LLM面面观之MoE

1. 背景

根据本qiang~最新的趋势观察，基于MoE架构的开源大模型越来越多，比如马斯克的Grok-1(314B), Qwen1.5-MoE-A2.7B等，因此想探究一下MoE里面的部分细节。

此文是本qiang~针对大语言模型的MoE的整理，包括原理、流程及部分源码。

2. MoE原理

MoE的流行源于”欧洲的OpenAI” Mistral AI发布的论文及模型《Mixtral of Experts》，评测集上的效果吊打众多开源模型，如Llama 2 70B和GPT3.5。

《Mixtral of Experts》基础模型使用的是Mistral AI自研的Mistral 7B，该模型的特点包括：滑窗注意力(Sliding Window Aattention), 滚动缓冲区缓存(Rolling Buffer Cache)以及预填充-分块(Pre-fill and Chunking)，具体细节可以查阅文末的论文地址。

本文以《Mixtral of Experts》为引子，探究MoE的相关细节，MoE的原理如下图所示：

图2.1 MoE的原理

(1) Transformers架构中的每一层中的FFN网络均替换为了8个FFN(专家)，且由一个网关路由(gate router)进行控制

(2) 针对每一个token，每一层的网关路由仅选择其中的2个FFN(专家)来处理当前状态并进行加权输出

(3) 结果就是，每一个token访问了47B参数，但是在推理阶段仅仅使用了13B的激活参数(即，只使用2个专家，冻结其他6个专家)。

(4) 与Dropout机制对比，Dropout让部分神经元失活，而MoE是让部分专家失活。

3. 源码

本qiang~研读并尝试执行了Mistral官网的github推理代码，该代码框架非常适合新手，无他，只因其几乎只是在torch上层做的封装，很少引擎其他第三方库，不像transformers，功能强大，但不适合新手研读代码…

为了普适性，下面的代码截取了transformers框架中的代码。

首先看下通用Transformers中FFN中的代码模块，代码位置在transformers.models.mistral.modeling_mistral, 主要流程是：

(1) 先经过gate_proj和up_proj的2个[hidden_size, intermediate_size]的线性转换

(2) 使用激活函数对gate_proj进行激活

(3) 二者的内积再经过down_proj线性转换。

class MistralMLP(nn.Module):def __init__(self, config):super().__init__()self.config = configself.hidden_size = config.hidden_sizeself.intermediate_size = config.intermediate_sizeself.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)self.act_fn = ACT2FN[config.hidden_act]def forward(self, x):return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))

再来看下MoE中的专家模块，代码位置在transformers.models.mixtral.modeling_mixtral，主要流程是：

(1) 首先经过网关路由self.gate

(2) 然后选择其中2个专家，并归一化

(3) 之后遍历每个专家网络，并按照expert_mask进行筛选

(4) 如果expert_mask有值，则选择指定部分的隐藏层进行FFN操作，且输出结果进行加权

(5) 最后原地增加先前初始化的最终结果变量final_hidden_states

class MixtralSparseMoeBlock(nn.Module):def __init__(self, config):super().__init__()self.hidden_dim = config.hidden_sizeself.ffn_dim = config.intermediate_sizeself.num_experts = config.num_local_expertsself.top_k = config.num_experts_per_tok# gatingself.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)self.experts = nn.ModuleList([MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)])def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:""" """batch_size, sequence_length, hidden_dim = hidden_states.shapehidden_states = hidden_states.view(-1, hidden_dim)# router_logits: (batch * sequence_length, n_experts)router_logits = self.gate(hidden_states)routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)routing_weights /= routing_weights.sum(dim=-1, keepdim=True)# we cast back to the input dtyperouting_weights = routing_weights.to(hidden_states.dtype)final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device)# One hot encode the selected experts to create an expert mask# this will be used to easily index which expert is going to be sollicitatedexpert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)# Loop over all available experts in the model and perform the computation on each expertfor expert_idx in range(self.num_experts):expert_layer = self.experts[expert_idx]idx, top_x = torch.where(expert_mask[expert_idx])if top_x.shape[0] == 0:continue# in torch it is faster to index using lists than torch tensorstop_x_list = top_x.tolist()idx_list = idx.tolist()# Index the correct hidden states and compute the expert hidden state for# the current expert. We need to make sure to multiply the output hidden# states by `routing_weights` on the corresponding tokens (top-1 and top-2)current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]# However `index_add_` only support torch tensors for indexing so we'll use# the `top_x` tensor here.final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)return final_hidden_states, router_logits

其中MixtralBlockSparseTop2MLP代码如下，可以看到和传统MistralMLP内容完全一致。

class MixtralBlockSparseTop2MLP(nn.Module):def __init__(self, config: MixtralConfig):super().__init__()self.ffn_dim = config.intermediate_sizeself.hidden_dim = config.hidden_sizeself.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)self.act_fn = ACT2FN[config.hidden_act]def forward(self, hidden_states):current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)current_hidden_states = self.w2(current_hidden_states)return current_hidden_states