File size: 13,847 Bytes
e679d69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
import copy
import logging
from typing import Dict, List, Optional, Union

from lagent.schema import ModelStatusCode
from .base_api import APITemplateParser
from .base_llm import BaseLLM

logger = logging.getLogger(__name__)


class HFTransformer(BaseLLM):
    """Model wrapper around HuggingFace general models.

    Adapted from Internlm (https://github.com/InternLM/InternLM/blob/main/
        chat/web_demo.py)

    Args:
        path (str): The name or path to HuggingFace's model.
        tokenizer_path (str): The path to the tokenizer. Defaults to None.
        tokenizer_kwargs (dict): Keyword arguments for the tokenizer.
            Defaults to {}.
        tokenizer_only (bool): If True, only the tokenizer will be initialized.
            Defaults to False.
        model_kwargs (dict): Keyword arguments for the model, used in loader.
            Defaults to dict(device_map='auto').
        meta_template (Dict, optional): The model's meta prompt
            template if needed, in case the requirement of injecting or
            wrapping of any meta instructions.
    """

    def __init__(self,
                 path: str,
                 tokenizer_path: Optional[str] = None,
                 tokenizer_kwargs: dict = dict(),
                 tokenizer_only: bool = False,
                 model_kwargs: dict = dict(device_map='auto'),
                 meta_template: Optional[Dict] = None,
                 stop_words_id: Union[List[int], int] = None,
                 **kwargs):
        super().__init__(
            path=path,
            tokenizer_only=tokenizer_only,
            meta_template=meta_template,
            **kwargs)
        if isinstance(stop_words_id, int):
            stop_words_id = [stop_words_id]
        self.gen_params.update(stop_words_id=stop_words_id)
        if self.gen_params['stop_words'] is not None and \
                self.gen_params['stop_words_id'] is not None:
            logger.warning('Both stop_words and stop_words_id are specified,'
                           'only stop_words_id will be used.')

        self._load_tokenizer(
            path=path,
            tokenizer_path=tokenizer_path,
            tokenizer_kwargs=tokenizer_kwargs)
        if not tokenizer_only:
            self._load_model(path=path, model_kwargs=model_kwargs)

        from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList  # noqa: E501
        self.logits_processor = LogitsProcessorList()
        self.stopping_criteria = StoppingCriteriaList()
        self.prefix_allowed_tokens_fn = None

        stop_words_id = []
        if self.gen_params.get('stop_words_id'):
            stop_words_id = self.gen_params.get('stop_words_id')
        elif self.gen_params.get('stop_words'):
            for sw in self.gen_params.get('stop_words'):
                stop_words_id.append(self.tokenizer(sw)['input_ids'][-1])
        self.additional_eos_token_id = stop_words_id

    def _load_tokenizer(self, path: str, tokenizer_path: Optional[str],
                        tokenizer_kwargs: dict):
        from transformers import AutoTokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(
            tokenizer_path if tokenizer_path else path,
            trust_remote_code=True,
            **tokenizer_kwargs)

        if self.tokenizer.pad_token_id is None:
            if self.tokenizer.eos_token is not None:
                logger.warning(
                    f'Using eos_token_id {self.tokenizer.eos_token} '
                    'as pad_token_id.')
                self.tokenizer.pad_token = self.tokenizer.eos_token
            else:
                from transformers.generation import GenerationConfig
                self.gcfg = GenerationConfig.from_pretrained(path)

                if self.gcfg.pad_token_id is not None:
                    logger.warning(
                        f'Using pad_token_id {self.gcfg.pad_token_id} '
                        'as pad_token_id.')
                    self.tokenizer.pad_token_id = self.gcfg.pad_token_id
                else:
                    raise ValueError(
                        'pad_token_id is not set for this tokenizer. Try to '
                        'set pad_token_id via passing '
                        '`pad_token_id={PAD_TOKEN_ID}` in model_cfg.')

    def _load_model(self, path: str, model_kwargs: dict):
        import torch
        from transformers import AutoModel
        model_kwargs.setdefault('torch_dtype', torch.float16)
        self.model = AutoModel.from_pretrained(
            path, trust_remote_code=True, **model_kwargs)
        self.model.eval()

    def tokenize(self, inputs: str):
        assert isinstance(inputs, str)
        inputs = self.tokenizer(
            inputs, return_tensors='pt', return_length=True)
        return inputs['input_ids'].tolist()

    def generate(
        self,
        inputs: Union[str, List[str]],
        do_sample: bool = True,
        **kwargs,
    ):
        """Return the chat completions in non-stream mode.

        Args:
            inputs (Union[str, List[str]]): input texts to be completed.
            do_sample (bool): do sampling if enabled
        Returns:
            (a list of/batched) text/chat completion
        """
        for status, chunk, _ in self.stream_generate(inputs, do_sample,
                                                     **kwargs):
            response = chunk
        return response

    def stream_generate(
        self,
        inputs: List[str],
        do_sample: bool = True,
        **kwargs,
    ):
        """Return the chat completions in stream mode.

        Args:
            inputs (Union[str, List[str]]): input texts to be completed.
            do_sample (bool): do sampling if enabled
        Returns:
            tuple(Status, str, int): status, text/chat completion,
            generated token number
        """
        import torch
        from torch import nn
        with torch.no_grad():
            batched = True
            if isinstance(inputs, str):
                inputs = [inputs]
                batched = False
            inputs = self.tokenizer(
                inputs, padding=True, return_tensors='pt', return_length=True)
            input_length = inputs['length']
            for k, v in inputs.items():
                inputs[k] = v.cuda()
            input_ids = inputs['input_ids']
            attention_mask = inputs['attention_mask']
            batch_size = input_ids.shape[0]
            input_ids_seq_length = input_ids.shape[-1]
            generation_config = self.model.generation_config
            generation_config = copy.deepcopy(generation_config)
            new_gen_params = self.update_gen_params(**kwargs)
            generation_config.update(**new_gen_params)
            generation_config.update(**kwargs)
            model_kwargs = generation_config.to_dict()
            model_kwargs['attention_mask'] = attention_mask
            _, eos_token_id = (  # noqa: F841  # pylint: disable=W0612
                generation_config.bos_token_id,
                generation_config.eos_token_id,
            )
            if eos_token_id is None:
                if self.gcfg.eos_token_id is not None:
                    eos_token_id = self.gcfg.eos_token_id
                else:
                    eos_token_id = []
            if isinstance(eos_token_id, int):
                eos_token_id = [eos_token_id]
            if self.additional_eos_token_id is not None:
                eos_token_id.extend(self.additional_eos_token_id)
            eos_token_id_tensor = torch.tensor(eos_token_id).to(
                input_ids.device) if eos_token_id is not None else None
            generation_config.max_length = (
                generation_config.max_new_tokens + input_ids_seq_length)
            # Set generation parameters if not already defined
            logits_processor = self.logits_processor
            stopping_criteria = self.stopping_criteria

            logits_processor = self.model._get_logits_processor(
                generation_config=generation_config,
                input_ids_seq_length=input_ids_seq_length,
                encoder_input_ids=input_ids,
                prefix_allowed_tokens_fn=self.prefix_allowed_tokens_fn,
                logits_processor=logits_processor,
            )

            stopping_criteria = self.model._get_stopping_criteria(
                generation_config=generation_config,
                stopping_criteria=stopping_criteria)
            logits_warper = self.model._get_logits_warper(generation_config)

            unfinished_sequences = input_ids.new(batch_size).fill_(1)
            scores = None
            while True:
                model_inputs = self.model.prepare_inputs_for_generation(
                    input_ids, **model_kwargs)
                # forward pass to get next token
                outputs = self.model(
                    **model_inputs,
                    return_dict=True,
                    output_attentions=False,
                    output_hidden_states=False,
                )

                next_token_logits = outputs.logits[:, -1, :]

