[PR #6] [MERGED] Rope update #80

New issue

Closed

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

bsakaguchi commented

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

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/6
Author: @contentis
Created: 1/6/2026
Status: ✅ Merged
Merged: 1/6/2026
Merged by: @comfyanonymous

Base: main ← Head: rope_update

📝 Commits (2)

3bc00dd improved rope function
142d357 allow for f16 freq

📊 Changes

9 files changed (+511 additions, -435 deletions)

View changed files

📝 comfy_kitchen/backends/cuda/__init__.py (+24 -4)
📝 comfy_kitchen/backends/cuda/dlpack_bindings.cpp (+92 -46)
📝 comfy_kitchen/backends/cuda/dtype_dispatch.cuh (+10 -0)
📝 comfy_kitchen/backends/cuda/ops/apply_rope.cu (+250 -220)
📝 comfy_kitchen/backends/eager/__init__.py (+2 -2)
📝 comfy_kitchen/backends/triton/__init__.py (+5 -20)
📝 comfy_kitchen/backends/triton/rope.py (+45 -44)
📝 tests/conftest.py (+10 -1)
📝 tests/test_rope.py (+73 -98)

📄 Description

Changes made:

Modify tirton and cuda kernel to allow for torch-like broadcasting, thereby support bothe BHND and BNHD (ZImage)
Extended testing to include both layouts
Optional input in cuda kernel for fused q,k instead of two kernel launches
Performance optimiations to support broader load+stores.
Extended input dtype to allow for f16/bf16 freq_cis

_{🔄 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/6 **Author:** [@contentis](https://github.com/contentis) **Created:** 1/6/2026 **Status:** ✅ Merged **Merged:** 1/6/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `rope_update` --- ### 📝 Commits (2) - [`3bc00dd`](https://github.com/Comfy-Org/comfy-kitchen/commit/3bc00ddcd520073056d3cfc6e3a8368bb0eda858) improved rope function - [`142d357`](https://github.com/Comfy-Org/comfy-kitchen/commit/142d357aa062b5ac092ab3fbfedddf82a4d1227f) allow for f16 freq ### 📊 Changes **9 files changed** (+511 additions, -435 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/backends/cuda/__init__.py` (+24 -4) 📝 `comfy_kitchen/backends/cuda/dlpack_bindings.cpp` (+92 -46) 📝 `comfy_kitchen/backends/cuda/dtype_dispatch.cuh` (+10 -0) 📝 `comfy_kitchen/backends/cuda/ops/apply_rope.cu` (+250 -220) 📝 `comfy_kitchen/backends/eager/__init__.py` (+2 -2) 📝 `comfy_kitchen/backends/triton/__init__.py` (+5 -20) 📝 `comfy_kitchen/backends/triton/rope.py` (+45 -44) 📝 `tests/conftest.py` (+10 -1) 📝 `tests/test_rope.py` (+73 -98) </details> ### 📄 Description Changes made: - Modify tirton and cuda kernel to allow for torch-like broadcasting, thereby support bothe BHND and BNHD (ZImage) - Extended testing to include both layouts - Optional input in cuda kernel for fused q,k instead of two kernel launches - Performance optimiations to support broader load+stores. - Extended input dtype to allow for f16/bf16 freq_cis --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

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