**✅ SUCCESS: oobabooga now works with ROCm 7.2 on Windows + AMD GPUs!**

[brubeedoobee]

Update: Reported to AMD team as ROCm Issue #5871

Sharing my experience getting oobabooga running on AMD 7900 XTX with Windows. Got very close but hit an AMD bug. Documenting for others.

AMD Radeon RX 7900 XTX + Windows 11 + ROCm 7.1.1: Model Loads Successfully, Crashes on Text Generation
TL;DR: Model loads to GPU perfectly, crashes immediately on first generation attempt. Root cause: HIP runtime bug in amdhip64_7.dll. Awaiting AMD fix.
Hardware/Software Configuration

GPU: AMD Radeon RX 7900 XTX (24GB VRAM)
OS: Windows 11
Driver: AMD Software PyTorch Edition 25.20.01.17 (driver store version 32.0.22001.17002)
Python: 3.12.10
PyTorch: 2.9.0+rocmsdk20251116 (ROCm 7.1.1)
oobabooga: Latest main branch (January 2026)

Installation Steps & Reproduction

1. Install Python 3.12

Download from python.org
During install, check "Add Python to PATH" and "Disable path length limit"

2. Clone oobabooga
   cmdcd D:
   git clone https://github.com/oobabooga/text-generation-webui
   cd text-generation-webui
3. Create virtual environment
   cmdpython -m venv venv
   venv\Scripts\activate.bat
4. Install oobabooga base requirements
   cmdpip install -r requirements\full\requirements.txt
5. Install ROCm SDK + PyTorch (in the activated venv)
   cmdpip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/rocm_sdk_core-0.1.dev0-py3-none-win_amd64.whl

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/rocm_sdk_devel-0.1.dev0-py3-none-win_amd64.whl

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/rocm_sdk_libraries_custom-0.1.dev0-py3-none-win_amd64.whl

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/rocm-0.1.dev0.tar.gz

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/torch-2.9.0+rocmsdk20251116-cp312-cp312-win_amd64.whl

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/torchaudio-2.9.0+rocmsdk20251116-cp312-cp312-win_amd64.whl

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.1.1/torchvision-0.24.0+rocmsdk20251116-cp312-cp312-win_amd64.whl
6. Verify PyTorch sees GPU
cmdpython -c "import torch; print(torch.cuda.is_available())"
python -c "import torch; print(torch.cuda.get_device_name(0))"
Expected output: True and AMD Radeon RX 7900 XTX
7. Start oobabooga
cmdpython server.py
8. Reproduce the crash

Open browser to http://127.0.0.1:7860/
Go to Model tab
In "Download model or LoRA", enter: mistralai/Mistral-7B-Instruct-v0.2
Click Download (wait for completion, ~15GB)
Select the model from dropdown
Model loader: Transformers
attn-implementation: eager (or sdpa, both crash)
Click Load → ✅ Succeeds ("Successfully loaded" message appears)
Go to Chat tab
Type any message (e.g., "Hello")
Click Generate → ❌ Python crashes silently
Check Windows Event Viewer for amdhip64_7.dll crash

What Works ✅

PyTorch correctly detects 7900 XTX
Model downloads successfully
Model loads to GPU successfully (completes in ~22 seconds)
Shows "Successfully loaded" message

What Fails ❌

Text generation crashes immediately on first GPU compute
Python process terminates with no console error
Windows Event Viewer shows amdhip64_7.dll access violation

Root Cause

Windows Event Viewer shows:

Faulting module: amdhip64_7.dll (version 10.0.3581.0)
Exception code: 0xc0000005 (Access Violation)
Fault offset: 0x00000000002ab8ba
Module path: D:\text-generation-webui\venv\Lib\site-packages\_rocm_sdk_core\bin\amdhip64_7.dll

This is AMD's HIP runtime library crashing during the first GPU compute operation.
Attempted Fixes That Don't Help

❌ Disabling Windows TDR (registry edits to TdrDelay, TdrLevel)
❌ Changing attention implementation (eager vs sdpa)
❌ Environment variables (HSA_OVERRIDE_GFX_VERSION, HSA_ENABLE_SDMA=0, AMD_DIRECT_DISPATCH=0)
❌ Different models or quantization levels
❌ Lowering truncation length

Status
This appears to be a known bug in ROCm 7.1.1 for Windows with RDNA3 GPUs (similar crashes reported in llama.cpp GitHub issues #17429).
The good news: We're extremely close - model loading works perfectly. This should be fixed in the next AMD driver/ROCm release.
Workarounds:

Wait for AMD driver update (25.30.x or ROCm 7.2+)
Dual-boot Linux - ROCm works perfectly on Linux
CPU-only mode - Set CUDA_VISIBLE_DEVICES=-1 (slow but functional)

For AMD/ROCm Team
If you need additional diagnostics (core dumps, debug logs, etc.) from a Windows 11 + 7900 XTX setup, happy to provide them.

