mirror of
https://github.com/remsky/Kokoro-FastAPI.git
synced 2025-08-05 16:48:53 +00:00
b064066b9e
Tested with rocm 6.3.1. Known issues with this commit: * The generated docker image becomes quite big * Suppressed MIOPEN warnings, which might not be a good idea * No windows support, at least not tested * Incomplete documentation * Not thoroughly tested
1.8 KiB
1.8 KiB
UV Setup
Deprecated notes for myself
Structure
docker/
├── cpu/
│ ├── pyproject.toml # CPU deps (torch CPU)
│ └── requirements.lock # CPU lockfile
├── gpu/
│ ├── pyproject.toml # GPU deps (torch CUDA)
│ └── requirements.lock # GPU lockfile
├── rocm/
│ ├── pyproject.toml # ROCM deps (torch ROCM)
│ └── requirements.lock # ROCM lockfile
└── shared/
└── pyproject.toml # Common deps
Regenerate Lock Files
CPU
cd docker/cpu
uv pip compile pyproject.toml ../shared/pyproject.toml --output-file requirements.lock
GPU
cd docker/gpu
uv pip compile pyproject.toml ../shared/pyproject.toml --output-file requirements.lock
ROCM
cd docker/rocm
uv pip compile pyproject.toml ../shared/pyproject.toml --output-file requirements.lock
Local Dev Setup
CPU
cd docker/cpu
uv venv
.venv\Scripts\activate # Windows
uv pip sync requirements.lock
GPU
cd docker/gpu
uv venv
.venv\Scripts\activate # Windows
uv pip sync requirements.lock --extra-index-url https://download.pytorch.org/whl/cu121 --index-strategy unsafe-best-match
ROCM
cd docker/rocm
uv venv
source .venv/bin/activate
# not tested on Windows
#.venv\Scripts\activate # Windows
uv pip sync requirements.lock --extra-index-url https://download.pytorch.org/whl/rocm6.2
Run Server
# From project root with venv active:
uvicorn api.src.main:app --reload
Docker
CPU
cd docker/cpu
docker compose up
GPU
cd docker/gpu
docker compose up
ROCM
cd docker/rocm
docker compose up
Known Issues
- Module imports: Run server from project root
- PyTorch CUDA: Always use --extra-index-url and --index-strategy for GPU env