from __future__ import annotations from typing import Any from transformers import PretrainedConfig _MROPE_ROPE_SCALING_KEYS = frozenset( {"interleaved", "mrope_interleaved", "mrope_section"} ) def _normalize_rope_scaling( rope_scaling: dict[str, Any] | None ) -> dict[str, Any] | None: if rope_scaling is None: return None normalized = dict(rope_scaling) if "type" in normalized and "rope_type" not in normalized: normalized["rope_type"] = normalized["type"] if _MROPE_ROPE_SCALING_KEYS.intersection(normalized.keys()): # SGLang 0.5.12 selects MRotaryEmbedding from rope_type/type="default" # plus mrope_section. Passing "mrope" reaches the generic rope switch # and fails before the talker can start. rope_type = normalized.get("rope_type", normalized.get("type", "default")) if rope_type == "mrope": rope_type = "default" normalized["rope_type"] = rope_type if "type" in normalized: normalized["type"] = rope_type return normalized class Qwen3OmniMoeAudioEncoderConfig(PretrainedConfig): 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, n_window_infer=400, conv_chunksize=500, downsample_hidden_size=480, **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 self.n_window_infer = n_window_infer self.conv_chunksize = conv_chunksize self.downsample_hidden_size = downsample_hidden_size class Qwen3OmniMoeVisionEncoderConfig(PretrainedConfig): def __init__( self, depth=27, hidden_size=1152, hidden_act="gelu_pytorch_tanh", intermediate_size=4304, num_heads=16, in_channels=3, patch_size=16, spatial_merge_size=2, temporal_patch_size=2, out_hidden_size=3584, num_position_embeddings=2304, deepstack_visual_indexes=[8, 16, 24], tokens_per_second=None, initializer_range=0.02, **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.out_hidden_size = out_hidden_size self.num_position_embeddings = num_position_embeddings self.initializer_range = initializer_range self.deepstack_visual_indexes = deepstack_visual_indexes self.tokens_per_second = tokens_per_second class Qwen3OmniMoeTextConfig(PretrainedConfig): def __init__( self, vocab_size=3584, hidden_size=2048, intermediate_size=18944, num_hidden_layers=28, num_attention_heads=28, num_key_value_heads=4, head_dim=None, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_theta=1000000.0, rope_scaling=None, partial_rotary_factor=1.0, attention_bias=False, sliding_window=None, attention_dropout=0, dual_chunk_attention_config=None, decoder_sparse_step=1, moe_intermediate_size=768, num_experts_per_tok=8, num_experts=128, norm_topk_prob=True, output_router_logits=False, router_aux_loss_coef=0.001, mlp_only_layers=None, **kwargs, ): super().__init__( tie_word_embeddings=tie_word_embeddings, **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.head_dim = ( head_dim if head_dim is not None else hidden_size // num_attention_heads ) self.sliding_window = sliding_window self.num_key_value_heads = num_key_value_heads self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = _normalize_rope_scaling(rope_scaling) self.partial_rotary_factor = partial_rotary_factor self.attention_bias = attention_bias self.attention_dropout = attention_dropout self.dual_chunk_attention_config = dual_chunk_attention_config # MoE arguments self.decoder_sparse_step = decoder_sparse_step self.moe_intermediate_size = moe_intermediate_size self.num_experts_per_tok = num_experts_per_tok self.num_experts = num_experts self.norm_topk_prob = norm_topk_prob self.output_router_logits = output_router_logits self.router_aux_loss_coef = router_aux_loss_coef self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers class Qwen3OmniMoeThinkerConfig(PretrainedConfig): def __init__( self, audio_config=None, vision_config=None, text_config=None, audio_token_id=151646, image_token_id=151655, video_token_id=151656, vision_start_token_id=151652, position_id_per_seconds=25, audio_start_token_id=151647, user_token_id=872, initializer_range=0.02, **kwargs, ): super().__init__(**kwargs) self.user_token_id = user_token_id self.vision_start_token_id = vision_start_token_id self.position_id_per_seconds = position_id_per_seconds self.audio_start_token_id = audio_start_token_id self.initializer_range = initializer_range if isinstance(vision_config, dict): vision_config = Qwen3OmniMoeVisionEncoderConfig(**vision_config) elif vision_config is None: vision_config = Qwen3OmniMoeVisionEncoderConfig() self.vision_config = vision_config if isinstance(audio_config, dict): audio_config = Qwen3OmniMoeAudioEncoderConfig(**audio_config) elif audio_config is None: audio_config = Qwen3OmniMoeAudioEncoderConfig() self.audio_config = audio_config if isinstance(text_config, dict): text_config = Qwen3OmniMoeTextConfig(**text_config) elif text_config is None: text_config = Qwen3OmniMoeTextConfig() self.text_config = text_config