ci

name	ci
description	Run K80-specific CI workflows for vLLM on the self-hosted runner. Use when triggering a remote build or smoke test, or deciding between local vs CI.

CI

Trigger build and smoke-test workflows via GitHub Actions on the self-hosted runner (labels self-hosted, vllm) that has K80 GPU access.

When to use

• You don't have K80 hardware on the current machine
• Validating a branch/PR before merging
• Reproducing a build in a clean container environment

Workflows

File	Purpose
.github/workflows/k80-build.yml	Docker build of vllm37-local image
.github/workflows/k80-runtime.yml	Start container, poll /v1/models, smoke-test /v1/completions
.github/workflows/k80-pipeline.yml	k80-build then k80-runtime

All run on runs-on: [self-hosted, vllm].

Trigger

GitHub UI → Actions tab → select workflow → Run workflow.

CLI:

gh workflow run k80-pipeline.yml -f tp_size=1 -f no_cache=false
gh workflow run k80-build.yml -f no_cache=true
gh workflow run k80-runtime.yml -f tp_size=1 -f model=TinyLlama/TinyLlama-1.1B-Chat-v1.0

Safety defaults (do not change casually)

• tp_size: 1 — TP=4 previously halted the system (power-related). TP=2 is known-working but only with verified PSU setup. TP=4 is unvalidated.
• NCCL_P2P_DISABLE=1 — required for any TP>1 on K80 (PCIe only). Baked into docker/k80/docker-compose.yml.
• Containers are torn down with if: always(); logs uploaded as artifact k80-runtime-logs for post-mortem.

Local vs CI

• K80 present → cd docker/k80 && make build-local && make run is fastest
• K80 absent → trigger the CI pipeline
• Builder image rebuild (~120 min) is only triggered via the rebuild_builder: true input on k80-build.yml; otherwise it pulls dogkeeper886/vllm37-builder:latest on first CI run.

Related

• /k80-build — build locally when you do have K80 hardware
• docker/k80/README.md — runtime configuration (model, TP size, etc.)