ComfyUI Installer for AMD Ryzen AI Max 395 (gfx1151)

This repository contains installation scripts and documentation for running ComfyUI with GPU acceleration on AMD Ryzen AI Max 395 APUs (Strix Halo architecture, gfx1151).

Overview
The Problem
The Solution
Requirements
Installation
Usage
Troubleshooting
Technical Details
Download Links & Wheel Sources
Sources & References

Overview

The AMD Ryzen AI Max 395 features an integrated Radeon GPU based on the RDNA 3.5 architecture, identified as gfx1151 (Strix Halo). This is a relatively new GPU architecture that requires specific PyTorch wheels compiled with gfx1151 support.

Hardware Specifications

Component	Details
APU	AMD Ryzen AI Max 395
GPU Architecture	RDNA 3.5 (gfx1151)
GPU Name	AMD Radeon Graphics (Strix Halo)
Device ID	0x1586
Shared Memory	Up to 96GB (system RAM)

The Problem

Symptoms

When attempting to run ComfyUI with standard ROCm PyTorch wheels, users encounter several errors:

Error 1: No GPU detected

RuntimeError: No HIP GPUs are available

Error 2: Architecture mismatch (after GPU is detected)

HIP error: invalid device function

Error 3: Missing library in nightly wheels

ModuleNotFoundError: Unknown rocm library 'hipsparselt'

Root Cause Analysis

The issue occurs at multiple levels of the ROCm stack:

Layer                          Status
─────────────────────────────────────────────────────
Kernel Driver (amdgpu)         ✅ Working
Display/OpenGL (Mesa)          ✅ Working
ROCm SMI (rocm-smi)            ✅ Working
HSA Runtime (rocminfo)         ⚠️  May show errors
HIP Compute                    ❌ "invalid device function"
PyTorch                        ❌ Kernels don't match gfx1151

The core problem: Official PyTorch ROCm wheels from https://download.pytorch.org/whl/rocm6.x are compiled for these GPU architectures:

gfx900, gfx906, gfx908, gfx90a (Vega/MI series)
gfx1030 (RDNA 2)
gfx1100 (RDNA 3)

gfx1151 (Strix Halo) is NOT included in the official wheels.

When PyTorch attempts to run a CUDA/HIP kernel on gfx1151 using wheels compiled for other architectures, it fails with invalid device function because the GPU instruction set doesn't match.

The hipsparselt Error

The CloudFront nightly wheels (d2awnip2yjpvqn.cloudfront.net/v2/gfx1151/) have a bug in the rocm_sdk package. It tries to preload hipsparselt, a library that doesn't exist for gfx1151:

# This fails because hipsparselt isn't available for gfx1151
ModuleNotFoundError: Unknown rocm library 'hipsparselt'

This is why we use the self-contained wheel from scottt instead.

The Solution

Self-Contained PyTorch Wheel (Recommended)

The install script uses @scottt's self-contained wheel which:

Bundles all required ROCm libraries internally
Doesn't rely on the broken rocm_sdk package
Is compiled specifically for gfx1151
Requires Python 3.11 (not 3.12)

Wheel Comparison

Wheel Source	Status	Python	Issue
`pytorch.org/whl/rocm6.x`	❌ Broken	Any	No gfx1151 kernels
`d2awnip2yjpvqn.cloudfront.net`	❌ Broken	3.12	hipsparselt missing
scottt/rocm-TheRock	✅ Working	3.11	Self-contained

The Fix Applied

The install script:

Requires Python 3.11 - The working wheel is only available for Python 3.11
Uses scottt's self-contained wheel - Avoids the broken rocm_sdk package
Prevents torch overwrite - ComfyUI's requirements.txt includes torch which would replace our working wheel
Verifies torch version - Checks and reinstalls if the wrong version was installed

Why Dependencies Break Things

Problem 1: requirements.txt

ComfyUI's requirements.txt includes:

torch
torchvision
torchaudio

When pip installs these, it pulls from PyPI/the default index, which gives you torch==2.6.0+rocm6.1 - the broken wheel without gfx1151 support.

