pytorch-finetuning: pin full fine-tuning model to one device (#687)

device_map=auto offloaded gemma-3-4b across GPU and CPU on halo; full
fine-tuning cannot train across that split and crashed. Pin to the GPU.
Fixes #672.

Author: lucbruni-amd

playbooks/supplemental/pytorch-finetuning/assets/train_full_finetuning.py

@@ -96,10 +96,11 @@ def format_chat(ex):
 print("Note: Model is stored as MXFP4 on Hugging Face but will be loaded as BF16 for training")
 print("(This is expected - the warning about MXFP4 is informational)\n")
 
+device = "cuda" if torch.cuda.is_available() else "cpu"
 model = AutoModelForCausalLM.from_pretrained(
     MODEL,
     dtype=torch.bfloat16,             # Use BF16 for better stability and ROCm support (dtype not torch_dtype)
-    device_map="auto",                # Automatically distribute across available GPUs
+    device_map=device,
     trust_remote_code=True,
     low_cpu_mem_usage=True            # Reduce CPU memory during loading
 )