Update streaming in npu examples (#12495)

* feat: add streaming

* Update readme accordingly

---------

Co-authored-by: Yuwen Hu <yuwen.hu@intel.com>

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md

@@ -136,6 +136,7 @@ Arguments info:
 - `--max-context-len MAX_CONTEXT_LEN`: Defines the maximum sequence length for both input and output tokens. It is default to be `1024`.
 - `--max-prompt-len MAX_PROMPT_LEN`: Defines the maximum number of tokens that the input prompt can contain. It is default to be `512`.
 - `--disable-transpose-value-cache`: Disable the optimization of transposing value cache.
+- `--disable-streaming`: Disable streaming mode of generation.
 - `--save-directory SAVE_DIRECTORY`: argument defining the path to save converted model. If it is a non-existing path, the original pretrained model specified by `REPO_ID_OR_MODEL_PATH` will be loaded, otherwise the lowbit model in `SAVE_DIRECTORY` will be loaded.
 
 ### Troubleshooting

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/baichuan2.py

@@ -20,7 +20,7 @@ import time
 import argparse
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, TextStreamer
 
 from transformers.utils import logging
 
@@ -56,6 +56,7 @@ if __name__ == "__main__":
     parser.add_argument("--max-context-len", type=int, default=1024)
     parser.add_argument("--max-prompt-len", type=int, default=512)
     parser.add_argument("--disable-transpose-value-cache", action="store_true", default=False)
+    parser.add_argument("--disable-streaming", action="store_true", default=False)
     parser.add_argument("--save-directory", type=str,
         required=True,
         help="The path of folder to save converted model, "
@@ -94,6 +95,10 @@ if __name__ == "__main__":
         )
         tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
+    if args.disable_streaming:
+        streamer = None
+    else:
+        streamer = TextStreamer(tokenizer=tokenizer, skip_special_tokens=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """
@@ -105,19 +110,19 @@ if __name__ == "__main__":
         for i in range(5):
             prompt = get_prompt(args.prompt, [], system_prompt=DEFAULT_SYSTEM_PROMPT)
             _input_ids = tokenizer.encode(prompt, return_tensors="pt")
+            print("-" * 20, "Input", "-" * 20)
             print("input length:", len(_input_ids[0]))
+            print(prompt)
+            print("-" * 20, "Output", "-" * 20)
             st = time.time()
             output = model.generate(
-                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict, streamer=streamer
             )
             end = time.time()
+            if args.disable_streaming:
+                output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+                print(output_str)
             print(f"Inference time: {end-st} s")
-            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
-            print("-" * 20, "Input", "-" * 20)
-            print(input_str)
-            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
-            print("-" * 20, "Output", "-" * 20)
-            print(output_str)
 
     print("-" * 80)
     print("done")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/llama2.py

@@ -20,7 +20,7 @@ import time
 import argparse
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, TextStreamer
 
 from transformers.utils import logging
 
@@ -56,6 +56,7 @@ if __name__ == "__main__":
     parser.add_argument("--max-context-len", type=int, default=1024)
     parser.add_argument("--max-prompt-len", type=int, default=512)
     parser.add_argument("--disable-transpose-value-cache", action="store_true", default=False)
+    parser.add_argument("--disable-streaming", action="store_true", default=False)
     parser.add_argument("--save-directory", type=str,
         required=True,
         help="The path of folder to save converted model, "
@@ -93,6 +94,10 @@ if __name__ == "__main__":
         )
         tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
+    if args.disable_streaming:
+        streamer = None
+    else:
+        streamer = TextStreamer(tokenizer=tokenizer, skip_special_tokens=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """
@@ -104,19 +109,19 @@ if __name__ == "__main__":
         for i in range(5):
             prompt = get_prompt(args.prompt, [], system_prompt=DEFAULT_SYSTEM_PROMPT)
             _input_ids = tokenizer.encode(prompt, return_tensors="pt")
+            print("-" * 20, "Input", "-" * 20)
             print("input length:", len(_input_ids[0]))
+            print(prompt)
+            print("-" * 20, "Output", "-" * 20)
             st = time.time()
             output = model.generate(
-                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict, streamer=streamer
             )
             end = time.time()
+            if args.disable_streaming:
+                output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+                print(output_str)
             print(f"Inference time: {end-st} s")
-            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
-            print("-" * 20, "Input", "-" * 20)
-            print(input_str)
-            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
-            print("-" * 20, "Output", "-" * 20)
-            print(output_str)
 
     print("-" * 80)
     print("done")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/llama3.py

@@ -20,7 +20,7 @@ import time
 import argparse
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, TextStreamer
 
 from transformers.utils import logging
 
@@ -57,6 +57,7 @@ if __name__ == "__main__":
     parser.add_argument("--max-context-len", type=int, default=1024)
     parser.add_argument("--max-prompt-len", type=int, default=512)
     parser.add_argument("--disable-transpose-value-cache", action="store_true", default=False)
+    parser.add_argument("--disable-streaming", action="store_true", default=False)
     parser.add_argument("--save-directory", type=str,
         required=True,
         help="The path of folder to save converted model, "
@@ -94,6 +95,10 @@ if __name__ == "__main__":
         )
         tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
+    if args.disable_streaming:
+        streamer = None
+    else:
+        streamer = TextStreamer(tokenizer=tokenizer, skip_special_tokens=True)
 
