[PR #76] [MERGED] Speed up native builds with portable compiler caching #124

New issue

Closed

opened 2026-10-26 22:20:36 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:20:36 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/76
Author: @contentis
Created: 7/14/2026
Status: ✅ Merged
Merged: 7/15/2026
Merged by: @comfyanonymous

Base: main ← Head: build-cache-version

📝 Commits (6)

7045bc1 Speed up native builds with portable compiler caching
bf9bf66 single version truth
07c0bfb Fix windows and add stats
59aaa41 pls
938cbb6 this should hit now
fc8b1fb harden cuda discovery

📊 Changes

5 files changed (+206 additions, -138 deletions)

View changed files

📝 .github/workflows/build-wheels.yml (+116 -88)
📝 comfy_kitchen/__init__.py (+0 -2)
📝 comfy_kitchen/backends/cuda/CMakeLists.txt (+44 -21)
📝 comfy_kitchen/backends/cuda/dlpack_bindings.cpp (+0 -2)
📝 setup.py (+46 -25)

📄 Description

This reduces CUDA wheel build time by using Ninja and ccache in Linux and Windows CI.

CUDA kernels now build as a separate object target. This keeps the expensive CUTLASS translation units independent of the Python/nanobind extension target, allowing ccache to reuse kernel fatbins across wheel builds when Python-specific build inputs change.

Also removes forced CMake-cache deletion so CMake can reuse its generated build graph.
Measured locally on Linux with the default fatbin architecture set:

Build from scratch: 447.30 s (7m 27s)
Rebuild after touching only ops/apply_rope.cu: 6.44 s

The RoPE-only rebuild compiled apply_rope.cu and relinked _C; all other CUDA translation units were unchanged.

CI cache-hit rates will be reported through ccache --show-stats.

_{🔄 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/76 **Author:** [@contentis](https://github.com/contentis) **Created:** 7/14/2026 **Status:** ✅ Merged **Merged:** 7/15/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `build-cache-version` --- ### 📝 Commits (6) - [`7045bc1`](https://github.com/Comfy-Org/comfy-kitchen/commit/7045bc18520dc06cdeaaec8c21a9a841092777d6) Speed up native builds with portable compiler caching - [`bf9bf66`](https://github.com/Comfy-Org/comfy-kitchen/commit/bf9bf6646bdaafbdea511811f23064bcfab09602) single version truth - [`07c0bfb`](https://github.com/Comfy-Org/comfy-kitchen/commit/07c0bfb6e973a5355bdcba9ab4e84b91ab415240) Fix windows and add stats - [`59aaa41`](https://github.com/Comfy-Org/comfy-kitchen/commit/59aaa411743711b1e509207bc9cb469018e98ae9) pls - [`938cbb6`](https://github.com/Comfy-Org/comfy-kitchen/commit/938cbb6baf2ce4738977e739e9ff6705047a63fa) this should hit now - [`fc8b1fb`](https://github.com/Comfy-Org/comfy-kitchen/commit/fc8b1fbe6c1f89a6eef80ab2bb70614c91f92442) harden cuda discovery ### 📊 Changes **5 files changed** (+206 additions, -138 deletions) <details> <summary>View changed files</summary> 📝 `.github/workflows/build-wheels.yml` (+116 -88) 📝 `comfy_kitchen/__init__.py` (+0 -2) 📝 `comfy_kitchen/backends/cuda/CMakeLists.txt` (+44 -21) 📝 `comfy_kitchen/backends/cuda/dlpack_bindings.cpp` (+0 -2) 📝 `setup.py` (+46 -25) </details> ### 📄 Description This reduces CUDA wheel build time by using Ninja and ccache in Linux and Windows CI. CUDA kernels now build as a separate object target. This keeps the expensive CUTLASS translation units independent of the Python/nanobind extension target, allowing ccache to reuse kernel fatbins across wheel builds when Python-specific build inputs change. Also removes forced CMake-cache deletion so CMake can reuse its generated build graph. Measured locally on Linux with the default fatbin architecture set: - Build from scratch: 447.30 s (7m 27s) - Rebuild after touching only ops/apply_rope.cu: 6.44 s The RoPE-only rebuild compiled apply_rope.cu and relinked _C; all other CUDA translation units were unchanged. CI cache-hit rates will be reported through ccache --show-stats. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:20:36 +00:00