crfm-helm 0.4.0__py3-none-any.whl → 0.5.10__py3-none-any.whl

helm/clients/audio_language/qwen_audiolm_client.py

@@ -0,0 +1,153 @@
+from threading import Lock
+from typing import Any, Dict, List, Optional
+
+from dataclasses import dataclass
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from helm.common.cache import CacheConfig
+from helm.common.gpu_utils import get_torch_device_name
+from helm.common.hierarchical_logger import hexception, hlog, htrack_block
+from helm.common.media_object import TEXT_TYPE
+from helm.common.request import Request, RequestResult, GeneratedOutput, Token
+from helm.common.request import wrap_request_time
+from helm.clients.client import CachingClient, generate_uid_for_multimodal_prompt
+
+
+@dataclass(frozen=True)
+class LoadedQwenModelProcessor:
+    """Loaded model and processor for Qwen."""
+
+    model: AutoModelForCausalLM
+    tokenizer: AutoTokenizer
+
+
+_models_lock: Lock = Lock()
+_models: Dict[str, Optional[LoadedQwenModelProcessor]] = {
+    "Qwen/Qwen-Audio-Chat": None,
+}
+
+
+class QwenAudioLMClient(CachingClient):
+    """
+    From https://huggingface.co/Qwen/Qwen-Audio,
+    Qwen-Audio-Chat (Qwen Large Audio Language Model) is the audio multimodal version of the large model series,
+    Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-Audio-Chat accepts audio, text as inputs, outputs text.
+    Alibaba released Qwen-Audio and Qwen-Audio-Chat, which is a chatbot model based on Qwen-Audio.
+    We for now integrated Qwen-Audio-Chat for instruction-following tasks.
+
+    Paper: https://arxiv.org/abs/2311.07919
+    """
+
+    def __init__(self, cache_config: CacheConfig):
+        super().__init__(cache_config=cache_config)
+        self._device: str = get_torch_device_name()
+
+    def _get_model(self, helm_model_name: str) -> LoadedQwenModelProcessor:
+        global _models_lock
+        global _models
+
+        model_name: str
+        if helm_model_name == "qwen-audio-chat":
+            model_name = "Qwen/Qwen-Audio-Chat"
+        else:
+            raise ValueError(f"Unhandled model name: {helm_model_name}")
+
+        # Ensure that only one thread is loading the model at a time
+        with _models_lock:
+            loaded_model_processor = _models[model_name]
+            if loaded_model_processor is None:
+                hlog(f"Loading model {model_name} and caching in memory...")
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_name,
+                    device_map=self._device,
+                    trust_remote_code=True,
+                    bf16=True,
+                ).eval()
+                tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+                _models[model_name] = LoadedQwenModelProcessor(model, tokenizer)
+                loaded_model_processor = _models[model_name]
+
+        assert loaded_model_processor is not None
+        return loaded_model_processor
+
+    def make_request(self, request: Request) -> RequestResult:
+        assert request.multimodal_prompt is not None, "Multimodal prompt is required"
+
+        loaded_model_processor: LoadedQwenModelProcessor = self._get_model(request.model_engine)
+        model = loaded_model_processor.model
+        tokenizer = loaded_model_processor.tokenizer
+
+        generation_args = {
+            "max_length": request.max_tokens,
+        }
+
+        query: List[Dict[str, str]] = []
+        prompt_text: str = ""
+
+        for media_object in request.multimodal_prompt.media_objects:
+            if media_object.is_type("audio") and media_object.location:
+                query.append({"audio": media_object.location})
+            elif media_object.is_type(TEXT_TYPE):
+                if media_object.text is None:
+                    raise ValueError("MediaObject of text type has missing text field value")
+
+                query.append({"text": media_object.text})
+                prompt_text += media_object.text
+            else:
+                raise ValueError(f"Unrecognized MediaObject type {media_object.type}")
+
+        completions: List[GeneratedOutput] = []
+        request_time: float = 0
+        request_datetime: Optional[int] = None
+        all_cached: bool = True
+
