import math import struct import inspect from .LMConfig import LMConfig from typing import Any, Optional, Tuple import numpy as np import torch import torch.nn.functional as F from torch import nn from transformers import PreTrainedModel from transformers.modeling_outputs import CausalLMOutputWithPast # 定义 RMSNorm 类，实现一种归一化方法，类似于 LayerNorm，但计算方式不同 class RMSNorm(torch.nn.Module): def __init__(self, dim: int, eps: float): super().__init__() self.eps = eps # 设置 epsilon，防止除零错误 self.weight = nn.Parameter(torch.ones(dim)) # 初始化权重参数 def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) # 计算 RMSNorm def forward(self, x): output = self._norm(x.float()).type_as(x) # 应用 RMSNorm return output * self.weight # 乘以权重参数 # 定义 precompute_pos_cis 函数，用于预计算位置编码的复数形式 def precompute_pos_cis(dim: int, end: int, theta: float = 10000.0): freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) # 计算频率 t = torch.arange(end, device=freqs.device) # 生成时间序列 freqs = torch.outer(t, freqs).float() # 计算外积 pos_cis = torch.polar(torch.ones_like(freqs), freqs) # 计算复数形式的位置编码 return pos_cis # 定义 apply_rotary_emb 函数，用于应用旋转位置编码 def apply_rotary_emb(xq, xk, pos_cis): def unite_shape(pos_cis, x): ndim = x.ndim assert 0 <= 1 < ndim assert pos_cis.shape == (x.shape[1], x.shape[-1]) shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] return pos_cis.view(*shape) xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) # 将 xq 转换为复数形式 xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) # 将 xk 转换为复数形式 pos_cis = unite_shape(pos_cis, xq_) # 调整 pos_cis 的形状 xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3) # 应用旋转位置编码 xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3) # 应用旋转位置编码 return xq_out.type_as(xq), xk_out.type_as(xk) # 返回结果 # 定义 repeat_kv 函数，用于重复 KV 头的值 def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor: """torch.repeat_interleave(x, dim=2, repeats=n_rep)""" bs, slen, n_kv_heads, head_dim = x.shape if n_rep == 1: return x return ( x[:, :, :, None, :] .expand(bs, slen, n_kv_heads, n_rep, head_dim) .reshape(bs, slen, n_kv_heads * n_rep, head_dim) ) # 定义 Attention 类，实现自注意力机制 class Attention(nn.Module): def __init__(self, args: LMConfig): super().__init__() self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads # 设置 KV 头的数量 assert args.n_heads % self.n_kv_heads == 0 # 确保 KV 头的数量是总头数的因数 self.n_local_heads = args.n_heads # 设置本地头的数量 self.n_local_kv_heads = self.n_kv_heads # 设置本地 KV 头的数量 self.n_rep = self.n_local_heads // self.n_local_kv_heads # 计算重复次数 self.head_dim = args.dim // args.n_heads # 计算每个头的维度 self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False) # 初始化 Q 矩阵 self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) # 初始化 K 矩阵 self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) # 初始化 V 矩阵 self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False) # 初始化输出矩阵 self.k_cache, self.v_cache = None, None # 初始化 KV 缓存 self.attn_dropout = nn.Dropout(args.dropout) # 初始化注意力 dropout self.resid_dropout = nn.Dropout(args.dropout) # 初始化残差 dropout self.dropout = args.dropout # 设置 dropout 概率 self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn # 判断是否使用 Flash Attention if not self.flash: # print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0") mask = torch.full((1, 1, args.max_seq_len, args.max_seq_len), float("-inf")) # 初始化掩码 mask = torch.triu(mask, diagonal=1) # 生成上三角掩码 self.register_buffer("mask", mask) # 注册掩码 def forward(self, x: torch.Tensor, pos_cis: torch.Tensor, use_kv_cache=False): bsz, seqlen, _ = x.shape if use_kv_cache and self.eval(): # 如果使用 KV 缓存且在评估模式下 if self.k_cache is None or self.k_cache.shape[1] != x.shape[1] - 1: xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) # 计算 Q, K, V else: token = x[:, -1:, :] # 获取最后一个 token xq = torch.cat((torch.zeros_like(x[:, :-1, :]), self.wq(token)), dim=1) # 更新 Q xk = torch.cat((self.k_cache, self.wk(token)), dim=1) # 更新 K xv = torch.cat((self.v_cache, self.wv(token)), dim=1) # 更新 V self.k_cache, self.v_cache = xk, xv # 更新 KV 缓存 else: xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) # 计算 Q, K, V xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim) # 调整 Q 的形状 xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) # 调整 K 的形状 xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) # 调整 V 的形状 xq, xk = apply_rotary_emb(xq, xk, pos_cis) # 应用旋转位置编码 xk = repeat_kv(xk, self.n_rep) # 重复 K 的值 xv = repeat_kv(xv, self.n_rep) # 重复 V 的值 xq = xq.transpose(1, 2) # 调整 Q 的形状 xk = xk.transpose(1, 2) # 调整 K 的形状 xv = xv.transpose(1, 2) # 调整 V 的形状 if self.flash: output = torch.nn.functional.scaled_dot_product_attention(xq, xk, xv, attn_mask=None, dropout_p=self.dropout if self.training else 0.0, is_causal=True) # 使用 Flash Attention else: scores = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_dim) # 计算注意力分数 assert hasattr(self, 'mask') scores = scores + self.mask[:, :, :seqlen, :seqlen] # 应用掩码 scores = F.softmax(scores.float(), dim=-1).type_as(xq) # 计算 softmax scores = self.attn_dropout(scores) # 应用注意力 dropout output = torch.matmul(scores, xv) # 计算输出 output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1) # 调整输出的形状 output = self.wo(output) # 应用输出矩阵 output = self.resid_dropout(output) # 应用残差 dropout return output # 返回输出 # 定义 FeedForward 类，实现前馈神经网络 class FeedForward(nn.Module): def __init__(self, dim: int, hidden_dim: int, multiple_of: int, dropout: float): super().__init__() if hidden_dim is None: hidden_dim = 4 * dim # 设置隐藏层维度 hidden_dim = int(2 * hidden_dim / 3) # 调整隐藏层维度 hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) # 调整隐藏层维度 self.w1 = nn.Linear(dim, hidden_dim, bias=False) # 初始化第一层线性变换 self.w2 = nn.Linear(hidden_dim, dim, bias=False) # 初始化第二层线性变换 self.w3 = nn.Linear(dim, hidden_dim, bias=False) # 初始化第三层线性变换 self.dropout = nn.Dropout(dropout) # 初始化 dropout def forward(self, x): return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x))) # 前向传播 # 定义 MoEGate 类，实现专家混合（MoE）的门控机制 class MoEGate(nn.Module): def __init__(self, config: LMConfig): super().__init__() self.config = config self.top_k = config.num_experts_per_tok # 设置每个 token 选择的专家数量 self.n_routed_experts = config.n_routed_experts # 设置路由专家的数量 self.scoring_func = config.scoring_func # 设置评分函数 self.alpha = config.aux_loss_alpha # 设置辅助损失的权重 self.seq_aux = config.seq_aux # 设置序列辅助损失 self.norm_topk_prob = config.norm_topk_prob # 设置是否归一化 top-k 概率 self.gating_dim = config.dim # 设置门控维度 self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) # 初始化权重参数 self.reset_parameters() # 重置参数 def reset_parameters(self) -> None: import torch.nn.init as init init.kaiming_uniform_(self.weight, a=math.sqrt(5)) # 使用 Kaiming 初始化权重 def forward(self, hidden_states): bsz, seq_len, h = hidden_states.shape hidden_states = hidden_states.view(-1, h) # 调整隐藏状态的形状 logits = F.linear(hidden_states, self.weight, None) # 计算 logits if self.scoring_func == 'softmax': scores = logits.softmax(dim=-1) # 计算 softmax 评分 else: raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}') topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False) # 选择 top-k 专家 if self.top_k > 1 and self.norm_topk_prob: denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20 # 计算归一化分母 topk_weight = topk_weight / denominator # 归一化 top-k 概率 if self.training and self.alpha > 0.0: scores_for_aux = scores aux_topk = self.top_k topk_idx_for_aux_loss = topk_idx.view(bsz, -1) if self.seq_aux: scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1) ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device) ce.scatter_add_(1, topk_idx_for_aux_loss, torch.ones(bsz, seq_len * aux_topk, device=hidden_states.device)).div_( seq_len * aux_topk / self.n_routed_experts) aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alpha else: mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts) ce = mask_ce.float().mean(0) Pi = scores_for_aux.mean(0) fi = ce * self.n_routed_experts aux_loss = (Pi * fi).sum() * self.alpha else: aux_loss = None return topk_idx, topk_weight, aux_loss # 返回 top-k 专家索引、权重和辅助损失 # 定义 MOEFeedForward 类，实现专家混合（MoE）的前馈神经网络 class MOEFeedForward(nn.Module): def __init__(self, config: LMConfig): super().