Pular para o conteúdo
Kitbay
Skill

spark-environment-setup

MITAtualizado em 18 de nov. de 202642.668Estrelas no GitHub

A skill from the dgx-spark-ops plugin in Seth Hobson's agents collection for setting up a machine learning environment on an NVIDIA DGX Spark, a desktop system with an Arm CPU, a GB10 GPU, 128 GB of unified memory and CUDA 13.

O que a skill ensina

A skill from the dgx-spark-ops plugin in Seth Hobson's agents collection for setting up a machine learning environment on an NVIDIA DGX Spark, a desktop system with an Arm CPU, a GB10 GPU, 128 GB of unified memory and CUDA 13. It recommends starting from an NVIDIA NGC PyTorch container, or an Unsloth container for fine-tuning, and using plain pip only with pinned, dated versions. It explains that most failures come from mixing CUDA 12 and 13 builds, gives an order for checking why the GPU is not visible, a workaround for Triton kernel builds, and the state of PyTorch, vLLM, xformers, flash-attn, TRL, PEFT and others on this hardware. Two reference files hold the version matrix.

O que faz

  • Choose between NGC containers and pip installs
  • Fix CUDA 12 and 13 library mismatches
  • Check that the GPU is visible to PyTorch
  • Pin container digests for repeatable runs