Skip to content

model_base

This module contains the base Model class

Model

Parent model class.

Source code in happy_vllm/model/model_base.py 
 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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
class Model:
    """Parent model class.
    """

    def __init__(self, **kwargs):
        '''Init. model class'''
        self._model = None
        self._tokenizer = None
        self._model_conf = None
        self._model_explainer = None
        self.openai_serving_chat = None
        self.openai_serving_completion = None
        self.openai_serving_tokenization = None
        self._loaded = False
        self.app_name = kwargs.get('app_name', "happy_vllm")

    def is_model_loaded(self):
        """return the state of the model"""
        return self._loaded

    async def loading(self, async_engine_client: MQLLMEngineClient, args: Namespace, **kwargs):
        """load the model"""
        await self._load_model(async_engine_client, args, **kwargs)
        self._loaded = True
        if args.with_launch_arguments:
            self.launch_arguments = vars(args)
        else:
            self.launch_arguments = {}

    async def _load_model(self, async_engine_client: MQLLMEngineClient, args: Namespace, **kwargs) -> None:
        """Load a model from a file

        Returns:
            Tuple[Any, dict]: A tuple containing the model and a dict of metadata about it.
        """

        self._model_conf = {'model_name': args.model_name}

        logger.info(f"Loading the model from {args.model}")
        if args.model_name != "TEST MODEL":
            self._model = async_engine_client
            model_config = await self._model.get_model_config()
            # Define the tokenizer differently if we have an AsyncLLMEngine
            if isinstance(self._model, AsyncLLMEngine):
                tokenizer_tmp = self._model.engine.tokenizer
            else:
                tokenizer_tmp = self._model.tokenizer
            if isinstance(tokenizer_tmp, TokenizerGroup): # type: ignore
                self._tokenizer = tokenizer_tmp.tokenizer # type: ignore
            else:
                self._tokenizer = tokenizer_tmp # type: ignore
            self._tokenizer_lmformatenforcer = build_token_enforcer_tokenizer_data(self._tokenizer)
            self.max_model_len = model_config.max_model_len # type: ignore
            # To take into account Mistral tokenizers
            try:
                self.original_truncation_side = self._tokenizer.truncation_side # type: ignore
            except:
                self.original_truncation_side = "left"
            if args.disable_log_requests:
                request_logger = None
            else:
                request_logger = RequestLogger(max_log_len=args.max_log_len)
            served_model_names = [args.model_name]
            base_model_paths = [
                BaseModelPath(name=name, model_path=args.model)
                for name in served_model_names
            ]
            resolved_chat_template = load_chat_template(args.chat_template)
            logger.info("Using supplied chat template:\n%s", resolved_chat_template)
            self.openai_serving_chat = OpenAIServingChat(cast(AsyncLLMEngine,self._model), model_config, base_model_paths,
                                                        args.response_role,
                                                        lora_modules=args.lora_modules,
                                                        prompt_adapters=args.prompt_adapters,
                                                        request_logger=request_logger,
                                                        chat_template=resolved_chat_template,
                                                        chat_template_content_format=args.chat_template_content_format,
                                                        return_tokens_as_token_ids=args.return_tokens_as_token_ids,
                                                        enable_auto_tools=args.enable_auto_tool_choice,
                                                        tool_parser=args.tool_call_parser,
                                                        enable_prompt_tokens_details=args.enable_prompt_tokens_details)
            self.openai_serving_completion = OpenAIServingCompletion(cast(AsyncLLMEngine,self._model), model_config, base_model_paths, 
                                                                    lora_modules=args.lora_modules,
                                                                    prompt_adapters=args.prompt_adapters,
                                                                    request_logger=request_logger,
                                                                    return_tokens_as_token_ids=args.return_tokens_as_token_ids)
            self.openai_serving_tokenization  = OpenAIServingTokenization(cast(AsyncLLMEngine,self._model), model_config, base_model_paths,
                                                                        lora_modules=args.lora_modules,
                                                                        request_logger=request_logger,
                                                                        chat_template=resolved_chat_template,
                                                                        chat_template_content_format=args.chat_template_content_format)

