Docker Image: NVIDIA vLLM 0.23.0 with Claude Code 2.1.195+ Compatibility

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
,
1

Hi everyone!

I wanted to share a project I’ve been working on for anyone using NVIDIA DGX Spark (GB10) with vLLM and Claude Code.

I published a Docker image that updates NVIDIA’s official 26.05.post1-py3 container to upstream vLLM 0.23.0 while preserving NVIDIA’s optimized CUDA, PyTorch and GB10-specific libraries.

What’s included

  • ✅ Upstream vLLM 0.23.0

  • ✅ NVIDIA CUDA/PyTorch/NCCL preserved

  • ✅ GB10 (sm_120) optimized build

  • ✅ OpenAI-compatible API

  • ✅ Minimal changes to the original NVIDIA image

Claude Code compatibility

One of the main motivations for this project was recent Claude Code releases.

Newer versions (for example, 2.1.195+) send Anthropic Messages API requests that aren’t accepted by current vLLM, causing requests to fail.

This image includes a small compatibility layer that normalizes those requests before they reach vLLM, allowing both newer and older Claude Code versions to work without modifying the client or downgrading Claude Code.

The goal was to keep the patch as small as possible while staying compatible with the upstream NVIDIA image.

Repository:

I’d really appreciate any feedback from other DGX Spark users. If you’ve tested different models or workloads on GB10, I’d love to hear about your experience or any suggestions for improving the image.

2

I am new and I like your idea. The only thing I wonder is that the Claude code integration would simply be perfect with a second vs coder (Devin formerly known as Windsurf) extension (that Claude code (may be referred to as Claw code) to run next to the current up-to-date Claude code (with Anthropic inside) however as a second extension (maybe based on the Claw code). I’d like to get 1 or two Sparks to run qwen3.x 122 or Nemo at the best available size for the spark and best available quality and speed for agents and Devin (formerly known as Windsurf) a MS vs coder clone.

Do you have recommendations for this constellation I’d like to use all on my MacBook, (and my colleagues as well (3-4 users) want to use it for their laptops and Devin’s…) does your project make sense for this type of project and is the Claude code applicable to everybody’s Mac, or does this only run on the Spark…

3

Why not fix or add to the Anthropic messages endpoints in vllm? While there’s a lot going on in vllm, the API endpoints are pretty straight forward.

4

Can this be integrated into @eugr_nv 's community repo? I had to pin my CC version to an older one for a while now (.153) and would love to have a fix in our community one :)

thanks!

5

I just tried Claude Code 2.1.195 with the latest spark-vllm-docker build and it worked just fine, so whatever was broken seems to be fixed already…

6

I’ll try again tonight, I’ve updated to 0..23 over the weekend and haven’t tried again.

7

Ok, awesome, I had a 4 min job done by Claude Code with vLLM 0.23.1rc1dev309 (community docker updated to a few days back) and worker perfectly.

Thanks @eugr_nv :)