Skip to content

DARC Blog

This blog is a place to learn how to solve Research Computing tasks at Stanford GSB.

description

  • in LLM
  • 9 min read

Self-Hosting LLMs with NVIDIA NIM on the Yens

NVIDIA NIM (NVIDIA Inference Microservices) lets you deploy optimized, production-ready large language models on your own infrastructure. Instead of sending data to a third-party API, you pull a pre-packaged container, start it on a GPU node, and query it through a standard OpenAI-compatible REST endpoint — all within Stanford's network.

This guide walks through deploying Google's Gemma 4 31B IT model on a single H200 GPU on the Yen cluster using Singularity containers.

H200 GPUs Required

NIM containers require H200 GPUs and do not work on older GPU architectures such as the A30 or A40. On the Yen cluster, the only node with H200 GPUs is yen-gpu4. All examples in this guide use yen-gpu4 as the target node.

  • in Yen
  • 2 min read

SSH Setup for the Yen Servers

When working on the Yen cluster, you may occasionally open multiple SSH sessions from your local computer or connect from one Yen node to another to check a running process or run a command on a different node.

You can make these workflows smoother by configuring SSH multiplexing on your local computer and setting up SSH key authentication between Yen nodes.

  • in LLM
  • 12 min read

Fine-Tuning Open Source Models

When large language models (LLMs) first appeared, they felt almost magical — you could ask them anything and they’d reply with surprisingly fluent text. But once you start applying them in research or production, the limitations show up quickly. The base model can sort of do your task, but not reliably enough. That’s where fine-tuning comes in.

  • in LLM
  • 6 min read

Running Ollama on Stanford Computing Clusters

Imagine running a notebook cell like this — and getting a full response from a large language model hosted on your own cluster:

LLM running on Yen Jupyter

With Ollama, you can host models like Llama 3 or DeepSeek on Stanford’s GPU clusters — no API keys, no external calls — and interact with them through your own code or notebooks.

This guide walks you through setting up Ollama across Stanford's GPU computing clusters — Yen, Sherlock, and Marlowe — to efficiently run large language models (LLMs).

Introduction to Using Redivis

Redivis is a powerful data querying and analysis platform built specifically with researchers in mind. It is currently the GSB's solution for hosting Big Data (data on the scale of TBs) for researchers at the school. At the time of this post, the StanfordGSBLibrary Redivis organization hosts more than 50 datasets consisting of over 100 TB of data with 300 organization members.

If you are a researcher who is just starting to use Redivis, or considering using the platform, this blog post will help you get started by covering some common use cases and helpful tips.

Train Machine Learning Models on Colab GPU

Google Colab

Google Colab enables you to run Jupyter notebooks in the cloud with the option to use a CPU or accelerate computations by adding GPU or TPU support. We will use the free Colab tier, but for longer training jobs or access to better GPUs (e.g., T4, P100, or V100), the paid Colab Pro or Colab Pro+ option may be a better choice. Navigate to Colab website and check out an example Jupyter notebook that uses a GPU for machine learning training.

Reproducible Research Essentials

This guide provides the foundational components needed to ensure reproducibility in your research. It focuses on:

  • Documentating fixed inputs and expected outputs
  • Making a README file
  • Managing computational environments
  • Summary with additional resources
  • Advanced topics