        # For test purpose
        else:
            self.max_model_len = 2048
            self.original_truncation_side = 'right'
            self._tokenizer = AutoTokenizer.from_pretrained(utils.TEST_TOKENIZER_NAME,
                                                     cache_dir=os.environ["TEST_MODELS_DIR"], truncation_side=self.original_truncation_side)
            self._tokenizer_lmformatenforcer = build_token_enforcer_tokenizer_data(self._tokenizer)
            self._model = MockModel(self._tokenizer, self.app_name)
            self.openai_serving_tokenization = MockOpenAIServingTokenization(self._tokenizer)
        logger.info(f"Model loaded")

    def tokenize(self, text: str) -> List[int]:
        """Tokenizes a text

        Args:
            text (str) : The text to tokenize

        Returns:
            list : The list of token ids
        """
        return list(utils.proper_tokenization(self._tokenizer, text))

    def split_text(self, text: str, num_tokens_in_chunk: int = 200, separators: Union[list, None] = None) -> List[str]:
        '''Splits a text in small texts containing at least num_tokens_in_chunk tokens and ending by a separator. note that the `separators` 
        used are the tokenization of the strings and not the strings themselves (which explains why we must for example 
        specify ' .' and '.' as two separate separators) 

        Args:
            text (str) : The text to split

        Kwargs:
            num_tokens_in_chunk (int) : The minimal number of tokens in the chunk
            separators (list) : The separators marking the end of a sentence

        Returns:
            A list of strings each string containing at least num_tokens_in_chunk tokens and ending by a separator
        '''
        if separators is None:
            separators = [".", "!", "?", "|", " .", " !", " ?", " |"]
        separators_tokens_ids = set()
        for separator in separators:
            separators_tokens_ids.add(utils.proper_tokenization(self._tokenizer, separator))
        tokens = list(utils.proper_tokenization(self._tokenizer, text))
        indices_separators = []
        for separator_tokens_ids in separators_tokens_ids:
            indices_separators += find_indices_sub_list_in_list(tokens, list(separator_tokens_ids))
        indices_separators.sort()

        chunks = []
        index_beginning_chunk = 0
        current_used_separator = 0
        while current_used_separator < len(indices_separators):
            index_current_used_separator = indices_separators[current_used_separator]
            if index_current_used_separator +1 - index_beginning_chunk >= num_tokens_in_chunk:
                chunks.append(tokens[index_beginning_chunk:index_current_used_separator + 1])
                index_beginning_chunk = index_current_used_separator + 1
            current_used_separator += 1
        chunks.append(tokens[index_beginning_chunk:])
        chunks_decoded = [utils.proper_decode(self._tokenizer, chunk) for chunk in chunks]
        chunks_decoded = [element for element in chunks_decoded if element!= ""]
        return chunks_decoded

    def extract_text_outside_truncation(self, text: str, truncation_side: Union[str, None] = None, max_length: Union[int, None] = None) -> str:
        """Extracts the part of the prompt not kept after truncation, which will not be infered by the model.
        First, we tokenize the prompt while applying truncation.
        We obtain a list of sequences of token ids padded, which are outside the truncation.
        Then we decode this list of tensors of token IDs containing special tokens to a string.

        Args:
            text (str) : The text we want to parse
            truncation_side (str) : The side of the truncation
            max_length (int) : The length above which the text will be truncated

        Returns:
            The part of the text which will be dropped by the truncation (str)
        """
        if max_length is None:
            max_length = self.max_model_len
        if truncation_side is None:
            truncation_side = self.original_truncation_side
        self._tokenizer.truncation_side = truncation_side
        list_tokens = self._tokenizer(text, truncation=True, add_special_tokens=False, max_length=max_length, return_overflowing_tokens=True)['input_ids']
        if len(list_tokens) <= 1:
            return ''
        not_truncated = list_tokens[0]
        truncated_tmp = list_tokens[1:]
        if self._tokenizer.truncation_side == 'left':
            truncated_tmp.reverse()
        truncated = []
        for truncated_tokens in truncated_tmp:
            truncated += truncated_tokens
        truncated_str = self._tokenizer.decode(truncated)
        self._tokenizer.truncation_side = self.original_truncation_side
        return truncated_str