+        with htrack_block(f"Generating for prompt: {prompt_text}"):
+            for completion_index in range(request.num_completions):
+                try:
+
+                    def do_it() -> Dict[str, Any]:
+                        completion, _ = model.chat(  # type: ignore
+                            tokenizer, query=tokenizer.from_list_format(query), history=None  # type: ignore
+                        )
+                        tokens: List[str] = tokenizer.tokenize(completion)  # type: ignore
+                        return {"output": (completion, tokens)}
+
+                    # Include the prompt and model name in the cache key
+                    cache_key = CachingClient.make_cache_key(
+                        raw_request={
+                            "completion_index": completion_index,
+                            "model": request.model,
+                            "prompt": generate_uid_for_multimodal_prompt(request.multimodal_prompt),
+                            **generation_args,
+                        },
+                        request=request,
+                    )
+                    result, cached = self.cache.get(cache_key, wrap_request_time(do_it))
+                except RuntimeError as model_error:
+                    hexception(model_error)
+                    return RequestResult(
+                        success=False, cached=False, error=str(model_error), completions=[], embedding=[]
+                    )
+
+                text, tokens = result["output"]
+
+                # Tokenize truncated text to get the list of tokens
+                completions.append(
+                    GeneratedOutput(
+                        text=text, logprob=0, tokens=[Token(text=str(token), logprob=0) for token in tokens]
+                    )
+                )
+
+                request_time += result["request_time"]
+                # Use the datetime from the first completion because that's when the request was fired
+                request_datetime = request_datetime or result.get("request_datetime")
+                all_cached = all_cached and cached
+
+        return RequestResult(
+            success=True,
+            cached=all_cached,
+            request_time=request_time,
+            request_datetime=request_datetime,
+            completions=completions,
+            embedding=[],
+        )

helm/clients/audio_language/qwen_omni/configuration_qwen2_5_omni.py

@@ -0,0 +1,519 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen2_5_omni.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+#
+# 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.
+
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2_5OmniVisionEncoderConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2_5OmniThinkerVision`].
+    It is used to instantiate a Qwen2.5-VL vision encoder according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a
+    similar configuration to that of the audio encoder of the Qwen2.5-VL architecture.
+
+    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the
+    model outputs. Read the documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        depth (`int`, *optional*, defaults to 32):
+            Number of layers (depth) in the model.
+        hidden_size (`int`, *optional*, defaults to 3584):
+            The size of the hidden layers.
+        hidden_act (`str`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function used in the model. Supported options include
+            `"quick_gelu"` and others as applicable.
+        mlp_ratio (`float`, *optional*, defaults to 4):
+            The ratio used to determine the size of the MLP (Multi-Layer Perceptron) hidden layer.
+        num_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer.
+        in_channels (`int`, *optional*, defaults to 3):
+            Number of input channels.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size of the patches extracted from the input.
+        spatial_merge_size (`int`, *optional*, defaults to 2):
+            The size used for merging spatial dimensions.
+        temporal_patch_size (`int`, *optional*, defaults to 2):
+            The size used for patches along the temporal dimension.
