An out-of-bound issue is fixed, which caused black or noise outputs, see #98

Head dim 256 is supported. It's needed for Ideogram 4, see thu-ml#329

If you feel brave, you can try https://github.com/woct0rdho/sageattention-autotune and see if it's faster.

Starting from this version, the wheels are built with bitwise reproducibility and attestation. This helps prevent supply chain attacks.

A big change in this version is that it supports torch.compile, see thu-ml#218 . Previously SageAttention must be put inside torch.compiler.disable and cause a graph break, now it's no longer the case.

To enable full graph compilation in ComfyUI, choose allow_compile in PatchSageAttentionKJ (or sageattn_compiled in WanVideoModelLoader), and choose fullgraph in TorchCompileModelAdvanced (or WanVideoTorchCompileSettings).

Note that we only recommend to apply 'full graph compilation' to transformer blocks, and treat each block as a 'full graph'. The TorchCompileModel node in native ComfyUI usually fails, because it tries to compile the whole model pipeline rather than the blocks.

There are reports that full graph compilation solves some black/noise outputs and lowers VRAM usage. However, it can fail or the compiling time can be very slow, if it's not correctly configured or it's applied to some complicated models. There are also reports that it causes some new black/noise outputs. Developers from multiple projects, including PyTorch, Triton, SageAttention, ComfyUI, Kijai and other custom nodes, are still working to make it easier.

Recently ComfyUI introduced comfy-aimdo but it breaks full graph compilation. You may disable it with --disable-dynamic-vram when starting ComfyUI.

If full graph compilation fails, you can still use this version of SageAttention without it.

Another change is that along with ABI3, it uses libtorch stable ABI. There is torch2.9.0andhigher in the filename (I hope we can come up with a better name scheme), so the wheel supports PyTorch >= 2.9 . Note again that cp39-abi3 means it uses Python ABI3 and supports all Python versions >= 3.9 .

(This is not SageAttention3. For SageAttention3 you can download the wheel at https://github.com/mengqin/SageAttention/releases , and see the discussion at #42 (comment) )

Fix GQA case for smooth_k, see thu-ml#252

Previously the SageAttention2++ kernels may not correctly fallback to the old SageAttention2 kernels on RTX 40xx and CUDA < 12.8 . Now it's fixed, see #46

The wheels for PyTorch 2.9 are published. CUDA 13.0 is supported since PyTorch 2.9 .

Some improvements are added to the Triton kernels, including better precision of quantized qk ( thu-ml#224 ), and attn_mask ( thu-ml#227 ).

The CUDA kernels for sm75 are removed. In post1 they caused errors when dispatching bf16, see #29 . We may need to build a separate binary for sm75 when we do this again.

RTX 20xx (sm75) is still supported and it will run the Triton kernels.

Now the wheels are built with the Python Stable ABI (also known as ABI3). This means we no longer need to build a different wheel for every Python version (although in principle we still need to build for every PyTorch and CUDA version). Later I'll make a PR to the official SageAttention repo.

There is cp39-abi3 in the wheel filenames, so they support Python >= 3.9 .

For RTX 20xx (sm75), this wheel will run the Triton kernels. The CUDA kernels for sm75 are bundled for testing but they do not fully work yet.

The wheel for PyTorch 2.8 is stable now.

SageAttention2++ kernels (sv_f8_accum_f16) are added. Compared to SageAttention 2.1, they improve the speed with almost no quality loss. They only support RTX 40xx (sm89) and 50xx (sm120) GPUs, and CUDA >= 12.8, therefore PyTorch >= 2.7 .

The SageAttention 2.2 wheels contain both the SageAttention2++ kernels and the old SageAttention2 kernels. On older GPU and CUDA, the old SageAttention2 kernels will be used, so it will still run but without the speedup.

For PyTorch 2.8, the nightly wheels are unstable, so the SageAttention wheels here may not work with the torch nightly wheel on any day. They're only tested with torch 2.8.0.dev20250627 .

The instructions to install are moved to the README: https://github.com/woct0rdho/SageAttention . Now the release page only contains changelogs.

For PyTorch 2.8, the nightly wheels are unstable, so the SageAttention wheels here may not work with the torch nightly wheel on any day. They're only tested with torch 2.8.0.dev20250415 .