[PR #3144] [CLOSED] Add h3_sage_amd: AMD gfx12 Memory-Efficient Sage Attention for MiniMax H3 #3050

New issue

Closed

opened 2026-10-28 21:14:47 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-28 21:14:47 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/ComfyUI-Manager/pull/3144
Author: @Newaiguy
Created: 8/9/2026
Status: ❌ Closed

Base: main ← Head: main

📝 Commits (1)

34c6b3a Add h3_sage_amd: AMD gfx12 Sage Attention for MiniMax H3

📊 Changes

1 file changed (+17 additions, -3 deletions)

View changed files

📝 custom-node-list.json (+17 -3)

📄 Description

New Custom Node

h3_sage_amd - AMD gfx12 Memory-Efficient Sage Attention for MiniMax H3

Description

AMD gfx12 (RX 9070 / 9070 XT) MiniMax H3 memory-efficient Sage Attention node. Mirrors KJNodes' MiniMaxH3MemoryEfficientSageAttentionPatch (NVIDIA) using native HIP kernel.

Uses sageattn_qk_int8_pv_gfx12_native (compiled HIP C++ kernel, equivalent to N卡 _qattn_smXX)
Saves 33% peak VRAM by freeing qkv buffer before attention
Includes Triton fallback when native module unavailable
Tested on RX 9070 XT: H3 10-step sampling ~350s

Repository

https://github.com/Newaiguy/ComfyUI-h3_sage_amd

Requirements

SageAttention 2.2.0+ with gfx12 native support (PR #368)
AMD gfx12 GPU (RX 9070 / 9070 XT)
ComfyUI with --use-sage-attention flag

_{🔄 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/ComfyUI-Manager/pull/3144 **Author:** [@Newaiguy](https://github.com/Newaiguy) **Created:** 8/9/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `main` --- ### 📝 Commits (1) - [`34c6b3a`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/34c6b3af1dc345067a597d5fb9480b611ac331dc) Add h3_sage_amd: AMD gfx12 Sage Attention for MiniMax H3 ### 📊 Changes **1 file changed** (+17 additions, -3 deletions) <details> <summary>View changed files</summary> 📝 `custom-node-list.json` (+17 -3) </details> ### 📄 Description ## New Custom Node **h3_sage_amd** - AMD gfx12 Memory-Efficient Sage Attention for MiniMax H3 ### Description AMD gfx12 (RX 9070 / 9070 XT) MiniMax H3 memory-efficient Sage Attention node. Mirrors KJNodes' `MiniMaxH3MemoryEfficientSageAttentionPatch` (NVIDIA) using native HIP kernel. - Uses `sageattn_qk_int8_pv_gfx12_native` (compiled HIP C++ kernel, equivalent to N卡 `_qattn_smXX`) - Saves 33% peak VRAM by freeing qkv buffer before attention - Includes Triton fallback when native module unavailable - Tested on RX 9070 XT: H3 10-step sampling ~350s ### Repository https://github.com/Newaiguy/ComfyUI-h3_sage_amd ### Requirements - SageAttention 2.2.0+ with gfx12 native support (PR #368) - AMD gfx12 GPU (RX 9070 / 9070 XT) - ComfyUI with `--use-sage-attention` flag --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-28 21:14:47 +00:00