revert

Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

Author: akoumpa

nemo_automodel/_transformers/auto_model.py

@@ -434,7 +434,7 @@ def from_pretrained(
         use_sdpa_patching: bool = True,
         sdpa_method: Optional[List[SDPBackend]] = None,
         torch_dtype="auto",
-        attn_implementation: str = "sdpa", #flash_attention_2",
+        attn_implementation: str = "flash_attention_2",
         quantization_config=None,
         force_hf: bool = False,
         **kwargs,