feat: use DeepSpeed ZeRO-3 for optimal 2-GPU training

training/scripts/train.py

@@ -34,16 +34,12 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # 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
+    # Load model - let the model's own quantization config handle it
+    # (Ornith uses CompressedTensors, not BitsAndBytes)
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
         torch_dtype=torch.bfloat16,
-        device_map="cpu",  # Load on CPU first
+        device_map="auto",  # Let transformers distribute across GPUs
     )
     
     # Add LoRA
@@ -93,13 +89,9 @@ def train(config_path):
         gradient_checkpointing=config.get("gradient_checkpointing", True),
     )
     
-    # Use FSDP for multi-GPU training
+    # SFT Trainer (DeepSpeed handles distributed training via config)
     from trl import SFTTrainer
     
-    # Prepare model for FSDP
-    model = accelerator.prepare(model)
-    
-    # SFT Trainer
     trainer = SFTTrainer(
         model=model,
         tokenizer=tokenizer,
@@ -109,9 +101,6 @@ 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()