Fix short prompt for IPEX_CPU speculative decoding cause no_attr error (#10783)

python/llm/example/CPU/Speculative-Decoding/baichuan2/speculative.py

@@ -70,6 +70,8 @@ if __name__ == '__main__':
         prompt = BAICHUAN_PROMPT_FORMAT.format(prompt=args.prompt)
         inputs = tokenizer(prompt, return_tensors='pt', padding=True)
         input_ids = inputs.input_ids.to(model.device)
+        actual_in_len = input_ids.shape[1]
+        print("actual input_ids length:" + str(actual_in_len))
         attention_mask = inputs.attention_mask.to(model.device)
 
         # warmup
@@ -90,7 +92,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
         
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/chatglm3/speculative.py

@@ -87,7 +87,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/llama2/speculative.py

@@ -16,6 +16,7 @@
 
 import torch
 from ipex_llm.transformers import AutoModel, AutoModelForCausalLM
+
 from transformers import LlamaTokenizer, AutoTokenizer
 import argparse
 import time
@@ -104,7 +105,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/mistral/speculative.py

@@ -97,7 +97,14 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")
+

python/llm/example/CPU/Speculative-Decoding/qwen/speculative.py

@@ -101,7 +101,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/starcoder/speculative.py

@@ -81,7 +81,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/vicuna/speculative.py

@@ -79,6 +79,8 @@ if __name__ == '__main__':
         prompt = Vicuna_PROMPT_FORMAT.format(prompt=args.prompt)
         inputs = tokenizer(prompt, return_tensors='pt', padding=True)
         input_ids = inputs.input_ids.to(model.device)
+        actual_in_len = input_ids.shape[1]
+        print("actual input_ids length:" + str(actual_in_len))
         attention_mask = inputs.attention_mask.to(model.device)
 
         # warmup
@@ -97,7 +99,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")

python/llm/example/CPU/Speculative-Decoding/ziya/speculative.py

@@ -81,7 +81,13 @@ if __name__ == '__main__':
         output_str = tokenizer.decode(output[0], skip_special_tokens=True)
         end = time.perf_counter()
 
-        print(output_str)
-        print(f"Tokens generated {model.n_token_generated}")
         print(f"E2E Generation time {(end - st):.4f}s")
-        print(f"First token latency {model.first_token_time:.4f}s")
+        print(output_str)
+
+        # When the IPEX_CPU optimized models recive short prompts(length < 256)
+        # it will use normal generate() and has not these attr
+        from ipex_llm.transformers.convert import get_enable_ipex
+        _enable_ipex = get_enable_ipex()
+        if not _enable_ipex or actual_in_len >= 256:
+            print(f"Tokens generated {model.n_token_generated}")
+            print(f"First token latency {model.first_token_time:.4f}s")