All pull requests

Pull requests

• [Triton/Gluon] : Add exclude list arg and allow list/dict/tuple of tensors in generate_rotating_buffers_pool

  Triton/Gluon
  #7121·rahulbatra85 opened Oct 29, 2026Contributor·

  ·
  1

• [Triton/Gluon] gfx1250 fused Q/K Gemma RMSNorm + MRoPE + KV cache write for Qwen3.5

  Triton/Gluon
  #7120·Boss2002n opened Oct 29, 2026Contributor·

  ·
  

• [FlyDSL] Fix gfx1250 GEMM output offset overflow

  FlyDSL
  #7119·chandankumr opened Oct 29, 2026·

  ·
  1

  1

• [Triton/Gluon] gfx1250 fused Q/K Gemma RMSNorm + MRoPE for Qwen3.5

  Triton/Gluon
  #7118·Boss2002n opened Oct 28, 2026Contributor·

  ·
  1

  

• [Triton/Gluon] [Config] a4w4: dispatch triton_preshuffle tuned rows, tune Qwen3.8-27B decode GEMMs on gfx950

  #7117·vorapolsiloai opened Oct 28, 2026Contributor·

  ·

• [Triton/Gluon] act_mul_and_mxfp4_quant: honour scaling_mode, add "roundup" and round_to_input_dtype

  #7116·vorapolsiloai opened Oct 28, 2026Contributor·

  ·
  1

• [Triton/Gluon] [Config] [gfx1250] Tune split-K for mhc_fused_post_pre_delayed_rmsnorm decode

  ConfigTriton/Gluon
  #7115·karverma-amd opened Oct 28, 2026Contributor·

  ·
  1

• [HIP] [Perf] fn-resident mHC pre GEMM and bf16 hi/lo split for gfx950 prefill

  #7114·anhminhnguyenhoang opened Oct 28, 2026Contributor·

  ·
  1

• [Triton] [gfx1250] sampling: register-resident top-p for the top-k-first fast path

  #7113·azaidy opened Oct 28, 2026Contributor·

  ·

• Alizaidy/gfx1250 qwen support

  #7111·Boss2002n opened Oct 28, 2026Contributor·1 task·

  ·
  1

• Fix for num_local_tokens

  #7110·JohnNikolay84 opened Oct 28, 2026Contributor·1 task·

  ·
  1

  

• [Config] Add Muse Spark Open 816B MXFP4 fused-MoE FlyDSL configs for gfx950 (MI350P, 128 CUs)

  Config
  #7109·mpashkovskii opened Oct 28, 2026Contributor·

  ·
  3

• [Feature] Allow explicit Even activation-scale selection for MXFP4 MoE

  #7108·XinyuJiangCMU opened Oct 28, 2026Contributor·

  ·
  4

• [Triton/Gluon] [gfx1250] Fused GDN speculative-verify step on ReplaySSM state

  #7107·ahmed-bsod opened Oct 28, 2026Contributor·

  ·
  1

• [Tuning] Skip run_config tests that have no tuned rows for the device

  #7106·RElbers opened Oct 28, 2026Contributor·

  ·
  1

• [Triton/Gluon] [CI] Configs per CU count: <arch>/<cu_num>cu/ trees

  CITriton/Gluon
  #7105·Boss2002n opened Oct 28, 2026Contributor·

  ·
  7

• [FlyDSL] moe_reduce: guard against -1 topk_ids padding sentinel

  FlyDSL
  #7104·rjrock opened Oct 28, 2026Contributor·1 task·

  ·
  1

• [JIT] Add -mllvm -ignore-tti-inline-compatible=true for the 10.1 LLVM compiler to avoid some performance loss in ROCm 10.1

  JIT
  #7103·rasmith opened Oct 28, 2026Contributor·1 task done·

  ·
  1

  1

• [CK] [Feature] MXFP4 MoE tuner: propose the native BK=128 gemm2

  CK
  #7102·yuychang opened Oct 28, 2026Contributor·1 task done·

  ·
  4

• [HIP] [Bugfix] Accept generic torch.Stream in torch_to_c_types

  ci:sglangHIP
  #7101·yuychang opened Oct 28, 2026Contributor·1 task done·

  ·
  2

• [Triton/Gluon] [GFX12] Add SPEC_DECODE mode to UA with tunned Qwen35 config

  Triton/Gluon
  #7100·k50112113 opened Oct 28, 2026Contributor·

  ·
  5

• [FlyDSL] [gfx1250] FMHA prefill m32x8: skip query tiles past a varlen sequence's end

  FlyDSL
  #7098·vgokhale opened Oct 28, 2026Contributor·

  ·
  1

• MHAv4: temporarily disable rotated_fp8_quantization_matches_reference on gfx942

  #7097·jcaraban opened Oct 28, 2026Contributor·

  ·
  1

• [Triton/Gluon] [CI] Load GEMM configs from <arch>/<cu_num>cu/ first

  CITriton/Gluon
  #7096·jiaryang opened Oct 28, 2026Contributor·

  ·
  22

  

• [FlyDSL] Reduce gfx1250 TP MoE routing contention with workgroup histograms

  FlyDSL
  #7095·xytpai opened Oct 28, 2026Contributor·

  ·
  1