[`FA2`/ `Mistral`] Revert previous behavior with right padding + forward (#27125)

Update modeling_mistral.py

src/transformers/models/mistral/modeling_mistral.py

@@ -820,6 +820,7 @@ class MistralModel(MistralPreTrainedModel):
             attention_mask is not None
             and hasattr(self.config, "_flash_attn_2_enabled")
             and self.config._flash_attn_2_enabled
+            and past_key_values is not None
         ):
             is_padding_right = attention_mask[:, -1].sum().item() != batch_size
             if is_padding_right: