菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| 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.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.
RunPod endpoint operations runbook — checking worker state, forcing scale-down, reading the spend query, and cold-start expectations. Use before and after cloud inference sessions. Forward-looking — RunPod tooling lives outside this repo.
基于 SOC 职业分类
