qwen2_audio.py

# SPDX-License-Identifier: Apache-2.0

# 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."""
from functools import cached_property
from typing import (Any, Iterable, Mapping, Optional, Set, Tuple, TypedDict,
                    Union)

import torch
import torch.nn as nn
from transformers import BatchFeature
from transformers.models.qwen2_audio import (Qwen2AudioConfig,
                                             Qwen2AudioEncoder,
                                             Qwen2AudioProcessor)
from transformers.models.whisper import WhisperFeatureExtractor

from vllm.config import VllmConfig
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
                                    NestedTensors)
from vllm.multimodal.parse import (AudioProcessorItems, MultiModalDataItems,
                                   MultiModalDataParser)
from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                        BaseProcessingInfo, PromptReplacement,
                                        PromptReplacementDetails)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsMultiModal, SupportsPP
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
                    maybe_prefix, merge_multimodal_embeddings)


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

    feature_attention_mask: torch.Tensor
    """Shape: `(num_audios, 3000)`"""


# === 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


# From Qwen2AudioEncoder._get_feat_extract_output_lengths
def _get_feat_extract_output_lengths(input_lengths: torch.Tensor):
    feat_lengths = (input_lengths - 1) // 2 + 1
    output_lengths = (feat_lengths - 2) // 2 + 1
    return feat_lengths, output_lengths


class Qwen2AudioProcessingInfo(BaseProcessingInfo):

    def get_hf_config(self):
        return self.ctx.get_hf_config(Qwen2AudioConfig)

    def get_hf_processor(
        self,
        *,
        # Ignored in initialization
        sampling_rate: Optional[int] = None,
        **kwargs: object,
    ) -> Qwen2AudioProcessor:
        return self.ctx.get_hf_processor(Qwen2AudioProcessor, **kwargs)

    def get_feature_extractor(
        self,
        *,
        # Ignored in initialization
        sampling_rate: Optional[int] = None,
    ) -> WhisperFeatureExtractor:
        hf_processor = self.get_hf_processor(sampling_rate=sampling_rate)
        feature_extractor = hf_processor.feature_extractor  # type: ignore
        assert isinstance(feature_extractor, WhisperFeatureExtractor)
        return feature_extractor

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

    def get_mm_max_tokens_per_item(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> Mapping[str, int]:
        hf_config = self.get_hf_config()
        max_source_positions = hf_config.audio_config.max_source_positions
        max_output_lengths = (max_source_positions - 2) // 2 + 1

        return {"audio": max_output_lengths}


class Qwen2AudioDummyInputsBuilder(
        BaseDummyInputsBuilder[Qwen2AudioProcessingInfo]):

    def get_dummy_processor_inputs(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> ProcessorInputs:
        feature_extractor = self.info.get_feature_extractor()

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

        mm_data = {
            "audio":
            self._get_dummy_audios(length=audio_len, num_audios=num_audios)
        }

        return ProcessorInputs(
            prompt_text="<|AUDIO|>" * num_audios,
            mm_data=mm_data,
        )


class Qwen2AudioMultiModalProcessor(
        BaseMultiModalProcessor[Qwen2AudioProcessingInfo]):

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

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, Any],
    ) -> BatchFeature:
        # Text-only input not supported in composite processor
        if not mm_data or not mm_data.get("audios", []):
            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,
        )

        return super()._call_hf_processor(
            prompt=prompt,
            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"),
        )

    def _get_prompt_replacements(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> list[PromptReplacement]:
        processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()

        # 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|>")

        audio_token_id = vocab[audio_token]
        audio_bos_id = vocab[audio_bos_token]
        audio_eos_id = vocab[audio_eos_token]

        feature_attention_mask = out_mm_kwargs.get("feature_attention_mask")
        if feature_attention_mask is None:
            audio_output_lengths = []
        else:
            assert isinstance(feature_attention_mask, torch.Tensor)
            _, audio_output_lens = _get_feat_extract_output_lengths(
                feature_attention_mask.sum(-1))

            audio_output_lengths = audio_output_lens.tolist()

        def get_replacement_qwen2_audio(item_idx: int):
            num_features = audio_output_lengths[item_idx]
            if num_features == 0:
                audios = mm_items.get_items("audio", AudioProcessorItems)
                audio = audios.get(item_idx)
                raise ValueError(
                    f"The audio {audio} (len={len(audio)}) is too short "
                    "to be represented inside the model")

            audio_tokens = [audio_token_id] * num_features

            return PromptReplacementDetails(
                full=[audio_bos_id] + audio_tokens + [audio_eos_id],
                features=audio_tokens,
            )

        return [
            PromptReplacement(
                modality="audio",
                target=audio_token,
                replacement=get_replacement_qwen2_audio,
            )
        ]


@MULTIMODAL_REGISTRY.register_processor(
    Qwen2AudioMultiModalProcessor,
    info=Qwen2AudioProcessingInfo,
    dummy_inputs=Qwen2AudioDummyInputsBuilder)
class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal,
                                         SupportsPP):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
        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

        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=["Qwen2ForCausalLM"],
        )

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

    @cached_property
    def sampler(self):
        if hasattr(self.language_model, "sampler"):
            return self.language_model.sampler

        return get_sampler()

    def _validate_and_reshape_mm_tensor(self, mm_input: object,
                                        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
        audio_output_lengths = audio_output_lengths.unsqueeze(1)
        audio_features_mask = torch.arange(max_audio_tokens).expand(
            num_audios, max_audio_tokens).to(
                audio_output_lengths.device) < audio_output_lengths
        masked_audio_features = audio_features[audio_features_mask].view(
            -1, embed_dim)

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

    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
        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,
        multimodal_embeddings: Optional[NestedTensors] = None,
    ) -> 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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:

        if intermediate_tensors is not None:
            inputs_embeds = None

        # 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

        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        return self.language_model.sample(logits, sampling_metadata)

    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)