Problem 2: Transitive Dependencies (kornia)

Even after filtering torch from requirements.txt, packages like kornia have torch>=2.0.0 as a dependency. When pip resolves kornia's dependencies, it downloads the default CUDA torch from PyPI:

Collecting torch>=2.0.0 (from kornia>=0.7.1...)
  Downloading torch-2.6.0-cp311-cp311-manylinux1_x86_64.whl

This results in torch-2.6.0+cu124 (CUDA version) being installed, which fails with:

RuntimeError: Found no NVIDIA driver on your system

The Solution

The script handles these issues by:

Installing scottt's wheel first
Using --no-deps when installing torchvision/torchaudio
Filtering torch packages from requirements.txt
Installing kornia with --no-deps to prevent torch dependency resolution
Force reinstalling scottt's wheel as the FINAL step - guarantees correct wheel
Verifying the correct version and failing if wrong

Requirements

System Requirements

OS: Ubuntu 22.04 / 24.04 LTS (or compatible Linux distribution)
Kernel: 6.x with amdgpu driver
Python: 3.11 (required - see installation below)
RAM: 32GB+ recommended (shared with GPU)
Storage: 20GB+ free space

Python 3.11 Installation (Ubuntu 24.04)

Ubuntu 24.04 ships with Python 3.12 by default. You need to install Python 3.11 from the deadsnakes PPA:

# Add the deadsnakes PPA
sudo add-apt-repository -y ppa:deadsnakes/ppa

# Update package lists
sudo apt update

# Install Python 3.11
sudo apt install -y python3.11 python3.11-venv python3.11-dev

# Verify installation
python3.11 --version
# Should output: Python 3.11.x

ROCm Requirements

The following ROCm tools should be installed and functional:

# Check ROCm tools
rocm-smi -i          # Should show your GPU
rocminfo             # May show warnings on gfx1151 (non-fatal)

Install ROCm (if not already installed)

# Add AMD ROCm repository (Ubuntu 24.04)
wget https://repo.radeon.com/amdgpu-install/latest/ubuntu/noble/amdgpu-install_6.3.60300-1_all.deb
sudo apt install ./amdgpu-install_6.3.60300-1_all.deb
sudo amdgpu-install --usecase=rocm

# Add user to required groups
sudo usermod -aG video,render $USER
# Log out and back in for group changes to take effect

Installation

Quick Start

# 1. Install Python 3.11 (if not already installed)
sudo add-apt-repository -y ppa:deadsnakes/ppa
sudo apt update
sudo apt install -y python3.11 python3.11-venv python3.11-dev

# 2. Clone this repository
git clone <this-repo> ~/Projects/comfyui
cd ~/Projects/comfyui

# 3. Remove any existing broken venv
rm -rf ~/ComfyUI/venv

# 4. Run the installer
chmod +x install_comfyui_amd_ryzen.sh
./install_comfyui_amd_ryzen.sh

What the Script Does

Checks Python version - Verifies Python 3.11 is available
Validates ROCm installation - Checks for rocminfo, rocm-smi
Tests GPU visibility - Ensures rocm-smi can see your GPU
Installs system dependencies - git, python3-venv, build-essential
Clones/updates ComfyUI - From official GitHub repository
Creates Python 3.11 virtual environment - Isolated from system Python
Removes conflicting packages - Old torch/CUDA/ROCm/NVIDIA wheels
Installs scottt's self-contained gfx1151 wheel - From GitHub releases
Installs dependencies with --no-deps - Prevents pip from pulling in CUDA torch
Removes any NVIDIA packages - Cleans up any accidentally installed CUDA dependencies
Force reinstalls the gfx1151 torch wheel - Ensures correct version after all deps installed
Creates launcher script - run_comfy_amd_ryzen_ai.sh
Validates PyTorch HIP - Tests GPU compute functionality

