| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from __future__ import annotations |
| 5 | |
| 6 | from transformers import CONFIG_MAPPING, PreTrainedConfig |
| 7 | |
| 8 | # NOTE: Temporary shim for FunAudioChat checkpoints. |
| 9 | # These checkpoints use `model_type="funaudiochat"`, which is not currently |
| 10 | # recognized by released Transformers, and the public checkpoint does not |
| 11 | # provide an `auto_map` to enable `trust_remote_code=True`. |
| 12 | # Remove this file once Transformers adds native support (or the checkpoint |
| 13 | # provides an `auto_map`) and vLLM can rely on `AutoConfig.from_pretrained()`. |
| 14 | |
| 15 | |
| 16 | class FunAudioChatAudioEncoderConfig(PreTrainedConfig): |
| 17 | model_type = "funaudiochat_audio_encoder" |
| 18 | |
| 19 | def __init__( |
| 20 | self, |
| 21 | _attn_implementation: str | None = None, |
| 22 | num_mel_bins: int = 128, |
| 23 | encoder_layers: int = 32, |
| 24 | encoder_attention_heads: int = 20, |
| 25 | encoder_ffn_dim: int = 5120, |
| 26 | d_model: int = 1280, |
| 27 | dropout: float = 0.0, |
| 28 | attention_dropout: float = 0.0, |
| 29 | activation_function: str = "gelu", |
| 30 | activation_dropout: float = 0.0, |
| 31 | scale_embedding: bool = False, |
| 32 | initializer_range: float = 0.02, |
| 33 | max_source_positions: int = 1500, |
| 34 | n_window: int = 100, |
| 35 | output_dim: int = 3584, |
| 36 | bos_token_id: int | None = None, |
| 37 | codebook_size: int | None = None, |
| 38 | continuous_features_mode: str = "replace", |
| 39 | crq_transformer_config: dict | None = None, |
| 40 | eos_token_id: int | None = None, |
| 41 | group_size: int = 5, |
| 42 | enable_audio_invert_tower: bool = True, |
| 43 | pad_token_id: int | None = None, |
| 44 | **kwargs, |
| 45 | ) -> None: |
| 46 | attn_impl = kwargs.pop("_attn_implementation", None) or _attn_implementation |
| 47 | super().__init__(**kwargs) |
| 48 | # Match HF default for attention implementation selection. |
| 49 | self._attn_implementation = attn_impl or "sdpa" |
| 50 | |
| 51 | self.num_mel_bins = num_mel_bins |
| 52 | self.d_model = d_model |
| 53 | self.encoder_layers = encoder_layers |
| 54 | self.encoder_attention_heads = encoder_attention_heads |
| 55 | self.encoder_ffn_dim = encoder_ffn_dim |
| 56 | self.dropout = dropout |
| 57 | self.attention_dropout = attention_dropout |
| 58 | self.activation_function = activation_function |
| 59 | self.activation_dropout = activation_dropout |
| 60 | self.num_hidden_layers = encoder_layers |
| 61 | self.initializer_range = initializer_range |
| 62 | self.scale_embedding = scale_embedding |
| 63 | self.max_source_positions = max_source_positions |
| 64 | self.n_window = n_window |
| 65 | self.output_dim = output_dim |
| 66 | |
| 67 | self.bos_token_id = bos_token_id |
| 68 | self.codebook_size = codebook_size |
| 69 | self.continuous_features_mode = continuous_features_mode |
| 70 | self.crq_transformer_config = crq_transformer_config |
| 71 | self.eos_token_id = eos_token_id |
| 72 | self.group_size = group_size |
| 73 | self.enable_audio_invert_tower = enable_audio_invert_tower |
| 74 | self.pad_token_id = pad_token_id |
| 75 | |
| 76 | |
| 77 | class FunAudioChatConfig(PreTrainedConfig): |
| 78 | model_type = "funaudiochat" |
| 79 | attribute_map = { |
| 80 | "audio_token_id": "audio_token_index", |
| 81 | } |
| 82 | |
| 83 | def __init__( |
| 84 | self, |
| 85 | audio_config: PreTrainedConfig | dict | None = None, |
| 86 | text_config: PreTrainedConfig | dict | None = None, |
| 87 | audio_token_index: int = 151646, |
| 88 | ignore_index: int = -100, |
| 89 | hidden_size: int | None = None, |
| 90 | **kwargs, |
| 91 | ) -> None: |
| 92 | self.audio_token_index = audio_token_index |
| 93 | self.ignore_index = ignore_index |
| 94 | |
| 95 | if audio_config is None: |
| 96 | self.audio_config = FunAudioChatAudioEncoderConfig() |
| 97 | elif isinstance(audio_config, dict): |
| 98 | default_model_type = FunAudioChatAudioEncoderConfig.model_type |
| 99 | audio_config.setdefault("model_type", default_model_type) |
| 100 | self.audio_config = FunAudioChatAudioEncoderConfig(**audio_config) |
| 101 | else: |
| 102 | self.audio_config = audio_config |
| 103 | |
| 104 | if text_config is None: |
| 105 | self.text_config = CONFIG_MAPPING["qwen2"]() |
| 106 | elif isinstance(text_config, dict): |
| 107 | # Default to qwen2 for backwards compatibility; FunAudioChat uses |
| 108 | # qwen3 in practice for recent checkpoints. |
| 109 | text_config.setdefault("model_type", "qwen2") |
| 110 | self.text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config) |
| 111 | else: |
| 112 | self.text_config = text_config |
| 113 | |
| 114 | self.hidden_size = ( |
| 115 | int(self.text_config.hidden_size) |
| 116 | if hidden_size is None |
| 117 | else int(hidden_size) |
| 118 | ) |
| 119 | |
| 120 | super().__init__(**kwargs) |