메뉴
Container debugging runbook for this inference stack — per-service healthcheck endpoints, common CUDA/driver failure signatures, log locations, and the safe restart procedure. Use when a service is unhealthy or failing to start.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | docker-debug |
| description | Container debugging runbook for this inference stack — per-service healthcheck endpoints, common CUDA/driver failure signatures, log locations, and the safe restart procedure. Use when a service is unhealthy or failing to start. |
For the llama.cpp / vLLM / open-webui stack defined in docker-compose.yaml.
Sandboxed execution: agents cannot reach the host Docker daemon, the services, or the GPU. Every diagnostic and restart command in this runbook is executed by the human on the workstation. The agent supplies the exact command, waits for the pasted output or logs, and matches the failure signatures below against what the human reports.
| Service | External port | Internal healthcheck |
|---|---|---|
| llama-cpu | 8001 | curl -f http://localhost:8001/v1/health |
| llama-gpu | 8004 | curl -f http://localhost:8002/v1/health (internal 8002) |
| vllm-gpu | 8005 | curl -f http://localhost:8003/health (internal 8003) |
| open-webui | 3000 | container :main, no explicit healthcheck |
| comfyui | 8188 | curl -f http://localhost:8188/system_stats |
Start with make status for the compose-level health summary and make health for the
detailed per-service check.
CUDA error: no kernel image is available for execution — image built for the wrong
compute capability; the RTX 5090 is Blackwell (sm_120). Rebuild the GPU image.CUDA driver version is insufficient for CUDA runtime version — host NVIDIA driver older
than the container's CUDA toolkit; update the host driver.could not select device driver "nvidia" — NVIDIA Container Toolkit not wired into
Docker; check nvidia-smi on the host and the runtime config.--gpu-memory-utilization, or quantization.make logs, or per-service make logs-gpu / make logs-cpu /
make logs-vllm / make logs-ui.docker logs <container> (llama-cpu, llama-gpu, vllm-gpu, open-webui).*.log files are git-ignored.make status); do not restart the whole stack for a
single failure.docker compose restart <service>.make down then make up (or make rebuild for a
from-scratch image build). make reset is the full stop-clean-rebuild cycle.docker system prune / docker volume rm to "fix" a hang — those are guarded
ask actions and destroy state; diagnose from logs first.Mechanics contract for running a multi-agent workflow that survives to completion - the durable-ledger rule, the fan-in partition rule, the output-budget rule, and the recovery procedure for a run that dies mid-flight. Read before authoring or resuming any Workflow script.
The handshake for verifying anything that needs the workstation — how to state acceptance criteria before running, hand the human an exact command block, and record the reported output as evidence. Use whenever a claim depends on the GPU, Docker, the running services, /mnt/ai-data, or a full make quality.
The method contract for a quality campaign — the phase ladder, the standing dimension list, the severity rubric, what counts as a confirmed finding, and the batch-and-push cadence. Read before running or participating in a bug bash or quality gate.
How to launch, resume, and serve model training runs on this stack — checkpoint layout, the memory-fit calculator, and the train-and-serve round trip. Skeleton reference; grows once the training tier lands.
The reproducible benchmark protocol for this workstation — which entry points to run, warm-up policy, how results are recorded and compared, and what counts as a regression. Use before reporting any latency/throughput/memory number.