Post-Installation Fixes

If you see import errors about missing modules after installation, install them manually:

source ~/ComfyUI/venv/bin/activate

# Common missing dependencies (due to --no-deps installation strategy)
pip install typing-inspection kornia_rs httpx

# If huggingface-hub version conflict:
pip install "huggingface-hub<1.0,>=0.34.0"

Environment Variables

You can customize the installation:

# Use a different install directory
COMFY_DIR=~/my-comfyui ./install_comfyui_amd_ryzen.sh

# Use a specific Python binary (must be 3.11)
PYTHON_BIN=/usr/bin/python3.11 ./install_comfyui_amd_ryzen.sh

Usage

Starting ComfyUI

cd ~/ComfyUI
./run_comfy_amd_ryzen_ai.sh

Then open your browser to: http://localhost:8188/

Launcher Script Environment

The launcher script (run_comfy_amd_ryzen_ai.sh) sets these environment variables:

unset CUDA_VISIBLE_DEVICES    # CRITICAL: Remove any empty CUDA_VISIBLE_DEVICES
HIP_VISIBLE_DEVICES=0         # Use first AMD GPU
AMD_LOG_LEVEL=0               # Reduce log noise
GPU_MAX_HEAP_SIZE=100         # Allow full VRAM usage
GPU_MAX_ALLOC_PERCENT=100     # Allow large allocations

⚠️ Important: The unset CUDA_VISIBLE_DEVICES is critical! PyTorch ROCm uses CUDA API names internally. If CUDA_VISIBLE_DEVICES is set to empty (even from your .bashrc), PyTorch won't see any GPUs. The run script now explicitly unsets this variable.

Debugging

If you encounter issues, enable debug mode by editing the launcher:

# Uncomment these lines in run_comfy_amd_ryzen_ai.sh:
export AMD_SERIALIZE_KERNEL=3
export HIP_LAUNCH_BLOCKING=1

Troubleshooting

"No HIP GPUs are available"

Cause: PyTorch can't see the GPU via HIP.

Solutions:

Check if CUDA_VISIBLE_DEVICES is set to empty - This is the most common cause!

# Check current value
echo $CUDA_VISIBLE_DEVICES

# If it shows empty (but the variable exists), that's the problem!
# BAD - This hides the GPU from PyTorch ROCm:
export CUDA_VISIBLE_DEVICES=""

# GOOD - Unset it completely:
unset CUDA_VISIBLE_DEVICES
export HIP_VISIBLE_DEVICES=0

Try a fresh terminal - Your .bashrc or other startup scripts might set CUDA_VISIBLE_DEVICES="". Open a completely new terminal and try again.

Check for the variable in startup files:

grep -r "CUDA_VISIBLE_DEVICES" ~/.bashrc ~/.profile ~/.zshrc 2>/dev/null

Check user groups: groups $USER should include video and render
Check device permissions: ls -la /dev/dri/
Verify ROCm: rocm-smi -i

# Fix permissions
sudo usermod -aG video,render $USER
# Then log out and back in

"invalid device function"

Cause: PyTorch wheels don't include gfx1151 kernels.

Solution: Ensure you're using scottt's self-contained wheel. Re-run the install script.

# Verify correct PyTorch version
source ~/ComfyUI/venv/bin/activate
python -c "import torch; print(torch.__version__)"
# Should show: 2.7.0a0+gitbfd8155 (self-contained wheel)
# NOT: 2.6.0+rocm6.1 or 2.7.1+rocm7.10.0a... (broken wheels)

"Unknown rocm library 'hipsparselt'"

Cause: Using the CloudFront nightly wheels which have a broken rocm_sdk package.

Solution: The script now uses scottt's self-contained wheel which doesn't have this issue. If you see this error:

# Remove the broken venv and reinstall
rm -rf ~/ComfyUI/venv
cd ~/Projects/comfyui
./install_comfyui_amd_ryzen.sh

"python3.11 not found"

Cause: Python 3.11 is not installed.

