Day 4 — Parallelization & Local LLMs
Day 4 scales yesterday’s single Slurm job into a full research pipeline. You’ll fan one script across many filings at once with job arrays, then meet local LLMs — models you run on the cluster yourself rather than calling over the internet — and see what it takes to serve one. From there, the ways LLMs fail, and what to do about them. The Day 4 Challenge closes the day, and the course, by pushing the array from a hundred filings to all of them.
Duration: ~3 hours
Day 3 Recap
- Profiled the extraction script to measure its real time and memory needs
- Wrote and submitted
slurm/extract_form_3_batch.slurmas a batch job, with#SBATCHdirectives grounded in those measurements - Monitored the job with
squeueandsacct, and documented the pipeline inREADME.md
Any questions about Day 1–3 before we move on?
Sections
Work through the sections in order — later ones build on earlier ones, and the Day 4 Challenge draws on everything you’ve learned.
| Section | Format | What you’ll learn |
|---|---|---|
| Parallelization Basics | 🖊️ Concept | What it means to run work in parallel, when it helps, and the three ways to split a job — across cores, across jobs, or both |
| Slurm Job Arrays | 🖊️💻 Concept + Hands-on | One script, many tasks: --array, SLURM_ARRAY_TASK_ID, and making a task safe to rerun so a partial failure costs you only what failed |
| Why Run LLMs on the Yens? | 🖊️ Concept | When to run a model yourself on the Yens vs. calling a cloud API — privacy, cost, reproducibility, and open vs. proprietary models |
| How to Run LLMs on the Yens | 🖊️💻 Concept + Hands-on | The three steps — loading a model, serving it, querying it — then querying a shared server yourself, and seeing the same model answer on a GPU and on a CPU |
| Handling LLM Failure Modes | 💻 Hands-on | The main ways LLMs fail — hallucination, inconsistency, a lack of guardrails — and how to make a pipeline more robust against them |
| Day 4 Challenge | 🔑 Capstone | Process every filing in the dataset with an array job — more filings than the scheduler allows tasks, so the mapping is yours to work out |
| Staying In Touch | 🏛️ Community | Where to get help after the course — the #gsb-yen-users Slack, the DARC team, and patterns to build on |