[jasonwang-runpod]

Problem

vLLM starts before the SDK, so unrecognized startup failures stay in worker logs instead of reaching queued requests.

Solution

Run startup under SDK prestart supervision while keeping jobs queued until startup settles. Preserve existing retries and startup_error responses for known configuration failures. Unrecognized failures now fail one queued job with prestart_failed, include vLLM output, and exit so the platform can restart the worker.

Blocked on a stable Runpod SDK 2.0.x release on PyPI.

Testing

• 147 worker tests, lint, and vLLM flag checks passed.
• Live A40 / vLLM 0.31.0 smoke against SDK main 4473f30a: queue prompt, chat, streaming, cold starts, and startup failure reporting.
• Hosted API smoke used a fake vLLM backend; existing local-input and realtime SDK limitations remain.