diff --git a/python/llm/dev/benchmark/all-in-one/README.md b/python/llm/dev/benchmark/all-in-one/README.md
index 24b22c7f..32d2f49b 100644
--- a/python/llm/dev/benchmark/all-in-one/README.md
+++ b/python/llm/dev/benchmark/all-in-one/README.md
@@ -20,6 +20,7 @@ test_api:
   - "transformer_int4"
   - "native_int4"
   - "optimize_model"
+  - "pytorch_autocast_bf16"
   # - "transformer_int4_gpu"  # on arc
   # - "optimize_model_gpu"  # on arc
 ```
diff --git a/python/llm/dev/benchmark/all-in-one/config.yaml b/python/llm/dev/benchmark/all-in-one/config.yaml
index 39662ad6..b89132ba 100644
--- a/python/llm/dev/benchmark/all-in-one/config.yaml
+++ b/python/llm/dev/benchmark/all-in-one/config.yaml
@@ -12,5 +12,6 @@ test_api:
   - "transformer_int4"
   - "native_int4"
   - "optimize_model"
+  - "pytorch_autocast_bf16"
   # - "transformer_int4_gpu"  # on arc
   # - "optimize_model_gpu"  # on arc
\ No newline at end of file
diff --git a/python/llm/dev/benchmark/all-in-one/run.py b/python/llm/dev/benchmark/all-in-one/run.py
index a985b233..6539fea1 100644
--- a/python/llm/dev/benchmark/all-in-one/run.py
+++ b/python/llm/dev/benchmark/all-in-one/run.py
@@ -45,6 +45,8 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
         result = run_transformer_int4_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials)
     elif test_api == 'optimize_model_gpu':
         result = run_optimize_model_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials)
+    elif test_api == 'pytorch_autocast_bf16':
+        result = run_pytorch_autocast_bf16(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials)
 
     for in_out_pair in in_out_pairs:
         results.append([repo_id,
@@ -106,7 +108,7 @@ def run_transformer_int4(repo_id,
                          warm_up,
                          num_trials):
     from bigdl.llm.transformers import AutoModel, AutoModelForCausalLM
-    from transformers import AutoTokenizer
+    from transformers import AutoTokenizer, LlamaTokenizer
 
     model_path = get_model_path(repo_id, local_model_hub)
     # Load model in 4 bit,
@@ -115,6 +117,18 @@ def run_transformer_int4(repo_id,
     if repo_id in ['THUDM/chatglm-6b', 'THUDM/chatglm2-6b']:
         model = AutoModel.from_pretrained(model_path, load_in_4bit=True, trust_remote_code=True, torch_dtype='auto')
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in ['meta-llama/Llama-2-70b-chat-hf']:
+        # Can be removed when issue https://github.com/analytics-zoo/nano/issues/563 is resolved.
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_4bit=True,
+                                                     trust_remote_code=True, optimize_model=False)
+        # Need to use LlamaTokenizer, reason please refer to issue: https://github.com/intel-analytics/BigDL/issues/8944
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in ['meta-llama/Llama-2-7b-chat-hf','meta-llama/Llama-2-13b-chat-hf',
+                     'meta-llama/Llama-2-70b-chat-hf','decapoda-research/llama-7b-hf',
+                     'decapoda-research/llama-65b-hf','lmsys/vicuna-7b-v1.5',
+                     'lmsys/vicuna-13b-v1.3','project-baize/merged-baize-30b']:
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_4bit=True, trust_remote_code=True)
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
     else:
         model = AutoModelForCausalLM.from_pretrained(model_path, load_in_4bit=True, trust_remote_code=True)
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
@@ -139,7 +153,7 @@ def run_transformer_int4(repo_id,
             result[in_out] = []
             for i in range(num_trials + warm_up):
                 st = time.perf_counter()
-                output_ids = model.generate(input_ids, do_sample=False, max_new_tokens=out_len)
+                output_ids = model.generate(input_ids, do_sample=False, max_new_tokens=out_len, use_cache=True)
                 end = time.perf_counter()
                 print("model generate cost: " + str(end - st))
                 output = tokenizer.batch_decode(output_ids)
@@ -148,6 +162,59 @@ def run_transformer_int4(repo_id,
                     result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time])
     return result
 
+def run_pytorch_autocast_bf16(repo_id,
+                         local_model_hub,
+                         in_out_pairs,
+                         warm_up,
+                         num_trials):
+    from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, LlamaTokenizer
+
+    model_path = get_model_path(repo_id, local_model_hub)
+    st = time.perf_counter()
+    if repo_id in ['THUDM/chatglm-6b', 'THUDM/chatglm2-6b']:
+        # TODO: need verify chatglm family run bf16.
+        model = AutoModel.from_pretrained(model_path, trust_remote_code=True).float()
+        #model = AutoModel.from_pretrained(model_path, trust_remote_code=True).bfloat()
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in ['meta-llama/Llama-2-7b-chat-hf','meta-llama/Llama-2-13b-chat-hf',
+                     'meta-llama/Llama-2-70b-chat-hf','decapoda-research/llama-7b-hf',
+                     'decapoda-research/llama-65b-hf','lmsys/vicuna-7b-v1.5',
+                     'lmsys/vicuna-13b-v1.3','project-baize/merged-baize-30b']:
+        model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+        # Need to use LlamaTokenizer, reason please refer to issue: https://github.com/intel-analytics/BigDL/issues/8944
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    else:
+        model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    end = time.perf_counter()
+    print(">> loading of model costs {}s".format(end - st))
+
+    model = BenchmarkWrapper(model)
+    result = {}
+    with torch.inference_mode(), torch.autocast("cpu"):
+        for in_out in in_out_pairs:
+            in_out_len = in_out.split("-")
+            in_len = int(in_out_len[0])
+            out_len = int(in_out_len[1])
+            input_str = open(f"prompt/{in_len}.txt", 'r').read()
+            # As different tokenizer has different encodings,
+            # slice the input_ids to ensure the prompt length is required length.
+            input_ids = tokenizer.encode(input_str, return_tensors="pt")
+            input_ids = input_ids[:, :in_len]
+            true_str = tokenizer.batch_decode(input_ids)[0]
+            input_ids = tokenizer.encode(true_str, return_tensors="pt")
+            result[in_out] = []
+            print("input tokens: {}".format(input_ids.shape[1]))
+            for i in range(num_trials + warm_up):
+                st = time.perf_counter()
+                output_ids = model.generate(input_ids, do_sample=False, max_new_tokens=out_len, use_cache=True)
+                end = time.perf_counter()
+                print("model generate cost: " + str(end - st))
+                output = tokenizer.batch_decode(output_ids)
+                print(output[0])
+                if i >= warm_up:
+                    result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time])
+    return result
 
 def run_optimize_model(repo_id,
                        local_model_hub,