feat: Add ShortGPT pruning algorithm

[sky-2002]

Description

Implements the ShortGPT pruning algorithm.
Refer #418

Related Issue

Fixes #418

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Additional Notes

I am figuring out how to write tests for this, found the test framework a little tricky. Would appreciate some help there. I did not understand how the run_full_integration gets the dataset required to run tests, tokenizer etc.

Also, I found the SmashConfigPrefixWrapper confusing, so I tested with an alternative way:

from transformers import AutoModelForCausalLM, AutoTokenizer
from pruna.algorithms.shortgpt import ShortGPT
from datasets import load_dataset
algo = ShortGPT()

llama_tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B", 
llama_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B", torch_dtype="auto")

dataset = load_dataset("wikitext", "wikitext-2-raw-v1", split="train[:100]")

dl = DataLoader(dataset, batch_size=2, shuffle=True)

qmodel = algo._apply(llama_model, {
    "tokenizer": llama_tokenizer,
    "device": "cpu",
    "prune_ratio": 0.25,
    "angular": True,
    "train_dataloader": dl,
})

[cursor]

Comment @cursor review or bugbot run to trigger another review on this PR

[cursor]

Bug: NaNs: The Silent Killer of Predictable Behavior

If the dataloader is empty or yields no batches, counts remains 0, causing bis /= counts to produce NaN values. These NaN values propagate through np.argsort, resulting in unpredictable pruning behavior instead of a clear error message.

 

[github-actions]

This PR has been inactive for 10 days and is now marked as stale.