main.py

# Ensure Apple Metal (MPS) is enabled
import torch
import os
from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed
from datasets import load_dataset
from peft import LoraConfig, TaskType
from trl import SFTConfig, SFTTrainer
from enum import Enum

# ✅ Set device to Metal Performance Shaders (MPS) for Mac M3
device = "mps" if torch.backends.mps.is_available() else "cpu"
print(f"Using device: {device}")

# ✅ Set seed for reproducibility
set_seed(42)

# ✅ Model and dataset
model_name = "google/gemma-2-2b-it"
dataset_name = "Jofthomas/hermes-function-calling-thinking-V1"
tokenizer = AutoTokenizer.from_pretrained(model_name, token=True)

# ✅ Adjust tokenizer with special tokens
class ChatmlSpecialTokens(str, Enum):
    tools = "<tools>"
    eotools = "</tools>"
    think = "<think>"
    eothink = "</think>"
    tool_call="<tool_call>"
    eotool_call="</tool_call>"
    tool_response="<tool_response>"
    eotool_response="</tool_response>"
    pad_token = "<pad>"
    eos_token = "<eos>"

    @classmethod
    def list(cls):
        return [c.value for c in cls]

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    pad_token=ChatmlSpecialTokens.pad_token.value,
    additional_special_tokens=ChatmlSpecialTokens.list()
)

# ✅ Load model and move it to MPS
model = AutoModelForCausalLM.from_pretrained(model_name, token=True, attn_implementation="eager")
model.resize_token_embeddings(len(tokenizer))
model.to(device)

# ✅ Data preprocessing function
def preprocess(sample):
    messages = sample["messages"]

    if not messages or not isinstance(messages, list):
        return {"text": ""}  # Return empty text if messages are missing

    first_message = messages[0]

    # Ensure system messages are merged with the first user message
    if first_message["role"] == "system":
        system_message_content = first_message.get("content", "")
        if len(messages) > 1 and messages[1]["role"] == "user":
            messages[1]["content"] = (
                    system_message_content
                    + "\n\nAlso, before making a call to a function, take the time to plan the function to take. "
                    + "Make that thinking process between <think>{your thoughts}</think>\n\n"
                    + messages[1].get("content", "")
            )
            messages.pop(0)  # Remove system message

    # Ensure the conversation alternates between "user" and "assistant"
    valid_roles = ["user", "assistant"]
    cleaned_messages = [
        msg for msg in messages if msg.get("role") in valid_roles and msg.get("content")
    ]

    # Check if messages are empty after cleanup
    if not cleaned_messages or cleaned_messages[0]["role"] != "user":
        return {"text": ""}  # Ensure the first message is always from the user

    # Apply chat template
    try:
        formatted_text = tokenizer.apply_chat_template(cleaned_messages, tokenize=False)
        return {"text": formatted_text}
    except Exception as e:
        print(f"Error processing message: {e}")
        return {"text": ""}

# ✅ Load dataset
dataset = load_dataset(dataset_name, cache_dir="/tmp")
dataset = dataset.rename_column("conversations", "messages")
dataset = dataset.map(preprocess, remove_columns=["messages"])
dataset = dataset["train"].train_test_split(0.1)

# ✅ Print dataset size before training
print(f"Training dataset size: {len(dataset['train'])} samples")
print(f"Evaluation dataset size: {len(dataset['test'])} samples")

# ✅ LoRA configuration
peft_config = LoraConfig(
    r=16,
    lora_alpha=64,
    lora_dropout=0.05,
    target_modules=["gate_proj", "q_proj", "lm_head", "o_proj", "k_proj", "embed_tokens", "down_proj", "up_proj", "v_proj"],
    task_type=TaskType.CAUSAL_LM,
    bias="none",
)

# ✅ Training configuration (adjusted for performance on Mac M3 Max)
num_train_epochs = 5  # ✅ Increase to 5 epochs for better training
max_steps = 1000  # ✅ Ensure at least 1000 training steps
learning_rate = 5e-5  # ✅ Reduce learning rate to prevent overfitting

training_arguments = SFTConfig(
    output_dir="gemma-2-2B-it-macM3",
    per_device_train_batch_size=2,  # ✅ Keep small if training on MPS
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=4,  # ✅ Helps fit larger batch sizes
    save_strategy="epoch",
    save_total_limit=2,
    save_safetensors=False,
    evaluation_strategy="epoch",
    logging_steps=5,
    learning_rate=learning_rate,
    max_grad_norm=1.0,
    weight_decay=0.1,
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    report_to="tensorboard",
    bf16=True,  # ✅ Efficient mixed precision training for Mac MPS
    push_to_hub=False,
    num_train_epochs=num_train_epochs,
    max_steps=max_steps,  # ✅ Ensure training runs for at least 1000 steps
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    packing=True,
    max_seq_length=1500,
)

# ✅ Trainer setup
trainer = SFTTrainer(
    model=model,
    args=training_arguments,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    processing_class=tokenizer,
    peft_config=peft_config,
)

# ✅ Start training (should work efficiently on Mac M3 Max)
trainer.train()
trainer.save_model()

print("Training complete! 🚀 Model saved successfully.")