qwen2_audio.py 16.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# Copyright 2024 The Qwen team.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen2-Audio model compatible with HuggingFace weights."""
25
from collections.abc import Iterable, Mapping, Sequence
26
from typing import Any, Optional, TypedDict, Union
27
28
29

import torch
import torch.nn as nn
30
from transformers import BatchFeature
31
32
33
34
from transformers.models.qwen2_audio import (Qwen2AudioConfig,
                                             Qwen2AudioEncoder,
                                             Qwen2AudioProcessor)
from transformers.models.whisper import WhisperFeatureExtractor
35

36
from vllm.config import VllmConfig
37
from vllm.model_executor.sampling_metadata import SamplingMetadata
38
from vllm.multimodal import MULTIMODAL_REGISTRY
39
40
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                    MultiModalKwargs)
41
42
from vllm.multimodal.parse import (AudioProcessorItems, MultiModalDataItems,
                                   MultiModalDataParser)
43
from vllm.multimodal.processing import (BaseMultiModalProcessor,
44
                                        BaseProcessingInfo, PromptReplacement,
45
                                        PromptUpdate, PromptUpdateDetails)
46
from vllm.multimodal.profiling import BaseDummyInputsBuilder
47
from vllm.sequence import IntermediateTensors
48

49
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
50
51
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
                    maybe_prefix, merge_multimodal_embeddings)
52
53
54
55
56


# # === Audio Inputs === #
class Qwen2AudioInputs(TypedDict):
    input_features: torch.Tensor
57
    """Shape: `(num_audios, num_mel_bins, 3000)`"""
58
59

    feature_attention_mask: torch.Tensor
60
    """Shape: `(num_audios, 3000)`"""
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76


# === Audio Encoder === #


class Qwen2AudioMultiModalProjector(nn.Module):

    def __init__(self, audio_hidden_size: int, text_hidden_size: int):
        super().__init__()
        self.linear = nn.Linear(audio_hidden_size, text_hidden_size, bias=True)

    def forward(self, audio_features):
        hidden_states = self.linear(audio_features)
        return hidden_states


77
# From Qwen2AudioEncoder._get_feat_extract_output_lengths
78
def _get_feat_extract_output_lengths(input_lengths: torch.Tensor):
79
80
81
    feat_lengths = (input_lengths - 1) // 2 + 1
    output_lengths = (feat_lengths - 2) // 2 + 1
    return feat_lengths, output_lengths
82
83


84
class Qwen2AudioProcessingInfo(BaseProcessingInfo):
85

86
    def get_hf_config(self):
87
88
        return self.ctx.get_hf_config(Qwen2AudioConfig)

89
    def get_hf_processor(
90
91
92
93
        self,
        *,
        # Ignored in initialization
        sampling_rate: Optional[int] = None,
94
        **kwargs: object,
95
    ) -> Qwen2AudioProcessor:
96
        return self.ctx.get_hf_processor(Qwen2AudioProcessor, **kwargs)
97

98
    def get_feature_extractor(
99
100
101
102
103
        self,
        *,
        # Ignored in initialization
        sampling_rate: Optional[int] = None,
    ) -> WhisperFeatureExtractor:
104
        hf_processor = self.get_hf_processor(sampling_rate=sampling_rate)
105
106
107
108
        feature_extractor = hf_processor.feature_extractor  # type: ignore
        assert isinstance(feature_extractor, WhisperFeatureExtractor)
        return feature_extractor

109
110
    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"audio": None}
111

112
113
114
115

class Qwen2AudioDummyInputsBuilder(
        BaseDummyInputsBuilder[Qwen2AudioProcessingInfo]):

116
117
118
119
120
121
122
123
124
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_audios = mm_counts.get("audio", 0)

        hf_processor = self.info.get_hf_processor()
        audio_token = hf_processor.audio_token

        return audio_token * num_audios

    def get_dummy_mm_data(
125
        self,
126
127
        seq_len: int,
        mm_counts: Mapping[str, int],
128
    ) -> MultiModalDataDict:
129
        feature_extractor = self.info.get_feature_extractor()
130
131
132
133
134

        sampling_rate = feature_extractor.sampling_rate
        audio_len = feature_extractor.chunk_length * sampling_rate
        num_audios = mm_counts.get("audio", 0)

135
        return {
136
137
138
139
            "audio":
            self._get_dummy_audios(length=audio_len, num_audios=num_audios)
        }

140

141
142
class Qwen2AudioMultiModalProcessor(
        BaseMultiModalProcessor[Qwen2AudioProcessingInfo]):
143

