Upgrade Sonar

Test an upgrade with the production model and configuration before you replace a running replica.

Record the current deployment

Save the Sonar version, commit, model revision, tokenizer revision, PyTorch version, CUDA or ROCm version, and complete server arguments.

Export a small set of representative requests with expected output properties. Exact generated text can vary because of floating-point and batching changes.

Upgrade a wheel installation

uv pip install --upgrade aphrodite-engine --torch-backend=cu130

For a nightly build:

uv pip install --upgrade aphrodite-engine \
  --extra-index-url https://sonar.dphn.ai/nightly

Use a fresh virtual environment when core dependencies change. Keep the old environment until the replacement passes validation.

Upgrade a source checkout

git fetch origin
git switch main
git pull --ff-only
uv pip install --editable . --torch-backend=cu130

Regenerate the CMake preset when Python, CUDA, the compiler, or the virtual environment changes. Rebuild native extensions after CMake or csrc/ changes.

Validate compatibility

Check startup warnings for renamed flags and deprecated values. Compare the generated server argument reference with the saved command.

Run chat, streaming, structured output, tool calling, LoRA, and multimodal requests that the service supports. Compare latency and memory with the saved baseline.

Roll back

Keep the old image or virtual environment and its model revision. Remove the failed replacement from traffic, then restore the old replica.

Do not reuse a compilation cache across incompatible PyTorch or CUDA builds when a rollback produces loader or symbol errors. Use separate cache keys for each build stack.