[`BC`] Fix BC for AWQ quant (#29965)

fix awq quant

src/transformers/models/cohere/modeling_cohere.py

@@ -963,7 +963,7 @@ class CohereModel(CoherePreTrainedModel):
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
-        if hasattr(self.layers[0].self_attn, "past_key_value"):  # static cache
+        if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (

src/transformers/models/gemma/modeling_gemma.py

@@ -971,7 +971,7 @@ class GemmaModel(GemmaPreTrainedModel):
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
-        if hasattr(self.layers[0].self_attn, "past_key_value"):  # static cache
+        if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (

src/transformers/models/llama/modeling_llama.py

@@ -1064,7 +1064,7 @@ class LlamaModel(LlamaPreTrainedModel):
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
-        if hasattr(self.layers[0].self_attn, "past_key_value"):  # static cache
+        if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (