All pull requests

Pull requests

• Add SM89/SM120 seqlens and paged KV decode path for CUDA graphs.

  #402·cuihangbin opened Sep 7, 2026·

  ·

• sageattn3 (Blackwell): GQA/MQA support, non-standard head dims, device-adaptive persistent grid, and in-place K mutation fix

  #401·Zeta-li opened Sep 7, 2026·3 tasks done·

  ·

• Add clean Blackwell MXFP8 D128 attention path

  #399·oy60 opened Sep 2, 2026·

  ·

• Fix Windows build with MSVC and CUDA 13

  #397·Evilander opened Aug 30, 2026·

  ·

• Fix SA3 long-sequence TMA stride overflow

  #394·258369977 opened Aug 26, 2026·

  ·
  1

• Use int64 element offsets in the Triton kernels

  #387·juju812 opened Aug 11, 2026·

  ·
  1

• Add ROCm (HIP) support for the Triton attention path

  #381·Scorp1o117 opened Aug 4, 2026·

  ·
  3

• Patch kernels launch - launch them in the current CUDA stream

  #380·Abusagit opened Jul 28, 2026·

  ·

• Fix hardcoded SM count (170) in Blackwell FP4 persistent kernel launch

  #379·sowmy86 opened Jul 22, 2026·

  ·
  1

  2

• fix: prevent NaN outputs for short/non-multiple sequence lengths via CUDA-level zero-fill

  #377·Onkitova opened Jul 17, 2026·

  ·

• feat: add Intel GPU (XPU) support via auto-round-lib dispatch

  #375·luoyu-intel opened Jul 6, 2026·

  ·

• Fix Sage2 causal attention for unequal qkv lengths

  #372·qqq-tao opened Jun 15, 2026·

  ·

• Launch sm90 attention and fused quant kernels on the current torch stream (CUDA-graph capture correctness)

  #371·shanai13 opened Jun 11, 2026·

  ·

• Add Windows support for SageAttention3 Blackwell

  #370·sepcnt opened May 29, 2026·

  ·
  1

• Add native gfx12 attention backend

  #368·jammm opened May 13, 2026·

  ·
  100

• fix: use num_stages=1 on ROCm to avoid Triton pipelining crash on RDNA4 (gfx12xx)

  #365·brosequist opened Apr 29, 2026·

  ·
  2

• Comprehensive fix for Windows MSVC build errors (C2872 std, LNK2001) and thread-safety

  #355·munder-sa opened Mar 20, 2026·

  ·

• Windows MSVC compile fix for SA2, SA3

  #352·loscrossos opened Feb 26, 2026·

  ·
  1

• add support of sparse:

  #350·STwangyingrui opened Feb 11, 2026·

  ·

• Add Windows build workflow for GitHub Actions

  #335·hust-wx opened Dec 25, 2025·

  ·

• SageAttention AMD version

  #332·eliotwang opened Dec 18, 2025·

  ·
  27

• Support head dim 256

  #329·woct0rdho opened Dec 13, 2025·

  ·

• Sage Attention supports minimal modifications when compiling MS Visual Studio on Windows.

  #323·mengqin opened Dec 8, 2025·

  ·
  5

• Add MSVC support to setup.py for Windows builds

  #293·sfinktah opened Nov 2, 2025·

  ·

• support TORCH_CUDA_ARCH_LIST and avoid link against libcuda.so at compile time

  #245·winggan opened Aug 24, 2025·

  ·
  2