diff --git a/python/llm/src/ipex_llm/transformers/npu_model.py b/python/llm/src/ipex_llm/transformers/npu_model.py
index 673a56ecfaa..b952b1aef8a 100644
--- a/python/llm/src/ipex_llm/transformers/npu_model.py
+++ b/python/llm/src/ipex_llm/transformers/npu_model.py
@@ -290,7 +290,8 @@ def optimize_npu_model(cls, *args, **kwargs):
             model.config.update({"group_size": quantization_group_size})
             model.config.update({"asym": qtype == "asym_int4_rtn"})
             optimize_llm_pre(model, qtype, mixed_precision,
-                             quantization_group_size=quantization_group_size)
+                             quantization_group_size=quantization_group_size,
+                             max_prompt_len=max_prompt_len)
             cls.load_convert(qtype, model, "cpu", modules_to_not_convert,
                              quantization_group_size, imatrix_data,
                              *args, **kwargs)
@@ -580,7 +581,7 @@ def load_low_bit(cls, pretrained_model_name_or_path: str, *model_args, **kwargs)
             with torch.no_grad():
                 optimize_llm_pre(model, qtype, mixed_precision,
                                  quantization_group_size=quantization_group_size,
-                                 load=bigdl_lcmu_enabled)
+                                 load=bigdl_lcmu_enabled, max_prompt_len=max_prompt_len)
                 cls.load_convert(qtype, model, quant_device, modules_to_not_convert,
                                  quantization_group_size, *model_args, **kwargs)
                 create_npu_kernels(llm)
@@ -804,7 +805,8 @@ def optimize_npu_model(cls, *args, **kwargs):
 
         with torch.no_grad():
             optimize_llm_pre(model, qtype, mixed_precision,
-                             quantization_group_size=quantization_group_size)
+                             quantization_group_size=quantization_group_size,
+                             max_prompt_len=max_prompt_len)
             cls.load_convert_fp16(qtype, model.encoder, "cpu", modules_to_not_convert,
                                   quantization_group_size)
             create_npu_kernels(model.encoder)
diff --git a/python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py b/python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py
index 64d6f30b160..12c53fc3701 100644
--- a/python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py
+++ b/python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py
@@ -31,7 +31,7 @@ def convert_forward(m, target_m, new_forward):
 
 
 def optimize_llm_pre(model: torch.nn.Module, qtype, mixed_precision,
-                     quantization_group_size=0, load=False):
+                     quantization_group_size=0, load=False, max_prompt_len=512):
     if model.config.model_type == "baichuan":
         # process NormHead module in Baichuan2 7B
         if hasattr(model, 'lm_head') and model.lm_head is not None:
@@ -48,6 +48,13 @@ def optimize_llm_pre(model: torch.nn.Module, qtype, mixed_precision,
 
     cpu_lm_head = os.environ.get("IPEX_LLM_CPU_LM_HEAD", "0") != "0"
 
+    # workaround for long input performance of llama3.2-3b and glm-edge-4b CW
+    if os.environ.get("IPEX_LLM_NPU_DISABLE_COMPILE_OPT") is None:
+        disable_compile_opt = model.config.model_type == "llama" and \
+            model.config.hidden_size == 3072 and max_prompt_len >= 1920 and \
+            quantization_group_size == 0
+        os.environ["IPEX_LLM_NPU_DISABLE_COMPILE_OPT"] = "1" if disable_compile_opt else "0"
+
     # workaround for MiniCPM-2B
     if model.config.model_type == "minicpm" and model.config.num_hidden_layers == 40:
         # 73440 is vocab_size of MiniCPM-1B