AMD ROCm Windows Helpers & Utilities

A curated collection of automation scripts, runtime patches, and technical guides designed to deploy and run modern generative AI workloads natively on Windows using AMD ROCm™.

Note

Project Context & Disclaimer:

• Proof of Concept: This collection of setup scripts, utilities, and technical documentation represents an early proof of concept created to overcome current Windows AMD ROCm hurdles. While these setups have been tested and verified on local AMD hardware (such as Radeon RX 7000 and RX 9000 series GPUs), hardware configurations, Windows driver versions, and environments can differ significantly—what works seamlessly in one test environment may require adjustments or further testing on yours. Community feedback, issue reports, and real-world testing are warmly welcomed!
• AI-Assisted Development: These scripts, patches, and documentation were created with AI assistance. In the spirit of complete transparency: if you prefer not to use AI-assisted code, please feel free to pass on these utilities and wait for official upstream Windows ROCm fixes from the respective project maintainers.

---

Overview

Historically, running advanced generative AI workflows on AMD Radeon™ hardware on Windows required dual-booting Linux, setting up heavy WSL2 virtual machines, or settling for constrained DirectML fallbacks.

With AMD's official release of native Windows ROCm PyTorch wheels (such as rocm-rel-7.2.1 / PyTorch 2.9.1+rocm7.2.1), native execution on Windows is now a reality. However, developers and creators frequently run into subtle ecosystem blockers:

1. PyPI Dependency Overwrite Traps: Standard package installations and requirements.txt files frequently overwrite ROCm-enabled PyTorch builds with standard PyPI CPU or CUDA wheels.
2. Missing rocm_sdk Windows Stubs: AMD's rocm_sdk Python module expects Linux-oriented shared libraries, causing process initialization to fail on Windows.
3. Missing Distributed (USE_DISTRIBUTED=0) Extensions: Windows ROCm PyTorch wheels lack torch._C._distributed_c10d, causing libraries that unconditionally import distributed tensors (like torchao, diffusers, and accelerate) to crash immediately.
4. Dynamic DLL Path Resolution: Windows cannot resolve ROCm runtime DLLs without explicit environment PATH configuration prior to Python execution.
5. Cross-Attention Crashes: Default split cross-attention implementations crash on Windows ROCm unless forced to native Scaled Dot-Product Attention (SDPA).

This repository provides dedicated PowerShell automation scripts and in-depth technical documentation to solve these problems cleanly and repeatably.

---

Utilities & Documentation

Component	Description	Documentation	Utility Script
ComfyUI Native ROCm	Native Windows ComfyUI manager with dependency pinning, runtime DLL injection, SDPA enforcement, and background execution.	ComfyUI Guide	comfyui.ps1
AI-Toolkit Standalone	Standalone LoRA training environment (Flux, Chroma, SDXL, SD 1.5) with distributed tensor emulation and torchao patching.	AI-Toolkit Guide	Setup-AmdAiToolkit.ps1

---

1. ComfyUI Native AMD ROCm Helper

• Documentation: comfyui-rocm-windows/AMD_Windows_ROCm_ComfyUI_Native.md
• Utility Script: comfyui-rocm-windows/comfyui.ps1

Key Capabilities:

• Automated Virtualenv Bootstrap: Automatically provisions Python 3.12 virtual environments and strictly pins official AMD ROCm 7.2.1 PyTorch wheels.
• In-Memory & SDK Patching: Inspects and safely stubs missing Windows library entries in rocm_sdk\_dist_info.py.
• Protected Node Updates: Safely updates ComfyUI and custom nodes while preventing PyPI from silently replacing ROCm PyTorch wheels with CPU/CUDA wheels.
• Git Ownership Remediation: Detects Git dubious ownership errors common across user migrations or multi-drive setups and applies ACL repairs (takeown / icacls).
• Background Daemon Management: Supports starting ComfyUI as a detached background service with PID tracking, graceful shutdown, and split stdout/stderr logging.

---

2. Standalone AI-Toolkit Setup & Patch Suite

• Documentation: amd-ai-toolkit-standalone/AMD_AI_TOOLKIT_STANDALONE.md
• Utility Script: amd-ai-toolkit-standalone/Setup-AmdAiToolkit.ps1

Key Capabilities:

• Distributed Tensor Emulation: Emulates missing c10d C++ distributed extensions on Windows, allowing torchao and diffusers to import cleanly without crashing.
• torchao & accelerate Runtime Patching: Automatically applies surgical patches to resolve Windows import blockers.
• Full Model Architecture Support: Enables training LoRAs for Flux.1, Chroma, SDXL, and SD 1.5 directly on Windows with AMD hardware.
• Zero-Wait Standalone Execution: Designed for engineers and creators who prefer working in terminal environments or standalone automation pipelines without waiting for upstream PR merges.

---

Hardware & Architecture Compatibility Matrix

GPU Model Family	Architecture	Target HSA_OVERRIDE_GFX_VERSION	Backend	Status
Radeon™ RX 9070 XT / 9070	RDNA 4 (Navi 48)	11.0.0	ROCm 7.2.1	Verified / Primary Test System
Radeon™ RX 7900 XTX / XT / GRE	RDNA 3 (Navi 31)	11.0.0	ROCm 7.2.1	Verified / Production
Radeon™ RX 7800 XT / 7700 XT	RDNA 3 (Navi 32)	11.0.1	ROCm 7.2.1	Supported
Radeon™ RX 7600 XT / 7600	RDNA 3 (Navi 33)	11.0.2	ROCm 7.2.1	Supported
Radeon™ RX 6950 XT / 6900 XT / 6800	RDNA 2 (Navi 21)	10.3.0	ROCm 7.2.1	Supported
Radeon™ RX 6700 XT / 6750 XT	RDNA 2 (Navi 22)	10.3.0	ROCm 7.2.1	Supported

---

System Prerequisites

Before using any of the helpers in this repository, ensure your Windows host meets the following prerequisites:

1. Operating System: Windows 10 (version 21H2 or newer) or Windows 11 (64-bit).
2. AMD Graphics Driver: AMD Software: Adrenalin Edition 24.x+ or official ROCm-supported driver.
3. AMD ROCm™ & HIP SDK for Windows:
   • Install AMD ROCm 7.2 (or compatible 7.x) & HIP SDK from the AMD Developer Portal.
   • Expected default installation path: C:\Program Files\AMD\ROCm\7.2\ or C:\Program Files\AMD\ROCm\.
4. Python: Python 3.12 (64-bit) from python.org. Ensure "Add python.exe to PATH" is enabled.
5. Git: Git for Windows from git-scm.com.
6. PowerShell: PowerShell 7+ (pwsh) is strongly recommended.

---

Quick Start

Running ComfyUI

To set up or launch ComfyUI with native AMD ROCm:

# Run from PowerShell 7+
./comfyui-rocm-windows/comfyui.ps1 -Help

Refer to the ComfyUI Native Guide for full commands (run, start, stop, update, fix-ownership).

Setting Up AI-Toolkit

To provision a standalone AI-Toolkit training environment:

# Run from PowerShell 7+
./amd-ai-toolkit-standalone/Setup-AmdAiToolkit.ps1

Refer to the AI-Toolkit Standalone Guide for training configuration examples and troubleshooting.