llm_toolkit/eval_logical_reasoning.py

import os
import sys
import torch
from dotenv import find_dotenv, load_dotenv

found_dotenv = find_dotenv(".env")

if len(found_dotenv) == 0:
    found_dotenv = find_dotenv(".env.example")
print(f"loading env vars from: {found_dotenv}")
load_dotenv(found_dotenv, override=False)

path = os.path.dirname(found_dotenv)
print(f"Adding {path} to sys.path")
sys.path.append(path)

from llm_toolkit.llm_utils import *
from llm_toolkit.logical_reasoning_utils import *

model_name = os.getenv("MODEL_NAME")
adapter_name_or_path = os.getenv("ADAPTER_NAME_OR_PATH")
load_in_4bit = os.getenv("LOAD_IN_4BIT") == "true"
data_path = os.getenv("LOGICAL_REASONING_DATA_PATH")
results_path = os.getenv("LOGICAL_REASONING_RESULTS_PATH")
use_english_datasets = os.getenv("USE_ENGLISH_DATASETS") == "true"
using_p1 = os.getenv("USING_P1_PROMPT_TEMPLATE") == "true"
test_data = os.getenv("TEST_DATA", None)
using_llama_factory = os.getenv("USING_LLAMA_FACTORY") == "true"
max_new_tokens = int(os.getenv("MAX_NEW_TOKENS", 16))
repetition_penalty = float(os.getenv("REPETITION_PENALTY", 1.0))
batch_size = int(os.getenv("BATCH_SIZE", 2))

dtype = (
    torch.float32
    if os.getenv("USE_FLOAT32_FOR_INFERENCE") == "true"
    else (
        torch.bfloat16
        if os.getenv("USE_BF16_FOR_INFERENCE") == "true"
        else torch.float16
    )
)

print(model_name, adapter_name_or_path, load_in_4bit, data_path, results_path)

gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(1) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")

model, tokenizer = load_model(
    model_name,
    load_in_4bit=load_in_4bit,
    adapter_name_or_path=adapter_name_or_path,
    using_llama_factory=using_llama_factory,
    dtype=dtype,
)

gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(2) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")

datasets = load_logical_reasoning_dataset(
    data_path,
    tokenizer=tokenizer,
    chinese_prompt=not use_english_datasets,
    using_p1=using_p1,
    test_data=test_data,
)

if len(sys.argv) > 1:
    num = int(sys.argv[1])
    if num > 0:
        print(f"--- evaluating {num} entries")
        datasets["test"] = datasets["test"].select(range(num))

print_row_details(datasets["test"].to_pandas(), indices=[0, -1])

print("Evaluating model: " + model_name)
predictions = eval_model(
    model,
    tokenizer,
    datasets["test"],
    max_new_tokens=max_new_tokens,
    repetition_penalty=repetition_penalty,
    batch_size=batch_size,
)

gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(3) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")

if adapter_name_or_path is not None:
    model_name += "/" + adapter_name_or_path.split("/")[-1]

save_results(
    (
        "answer"
        if test_data
        else f"{model_name}_{dtype}{'_4bit' if load_in_4bit else ''}{'_lf' if using_llama_factory else ''}"
    ),
    results_path,
    datasets["test"],
    predictions,
    debug=True,
)

if not test_data:
    metrics = calc_metrics(datasets["test"]["label"], predictions, debug=True)
    print(metrics)