self.audio_token_id = audio_token_id self.image_token_id = image_token_id self.video_token_id = video_token_id # --------------------------------------------------------------------------- # Talker configs # --------------------------------------------------------------------------- class Qwen3OmniMoeTalkerTextConfig(Qwen3OmniMoeTextConfig): """Text config for the talker MoE backbone (20-layer, shared expert). Inherits from Qwen3OmniMoeTextConfig, adds shared_expert_intermediate_size. """ model_type = "qwen3_omni_moe_talker_text" def __init__( self, vocab_size=3072, hidden_size=1024, intermediate_size=2048, num_hidden_layers=20, num_attention_heads=16, num_key_value_heads=2, head_dim=128, hidden_act="silu", max_position_embeddings=65536, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_theta=1000000.0, rope_scaling=None, attention_bias=False, sliding_window=None, attention_dropout=0, decoder_sparse_step=1, moe_intermediate_size=384, shared_expert_intermediate_size=768, # Talker-specific num_experts_per_tok=6, num_experts=128, norm_topk_prob=True, output_router_logits=False, router_aux_loss_coef=0.001, mlp_only_layers=None, **kwargs, ): # Call parent (Qwen3OmniMoeTextConfig) with all standard MoE params super().__init__( vocab_size=vocab_size, hidden_size=hidden_size, intermediate_size=intermediate_size, num_hidden_layers=num_hidden_layers, num_attention_heads=num_attention_heads, num_key_value_heads=num_key_value_heads, hidden_act=hidden_act, max_position_embeddings=max_position_embeddings, initializer_range=initializer_range, rms_norm_eps=rms_norm_eps, use_cache=use_cache, tie_word_embeddings=tie_word_embeddings, rope_theta=rope_theta, rope_scaling=rope_scaling, attention_bias=attention_bias, sliding_window=sliding_window, attention_dropout=attention_dropout, decoder_sparse_step=decoder_sparse_step, moe_intermediate_size=moe_intermediate_size, num_experts_per_tok=num_experts_per_tok, num_experts=num_experts, norm_topk_prob=norm_topk_prob, output_router_logits=output_router_logits, router_aux_loss_coef=router_aux_loss_coef, mlp_only_layers=mlp_only_layers, **kwargs, ) # Add Talker-specific field self.head_dim = head_dim self.shared_expert_intermediate_size = shared_expert_intermediate_size class Qwen3OmniMoeTalkerCodePredictorConfig(PretrainedConfig): """Config for the CodePredictor (5-layer dense transformer).""" model_type = "qwen3_omni_moe_talker_code_predictor" def __init__( self, vocab_size=2048, hidden_size=1024, intermediate_size=3072, num_hidden_layers=5, num_attention_heads=16, num_key_value_heads=8, head_dim=128, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_theta=1000000.0, rope_scaling=None, attention_bias=False, sliding_window=None, attention_dropout=0, num_code_groups=16, **kwargs, ): super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs) self.vocab_size = vocab_size 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.num_key_value_heads = num_key_value_heads self.head_dim = head_dim self.hidden_act = hidden_act self.max_position_embeddings = max_position_embeddings self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = _normalize_rope_scaling(rope_scaling) self.attention_bias = attention_bias self.sliding_window = sliding_window self.attention_dropout = attention_dropout self.num_code_groups = num_code_groups class Qwen3OmniMoeTalkerConfig(PretrainedConfig): """Top-level talker config combining text model + code predictor.""" model_type = "qwen3_omni_moe_talker" def __init__( self, text_config: dict | Qwen3OmniMoeTalkerTextConfig | None = None, code_predictor_config: ( dict | Qwen3OmniMoeTalkerCodePredictorConfig | None ) = None, num_code_groups=16, thinker_hidden_size=2048, accept_hidden_layer=24, codec_eos_token_id=2150, codec_nothink_id=2155, codec_think_bos_id=2156, codec_think_eos_id=2157, codec_pad_id=2148, codec_bos_id=2149, **kwargs, ): super().__init__(**kwargs) if text_config is None: self.text_config = Qwen3OmniMoeTalkerTextConfig() elif isinstance(text_config, dict): self.text_config = Qwen3OmniMoeTalkerTextConfig(**text_config) else: self.text_config = text_config if code_predictor_config is None: self.code_predictor_config = Qwen3OmniMoeTalkerCodePredictorConfig() elif isinstance(code_predictor_config, dict): self.code_predictor_config = Qwen3OmniMoeTalkerCodePredictorConfig( **code_predictor_config ) else: self.code_predictor_config = code_predictor_config self.num_code_groups = num_code_groups self.thinker_hidden_size = thinker_hidden_size self.accept_hidden_layer = accept_hidden_layer self.codec_eos_token_id = codec_eos_token_id self.codec_nothink_id = codec_nothink_id self.codec_think_bos_id = codec_think_bos_id self.codec_think_eos_id = codec_think_eos_id self.codec_pad_id = codec_pad_id self.codec_bos_id = codec_bos_id