File size: 6,820 Bytes
9f341cc
 
 
3f608c6
3125c87
babcd78
3f608c6
 
bf8c5bd
 
 
 
 
 
3125c87
9f341cc
 
 
3125c87
9f341cc
3f608c6
6aa8b86
 
5dc8ebf
3f608c6
cd6b52a
3f608c6
 
 
 
 
 
 
9f341cc
 
 
 
 
eb00725
 
 
 
9f341cc
 
1b9f698
 
 
 
 
 
d2b20f2
9f341cc
 
e2b245b
403b8cf
1b9f698
2da6968
e2b245b
9f341cc
e820e51
 
9f341cc
 
 
 
 
 
2da6968
 
 
 
 
 
 
8ba223c
 
64645f0
eb00725
64645f0
 
 
8ba223c
fa3bfb0
bf8c5bd
1b9f698
 
 
 
 
 
 
 
 
395ee29
 
 
 
 
e820e51
e2b245b
9f341cc
 
 
8ba223c
403b8cf
9f341cc
 
 
e2b245b
 
 
d2b20f2
9f341cc
4ba2ca6
bf8c5bd
 
4ba2ca6
 
 
 
b96cef7
d2b20f2
9f341cc
 
 
babcd78
d2b20f2
9f341cc
d2b20f2
047008b
 
 
 
 
d2b20f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ba2ca6
d2b20f2
 
 
 
 
4ba2ca6
d2b20f2
 
 
 
4ba2ca6
 
bf8c5bd
4ba2ca6
d2b20f2
c95d47e
4ba2ca6
d2b20f2
 
 
 
c95d47e
d2b20f2
c95d47e
 
 
9f341cc
 
 
 
4ba2ca6
9f341cc
395ee29
d2b20f2
9f341cc
 
c95d47e
 
047008b
9f341cc
d2b20f2
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
import json
import re
import requests


from tclogger import logger
from transformers import AutoTokenizer

from constants.models import (
    MODEL_MAP,
    STOP_SEQUENCES_MAP,
    TOKEN_LIMIT_MAP,
    TOKEN_RESERVED,
)
from constants.envs import PROXIES
from messagers.message_outputer import OpenaiStreamOutputer


class HuggingfaceStreamer:
    def __init__(self, model: str):
        if model in MODEL_MAP.keys():
            self.model = model
        else:
            self.model = "mixtral-8x7b"
        self.model_fullname = MODEL_MAP[self.model]
        self.message_outputer = OpenaiStreamOutputer(model=self.model)

        if self.model == "gemma-7b":
            # this is not wrong, as repo `google/gemma-7b-it` is gated and must authenticate to access it
            # so I use mistral-7b as a fallback
            self.tokenizer = AutoTokenizer.from_pretrained(MODEL_MAP["mistral-7b"])
        else:
            self.tokenizer = AutoTokenizer.from_pretrained(self.model_fullname)

    def parse_line(self, line):
        line = line.decode("utf-8")
        line = re.sub(r"data:\s*", "", line)
        data = json.loads(line)
        try:
            content = data["token"]["text"]
        except:
            logger.err(data)
        return content

    def count_tokens(self, text):
        tokens = self.tokenizer.encode(text)
        token_count = len(tokens)
        logger.note(f"Prompt Token Count: {token_count}")
        return token_count

    def chat_response(
        self,
        prompt: str = None,
        temperature: float = 0.5,
        top_p: float = 0.95,
        max_new_tokens: int = None,
        api_key: str = None,
        use_cache: bool = False,
    ):
        # https://huggingface.co/docs/api-inference/detailed_parameters?code=curl
        # curl --proxy http://<server>:<port> https://api-inference.huggingface.co/models/<org>/<model_name> -X POST -d '{"inputs":"who are you?","parameters":{"max_new_token":64}}' -H 'Content-Type: application/json' -H 'Authorization: Bearer <HF_TOKEN>'
        self.request_url = (
            f"https://api-inference.huggingface.co/models/{self.model_fullname}"
        )
        self.request_headers = {
            "Content-Type": "application/json",
        }