144
    def _get_data_parser(self) -> MultiModalDataParser:
145
        feature_extractor = self.info.get_feature_extractor()
146
        return MultiModalDataParser(target_sr=feature_extractor.sampling_rate)
147

148
149
150
    def _call_hf_processor(
        self,
        prompt: str,
151
        mm_data: Mapping[str, object],
152
        mm_kwargs: Mapping[str, Any],
153
    ) -> BatchFeature:
154
155
156
157
158
159
160
        # NOTE - we rename audios -> audio in mm data because transformers has
        # deprecated audios for the qwen2audio processor and will remove
        # support for it in transformers 4.54.
        audios = mm_data.pop("audios", [])
        if audios:
            mm_data["audio"] = audios

161
        # Text-only input not supported in composite processor
162
        if not mm_data.get("audio", []):
163
164
165
166
167
168
169
170
171
            prompt_ids = self.info.get_tokenizer().encode(prompt)
            prompt_ids = self._apply_hf_processor_tokens_only(prompt_ids)
            return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt")

        feature_extractor = self.info.get_feature_extractor(**mm_kwargs)
        mm_kwargs = dict(
            **mm_kwargs,
            sampling_rate=feature_extractor.sampling_rate,
        )
172

173
        return super()._call_hf_processor(
174
            prompt=prompt,
175
176
177
178
179
180
181
182
183
184
185
186
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
        )

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return dict(
            input_features=MultiModalFieldConfig.batched("audio"),
            feature_attention_mask=MultiModalFieldConfig.batched("audio"),
187
188
        )

189
    def _get_prompt_updates(
190
191
        self,
        mm_items: MultiModalDataItems,
192
193
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
194
    ) -> Sequence[PromptUpdate]:
195
196
197
        processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()
198
199
200
201
202
203
204

        # Use getattr with default to be compatible with transformers<4.48
        audio_token = getattr(processor, "audio_token", "<|AUDIO|>")
        audio_bos_token = getattr(processor, "audio_bos_token",
                                  "<|audio_bos|>")
        audio_eos_token = getattr(processor, "audio_eos_token",
                                  "<|audio_eos|>")
205

206
207
208
209
        audio_token_id = vocab[audio_token]
        audio_bos_id = vocab[audio_bos_token]
        audio_eos_id = vocab[audio_eos_token]

210
        feature_attention_mask = out_mm_kwargs.get("feature_attention_mask")
211
212
213
        if feature_attention_mask is None:
            audio_output_lengths = []
        else:
214
            assert isinstance(feature_attention_mask, torch.Tensor)
215
            _, audio_output_lens = _get_feat_extract_output_lengths(
216
217
                feature_attention_mask.sum(-1))

218
219
            audio_output_lengths = audio_output_lens.tolist()

220
        def get_replacement_qwen2_audio(item_idx: int):
221
222
            num_features = audio_output_lengths[item_idx]
            if num_features == 0:
223
                audios = mm_items.get_items("audio", AudioProcessorItems)
224
225
226
227
                audio_len = audios.get_audio_length(item_idx)

                raise ValueError(f"The audio (len={audio_len}) is too short "
                                 "to be represented inside the model")
228

229
            audio_tokens = [audio_token_id] * num_features
230

231
232
233
            return PromptUpdateDetails.select_token_id(
                [audio_bos_id] + audio_tokens + [audio_eos_id],
                embed_token_id=audio_token_id,
234
            )
235
236
237
238

        return [
            PromptReplacement(
                modality="audio",
239
                target=audio_token,
240
241
                replacement=get_replacement_qwen2_audio,
            )
242
        ]
243
244


245
246
247
248
@MULTIMODAL_REGISTRY.register_processor(
    Qwen2AudioMultiModalProcessor,
    info=Qwen2AudioProcessingInfo,
    dummy_inputs=Qwen2AudioDummyInputsBuilder)
249
250
251
class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal,
                                         SupportsPP):

252
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
253
        super().__init__()
254
255
256
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
257
258
259
260
261
262
263
264
265
        self.config = config
        self.multimodal_config = multimodal_config

        self.audio_tower = Qwen2AudioEncoder(config.audio_config)
        self.multi_modal_projector = Qwen2AudioMultiModalProjector(
            config.audio_config.d_model, config.text_config.hidden_size)

        self.quant_config = quant_config

266
267
268
269
270
271
        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=["Qwen2ForCausalLM"],
        )
272
273
274
275

