| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright 2025 The Qwen Team and The HuggingFace Inc. team. |
| 4 | # All rights reserved. |
| 5 | # |
| 6 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 7 | # you may not use this file except in compliance with the License. |
| 8 | # You may obtain a copy of the License at |
| 9 | # |
| 10 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 11 | # |
| 12 | # Unless required by applicable law or agreed to in writing, software |
| 13 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 14 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 15 | # See the License for the specific language governing permissions and |
| 16 | # limitations under the License. |
| 17 | """Qwen3.5 model configuration.""" |
| 18 | |
| 19 | from transformers.configuration_utils import PreTrainedConfig |
| 20 | |
| 21 | |
| 22 | class Qwen3_5TextConfig(PreTrainedConfig): |
| 23 | model_type = "qwen3_5_text" |
| 24 | keys_to_ignore_at_inference = ["past_key_values"] |
| 25 | |
| 26 | base_model_tp_plan = { |
| 27 | "layers.*.self_attn.q_proj": "colwise", |
| 28 | "layers.*.self_attn.k_proj": "colwise", |
| 29 | "layers.*.self_attn.v_proj": "colwise", |
| 30 | "layers.*.self_attn.o_proj": "rowwise", |
| 31 | "layers.*.mlp.gate_proj": "colwise", |
| 32 | "layers.*.mlp.up_proj": "colwise", |
| 33 | "layers.*.mlp.down_proj": "rowwise", |
| 34 | } |
| 35 | base_model_pp_plan = { |
| 36 | "embed_tokens": (["input_ids"], ["inputs_embeds"]), |
| 37 | "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), |
| 38 | "norm": (["hidden_states"], ["hidden_states"]), |
| 39 | } |
| 40 | base_config_key = "text_config" |
| 41 | |
| 42 | def __init__( |
| 43 | self, |
| 44 | vocab_size=248320, |
| 45 | hidden_size=4096, |
| 46 | intermediate_size=12288, |
| 47 | num_hidden_layers=32, |
| 48 | num_attention_heads=16, |
| 49 | num_key_value_heads=4, |
| 50 | hidden_act="silu", |
| 51 | max_position_embeddings=32768, |
| 52 | initializer_range=0.02, |
| 53 | rms_norm_eps=1e-6, |
| 54 | use_cache=True, |
| 55 | tie_word_embeddings=False, |
| 56 | rope_parameters=None, |
| 57 | attention_bias=False, |
| 58 | attention_dropout=0.0, |
| 59 | head_dim=256, |
| 60 | linear_conv_kernel_dim=4, |
| 61 | linear_key_head_dim=128, |
| 62 | linear_value_head_dim=128, |
| 63 | linear_num_key_heads=16, |
| 64 | linear_num_value_heads=32, |
| 65 | layer_types=None, |
| 66 | pad_token_id=None, |
| 67 | bos_token_id=None, |
| 68 | eos_token_id=None, |
| 69 | **kwargs, |
| 70 | ): |
| 71 | self.vocab_size = vocab_size |
| 72 | self.max_position_embeddings = max_position_embeddings |
| 73 | self.hidden_size = hidden_size |
| 74 | self.intermediate_size = intermediate_size |
| 75 | self.num_hidden_layers = num_hidden_layers |
| 76 | self.num_attention_heads = num_attention_heads |
| 77 | self.num_key_value_heads = num_key_value_heads |
| 78 | self.hidden_act = hidden_act |
| 79 | self.initializer_range = initializer_range |
| 80 | self.rms_norm_eps = rms_norm_eps |
| 81 | self.use_cache = use_cache |
| 82 | self.attention_bias = attention_bias |
| 83 | self.attention_dropout = attention_dropout |
| 84 | self.head_dim = head_dim |
| 85 | self.rope_parameters = rope_parameters |
| 86 | kwargs.setdefault("partial_rotary_factor", 0.25) |
| 87 | |
| 88 | self.layer_types = layer_types |
| 89 | if self.layer_types is None: |
| 90 | interval_pattern = kwargs.get("full_attention_interval", 4) |
| 91 | self.layer_types = [ |
| 92 | "linear_attention" |
| 93 | if bool((i + 1) % interval_pattern) |
| 94 | else "full_attention" |
| 95 | for i in range(self.num_hidden_layers) |
| 96 | ] |
| 97 | kwargs["ignore_keys_at_rope_validation"] = { |
