doing lm_eval's work

eval_script.py

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer

from lm_eval.models.huggingface import HFLM
from lm_eval.evaluator import evaluate
from lm_eval.tasks import get_task_dict


path_to_hf_checkpoint = "/home/cdhernandez/local/gpt-fast/checkpoints/meta-llama/Meta-Llama-3-8B"
task_list = ["wikitext"]
device = "cuda"
precision = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(path_to_hf_checkpoint)
model = AutoModelForCausalLM.from_pretrained(path_to_hf_checkpoint).to(device="cuda", dtype=precision)

from torchao.quantization.quant_api import change_linear_weights_to_int4_woqtensors

# your API Here
change_linear_weights_to_int4_woqtensors(model)

with torch.no_grad():
    result = evaluate(
        HFLM(pretrained=model, tokenizer=tokenizer),
        get_task_dict(task_list),
        limit = 10
    )
for task, res in result["results"].items():
    print(f"{task}: {res}")

• Remove path to checkpoint: we can just download hte model directly from the path from hf
• Explicitly list out public APIs that people can try (it seems like there's only 4)
• Put this in docs/ folder in ao titled how to run an eval or create a new folder called scripts/ and call this eval.py