[PR #251] [MERGED] Add SiTU input activation to the INT8 ConvRot quantizer #257

New issue

Closed

opened 2026-10-26 22:22:11 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:22:11 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/251
Author: @kijai
Created: 10/16/2026
Status: ✅ Merged
Merged: 10/16/2026
Merged by: @comfyanonymous

Base: main ← Head: situ-input-act

📝 Commits (2)

e8e9d27 Add SiTU input activation to the INT8 ConvRot quantizer
f901854 Fix SiTU precision near zero in the INT8 quantizer

📊 Changes

8 files changed (+119 additions, -41 deletions)

View changed files

📝 comfy_kitchen/__init__.py (+1 -1)
📝 comfy_kitchen/backends/_activations.py (+13 -2)
📝 comfy_kitchen/backends/ascend/__init__.py (+2 -1)
📝 comfy_kitchen/backends/cuda/__init__.py (+2 -2)
📝 comfy_kitchen/backends/cuda/dlpack_bindings.cpp (+7 -7)
📝 comfy_kitchen/backends/cuda/input_act_codes.h (+8 -3)
📝 comfy_kitchen/backends/cuda/ops/int8_linear.cu (+30 -6)
📝 tests/test_int8_input_act.py (+56 -19)

📄 Description

Adds input_act="situ"

It is folded into the INT8 ConvRot quantizer the same way as swiglu, in the row quantizer and the M=1 fused GEMV, so the MLP down-projection never writes the activation to HBM.

MOSS-Image int8 forward at 1024x1024 on a 4090: 175.7 -> 128.5 ms. situ ran as ~9 elementwise passes and was about a third of the forward.

_{🔄 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/251 **Author:** [@kijai](https://github.com/kijai) **Created:** 10/16/2026 **Status:** ✅ Merged **Merged:** 10/16/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `situ-input-act` --- ### 📝 Commits (2) - [`e8e9d27`](https://github.com/Comfy-Org/comfy-kitchen/commit/e8e9d2787813be046ac4fd7fef490da1db40c48c) Add SiTU input activation to the INT8 ConvRot quantizer - [`f901854`](https://github.com/Comfy-Org/comfy-kitchen/commit/f901854de928f9fd31483bf897ba86221ee9199a) Fix SiTU precision near zero in the INT8 quantizer ### 📊 Changes **8 files changed** (+119 additions, -41 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/__init__.py` (+1 -1) 📝 `comfy_kitchen/backends/_activations.py` (+13 -2) 📝 `comfy_kitchen/backends/ascend/__init__.py` (+2 -1) 📝 `comfy_kitchen/backends/cuda/__init__.py` (+2 -2) 📝 `comfy_kitchen/backends/cuda/dlpack_bindings.cpp` (+7 -7) 📝 `comfy_kitchen/backends/cuda/input_act_codes.h` (+8 -3) 📝 `comfy_kitchen/backends/cuda/ops/int8_linear.cu` (+30 -6) 📝 `tests/test_int8_input_act.py` (+56 -19) </details> ### 📄 Description Adds `input_act="situ"` It is folded into the INT8 ConvRot quantizer the same way as `swiglu`, in the row quantizer and the M=1 fused GEMV, so the MLP down-projection never writes the activation to HBM. MOSS-Image int8 forward at 1024x1024 on a 4090: 175.7 -> 128.5 ms. `situ` ran as ~9 elementwise passes and was about a third of the forward. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:22:11 +00:00