k80-build

name	k80-build
description	Build vLLM for Tesla K80 (sm_37) using the Docker builder/runtime pattern. Use when you need to produce a `vllm37-local` image, or are considering a host build (don't).
argument-hint	["local|github|builder"]

K80 Build

Build vLLM for Tesla K80 via Docker. Never build on the host — CUDA 11.4 + GCC 10 + PyTorch 2.0.1 + sm_37 combination is pinned inside the builder image.

When to use

• Producing a runtime image to test code changes
• Rebuilding after touching csrc/, setup.py, or requirements/cuda_k80.txt
• Building on a clean machine (pull dogkeeper886/vllm37-builder first)

Environment

• CUDA 11.4.4, GCC 10.5, CMake 4, Python 3.10, PyTorch 2.0.1
• Target: Tesla K80 (compute capability 3.7)
• Host toolchain: ignored — everything runs in Docker

Build types

• builder — make build-builder (CUDA + GCC + PyTorch 2.0.1 base, ~120 min first time)
• local — make build-local (compile current checkout against the builder, ~10 min)
• github — make build-runtime (clone $VLLM_REPO@$VLLM_BRANCH in-container)

Commands

cd docker/k80

# One-time: get the builder (prefer pull over rebuild)
docker pull dogkeeper886/vllm37-builder:latest
docker tag dogkeeper886/vllm37-builder:latest vllm37-builder:latest
# OR: make build-builder  (~120 min)

# Iterate:
make build-local          # ~10 min incremental
make build-local JOBS=8   # override parallelism via .env

Key build flags (already set in Dockerfiles)

• TORCH_CUDA_ARCH_LIST="3.7" — K80 compute capability
• VLLM_BUILD_LEGACY_CUDA=1 — skip CUTLASS kernels (require sm_70+)
• --no-deps on pip install . — prevents PyTorch 2.7 from overwriting 2.0.1

Outputs

• vllm37-builder:latest — base image with toolchain
• vllm37-local:latest — runtime image (from local source)
• vllm37:latest — runtime image (from GitHub)

Related

• /ci — trigger the build remotely on the self-hosted runner
• docker/k80/README.md — architecture and limitations
• docker/k80/Makefile — all build targets