[GH-ISSUE #147] Build managed SageAttention 2 wheels for Windows and Linux #47

New issue

Closed

opened 2026-10-26 22:19:40 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 22:19:40 +00:00

Owner

Originally created by @Kosinkadink on GitHub (Sep 1, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/147

Goal

Publish official SageAttention 2.2.0 as prebuilt Windows and Linux wheels from Comfy-Org infrastructure so ComfyUI users do not compile or hand-install it.

Scope

Build a companion comfy-kitchen-sageattention distribution that provides the upstream sageattention import namespace without coupling Torch ABI-sensitive extensions to the stable-ABI comfy-kitchen wheel.
Pin upstream tag v2.2.0 at commit eb615cf6cf4d221338033340ee2de1c37fbdba4a and verify the source archive sha256 2c8232be47d35d20f095b1e38a052c6aa7b79dfd8ae4ab733d4df99c301ca7d0 before building.
Inventory and ship the Apache-2.0 license and attribution. Record that the pinned upstream source has no NOTICE file.
Build Python 3.12 binary wheels for manylinux x86_64 and win_amd64 against Torch 2.13.0 + CUDA 13.0, covering upstream SM80, SM86, SM89, SM90, and SM120 dispatch.
Declare the matching Triton dependency on Linux and triton-windows dependency on Windows.
Verify wheel tags, package metadata, compiled members, external linkage, source identity, and import behavior in clean environments.
Publish only from a dedicated SageAttention release tag after the wheel jobs pass.

Acceptance criteria

pip install comfy-kitchen-sageattention selects a binary wheel on supported Windows and Linux hosts and never invokes a local compiler.
from sageattention import sageattn succeeds with the pinned Torch/CUDA/Triton line.
Every published wheel is traceable to the exact upstream commit, source digest, build toolchain, and packaging revision.
No unsupported architecture is advertised as executable, and consumers can authenticate the companion distribution separately from an unmanaged upstream install.

Originally created by @Kosinkadink on GitHub (Sep 1, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/147 ## Goal Publish official SageAttention 2.2.0 as prebuilt Windows and Linux wheels from Comfy-Org infrastructure so ComfyUI users do not compile or hand-install it. ## Scope - Build a companion `comfy-kitchen-sageattention` distribution that provides the upstream `sageattention` import namespace without coupling Torch ABI-sensitive extensions to the stable-ABI `comfy-kitchen` wheel. - Pin upstream tag `v2.2.0` at commit `eb615cf6cf4d221338033340ee2de1c37fbdba4a` and verify the source archive sha256 `2c8232be47d35d20f095b1e38a052c6aa7b79dfd8ae4ab733d4df99c301ca7d0` before building. - Inventory and ship the Apache-2.0 license and attribution. Record that the pinned upstream source has no NOTICE file. - Build Python 3.12 binary wheels for manylinux x86_64 and win_amd64 against Torch 2.13.0 + CUDA 13.0, covering upstream SM80, SM86, SM89, SM90, and SM120 dispatch. - Declare the matching Triton dependency on Linux and `triton-windows` dependency on Windows. - Verify wheel tags, package metadata, compiled members, external linkage, source identity, and import behavior in clean environments. - Publish only from a dedicated SageAttention release tag after the wheel jobs pass. ## Acceptance criteria - `pip install comfy-kitchen-sageattention` selects a binary wheel on supported Windows and Linux hosts and never invokes a local compiler. - `from sageattention import sageattn` succeeds with the pinned Torch/CUDA/Triton line. - Every published wheel is traceable to the exact upstream commit, source digest, build toolchain, and packaging revision. - No unsupported architecture is advertised as executable, and consumers can authenticate the companion distribution separately from an unmanaged upstream install.

bsakaguchi closed this issue

2026-10-26 22:19:41 +00:00

bsakaguchi commented

2026-10-26 22:19:41 +00:00

Author

Owner

@Kosinkadink commented on GitHub (Sep 1, 2026):

Implementation is in #148. The branch pins SageAttention v2.2.0 source and license provenance, builds Python 3.12 / Torch 2.13 / CUDA 13 binary wheels for Linux and Windows, and verifies package metadata plus each extension's cubin coverage.

Local evidence on RipperPC:

final Linux wheel: 38,397,300 bytes, sha256 004ad09adb225ec0971a90608b05de807a706acb0e2755f7776e9cf3b70ca132
_fused: SM80/86/89/90/120
_qattn_sm80: SM80/86/89/90/120
_qattn_sm89: SM89/90/120
_qattn_sm90: SM90
locked RTX PRO 6000 Blackwell SM120 causal and non-causal smoke passed
521 CPU tests passed, 1361 skipped; Ruff clean
exact final diff independently reviewed with no remaining blocker

Free GitHub manylinux and Windows artifact builds are running. No publication tag has been created.

@Kosinkadink commented on GitHub (Sep 1, 2026): Implementation is in #148. The branch pins SageAttention v2.2.0 source and license provenance, builds Python 3.12 / Torch 2.13 / CUDA 13 binary wheels for Linux and Windows, and verifies package metadata plus each extension's cubin coverage. Local evidence on RipperPC: - final Linux wheel: 38,397,300 bytes, sha256 `004ad09adb225ec0971a90608b05de807a706acb0e2755f7776e9cf3b70ca132` - `_fused`: SM80/86/89/90/120 - `_qattn_sm80`: SM80/86/89/90/120 - `_qattn_sm89`: SM89/90/120 - `_qattn_sm90`: SM90 - locked RTX PRO 6000 Blackwell SM120 causal and non-causal smoke passed - 521 CPU tests passed, 1361 skipped; Ruff clean - exact final diff independently reviewed with no remaining blocker Free GitHub manylinux and Windows artifact builds are running. No publication tag has been created.

bsakaguchi commented

2026-10-26 22:19:42 +00:00

Author

Owner

@Kosinkadink commented on GitHub (Sep 1, 2026):

Owner update at PR head bf64ba6:

Windows now completes the full CUDA build with C++20. Its first verification exposed that Windows cuobjdump names Hopper cubins as sm_90a; the verifier now accepts that architecture-specific suffix while preserving exact architecture-set checks.
Linux source/wheel/cubin verification and SM120 execution pass locally. Two hosted Linux builds were terminated with exit 143 at MAX_JOBS=8, so Linux compiler concurrency is now limited to 2 to fit runner memory.
Local validation: Ruff clean; 521 passed, 1361 skipped; all five managed architecture arms verified; managed wheel Sage and masked-SDPA adapter execution passed on SM120.

Exact-head Windows and Linux CI is running.

@Kosinkadink commented on GitHub (Sep 1, 2026): Owner update at PR head `bf64ba6`: - Windows now completes the full CUDA build with C++20. Its first verification exposed that Windows `cuobjdump` names Hopper cubins as `sm_90a`; the verifier now accepts that architecture-specific suffix while preserving exact architecture-set checks. - Linux source/wheel/cubin verification and SM120 execution pass locally. Two hosted Linux builds were terminated with exit 143 at `MAX_JOBS=8`, so Linux compiler concurrency is now limited to 2 to fit runner memory. - Local validation: Ruff clean; 521 passed, 1361 skipped; all five managed architecture arms verified; managed wheel Sage and masked-SDPA adapter execution passed on SM120. Exact-head Windows and Linux CI is running.

bsakaguchi referenced this issue

2026-10-26 22:20:24 +00:00

[PR #47] [ROCm] triton: skip dequantize_nvfp4 #101