File size: 13,231 Bytes
cb105bf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import copy

import torch
from transformers import BaseImageProcessor, PreTrainedTokenizer
from transformers.feature_extraction_utils import BatchFeature
from transformers.processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack

try:
    from .processor_core import (
        CHAT_TEMPLATE,
        CHAT_TEMPLATE_FAKE_THINKING,
        make_image_config_from_processor,
        process_images,
    )
except ImportError:
    from processor_core import (  # type: ignore[no-redef]
        CHAT_TEMPLATE,
        CHAT_TEMPLATE_FAKE_THINKING,
        make_image_config_from_processor,
        process_images,
    )


class ModRWKVProcessorKwargs(ProcessingKwargs, total=False):
    _defaults = {
        "text_kwargs": {
            "padding": False,
            "return_token_type_ids": False,
        },
        "images_kwargs": {},
    }


class ModRWKVProcessor(ProcessorMixin):
    attributes = ["image_processor", "tokenizer"]
    tokenizer_class = "RwkvTokenizer"
    user_image_tag = "<image>"

    def __init__(
        self,
        tokenizer: PreTrainedTokenizer = None,
        image_processor: BaseImageProcessor = None,
        chat_template=None,
        auto_insert_image_tags: bool = True,
        total_pixels_budget: bool = True,
    ):
        chat_template = CHAT_TEMPLATE if chat_template is None else chat_template
        super().__init__(
            tokenizer=tokenizer,
            image_processor=image_processor,
            chat_template=chat_template,
        )
        self.auto_insert_image_tags = auto_insert_image_tags
        self.total_pixels_budget = total_pixels_budget
        self.image_token = getattr(tokenizer, "image_token", "<|image_pad|>")
        self.vision_start_token = getattr(tokenizer, "vision_start_token", "<|vision_start|>")
        self.vision_end_token = getattr(tokenizer, "vision_end_token", "<|vision_end|>")
        self.image_token_id = self.tokenizer.convert_tokens_to_ids(self.image_token)
        self.vision_start_token_id = self.tokenizer.convert_tokens_to_ids(self.vision_start_token)
        self.vision_end_token_id = self.tokenizer.convert_tokens_to_ids(self.vision_end_token)
        self.vision_image_token = (
            f"{self.vision_start_token}{self.image_token}{self.vision_end_token}"
        )

    def to_dict(self):
        output = {}
        if self.image_processor is not None:
            output["image_processor"] = self.image_processor.to_dict()
        if getattr(self, "auto_map", None) is not None:
            output["auto_map"] = copy.deepcopy(self.auto_map)
        output["processor_class"] = self.__class__.__name__
        if not self.auto_insert_image_tags:
            output["auto_insert_image_tags"] = False
        output["total_pixels_budget"] = self.total_pixels_budget
        return output

    def _flatten_images(self, images):
        if images is None:
            return []
        if not isinstance(images, (list, tuple)):
            return [images]

        flat_images = []
        for item in images:
            if isinstance(item, (list, tuple)):
                flat_images.extend(self._flatten_images(item))
            else:
                flat_images.append(item)
        return flat_images

    def _get_num_images_per_text_sample(self, images, batch_size):
        if images is None:
            return [0] * batch_size
        if batch_size == 1:
            return [len(self._flatten_images(images))]
        if isinstance(images, (list, tuple)) and len(images) == batch_size:
            return [len(self._flatten_images(sample_images)) for sample_images in images]
        return None

    def _get_images_per_text_sample(self, images, batch_size):
        if images is None:
            return [[] for _ in range(batch_size)]
        if batch_size == 1:
            return [self._flatten_images(images)]
        if isinstance(images, (list, tuple)) and len(images) == batch_size:
            return [self._flatten_images(sample_images) for sample_images in images]
        return None

    def _process_images(self, images, batch_size, images_kwargs):
        image_groups = self._get_images_per_text_sample(images, batch_size)
        if image_groups is None:
            image_groups = [self._flatten_images(images)]
            num_images_per_sample = None
        else:
            num_images_per_sample = [len(group) for group in image_groups]

        image_config = make_image_config_from_processor(
            self.image_processor,
            **images_kwargs,
        )
        processed_groups = [process_images(group, image_config) for group in image_groups]
        num_image_tokens = [
            count
            for processed in processed_groups
            for count in processed.image_token_counts
        ]
        if not num_image_tokens:
            return {}, None, None, num_images_per_sample

        pixel_values = torch.cat(
            [
                processed.flat_patches
                for processed in processed_groups
                if processed.flat_patches.numel() > 0
            ],
            dim=0,
        )
        image_grid_thw = torch.cat(
            [
                processed.grid_thw
                for processed in processed_groups
                if processed.grid_thw.numel() > 0
            ],
            dim=0,
        )
        return (
            {
                "pixel_values": pixel_values,
                "image_grid_thw": image_grid_thw,
            },
            image_grid_thw,
            num_image_tokens,
            num_images_per_sample,
        )

    def _normalize_image_tags(self, text):
        return text.replace(self.user_image_tag, self.vision_image_token)

    def _strip_excess_image_tags(self, text, num_allowed):
        tag = self.user_image_tag
        count = text.count(tag)
        if count <= num_allowed:
            return text
        parts = text.split(tag)
        kept = tag.join(parts[: num_allowed + 1])
        rest = "".join(parts[num_allowed + 1 :])
        return kept + rest

    def _append_missing_image_tags(self, text, num_missing_images):
        if num_missing_images <= 0:
            return text
        return text + self.vision_image_token * num_missing_images

    def _get_num_multimodal_tokens(self, image_grid_thw=None, **kwargs):
        vision_data = {}
        if image_grid_thw is not None:
            processor_defaults = getattr(self.image_processor, "_defaults", {})
            images_kwargs = dict(processor_defaults.get("images_kwargs", {}))
            images_kwargs.update(kwargs)
            merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size

            num_image_patches = [int(grid[0] * grid[1] * grid[2]) for grid in image_grid_thw]
            num_image_tokens = [num_patches // merge_size**2 for num_patches in num_image_patches]
            vision_data.update(
                {
                    "num_image_tokens": num_image_tokens,
                    "num_image_patches": num_image_patches,
                }
            )
        return MultiModalData(**vision_data)

    def _count_token_occurrences(self, input_ids, token_id):
        return [sum(1 for token in sample_ids if token == token_id) for sample_ids in input_ids]

    def _validate_image_token_alignment(self, text_inputs, expected_image_tokens, expected_num_images):
        input_ids = text_inputs["input_ids"]
        actual_image_tokens = self._count_token_occurrences(input_ids, self.image_token_id)
        actual_vision_starts = self._count_token_occurrences(input_ids, self.vision_start_token_id)
        actual_vision_ends = self._count_token_occurrences(input_ids, self.vision_end_token_id)

        if actual_image_tokens != expected_image_tokens:
            raise ValueError(
                "Image token count does not match image_grid_thw-derived token count: "
                f"expected {expected_image_tokens}, got {actual_image_tokens}."
            )
        if actual_vision_starts != expected_num_images or actual_vision_ends != expected_num_images:
            raise ValueError(
                "Vision boundary token count does not match the number of image placeholders: "
                f"expected {expected_num_images}, got starts={actual_vision_starts}, ends={actual_vision_ends}."
            )

    def __call__(self, images=None, text=None, **kwargs: Unpack[ModRWKVProcessorKwargs]):
        output_kwargs = self._merge_kwargs(
            ModRWKVProcessorKwargs,
            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
            **kwargs,
        )

        if not isinstance(text, list):
            text = [text] if text is not None else None

        batch_size = len(text) if text is not None else 1
        if images is not None:
            (
                image_inputs,
                image_grid_thw,
                num_image_tokens,
                num_images_per_sample,
            ) = self._process_images(
                images,
                batch_size,
                output_kwargs["images_kwargs"],
            )
        else:
            image_inputs = {}
            image_grid_thw = None
            num_image_tokens = None
            num_images_per_sample = None

        if text is None:
            return BatchFeature(data=image_inputs)

        text = text.copy()
        expected_image_tokens = [0 for _ in text]
        expected_num_images = [0 for _ in text]
        if image_grid_thw is not None:
            index = 0
            for i in range(len(text)):
                if not self.auto_insert_image_tags:
                    text[i] = text[i].replace(self.user_image_tag, " ")
                else:
                    if num_images_per_sample is not None:
                        text[i] = self._strip_excess_image_tags(text[i], num_images_per_sample[i])
                    text[i] = self._normalize_image_tags(text[i])

                if self.auto_insert_image_tags and num_images_per_sample is not None:
                    missing = num_images_per_sample[i] - text[i].count(self.image_token)
                    text[i] = self._append_missing_image_tags(text[i], missing)

                if self.auto_insert_image_tags:
                    placeholder_count = text[i].count(self.vision_image_token)
                    if index + placeholder_count > len(num_image_tokens):
                        raise ValueError(
                            "Number of image placeholders in text exceeds provided images: "
                            f"consumed {index + placeholder_count}, available {len(num_image_tokens)}."
                        )
                    sample_counts = num_image_tokens[index : index + placeholder_count]
                    text[i] = self.tokenizer.expand_image_placeholders(
                        text[i],
                        sample_counts,
                    )
                    expected_image_tokens[i] += sum(sample_counts)
                    expected_num_images[i] += len(sample_counts)
                    index += placeholder_count
                else:
                    while self.image_token in text[i]:
                        if index >= len(num_image_tokens):
                            raise ValueError(
                                "Number of image placeholders in text exceeds provided images: "
                                f"consumed {index + 1}, available {len(num_image_tokens)}."
                            )
                        text[i] = text[i].replace(
                            self.image_token,
                            "<|placeholder|>" * num_image_tokens[index],
                            1,
                        )
                        expected_image_tokens[i] += num_image_tokens[index]
                        expected_num_images[i] += 1
                        index += 1
                    text[i] = text[i].replace("<|placeholder|>", self.image_token)

            if self.auto_insert_image_tags and index != len(num_image_tokens):
                raise ValueError(
                    "Number of image placeholders in text does not match provided images: "
                    f"consumed {index}, available {len(num_image_tokens)}."
                )
        else:
            for i in range(len(text)):
                text[i] = text[i].replace(self.user_image_tag, "")

        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
        if image_grid_thw is not None:
            self._validate_image_token_alignment(
                text_inputs,
                expected_image_tokens,
                expected_num_images,
            )
        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)

    def apply_chat_template(self, conversation, chat_template=None, **kwargs):
        kwargs.setdefault("return_dict", True)
        return super().apply_chat_template(
            conversation,
            chat_template=chat_template,
            **kwargs,
        )


ModRWKVProcessor.register_for_auto_class("AutoProcessor")