- The canonical local loop is image-first:
make vllm-sr-devvllm-sr serve --image-pull-policy never
- AMD loop:
make vllm-sr-dev VLLM_SR_PLATFORM=amdvllm-sr serve --image-pull-policy never --platform amd
- AMD real-model deployment reference:
- If the task is about deploying a real AMD backend with alias-based routing instead of smoke validation, prefer:
make agent-serve-local ENV=amd AGENT_SERVE_CONFIG=deploy/recipes/balance.yaml- plus the backend/network setup from
deploy/amd/README.md
- Feature-gate smoke defaults to the env-specific manifest config under
e2e/config/config.agent-smoke.*.yaml - Use
AGENT_SERVE_CONFIGonly when you need to override that default - Treat
src/vllm-sr/**,tools/make/docker.mk, andsrc/vllm-sr/Dockerfile*as startup-chain changes - For these changes, run:
- CLI fast tests
- CLI integration tests
- local smoke
- affected E2E profiles