@elizaos/training 2.0.0-alpha.10

package/python/scripts/run_tinker_training.py

@@ -0,0 +1,192 @@
+#!/usr/bin/env python3
+"""
+ElizaOS Tinker Training Script
+
+Run GRPO training using Tinker API (cloud-based, no local GPU required).
+
+Prerequisites:
+1. Set TINKER_API_KEY environment variable
+2. Set DATABASE_URL environment variable
+3. Set OPENAI_API_KEY for RLAIF judge
+
+Usage:
+    python scripts/run_tinker_training.py --steps 100 --model Qwen/Qwen3-30B-A3B-Instruct
+
+For help:
+    python scripts/run_tinker_training.py --help
+"""
+
+import asyncio
+import logging
+import os
+import sys
+from pathlib import Path
+
+# Add src to path for local development
+src_path = Path(__file__).parent.parent / "src"
+sys.path.insert(0, str(src_path))
+
+
+def check_environment() -> bool:
+    """Check required environment variables"""
+    missing = []
+    
+    if not os.environ.get("TINKER_API_KEY"):
+        missing.append("TINKER_API_KEY")
+    
+    if not os.environ.get("DATABASE_URL"):
+        missing.append("DATABASE_URL")
+    
+    if not os.environ.get("OPENAI_API_KEY"):
+        missing.append("OPENAI_API_KEY")
+    
+    if missing:
+        print("=" * 60)
+        print("  MISSING ENVIRONMENT VARIABLES")
+        print("=" * 60)
+        for var in missing:
+            print(f"  - {var}")
+        print()
+        print("Please set these before running:")
+        print("  export TINKER_API_KEY=your_key_here")
+        print("  export DATABASE_URL=postgresql://...")
+        print("  export OPENAI_API_KEY=sk-...")
+        print("=" * 60)
+        return False
+    
+    return True
+
+
+async def main() -> int:
+    """Main entry point"""
+    import argparse
+    
+    # Setup logging
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+    
+    parser = argparse.ArgumentParser(
+        description="ElizaOS Tinker Training",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+  # Basic training run
+  python scripts/run_tinker_training.py --steps 100
+  
+  # Use larger model
+  python scripts/run_tinker_training.py --model Qwen/Qwen3-235B-A22B-Instruct
+  
+  # Adjust hyperparameters
+  python scripts/run_tinker_training.py --lr 1e-5 --group-size 8 --lora-rank 64
+        """,
+    )
+    
+    parser.add_argument(
+        "--model",
+        default="Qwen/Qwen3-30B-A3B-Instruct",
+        help="Base model to train (default: Qwen/Qwen3-30B-A3B-Instruct)",
+    )
+    parser.add_argument(
+        "--steps",
+        type=int,
+        default=100,
+        help="Number of training steps (default: 100)",
+    )
+    parser.add_argument(
+        "--group-size",
+        type=int,
+        default=4,
+        help="GRPO group size - trajectories compared per step (default: 4)",
+    )
+    parser.add_argument(
+        "--lr",
+        type=float,
+        default=4e-5,
+        help="Learning rate (default: 4e-5)",
+    )
+    parser.add_argument(
+        "--lora-rank",
+        type=int,
+        default=32,
+        help="LoRA rank (default: 32)",
+    )
+    parser.add_argument(
+        "--weight-sync-interval",
+        type=int,
+        default=5,
+        help="Steps between weight syncs (default: 5)",
+    )
+    parser.add_argument(
+        "--log-file",
+        default="./logs/tinker_training_metrics.jsonl",
+        help="Metrics log file path",
+    )
+    parser.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Check environment without running training",
+    )
+    
+    args = parser.parse_args()
+    
+    # Check environment
+    if not check_environment():
+        return 1
+    
+    if args.dry_run:
+        print("\n✓ Environment check passed. Ready to train.")
+        return 0
+    
+    # Import trainer (after environment check)
+    from training.tinker_trainer import (
+        TinkerTrainer,
+        TinkerTrainingConfig,
+    )
+    
+    # Create config
+    config = TinkerTrainingConfig(
+        base_model=args.model,
+        training_steps=args.steps,
+        group_size=args.group_size,
+        learning_rate=args.lr,
+        lora_rank=args.lora_rank,
+        weight_sync_interval=args.weight_sync_interval,
+        database_url=os.environ["DATABASE_URL"],
+        log_file=args.log_file,
+    )
+    
+    # Run training
+    print("\n" + "=" * 60)
+    print("  ELIZAOS TINKER TRAINING")
+    print("=" * 60)
+    print(f"  Model: {config.base_model}")
+    print(f"  Steps: {config.training_steps}")
+    print(f"  Group size: {config.group_size}")
+    print(f"  Learning rate: {config.learning_rate}")
+    print(f"  LoRA rank: {config.lora_rank}")
+    print("=" * 60 + "\n")
+    
+    trainer = TinkerTrainer(config)
+    result = await trainer.train()
+    
+    if result.get("success"):
+        print("\n" + "=" * 60)
+        print("  ✓ TRAINING COMPLETE")
+        print("=" * 60)
+        print(f"  Run ID: {result['run_id']}")
+        print(f"  Steps completed: {result['steps']}")
+        print(f"  Windows processed: {result['windows_processed']}")
+        print(f"  Final weights: {result['final_weights']}")
+        if result.get("metrics_file"):
+            print(f"  Metrics: {result['metrics_file']}")
+        print("=" * 60)
+        return 0
+    else:
+        print("\n✗ Training failed")
+        return 1
+
+
+if __name__ == "__main__":
+    sys.exit(asyncio.run(main()))