Environment: commit 2c1e326; macOS 15 (Darwin 25.3.0) / Python 3.13.7 /
Docker 29.1.5; no Kubernetes cluster, no GPU; helm, helmfile, yq, kind, podman
absent. Reproduced in a venv with pip install -e . plus
llm-d-planner v0.1.0.
Repro (the scout used a locally built image whose vllm entrypoint never
serves, to isolate this from lack of a GPU; the real-world trigger is the one the
docs name, "vLLM container exits at load"):
llmdbenchmark --spec <nok8s-spec> --base-dir . standup --nok8s-deploy-timeout 15
Expected: the failing container's logs are captured, since that is the only
place the cause (bad HF token, model too large for VRAM, wrong accelerator
image) is visible.
Actual: the error names the vLLM port, but the only artifact written is
nok8s-envoy.log, containing an unrelated Envoy startup warning:
[06] Stack 'nok8s-single': FAILED: [06] nok8s_deploy (nok8s-single): FAILED - Timed out waiting for http://localhost:8000/v1/models after 15s
$ cat ws9/.../setup/logs/nok8s-envoy.log
[2026-07-30 09:59:20.059][1][warning][main] [source/server/server.cc:959] There is no configured limit to the number of allowed active downstream connections. Configure a limit in `envoy.resource_monitors.global_downstream_max_connections` resource monitor.
No nok8s-vllm-0.log is produced.
Cause: step_06_nok8s_deploy.py:99-104
ready_err = self._wait_ready(cmd, runtime, spec, context)
if ready_err:
self._dump_logs(cmd, runtime, "envoy", context)
return self._fail(stack_path, ready_err, [ready_err])"envoy" is hard-coded, whereas _wait_ready (:247-269) already knows which
port timed out. docs/nok8s.md:178-179 compensates by telling users to run
docker logs vllm-0 by hand. _dump_logs already exists and is used correctly
for launch failures at :92.
Suggested fix: have _wait_ready return the failing container name and dump
that container's logs, or dump all of them on any readiness failure.
Repro (the scout used a locally built image whose
vllmentrypoint neverserves, to isolate this from lack of a GPU; the real-world trigger is the one the
docs name, "vLLM container exits at load"):
Expected: the failing container's logs are captured, since that is the only
place the cause (bad HF token, model too large for VRAM, wrong accelerator
image) is visible.
Actual: the error names the vLLM port, but the only artifact written is
nok8s-envoy.log, containing an unrelated Envoy startup warning:No
nok8s-vllm-0.logis produced.Cause:
step_06_nok8s_deploy.py:99-104"envoy"is hard-coded, whereas_wait_ready(:247-269) already knows whichport timed out.
docs/nok8s.md:178-179compensates by telling users to rundocker logs vllm-0by hand._dump_logsalready exists and is used correctlyfor launch failures at
:92.Suggested fix: have
_wait_readyreturn the failing container name and dumpthat container's logs, or dump all of them on any readiness failure.