[brubeedoobee]

Update: oobabooga now works with ROCm 7.2 on Windows!*

I successfully got text-generation-webui running on AMD hardware with the latest ROCm release. Here's the current status:

Working Configuration

Hardware:

• AMD Radeon AI PRO R9700 (32GB VRAM) - also tested with RX 7900 XTX (24GB)
• Windows 11
• Ryzen 9 9950X

Software:

• ROCm 7.2 (rocm-rel-7.2)
• PyTorch 2.9.1+rocmsdk20260116
• Python 3.12
• oobabooga main branch (January 2026)

Installation Steps

1. Create virtual environment:

python -m venv rocm_env
rocm_env\Scripts\activate

2. Install ROCm 7.2 + PyTorch:

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_core-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_devel-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_libraries_custom-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm-7.2.0.dev0.tar.gz

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torch-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchaudio-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchvision-0.24.1+rocmsdk20260116-cp312-cp312-win_amd64.whl

3. Install oobabooga requirements:

cd text-generation-webui
pip install -r requirements/full/requirements_amd.txt

4. Remove torchao (compatibility issue):

pip uninstall torchao -y

5. Launch:

python server.py --listen --api

Key Settings for Optimal Performance

• Model loader: Transformers (ExLlamaV2 doesn't work on ROCm Windows yet)
• attn-implementation: sdpa (NOT eager - this is critical for GPU utilization)
• Avoid Mixtral models - they have tensor size mismatches on ROCm Windows

Performance

Achieving ~22-24 tokens/s on 7B models with proper GPU utilization using SDPA attention.

Known Issues

1. Dual-GPU bug: If you have integrated graphics (AMD APU + discrete GPU), ROCm crashes. Workaround: Disable iGPU in BIOS. Reported to AMD: ROCm Issue #5909

2. torchao incompatibility: Must be uninstalled - causes ModuleNotFoundError: No module named 'torch._C._distributed_c10d'

3. Mixtral models crash with RuntimeError: The size of tensor a (32) must match the size of tensor b (0)

Bottom Line

ROCm 7.2 + oobabooga works on Windows with AMD GPUs! Use Transformers loader with SDPA attention for best results.
Good to post?do you mention we tested on the 7900 and r9700?2:37 PMYou're right - let me revise the hardware section to be clearer:
markdownHardware:

• Tested on two GPUs:
  • AMD Radeon RX 7900 XTX (24GB VRAM) ✅
  • AMD Radeon AI PRO R9700 (32GB VRAM) ✅
• Windows 11
• Ryzen 9 9950X (with integrated graphics)
  And update the dual-GPU section to:
  markdown## Known Issues

1. Dual-GPU bug: Both the RX 7900 XTX and R9700 crash when integrated graphics (iGPU) is enabled. ROCm's HIP runtime has an enumeration bug with AMD APU + discrete GPU combinations on Windows. Workaround: Disable iGPU in BIOS. The HIP_VISIBLE_DEVICES environment variable does not work as a workaround on Windows. Reported to AMD: ROCm Issue #5909
   This makes it clear you tested both cards and both had the same dual-GPU issue. Better?can you integrate it in? It's already hard to copy2:38 PMHere's the complete integrated post ready to copy:
   markdownUpdate: oobabooga now works with ROCm 7.2 on Windows!

I successfully got text-generation-webui running on AMD hardware with the latest ROCm release. Here's the current status:

Working Configuration

Hardware:

• Tested on two GPUs:
  • AMD Radeon RX 7900 XTX (24GB VRAM) ✅
  • AMD Radeon AI PRO R9700 (32GB VRAM) ✅
• Windows 11
• Ryzen 9 9950X (with integrated graphics)

Software:

• ROCm 7.2 (rocm-rel-7.2)
• PyTorch 2.9.1+rocmsdk20260116
• Python 3.12
• oobabooga main branch (January 2026)

Installation Steps

1. Create virtual environment:

python -m venv rocm_env
rocm_env\Scripts\activate

2. Install ROCm 7.2 + PyTorch:

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_core-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_devel-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_libraries_custom-7.2.0.dev0-py3-none-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm-7.2.0.dev0.tar.gz

pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torch-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchaudio-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchvision-0.24.1+rocmsdk20260116-cp312-cp312-win_amd64.whl

3. Install oobabooga requirements:

cd text-generation-webui
pip install -r requirements/full/requirements_amd.txt

4. Remove torchao (compatibility issue):

pip uninstall torchao -y

5. Launch:

python server.py --listen --api

Key Settings for Optimal Performance

• Model loader: Transformers (ExLlamaV2 doesn't work on ROCm Windows yet)
• attn-implementation: sdpa (NOT eager - this is critical for GPU utilization)
• Avoid Mixtral models - they have tensor size mismatches on ROCm Windows

Performance

Achieving ~22-24 tokens/s on 7B models with proper GPU utilization using SDPA attention.

Known Issues

1. Dual-GPU bug: Both the RX 7900 XTX and R9700 crash when integrated graphics (iGPU) is enabled. ROCm's HIP runtime has an enumeration bug with AMD APU + discrete GPU combinations on Windows. Workaround: Disable iGPU in BIOS. The HIP_VISIBLE_DEVICES environment variable does not work as a workaround on Windows. Reported to AMD: ROCm Issue #5909

2. torchao incompatibility: Must be uninstalled - causes ModuleNotFoundError: No module named 'torch._C._distributed_c10d'

3. Mixtral models crash with RuntimeError: The size of tensor a (32) must match the size of tensor b (0)

Bottom Line

ROCm 7.2 + oobabooga works on Windows with AMD GPUs! Use Transformers loader with SDPA attention for best results.