diff --git a/python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md b/python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md
index 246053c793a..53f47df7946 100644
--- a/python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md
+++ b/python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md
@@ -37,8 +37,8 @@ pip install timm torch==2.1.2 torchvision==0.16.2
 pip install BCEmbedding==0.1.5 transformers==4.40.0
 
 # [optional] for Speech_Paraformer-Large
-pip install -U funasr
-pip install modelscope torch==2.1.2 torchaudio==2.1.2
+pip install funasr==1.1.14
+pip install modelscope==1.20.1 torch==2.1.2 torchaudio==2.1.2
 ```
 
 ### 2. Runtime Configurations
diff --git a/python/llm/src/ipex_llm/transformers/npu_model.py b/python/llm/src/ipex_llm/transformers/npu_model.py
index 51dcb17ccae..850f9009ee5 100644
--- a/python/llm/src/ipex_llm/transformers/npu_model.py
+++ b/python/llm/src/ipex_llm/transformers/npu_model.py
@@ -162,11 +162,13 @@ def from_pretrained(cls, *args, **kwargs):
                 model = cls.HF_Model.from_pretrained(*args, **kwargs)
             else:
                 model = cls.HF_Model(*args, **kwargs)
-            model.config.update({"bigdl_lcmu_enabled": False})
+            if hasattr(model, "config"):
+                model.config.update({"bigdl_lcmu_enabled": False})
 
         logger.info(f"Converting model, it may takes up to several minutes ...")
 
-        model.config.update({"optimize_model": optimize_model})
+        if hasattr(model, "config"):
+            model.config.update({"optimize_model": optimize_model})
 
         if mock_device == "cpu":
             with torch.no_grad():
diff --git a/python/llm/src/ipex_llm/transformers/npu_models/paraformer_mp.py b/python/llm/src/ipex_llm/transformers/npu_models/paraformer_mp.py
index 8d72cec50e9..a315f6a635d 100644
--- a/python/llm/src/ipex_llm/transformers/npu_models/paraformer_mp.py
+++ b/python/llm/src/ipex_llm/transformers/npu_models/paraformer_mp.py
@@ -294,17 +294,17 @@ def forward(
             torch.Tensor: result
         """
         backend_cls = self.backend_cls_prefill
-        inputs = (x,
-                  masks,
-                  self.layer_norm_0_weight,
-                  self.layer_norm_0_bias,
-                  self.layer_norm_1_weight,
-                  self.layer_norm_1_bias,
-                  self.fsmn_weight,
-                  self.qkv_bias,
-                  self.out_bias,
-                  self.w1_bias,
-                  self.w2_bias,
+        inputs = (x.to(torch.float16),
+                  masks.to(torch.float16),
+                  self.layer_norm_0_weight.to(torch.float16),
+                  self.layer_norm_0_bias.to(torch.float16),
+                  self.layer_norm_1_weight.to(torch.float16),
+                  self.layer_norm_1_bias.to(torch.float16),
+                  self.fsmn_weight.to(torch.float16),
+                  self.qkv_bias.to(torch.float16),
+                  self.out_bias.to(torch.float16),
+                  self.w1_bias.to(torch.float16),
+                  self.w2_bias.to(torch.float16),
                   )
 
         outputs = run_model(
@@ -431,6 +431,8 @@ def forward(
         args = (xs_pad, masks)
         self.prefill_input_queue.put(args)
         xs_pad, masks = self.prefill_result_queue.get()
+        xs_pad = xs_pad.to(torch.float32)
+        masks = masks.to(torch.float32)
         return xs_pad, masks
 
     def shutdown(self):
@@ -639,7 +641,7 @@ def __init__(
     ):
         super().__init__()
 
-        self.do_print = True
+        self.do_print = do_print
 
         op_parameters = []
         for w in parameters: