[`Llama + AWQ`] fix `prepare_inputs_for_generation` 🫠 (#29381)

* use the generation config 🫠

* fixup

src/transformers/models/gemma/modeling_gemma.py

@@ -1161,7 +1161,7 @@ class GemmaForCausalLM(GemmaPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
-        if getattr(self.model.layers[0].self_attn, "past_key_value", None) is not None:
+        if self.generation_config.cache_implementation == "static":
             # generation with static cache
             cache_position = kwargs.get("cache_position", None)
             if cache_position is None:

src/transformers/models/llama/modeling_llama.py

@@ -1277,7 +1277,7 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
-        if getattr(self.model.layers[0].self_attn, "past_key_value", None) is not None:
+        if self.generation_config.cache_implementation == "static":
             # generation with static cache
             cache_position = kwargs.get("cache_position", None)
             if cache_position is None: