All pull requests

Pull requests

• [Triton/Gluon] Mask OOB e8m0 scales on the ragged-K path of fused_gemm_afp4wfp4_mul_add and split_cat

  Triton/Gluon
  #7281·Ragua1 opened Nov 1, 2026Contributor·3 of 10 tasks·

  ·
  2

• [Perf] mha: flash_attn_varlen_fp8_pertensor_func writes into a caller's output buffer

  #7280·vorapolsiloai opened Nov 1, 2026Contributor·4 of 10 tasks·

  ·
  2

• [Perf] act_mul_and_mxfp4_quant: Round to E2M1 with the gfx950 instruction

  #7279·vorapolsiloai opened Nov 1, 2026Contributor·3 of 10 tasks·

  ·
  2

• [Triton] MHA fp8 prefill: run the fp16 kernel on archs with software-emulated fp8

  Triton/Gluon
  #7277·Ragua1 opened Nov 1, 2026Contributor·4 of 10 tasks·

  ·
  8

• [gfx1250] MXFP8 sparse MLA attention kernel for DSV4.1 (FP8 KV cache)

  #7276·yhl-amd opened Nov 1, 2026Contributor·

  ·
  2

• [FlyDSL] [Perf] Use wave32 ballots for small persistent expert lookup

  FlyDSL
  #7275·xytpai opened Nov 1, 2026Contributor·

  ·
  4

• [FlyDSL] [Test] Honor the requested grouped MoE activation format

  FlyDSL
  #7274·xytpai opened Nov 1, 2026Contributor·

  ·
  2

• [FlyDSL] optimize dsv4 flash perf

  FlyDSL
  #7272·binding7012 opened Nov 1, 2026Contributor·

  ·
  2

• [FlyDSL] Cache HERD host launches and retire legacy BMM entry points

  FlyDSL
  #7271·coderfeli opened Nov 1, 2026Collaborator·4 of 10 tasks·

  ·
  2

• [Triton] gfx1151: tune gemm_a8w8_preshuffled configs (int8, 1.24x vs main)

  #7270·klin2024 opened Nov 1, 2026Contributor·

  ·
  2

• [FlyDSL] Generalize gfx1250 PA wave query tiles to two through four

  #7269·zhiding512 opened Nov 1, 2026Contributor·4 of 6 tasks·

  ·

• [Triton] ep_sort_routing: grid size at run time, no compile per prefill shape

  #7268·Boss2002n opened Nov 1, 2026Contributor·

  ·
  

• [Triton/Gluon] [gfx1250] a4w4 MoE configs for Qwen3.5-35B-A3B at EP4

  #7267·Boss2002n opened Nov 1, 2026Contributor·

  ·
  2

  

• [Triton] ep_sort_routing: barrier before the last-CTA ticket (scratch race on gfx1250)

  #7266·Boss2002n opened Nov 1, 2026Contributor·

  ·
  2

  

• [HIP] perf(custom_ar): gfx1250 LL all-gather for messages up to 4 MiB a rank

  HIP
  #7265·XiaobingSuper opened Nov 1, 2026Contributor·

  ·
  4

• [FlyDSL] [JIT] optimize mega moe tp for small token

  FlyDSLJIT
  #7264·binding7012 opened Nov 1, 2026Contributor·

  ·
  2

• [FlyDSL] [Fix] Use legal buffer access widths for vec8 MoE gather-reduce

  FlyDSL
  #7263·xytpai opened Nov 1, 2026Contributor·

  ·
  3

• [FlyDSL] [Perf] Avoid allocating an unread MoE quantizer placeholder

  FlyDSL
  #7262·xytpai opened Nov 1, 2026Contributor·

  ·
  3

• [FlyDSL] [Perf] Cache immutable grouped MoE integer config parsing

  FlyDSL
  #7261·xytpai opened Nov 1, 2026Contributor·

  ·
  3

• [Triton/Gluon] [Config] Add gfx1150 gemm_a16w4 configs

  #7260·klin2024 opened Nov 1, 2026Contributor·

  ·
  2

• [Config] [Perf] [gfx950] MiniMax-M3 MoE tuned rows: TP2 a4w4 prefill re-tune, TP8 a16w4 inter 384

  #7259·vorapolsiloai opened Oct 31, 2026Contributor·

  ·
  2

• [FlyDSL] QuickAllReduce INT4: E5M2 group scale so groups above 480 no longer clip

  #7258·vorapolsiloai opened Oct 31, 2026Contributor·

  ·
  2

• [Triton/Gluon] MHA prefill: use the gfx1151 attn_fwd config on gfx1150

  Triton/Gluon
  #7257·Ragua1 opened Oct 31, 2026Contributor·3 of 10 tasks·

  ·
  3

• [Triton/Gluon] [gfx1250] fused_rmsnorm_add: row kernel, gemma_norm, move to fusions

  ci:triton-300xTriton/Gluon
  #7256·Boss2002n opened Oct 31, 2026Contributor·

  ·
  4

  

• [Triton/Gluon] [Config] [gfx950] unified_attention: num_stages 2 for the head-512 fp8 prefill tile

  #7255·amd-sriram opened Oct 31, 2026Contributor·4 of 7 tasks·

  ·
  2