Docker-vLLM (GPT OSS)-VSC: struggling to figure out how to get tool use functioning

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

I see vids/GIFs of GPT OSS being agentic, using the CLI, reading/writing files, etc.

So in VS Codium, I tried the Cline extension. I also tried the Void IDE, a VSC fork. Neither seem to be able to decipher GPT OSS’s tool use output attempts for some reason.

This is the command I’m using:

docker run -it --gpus all --ipc=host --platform "linux/arm64" \
--ulimit memlock=-1 --ulimit stack=67108864 -p 8000:8000 \
nvcr.io/nvidia/vllm:25.12.post1-py3 vllm serve "openai/gpt-oss-120b" \
--gpu-memory-utilization=0.8 --enable-auto-tool-choice \
--tool-call-parser openai

I tried without the two *tool* arguments, as well as ulimit arguments.

The speed and chatting seem to work perfectly fine, but the error message is about Harmony format (which GPT OSS uses) for tool use, and unsure if anyone has any open source solution for this, whether it be a VSC extension or a fork, or something else.

2

While NVIDIA docker should work with gpt-oss just fine, try our community build instead - it’s up to date unlike NVIDIA one: GitHub - eugr/spark-vllm-docker: Docker configuration for running VLLM on dual DGX Sparks

It include tiktoken encodings that are necessary to run gpt-oss out of the box.

For gpt-oss specifically you want to run with the following parameters - works with Cline without any issues:

vllm serve openai/gpt-oss-120b \
        --host 0.0.0.0 \
        --port 8000 \
        --enable-auto-tool-choice \
        --tool-call-parser=openai \
        --reasoning-parser=openai_gptoss \
        --gpu-memory-utilization 0.70 \
        --enable-prefix-caching \
        --load-format fastsafetensors