Homework 3

The homework has two parts.

Part 1: Concepts component - Handout

This part is about sparse and efficient transformers as well as a new hardware efficiency method called Prompt Cache.
Download the homework handout from the following link: [Download]

Complete the handout in latex and submit to Gradescope.

Part 2: Hands-on excercise: Implementing a Diffusion Model

The second part is implementation of a Diffusion Model and specifically the Denoising Diffusion Probabilistic Models.
A Colab notebook is provided to guide you through the process of implementing the DDPM model from scratch and training it on a toy data sample.
Please see the instructions and the notebook here: Colab.

Follow the notebook, complete the exercises, and download your final .ipynb file and submit it through Gradescope.

Homework 2

The homework has two parts.

Please download the files from the following link: Download

Prompting and in-context learning

This part focuses on in-context learning using an open-source Large Language Model.
You will be working with Mistral-7B-0.1, a recently-released large language model, with the goal of finding a prompting technique that causes the model to obtain the best performance on GSM8K, a dataset of middle-school-level mathematics word problems.

Parameter-efficient finetuning using IA3

In this focuses on the IA3 algorithm for parameter-efficient fine-tuning (PEFT).
You will be using the HuggingFace Transformers library to fine-tune a pre-trained model on the XSum dataset for abstractive summarization. The goal is to compare the performance of the IA3 PEFT method with the baseline full fine-tuning approach.

The detailed instructions for completing each homework is in the corresponding README file.

Homework 1

This homework has two parts. One hand out and one coding section.
Download the homework from the following link: [Download]

For each part instructions are in the file.


- CPSC_488_Fall_2023_HW_1.pdf   # the handout
- tensor.py                     # implementing auto differentiation and back propagation
- tokenizer.py                  # implementing a tokenizer
- transformer.ipynb                # implementing GQA
- tokenizer_test.py             # test file for tokenizer
- tensor_test.py                # test file for the tensor implementation

For the handout, complete the solutions in latex and return the pdf file.

For the tensor.py and tokenizer.py file, the instructions are in the file.

The transformer.ipynb also includes detailed instructions.

You can submit the assignment in a zip file with the following format: lastname-firstname-hw1.zip.


  • Handout: 30 points (each Q gets 6 points)
  • Backpropagation: 20 points
  • Tokenizer: 40 points
  • Transformer GQA: 10 points