Introduces a third hardware mode alongside CUDA and CPU: ROCm (AMD GPU).
AMD GPUs present as CUDA devices under PyTorch ROCm, so the existing GPU
path is reused with minimal changes — the main additions are wheel management,
device detection, and suppressing flash_attn (unsupported on ROCm).
- server/vibevoice_server.py: extend _resolve_device() to recognise 'rocm'
(auto-detected via torch.version.hip); add _torch_device() helper that maps
'rocm' → 'cuda' for all PyTorch API calls; apply GPU optimisations for both
cuda and rocm in _init_model(); always use sdpa on ROCm; propagate
_torch_device() to _load_voice_presets() map_location.
- server/start.sh: add --rocm flag; sync .venv-rocm with uv sync --no-sources
then replace torch with the ROCm 6.2 wheel via uv pip install; set
VIBEPOD_DEVICE=rocm for uvicorn.
- server/pyproject.toml: register pytorch-rocm62 index (explicit); add
.venv-rocm to ruff excludes.
- package.json: add dev:rocm and dev:server:rocm scripts.
- README.md: document ROCm mode, prerequisites (RX 6000+, ROCm 6.2+, Linux),
and new commands; expand CUDA vs CPU section to CUDA vs CPU vs ROCm.
https://claude.ai/code/session_0168pSswiaoEf6LEx6UQWfBu
- Add ruff>=0.11.0 as dev dependency via [dependency-groups]
- Configure [tool.ruff]: line-length=100, py310 target, LF line endings
- Lint rules: E, F, UP, B, SIM, I (ignoring E501/B905)
- Add lint:server and lint:server:fix scripts to root package.json
- Update format/format:check to also run ruff for server/
Switch vibevoice dependency from microsoft/VibeVoice to JezzWTF/VibeVoice
fork (commit e76701f) which contains the async decode + parallel CFG
optimisations directly in generate(). Removes the instance-method
patching approach (vibevoice_generate_patch.py deleted).
server/vibevoice_server.py:
- Add _cfg_executor (ThreadPoolExecutor, 1 worker) alongside _decode_executor
- _install_cpu_pipeline_optimizations now sets both executors directly as
model._vibepod_decode_executor and model._vibepod_cfg_executor
- Both executors shut down in lifespan on exit
- Remove vibevoice_generate_patch import/install (no longer needed)
server/pyproject.toml:
- vibevoice source changed to git+https://github.com/JezzWTF/VibeVoice.git
- No machine-local paths; works identically on any clone
Claude
LyAhn