Skip to content

vllm_omni.transformers_utils.configs.ming3_audio

Native Hugging Face configs for inclusionAI/Ming-3.0-audio-flash.

BailingMoeV3AudioConfig

Bases: PretrainedConfig

audio_config instance-attribute 

audio_config = (
    audio_config or BailingMoeV3AudioEncoderConfig()
)

audio_token_id instance-attribute 

audio_token_id = audio_token_id

is_composition class-attribute instance-attribute 

is_composition = True

model_type class-attribute instance-attribute 

model_type = 'bailing_moe_v3_audio'

norm_query_embeds instance-attribute 

norm_query_embeds = norm_query_embeds

sub_configs class-attribute instance-attribute 

sub_configs: ClassVar = {
    "text_config": BailingMoeV3TextConfig,
    "audio_config": BailingMoeV3AudioEncoderConfig,
}

text_config instance-attribute 

text_config = text_config or BailingMoeV3TextConfig()

get_text_config 

get_text_config(decoder: bool = False) -> PretrainedConfig

BailingMoeV3AudioEncoderConfig

Bases: PretrainedConfig

downsampler_kernel_size_list instance-attribute 

downsampler_kernel_size_list = (
    downsampler_kernel_size_list or [3]
)

downsampler_padding_list instance-attribute 

downsampler_padding_list = downsampler_padding_list or [
    kernel // 2
    for kernel in self.downsampler_kernel_size_list
]

downsampler_stride_list instance-attribute 

downsampler_stride_list = downsampler_stride_list or [2]

model_type class-attribute instance-attribute 

model_type = 'bailing_moe_v3_audio_encoder'

n_audio_ctx instance-attribute 

n_audio_ctx = n_audio_ctx

n_audio_head instance-attribute 

n_audio_head = n_audio_head

n_audio_layer instance-attribute 

n_audio_layer = n_audio_layer

n_audio_state instance-attribute 

n_audio_state = n_audio_state

n_mels instance-attribute 

n_mels = n_mels

whisper_chunk_size instance-attribute 

whisper_chunk_size = whisper_chunk_size

BailingMoeV3TextConfig

Bases: PretrainedConfig

Native config adapter for the Bailing MoE v3 text backbone.

The released checkpoint deliberately stores an empty model_type in text_config. Normalize it to vLLM's canonical Bailing hybrid type so the shared architecture converter recognizes MLA and sizes its 576-wide latent KV cache correctly.

model_type class-attribute instance-attribute 

model_type = 'bailing_hybrid'