Trouble to find CUDA at DGX

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

Hi Community,

I have getting trouble to setup my new DGX machine and would like some help here.

For both uv and conda virtual env, I am getting the error for either CUDA_HOME is not found or “ValueError: Unsupported platform, please use CUDA, ROCm, Neuron, OpenVINO, or CPU.”. However, both of my nvidia driver is active and nvcc is available, and the variable is set:
nvidia-smi
Thu Feb 26 14:42:04 2026
±----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 |
±----------------------------------------±-----------------------±---------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GB10 On | 0000000F:01:00.0 On | N/A |
| N/A 40C P8 5W / N/A | Not Supported | 0% Default |
| | | N/A |
±----------------------------------------±-----------------------±---------------------

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:57:39_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88
Build cuda_13.0.r13.0/compiler.36424714_0

echo $CUDA_HOME
/usr/local/cuda

wondering is there anything wrongly installed, or anywhere I can investigate?

Thanks!

2

Hi @sw3237 — I hit exactly the same issue on my DGX Spark. The root problem is that the GB10 runs CUDA 13.0, and no standard pip release of vLLM (or most other ML frameworks) supports CUDA 13.0 yet. So any attempt to install vLLM into a conda or uv environment will fail with that platform error regardless of whether CUDA_HOME is set correctly — it’s not an environment variable problem, the wheels simply don’t exist for this CUDA version.

The solution is to use @eugr’s spark-vllm-docker, which is a pre-built Docker image specifically compiled for the GB10 and CUDA 13.0:

It handles all the CUDA 13.0 compatibility, unified memory architecture, and vLLM configuration for you. You run the container and serve models from inside it rather than from a host Python environment.

Setup is roughly:

  1. Clone the repo on your DGX host
  2. Follow the build instructions (takes ~35 min first time)
  3. Start the container
  4. Download your model inside the container and serve with vLLM

Your nvidia-smi and nvcc output look completely normal for a healthy DGX Spark — nothing is wrongly installed, you just can’t use the host Python environment for this yet. Docker is the way to go for now.

Hope that helps.

4

Thank you so much for the info! this is super helpful!