        if api_key:
            logger.note(
                f"Using API Key: {api_key[:3]}{(len(api_key)-7)*'*'}{api_key[-4:]}"
            )
            self.request_headers["Authorization"] = f"Bearer {api_key}"

        if temperature is None or temperature < 0:
            temperature = 0.0
        # temperature must  0 < and < 1 for HF LLM models
        temperature = max(temperature, 0.01)
        temperature = min(temperature, 0.99)
        top_p = max(top_p, 0.01)
        top_p = min(top_p, 0.99)

        token_limit = int(
            TOKEN_LIMIT_MAP[self.model] - TOKEN_RESERVED - self.count_tokens(prompt)
        )
        if token_limit <= 0:
            raise ValueError("Prompt exceeded token limit!")

        if max_new_tokens is None or max_new_tokens <= 0:
            max_new_tokens = token_limit
        else:
            max_new_tokens = min(max_new_tokens, token_limit)

        # References:
        #   huggingface_hub/inference/_client.py:
        #     class InferenceClient > def text_generation()
        #   huggingface_hub/inference/_text_generation.py:
        #     class TextGenerationRequest > param `stream`
        # https://huggingface.co/docs/text-generation-inference/conceptual/streaming#streaming-with-curl
        # https://huggingface.co/docs/api-inference/detailed_parameters#text-generation-task
        self.request_body = {
            "inputs": prompt,
            "parameters": {
                "temperature": temperature,
                "top_p": top_p,
                "max_new_tokens": max_new_tokens,
                "return_full_text": False,
            },
            "options": {
                "use_cache": use_cache,
            },
            "stream": True,
        }

        if self.model in STOP_SEQUENCES_MAP.keys():
            self.stop_sequences = STOP_SEQUENCES_MAP[self.model]
        #     self.request_body["parameters"]["stop_sequences"] = [
        #         self.STOP_SEQUENCES[self.model]
        #     ]

        logger.back(self.request_url)
        stream_response = requests.post(
            self.request_url,
            headers=self.request_headers,
            json=self.request_body,
            proxies=PROXIES,
            stream=True,
        )
        status_code = stream_response.status_code
        if status_code == 200:
            logger.success(status_code)
        else:
            logger.err(status_code)

        return stream_response

    def chat_return_dict(self, stream_response):
        # https://platform.openai.com/docs/guides/text-generation/chat-completions-response-format
        final_output = self.message_outputer.default_data.copy()
        final_output["choices"] = [
            {
                "index": 0,
                "finish_reason": "stop",
                "message": {
                    "role": "assistant",
                    "content": "",
                },
            }
        ]
        logger.back(final_output)

        final_content = ""
        for line in stream_response.iter_lines():
            if not line:
                continue
            content = self.parse_line(line)

            if content.strip() == self.stop_sequences:
                logger.success("\n[Finished]")
                break
            else:
                logger.back(content, end="")
                final_content += content

        if self.model in STOP_SEQUENCES_MAP.keys():
            final_content = final_content.replace(self.stop_sequences, "")

        final_content = final_content.strip()
        final_output["choices"][0]["message"]["content"] = final_content
        return final_output

    def chat_return_generator(self, stream_response):
        is_finished = False
        line_count = 0
        for line in stream_response.iter_lines():
            if line:
                line_count += 1
            else:
                continue

            content = self.parse_line(line)

            if content.strip() == self.stop_sequences:
                content_type = "Finished"
                logger.success("\n[Finished]")
                is_finished = True
            else:
                content_type = "Completions"
                if line_count == 1:
                    content = content.lstrip()
                logger.back(content, end="")

            output = self.message_outputer.output(
                content=content, content_type=content_type
            )
            yield output

        if not is_finished:
            yield self.message_outputer.output(content="", content_type="Finished")