Views
No views yet
transformers library:1from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
2import torch
3import os
4import signal
5import random
6import numpy as np
7import time
8from collections import Counter
9
10cpu_count = os.cpu_count()
11print(f"Number of CPU cores in the system: {cpu_count}")
12half_cpu_count = cpu_count // 2
13os.environ["MKL_NUM_THREADS"] = str(half_cpu_count)
14os.environ["OMP_NUM_THREADS"] = str(half_cpu_count)
15torch.set_num_threads(half_cpu_count)
16
17print(f"PyTorch threads: {torch.get_num_threads()}")
18print(f"MKL threads: {os.getenv('MKL_NUM_THREADS')}")
19print(f"OMP threads: {os.getenv('OMP_NUM_THREADS')}")
20
21# Load the model and tokenizer
22NEW_MODEL_ID = "huihui-ai/Huihui-LFM2-2.6B-Exp-abliterated"
23print(f"Load Model {NEW_MODEL_ID} ... ")
24
25model = AutoModelForCausalLM.from_pretrained(
26 NEW_MODEL_ID,
27 device_map="auto",
28 trust_remote_code=True,
29 torch_dtype=torch.bfloat16,
30)
31
32tokenizer = AutoTokenizer.from_pretrained(NEW_MODEL_ID, trust_remote_code=True)
33
34messages = []
35skip_prompt=True
36skip_special_tokens=True
37
38class CustomTextStreamer(TextStreamer):
39 def __init__(self, tokenizer, skip_prompt=True, skip_special_tokens=True):
40 super().__init__(tokenizer, skip_prompt=skip_prompt, skip_special_tokens=skip_special_tokens)
41 self.generated_text = ""
42 self.stop_flag = False
43 self.init_time = time.time() # Record initialization time
44 self.end_time = None # To store end time
45 self.first_token_time = None # To store first token generation time
46 self.token_count = 0 # To track total tokens
47
48 def on_finalized_text(self, text: str, stream_end: bool = False):
49 if self.first_token_time is None and text.strip(): # Set first token time on first non-empty text
50 self.first_token_time = time.time()
51 if stream_end:
52 self.end_time = time.time() # Record end time when streaming ends
53 self.generated_text += text
54 self.token_count += 1
55 print(text, end="", flush=True)
56 if stream_end:
57 self.end_time = time.time() # Record end time when streaming ends
58 if self.stop_flag:
59 raise StopIteration
60
61 def stop_generation(self):
62 self.stop_flag = True
63 self.end_time = time.time() # Record end time when generation is stopped
64
65 def get_metrics(self):
66 """Returns initialization time, first token time, first token latency, end time, total time, total tokens, and tokens per second."""
67 if self.end_time is None:
68 self.end_time = time.time() # Set end time if not already set
69 total_time = self.end_time - self.init_time # Total time from init to end
70 tokens_per_second = self.token_count / total_time if total_time > 0 else 0
71 first_token_latency = (self.first_token_time - self.init_time) if self.first_token_time is not None else None
72 metrics = {
73 "init_time": self.init_time,
74 "first_token_time": self.first_token_time,
75 "first_token_latency": first_token_latency,
76 "end_time": self.end_time,
77 "total_time": total_time, # Total time in seconds
78 "total_tokens": self.token_count,
79 "tokens_per_second": tokens_per_second
80 }
81 return metrics
82
83def generate_stream(model, tokenizer, messages, skip_prompt, skip_special_tokens, max_new_tokens):
84 input_ids = tokenizer.apply_chat_template(
85 messages,
86 tokenize=True,
87 add_generation_prompt=True,
88 return_dict=True,
89 return_tensors="pt",
90 ).to(model.device)
91
92 streamer = CustomTextStreamer(tokenizer, skip_prompt=skip_prompt, skip_special_tokens=skip_special_tokens)
93
94 def signal_handler(sig, frame):
95 streamer.stop_generation()
96 print("\n[Generation stopped by user with Ctrl+C]")
97
98 signal.signal(signal.SIGINT, signal_handler)
99
100 print("Response: ", end="", flush=True)
101 try:
102 generated_ids = model.generate(
103 **input_ids,
104 max_new_tokens=max_new_tokens,
105 do_sample=True,
106 temperature=0.3,
107 min_p=0.15,
108 repetition_penalty=1.05,
109 streamer=streamer,
110 )
111 del generated_ids
112 except StopIteration:
113 print("\n[Stopped by user]")
114
115 del input_ids
116 torch.cuda.empty_cache()
117 signal.signal(signal.SIGINT, signal.SIG_DFL)
118
119 return streamer.generated_text, streamer.stop_flag, streamer.get_metrics()
120
121while True:
122 print(f"skip_prompt: {skip_prompt}")
123 print(f"skip_special_tokens: {skip_special_tokens}")
124
125 user_input = input("User: ").strip()
126 if user_input.lower() == "/exit":
127 print("Exiting chat.")
128 break
129 if user_input.lower() == "/clear":
130 messages = []
131 print("Chat history cleared. Starting a new conversation.")
132 continue
133 if user_input.lower() == "/skip_prompt":
134 skip_prompt = not skip_prompt
135 continue
136 if user_input.lower() == "/skip_special_tokens":
137 skip_special_tokens = not skip_special_tokens
138 continue
139 if not user_input:
140 print("Input cannot be empty. Please enter something.")
141 continue
142
143 messages.append({"role": "user", "content": user_input})
144
145 response, stop_flag, metrics = generate_stream(model, tokenizer, messages, skip_prompt, skip_special_tokens, 40960)
146 print("\n\nMetrics:")
147 for key, value in metrics.items():
148 print(f" {key}: {value}")
149
150 print("", flush=True)
151 if stop_flag:
152 continue
153 messages.append({"role": "assistant", "content": response}) bc1qqnkhuchxw0zqjh2ku3lu4hq45hc6gy84uk70ge