mirror of
https://github.com/Comfy-Org/comfy-kitchen.git
synced 2026-11-04 23:37:42 +00:00
[PR #255] [MERGED] Fuse input activations into the W4A8/W6A8 quantizer #261
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Comfy-Org/comfy-kitchen#261
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
📋 Pull Request Information
Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/255
Author: @kijai
Created: 10/18/2026
Status: ✅ Merged
Merged: 10/18/2026
Merged by: @comfyanonymous
Base:
main← Head:w4a8-fused-input-act📝 Commits (3)
4a03b0aFuse input activations into the W4A8/W6A8 quantizer9b0526cMatch the ConvRot quantizer's shared-memory estimate to its launch, test the fused prefill activation2a83b87Fix test📊 Changes
3 files changed (+88 additions, -23 deletions)
View changed files
📝
comfy_kitchen/backends/cuda/__init__.py(+63 -21)📝
tests/test_convrot_w4a4.py(+2 -2)📝
tests/test_w4a8_mma.py(+23 -0)📄 Description
W4A8/W6A8 linears applied the input activation (swiglu, situ, ...) as a separate pass before quantizing, except for M<=8. The quantizer they use already supports fusing it.
MOSS-Image mixed file now matches int8 speed, and MiniMax-H3 is slightly faster as well.
🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.