All pull requests

Pull requests

• Add FP16 tensor-core fallback for int8_linear on devices without INT8 tensor cores

  #277·seningpd opened Oct 24, 2026·

  ·
  6

• Issue HIP INT8 attention K/V staging loads together

  #276·qweqweewqe7-create opened Oct 23, 2026·

  ·
  3

• Add ck.input_act with Metal swiglu/situ for Apple GPUs (depends on #270), based on main @ f8b7e97 (v0.2.40)

  #275·LXD-8 opened Oct 23, 2026·

  ·
  23

• Decline fp16-accumulate CUTLASS kernels on sub-sm_80 GPUs

  #272·potoior opened Oct 23, 2026·

  ·
  1

  5

• sol_attn: caller-supplied block lists (block-sparse attention)

  #271·kijai opened Oct 22, 2026Contributor·

  ·
  1

• Add Metal RoPE and AdaLN kernels for Apple GPUs (mps backend), based on main @ f8b7e97 (v0.2.40)

  #270·LXD-8 opened Oct 22, 2026·

  ·
  44

• Fix singleton strides in CUDA flash decode

  #269·curealb opened Oct 22, 2026·

  ·
  1

• Add closest_hit_on_mesh CUDA op

  #268·kijai opened Oct 22, 2026Contributor·

  ·
  1

• Add long key sequence tests for int8 attention

  #267·potoior opened Oct 22, 2026·

  ·
  1

• Compile for sm_86 and keep its native int<->float conversions

  #264·Many0therFunctions opened Oct 20, 2026·

  ·
  2

• Optimize HIP SVDQuant W4A4 kernels

  #262·0xDELUXA opened Oct 19, 2026Contributor·

  ·
  7

• cuda: mark attn_mask .none() on sage_sdpa and sage_sdpa_prequantized for nanobind < 2.11

  #256·fblissjr opened Oct 18, 2026·

  ·
  1

• Vectorize HIP rms_adaln and split-half rms_rope

  #254·qweqweewqe7-create opened Oct 17, 2026·

  ·
  8

• Route INT8 linear_input_act through the int8_linear custom op

  #253·qweqweewqe7-create opened Oct 17, 2026·

  ·
  4

• Add an mps backend: weight-only INT8 and W4A8/W6A8 linears for Apple GPUs

  #249·roblennon opened Oct 13, 2026·

  ·
  1

  11

• Linux CUDA runtime interoperability: matched shared CUDART and relocatable wheels

  #246·xmarre opened Oct 12, 2026·

  ·
  1

  4

• Fix source-wheel CMake cache invalidation for CUDA and Python toolchains

  #244·xmarre opened Oct 12, 2026·

  ·
  4

• CUDA Sol-Attn: deterministic overflow retry with exact-tail preservation

  #242·xmarre opened Oct 12, 2026·

  ·
  1

  10

• Add quad remesh kernels (CORE-607)

  #240·kijai opened Oct 10, 2026Contributor·

  ·
  1

• [Feat] [RDNA1] Add RDNA1 INT8 Attention backend separate from the existing HIP backend

  #238·Zakhrov opened Oct 10, 2026·

  ·
  28

• sm86 INT8 GEMM tuning: measured tile chooser, per-shape cfg table, and path logging

  #237·Many0therFunctions opened Oct 10, 2026·

  ·
  9

• Validate sol_attn binding arguments on CUDA like the HIP binding does

  #236·Many0therFunctions opened Oct 10, 2026·

  ·
  1

• Optimize NVFP4 and MXFP8 quantization in Triton backend

  #233·nukuihayu opened Oct 8, 2026·

  ·
  1

• Tune SM120 V quantization for large D128 inputs

  #231·Tokha233 opened Oct 1, 2026·

  ·
  4

• Tune unmasked SM120 D64 attention with smaller query tiles

  #227·Tokha233 opened Oct 1, 2026·

  ·
  7