DGX-Spark-Finetune-LLM

1 comments

• waybarrios 5 hours ago
  I built a toolkit to fine-tune LLMs using LoRA + native 4-bit quantization on NVIDIA's new Blackwell GPUs (DGX Spark with GB10).

    Key features:
    - NVFP4 (4-bit) via Transformer Engine - fastest option
    - MXFP8 (8-bit) for higher precision
    - bitsandbytes FP4 fallback for any CUDA GPU
    - ~240MB LoRA adapters instead of ~6GB full models
  
    Tested on DGX Spark (128GB VRAM). Training SmolLM3-3B takes ~70GB VRAM with NVFP4.