INSPIRE 2026 · STUDENT WORKSHOP

Fine-Tune It,
Then Make It Faster

Teach a language model to sort real complaints on a supercomputer, then make it faster.

Portal:

Before class

Log in at accounts.tacc.utexas.edu and set up MFA with an authenticator app (Okta Verify or Google Authenticator), not SMS. Take a screenshot of your account page and submit it. That is your pre-work.

After class, download your finished notebook (right-click it in Jupyter's file browser, then Download) and submit it.

Session 1: Fine-tune a model

Everyone does this on their own. One GPU each.

01

Log in to TAP

Go to tap.tacc.utexas.edu and click MFA Log In to TAP.

Click MFA Log In to TAP
Click MFA Log In to TAP

Enter your TACC username and password, then click Log In.

Enter your TACC username and password
Enter your TACC username and password

Type the code from your authenticator app and click Verify.

Enter the code from your authenticator app
Enter the code from your authenticator app
02

Start your GPU session

Fill in Submit New Job:

SettingChoose
SystemVista
ApplicationJupyter notebook
ProjectTRA25001
Queuegh (the GPU queue)
Nodes / Tasks1 / 1
Time Limit2:0:0 (two hours)
ReservationSWE1102

Then click Submit.

Submit New Job: Vista, Jupyter notebook, TRA25001, gh, and reservation SWE1102
Submit New Job: Vista, Jupyter notebook, TRA25001, gh, and reservation SWE1102
03

Open Jupyter

TAP shows your job status. It says PENDING while it waits for a GPU and refreshes on its own. When it is running, click Connect. Jupyter opens in your browser.

TAP Job Status: wait for PENDING to change to running
TAP Job Status: wait for PENDING to change to running
04

Get the workshop

In Jupyter, open a Terminal from the Launcher and type:

cd $HOME
git clone https://github.com/ashleyscruse/hpc-finetune-workshop

This puts the workshop in your home folder, so it shows up in Jupyter's file browser on the left.

05

Run it

In the file browser, open hpc-finetune-workshop/notebooks/01-fine-tune.ipynb.

Click a cell and press Shift + Enter. Go top to bottom. A cell is done when the [*] beside it turns into a number.

06

When you're done: end your session

Go back to TAP and click End Job so the GPU goes back. Your results are saved.

01

Log in to the supercomputer portal

Go to morehouse.tapis.io and click Log in with single sign-on.

Click Log in with single sign-on
Click Log in with single sign-on

Enter your TACC username and password, then click Log In.

Your TACC username and password
Your TACC username and password

Enter the code from your authenticator app and click Submit.

Your MFA code
Your MFA code
02

Connect your account to Vista

On the dashboard, click Systems in the left panel.

Click Systems
Click Systems

Click vista. If it says you are not yet authenticated, click Authenticate with TMS Keys.

Check it worked: AUTH says hasCredentials: true, ACCESS says listing works, and Browse files shows a list of your folders.

vista, connected: listing works, and your files show below
vista, connected: listing works, and your files show below
03

Start your GPU session

a. Pick the app

Click Apps, then click jupyter-hpc-native.

Apps, then jupyter-hpc-native
Apps, then jupyter-hpc-native

Click submit job at the top right.

submit job
submit job

b. Choose the guided launcher

Click Guided job launcher (v2).

Guided job launcher (v2)
Guided job launcher (v2)

c. Job Basics

Don't change anything. Click Continue at the bottom left.

Job Basics: click Continue
Job Basics: click Continue

d. Execution Options

Job Type: Batch. Batch Logical Queue: change it to gh (it starts on gh-dev). Node Count: 1. Maximum Minutes: 120. Then Continue.

Job Type, Batch Logical Queue, Node Count, Maximum Minutes
Job Type, Batch Logical Queue, Node Count, Maximum Minutes

e. File Inputs, Arguments, Environment Variables

Click Continue on each. Don't change anything.

f. Scheduler Options

Make sure the -A allocation and --reservation boxes are checked. Type TRA25001 in the allocation box and SWE1102 in the reservation box. Leave the job name row as it is. Then Continue.

Allocation TRA25001, reservation SWE1102
Allocation TRA25001, reservation SWE1102

g. Submit

On Archiving, click Continue. On Review & Submit, check that it says The job is ready for submission, then click Submit Job.

Review & Submit: click Submit Job
Review & Submit: click Submit Job
04

Open Jupyter

Click Jobs in the left panel and click your job. Check that it shows -A TRA25001 and rsv SWE1102. Its STATUS will change from staging to running.

Your job: our allocation and reservation, and its status
Your job: our allocation and reservation, and its status

When it is running, click the button to open Jupyter. No button? Open the output file tapisjob.out, copy the address after JUPYTER_URL, and open it in a new tab.

05

Get the workshop

In Jupyter, click File → New → Terminal and type:

cd $WORK
git clone https://github.com/ashleyscruse/hpc-finetune-workshop

Then open hpc-finetune-workshop/notebooks/01-fine-tune.ipynb.

06

Run it

Click a cell and press Shift + Enter. Go top to bottom. A cell is done when the [*] beside it turns into a number.

Stuck?

Raise your hand. Most problems are one of these:

What you seeLikely cause
Your session waits and never startsThe wrong project or reservation, or you still have another session running
"multiple projects" errorThe project was not set. It should be TRA25001
Can't log in to TAPCheck your TACC username and password, and wait for a fresh code from your authenticator app
Your job waits and never startsA typo in the allocation or reservation, or you still have another job running
"multiple projects" errorThe allocation box is empty or unchecked. It should say TRA25001
destination path already existsYou already have the workshop folder. Just open it
Can't find the workshop folder in the file browserIt was cloned somewhere else. Redo step 04, starting with cd $HOME
Can't submit, or Browse Files shows nothingYour TMS keys aren't set up. Redo step 02
"Can't see the shared folder"Raise your hand