Issue evidence · Serve · vllm-project/vllm
GitHub issue: https://github.com/vllm-project/vllm/issues/39072
Not run
Gemma 4 31B omits the required path argument when calling the edit tool through vLLM's gemma4 tool parser.
Tried to serve google/gemma-4-31B-it (31B, about 62 GB, not gated) on an A100 80GB through Badgr with vLLM 0.30.0 and --tool-call-parser gemma4, to send 30 requests with the PI agent's edit tool schema. The deployment failed on provider capacity (provider_capacity_failed) before the server became healthy.
Not run to a result: the A100 route never became healthy. A rerun on any 80 GB GPU with the same 30-request script would settle it.
Checked 2026-10-15 in Badgr’s local development environment.
← All issue evidence