import math import struct import inspect import time 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 class RMSNorm(torch.nn.Module): def __init__(self, dim: int, eps: float): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) def forward(self, x): output = self._norm(x.float()).type_as(x) return output * self.weight 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)) t = torch.arange(end, device=freqs.device) # type: ignore freqs = torch.outer(t, freqs).float() # type: ignore pos_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 return pos_cis 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)) xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) pos_cis = unite_shape(pos_cis, xq_) xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3) xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3) return xq_out.type_as(xq), xk_out.type_as(xk) 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) ) 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 assert args.n_heads % self.n_kv_heads == 0 self.n_local_heads = args.n_heads self.n_local_kv_heads = self.n_kv_heads self.n_rep = self.n_local_heads // self.n_local_kv_heads self.head_dim = args.dim // args.n_heads self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False) self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False) self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False) self.k_cache, self.v_cache = None, None self.attn_dropout = nn.Dropout(args.dropout) self.resid_dropout = nn.Dropout(args.dropout) self.dropout = args.dropout self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn # 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")) mask = torch.triu(mask, diagonal=1) self.register_buffer("mask", mask) def forward(self, x: torch.Tensor, pos_cis: torch.Tensor, kv_cache=False): bsz, seqlen, _ = x.shape xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim) xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim) xq, xk = apply_rotary_emb(xq, xk, pos_cis) # 更高效的kv_cache实现 if kv_cache and self.eval(): if seqlen == 1 and all(cache is not None for cache in (self.k_cache, self.v_cache)): xk = torch.cat((self.k_cache, xk), dim=1) xv = torch.cat((self.v_cache, xv), dim=1) self.k_cache, self.v_cache = xk, xv xk = repeat_kv(xk, self.n_rep) # (bs, seqlen, n_local_heads, head_dim) xv = repeat_kv(xv, self.n_rep) # (bs, seqlen, n_local_heads, head_dim) xq = xq.transpose(1, 2) xk = xk.transpose(1, 2) xv = xv.transpose(1, 2) if self.flash and seqlen != 1: 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) else: scores = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_dim) scores = scores + self.mask[:, :, :seqlen, :seqlen] # (bs, n_local_heads, seqlen, cache_len + seqlen) scores = F.softmax(scores.float(), dim=-1).type_as(xq) scores = self.attn_dropout(scores) output = torch.matmul(scores, xv) # (bs, n_local_heads, seqlen, head_dim) output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1) output = self.wo(output) output = self.resid_dropout(output) return output 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 hidden_dim = int(2 * hidden_dim / 3) hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) self.w1 = nn.Linear(dim, hidden_dim, bias=False) self.w2 = nn.Linear(hidden_dim, dim, bias=False) self.w3 = nn.Linear(dim, hidden_dim, bias=False) self.dropout = nn.Dropout(dropout) def forward(self, x): return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x))) class MoEGate(nn.Module): def __init__(self, config: LMConfig): super().__init__() self.config = config self.top_k = config.num_experts_per_tok self.n_routed_experts = config.n_routed_experts self.scoring_func = config.scoring_func self.alpha = config.aux_loss_alpha self.seq_aux = config.seq_aux self.norm_topk_prob = config.norm_topk_prob self.gating_dim = config.dim self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) self.reset_parameters() def reset_parameters(self) -> None: import torch.nn.init as init init.kaiming_uniform_(self.weight, a=math.sqrt(5)) def forward(self, hidden_states): bsz, seq_len, h = hidden_states.shape hidden_states = hidden_states.view(-1, h) logits = F.linear(hidden_states, self.weight, None) if self.scoring_func == 'softmax': scores = logits.softmax(dim=-1) 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) if self.top_k > 1 and self.norm_topk_prob: denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20 topk_weight = topk_weight / denominator 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 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) ]) self.gate = MoEGate(config) 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, ) 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 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) self.layer_id = layer_id self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps) self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps) if args.use_moe: self.feed_forward = MOEFeedForward(args) else: self.feed_forward = FeedForward( dim=args.dim, hidden_dim=args.hidden_dim, multiple_of=args.multiple_of, dropout=args.dropout, ) def forward(self, x, pos_cis, kv_cache=False): h = x + self.attention(self.attention_norm(x), pos_cis, kv_cache) out = h + self.feed_forward(self.ffn_norm(h)) return out 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) self.dropout = nn.Dropout(params.dropout) self.layers = torch.nn.ModuleList() for layer_id in range(self.n_layers): self.layers.append(TransformerBlock(layer_id, params)) self.norm = RMSNorm(params.dim, eps=params.norm_eps) self.output = nn.Linear(params.dim, params.vocab_size, bias=False) self.tok_embeddings.weight = self.output.weight pos_cis = precompute_pos_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len) self.register_buffer("pos_cis", pos_cis, persistent=False) self.apply(self._init_weights) 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)) self.last_loss = None self.OUT = CausalLMOutputWithPast() def _init_weights(self, module): if isinstance(module, nn.Linear): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) if module.bias is not None: torch.nn.init.zeros_(module.bias) elif isinstance(module, nn.Embedding): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) def forward(self, tokens: Optional[torch.Tensor] = None, targets: Optional[torch.Tensor] = None, kv_cache=False, **keyargs): current_idx = 0 if 'input_ids' in keyargs: tokens = keyargs['input_ids'] if 'attention_mask' in keyargs: targets = keyargs['attention_mask'] if 'current_idx' in keyargs: current_idx = int(keyargs['current_idx']) _bsz, seqlen = tokens.shape h = self.tok_embeddings(tokens) h = self.dropout(h) pos_cis = self.pos_cis[current_idx:current_idx + seqlen] for idx, layer in enumerate(self.layers): h = layer(h, pos_cis, kv_cache) h = self.norm(h) if targets is not None: logits = self.output(h) self.last_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1) else: logits = self.output(h[:, [-1], :]) self.last_loss = None self.OUT.__setitem__('logits', logits) self.OUT.__setitem__('last_loss', self.last_loss) return self.OUT @torch.inference_mode() def generate(self, idx, eos, max_new_tokens, temperature=0.7, top_k=8, stream=True, rp=1., kv_cache=True): # rp: repetition_penalty index = idx.shape[1] init_inference = True while idx.shape[1] < max_new_tokens - 1: if init_inference or not kv_cache: inference_res, init_inference = self(idx, kv_cache=kv_cache), False else: inference_res = self(idx[:, -1:], kv_cache=kv_cache, current_idx=idx.shape[1] - 1) logits = inference_res.logits logits = logits[:, -1, :] for token in set(idx.tolist()[0]): logits[:, token] /= rp if temperature == 0.0: _, idx_next = torch.topk(logits, k=1, dim=-1) else: logits = logits / temperature if top_k is not None: v, _ = torch.topk(logits, min(top_k, logits.size(-1))) logits[logits < v[:, [-1]]] = -float('Inf') probs = F.softmax(logits, dim=-1) idx_next = torch.multinomial(probs, num_samples=1, generator=None) if idx_next == eos: break idx = torch.cat((idx, idx_next), dim=1) if stream: yield idx[:, index:] if not stream: yield idx[:, index:] @torch.inference_mode() def eval_answer(self, idx): idx_cond = idx if idx.size(1) <= self.params.max_seq_len else idx[:, -self.params.max_seq_len:] inference_res = self(idx_cond) logits = inference_res.logits logits = logits[:, -1, :] return logits