@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
py 205 lines 7.42 KB
Raw
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-MoE model configuration."""
18
19 from transformers.configuration_utils import PreTrainedConfig
20
21
22 class Qwen3_5MoeTextConfig(PreTrainedConfig):
23 model_type = "qwen3_5_moe_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.experts.gate_up_proj": "packed_colwise",
32 "layers.*.mlp.experts.down_proj": "rowwise",
33 "layers.*.mlp.shared_expert.gate_proj": "colwise",
34 "layers.*.mlp.shared_expert.up_proj": "colwise",
35 "layers.*.mlp.shared_expert.down_proj": "rowwise",
36 }
37 base_model_pp_plan = {
38 "embed_tokens": (["input_ids"], ["inputs_embeds"]),
39 "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
40 "norm": (["hidden_states"], ["hidden_states"]),
41 }
42 base_config_key = "text_config"
43
44 def __init__(
45 self,
46 vocab_size=248320,
47 hidden_size=2048,
48 num_hidden_layers=40,
49 num_attention_heads=16,
50 num_key_value_heads=2,
51 hidden_act="silu",
52 max_position_embeddings=32768,
53 initializer_range=0.02,
54 rms_norm_eps=1e-6,
55 use_cache=True,
56 tie_word_embeddings=False,
57 rope_parameters=None,
58 attention_bias=False,
59 attention_dropout=0.0,
60 head_dim=256,
61 linear_conv_kernel_dim=4,
62 linear_key_head_dim=128,
63 linear_value_head_dim=128,
64 linear_num_key_heads=16,
65 linear_num_value_heads=32,
66 moe_intermediate_size=512,
67 shared_expert_intermediate_size=512,
68 num_experts_per_tok=8,
69 num_experts=256,
70 output_router_logits=False,
71 router_aux_loss_coef=0.001,
72 layer_types=None,
73 pad_token_id=None,
74 bos_token_id=None,
75 eos_token_id=None,
76 **kwargs,
77 ):
78 self.vocab_size = vocab_size
79 self.max_position_embeddings = max_position_embeddings
80 self.hidden_size = hidden_size
81 self.num_hidden_layers = num_hidden_layers
82 self.num_attention_heads = num_attention_heads
83 self.num_key_value_heads = num_key_value_heads
84 self.hidden_act = hidden_act
85 self.initializer_range = initializer_range
86 self.rms_norm_eps = rms_norm_eps
87 self.use_cache = use_cache
88 self.attention_bias = attention_bias
89 self.attention_dropout = attention_dropout
90 self.head_dim = head_dim
91 self.rope_parameters = rope_parameters
92 kwargs.setdefault("partial_rotary_factor", 0.25)
93
94 self.layer_types = layer_types
95 if self.layer_types is None:
96 interval_pattern = kwargs.get("full_attention_interval", 4)
97 self.layer_types = [
98 "linear_attention"
99 if bool((i + 1) % interval_pattern)
100 else "full_attention"
101 for i in range(self.num_hidden_layers)
102 ]
103 kwargs["ignore_keys_at_rope_validation"] = {
104 "mrope_section",
105 "mrope_interleaved",
106 }
107 self.validate_layer_type()
108
109 # linear attention part
110 self.linear_conv_kernel_dim = linear_conv_kernel_dim
111 self.linear_key_head_dim = linear_key_head_dim
112 self.linear_value_head_dim = linear_value_head_dim
113 self.linear_num_key_heads = linear_num_key_heads
114 self.linear_num_value_heads = linear_num_value_heads
115 self.moe_intermediate_size = moe_intermediate_size
116 self.shared_expert_intermediate_size = shared_expert_intermediate_size
117 self.num_experts_per_tok = num_experts_per_tok
118 self.num_experts = num_experts
119 self.output_router_logits = output_router_logits
120 self.router_aux_loss_coef = router_aux_loss_coef
121 super().__init__(**kwargs)
122 # Set these AFTER super().__init__() because transformers v4's
123 # PreTrainedConfig.__init__ has these as explicit params with different
124 # defaults (e.g. tie_word_embeddings=True) that would overwrite our values.
125 self.pad_token_id = pad_token_id
126 self.bos_token_id = bos_token_id
127 self.eos_token_id = eos_token_id
128 self.tie_word_embeddings = tie_word_embeddings
129
130
131 class Qwen3_5MoeVisionConfig(PreTrainedConfig):
132 model_type = "qwen3_5_moe"
133 base_config_key = "vision_config"
134
135 def __init__(
136 self,
137 depth=27,
138 hidden_size=1152,
139 hidden_act="gelu_pytorch_tanh",
140 intermediate_size=4304,
141 num_heads=16,
142 in_channels=3,
143 patch_size=16,
144 spatial_merge_size=2,
145 temporal_patch_size=2,
146 out_hidden_size=3584,
147 num_position_embeddings=2304,
148 initializer_range=0.02,
149 **kwargs,
150 ):
151 super().__init__(**kwargs)
152
153 self.depth = depth
154 self.hidden_size = hidden_size
155 self.hidden_act = hidden_act
156 self.intermediate_size = intermediate_size
157 self.num_heads = num_heads
158 self.in_channels = in_channels
159 self.patch_size = patch_size
160 self.spatial_merge_size = spatial_merge_size
161 self.temporal_patch_size = temporal_patch_size
162 self.out_hidden_size = out_hidden_size
163 self.num_position_embeddings = num_position_embeddings
164 self.initializer_range = initializer_range
165
166
167 class Qwen3_5MoeConfig(PreTrainedConfig):
168 model_type = "qwen3_5_moe"
169 sub_configs = {
170 "vision_config": Qwen3_5MoeVisionConfig,
171 "text_config": Qwen3_5MoeTextConfig,
172 }
173 keys_to_ignore_at_inference = ["past_key_values"]
174
175 def __init__(
176 self,
177 text_config=None,
178 vision_config=None,
179 image_token_id=248056,
180 video_token_id=248057,
181 vision_start_token_id=248053,
182 vision_end_token_id=248054,
183 tie_word_embeddings=False,
184 **kwargs,
185 ):
186 if isinstance(vision_config, dict):
187 self.vision_config = self.sub_configs["vision_config"](**vision_config)
188 elif vision_config is None:
189 self.vision_config = self.sub_configs["vision_config"]()
190
191 if isinstance(text_config, dict):
192 self.text_config = self.sub_configs["text_config"](**text_config)
193 elif text_config is None:
194 self.text_config = self.sub_configs["text_config"]()
195
196 self.image_token_id = image_token_id
197 self.video_token_id = video_token_id
198 self.vision_start_token_id = vision_start_token_id
199 self.vision_end_token_id = vision_end_token_id
200 super().__init__(**kwargs)
201 # Set after super().__init__() to avoid v4 PreTrainedConfig overwrite
202 self.tie_word_embeddings = tie_word_embeddings
203
204
205 __all__ = ["Qwen3_5MoeConfig", "Qwen3_5MoeTextConfig"]