Remove flash attention3 as the default attention backend during training

Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com>

Author: pthombre

dfm/src/automodel/recipes/train.py

@@ -171,7 +171,7 @@ def setup(self):
         self.rng = StatefulRNG(seed=self.seed, ranked=True)
 
         self.model_id = self.cfg.get("model.pretrained_model_name_or_path")
-        self.attention_backend = self.cfg.get("model.attention_backend", "_flash_3_hub")
+        self.attention_backend = self.cfg.get("model.attention_backend")
         self.learning_rate = self.cfg.get("optim.learning_rate", 5e-6)
         self.bf16 = torch.bfloat16