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

276
    def _validate_and_reshape_mm_tensor(self, mm_input: object,
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
                                        name: str) -> torch.Tensor:
        if not isinstance(mm_input, (torch.Tensor, list)):
            raise ValueError(f"Incorrect type of {name}. "
                             f"Got type: {type(mm_input)}")
        if isinstance(mm_input, torch.Tensor):
            return torch.concat(list(mm_input))
        else:
            return torch.concat(mm_input)

    def _parse_and_validate_audio_input(
            self, **kwargs: object) -> Optional[Qwen2AudioInputs]:
        input_features = kwargs.pop('input_features', None)
        feature_attention_mask = kwargs.pop('feature_attention_mask', None)
        if input_features is None:
            return None
        input_features = self._validate_and_reshape_mm_tensor(
            input_features, 'input_features')
        feature_attention_mask = self._validate_and_reshape_mm_tensor(
            feature_attention_mask, 'feature_attention_mask')
        if not isinstance(input_features, (torch.Tensor, list)):
            raise ValueError("Incorrect type of audio input features. "
                             f"Got type: {type(input_features)}")
        return Qwen2AudioInputs(input_features=input_features,
                                feature_attention_mask=feature_attention_mask)

    def _process_audio_input(self,
                             audio_input: Qwen2AudioInputs) -> torch.Tensor:

        input_features = audio_input["input_features"]
        feature_attention_mask = audio_input["feature_attention_mask"]

        audio_feat_lengths, audio_output_lengths = (
            self.audio_tower._get_feat_extract_output_lengths(
                feature_attention_mask.sum(-1)))

        batch_size, _, max_mel_seq_len = input_features.shape
        max_seq_len = (max_mel_seq_len - 2) // 2 + 1
        # Create a sequence tensor of shape (batch_size, max_seq_len)
        seq_range = (torch.arange(
            0,
            max_seq_len,
            dtype=audio_feat_lengths.dtype,
            device=audio_feat_lengths.device).unsqueeze(0).expand(
                batch_size, max_seq_len))
        lengths_expand = audio_feat_lengths.unsqueeze(-1).expand(
            batch_size, max_seq_len)
        # Create mask
        padding_mask = seq_range >= lengths_expand

        audio_attention_mask_ = padding_mask.view(
            batch_size, 1, 1, max_seq_len).expand(batch_size, 1, max_seq_len,
                                                  max_seq_len)
        audio_attention_mask = audio_attention_mask_.to(
            dtype=self.audio_tower.conv1.weight.dtype,
            device=self.audio_tower.conv1.weight.device)
        audio_attention_mask[audio_attention_mask_] = float("-inf")

        audio_outputs = self.audio_tower(input_features,
                                         attention_mask=audio_attention_mask)
        selected_audio_feature = audio_outputs.last_hidden_state
        audio_features = self.multi_modal_projector(selected_audio_feature)
        num_audios, max_audio_tokens, embed_dim = audio_features.shape
339
        audio_output_lengths = audio_output_lengths.unsqueeze(1)
340
        audio_features_mask = torch.arange(max_audio_tokens).expand(
341
342
            num_audios, max_audio_tokens).to(
                audio_output_lengths.device) < audio_output_lengths
343
344
345
        masked_audio_features = audio_features[audio_features_mask].view(
            -1, embed_dim)

346
347
348
        # Split to tuple of embeddings for individual audio input.
        return torch.split(masked_audio_features,
                           audio_output_lengths.flatten().tolist())
349

350
351
352
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

353
    def get_multimodal_embeddings(
354
            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
355
356
357
358
359
360
361
362
363
        audio_input = self._parse_and_validate_audio_input(**kwargs)
        if audio_input is None:
            return None
        masked_audio_features = self._process_audio_input(audio_input)
        return masked_audio_features

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
364
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
365
366
367
368
369
370
371
372
    ) -> torch.Tensor:
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
        if multimodal_embeddings is not None:
            inputs_embeds = merge_multimodal_embeddings(
                input_ids, inputs_embeds, multimodal_embeddings,
                self.config.audio_token_index)
        return inputs_embeds

373
374
375
376
377
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
378
        inputs_embeds: Optional[torch.Tensor] = None,
379
380
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:
381

382
383
        if intermediate_tensors is not None:
            inputs_embeds = None
384
385
386
387
388
389
390
391
392

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility.
        elif inputs_embeds is None:
            multimodal_embeddings = self.get_multimodal_embeddings(**kwargs)
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      multimodal_embeddings)
            input_ids = None

393
394
395
396
        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
397
398
        return hidden_states

399
400
401
402
403
404
405
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
406

407
408
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
409
410
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)