- Dockerfile 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
|
||
| .github/workflows | ||
| Dockerfile | ||
| LICENSE | ||
| README.md | ||
| renovate.json | ||
vLLM Patched Image
This repository provides a custom build of vLLM that includes a patched version of the transformers library installed directly from source.
Intent
The primary purpose is to maintain a Docker image compatible with vllm/vllm-openai but with the latest transformers features or fixes from the huggingface/transformers main branch. This is useful when the served model requires a specific transformer version that hasn't been officially released or integrated into the upstream vLLM images yet. An example of such a model is GLM-OCR.
How it works
The Dockerfile uses an ARG BASE_TAG to pull a specific version of vllm/vllm-openai and then updates the transformers package using uv.
CI/CD Workflow
The GitHub Actions workflow Build and Publish Patched vLLM can be triggered manually via workflow_dispatch.
Inputs
vllm_tag: The upstream vLLM image tag to use as a base (e.g.,v0.7.2-cu124). Defaults tolatest.
Outputs
The workflow builds and pushes the image to GitHub Container Registry (GHCR):
ghcr.io/DCC-BS/vllm:<vllm_tag>ghcr.io/DCC-BS/vllm:latest
Usage
You can pull the latest patched image:
docker pull ghcr.io/DCC-BS/vllm:latest
Or a specific version:
docker pull ghcr.io/DCC-BS/vllm:v0.16.0-cu130