__init__() self.config = config self.experts = nn.ModuleList([ FeedForward( dim=config.dim, hidden_dim=config.hidden_dim, multiple_of=config.multiple_of, dropout=config.dropout, ) for _ in range(config.n_routed_experts) ]) # 初始化专家列表 self.gate = MoEGate(config) # 初始化门控机制 if config.n_shared_experts is not None: self.shared_experts = FeedForward( dim=config.dim, hidden_dim=config.hidden_dim, multiple_of=config.multiple_of, dropout=config.dropout, ) # 初始化共享专家 def forward(self, x): identity = x orig_shape = x.shape bsz, seq_len, _ = x.shape # 使用门控机制选择专家 topk_idx, topk_weight, aux_loss = self.gate(x) x = x.view(-1, x.shape[-1]) flat_topk_idx = topk_idx.view(-1) if self.training: # 训练模式下，重复输入数据 x = x.repeat_interleave(self.config.num_experts_per_tok, dim=0) y = torch.empty_like(x, dtype=torch.float16) for i, expert in enumerate(self.experts): y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]) y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1) y = y.view(*orig_shape) else: # 推理模式下，只选择最优专家 y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape) if self.config.n_shared_experts is not None: y = y + self.shared_experts(identity) return y @torch.no_grad() def moe_infer(self, x, flat_expert_indices, flat_expert_weights): expert_cache = torch.zeros_like(x) idxs = flat_expert_indices.argsort() tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0) token_idxs = idxs // self.config.num_experts_per_tok # 例如当tokens_per_expert=[6, 15, 20, 26, 33, 38, 46, 52] # 当token_idxs=[3, 7, 19, 21, 24, 25, 4, 5, 6, 10, 11, 12...] # 意味着当token_idxs[:6] -> [3, 7, 19, 21, 24, 25, 4]位置的token都由专家0处理，token_idxs[6:15]位置的token都由专家1处理...... for i, end_idx in enumerate(tokens_per_expert): start_idx = 0 if i == 0 else tokens_per_expert[i - 1] if start_idx == end_idx: continue expert = self.experts[i] exp_token_idx = token_idxs[start_idx:end_idx] expert_tokens = x[exp_token_idx] expert_out = expert(expert_tokens) expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]]) # 使用 scatter_add_ 进行 sum 操作 expert_cache.scatter_add_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out) return expert_cache # 定义 TransformerBlock 类，实现 Transformer 的一个块，包括自注意力和前馈神经网络 class TransformerBlock(nn.Module): def __init__(self, layer_id: int, args: LMConfig): super().__init__() self.n_heads = args.n_heads self.dim = args.dim self.head_dim = args.dim // args.n_heads self.attention = Attention(args) # 初始化自注意力机制 self.layer_id = layer_id self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps) # 初始化注意力归一化 self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps) # 初始化前馈神经网络归一化 if args.use_moe: self.feed_forward = MOEFeedForward(args) # 初始化专家混合前馈神经网络 else: self.feed_forward = FeedForward( dim=args.dim, hidden_dim=args.hidden_dim, multiple_of=args.multiple_of, dropout=args.dropout, ) # 初始化前馈神经网络 def forward(self, x, pos_cis, use_kv_cache=False): h = x + self.attention(self.attention_norm(x), pos_cis, use_kv_cache) # 计算自注意力 out = h + self.feed_forward(self.ffn_norm(h)) # 计算前馈神经网络 return out # 返回输出 # 定义 Transformer 类，实现整个 Transformer 模型 class Transformer(PreTrainedModel): config_class = LMConfig last_loss: Optional[torch.Tensor] def __init__(self, params: LMConfig = None): super().__init__(params) if not params: params = LMConfig() self.params = params self.vocab_size = params.vocab_size self.n_layers = params.n_layers class Transformer(PreTrainedModel): config_class = LMConfig last_loss: Optional[torch.Tensor] def __init__(self, params: LMConfig = None): super().