Solution: Install from deadsnakes PPA:

sudo add-apt-repository -y ppa:deadsnakes/ppa
sudo apt update
sudo apt install -y python3.11 python3.11-venv python3.11-dev

Wrong torch version installed

Cause: Dependencies (requirements.txt or packages like kornia) overwrote the gfx1151 wheel.

Symptoms:

python -c "import torch; print(torch.__version__)"
# Shows: 2.6.0+rocm6.1  (WRONG - standard ROCm, no gfx1151)
# Shows: 2.6.0+cu124    (WRONG - CUDA version, needs NVIDIA)
# Should show: 2.7.0a0+gitbfd8155  (CORRECT - scottt's gfx1151 wheel)

Solution: The updated script now force-reinstalls the correct wheel as the final step. If it still fails:

source ~/ComfyUI/venv/bin/activate

# Uninstall all torch packages
pip uninstall -y torch torchvision torchaudio

# Reinstall scottt's wheel
pip install --force-reinstall --no-deps \
  https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl

# Reinstall torchvision/torchaudio without dependencies
pip install --no-deps torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1

"Found no NVIDIA driver on your system"

Cause: CUDA version of torch was installed instead of ROCm/HIP version.

Symptoms:

python -c "import torch; print(torch.__version__)"
# Shows: 2.6.0+cu124 or similar with "cu" in version

Solution: Same as above - reinstall scottt's wheel. The "cu" in the version indicates CUDA.

"rocminfo" shows errors

Cause: HSA runtime issues on gfx1151 (common, often non-fatal).

Note: This may not prevent PyTorch from working. The gfx1151 wheels handle HSA differently.

torch.cuda.is_available() returns False (but worked before)

Cause: The virtual environment got corrupted by conflicting packages. This often happens when pip installs dependencies that pull in CUDA torch, overwriting the gfx1151 wheel.

Symptoms:

python -c "import torch; print(torch.cuda.is_available())"
# Returns: False (even though torch.__version__ looks correct)

Solution: Create a fresh virtual environment:

# Remove corrupted venv
rm -rf ~/ComfyUI/venv

# Create fresh venv with Python 3.11
cd ~/ComfyUI
python3.11 -m venv venv
source venv/bin/activate

# Install scottt's wheel first
pip install --upgrade pip wheel setuptools
pip install https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl

# Install other dependencies with --no-deps to prevent torch replacement
pip install --no-deps torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1
pip install numpy pillow scipy aiohttp yarl pyyaml tqdm psutil requests
pip install --no-deps kornia spandrel einops transformers tokenizers sentencepiece safetensors torchsde
pip install typing-inspection kornia_rs httpx "huggingface-hub<1.0,>=0.34.0" packaging regex trampoline
pip install pydantic pydantic-settings annotated-types pydantic-core python-dotenv
pip install alembic SQLAlchemy greenlet Mako av opencv-python-headless
pip install comfyui-frontend-package comfyui-workflow-templates comfyui-embedded-docs

# Verify
python -c "import torch; print('CUDA:', torch.cuda.is_available())"
# Should return: CUDA: True

torchaudio errors (libhipblas.so.2 not found)

Cause: The ROCm 6.1 torchaudio package requires libhipblas.so.2, but scottt's torch wheel uses ROCm 6.5 which has a different library version.

Symptoms:

OSError: libhipblas.so.2: cannot open shared object file: No such file or directory
Cannot import nodes_audio.py module for custom nodes

Impact: Audio-related features in ComfyUI won't work (ACE model, MMAudio VAE, Whisper). All image/video generation features work normally.

Solution: This is expected and can be safely ignored. To suppress the warnings:

source ~/ComfyUI/venv/bin/activate
pip uninstall -y torchaudio

Out of Memory (OOM) errors

Cause: Strix Halo shares system RAM with the GPU.

