Issues

• sageattn() silently drops smooth_k (and every other kwarg): ComfyUI's smooth_k=False never reaches the kernel

  thu-ml/SageAttention#404 · Filter by author JoaoZaokk opened on Oct 2, 2026

• MiniMax H3 Memory Efficient SageAttention hard-crashes on NVIDIA L20 / SM89 with FP8 path; FP16 SageAttention works

  thu-ml/SageAttention#403 · Filter by author thr1f1sh opened on Sep 9, 2026

• Support for PyTorch 2.14

  thu-ml/SageAttention#400 · Filter by author cardamon opened on Sep 5, 2026
  1

• win10-CUDA-12.6/13.0 Python-3.12.10/3.13.12：Successfully installed sageattention-2.2.0

  thu-ml/SageAttention#393 · Filter by author Dao28 opened on Aug 25, 2026

• sm120 (Blackwell): sageattn produces silently-wrong output when a captured CUDA graph is replayed with new input values (no NaN)

  thu-ml/SageAttention#392 · Filter by author keeganmccallum opened on Aug 18, 2026
  2

• [Bug][sm_120][Windows] Driver-level GPU lost during MiniMax H3 generation with int8-QK/FP8-PV kernel on RTX 5060 Ti; native attention control passes at higher load

  thu-ml/SageAttention#391 · Filter by author leimuxiao opened on Aug 16, 2026
  2

• gfx12 int8 kernel: fp16-V slower than fp16 sageattn at long sequences; fp8-V wins micro-benchmarks but no real-workload gain (RX 9070 XT)

  thu-ml/SageAttention#389 · Filter by author Newaiguy opened on Aug 15, 2026
  2

• [Bug] FP8 PV kernel produces silent noise corruption above ~160k sequence length on sm_120 (RTX PRO 6000 Blackwell)

  thu-ml/SageAttention#388 · Filter by author anitman opened on Aug 14, 2026
  1

• int32 overflow in Triton kernels silently corrupts output (and sometimes crashes) for non-contiguous inputs at long sequence lengths

  thu-ml/SageAttention#386 · Filter by author juju812 opened on Aug 11, 2026
  1

• [SM89][Windows] Driver reset on very long HND attention; bounded-memory workaround

  thu-ml/SageAttention#384 · Filter by author Axcyn opened on Aug 8, 2026

• per_channel_fp8 pads kv_len to 64 while sm90 kernel asserts pad to CTA_K=128 — direct callers of _qattn_sm90 abort at runtime

  thu-ml/SageAttention#383 · Filter by author buaasun opened on Aug 7, 2026

• SageAttention3 Blackwell per-block mean fails to initialize TMA descriptor at ~74k tokens on sm_120

  thu-ml/SageAttention#382 · Filter by author avp1598 opened on Aug 5, 2026
  1

• SageAttention3 (Blackwell FP4): low speedup on RTX PRO 5000 (sm_120, 110 SMs)

  thu-ml/SageAttention#378 · Filter by author CV-poo opened on Jul 21, 2026

• Hasn‘t the implementation of int4 qk in sageattention2 been given?

  thu-ml/SageAttention#374 · Filter by author xinglzx opened on Jun 30, 2026

• 一直去不掉宏污染错误 , 2.2可以编译成功 , 但是sageatten3在comfyui 0.25.1怎么都编译不了 , 怎么解决

  thu-ml/SageAttention#373 · Filter by author DarkNightMeteor opened on Jun 22, 2026

• Sageattention 3 reduce the generated video quality (Wan2.2 model) but yes it is much faster

  thu-ml/SageAttention#369 · Filter by author kananaestate opened on May 19, 2026

• [Bug Report] Video Temporal Flickering when using SageAttention in ComfyUI

  thu-ml/SageAttention#367 · Filter by author voidyear0320 opened on May 12, 2026

• Private security contact for SageAttention vulnerability report

  thu-ml/SageAttention#366 · Filter by author igorsantosgrueiro opened on May 5, 2026

• Can we get modern install?

  thu-ml/SageAttention#363 · Filter by author Dunc4n1dah0 opened on Apr 22, 2026
  1

• How to compile with Torch 2.11.0, Python 3.14, Linux?

  thu-ml/SageAttention#361 · Filter by author noobmaster777 opened on Apr 11, 2026
  3

• _qattn_sm80 extension fails on Ada Lovelace (sm89/L40S) — illegal instruction at runtime or compilation failure

  thu-ml/SageAttention#360 · Filter by author ShaiDiamant opened on Apr 10, 2026

• pytorch2.11.0+cuda13.0 5090无法编译

  thu-ml/SageAttention#359 · Filter by author tonylee2022 opened on Apr 7, 2026

• Supporting B300 sm103

  thu-ml/SageAttention#358 · Filter by author asbevz opened on Mar 29, 2026

• [Bug] sageattn3_blackwell: CUDA error: misaligned address on RTX 5060 Ti (Blackwell)

  thu-ml/SageAttention#357 · Filter by author gzsiang opened on Mar 29, 2026
  3

• implementation for sageattn_qk_int8_pv_int8_cuda

  thu-ml/SageAttention#356 · Filter by author AIgiraffe opened on Mar 26, 2026