Troubleshooting

Regional PyPI mirrors

Symptom. A configured PyPI mirror may be missing a required package version or lag behind PyPI. For example, if the mirror does not provide mooncake-transfer-engine-cuda13>=0.3.10, installation fails to resolve the requirement in pyproject.toml. This is an index availability issue; the mirror does not have the required version.

Check. If uv reports that the requirement was not found in the package registry, check the configured mirror for the required wheel. To see whether PyPI has a wheel for your environment, use the download command below.

Fix. Download that one wheel from PyPI, then use --find-links while keeping your existing uv mirror configuration for the other packages. With the Python 3.12 virtual environment from the installation guide active, install pip into it so the downloaded wheel matches that environment:

uv pip install pip
python -m pip download --no-deps --index-url https://pypi.org/simple \
    "mooncake-transfer-engine-cuda13>=0.3.10" -d ./wheels
uv pip install --prerelease=allow --find-links ./wheels -e .   # or "sglang-omni==<version>" in place of -e .

Expected output of uv pip check

Symptom. A CUDA installation finishes, but the dependency check reports version or platform conflicts.

Check. With the SGLang-Omni environment active, run:

uv pip check

Interpret the output. uv pip check and python -m pip check compare the requirements every installed distribution declares with the versions installed next to it. The CUDA install paths leave a few declared ranges unmet on purpose, so the lines in this table are expected and the code they concern runs. Investigate any other line before serving a model.

Install path	Reported line	Why the runtime still works
Every CUDA path	grpcio-health-checking and grpcio-reflection require protobuf>=7.35.1,<8.0.0, protobuf 6.x is installed	SGLang pulls both through smg-grpc-servicer, and tool.uv.override-dependencies in pyproject.toml keeps protobuf in the 6.x series, so import grpc_health.v1.health_pb2 fails with a protobuf VersionError (gencode 7.35.1, runtime 6.33.6). Neither SGLang-Omni nor any module of the installed sglang package imports grpc_health or grpc_reflection.
After uv pip install --no-deps qwen-tts==0.1.1	qwen-tts requires transformers==4.57.3	The project runs Transformers 5.19.0, and sglang_omni/models/qwen3_tts/compat.py shims the API differences, see the Qwen3-TTS cookbook.
Same	qwen-tts requires accelerate==1.12.0	qwen-tts 0.1.1 never imports accelerate.
Same	qwen-tts requires onnxruntime, which is not installed	The CUDA path installs onnxruntime-gpu, which provides the onnxruntime module the speech tokenizer imports.
[fishaudio] extra and the CI image	descript-audiotools requires protobuf<3.20,>=3.9.2	The override keeps protobuf 6.x. audiotools reaches protobuf only through torch.utils.tensorboard, which the installed TensorBoard supports, and the CI import probe imports dac on every run.
Every CUDA path	decord 0.6.0 is not supported on this platform (pip) or The package decord was built for a different platform (uv)	The wheel is published as py3-none-manylinux2010_x86_64, but its WHEEL file records the tag cp36-cp36m. decord loads libdecord.so through ctypes, and the MiniCPM-o video frontend decodes with it on Python 3.12.

The two tools word the lines differently, and uv pip check exits with status 1 whenever it prints one, so its exit status is not a gate. A source checkout has a script that runs the same comparison with the interpreter it is started with and fails only on a line outside this table. The PyPI package does not include it, so run it from the root of the checkout, for example after installing from source, with the environment’s Python:

cd sglang-omni   # the root of the source checkout
python .github/scripts/check_omni_installed_requirements.py

The install commands keep --prerelease=allow from the releases that pinned cuda-tile==1.6.0rc5; SGLang 0.5.23 pins no pre-release package and the CUDA path resolves without the flag. With the flag, uv may take a pre-release of any package when one is newer than the latest release on PyPI, which is why uv pip freeze can list development or release-candidate versions of packages such as einops, safetensors or grpcio.

SGLang 0.5.23 ships manylinux_2_34 wheels, so the CUDA path needs glibc 2.34 or newer, as on Ubuntu 22.04. On an older system uv stops with sglang==0.5.23 has no wheels with a matching platform tag.