Fix the issue on latest transformers update

For the first iteration where we we are using `input_embeds` to start the generation, the conditional statement fails as the past_key_values now does not return none but returns `Dynamic_Cache()` according to the latest transformers version. Using `input_ids

modeling_phi.py

@@ -1161,7 +1161,7 @@ class PhiForCausalLM(PhiPreTrainedModel):
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
-        if inputs_embeds is not None and past_key_values is None:
+        if inputs_embeds is not None and (input_ids is None or input_ids.shape[1] == 0):
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
             model_inputs = {"input_ids": input_ids}