ASMFlyDSLHIPTriton/Gluon
  #6631·akao-amd opened Oct 16, 2026Contributor·1 task·

  ·
  7

• [Triton/Gluon] [Bugfix] Fix gfx1250 gemm_a16w16 dropping output rows past 2**31 bytes

  Triton/Gluon
  #6629·yichiche opened Oct 16, 2026Contributor·3 tasks done·

  ·
  3

• Optimize mxfp4 gfx1250 gluon

  #6628·azaidy opened Oct 16, 2026Contributor·

  ·
  8

• [Triton/Gluon] [gfx1250] paged MQA logits: 8-token pages; enable group32 FP8 GEMM

  Triton/Gluon
  #6626·yhl-amd opened Oct 16, 2026Contributor·3 of 4 tasks·

  ·
  3

• [Triton] Harden shuffled per-token FP8 paged decode

  #6625·XiaobingSuper opened Oct 16, 2026Contributor·

  ·
  1

• fix(dist): log MORI all2all handle creation

  #6624·JiaoliangYu opened Oct 16, 2026Contributor·

  ·
  1

• [Config][GEMM][gfx950] Qwen3.8-27B ViT BF16 GEMM shapes (tuning pending)

  #6623·vorapolsiloai opened Oct 16, 2026Contributor·2 of 5 tasks·

  ·
  1

• [Config] Add v41 tp2 moe config for shared expert fusion on 8k/16k input

  Config
  #6622·1am9trash opened Oct 16, 2026Member·1 task·

  ·
  1

• [HIP] [FlyDSL] [JIT] [gfx1250] Add MXFP4 a4w4 GEMM with tuned DSR1 configs

  FlyDSLHIPJIT
  #6621·aoli26 opened Oct 16, 2026Contributor·1 task done·

  ·
  7

• [MORI] Manage v2 CCO through the all2all manager

  #6618·JiaoliangYu opened Oct 16, 2026Contributor·

  ·
  1

• [FlyDSL] Extend D256 PA decode pipe to transposed V and packed/NHD sliding windows

  #6617·fsx950223 opened Oct 16, 2026Contributor·

  ·
  3

• [MHA] Fix -DONLY_FAV3=0 quoting that breaks the backward JIT build

  #6616·menglcai opened Oct 16, 2026·1 task done·

  ·
  1

• Fusing dvdk and dq with split-K, using pingpong pipeline

  #6614·hellozhuo-amd opened Oct 16, 2026Contributor·

  ·