Issue evidence · Serve · vllm-project/vllm

Startup memory-profiling assertion aborts whenever free memory grows

GitHub issue: https://github.com/vllm-project/vllm/issues/56830

Reproduced

If another process on the same GPU releases memory while vLLM is profiling at startup, vLLM aborts with an AssertionError.

What Badgr ran

Ran the reporter's co-tenant scenario on an RTX 4090 through Badgr in one container (vLLM 0.30.0, facebook/opt-125m, --gpu-memory-utilization 0.2): a PyTorch process held 8 GiB, vLLM started, and the holder released its memory right after "Model loading took". A control run did not release.

What came back

  • Release run: AssertionError "Error in memory profiling. Initial free memory 14.69 GiB, current free memory 22.27 GiB"; startup aborted.
  • Control run (no release): the server started normally.

The issue used v0.29.0 on an H100 NVL with two containers; this was one container on a 4090 with v0.30.0. A safer deployment is to keep the GPU's other tenants from releasing memory during startup, or to start vLLM alone.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence