File size: 9,895 Bytes
b0c0df0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import warnings
from typing import List, Optional, Tuple, Union

import torch
from accelerate import Accelerator, DistributedType
from accelerate.state import AcceleratorState
from tqdm import tqdm
from transformers import AutoModel, AutoTokenizer

from lmms_eval import utils
from lmms_eval.api.instance import Instance
from lmms_eval.api.model import lmms
from lmms_eval.api.registry import register_model

warnings.filterwarnings("ignore")

from loguru import logger as eval_logger


@register_model("minicpm_v")
class MiniCPM_V(lmms):
    """
    MiniCPM_V Model
    """

    def __init__(
        self,
        pretrained: str = "openbmb/MiniCPM-V",
        device: Optional[str] = "cuda",
        dtype: Optional[Union[str, torch.dtype]] = torch.bfloat16,
        batch_size: Optional[Union[int, str]] = 1,
        trust_remote_code: Optional[bool] = True,
        **kwargs,
    ) -> None:
        super().__init__()
        # Do not use kwargs for now
        assert kwargs == {}, f"Unexpected kwargs: {kwargs}"

        accelerator = Accelerator()
        if accelerator.num_processes > 1:
            self._device = torch.device(f"cuda:{accelerator.local_process_index}")
        else:
            self._device = device
        self._model = AutoModel.from_pretrained(pretrained, trust_remote_code=trust_remote_code, torch_dtype=dtype, device_map=self._device).to(dtype)
        self._tokenizer = AutoTokenizer.from_pretrained(pretrained, trust_remote_code=trust_remote_code)
        self._config = self._model.config
        self.model.eval()
        self.model.tie_weights()
        self.batch_size_per_gpu = int(batch_size)
        if accelerator.num_processes > 1:
            assert accelerator.distributed_type in [DistributedType.FSDP, DistributedType.MULTI_GPU, DistributedType.DEEPSPEED], "Unsupported distributed type provided. Only DDP and FSDP are supported."
            # If you want to use DistributedType.DEEPSPEED, you have to run accelerate config before using the model
            # Also, you have to select zero stage 0 (equivalent to DDP) in order to make the prepare model works
            # I tried to set different parameters in the kwargs to let default zero 2 stage works, but it didn't work.
            if accelerator.distributed_type == DistributedType.DEEPSPEED:
                kwargs = {
                    "train_micro_batch_size_per_gpu": self.batch_size_per_gpu,
                    "train_batch_size": self.batch_size_per_gpu * accelerator.num_processes,
                }
                AcceleratorState().deepspeed_plugin.deepspeed_config_process(must_match=True, **kwargs)
                eval_logger.info("Detected that you are using DistributedType.DEEPSPEED. Make sure you run `accelerate config` and set zero stage to 0")
            if accelerator.distributed_type == DistributedType.FSDP or accelerator.distributed_type == DistributedType.DEEPSPEED:
                self._model = accelerator.prepare(self.model)
            else:
                self._model = accelerator.prepare_model(self.model, evaluation_mode=True)
            self.accelerator = accelerator
            if self.accelerator.is_local_main_process:
                eval_logger.info(f"Using {accelerator.num_processes} devices with data parallelism")
            self._rank = self.accelerator.local_process_index
            self._world_size = self.accelerator.num_processes
        else:
            self.model.to(self._device)
            self._rank = 0
            self._world_size = 1

    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

    @property
    def tokenizer(self):
        return self._tokenizer

    @property
    def model(self):
        # returns the model, unwrapping it if using Accelerate
        if hasattr(self, "accelerator"):
            return self.accelerator.unwrap_model(self._model)
        else:
            return self._model

    @property
    def eot_token_id(self):
        # we use EOT because end of *text* is more accurate for what we're doing than end of *sentence*
        return self.tokenizer.eos_token_id

    @property
    def max_length(self):
        return self._max_length

    @property
    def batch_size(self):
        return self.batch_size_per_gpu

    @property
    def device(self):
        return self._device

    @property
    def rank(self):
        return self._rank

    @property
    def world_size(self):
        return self._world_size

    def tok_encode(self, string: str, left_truncate_len=None, add_special_tokens=None) -> List[int]:
        """ """
        add_special_tokens = False if add_special_tokens is None else add_special_tokens
        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
        # left-truncate the encoded context to be at most `left_truncate_len` tokens long
        if left_truncate_len:
            encoding = encoding[-left_truncate_len:]
        return encoding

    def tok_decode(self, tokens):
        return self.tokenizer.decode(tokens)

    def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
        # TODO
        assert False, "We have not implemented this function for MiniCPM_V yet"

    def flatten(self, input):
        new_list = []
        for i in input:
            for j in i:
                new_list.append(j)
        return new_list

    def generate_until(self, requests: List[Instance]) -> List[str]:
        res = []

        def _collate(x):
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
            # - any OOMs will happen right away rather than near the end
            toks = self.tok_encode(x[0])
            return -len(toks), x[0]

        # we group requests by their generation_kwargs,
        # so that we don't try to execute e.g. greedy sampling and temp=0.8 sampling
        # in the same batch.
        re_ords = utils.Collator([reg.args for reg in requests], _collate, grouping=True)
        chunks = re_ords.get_batched(n=self.batch_size, batch_fn=None)
        num_iters = len(requests) // self.batch_size if len(requests) % self.batch_size == 0 else len(requests) // self.batch_size + 1
        pbar = tqdm(total=num_iters, disable=(self.rank != 0), desc="Model Responding")
        for chunk in chunks:
            contexts, all_gen_kwargs, doc_to_visual, doc_id, task, split = zip(*chunk)
            task = task[0]
            split = split[0]
            visuals = [doc_to_visual[0](self.task_dict[task][split][ids]) for ids in doc_id]
            visuals = self.flatten(visuals)
            # we assume all gen kwargs in the batch are the same
            # this is safe to assume because the `grouper` object ensures it.
            gen_kwargs = all_gen_kwargs[0]

            # Set default values for until and max_new_tokens
            until = [self.tok_decode(self.eot_token_id)]

            # Update values from gen_kwargs if present
            if "until" in gen_kwargs:
                until = gen_kwargs.pop("until")
                if isinstance(until, str):
                    until = [until]
                elif not isinstance(until, list):
                    raise ValueError(f"Expected `gen_kwargs['until']` to be of type Union[str,list] but got {type(until)}")
            assert self.batch_size_per_gpu == 1, "Do not support batch_size_per_gpu > 1 for now"
            assert len(visuals) == 1, "MiniCPM_V interface does not support bn_image > 1 for now"
            context = contexts[0]
            if "<image>" in context:
                # minicpm does not expect the <image> tag
                context = context.replace("<image>", "")
            msgs = [{"role": "user", "content": context}]

            gen_kwargs["image_sizes"] = [visuals[idx].size for idx in range(len(visuals))]
            if "max_new_tokens" not in gen_kwargs:
                gen_kwargs["max_new_tokens"] = 1024
            if "temperature" not in gen_kwargs:
                gen_kwargs["temperature"] = 0
            if "top_p" not in gen_kwargs:
                gen_kwargs["top_p"] = None
            if "num_beams" not in gen_kwargs:
                gen_kwargs["num_beams"] = 1
            try:
                # ominicpm does not give much information on how they do eval so I just use the chat format.
                response, context, _ = self.model.chat(
                    image=visuals[0],
                    msgs=msgs,
                    context=None,
                    tokenizer=self.tokenizer,
                    sampling=True if gen_kwargs["temperature"] > 0 else False,
                    temperature=gen_kwargs["temperature"],
                    top_p=gen_kwargs["top_p"],
                    num_beams=gen_kwargs["num_beams"],
                    max_new_tokens=gen_kwargs["max_new_tokens"],
                )
            except Exception as e:
                eval_logger.error(f"Error {e} in generating")
                cont = ""
            res.append(response)
            self.cache_hook.add_partial("generate_until", (context, gen_kwargs), response)
            pbar.update(1)
            # reorder this group of results back to original unsorted form
        res = re_ords.get_original(res)

        pbar.close()
        return res

    def generate_until_multi_round(self, requests) -> List[str]:
        raise NotImplementedError("TODO: Implement multi-round generation")