[zhongdaor-nv]

Summary

When a cluster model alias resolves to local weights, TensorRT-LLM loads /model while clients request engine.served_model_name (for example, Qwen/Qwen3-0.6B). The Dynamo sidecar previously registered /model, so requests using the configured served name failed with 404 unknown_model.

Use the existing served-name resolver for the sidecar's --model-path, retaining the local-path and pinned-HF-ID fallbacks when no served name is configured. Add regression coverage for aggregate, prefill, and decode launches and document that this name must resolve to matching tokenizer/config artifacts; arbitrary display-only aliases are unsupported by this sidecar interface.

Verified that the ai-dynamo-runtime==1.6.0.dev20261016 sidecar CLI has no separate served-name option. The upstream registration code inspected at 27541b1 uses model_path as the registration name.

Validation

• Regression tests failed with the original implementation and passed with the fix.
• make check: Ruff, formatting, type checking, and generated-doc checks passed; 5,181 tests passed, 3 skipped, 6 deselected, and 11 xfailed, including launch snapshot checks.
• Ran make snapshots and reviewed the generated launch plans; no snapshot changes are needed for the existing examples.
• uv run srtctl dry-run -f examples/trtllm/dynamo-sidecar-disagg.yaml passed.
• No GPU or Slurm cluster execution was performed.

[github-actions]

src/srtctl diff: +3 -3