| # Copyright 2024 Bytedance Ltd. and/or its affiliates | |
| # | |
| # 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. | |
| # To support different vLLM versions, we add the model into SUPPORTED_MOE_MODELS separately to avoid triggering unsupported issues. | |
| SUPPORTED_MOE_MODELS = [] | |
| try: | |
| from vllm.model_executor.models.deepseek_v2 import DeepseekV2ForCausalLM, DeepseekV3ForCausalLM | |
| SUPPORTED_MOE_MODELS.append(DeepseekV2ForCausalLM) | |
| SUPPORTED_MOE_MODELS.append(DeepseekV3ForCausalLM) | |
| except ImportError: | |
| pass | |
| try: | |
| from vllm.model_executor.models.mixtral import MixtralForCausalLM | |
| SUPPORTED_MOE_MODELS.append(MixtralForCausalLM) | |
| except ImportError: | |
| pass | |
| try: | |
| from vllm.model_executor.models.qwen2_moe import Qwen2MoeForCausalLM | |
| SUPPORTED_MOE_MODELS.append(Qwen2MoeForCausalLM) | |
| except ImportError: | |
| pass | |
| try: | |
| from vllm.model_executor.models.qwen3_moe import Qwen3MoeForCausalLM | |
| SUPPORTED_MOE_MODELS.append(Qwen3MoeForCausalLM) | |
| except ImportError: | |
| pass | |
| def patch_vllm_moe_model_weight_loader(model): | |
| # this is a work around to load the weight of vllm fused moe model | |
| # it is from a bug from vllm 0.8.2 | |
| # all the weights are supposed to have a weight_loader, but the moe weights | |
| # do not have a weight_loader, so we need to patch it | |
| # (True, 'model.embed_tokens.weight') | |
| # (True, 'model.layers.0.self_attn.qkv_proj.weight') | |
| # (True, 'model.layers.0.self_attn.qkv_proj.bias') | |
| # (True, 'model.layers.0.self_attn.o_proj.weight') | |
| # (True, 'model.layers.0.mlp.gate.weight') | |
| # (True, 'model.layers.0.mlp.shared_expert.gate_up_proj.weight') | |
| # (True, 'model.layers.0.mlp.shared_expert.down_proj.weight') | |
| # (False, 'model.layers.0.mlp.shared_expert_gate.weight') use default | |
| # (False, 'model.layers.0.input_layernorm.weight') use default | |
| # (False, 'model.layers.0.post_attention_layernorm.weight') use default | |
| # (False, 'model.layers.0.mlp.experts.w13_weight') use mlp.experts.weight_loader | |
| # (False, 'model.layers.0.mlp.experts.w2_weight') use mlp.experts.weight_loader | |
| # Define MLP attribute mapping for different model types | |
| MLP_ATTR_MAPPING = { | |
| MixtralForCausalLM: "block_sparse_moe", | |
| } | |
| DEFAULT_MLP_ATTR = "mlp" | |
| if not isinstance(model, tuple(SUPPORTED_MOE_MODELS)): | |
| return | |
| for layer in model.model.layers: | |
| mlp_attr = MLP_ATTR_MAPPING.get(type(model), DEFAULT_MLP_ATTR) | |
| mlp = getattr(layer, mlp_attr) | |
| param_dict = dict(mlp.named_parameters()) | |
| for name, param in param_dict.items(): | |
| if "w13_weight" in name or "w2_weight" in name: | |
| param.weight_loader = mlp.experts.weight_loader | |