Skip to content

vllm_omni.transformers_utils.configs.yue2

HuggingFace config for m-a-p/YuE2-3B.

The checkpoint's own config.json already carries every field the AR backbone needs (it is Qwen3-1.7B-shaped with q_norm/k_norm and an extended 184,704-entry vocabulary), plus the acoustic-side fields (latent_dim, max_latent_frames, timestep_shift). Registering this class lets vLLM build a ModelConfig without trust_remote_code: the package's own Yue2ForCausalLM replaces the repository's auto_map modeling code, and the deploy YAML pins hf_overrides.architectures to it.

Yue2Config

Bases: PretrainedConfig

Config for the YuE2-3B AR–NAR Mixture-of-Transformers checkpoint.

attention_bias instance-attribute 

attention_bias = bool(kwargs.pop('attention_bias', False))

attention_dropout instance-attribute 

attention_dropout = float(
    kwargs.pop("attention_dropout", 0.0)
)

head_dim instance-attribute 

head_dim = head_dim

hidden_act instance-attribute 

hidden_act = str(kwargs.pop('hidden_act', 'silu'))

hidden_size instance-attribute 

hidden_size = hidden_size

intermediate_size instance-attribute 

intermediate_size = intermediate_size

latent_dim instance-attribute 

latent_dim = latent_dim

max_latent_frames instance-attribute 

max_latent_frames = max_latent_frames

max_position_embeddings instance-attribute 

max_position_embeddings = max_position_embeddings

max_window_layers instance-attribute 

max_window_layers = int(kwargs.pop('max_window_layers', 0))

model_type class-attribute instance-attribute 

model_type = 'yue2'

num_attention_heads instance-attribute 

num_attention_heads = num_attention_heads

num_hidden_layers instance-attribute 

num_hidden_layers = num_hidden_layers

num_key_value_heads instance-attribute 

num_key_value_heads = num_key_value_heads

rms_norm_eps instance-attribute 

rms_norm_eps = rms_norm_eps

rope_theta instance-attribute 

rope_theta = rope_theta

sliding_window instance-attribute 

sliding_window = kwargs.pop('sliding_window', None)

timestep_shift instance-attribute 

timestep_shift = timestep_shift

use_sliding_window instance-attribute 

use_sliding_window = bool(
    kwargs.pop("use_sliding_window", False)
)

vocab_size instance-attribute 

vocab_size = vocab_size