Solutions:

Close other applications
Use smaller models
Reduce batch sizes in ComfyUI
Increase swap space

Technical Details

Architecture Comparison

Architecture	Family	Examples	ROCm Support
gfx900	Vega	Vega 56/64	Official
gfx906	Vega	MI50	Official
gfx908	CDNA	MI100	Official
gfx90a	CDNA2	MI210/250	Official
gfx1030	RDNA2	RX 6800	Official
gfx1100	RDNA3	RX 7900	Official
gfx1151	RDNA3.5	Strix Halo	Community

Why Python 3.11?

The self-contained wheel from scottt is only built for Python 3.11:

torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl

The Python 3.12 wheels from the CloudFront nightly index have the broken rocm_sdk package that tries to load hipsparselt. Until this is fixed upstream, Python 3.11 is required.

Self-Contained vs SDK-Based Wheels

Aspect	Self-Contained (scottt)	SDK-Based (CloudFront)
ROCm Libraries	Bundled in wheel	External rocm_sdk package
hipsparselt	Not required	Required (but missing)
Size	Larger (~700MB)	Smaller wheel + SDK
Python	3.11 only	3.12
Status	✅ Working	❌ Broken

AOTriton Support

The gfx1151 wheels include AOTriton with experimental gfx1151 support, enabling:

torch.nn.functional.scaled_dot_product_attention
Flash Attention (experimental)
Optimized transformer operations

Download Links & Wheel Sources

Recommended: scottt's Self-Contained Wheels (Working)

These wheels bundle all ROCm libraries internally and don't require the broken rocm_sdk package.

Platform	Python	Wheel	Download
Linux	3.11	torch-2.7.0a0+gitbfd8155	Download
Windows	3.12	torch-2.7.0a0+git3f903c3	Download
Windows	3.12	torchvision-0.22.0+9eb57cd	Download
Windows	3.12	torchaudio-2.6.0a0+1a8f621	Download

Release Page: github.com/scottt/rocm-TheRock/releases/tag/v6.5.0rc-pytorch

CloudFront Nightly Wheels (Broken - hipsparselt issue)

⚠️ WARNING: These wheels have a bug in the rocm_sdk package that tries to load hipsparselt which doesn't exist for gfx1151. Use scottt's wheels instead.

Index URL: https://d2awnip2yjpvqn.cloudfront.net/v2/gfx1151/

Available packages (if the hipsparselt issue gets fixed):

Package	Versions Available	Python
torch	2.7.0a0 - 2.10.0a0	3.11, 3.12, 3.13
torchvision	0.22.0 - 0.25.0a0	3.11, 3.12, 3.13
torchaudio	2.7.0 - 2.10.0a0	3.11, 3.12, 3.13
pytorch-triton-rocm	Various	3.11, 3.12, 3.13
rocm-sdk	Various	3.11, 3.12

Browse packages: d2awnip2yjpvqn.cloudfront.net/v2/gfx1151/

Official PyTorch ROCm Wheels (No gfx1151 support)

❌ NOT WORKING FOR gfx1151: These wheels don't include gfx1151 kernels.

Index URL: https://download.pytorch.org/whl/rocm6.x

Only useful for torchvision/torchaudio when installed with --no-deps.

Direct Download Commands

# Linux (Python 3.11) - RECOMMENDED
pip install https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl

# Windows (Python 3.12)
pip install https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torch-2.7.0a0+git3f903c3-cp312-cp312-win_amd64.whl
pip install https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torchvision-0.22.0+9eb57cd-cp312-cp312-win_amd64.whl
pip install https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch/torchaudio-2.6.0a0+1a8f621-cp312-cp312-win_amd64.whl

# CloudFront nightly (if hipsparselt issue gets fixed)
pip install torch torchvision torchaudio --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx1151/

Sources & References

Official Documentation

Community Resources (gfx1151 Support)

ROCm/TheRock Discussion #655 - Self-contained gfx1151 PyTorch wheels
ROCm/TheRock Discussion #244 - gfx1151 PyTorch working with hipBLASLt
scottt/rocm-TheRock Releases - Working wheel download
weiziqian/rocm_pytorch_docker_gfx1151 - Docker setup for gfx1151
CloudFront gfx1151 Index - Nightly wheels (currently broken)

