[PR #17] [MERGED] Runtime cuBLAS linking #84

New issue

Closed

opened 2026-10-26 22:20:12 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:20:12 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/17
Author: @contentis
Created: 1/7/2026
Status: ✅ Merged
Merged: 1/7/2026
Merged by: @comfyanonymous

Base: main ← Head: runtime_linking

📝 Commits (1)

b4a2a66 runtime linking of cublas, readme update, turing support

📊 Changes

10 files changed (+371 additions, -108 deletions)

View changed files

📝 README.md (+16 -22)
📝 comfy_kitchen/backends/cuda/CMakeLists.txt (+4 -1)
📝 comfy_kitchen/backends/cuda/__init__.py (+35 -30)
➕ comfy_kitchen/backends/cuda/cublaslt_runtime.h (+248 -0)
📝 comfy_kitchen/backends/cuda/dlpack_bindings.cpp (+10 -0)
📝 comfy_kitchen/backends/cuda/ops/cublas_gemm_nvfp4.cu (+39 -35)
📝 comfy_kitchen/backends/cuda/ops/per_tensor_quantize.cu (+7 -7)
📝 comfy_kitchen/backends/cuda/ops/quantize_nvfp4.cu (+7 -7)
📝 comfy_kitchen/tensor/base.py (+3 -1)
📝 setup.py (+2 -5)

📄 Description

Use dlopen/LoadLibrary to load cublas at runtime and only disable features that require it (NVFP4 GEMM)
Furthermore clean up c namespaces, add turing support (oldest possible compile target with CU13), update readme accoringly.

_{🔄 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/17 **Author:** [@contentis](https://github.com/contentis) **Created:** 1/7/2026 **Status:** ✅ Merged **Merged:** 1/7/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `runtime_linking` --- ### 📝 Commits (1) - [`b4a2a66`](https://github.com/Comfy-Org/comfy-kitchen/commit/b4a2a661518bfd94a8fa4ce205f2ecc1deccf405) runtime linking of cublas, readme update, turing support ### 📊 Changes **10 files changed** (+371 additions, -108 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+16 -22) 📝 `comfy_kitchen/backends/cuda/CMakeLists.txt` (+4 -1) 📝 `comfy_kitchen/backends/cuda/__init__.py` (+35 -30) ➕ `comfy_kitchen/backends/cuda/cublaslt_runtime.h` (+248 -0) 📝 `comfy_kitchen/backends/cuda/dlpack_bindings.cpp` (+10 -0) 📝 `comfy_kitchen/backends/cuda/ops/cublas_gemm_nvfp4.cu` (+39 -35) 📝 `comfy_kitchen/backends/cuda/ops/per_tensor_quantize.cu` (+7 -7) 📝 `comfy_kitchen/backends/cuda/ops/quantize_nvfp4.cu` (+7 -7) 📝 `comfy_kitchen/tensor/base.py` (+3 -1) 📝 `setup.py` (+2 -5) </details> ### 📄 Description Use dlopen/LoadLibrary to load cublas at runtime and only disable features that require it (NVFP4 GEMM) Furthermore clean up c namespaces, add turing support (oldest possible compile target with CU13), update readme accoringly. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:20:12 +00:00