fix: use torchrun for distributed training, load model on CPU

training/scripts/train.py

@@ -36,10 +36,11 @@ def train(config_path):
     
     # Load model - let the model's own quantization config handle it
     # (Ornith uses CompressedTensors, not BitsAndBytes)
+    # Load on CPU first, then DeepSpeed will distribute
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        torch_dtype=torch.bfloat16,
-        device_map="auto",  # Let transformers distribute across GPUs
+        dtype=torch.bfloat16,
+        device_map="cpu",  # Load on CPU, DeepSpeed distributes
     )
     
     # Add LoRA