[PR #260] [MERGED] VAE speedups: Add stats override and row stride to group_norm_silu_pad3d #265

New issue

Closed

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

bsakaguchi commented

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

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/260
Author: @kijai
Created: 10/19/2026
Status: ✅ Merged
Merged: 10/21/2026
Merged by: @comfyanonymous

Base: main ← Head: group_norm_stats

📝 Commits (3)

17ebcf3 Add stats override and row stride to group_norm_silu_pad3d
1026d95 Signature order fix
ca0d637 Add stats override and row stride to group_norm_silu_pad3d

📊 Changes

8 files changed (+345 additions, -63 deletions)

View changed files

📝 comfy_kitchen/__init__.py (+14 -4)
📝 comfy_kitchen/backends/cuda/__init__.py (+65 -12)
📝 comfy_kitchen/backends/cuda/dlpack_bindings.cpp (+59 -8)
📝 comfy_kitchen/backends/cuda/ops/group_norm_pad3d.cu (+59 -24)
📝 comfy_kitchen/backends/eager/__init__.py (+8 -1)
📝 comfy_kitchen/backends/eager/group_norm_pad3d.py (+43 -7)
📝 comfy_kitchen/backends/hip/__init__.py (+7 -6)
📝 tests/test_group_norm_silu_pad3d.py (+90 -1)

📄 Description

Lets a caller normalize row strips of a frame with that frame's whole-image GroupNorm statistics, so a conv stack can run strip-wise without the full-size normalized intermediate. Needed for the fused SD/Flux2 VAE path.

Both ops also read row-interleaved frames in place through a row stride, and any multiple of 8 channels up to 2048 is served. Packed input is addressed exactly as before.

Core PR https://github.com/Comfy-Org/ComfyUI/pull/17196

_{🔄 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/260 **Author:** [@kijai](https://github.com/kijai) **Created:** 10/19/2026 **Status:** ✅ Merged **Merged:** 10/21/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `group_norm_stats` --- ### 📝 Commits (3) - [`17ebcf3`](https://github.com/Comfy-Org/comfy-kitchen/commit/17ebcf3be24a3e6f7e142e874d983a2b6129297e) Add stats override and row stride to group_norm_silu_pad3d - [`1026d95`](https://github.com/Comfy-Org/comfy-kitchen/commit/1026d950a1d3b68c15877b811b4a4a470c5fc147) Signature order fix - [`ca0d637`](https://github.com/Comfy-Org/comfy-kitchen/commit/ca0d637235561fbc69f205192a2eee17093ee67b) Add stats override and row stride to group_norm_silu_pad3d ### 📊 Changes **8 files changed** (+345 additions, -63 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/__init__.py` (+14 -4) 📝 `comfy_kitchen/backends/cuda/__init__.py` (+65 -12) 📝 `comfy_kitchen/backends/cuda/dlpack_bindings.cpp` (+59 -8) 📝 `comfy_kitchen/backends/cuda/ops/group_norm_pad3d.cu` (+59 -24) 📝 `comfy_kitchen/backends/eager/__init__.py` (+8 -1) 📝 `comfy_kitchen/backends/eager/group_norm_pad3d.py` (+43 -7) 📝 `comfy_kitchen/backends/hip/__init__.py` (+7 -6) 📝 `tests/test_group_norm_silu_pad3d.py` (+90 -1) </details> ### 📄 Description Lets a caller normalize row strips of a frame with that frame's whole-image GroupNorm statistics, so a conv stack can run strip-wise without the full-size normalized intermediate. Needed for the fused SD/Flux2 VAE path. Both ops also read row-interleaved frames in place through a row stride, and any multiple of 8 channels up to 2048 is served. Packed input is addressed exactly as before. Core PR https://github.com/Comfy-Org/ComfyUI/pull/17196 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

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