running kev on dgx-spark

#2
by pawlz - opened

Hi,

has anyone successfully run Kev-27B on NVIDIA DGX Spark / GB10?

I’m looking for a known-working Docker image or reproducible setup for Spark (ARM64, CUDA 13, unified memory).
I’ve tried running the official Kev server with the full BF16 checkpoint, but after:
Loading weights: 100%
the process appears to hang indefinitely. GPU memory is allocated, but GPU utilization stays at 0%.
I also tried a few custom Docker setups with pinned PyTorch/Transformers versions and different loading strategies, but I haven’t found a stable configuration yet.

Does anyone have:

  • a working Dockerfile/image for DGX Spark,
  • exact versions of PyTorch / Transformers / safetensors / CUDA that work,
  • or confirmation that Kev-27B is currently supported on GB10?
    I’d be happy to test a reference setup and report back.

Hardware: NVIDIA DGX Spark, GB10, 128 GB unified memory

Thanks.
Pawlz

Sign up or log in to comment