Comfy-Org/comfy-kitchen

Fork 0

mirror of https://github.com/Comfy-Org/comfy-kitchen.git synced 2026-11-04 23:37:42 +00:00

[PR #47] [ROCm] triton: skip `dequantize_nvfp4` #101

New issue

Open

opened 2026-10-26 22:20:24 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:20:24 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/47
Author: @Apophis3158
Created: 6/5/2026
Status: 🔄 Open

Base: main ← Head: rocm-skip-deq_nvfp4

📝 Commits (1)

54ae117 [ROCm] triton: skip dequantize_nvfp4

📊 Changes

1 file changed (+22 additions, -16 deletions)

View changed files

📝 comfy_kitchen/backends/triton/__init__.py (+22 -16)

📄 Description

AMD GPU will crash with triton backend:

[DEBUG] Backend triton selected for dequantize_per_tensor_fp8
[DEBUG] Backend triton selected for dequantize_per_tensor_fp8
[DEBUG] Backend triton selected for dequantize_nvfp4
error: couldn't allocate input reg for constraint 'r'

Add a way to skip dequantize_nvfp4 on ROCm.

Tested with flux 2 klein, successfully fallback to eager:

[DEBUG] Backend triton selected for dequantize_per_tensor_fp8
[DEBUG] Backend eager selected for dequantize_nvfp4
[DEBUG] Backend eager selected for dequantize_nvfp4
[DEBUG] Backend eager selected for dequantize_nvfp4

_{🔄 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/47 **Author:** [@Apophis3158](https://github.com/Apophis3158) **Created:** 6/5/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `rocm-skip-deq_nvfp4` --- ### 📝 Commits (1) - [`54ae117`](https://github.com/Comfy-Org/comfy-kitchen/commit/54ae117c5492b55ddf0bc269abaebb60f736e39d) [ROCm] triton: skip dequantize_nvfp4 ### 📊 Changes **1 file changed** (+22 additions, -16 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/backends/triton/__init__.py` (+22 -16) </details> ### 📄 Description AMD GPU will crash with triton backend: ``` [DEBUG] Backend triton selected for dequantize_per_tensor_fp8 [DEBUG] Backend triton selected for dequantize_per_tensor_fp8 [DEBUG] Backend triton selected for dequantize_nvfp4 error: couldn't allocate input reg for constraint 'r' ``` Add a way to skip `dequantize_nvfp4` on ROCm. Tested with flux 2 klein, successfully fallback to eager: ``` [DEBUG] Backend triton selected for dequantize_per_tensor_fp8 [DEBUG] Backend eager selected for dequantize_nvfp4 [DEBUG] Backend eager selected for dequantize_nvfp4 [DEBUG] Backend eager selected for dequantize_nvfp4 ``` --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-26 22:20:24 +00:00

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

Comfy-Org/comfy-kitchen#101

No description provided.

Rows
Columns

[PR #47] [ROCm] triton: skip dequantize_nvfp4 #101

📋 Pull Request Information

📝 Commits (1)

📊 Changes

📄 Description

[PR #47] [ROCm] triton: skip `dequantize_nvfp4` #101