# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # Copyright 2025 The Qwen Team 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. """Qwen3.5-MoE model configuration""" from transformers.configuration_utils import PretrainedConfig, layer_type_validation class Qwen3_5MoeTextConfig(PretrainedConfig): model_type = "qwen3_5_moe_text" keys_to_ignore_at_inference = ["past_key_values"] base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.experts.gate_up_proj": "packed_colwise", "layers.*.mlp.experts.down_proj": "rowwise", "layers.*.mlp.shared_expert.gate_proj": "colwise", "layers.*.mlp.shared_expert.up_proj": "colwise", "layers.*.mlp.shared_expert.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } base_config_key = "text_config" def __init__( self, vocab_size=248320, hidden_size=2048, num_hidden_layers=40, num_attention_heads=16, num_key_value_heads=2, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_parameters=None, attention_bias=False, attention_dropout=0.0, head_dim=256, linear_conv_kernel_dim=4, linear_key_head_dim=128, linear_value_head_dim=128, linear_num_key_heads=16, linear_num_value_heads=32, moe_intermediate_size=512, shared_expert_intermediate_size=512, num_experts_per_tok=8, num_experts=256, output_router_logits=False, router_aux_loss_coef=0.001, layer_types=None, pad_token_id=None, bos_token_id=None, eos_token_id=None, **kwargs, ): kwargs["ignore_keys_at_rope_validation"] = [ "mrope_section", "mrope_interleaved", ] self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_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.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.attention_bias = attention_bias self.attention_dropout = attention_dropout self.head_dim = head_dim self.rope_parameters = rope_parameters kwargs.setdefault("partial_rotary_factor", 0.25) self.layer_types = layer_types if self.layer_types is None: interval_pattern = kwargs.get("full_attention_interval", 4) self.layer_types = [ "linear_attention" if bool((i + 1) % interval_pattern) else "full_attention" for i in range(self.num_hidden_layers) ] layer_type_validation(self.layer_types, self.num_hidden_layers) # linear attention part self.linear_conv_kernel_dim = linear_conv_kernel_dim self.linear_key_head_dim = linear_key_head_dim self.linear_value_head_dim = linear_value_head_dim self.linear_num_key_heads = linear_num_key_heads self.linear_num_value_heads = linear_num_value_heads self.moe_intermediate_size = moe_intermediate_size self.shared_expert_intermediate_size = shared_expert_intermediate_size self.num_experts_per_tok = num_experts_per_tok self.num_experts = num_experts self.output_router_logits = output_router_logits self.router_aux_loss_coef = router_aux_loss_coef super().__init__(**kwargs) # Set these AFTER super().__init__() because transformers v4's # PretrainedConfig.__init__ has these as explicit params with different # defaults (e.g. tie_word_embeddings=True) that would overwrite our values. self.pad_token_id = pad_token_id self.bos_token_id = bos_token_id self.eos_token_id = eos_token_id self.tie_word_embeddings = tie_word_embeddings class Qwen3_5MoeVisionConfig(PretrainedConfig): model_type = "qwen3_5_moe" base_config_key = "vision_config" 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, 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 class Qwen3_5MoeConfig(PretrainedConfig): model_type = "qwen3_5_moe" sub_configs = { "vision_config": Qwen3_5MoeVisionConfig, "text_config": Qwen3_5MoeTextConfig, } keys_to_ignore_at_inference = ["past_key_values"] def __init__( self, text_config=None, vision_config=None, image_token_id=248056, video_token_id=248057, vision_start_token_id=248053, vision_end_token_id=248054, tie_word_embeddings=False, **kwargs, ): if isinstance(vision_config, dict): self.vision_config = self.sub_configs["vision_config"](**vision_config) elif vision_config is None: self.vision_config = self.sub_configs["vision_config"]() if isinstance(text_config, dict): self.text_config = self.sub_configs["text_config"](**text_config) elif text_config is None: self.text_config = self.sub_configs["text_config"]() self.image_token_id = image_token_id self.video_token_id = video_token_id self.vision_start_token_id = vision_start_token_id self.vision_end_token_id = vision_end_token_id super().__init__(**kwargs) # Set after super().__init__() to avoid v4 PretrainedConfig overwrite self.tie_word_embeddings = tie_word_embeddings __all__ = ["Qwen3_5MoeConfig", "Qwen3_5MoeTextConfig"]