ASUS Ascent GX10 / DGX Spark: Nemotron 3 Super NIM fails with NGC entitlement / org context error, not a local hardware issue

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

I’m posting this here because NVIDIA customer support bounced me around and sent me to a generic Linux graphics forum, but after a lot of troubleshooting (2 Days) this does not appear to be a local graphics or hardware problem.

System:

  • ASUS Ascent GX10

  • NVIDIA GB10 / Spark-class system

  • Driver: 580.126.09

  • nvidia-smi works on host

  • Docker GPU access works

  • CUDA toolkit installed: 13.2

  • nvcc --version reports 13.2

  • nvidia-smi reports CUDA 13.0

What I was trying to run:

  • nvcr.io/nim/nvidia/nemotron-3-super-120b-a12b:latest

What I verified successfully:

  • Host GPU is visible with nvidia-smi

  • Docker GPU passthrough works with:

docker run --rm --gpus all ubuntu nvidia-smi

  • Running headless with display-manager stopped does not change the core failure

  • This does not appear to be low RAM, disk-full, or a basic Docker/CUDA issue

Important note:

  • Docker on this system does not use --runtime=nvidia

  • --gpus all works correctly

The actual failure:

The NIM container starts, reads cached files, then fails when requesting the manifest/file map from NGC for:

nemotron-3-super-120b-a12b:rl-030326-fp8

Container error:

ManifestDownloadError: Error downloading manifest: HTTP status client error (400 Bad Request) for url:
https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/nemotron-3-super-120b-a12b/rl-030326-fp8/files

I also tested the NGC API directly with my API key.

Results:

  1. Model endpoint:
https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/nemotron-3-super-120b-a12b

returns:

403 Forbidden
Access Denied
  1. File manifest endpoint:
https://api.ngc.nvidia.com/v2/org/nim/team/nvidia/models/nemotron-3-super-120b-a12b/rl-030326-fp8/files

returns:

400 Invalid Request
Validation failed - [Org contex missing in uam request for <GET ...>, API Key does not have correct permissions (or subscriptions) to perform <GET ...>]

What seems to be going on:

This looks like an NGC entitlement / org context / UAM mapping issue, not a local GX10 problem.

My API key is valid and loaded, but I am signed up as an individual account using my business email/business name. I suspect that may be part of the issue if this model requires access through a properly entitled NGC org/team rather than just an individual login.

Questions:

  1. Has anyone else on GX10 / DGX Spark hit this exact Org context missing / permissions error with Nemotron 3 Super NIM?

  2. Does this model require a specific NGC org entitlement beyond just generating a Personal API key?

  3. Is there a known issue with the nemotron-3-super-120b-a12b:rl-030326-fp8 profile or its NGC manifest path?

  4. Has anyone successfully run this exact NIM image on GX10 / DGX Spark, and if so, what account/org setup was required?

At this point, I’ve ruled out:

  • GPU not detected

  • Docker GPU runtime failure

  • display-manager / GUI crash as the root cause

  • simple RAM exhaustion

  • simple disk exhaustion

Any help from someone who has actually gotten this NIM running on GX10 / DGX Spark would be appreciated. Just spent 8K on Nvidia Grahics cards and computers and Installing Nemotron 3 Super on a G10 Spark should be way less difficult. Thanks in advance for your time.

2

nvidia-bug-report.log.gz (397.5 KB)

3

ANYONE from NVIDIA???

4

Hi jeremy153,

Sorry to see you’re running into this issue. Would you be able to generate a new API key and try one more time? Please let me know what you see after that.

5

Thanks. I can generate a new Personal API key and retry. Before I do, can you confirm which scopes/options the key must include for nemotron-3-super-120b-a12b, and whether access requires membership in a specifically entitled NGC org/team rather than an individual account? My current key is valid, but direct API tests return 403 Access Denied and Org context missing for this model.

6

Hi jeremy153,

After doing more investigation, I can report to you that the Nemotron 3 Super NIM is not yet supported on the DGX Spark.

You still have options to run this model, one such being vLLM. Take a look at the community build: GitHub - eugr/spark-vllm-docker: Docker configuration for running VLLM on dual DGX Sparks · GitHub

Please give this a try and don’t hesitate to reach out here if anything comes up.

7

Running it on the Spark on Ollama now… Any idea when it will be officially supported for DIM?

8

Has something changed? A few days ago I was able to pull/start it as I posted here, but it crashed after loading the weights. This seems like a different failure to what’s reported above.

9

I unfortunately can’t speak to a concrete time it will be supported. Along those lines, I don’t know if anything specifically changed, just that it is not supported on Spark yet.