#!/usr/bin/env python3
# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "accelerate>=0.24.0",
#     "trackio",
# ]
# ///

from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import trackio
import os

print("🚀 Medium-Scale SFT Training with Trackio")
print("=" * 60)

# Initialize Trackio with Space sync
print("\n📊 Initializing Trackio...")
trackio.init(
    project="medium-sft-training",
    space_id="evalstate/trl-trackio-dashboard",
    config={
        "model": "Qwen/Qwen2.5-0.5B",
        "dataset": "trl-lib/Capybara",
        "dataset_size": 1000,
        "num_epochs": 3,
        "learning_rate": 2e-5,
        "batch_size": 4,
        "gradient_accumulation": 4,
        "lora_r": 16,
        "lora_alpha": 32,
        "hardware": "a10g-large",
    }
)
print("✅ Trackio initialized!")
print("📈 Dashboard: https://huggingface.co/spaces/evalstate/trl-trackio-dashboard")

# Load dataset - 1000 examples
print("\n📊 Loading dataset...")
dataset = load_dataset("trl-lib/Capybara", split="train[:1000]")
print(f"✅ Dataset loaded: {len(dataset)} examples")

# Get username
username = os.environ.get("HF_USERNAME", "evalstate")

# Training configuration - production settings
print("\n⚙️  Configuring training...")
config = SFTConfig(
    # Output and Hub settings
    output_dir="qwen-capybara-medium",
    push_to_hub=True,
    hub_model_id=f"{username}/qwen-capybara-medium",
    hub_strategy="every_save",  # Push all checkpoints

    # Training parameters - 3 epochs on 1K examples
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,  # Effective batch size = 16

    # Learning rate and schedule
    learning_rate=2e-5,
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",

    # Logging and checkpointing
    logging_steps=10,  # Log every 10 steps
    save_strategy="steps",
    save_steps=50,  # Save every 50 steps
    save_total_limit=3,  # Keep only 3 latest checkpoints

    # Evaluation
    eval_strategy="steps",
    eval_steps=50,

    # Optimization
    bf16=True,  # Use bfloat16 for A10G
    gradient_checkpointing=True,  # Save memory

    # Trackio monitoring
    report_to="trackio",
)

# LoRA configuration - larger than demo
print("🔧 Setting up LoRA (r=16)...")
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],  # More modules
)

# Create eval split
print("\n🔀 Creating train/eval split...")
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]
print(f"   Train: {len(train_dataset)} examples")
print(f"   Eval: {len(eval_dataset)} examples")

# Initialize trainer
print("\n🎯 Initializing trainer...")
trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
)

# Calculate training info
total_steps = (len(train_dataset) // (4 * 4)) * 3  # samples / (batch * grad_accum) * epochs
print(f"\n📊 Training Info:")
print(f"   Total steps: ~{total_steps}")
print(f"   Epochs: 3")
print(f"   Effective batch size: 16")
print(f"   Expected time: ~45-60 minutes")
print(f"   Checkpoints saved every 50 steps")

# Train!
print("\n🏃 Starting training...")
print("📈 Watch live metrics: https://huggingface.co/spaces/evalstate/trl-trackio-dashboard")
print("-" * 60)
trainer.train()

# Save to Hub
print("\n💾 Pushing final model to Hub...")
trainer.push_to_hub()

# Finish Trackio
print("\n📊 Finalizing Trackio metrics...")
trackio.finish()

print("\n" + "=" * 60)
print("✅ Training complete!")
print(f"📦 Model: https://huggingface.co/{username}/qwen-capybara-medium")
print(f"📊 Metrics: https://huggingface.co/spaces/evalstate/trl-trackio-dashboard")
print(f"💡 Try the model with:")
print(f'   from transformers import pipeline')
print(f'   generator = pipeline("text-generation", model="{username}/qwen-capybara-medium")')
print("=" * 60)