llm_toolkit/eval.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.translation_engine import *
from llm_toolkit.translation_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("DATA_PATH")
results_path = os.getenv("RESULTS_PATH")

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
)

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_translation_dataset(data_path, tokenizer)

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

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(
    model_name,
    results_path,
    datasets["test"],
    predictions,
    debug=True,
)

metrics = calc_metrics(datasets["test"]["english"], predictions, debug=True)
print(metrics)