DGX Spark taking lot of time to serve llm using vlllm

Hi Comm ,
I am new with DGX Spark and I am trying vLLM.
I tried loading gpt-oss 20b through latest vllm image , it was success but it took more than hour and half to load.
Now same problem with Qwen/Qwen3-Coder-30B-A3B-Instruct, It’s been 2 hrs and it still is loading.

This seems frustrating as beginner, i am caught up in the loop of trying models and then waiting hours for setup.
max model length 8192 or higher, everything takes more time
Am I doing something wrong?

Definitely something is wrong,. Use @eugr_nv 's repo, clone it from here GitHub - eugr/spark-vllm-docker: Docker configuration for running VLLM on dual DGX Sparks · GitHub build the image and use his recipes. You should be up and running in a few minutes.

If you don’t have the model already downloaded, then it will download it during startup. This can take a long time depending on your internet speed. You can also set the huggingface token so hf doesn’t throttle you as much.

another beginner here ;)
I can second azampatti’s recommendation for eugr/spark-vllm-docker.

I assume you use docker? You can also use “docker logs -f your_container_name” to check the logs after starting the container. If you can’t remember the name use “docker stats” . If it is a download, or any other issue it should be in the logs.

model download is fine but it takes time after that. That really tests my patience.

docker run -d

–name qwen3-coder
–gpus all
–ipc host
–net host
–ulimit memlock=-1
–ulimit stack=67108864
-e HF_TOKEN=“hf_xxx”
-e HF_HOME=/root/.cache/huggingface
-v ~/.cache/huggingface:/root/.cache/huggingface
vllm/vllm-openai:latest
Qwen/Qwen3-Coder-30B-A3B-Instruct
–gpu-memory-utilization 0.80
–max-model-len 8192
–enable-auto-tool-choice
–tool-call-parser qwen3_coder

but what is wrong with this?

That, the vllm image you’re using is far from the optimal for these boxes. Go Eugr’s! :)

Also, Qwen3-Coder isn’t the best for coding anymore (in most situations), try Qwen3.6-35B-A3B for a fast one or Qwen3.8-27B for a slower but higher accuracy one :)

I agree. Go with eugr’s stack. One command and you’ll have an almost perfectly optimized container running one of the newer models.