fix: use FSDP for 2-GPU training, reduce batch size

training/scripts/train.py

@@ -34,12 +34,16 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model - let the model's own quantization config handle it
-    # (Ornith uses CompressedTensors, not BitsAndBytes)
+    # Load model with distributed training support
+    # Use FSDP for multi-GPU training
+    from accelerate import Accelerator
+    accelerator = Accelerator()
+    
+    # Load model on CPU first, then distribute
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        device_map="auto",
         torch_dtype=torch.bfloat16,
+        device_map="cpu",  # Load on CPU first
     )
     
     # Add LoRA
@@ -89,6 +93,12 @@ def train(config_path):
         gradient_checkpointing=config.get("gradient_checkpointing", True),
     )
     
+    # Use FSDP for multi-GPU training
+    from trl import SFTTrainer
+    
+    # Prepare model for FSDP
+    model = accelerator.prepare(model)
+    
     # SFT Trainer
     trainer = SFTTrainer(
         model=model,
@@ -99,6 +109,9 @@ def train(config_path):
         max_seq_length=config["train_params"]["max_seq_length"],
     )
     
+    # Prepare trainer for distributed training
+    trainer = accelerator.prepare(trainer)
+    
     # Train
     print("Starting training...")
     trainer.train()