Hugging Face Train on DGX Cloud

Hugging Face and NVIDIA have launched Train on DGX Cloud, a service integrated into the Hugging Face Hub that allows Enterprise Hub organizations to fine-tune open Generative AI models using NVIDIA's accelerated compute infrastructure. This service addresses GPU scarcity and the complexity of writing training scripts by providing a no-code interface for accessing high-performance GPUs.

No-Code Model Fine-Tuning

Train on DGX Cloud enables users to fine-tune popular open models with a few clicks. The service is powered by Hugging Face AutoTrain and Hugging Face Spaces, removing the need for manual script writing, testing, and debugging.

Supported model architectures include:

  • Llama
  • Falcon
  • Mistral
  • Mixtral
  • T5
  • Gemma
  • Stable Diffusion
  • Stable Diffusion XL

Technical Workflow and Hardware Options

Users with a Hugging Face Enterprise subscription can initiate training jobs directly from the model page of supported architectures. The process involves selecting an Enterprise Organization and creating an AutoTrain Space to configure the job.

Hardware Configurations

Training jobs can be executed on the following GPU instances:

  • NVIDIA H100 Tensor Core GPUs: Available in 1x, 2x, 4x, and 8x instances.
  • NVIDIA L40S GPUs: Also supported for training tasks.

Training Process

  1. Configuration: Users select the hardware, base model, task, and training parameters (editable via JSON configuration).
  2. Data Upload: Training datasets must be uploaded as CSV or JSON files.
  3. Execution: AutoTrain validates the dataset and starts the training job.
  4. Output: Once complete, the fine-tuned model is uploaded to a private repository within the user's selected namespace on the Hugging Face Hub.

Pricing Model

Train on DGX Cloud operates on a pay-as-you-go basis, billed by the minute of GPU instance usage.

  • NVIDIA H100 instances: $8.25 per GPU hour.
  • NVIDIA L40S instances: $2.75 per GPU hour.

As an example, fine-tuning Mistral 7B on 1,500 samples using a single NVIDIA L40S GPU takes approximately 10 minutes and costs roughly $0.45.

Broader NVIDIA and Hugging Face Collaboration

Train on DGX Cloud is part of a larger strategic partnership between Hugging Face and NVIDIA to democratize accelerated machine learning. Other collaborative efforts include:

  • Open Science: The training of StarCoder 2 15B, a code LLM trained on over 600 languages via BigCode.
  • Open Source: The development of the optimum-nvidia library, which accelerates LLM inference on NVIDIA GPUs, achieving up to 1,200 tokens per second with Llama 2.
  • Inference Optimization: Ongoing work leveraging NVIDIA TensorRT-LLM and the integration of popular open models into NVIDIA NIM microservices.

Note: This service was deprecated and became unavailable as of April 10th, 2025.

Sources