160 lines
4.4 KiB
Markdown
160 lines
4.4 KiB
Markdown
# WhisperX ROCm Fork
|
|
|
|
This is a fork of [m-bain/whisperX](https://github.com/m-bain/whisperX) configured for AMD ROCm GPUs.
|
|
|
|
## Tested Configuration
|
|
|
|
- **OS**: Debian Sid (trixie/forky)
|
|
- **GPU**: AMD Radeon RX 7700 XT (gfx1101, 12GB VRAM)
|
|
- **ROCm**: 7.1.1
|
|
- **Python**: 3.10
|
|
- **PyTorch**: 2.11.0+rocm7.0 (nightly)
|
|
- **CTranslate2**: 4.6.2 (ROCm build from [paralin/ctranslate2-rocm](https://github.com/paralin/ctranslate2-rocm))
|
|
|
|
## Prerequisites
|
|
|
|
1. ROCm 7.1+ installed at `/opt/rocm`
|
|
2. CTranslate2 built with ROCm support (see [paralin/ctranslate2-rocm](https://github.com/paralin/ctranslate2-rocm))
|
|
|
|
## Installation (Debian Sid)
|
|
|
|
### 1. Clone and setup
|
|
|
|
```bash
|
|
git clone https://github.com/paralin/whisperX-rocm.git ~/whisperx
|
|
cd ~/whisperx
|
|
git checkout rocm
|
|
```
|
|
|
|
### 2. Create venv with uv
|
|
|
|
```bash
|
|
uv venv
|
|
uv pip install -e .
|
|
```
|
|
|
|
### 3. Install ROCm PyTorch
|
|
|
|
```bash
|
|
uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0
|
|
```
|
|
|
|
### 4. Install ROCm CTranslate2
|
|
|
|
First build CTranslate2 with ROCm support (see [paralin/ctranslate2-rocm](https://github.com/paralin/ctranslate2-rocm)), then:
|
|
|
|
```bash
|
|
# Remove PyPI ctranslate2 (has CUDA binaries)
|
|
rm -rf .venv/lib/python3.10/site-packages/ctranslate2*
|
|
|
|
# Install ROCm build
|
|
export CTRANSLATE2_ROOT=/usr/local
|
|
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
|
|
uv pip install --reinstall pybind11 ~/ctranslate2/python
|
|
```
|
|
|
|
## Environment Variables
|
|
|
|
These must be set before running whisperx:
|
|
|
|
```bash
|
|
export HSA_OVERRIDE_GFX_VERSION=11.0.1 # for gfx1101
|
|
export AMDGPU_TARGETS=gfx1101
|
|
export ROCM_PATH=/opt/rocm
|
|
export HIP_VISIBLE_DEVICES=0
|
|
export ROCR_VISIBLE_DEVICES=0
|
|
export LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm/lib:/opt/rocm/lib/llvm/lib:$LD_LIBRARY_PATH
|
|
```
|
|
|
|
## Usage
|
|
|
|
```bash
|
|
# Set environment (add to ~/.bashrc for convenience)
|
|
export HSA_OVERRIDE_GFX_VERSION=11.0.1
|
|
export ROCM_PATH=/opt/rocm
|
|
export HIP_VISIBLE_DEVICES=0
|
|
export LD_LIBRARY_PATH=/usr/local/lib:/opt/rocm/lib:/opt/rocm/lib/llvm/lib:$LD_LIBRARY_PATH
|
|
|
|
cd ~/whisperx
|
|
uv run whisperx audio.wav \
|
|
--language en \
|
|
--model large-v3 \
|
|
--compute_type float16 \
|
|
--device cuda \
|
|
--batch_size 8 \
|
|
--vad_method silero \
|
|
--output_dir ./output \
|
|
--output_format all
|
|
```
|
|
|
|
Note: We use `--device cuda` because ROCm's HIP layer translates CUDA API calls to AMD GPU.
|
|
|
|
## Verify Installation
|
|
|
|
```bash
|
|
# Check PyTorch sees the GPU
|
|
python -c "import torch; print(CUDA:, torch.cuda.is_available()); print(Device:, torch.cuda.get_device_name(0))"
|
|
|
|
# Check CTranslate2
|
|
python -c "import ctranslate2; print(ctranslate2.__version__); print(ctranslate2.get_supported_compute_types(cuda))"
|
|
```
|
|
|
|
Expected output:
|
|
```
|
|
CUDA: True
|
|
Device: AMD Radeon RX 7700 XT
|
|
|
|
4.6.2
|
|
{int8_float16, int8_bfloat16, bfloat16, int8_float32, int8, float16, float32}
|
|
```
|
|
|
|
## GPU Architecture
|
|
|
|
Set `HSA_OVERRIDE_GFX_VERSION` based on your GPU:
|
|
|
|
| GPU | Architecture | HSA_OVERRIDE_GFX_VERSION |
|
|
|-----|--------------|--------------------------|
|
|
| RX 7900 XTX/XT | gfx1100 | 11.0.0 |
|
|
| RX 7800 XT | gfx1101 | 11.0.1 |
|
|
| RX 7700 XT | gfx1101 | 11.0.1 |
|
|
| RX 7600 | gfx1102 | 11.0.2 |
|
|
| RX 6900/6800/6700 | gfx1030 | 10.3.0 |
|
|
| RX 6600 | gfx1032 | 10.3.2 |
|
|
|
|
## Upstream
|
|
|
|
- Original: [m-bain/whisperX](https://github.com/m-bain/whisperX)
|
|
|
|
## Known Issues
|
|
|
|
### Memory Access Fault on Long Audio
|
|
|
|
When transcribing longer audio files (>60s), you may encounter:
|
|
|
|
```
|
|
Memory access fault by GPU node-1 (Agent handle: 0x...) on address 0x...
|
|
Reason: Page not present or supervisor privilege.
|
|
```
|
|
|
|
**Status**: Under investigation. Short clips (~60s) work fine at ~28x realtime with small model.
|
|
|
|
**Workaround**: Process audio in chunks, or use CPU mode for long files.
|
|
|
|
**Working example** (first 60s):
|
|
```python
|
|
from faster_whisper import WhisperModel
|
|
model = WhisperModel("small", device="cuda", compute_type="float16")
|
|
segments, info = model.transcribe("audio.wav", language="en", clip_timestamps=[0, 60])
|
|
```
|
|
|
|
**Search terms for updates**:
|
|
- `"Memory access fault by GPU node" "Page not present or supervisor privilege" ROCm 7.1 PyTorch site:github.com`
|
|
- `"Memory access fault" ROCm CTranslate2 faster-whisper gfx1101`
|
|
|
|
This may be related to:
|
|
- ROCm 7.1.1 + PyTorch nightly (2.11.0+rocm7.0) incompatibility
|
|
- GPU memory fragmentation with longer sequences
|
|
- HIP/ROCm memory management issues with certain operations
|
|
|
|
**Related issue**: https://github.com/ROCm/ROCm/issues/5616 (gfx1103, same error pattern)
|