scripts/train.py

#!/usr/bin/env python3
import torch
from torch.utils.data import DataLoader
from transformers import (
    AutoTokenizer,
    TrainingArguments,
    Trainer,
    default_data_collator,
)
from datasets import load_dataset
from myolmoe import MyOlmoeForCausalLM, OlmoeConfig
import os
from transformers import TrainerCallback
import subprocess

def main():
    print("Starting my COOL OLMoE training script for small experts")
    # Load config - first try from local file, then from pretrained
    config_path = os.path.join("myolmoe", "config.json")
    if os.path.exists(config_path):
        config = OlmoeConfig.from_json_file(config_path)
    else:
        config = OlmoeConfig.from_pretrained("myolmoe")
    
    # Load model
    model = MyOlmoeForCausalLM.from_pretrained(
        "myolmoe",
        config=config,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        ignore_mismatched_sizes=True
    )
    
    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained("myolmoe")
    tokenizer.pad_token = tokenizer.eos_token
    
    # Load dataset
    dataset = load_dataset("allenai/tulu-v2-sft-mixture", split="train")
    
    def tokenize_function(examples):
        texts = []
        for message_list in examples["messages"]:
            formatted = ""
            for msg in message_list:
                role = msg["role"]
                content = msg["content"]
                if role == "user":
                    formatted += f"User: {content}\n"
                elif role == "assistant":
                    formatted += f"Assistant: {content}\n"
                else:
                    formatted += f"{role.capitalize()}: {content}\n"
            texts.append(formatted)

        tokenized = tokenizer(
            texts,
            truncation=True,
            max_length=4096,
            padding="max_length"
        )
        tokenized["labels"] = tokenized["input_ids"].copy()
        return tokenized

    tokenized_dataset = dataset.map(
        tokenize_function,
        batched=True,
        remove_columns=dataset.column_names,
        num_proc=4
    )
    
    # Training arguments
    training_args = TrainingArguments(
        output_dir="./checkpoints",
        per_device_train_batch_size=2,
        gradient_accumulation_steps=8,
        learning_rate=1e-4,
        num_train_epochs=3,
        logging_dir="./logs",
        logging_steps=10,
        save_steps=2000,
        save_total_limit=2,
        bf16=True,
        gradient_checkpointing=False,  # Disabled for now
        report_to="tensorboard",
        optim="adamw_torch",
        lr_scheduler_type="cosine",
        warmup_ratio=0.1,
        max_grad_norm=1.0,
    )
    
    # Freeze all parameters first
    for param in model.parameters():
        param.requires_grad = False
    
    # Unfreeze only the small experts and their gating networks
    trainable_params = []
    for name, param in model.named_parameters():
        if (
            "small_experts" in name or
            "small_gate" in name
        ):
            param.requires_grad = True
            trainable_params.append(name)

    
    print(f"Total trainable parameters: {len(trainable_params)}")
    
    # Verify gradient requirements
    for name, param in model.named_parameters():
        if param.requires_grad:
            print(f"Parameter {name} requires grad: {param.requires_grad}")

    # Custom data collator
    def data_collator(features):
        batch = default_data_collator(features)
        batch["output_router_logits"] = True
        return batch

    # Fixed CustomTrainer class that handles all possible arguments
    class CustomTrainer(Trainer):
        def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
            # Remove any unexpected arguments
            inputs = {k: v for k, v in inputs.items() if k not in ['num_items_in_batch']}
            
            # Ensure we're in training mode
            model.train()
            
            # Forward pass with gradients
            with torch.set_grad_enabled(True):
                outputs = model(**inputs)
                loss = outputs.loss
                
                if not loss.requires_grad:
                    raise RuntimeError("Loss doesn't require gradients. Check model parameters.")
                
                return (loss, outputs) if return_outputs else loss
            
    class GitPushCallback(TrainerCallback):
        def on_save(self, args, state, control, **kwargs):
            try:
                print("Saving checkpoint to Git repo...")
                
                # Add all changes (you can scope this to ./checkpoints/ if desired)
                subprocess.run(["git", "add", "."], check=True)

                # Skip commit if no changes
                result = subprocess.run(["git", "diff", "--cached", "--quiet"])
                if result.returncode == 0:
                    print("No changes to commit.")
                    return

                subprocess.run(["git", "commit", "-m", f'Checkpoint at step {state.global_step}'], check=True)
                subprocess.run(["git", "push"], check=True)
                print("Checkpoint pushed successfully.")
            except subprocess.CalledProcessError as e:
                print(f"Git push failed: {e}")

    # Initialize trainer
    trainer = CustomTrainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_dataset,
        data_collator=data_collator,
        callbacks=[GitPushCallback()]
    )
    
    # Test forward/backward pass before training
    print("Testing gradient flow...")
    test_loader = DataLoader(tokenized_dataset, batch_size=1, collate_fn=data_collator)
    test_batch = next(iter(test_loader))
    
    # Move batch to model's device
    device = next(model.parameters()).device
    test_batch = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in test_batch.items()}
    
    model.train()
    outputs = model(**test_batch)
    loss = outputs.loss
    print(f"Initial loss: {loss.item()}")
    
    loss.backward()
    print("Gradients computed successfully")
    
    # Check which parameters received gradients
    for name, param in model.named_parameters():
        if param.grad is not None:
            print(f"Parameter {name} received gradients")
    
    # Reset gradients
    model.zero_grad()

    # Check for existing checkpoint
    import re

    checkpoint_dir = None
    if os.path.isdir(training_args.output_dir):
        checkpoints = [
            os.path.join(training_args.output_dir, d)
            for d in os.listdir(training_args.output_dir)
            if re.match(r"checkpoint-\d+", d)
        ]
        if checkpoints:
            # Extract step numbers and find the highest
            checkpoint_dir = max(checkpoints, key=lambda x: int(x.split('-')[-1]))
            print(f"Resuming from checkpoint: {checkpoint_dir}")


    # Train
    print("Starting training...")
    trainer.train(resume_from_checkpoint=checkpoint_dir)

    # Save only the small experts and gates
    print("Saving small experts and gates...")
    small_expert_state_dict = {
        name: param for name, param in model.named_parameters()
        if name in trainable_params
    }
    
    os.makedirs("./final_model", exist_ok=True)
    torch.save(small_expert_state_dict, "./final_model/small_experts_and_gates.bin")
    config.save_pretrained("./final_model")

if __name__ == "__main__":
    main()