All pull requests

Pull requests

• epilogue: benchmark suite for the gfx950 GEMM-epilogue family (stacked on #90)

  #92·AviralGoelAMD opened Oct 6, 2026Contributor·

  ·

• Support for 9070XT (gfx1201/rdna4)

  #91·ryanswann-amd opened Oct 1, 2026·

  ·

• GEMM epilogue fusion for gfx950: templated bf16 GEMM + 11 epilogues + rms_reduce + hk API

  #90·AviralGoelAMD opened Sep 29, 2026Contributor·

  ·
  1

• gfx942 FP8 MoE Stage 1 kernel

  #89·Parxd opened Sep 27, 2026Contributor·

  ·

• gfx1250 multicast-free GEMM ladder (non-multicast cluster path)

  #85·bethune-bryant opened Aug 27, 2026·3 tasks·

  ·

• Use builtins for loads and async marks on CDNA4.

  #79·carlobertolli opened Aug 18, 2026·

  ·

• blockwise fp8 gemm

  #76·asdfvg123 opened Jul 30, 2026·

  ·

• Lane-ordering for MXFP8 GEMM scales

  #75·alextmagro opened Jul 28, 2026Collaborator·

  ·

• Add gfx1250 st_16x64 shared tile shape

  #71·RattataKing opened Jul 7, 2026·

  ·
  1

• Track header dependencies in kernel builds (-MMD -MP)

  #66·AviralGoelAMD opened Jun 23, 2026Contributor·

  ·

• Guard unsupported shapes in base GEMM dispatch (M,N % 256; K % 128)

  #65·AviralGoelAMD opened Jun 23, 2026Contributor·

  ·

• Fix unsatisfiable bf16 GEMM test tolerance (fp32 ref + atol floor)

  #64·AviralGoelAMD opened Jun 23, 2026Contributor·

  ·

• Add FP4 tile + vector types and memory operations

  #54·ccs1112 opened Apr 19, 2026·

  ·
  1

  2

• add CDNA3 flash attention forward kernels

  #33·Jayluci4 opened Feb 23, 2026·4 tasks done·

  ·
  1