iscra benchmark llm

gistfile1.txt

import os
import json
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from datasets import load_dataset
import torch
import time
import argparse

# Function to log training details for benchmarking
def log_training_details(training_args, logs, num_gpus, num_cpus, batch_time, epochs):
    details = {
        "num_gpus": num_gpus,
        "num_cpus": num_cpus,
        "batch_size_per_device": training_args.per_device_train_batch_size,
        "gradient_accumulation_steps": training_args.gradient_accumulation_steps,
        "effective_batch_size": training_args.per_device_train_batch_size * num_gpus * training_args.gradient_accumulation_steps,
        "learning_rate": training_args.learning_rate,
        "epochs": training_args.num_train_epochs,
        "train_runtime": logs["train_runtime"],
        "train_samples_per_second": logs["train_samples_per_second"],
        "batch_data_processing_time":batch_time,
        "num_epochs":epochs
    }
    tbs = training_args.per_device_train_batch_size * num_gpus * training_args.gradient_accumulation_steps
    with open(f"training_log_cpu_bs_{tbs}_epochs_{epochs}_{num_cpus}_gpu_{num_gpus}.json", "w") as f:
        json.dump(details, f, indent=4)

# Load BookCorpus dataset
# Load dataset
def load_custom_dataset():
    print("Loading dataset...")
    dataset = load_dataset("openwebtext", cache_dir="custom_dataset", split="train[:10000]")
    print("Dataset loaded successfully!")
    return dataset

# Tokenize dataset
def preprocess_data(examples):
    tokenized = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=128)
    # Add labels for causal language modeling
    tokenized["labels"] = tokenized["input_ids"].copy()
    return tokenized

# Load GPT-2 model and tokenizer
model_name = "gpt2"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
raw_dataset = load_custom_dataset()
tokenized_dataset = raw_dataset.map(preprocess_data, batched=True, num_proc=os.cpu_count())
tokenized_dataset = tokenized_dataset.remove_columns(["text"]).with_format("torch")
print("Data ready!!!!")
# Argument controls
def train_gpt2(num_gpus, num_cpus, batch_size, epochs, scaling_type="weak"):
    # Set environment variables for GPUs and CPUs
    os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(map(str, range(num_gpus)))
    torch.set_num_threads(num_cpus)

    training_args = TrainingArguments(
        output_dir="./results",
        evaluation_strategy="no",
        save_strategy="epoch",
        per_device_train_batch_size=batch_size,
        gradient_accumulation_steps=2,
        learning_rate=5e-5,
        num_train_epochs=epochs,
        save_total_limit=2,
        load_best_model_at_end=False,
        logging_dir="./logs",
        logging_steps=10,
        fp16=True if torch.cuda.is_available() else False,  # Mixed precision for GPUs
        report_to="none",
        dataloader_num_workers=num_cpus,
        ddp_find_unused_parameters=False,
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_dataset,
        eval_dataset=None,
    )


    # Measure batch processing time
    data_loader = trainer.get_train_dataloader()
    start_time = time.time()
    for i, batch in enumerate(data_loader):
        if i == 10:  # Benchmark on the first 10 batches
            break
        pass
    end_time = time.time()
    batch_time = (end_time - start_time) / 10
    # print(batch_time)
    # Train the model and log details
    logs = trainer.train()
    log_training_details(training_args, logs.metrics, num_gpus, num_cpus, batch_time, epochs)

# Argument parser
def parse_args():
    parser = argparse.ArgumentParser(description="Train GPT-2 with customizable parameters")
    parser.add_argument("--num_gpus", type=int, default=1, help="Number of GPUs to use")
    parser.add_argument("--num_cpus", type=int, default=8, help="Number of CPUs to use")
    parser.add_argument("--batch_size", type=int, default=128, help="Batch size per device")
    parser.add_argument("--epochs", type=int, default=10, help="Epochs")
    parser.add_argument("--scaling_type", type=str, default="weak", help="type of scaling")
    return parser.parse_args()

if __name__ == "__main__":
    args = parse_args()
    train_gpt2(num_gpus=args.num_gpus, num_cpus=args.num_cpus, batch_size=args.batch_size, epochs=args.epochs)