[PR #148] [MERGED] Build managed SageAttention wheels for CUDA 13 #177

New issue

Closed

opened 2026-10-26 22:21:15 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:21:15 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/148
Author: @Kosinkadink
Created: 9/1/2026
Status: ✅ Merged
Merged: 9/1/2026
Merged by: @Kosinkadink

Base: main ← Head: feature/147-sageattention-wheels

📝 Commits (8)

291a1e2 Build managed SageAttention binary wheels
2764c4f Pin SageAttention wheel build tools
100c2f0 Preserve SageAttention patch line endings
24b7c7b Correct SageAttention dispatch in managed wheels
3e25f63 Build SageAttention wheels with C++20
bf64ba6 Stabilize SageAttention wheel verification
9199773 Install cuSPARSE headers for SageAttention wheels
3c4182e Install CUDA math headers for SageAttention wheels

📊 Changes

8 files changed (+793 additions, -0 deletions)

View changed files

➕ .gitattributes (+1 -0)
➕ .github/workflows/build-sageattention-wheels.yml (+198 -0)
➕ sageattention/README.md (+35 -0)
➕ sageattention/packaging.patch (+205 -0)
➕ sageattention/prepare_source.py (+117 -0)
➕ sageattention/source.json (+47 -0)
➕ sageattention/verify_architectures.py (+58 -0)
➕ sageattention/verify_wheel.py (+132 -0)

📄 Description

Summary

pin SageAttention 2.2.0 source, archive digest, license, and reproducible build inputs
build Python 3.12 / Torch 2.13 / CUDA 13 wheels for manylinux x86_64 and Windows x86_64
package the upstream sageattention namespace as comfy-kitchen-sageattention
verify metadata, provenance, licenses, native files, imports, and per-extension SM80/86/89/90/120 cubin coverage
gate PyPI publication on the exact sageattention-v2.2.0.post1 tag after both platform builds pass

Validation

ruff check .
ruff format --check sageattention
CUDA_VISIBLE_DEVICES='' python -m pytest -q: 521 passed, 1361 skipped
pinned source preparation and packaging patch verification
local CUDA 13 wheel build: 38,397,300 bytes
per-extension cuobjdump --list-elf inventory passed
wheel metadata, provenance, license, and import checks passed
locked RTX PRO 6000 Blackwell SM120 smoke passed for causal and non-causal attention

Refs #147

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/comfy-kitchen/pull/148 **Author:** [@Kosinkadink](https://github.com/Kosinkadink) **Created:** 9/1/2026 **Status:** ✅ Merged **Merged:** 9/1/2026 **Merged by:** [@Kosinkadink](https://github.com/Kosinkadink) **Base:** `main` ← **Head:** `feature/147-sageattention-wheels` --- ### 📝 Commits (8) - [`291a1e2`](https://github.com/Comfy-Org/comfy-kitchen/commit/291a1e21f1e2cc6c3a0752db412a6df6cfe46842) Build managed SageAttention binary wheels - [`2764c4f`](https://github.com/Comfy-Org/comfy-kitchen/commit/2764c4fd6c02713c6fbce2d67ceb4befd1d13a5f) Pin SageAttention wheel build tools - [`100c2f0`](https://github.com/Comfy-Org/comfy-kitchen/commit/100c2f0d28b0085d6e7329049248fd3a5e987230) Preserve SageAttention patch line endings - [`24b7c7b`](https://github.com/Comfy-Org/comfy-kitchen/commit/24b7c7b464fdb8079cc333eaef94338487a6688f) Correct SageAttention dispatch in managed wheels - [`3e25f63`](https://github.com/Comfy-Org/comfy-kitchen/commit/3e25f63be5581b758db5687789c5dfbe734b7adf) Build SageAttention wheels with C++20 - [`bf64ba6`](https://github.com/Comfy-Org/comfy-kitchen/commit/bf64ba696aec34c2970ffb7ed098e7ba51ab22dc) Stabilize SageAttention wheel verification - [`9199773`](https://github.com/Comfy-Org/comfy-kitchen/commit/9199773c333cb5d656b2bea35239738ad9ea6461) Install cuSPARSE headers for SageAttention wheels - [`3c4182e`](https://github.com/Comfy-Org/comfy-kitchen/commit/3c4182e01b5eb07a1cdc6f5b42d50e94a7db0b6d) Install CUDA math headers for SageAttention wheels ### 📊 Changes **8 files changed** (+793 additions, -0 deletions) <details> <summary>View changed files</summary> ➕ `.gitattributes` (+1 -0) ➕ `.github/workflows/build-sageattention-wheels.yml` (+198 -0) ➕ `sageattention/README.md` (+35 -0) ➕ `sageattention/packaging.patch` (+205 -0) ➕ `sageattention/prepare_source.py` (+117 -0) ➕ `sageattention/source.json` (+47 -0) ➕ `sageattention/verify_architectures.py` (+58 -0) ➕ `sageattention/verify_wheel.py` (+132 -0) </details> ### 📄 Description ## Summary - pin SageAttention 2.2.0 source, archive digest, license, and reproducible build inputs - build Python 3.12 / Torch 2.13 / CUDA 13 wheels for manylinux x86_64 and Windows x86_64 - package the upstream `sageattention` namespace as `comfy-kitchen-sageattention` - verify metadata, provenance, licenses, native files, imports, and per-extension SM80/86/89/90/120 cubin coverage - gate PyPI publication on the exact `sageattention-v2.2.0.post1` tag after both platform builds pass ## Validation - `ruff check .` - `ruff format --check sageattention` - `CUDA_VISIBLE_DEVICES='' python -m pytest -q`: 521 passed, 1361 skipped - pinned source preparation and packaging patch verification - local CUDA 13 wheel build: 38,397,300 bytes - per-extension `cuobjdump --list-elf` inventory passed - wheel metadata, provenance, license, and import checks passed - locked RTX PRO 6000 Blackwell SM120 smoke passed for causal and non-causal attention Refs #147 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:21:15 +00:00

closed this issue
added the
pull-request
label

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

Comfy-Org/comfy-kitchen#177

No description provided.

Rows
Columns