__init__(**kwargs)

Init. model class

Source code in happy_vllm/model/model_base.py 
51
52
53
54
55
56
57
58
59
60
61
def __init__(self, **kwargs):
    '''Init. model class'''
    self._model = None
    self._tokenizer = None
    self._model_conf = None
    self._model_explainer = None
    self.openai_serving_chat = None
    self.openai_serving_completion = None
    self.openai_serving_tokenization = None
    self._loaded = False
    self.app_name = kwargs.get('app_name', "happy_vllm")

extract_text_outside_truncation(text, truncation_side=None, max_length=None)

Extracts the part of the prompt not kept after truncation, which will not be infered by the model. First, we tokenize the prompt while applying truncation. We obtain a list of sequences of token ids padded, which are outside the truncation. Then we decode this list of tensors of token IDs containing special tokens to a string.

Parameters:

Name Type Description Default
text str)

The text we want to parse

 required
truncation_side str)

The side of the truncation

 None
max_length int)

The length above which the text will be truncated

 None

Returns:

Type Description
str

The part of the text which will be dropped by the truncation (str)
Source code in happy_vllm/model/model_base.py 
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
def extract_text_outside_truncation(self, text: str, truncation_side: Union[str, None] = None, max_length: Union[int, None] = None) -> str:
    """Extracts the part of the prompt not kept after truncation, which will not be infered by the model.
    First, we tokenize the prompt while applying truncation.
    We obtain a list of sequences of token ids padded, which are outside the truncation.
    Then we decode this list of tensors of token IDs containing special tokens to a string.

    Args:
        text (str) : The text we want to parse
        truncation_side (str) : The side of the truncation
        max_length (int) : The length above which the text will be truncated

    Returns:
        The part of the text which will be dropped by the truncation (str)
    """
    if max_length is None:
        max_length = self.max_model_len
    if truncation_side is None:
        truncation_side = self.original_truncation_side
    self._tokenizer.truncation_side = truncation_side
    list_tokens = self._tokenizer(text, truncation=True, add_special_tokens=False, max_length=max_length, return_overflowing_tokens=True)['input_ids']
    if len(list_tokens) <= 1:
        return ''
    not_truncated = list_tokens[0]
    truncated_tmp = list_tokens[1:]
    if self._tokenizer.truncation_side == 'left':
        truncated_tmp.reverse()
    truncated = []
    for truncated_tokens in truncated_tmp:
        truncated += truncated_tokens
    truncated_str = self._tokenizer.decode(truncated)
    self._tokenizer.truncation_side = self.original_truncation_side
    return truncated_str

is_model_loaded()

return the state of the model

Source code in happy_vllm/model/model_base.py 
63
64
65
def is_model_loaded(self):
    """return the state of the model"""
    return self._loaded

loading(async_engine_client, args, **kwargs) async

load the model

Source code in happy_vllm/model/model_base.py 
67
68
69
70
71
72
73
74
async def loading(self, async_engine_client: MQLLMEngineClient, args: Namespace, **kwargs):
    """load the model"""
    await self._load_model(async_engine_client, args, **kwargs)
    self._loaded = True
    if args.with_launch_arguments:
        self.launch_arguments = vars(args)
    else:
        self.launch_arguments = {}

split_text(text, num_tokens_in_chunk=200, separators=None)

Splits a text in small texts containing at least num_tokens_in_chunk tokens and ending by a separator. note that the separators used are the tokenization of the strings and not the strings themselves (which explains why we must for example specify ' .' and '.' as two separate separators)

Parameters:

Name Type Description Default
text str)

