Repository Issue Activity (beta)

#262 Support request for latest CUDA + Python + Torch

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

#400 Support for PyTorch 2.14

#361 How to compile with Torch 2.11.0, Python 3.14, Linux?

#389 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)

#398 (on linux) SageAttention 2.2.0 isn't on pip and cannot compile it because of dependency downgrade hell

#344 sageattention-2.2.0+cu130torch2.9.1-cp313-cp313-win_amd64.whl

#321 sage-attention 3 mosaic artifacts and not faster than sage-attention 2.2 on dgx spark

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

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

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

#391 [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

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

#126 Circular Import when using sageattn

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

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

#383 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

#229 SageBwd Release

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

#176 Release SageAttention3 on Hugging Face

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

#221 Sage Attention with WAN FP8 model (or FP8 quantization) causes black output. So I'm stuck either doing FP16 with Sage but maxing VRAM or using FP8 but getting no Sage speedup

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

#373 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

#47 Only `head_dim==128` and `headdim==64` are supported?