Commit
·
8f0b71b
1
Parent(s):
b058d61
Create handler.py
Browse files- handler.py +123 -0
handler.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, List
|
2 |
+
import os
|
3 |
+
import logging
|
4 |
+
import re
|
5 |
+
import torch
|
6 |
+
import transformers
|
7 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
8 |
+
from collections import deque
|
9 |
+
|
10 |
+
logging.basicConfig(level=logging.DEBUG)
|
11 |
+
|
12 |
+
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
13 |
+
|
14 |
+
|
15 |
+
class EndpointHandler:
|
16 |
+
def __init__(self, path=""):
|
17 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
|
18 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
19 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
20 |
+
path,
|
21 |
+
return_dict=True,
|
22 |
+
device_map="auto",
|
23 |
+
load_in_8bit=True,
|
24 |
+
torch_dtype=dtype,
|
25 |
+
trust_remote_code=True,
|
26 |
+
)
|
27 |
+
generation_config = self.model.generation_config
|
28 |
+
generation_config.max_new_tokens = 140
|
29 |
+
generation_config.temperature = 0.7
|
30 |
+
generation_config.top_p = 0.7
|
31 |
+
generation_config.num_return_sequences = 1
|
32 |
+
generation_config.pad_token_id = self.tokenizer.eos_token_id
|
33 |
+
generation_config.eos_token_id = self.tokenizer.eos_token_id
|
34 |
+
generation_config.early_stopping = True
|
35 |
+
self.generate_config = generation_config
|
36 |
+
|
37 |
+
self.pipeline = transformers.pipeline(
|
38 |
+
"text-generation", model=self.model, tokenizer=self.tokenizer
|
39 |
+
)
|
40 |
+
|
41 |
+
def _ensure_token_limit(self, tokens: List[int]) -> List[int]:
|
42 |
+
MAX_TOKEN_COUNT = 1024
|
43 |
+
if len(tokens) > MAX_TOKEN_COUNT:
|
44 |
+
# Keep only the last 2048 tokens
|
45 |
+
return tokens[-MAX_TOKEN_COUNT:]
|
46 |
+
return tokens
|
47 |
+
|
48 |
+
def _extract_response(self, text: str) -> str:
|
49 |
+
# Check for the start of the bot's or assistant's response
|
50 |
+
bot_start = text.find("<bot> response:") + len("<bot> response:")
|
51 |
+
assistant_start = text.find("<assistant> response:") + len("<assistant> response:")
|
52 |
+
response_start = max(bot_start, assistant_start)
|
53 |
+
|
54 |
+
# If neither bot nor assistant start marker is found, set to the beginning of the text
|
55 |
+
if response_start == -1 or (assistant_start == len("<assistant> response:") and bot_start == len("<bot> response:")):
|
56 |
+
response_start = 0
|
57 |
+
|
58 |
+
# Extract everything after the bot's or assistant's start marker until the next "User:" content
|
59 |
+
user_response_start = text.find("User:", response_start)
|
60 |
+
if user_response_start != -1:
|
61 |
+
end_point = user_response_start
|
62 |
+
else:
|
63 |
+
end_point = len(text)
|
64 |
+
|
65 |
+
# Return only the bot's or assistant's response, removing "User:" content
|
66 |
+
bot_response = text[response_start:end_point].strip()
|
67 |
+
return bot_response
|
68 |
+
|
69 |
+
def _truncate_conversation(self, conversation: str, max_tokens: int = 512) -> str:
|
70 |
+
# Split the conversation into exchanges
|
71 |
+
exchanges = re.split(r'(?=User:|Assistant:)', conversation)
|
72 |
+
while len(exchanges) > 0:
|
73 |
+
tokenized_conv = self.tokenizer.encode(' '.join(exchanges), truncation=False)
|
74 |
+
if len(tokenized_conv) <= max_tokens:
|
75 |
+
return ' '.join(exchanges)
|
76 |
+
exchanges.pop(0) # Remove the oldest exchange
|
77 |
+
return "" # If all exchanges are removed, return an empty string.
|
78 |
+
|
79 |
+
|
80 |
+
def generate_response(self, user_prompt, additional_context=None):
|
81 |
+
if additional_context:
|
82 |
+
truncated_conversation = self._truncate_conversation(additional_context)
|
83 |
+
else:
|
84 |
+
truncated_conversation = ""
|
85 |
+
permanent_context = ("<context>: You are a life coaching bot with the goal of providing guidance, improving understanding, reducing suffering and improving life. Gain as much understanding of the user before providing guidance with detailed actionable steps.")
|
86 |
+
structured_prompt = f"{permanent_context}\n{truncated_conversation}\n<user>: {user_prompt}"
|
87 |
+
|
88 |
+
structured_prompt += "<bot> response:"
|
89 |
+
|
90 |
+
input_ids = self.tokenizer.encode(structured_prompt, return_tensors="pt")
|
91 |
+
stop_token_ids = [self.tokenizer.encode(token)[0] for token in ['<bot>', 'User ']] # assuming these tokens are single tokens in your tokenizer
|
92 |
+
|
93 |
+
max_length = 1024
|
94 |
+
outputs = input_ids
|
95 |
+
|
96 |
+
while len(outputs[0]) < max_length:
|
97 |
+
# Generate next token
|
98 |
+
next_token_logits = self.model(outputs).logits[:, -1, :]
|
99 |
+
next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
|
100 |
+
|
101 |
+
# Check if the token is in stop_tokens list
|
102 |
+
if any(token.item() in stop_token_ids for token in next_token):
|
103 |
+
break
|
104 |
+
|
105 |
+
# Append the next_token to the outputs
|
106 |
+
outputs = torch.cat([outputs, next_token], dim=-1)
|
107 |
+
|
108 |
+
response_text = self._extract_response(self.tokenizer.decode(outputs[0])).strip()
|
109 |
+
|
110 |
+
return response_text
|
111 |
+
|
112 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
113 |
+
try:
|
114 |
+
debug_info = "No debug info available."
|
115 |
+
user_prompt = data.get("inputs", data)
|
116 |
+
|
117 |
+
self.prev_user_message = user_prompt
|
118 |
+
response_text = self.generate_response(user_prompt)
|
119 |
+
|
120 |
+
return [{"generated_text": response_text, "debug_info": debug_info}]
|
121 |
+
except Exception as e:
|
122 |
+
logging.error(f"An error occurred in __call__ method: {e}")
|
123 |
+
return [{"generated_text": str(e), "debug_info": debug_info}]
|