Hi NVIDIA Cosmos team,
I’m unable to deploy the Cosmos Transfer 2.5 2B NIM container. The container starts successfully but crashes when downloading the guardrail model files.
Error:
ERROR: Download failed after 1 attempts. Last exception: HTTP status client error (400 Bad Request) for url (cosmos-transfer2.5-2b/1.0.0-guardrail/files)
nimlib.exceptions.ManifestDownloadError: Error downloading manifest: HTTP status client error (400 Bad Request)
Environment:
-
GPU: NVIDIA L4 (compute capability 8.9, 23GB VRAM)
-
Host: Google Cloud g2-standard-16, Ubuntu 22.04
-
Driver: 570.211.01, CUDA: 12.8
-
nimlib: 0.13.4, nim_sdk: 0.9.12
What works:
-
docker login to nvcr.io succeeds
-
Container image pull succeeds (30.44 GB)
-
Compute capability check passes (8.9)
-
Profile selected: 9c21d6c8c3d6ca18c6a0769c04db7bb5a96d36dcd2628379fcdf8927aa9a9027
What fails:
-
Downloading manifest profile 1.0.0-guardrail returns 400 Bad Request
-
Tried two different freshly generated NGC API keys
-
NGC account has Developer Program and AI Enterprise Essentials access
The 1.0.0-guardrail artifact appears to not be accessible. Can you verify it is published correctly?
Thank you,
Alan Healey-Greene
alan at chionyx dot com