All pull requests

Pull requests

• [FlyDSL] FP8 MQA logits improvements for gfx950

  #7412·vpietila-amd opened Nov 3, 2026Contributor·10 tasks·

  ·
  2

• [Config] MiniMax-M3 MXFP8: tune N=1280 K=6144 (DSpark context KV projection)

  Config
  #7411·micah-wil opened Nov 3, 2026Contributor·10 tasks·

  ·
  2

• [FlyDSL] perf: integrate APRE M256 kernels into gfx1250 MegaMoE S1F

  FlyDSL
  #7410·yanguahe opened Nov 3, 2026Contributor·3 of 10 tasks·

  ·
  2

• [FlyDSL] gfx942 a16wi4: faster int4 MoE dequant via v_perm_b32

  FlyDSL
  #7409·xaguilar-amd opened Nov 3, 2026Contributor·4 of 10 tasks·

  ·
  2

• [WIP] perf(gfx1250): integrate MiniMax-M3 fusion stack

  #7408·lirui927 opened Nov 3, 2026Contributor·

  ·

• [ASM] [HIP] [JIT] Add gfx1250 MLA v4.1 a8w4 fused sparse prefill kernel

  ASMHIPJIT
  #7407·tingchen988 opened Nov 3, 2026Contributor·

  ·
  2

• [Triton/Gluon] Step-5 CSA paged MQA logits kernel for gfx942

  ci:triton-300xTriton/Gluon
  #7406·akii96 opened Nov 3, 2026Contributor·5 tasks done·

  ·
  5

• perf(flydsl): fuse TP MoE GEMM2 weighted reduction

  #7405·lirui927 opened Nov 3, 2026Contributor·

  ·

• test(gemm): document gfx1250 F2 split-K reduction result

  #7404·lirui927 opened Nov 3, 2026Contributor·

  ·

• [Triton/Gluon] [gfx1250] Add a16w16 MoE GEMM

  Triton/Gluon
  #7403·xiaohuguo2023 opened Nov 3, 2026Member·7 tasks done·

  ·
  8

• [FlyDSL] Remove TP prefill MoE route counter fill

  #7402·lirui927 opened Nov 3, 2026Contributor·

  ·

• feat(flydsl): fuse sparse PA output quantization

  #7401·lirui927 opened Nov 3, 2026Contributor·

  ·

• feat(dist): lease gfx1250 custom AR input buffers

  #7400·lirui927 opened Nov 3, 2026Contributor·

  ·

• [ASM] [HIP] gfx942 mha bwd: hd192 x hdv {192, 128} 7-GEMM kernels, 64-bit odo addressing

  ASMHIP
  #7399·slippedJim opened Nov 3, 2026Contributor·5 of 10 tasks·

  ·
  17

• [Triton/Gluon] Fuse gfx1250 hidden-6144 AR RMSNorm quant

  #7398·lirui927 opened Nov 3, 2026Contributor·

  ·

• [HIP] Fold gfx1250 fused AR end barriers into kernels

  #7397·lirui927 opened Nov 3, 2026Contributor·

  ·

• [HIP] Allow gfx1250 LL quant with QuickReduce enabled

  #7396·lirui927 opened Nov 3, 2026Contributor·

  ·

• [ASM] [HIP] [JIT] Clamp bf16 MLA decode kv_indices and move gfx950 kernels to opus/gfx950/

  ASMHIPJIT
  #7395·kaiyang-1 opened Nov 3, 2026Contributor·

  ·
  5

• MHAv4 gfx942 parity

  #7394·nsakkine opened Nov 3, 2026Contributor·4 of 10 tasks·

  ·
  2

• [Triton/Gluon] gfx950 a16w16: Stream-K for compute-bound shapes (draft)

  #7391·jamesbowley opened Nov 3, 2026·

  ·

• [ASM] [HIP] Fold gfx1250 MLA v4.1 sparse prefill split-K in-kernel with workgroup clusters

  ASMHIP
  #7390·kaiyang-1 opened Nov 3, 2026Contributor·

  ·
  5

• [Config] [Perf] Add DeepSeek V4.1 EP16 A4W4 MoE tuning configs

  Config
  #7389·james-huang09 opened Nov 3, 2026Collaborator·

  ·
  2

• [FlyDSL] [Misc] pa_decode: key the tile kernel on FlyDSL dtypes, not dtype strings and flags

  #7388·fsx950223 opened Nov 3, 2026Contributor·4 of 10 tasks·

  ·

• [HIP] [OPUS] Add OPUS GEMM BlockScale_Mxfp8 GEMM for M <=1024 Shape.

  HIPOPUS
  #7387·Fyzyukk opened Nov 3, 2026Contributor·

  ·
  2

• [Bugfix] Allow gfx942 FlyDSL a16w4 and a16wi4 MoE tuning

  #7386·amd-oshkarav opened Nov 3, 2026·2 of 10 tasks·

  ·
  2