From 114e25fcfb0a0ee79861ac8c36b52c00dcb9fe2e Mon Sep 17 00:00:00 2001
From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com>
Date: Tue, 30 Jun 2026 23:45:57 -0400
Subject: [PATCH] fix: use device_map=auto to distribute model across GPUs

---
 training/scripts/train.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/training/scripts/train.py b/training/scripts/train.py
index 7420404..92c6aff 100755
--- a/training/scripts/train.py
+++ b/training/scripts/train.py
@@ -34,10 +34,11 @@ def train(config_path):
     
     # Load model in bf16 (fp8 not supported for training)
     # Disable quantization config to train full precision
+    # Use device_map="auto" to distribute across GPUs from the start
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
         torch_dtype=torch.bfloat16,
-        device_map="cpu",
+        device_map="auto",  # Distribute across GPUs
         quantization_config=None,  # Override any fp8 quantization
         trust_remote_code=True,
     )