+
+    Example:
+        >>> # xdoctest: +SKIP
+        >>> from transformers import Qwen2_5OmniVisionEncoderConfig, Qwen2_5OmniVisionEncoder
+        >>> # Initializing a Qwen2_5OmniVisionEncoderConfig
+        >>> configuration = Qwen2_5OmniVisionEncoderConfig()
+        >>> # Initializing a Qwen2_5OmniVisionEncoder (with random weights)
+        >>> model = Qwen2_5OmniVisionEncoder(configuration)
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "qwen2_5_omni_vision_encoder"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        depth=32,
+        hidden_size=3584,
+        hidden_act="silu",
+        intermediate_size=3420,
+        num_heads=16,
+        in_channels=3,
+        patch_size=14,
+        spatial_merge_size=2,
+        temporal_patch_size=2,
+        window_size=112,
+        out_hidden_size=3584,
+        fullatt_block_indexes=[7, 15, 23, 31],
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.depth = depth
+        self.hidden_size = hidden_size
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.num_heads = num_heads
+        self.in_channels = in_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+        self.window_size = window_size
+        self.fullatt_block_indexes = fullatt_block_indexes
+        self.out_hidden_size = out_hidden_size
+
+
+class Qwen2_5OmniAudioEncoderConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_audio_encoder"
+
+    def __init__(
+        self,
+        num_mel_bins=128,
+        encoder_layers=32,
+        encoder_attention_heads=20,
+        encoder_ffn_dim=5120,
+        d_model=1280,
+        dropout=0,
+        attention_dropout=0,
+        activation_function="gelu",
+        activation_dropout=0,
+        scale_embedding=False,
+        initializer_range=0.02,
+        max_source_positions=1500,
+        n_window=100,
+        output_dim=3584,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.num_mel_bins = num_mel_bins
+        self.d_model = d_model
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_function = activation_function
+        self.activation_dropout = activation_dropout
+        self.num_hidden_layers = encoder_layers
+        self.initializer_range = initializer_range
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        self.max_source_positions = max_source_positions
+        self.n_window = n_window
+        self.output_dim = output_dim
+
+
+class Qwen2_5OmniTextConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_text"
+    is_composition = False
+
+    def __init__(
+        self,
+        vocab_size=152064,
+        hidden_size=3584,
+        intermediate_size=18944,
+        num_hidden_layers=28,
+        num_attention_heads=28,
+        num_key_value_heads=4,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        rms_norm_eps=1e-06,
+        use_cache=True,
+        rope_theta=1000000.0,
+        use_sliding_window=False,
+        sliding_window=32768,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        rope_scaling=None,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.rope_scaling = rope_scaling
+        if self.rope_scaling is None:
+            self.rope_scaling = {"mrope_section": [16, 24, 24], "rope_type": "default", "type": "default"}
+        self.initializer_range = initializer_range
+
+        super().__init__(**kwargs)
+
+
+class Qwen2_5OmniThinkerConfig(PretrainedConfig):
+    model_type = "qwen2_5_omni_thinker"
+    sub_configs = {
+        "audio_config": Qwen2_5OmniAudioEncoderConfig,
+        "vision_config": Qwen2_5OmniVisionEncoderConfig,
+        "text_config": Qwen2_5OmniTextConfig,
+    }
+    is_composition = True
+
+    def __init__(
+        self,
+        audio_config=None,
+        vision_config=None,
+        text_config=None,
+        audio_token_index=151646,
+        image_token_index=151655,
+        video_token_index=151656,
+        tie_word_embeddings=False,
+        position_id_per_seconds=25,
+        seconds_per_chunk=2,
+        audio_start_token_id=151647,
+        audio_end_token_id=151648,
+        user_token_id=872,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.audio_token_index = audio_token_index
+        self.image_token_index = image_token_index
+        self.video_token_index = video_token_index
+        # 2025.02.20 the add
+        self.user_token_id = user_token_id
+        self.position_id_per_seconds = position_id_per_seconds
+        self.seconds_per_chunk = seconds_per_chunk
+        self.audio_start_token_id = audio_start_token_id
+        self.audio_end_token_id = audio_end_token_id
+        self.initializer_range = initializer_range
+
+        if isinstance(vision_config, dict):
+            vision_config = Qwen2_5OmniVisionEncoderConfig(**vision_config)
+        elif vision_config is None:
+            vision_config = Qwen2_5OmniVisionEncoderConfig()
+        self.vision_config = vision_config
+
+        if isinstance(audio_config, dict):
+            audio_config = Qwen2_5OmniAudioEncoderConfig(**audio_config)
+        elif audio_config is None:
+            audio_config = Qwen2_5OmniAudioEncoderConfig()
+        self.audio_config = audio_config
+
+        if isinstance(text_config, dict):
+            text_config = Qwen2_5OmniTextConfig(**text_config)
+        elif text_config is None:
+            text_config = Qwen2_5OmniTextConfig()
+        self.text_config = text_config
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+class Qwen2_5OmniTalkerConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_talker"
+    is_composition = False
+
+    def __init__(
+        self,
+        audio_token_index=151646,
+        image_token_index=151655,
+        video_token_index=151656,
+        vocab_size=8448,
+        tts_text_start_token_id=151860,
+        tts_text_end_token_id=151861,
+        tts_text_pad_token_id=151859,
+        tts_codec_start_token_id=8293,
+        tts_codec_end_token_id=8294,
+        tts_codec_pad_token_id=8292,
+        tts_codec_mask_token_id=8296,
+        vision_start_token_id=151652,
+        vision_end_token_id=151653,
+        embedding_size=3584,
+        hidden_size=3584,
+        intermediate_size=18944,
+        num_hidden_layers=28,
+        num_attention_heads=28,
+        num_key_value_heads=4,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        rms_norm_eps=1e-06,
+        head_dim=128,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=1000000.0,
+        use_sliding_window=False,
+        sliding_window=32768,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        rope_scaling=None,
+        position_id_per_seconds=25,
+        seconds_per_chunk=2,
+        audio_start_token_id=151647,
+        audio_end_token_id=151648,
+        initializer_range=0.02,
+        spatial_merge_size=2,
+        **kwargs,
+    ):
+        self.audio_token_index = audio_token_index
+        self.image_token_index = image_token_index
+        self.video_token_index = video_token_index
+
+        self.tts_text_start_token_id = tts_text_start_token_id
+        self.tts_text_end_token_id = tts_text_end_token_id
+        self.tts_text_pad_token_id = tts_text_pad_token_id
+        self.tts_codec_start_token_id = tts_codec_start_token_id
+        self.tts_codec_end_token_id = tts_codec_end_token_id
+        self.tts_codec_pad_token_id = tts_codec_pad_token_id
+
+        self.tts_codec_mask_token_id = tts_codec_mask_token_id
+
+        self.vision_start_token_id = vision_start_token_id
+        self.vision_end_token_id = vision_end_token_id
+
+        self.vocab_size = vocab_size
+        self.head_dim = head_dim
+        self.embedding_size = embedding_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.rope_scaling = rope_scaling
+        self.position_id_per_seconds = position_id_per_seconds  # zf
+        self.seconds_per_chunk = seconds_per_chunk  # zf
+        self.audio_start_token_id = audio_start_token_id  # zf
+        self.audio_end_token_id = audio_end_token_id  # zf
+
+        self.initializer_range = initializer_range
+        self.spatial_merge_size = spatial_merge_size
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+class Qwen2_5OmniDiTConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_dit"
+
+    def __init__(
+        self,
+        hidden_size=1024,
+        num_hidden_layers=22,
+        num_attention_heads=16,
+        ff_mult=2,
+        emb_dim=512,
+        head_dim=64,
+        rope_theta=10000.0,
+        max_position_embeddings=32768,
+        block_size=24,
+        look_ahead_layers=[10],
+        look_backward_layers=[0, 20],
+        repeats=2,
+        num_embeds=8193,
+        mel_dim=80,
+        dropout=0.1,
+        enc_emb_dim=192,
+        enc_dim=128,
+        enc_channels=[256, 256, 256, 256, 768],
+        enc_kernel_sizes=[5, 3, 3, 3, 1],
+        enc_dilations=[1, 2, 3, 4, 1],
+        enc_attention_channels=64,
+        enc_res2net_scale=2,
+        enc_se_channels=64,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.ff_mult = ff_mult
+        self.emb_dim = emb_dim
+        self.head_dim = head_dim
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.block_size = block_size
+        self.look_ahead_layers = look_ahead_layers
+        self.look_backward_layers = look_backward_layers
+        self.repeats = repeats
+        self.num_embeds = num_embeds
+        self.mel_dim = mel_dim
+        self.dropout = dropout
+        self.enc_emb_dim = enc_emb_dim
+        self.enc_dim = enc_dim
+        self.enc_channels = enc_channels
+        self.enc_kernel_sizes = enc_kernel_sizes
+        self.enc_dilations = enc_dilations
+        self.enc_attention_channels = enc_attention_channels
+        self.enc_res2net_scale = enc_res2net_scale
+        self.enc_se_channels = enc_se_channels
+        super().__init__(**kwargs)
+
+
+class Qwen2_5OmniBigVGANConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_bigvgan"
+
+    def __init__(
+        self,
+        mel_dim=80,
+        upsample_initial_channel=1536,
+        resblock_kernel_sizes=[3, 7, 11],
+        resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+        upsample_rates=[5, 3, 2, 2, 2, 2],
+        upsample_kernel_sizes=[11, 7, 4, 4, 4, 4],
+        **kwargs,
+    ):
+        self.mel_dim = mel_dim
+        self.upsample_initial_channel = upsample_initial_channel
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        super().__init__(**kwargs)
+
+
+class Qwen2_5OmniToken2WavConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni_token2wav"
+    sub_configs = {
+        "dit_config": Qwen2_5OmniDiTConfig,
+        "bigvgan_config": Qwen2_5OmniBigVGANConfig,
+    }
+    is_composition = True
+
+    def __init__(self, dit_config=None, bigvgan_config=None, **kwargs):
+        if dit_config is None:
+            dit_config = {}
+        if bigvgan_config is None:
+            bigvgan_config = {}
+        self.dit_config = Qwen2_5OmniDiTConfig(**dit_config)
+        self.bigvgan_config = Qwen2_5OmniBigVGANConfig(**bigvgan_config)
+        super().__init__(**kwargs)
+
+
+class Qwen2_5OmniConfig(PretrainedConfig):
+
+    model_type = "qwen2_5_omni"
+    sub_configs = {
+        "thinker_config": Qwen2_5OmniThinkerConfig,
+        "talker_config": Qwen2_5OmniTalkerConfig,
+        "token2wav_config": Qwen2_5OmniToken2WavConfig,
+    }
+    is_composition = True
+
+    def __init__(
+        self,
+        thinker_config=None,
+        talker_config=None,
+        token2wav_config=None,
+        enable_audio_output: bool = True,
+        **kwargs,
+    ):
+        if thinker_config is None:
+            thinker_config = {}
+            logger.info("thinker_config is None. Initializing thinker model with default values")
+
+        if talker_config is None:
+            talker_config = {}
+            logger.info("talker_config is None. Initializing talker model with default values")
+
+        if token2wav_config is None:
+            token2wav_config = {}
+            logger.info("token2wav_config is None. Initializing token2wav model with default values")
+
+        self.thinker_config = Qwen2_5OmniThinkerConfig(**thinker_config)
+        self.talker_config = Qwen2_5OmniTalkerConfig(**talker_config)
+        self.token2wav_config = Qwen2_5OmniToken2WavConfig(**token2wav_config)
+        self.enable_audio_output = enable_audio_output
+
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_sub_model_configs(
+        cls,
+        thinker_config: Qwen2_5OmniThinkerConfig,
+        talker_config: Qwen2_5OmniTalkerConfig,
+        token2wav_config: Qwen2_5OmniToken2WavConfig,
+        enable_audio_output: bool = True,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`Qwen2_5OmniConfig`] (or a derived class) from sub-models configuration.
+
+        Returns:
+            [`Qwen2_5OmniConfig`]: An instance of a configuration object
+        """
+        return cls(
+            thinker_config=thinker_config.to_dict(),
+            talker_config=talker_config.to_dict(),
+            token2wav_config=token2wav_config.to_dict(),
+            enable_audio_output=enable_audio_output,
+            **kwargs,
+        )
+
+
+__all__ = ["Qwen2_5OmniConfig", "Qwen2_5OmniThinkerConfig", "Qwen2_5OmniTalkerConfig", "Qwen2_5OmniToken2WavConfig"]