     DEFAULT_SYSTEM_PROMPT = """\
     """
@@ -105,19 +110,19 @@ if __name__ == "__main__":
         for i in range(5):
             prompt = get_prompt(args.prompt, [], system_prompt=DEFAULT_SYSTEM_PROMPT)
             _input_ids = tokenizer.encode(prompt, return_tensors="pt")
+            print("-" * 20, "Input", "-" * 20)
             print("input length:", len(_input_ids[0]))
+            print(prompt)
+            print("-" * 20, "Output", "-" * 20)
             st = time.time()
             output = model.generate(
-                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict, streamer=streamer
             )
             end = time.time()
+            if args.disable_streaming:
+                output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+                print(output_str)
             print(f"Inference time: {end-st} s")
-            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
-            print("-" * 20, "Input", "-" * 20)
-            print(input_str)
-            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
-            print("-" * 20, "Output", "-" * 20)
-            print(output_str)
 
     print("-" * 80)
     print("done")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/minicpm.py

@@ -20,7 +20,7 @@ import time
 import argparse
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, TextStreamer
 
 from transformers.utils import logging
 
@@ -43,6 +43,7 @@ if __name__ == "__main__":
     parser.add_argument("--max-context-len", type=int, default=1024)
     parser.add_argument("--max-prompt-len", type=int, default=512)
     parser.add_argument("--disable-transpose-value-cache", action="store_true", default=False)
+    parser.add_argument("--disable-streaming", action="store_true", default=False)
     parser.add_argument("--save-directory", type=str,
         required=True,
         help="The path of folder to save converted model, "
@@ -80,26 +81,32 @@ if __name__ == "__main__":
         )
         tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
+    if args.disable_streaming:
+        streamer = None
+    else:
+        streamer = TextStreamer(tokenizer=tokenizer, skip_special_tokens=True)
+
     print("-" * 80)
     print("done")
     with torch.inference_mode():
 
         print("finish to load")
         for i in range(5):
-            _input_ids = tokenizer.encode("<用户>{}<AI>".format(args.prompt), return_tensors="pt")
+            prompt = "<用户>{}<AI>".format(args.prompt)
+            _input_ids = tokenizer.encode(prompt, return_tensors="pt")
+            print("-" * 20, "Input", "-" * 20)
             print("input length:", len(_input_ids[0]))
+            print(prompt)
+            print("-" * 20, "Output", "-" * 20)
             st = time.time()
             output = model.generate(
-                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict, streamer=streamer
             )
             end = time.time()
+            if args.disable_streaming:
+                output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+                print(output_str)
             print(f"Inference time: {end-st} s")
-            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
-            print("-" * 20, "Input", "-" * 20)
-            print(input_str)
-            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
-            print("-" * 20, "Output", "-" * 20)
-            print(output_str)
 
     print("-" * 80)
     print("done")

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/qwen.py

@@ -20,7 +20,7 @@ import time
 import argparse
 
 from ipex_llm.transformers.npu_model import AutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, TextStreamer
 
 from transformers.utils import logging
 
@@ -45,6 +45,7 @@ if __name__ == "__main__":
     parser.add_argument("--quantization_group_size", type=int, default=0)
     parser.add_argument('--low-bit', type=str, default="sym_int4",
                         help='Load in low bit to use')
+    parser.add_argument("--disable-streaming", action="store_true", default=False)
     parser.add_argument("--disable-transpose-value-cache", action="store_true", default=False)
     parser.add_argument("--save-directory", type=str,
         required=True,
@@ -84,6 +85,10 @@ if __name__ == "__main__":
         )
         tokenizer = AutoTokenizer.from_pretrained(args.save_directory, trust_remote_code=True)        
 
+    if args.disable_streaming:
+        streamer = None
+    else:
+        streamer = TextStreamer(tokenizer=tokenizer, skip_special_tokens=True)
 
     print("-" * 80)
     print("done")
@@ -96,19 +101,19 @@ if __name__ == "__main__":
         print("finish to load")
         for i in range(3):
             _input_ids = tokenizer([text], return_tensors="pt").input_ids
+            print("-" * 20, "Input", "-" * 20)
             print("input length:", len(_input_ids[0]))
+            print(text)
+            print("-" * 20, "Output", "-" * 20)
             st = time.time()
             output = model.generate(
-                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict
+                _input_ids, num_beams=1, do_sample=False, max_new_tokens=args.n_predict, streamer=streamer
             )
             end = time.time()
+            if args.disable_streaming:
+                output_str = tokenizer.decode(output[0], skip_special_tokens=False)
+                print(output_str)
             print(f"Inference time: {end-st} s")
-            input_str = tokenizer.decode(_input_ids[0], skip_special_tokens=False)
-            print("-" * 20, "Input", "-" * 20)
-            print(input_str)
-            output_str = tokenizer.decode(output[0], skip_special_tokens=False)
-            print("-" * 20, "Output", "-" * 20)
-            print(output_str)
 
     print("-" * 80)
     print("done")