                # pre-process distribution
                next_token_scores = logits_processor(input_ids,
                                                     next_token_logits)
                next_token_scores = logits_warper(input_ids, next_token_scores)

                # sample
                probs = nn.functional.softmax(next_token_scores, dim=-1)
                if do_sample:
                    next_tokens = torch.multinomial(
                        probs, num_samples=1).squeeze(1)
                else:
                    next_tokens = torch.argmax(probs, dim=-1)

                # update generated ids, model inputs,
                # and length for next step
                input_ids = torch.cat([input_ids, next_tokens[:, None]],
                                      dim=-1)
                model_kwargs = self.model._update_model_kwargs_for_generation(  # noqa: E501
                    outputs,
                    model_kwargs,
                    is_encoder_decoder=False)
                unfinished_sequences = unfinished_sequences.mul(
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(
                        eos_token_id_tensor.unsqueeze(1)).prod(dim=0))
                output_token_ids = input_ids.cpu().tolist()
                for i in range(len(output_token_ids)):
                    output_token_ids[i] = output_token_ids[i][:][
                        input_length[i]:]
                    # Find the first occurrence of
                    # an EOS token in the sequence
                    first_eos_idx = next(
                        (idx
                         for idx, token_id in enumerate(output_token_ids[i])
                         if token_id in eos_token_id), None)
                    # If an EOS token is found, only the previous
                    # part of it is retained
                    if first_eos_idx is not None:
                        output_token_ids[i] = output_token_ids[
                            i][:first_eos_idx]

                response = self.tokenizer.batch_decode(output_token_ids)
                # print(response)
                if not batched:
                    response = response[0]
                yield ModelStatusCode.STREAM_ING, response, None
                # stop when each sentence is finished,
                # or if we exceed the maximum length
                if (unfinished_sequences.max() == 0
                        or stopping_criteria(input_ids, scores)):
                    break
            yield ModelStatusCode.END, response, None

    def stream_chat(
        self,
        inputs: List[dict],
        do_sample: bool = True,
        **kwargs,
    ):
        """Return the chat completions in stream mode.

        Args:
            inputs (List[dict]): input messages to be completed.
            do_sample (bool): do sampling if enabled
        Returns:
            the text/chat completion
        """
        prompt = self.template_parser(inputs)
        yield from self.stream_generate(prompt, do_sample, **kwargs)


class HFTransformerCasualLM(HFTransformer):

    def _load_model(self, path: str, model_kwargs: dict):
        import torch
        from transformers import AutoModelForCausalLM
        model_kwargs.setdefault('torch_dtype', torch.float16)
        self.model = AutoModelForCausalLM.from_pretrained(
            path, trust_remote_code=True, **model_kwargs)
        self.model.eval()


class HFTransformerChat(HFTransformerCasualLM):

    def __init__(self, template_parser=APITemplateParser, **kwargs):
        super().__init__(template_parser=template_parser, **kwargs)

    def chat(self,
             inputs: Union[List[dict], List[List[dict]]],
             do_sample: bool = True,
             **kwargs):
        """Return the chat completions in stream mode.

        Args:
            inputs (Union[List[dict], List[List[dict]]]): input messages to be completed.
            do_sample (bool): do sampling if enabled
        Returns:
            the text/chat completion
        """
        # handle batch inference with vanilla for loop
        if isinstance(inputs[0], list):
            resps = []
            for input in inputs:
                resps.append(self.chat(input, do_sample, **kwargs))
            return resps
        prompt = self.template_parser(inputs)
        query = prompt[-1]['content']
        history = prompt[:-1]
        try:
            response, history = self.model.chat(
                self.tokenizer, query, history=history)
        except Exception as e:
            # handle over-length input error
            logger.warning(str(e))
            response = ''
        return response