The text to split

 required
Kwargs

num_tokens_in_chunk (int) : The minimal number of tokens in the chunk separators (list) : The separators marking the end of a sentence

Returns:

Type Description
List[str]

A list of strings each string containing at least num_tokens_in_chunk tokens and ending by a separator
Source code in happy_vllm/model/model_base.py 
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
def split_text(self, text: str, num_tokens_in_chunk: int = 200, separators: Union[list, None] = None) -> List[str]:
    '''Splits a text in small texts containing at least num_tokens_in_chunk tokens and ending by a separator. note that the `separators` 
    used are the tokenization of the strings and not the strings themselves (which explains why we must for example 
    specify ' .' and '.' as two separate separators) 

    Args:
        text (str) : The text to split

    Kwargs:
        num_tokens_in_chunk (int) : The minimal number of tokens in the chunk
        separators (list) : The separators marking the end of a sentence

    Returns:
        A list of strings each string containing at least num_tokens_in_chunk tokens and ending by a separator
    '''
    if separators is None:
        separators = [".", "!", "?", "|", " .", " !", " ?", " |"]
    separators_tokens_ids = set()
    for separator in separators:
        separators_tokens_ids.add(utils.proper_tokenization(self._tokenizer, separator))
    tokens = list(utils.proper_tokenization(self._tokenizer, text))
    indices_separators = []
    for separator_tokens_ids in separators_tokens_ids:
        indices_separators += find_indices_sub_list_in_list(tokens, list(separator_tokens_ids))
    indices_separators.sort()

    chunks = []
    index_beginning_chunk = 0
    current_used_separator = 0
    while current_used_separator < len(indices_separators):
        index_current_used_separator = indices_separators[current_used_separator]
        if index_current_used_separator +1 - index_beginning_chunk >= num_tokens_in_chunk:
            chunks.append(tokens[index_beginning_chunk:index_current_used_separator + 1])
            index_beginning_chunk = index_current_used_separator + 1
        current_used_separator += 1
    chunks.append(tokens[index_beginning_chunk:])
    chunks_decoded = [utils.proper_decode(self._tokenizer, chunk) for chunk in chunks]
    chunks_decoded = [element for element in chunks_decoded if element!= ""]
    return chunks_decoded

tokenize(text)

Tokenizes a text

Parameters:

Name Type Description Default
text str)

The text to tokenize

 required

Returns:

Name Type Description
list List[int]

The list of token ids
Source code in happy_vllm/model/model_base.py 
149
150
151
152
153
154
155
156
157
158
def tokenize(self, text: str) -> List[int]:
    """Tokenizes a text

    Args:
        text (str) : The text to tokenize

    Returns:
        list : The list of token ids
    """
    return list(utils.proper_tokenization(self._tokenizer, text))

find_indices_sub_list_in_list(big_list, sub_list)

Find the indices of the presence of a sub list in a bigger list. For example if big_list = [3, 4, 1, 2, 3, 4, 5, 6, 3, 4] and sub_list = [3, 4], the result will be [1, 5, 9]

Parameters:

Name Type Description Default
big_list list)

The list in which we want to find the sub_list

 required
sub_list list

The list we want the indices of in the big_list

 required

Returns:

Name Type Description
list list

The list of indices of where the sub_list is in the big_list
Source code in happy_vllm/model/model_base.py 
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
def find_indices_sub_list_in_list(big_list: list, sub_list: list) -> list:
    """Find the indices of the presence of a sub list in a bigger list. For example
    if big_list = [3, 4, 1, 2, 3, 4, 5, 6, 3, 4] and sub_list = [3, 4],
    the result will be [1, 5, 9]

    Args:
        big_list (list) : The list in which we want to find the sub_list
        sub_list (list): The list we want the indices of in the big_list

    Returns:
        list : The list of indices of where the sub_list is in the big_list 
    """
    len_sub_list = len(sub_list)
    indices = []
    for index in range(len(big_list)):
        if big_list[index - len_sub_list + 1: index + 1] == sub_list:
            indices.append(index)
    return indices