eval_codet5.py

from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    PreTrainedModel,
    PreTrainedTokenizer,
)
from core import run_eval, filter_code, fix_indents, standard_prompt
import os
import torch

# TODO: move to python-dotenv
# add hugging face access token here
TOKEN = ""


@torch.inference_mode()
def generate_batch_completion(
    model: PreTrainedModel, tokenizer: PreTrainedTokenizer, prompt, batch_size
) -> list[str]:
    prompt_input = standard_prompt(prompt)
    input_batch = [prompt for _ in range(batch_size)]
    inputs = tokenizer(input_batch, return_tensors="pt").to(model.device)

    generated_ids = model.generate(
        **inputs,
        use_cache=True,
        max_new_tokens=512,
        temperature=0.2,
        top_p=0.95,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )

    batch_completions = tokenizer.batch_decode(
        generated_ids,
        skip_special_tokens=True,
    )

    return [filter_code(fix_indents(completion)) for completion in batch_completions]


if __name__ == "__main__":
    # adjust for n = 10 etc
    num_samples_per_task = 10
    out_path = "results/codet5p_770/eval.jsonl"
    os.makedirs("results/codet5p_770", exist_ok=True)

    tokenizer = AutoTokenizer.from_pretrained(
        "Salesforce/codet5p-770m",
        trust_remote_code=True,
        use_auth_token=TOKEN,
    )

    model = torch.compile(
        AutoModelForSeq2SeqLM.from_pretrained(
            "Salesforce/codet5p-770m",
            torch_dtype=torch.bfloat16,
            trust_remote_code=True,
            use_auth_token=TOKEN,
        )
        .eval()
        .to("cuda:5")
    )

    run_eval(
        model,
        tokenizer,
        num_samples_per_task,
        out_path,
        generate_batch_completion,
        True,
    )