H100 HBM3 serverless/container CUDA init failure (error 802)
We’re seeing a host/runtime issue on H100 HBM3.
Same worker image works on:
- H100 NVL
- RTX PRO 6000 Blackwell Server Edition
But fails on:
- H100 HBM3
Inside the container:
- nvidia-smi sees the GPU
- Driver Version: 570.211.01
- CUDA Version: 12.8
- torch: 2.10.0+cu128
But torch.cuda.is_available() is False, and ComfyUI startup fails with:
RuntimeError: Unexpected error from cudaGetDeviceCount()
Error 802: system not yet initialized
This looks like a host/runtime configuration issue on the H100 HBM3 pool rather than an image issue, since the same image works on H100 NVL.
Can you check the H100 HBM3 pool configuration?
Same worker image works on:
- H100 NVL
- RTX PRO 6000 Blackwell Server Edition
But fails on:
- H100 HBM3
Inside the container:
- nvidia-smi sees the GPU
- Driver Version: 570.211.01
- CUDA Version: 12.8
- torch: 2.10.0+cu128
But torch.cuda.is_available() is False, and ComfyUI startup fails with:
RuntimeError: Unexpected error from cudaGetDeviceCount()
Error 802: system not yet initialized
This looks like a host/runtime configuration issue on the H100 HBM3 pool rather than an image issue, since the same image works on H100 NVL.
Can you check the H100 HBM3 pool configuration?
