build-and-copy.sh: --vllm-source-dir no longer disables B12x

[nicjansma]

Fixes #363: A recent source checkout of vLLM 0.28.0+ (or https://github.com/local-inference-lab/vllm) won't complile all of the necessary B12X components for e.g. DS4F.

After investigating, vLLM 0.28.x doesn't (yet) have all of the local-inference-lab changes from @lukealonso merged in for B12X, i.e. vllm-project/vllm#52017, so the dual Spark DS4F recipe doesn't work.

local-inference-lab is very close to 0.28.x, including some fixes I wanted for DS4F stability, so I switched to building off that. However, that tree was failing builds as well since there were some changes with upstream b12x packages.

Changes

build-and-copy.sh changes:

• --vllm-source-dir no longer disables B12x
• Adds --with-b12x option to force B12x

Fixes

prepare_local_vllm_source() sets VLLM_REPO="local-source", which the B12X eligibility gate compares against two literal URLs - so a local clone of upstream vLLM never got the B12X kernel package, and --linear-backend b12x failed at runtime with Failed to find a kernel that can implement the ScaledMM linear layer.

The metadata recorded only b12x_repo: "disabled".

The gate now falls back to the checkout's origin remote, compared via a new normalize_git_url() that matches https, ssh, and git@ spellings. --with-b12x forces the package in for mirrors, forks, and checkouts without an origin. An unrecognised checkout now warns explicitly instead of failing silently at serve time.

Results

This build command now works:

cd /opt

git clone https://github.com/eugr/spark-vllm-docker.git
git clone https://github.com/local-inference-lab/vllm.git

cd spark-vllm-docker

./build-and-copy.sh --rebuild-vllm --tag vllm-node:latest-b12x --vllm-source-dir /opt/vllm/

AI Disclosure

This PR was investigated, codified and verified by an AI agent. I have manually reviewed, verified and tweaked the output.