__init__(params) if not params: params = LMConfig() self.params = params self.vocab_size = params.vocab_size self.n_layers = params.n_layers self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim) # 初始化词嵌入层 self.dropout = nn.Dropout(params.dropout) # 初始化 dropout 层 self.layers = torch.nn.ModuleList() # 初始化 Transformer 块列表 for layer_id in range(self.n_layers): self.layers.append(TransformerBlock(layer_id, params)) # 添加 Transformer 块 self.norm = RMSNorm(params.dim, eps=params.norm_eps) # 初始化归一化层 self.output = nn.Linear(params.dim, params.vocab_size, bias=False) # 初始化输出层 self.tok_embeddings.weight = self.output.weight # 共享词嵌入和输出层的权重 pos_cis = precompute_pos_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len) # 预计算位置编码 self.register_buffer("pos_cis", pos_cis, persistent=False) # 注册位置编码缓冲区 self.apply(self._init_weights) # 初始化模型权重 for pn, p in self.named_parameters(): if pn.endswith('w3.weight') or pn.endswith('wo.weight'): torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * params.n_layers)) # 对特定权重进行初始化 self.last_loss = None # 初始化最后一个损失 self.OUT = CausalLMOutputWithPast() # 初始化输出对象 def _init_weights(self, module): if isinstance(module, nn.Linear): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) # 初始化线性层的权重 if module.bias is not None: torch.nn.init.zeros_(module.bias) # 初始化线性层的偏置 elif isinstance(module, nn.Embedding): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) # 初始化嵌入层的权重 def forward(self, tokens: Optional[torch.Tensor] = None, targets: Optional[torch.Tensor] = None, use_kv_cache=False, **keyargs): if 'input_ids' in keyargs: tokens = keyargs['input_ids'] # 如果传入了 input_ids，则使用 input_ids if 'attention_mask' in keyargs: targets = keyargs['attention_mask'] # 如果传入了 attention_mask，则使用 attention_mask _bsz, seqlen = tokens.shape # 获取批量大小和序列长度 h = self.tok_embeddings(tokens) # 获取词嵌入 h = self.dropout(h) # 应用 dropout pos_cis = self.pos_cis[:seqlen] # 获取对应序列长度的位置编码 for idx, layer in enumerate(self.layers): h = layer(h, pos_cis, use_kv_cache) # 逐层应用 Transformer 块 h = self.norm(h) # 应用归一化 if targets is not None: logits = self.output(h) # 计算 logits self.last_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1) # 计算交叉熵损失 else: logits = self.output(h[:, [-1], :]) # 计算最后一个 token 的 logits self.last_loss = None # 没有目标时，损失为 None self.OUT.__setitem__('logits', logits) # 设置输出对象的 logits self.OUT.__setitem__('last_loss', self.last_loss) # 设置输出对象的 last_loss return self.OUT # 返回输出对象 @torch.inference_mode() # 推理模式 def generate(self, idx, eos, max_new_tokens, temperature=0.7, top_k=None, stream=True, repetition_penalty=1., use_kv_cache=True): index = idx.shape[1] # 获取当前序列长度 while idx.shape[1] < max_new_tokens - 1: # 当生成的 token 数量小于最大数量时 inference_res = self(idx, use_kv_cache=use_kv_cache) # 进行前向传播 logits = inference_res.logits # 获取 logits logits = logits[:, -1, :] # 获取最后一个 token 的 logits for token in set(idx.tolist()[0]): # 对重复 token 进行惩罚 logits[:, token] /= repetition_penalty if temperature == 0.0: # 如果温度为 0，直接选择概率最高的 token _, idx_next = torch.topk(logits, k=1, dim=-1) else: logits = logits / temperature # 调整 logits if top_k is not None: # 如果设置了 top-k 采样 v, _ = torch.topk(logits, min(top_k, logits.size(-1))) logits[logits < v[:, [-1]]] = -float('Inf') # 将小于 top-k 的 logits 设为负无穷 probs = F.softmax(logits, dim=-1) # 计算概率 idx_next = torch.multinomial(probs, num_samples=1, generator=None) # 采样下一个 token if idx_next == eos: # 如果生成的 token 是结束符，停止生成 break idx = torch.cat((idx, idx_next), dim=1) # 将生成的 token 添加到序列中 if stream: # 如果需要流式输出 yield idx[:, index:] # 返回生成的 token if not stream: # 如果不需要流式输出 yield idx[:, index:] # 返回生成的 token @torch.inference_mode() # 推理模式 def eval_answer(self, idx): idx_cond = idx if idx.size(1) <= self.params.max_seq_len else idx[:, -self.params.max_seq_len:] # 截取序列 inference_res = self(idx_cond) # 进行前向传播 logits = inference_res.logits # 获取 logits logits = logits[:, -1, :] # 获取最后一个 token 的 logits return logits # 返回 logits