vllm_omni.transformers_utils.configs.mammoth_moda2 ¶

architectures = ['Mammothmoda2Model']

gen_axes_dim_rope = gen_axes_dim_rope or [40, 40, 40]

gen_axes_lens = gen_axes_lens or [10000, 10000, 10000]

gen_condition_mode = gen_condition_mode

gen_dit_config = gen_dit_config

gen_image_condition_refiner_config = (
    gen_image_condition_refiner_config
)

gen_transport_config = gen_transport_config or {}

gen_vae_config = gen_vae_config

image_token_id: int

initializer_range = initializer_range

is_composition = True

llm_config = (
    AutoConfig.for_model(**llm_config)
    if llm_config is not None
    else None
)

model_type = 'mammothmoda2'

sub_configs: ClassVar = {'llm_config': AutoConfig}

tokenizer_class = 'MammothUTokenizer'

video_token_id: int

vision_config

vision_end_token_id: int

vision_start_token_id: int

extra_gen_vocab = getattr(
    self.text_config, "extra_gen_vocab", extra_gen_vocab
)

gen_vocab_size = getattr(
    self.text_config, "gen_vocab_size", gen_vocab_size
)

gen_vocab_start_index = getattr(
    self.text_config,
    "gen_vocab_start_index",
    gen_vocab_start_index,
)

image_token_id = image_token_id

keys_to_ignore_at_inference = ['past_key_values']

model_type = 'mammothmoda2_qwen2_5_vl'

moe_type = getattr(self.text_config, 'moe_type', moe_type)

sub_configs = {
    "vision_config": Mammothmoda2Qwen2_5_VLVisionConfig,
    "text_config": Mammothmoda2Qwen2_5_VLTextConfig,
}

text_config = text_config

tokenizer_class = 'MammothUTokenizer'

video_token_id = video_token_id

vision_config = vision_config

base_config_key = 'text_config'

base_model_pp_plan = Qwen2_5_VLTextConfig.base_model_pp_plan

base_model_tp_plan = Qwen2_5_VLTextConfig.base_model_tp_plan

extra_gen_vocab = extra_gen_vocab

gen_vocab_size = gen_vocab_size

gen_vocab_start_index = (
    self.vocab_size
    if self.extra_gen_vocab
    else self.vocab_size - self.gen_vocab_size
)

image_token_id = image_token_id

keys_to_ignore_at_inference = ['past_key_values']

model_type = 'mammothmoda2_qwen2_5_vl_text'

moe_type = moe_type

video_token_id = video_token_id

vocab_size = int(self.gen_vocab_start_index) + int(
    self.gen_vocab_size
)

base_config_key = 'vision_config'

model_type = 'mammothmoda2_qwen2_5_vl_vision'