| 98 | "mrope_section", |
| 99 | "mrope_interleaved", |
| 100 | } |
| 101 | self.validate_layer_type() |
| 102 | |
| 103 | # linear attention part |
| 104 | self.linear_conv_kernel_dim = linear_conv_kernel_dim |
| 105 | self.linear_key_head_dim = linear_key_head_dim |
| 106 | self.linear_value_head_dim = linear_value_head_dim |
| 107 | self.linear_num_key_heads = linear_num_key_heads |
| 108 | self.linear_num_value_heads = linear_num_value_heads |
| 109 | super().__init__(**kwargs) |
| 110 | # Set these AFTER super().__init__() because transformers v4's |
| 111 | # PreTrainedConfig.__init__ has these as explicit params with different |
| 112 | # defaults (e.g. tie_word_embeddings=True) that would overwrite our values. |
| 113 | self.pad_token_id = pad_token_id |
| 114 | self.bos_token_id = bos_token_id |
| 115 | self.eos_token_id = eos_token_id |
| 116 | self.tie_word_embeddings = tie_word_embeddings |
| 117 | |
| 118 | |
| 119 | class Qwen3_5VisionConfig(PreTrainedConfig): |
| 120 | model_type = "qwen3_5" |
| 121 | base_config_key = "vision_config" |
| 122 | |
| 123 | def __init__( |
| 124 | self, |
| 125 | depth=27, |
| 126 | hidden_size=1152, |
| 127 | hidden_act="gelu_pytorch_tanh", |
| 128 | intermediate_size=4304, |
| 129 | num_heads=16, |
| 130 | in_channels=3, |
| 131 | patch_size=16, |
| 132 | spatial_merge_size=2, |
| 133 | temporal_patch_size=2, |
| 134 | out_hidden_size=3584, |
| 135 | num_position_embeddings=2304, |
| 136 | initializer_range=0.02, |
| 137 | **kwargs, |
| 138 | ): |
| 139 | super().__init__(**kwargs) |
| 140 | |
| 141 | self.depth = depth |
| 142 | self.hidden_size = hidden_size |
| 143 | self.hidden_act = hidden_act |
| 144 | self.intermediate_size = intermediate_size |
| 145 | self.num_heads = num_heads |
| 146 | self.in_channels = in_channels |
| 147 | self.patch_size = patch_size |
| 148 | self.spatial_merge_size = spatial_merge_size |
| 149 | self.temporal_patch_size = temporal_patch_size |
| 150 | self.out_hidden_size = out_hidden_size |
| 151 | self.num_position_embeddings = num_position_embeddings |
| 152 | self.initializer_range = initializer_range |
| 153 | |
| 154 | |
| 155 | class Qwen3_5Config(PreTrainedConfig): |
| 156 | model_type = "qwen3_5" |
| 157 | sub_configs = { |
| 158 | "vision_config": Qwen3_5VisionConfig, |
| 159 | "text_config": Qwen3_5TextConfig, |
| 160 | } |
| 161 | keys_to_ignore_at_inference = ["past_key_values"] |
| 162 | |
| 163 | def __init__( |
| 164 | self, |
| 165 | text_config=None, |
| 166 | vision_config=None, |
| 167 | image_token_id=248056, |
| 168 | video_token_id=248057, |
| 169 | vision_start_token_id=248053, |
| 170 | vision_end_token_id=248054, |
| 171 | tie_word_embeddings=False, |
| 172 | **kwargs, |
| 173 | ): |
| 174 | if isinstance(vision_config, dict): |
| 175 | self.vision_config = self.sub_configs["vision_config"](**vision_config) |
| 176 | elif vision_config is None: |
| 177 | self.vision_config = self.sub_configs["vision_config"]() |
| 178 | |
| 179 | if isinstance(text_config, dict): |
| 180 | self.text_config = self.sub_configs["text_config"](**text_config) |
| 181 | elif text_config is None: |
| 182 | self.text_config = self.sub_configs["text_config"]() |
| 183 | |
| 184 | self.image_token_id = image_token_id |
| 185 | self.video_token_id = video_token_id |
| 186 | self.vision_start_token_id = vision_start_token_id |
| 187 | self.vision_end_token_id = vision_end_token_id |
| 188 | super().__init__(**kwargs) |
| 189 | # Set after super().__init__() to avoid v4 PreTrainedConfig overwrite |
| 190 | self.tie_word_embeddings = tie_word_embeddings |
| 191 | |
| 192 | |
| 193 | __all__ = ["Qwen3_5Config", "Qwen3_5TextConfig"] |