Issue Trackers

ROCm Issue #5643 - hipBLASLt on gfx1151
ROCm Issue #5404 - AOTriton for gfx1151
hipBLASLt Issue #1243 - Architecture support

Articles

PyTorch with ROCm 7 for Windows on Strix Halo - Wei Lu on Medium
Strix Halo LLM Tracker - Comprehensive status tracking

Files in This Repository

File	Description
`install_comfyui_amd_ryzen.sh`	Main installation script
`run_comfy_amd_ryzen_ai.sh`	Generated launcher (after install)
`rocm_comfyui_diagnostics.sh`	ROCm diagnostic tool
`cleanup_rocm_pytorch_comfyui.sh`	Cleanup script
`README.md`	This documentation

Known Limitations

Python 3.11 required - The working wheel is only available for Python 3.11
VRAM limitation - Some users report PyTorch can only use ~32GB even with 64GB available
hipBLASLt fallback - Performance may be reduced as hipBLASLt falls back to hipBLAS
Flash Attention - Experimental support, may not work for all models
torchaudio incompatibility - The ROCm 6.1 torchaudio requires libhipblas.so.2 which may not be available; audio features (ACE, MMAudio, Whisper) will be disabled but image/video generation works fine
CUDA_VISIBLE_DEVICES interference - If your shell has CUDA_VISIBLE_DEVICES="" set (common for NVIDIA setups), it will hide the GPU from PyTorch ROCm. The run script now uses unset CUDA_VISIBLE_DEVICES to handle this
Venv corruption - Installing packages can sometimes corrupt the venv by replacing the gfx1151 torch wheel with CUDA torch. Solution: delete venv and recreate from scratch

Contributing

If you find issues or improvements for gfx1151 support:

Test changes on actual Ryzen AI Max hardware
Document any new findings
Submit issues to the appropriate upstream projects

License

Scripts in this repository are provided as-is for educational and personal use. ComfyUI and PyTorch have their own respective licenses.

Last Updated: December 2025 Tested On: AMD Ryzen AI Max 395, Ubuntu 24.04, ROCm 6.5/7.1, Python 3.11 Working Wheel: torch-2.7.0a0+gitbfd8155-cp311-cp311-linux_x86_64.whl from scottt Status: ✅ ComfyUI running with GPU acceleration on gfx1151

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

ComfyUI Installer for AMD Ryzen AI Max 395 (gfx1151)

Table of Contents

Overview

Hardware Specifications

The Problem

Symptoms

Root Cause Analysis

The hipsparselt Error

The Solution

Self-Contained PyTorch Wheel (Recommended)

Wheel Comparison

The Fix Applied

Why Dependencies Break Things

Problem 1: requirements.txt

Problem 2: Transitive Dependencies (kornia)

The Solution

Requirements

System Requirements

Python 3.11 Installation (Ubuntu 24.04)

ROCm Requirements

Install ROCm (if not already installed)

Installation

Quick Start

What the Script Does

Post-Installation Fixes

Environment Variables

Usage

Starting ComfyUI

Launcher Script Environment

Debugging

Troubleshooting

"No HIP GPUs are available"

"invalid device function"

"Unknown rocm library 'hipsparselt'"

"python3.11 not found"

Wrong torch version installed

"Found no NVIDIA driver on your system"

"rocminfo" shows errors

torch.cuda.is_available() returns False (but worked before)

torchaudio errors (libhipblas.so.2 not found)

Out of Memory (OOM) errors

Technical Details

Architecture Comparison

Why Python 3.11?

Self-Contained vs SDK-Based Wheels

AOTriton Support

Download Links & Wheel Sources

Recommended: scottt's Self-Contained Wheels (Working)

CloudFront Nightly Wheels (Broken - hipsparselt issue)

Official PyTorch ROCm Wheels (No gfx1151 support)

Direct Download Commands

Sources & References

Official Documentation

Community Resources (gfx1151 Support)

Issue Trackers

Articles

Files in This Repository

Known Limitations

Contributing

License