refactor CPU llama inference code

optimum/exporters/ipex/model_patcher.py

@@ -13,7 +13,6 @@
 #  limitations under the License.
 
 from transformers.models.llama.modeling_llama import (
-    LlamaAttention,
     LlamaDecoderLayer,
     LlamaForCausalLM,
     LlamaModel,
@@ -23,8 +22,7 @@
 from optimum.intel.utils.import_utils import is_ipex_version
 
 from .modeling_utils import (
-    _IPEXLlamaDecoderLayerRef,
-    _llama_attn_forward,
+    _IPEXLlamaDecoderLayer,
     _llama_layer_norm_forward,
     _llama_model_forward,
 )
@@ -62,26 +60,12 @@ def patch_op(m, target_m, new_op_name, new_op):
 
 
 def _patch_llama_model(model):
-    if is_ipex_version("<", "2.5.0"):
-        raise ImportError("Only ipex version > 2.3.0 supports RotaryEmbedding and IndirectAccessKVCache")
-
-    from intel_extension_for_pytorch.llm.modules import IndirectAccessKVCache, RotaryEmbedding
-
-    ipex_rope = RotaryEmbedding(
-        model.config.max_position_embeddings,
-        model.config.hidden_size // model.config.num_attention_heads,
-        model.config.rope_theta,
-        model.config.architectures[0],
-    )
-    ipex_scale_dot_product = IndirectAccessKVCache(text_max_length=model.config.max_position_embeddings)
-    patch_op(model, LlamaAttention, "ipex_rope", ipex_rope)
-    patch_op(model, LlamaAttention, "ipex_scale_dot_product", ipex_scale_dot_product)
+    if is_ipex_version("<", "2.3.0"):
+        raise ImportError("Only ipex version >= 2.3.0 supports llama model patching")
 
     convert_functions(model, LlamaModel, "forward", _llama_model_forward)
-    convert_functions(model, LlamaAttention, "forward", _llama_attn_forward)
     convert_functions(model, LlamaRMSNorm, "forward", _llama_layer_norm_forward)
-
-    convert_class(model, LlamaDecoderLayer, _IPEXLlamaDecoderLayerRef, model.config)
+    convert_class(model, LlamaDecoderLayer, _IPEXLlamaDecoderLayer, model.config)
     return model