import math
import struct
import inspect
import time
from .LMConfig import LMConfig
from typing import Any, Optional, Tuple, List, Union
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
from torch import nn, einsum
from einops import rearrange, repeat
def exists(val):
return val is not None
# RMSNorm 类定义了一个用于归一化输入张量的模块。
class RMSNorm(torch.nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
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):
return self.weight * self._norm(x.float()).type_as(x)
# precompute_pos_cis 函数用于预计算位置编码（复数版本）。
def precompute_pos_cis(dim: int, end: int = int(32 * 1024), theta: float = 1e6):
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
# apply_rotary_emb 函数用于应用旋转位置编码（复数版本）。
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)
# precompute_pos_cis_real 函数用于预计算位置编码（实数版本）。
def precompute_pos_cis_real(dim: int, end: int = int(32 * 1024), theta: float = 1e6):
"""使用实数张量实现位置编码，避免使用复数张量
这个函数与precompute_pos_cis完全等价，但使用实数张量而非复数张量。
原始函数生成形状为[seq_len, dim//2]的复数张量，其中实部全为1，虚部为旋转角度。
这个函数生成形状为[seq_len, dim]的实数张量，其中偶数索引是cos(角度)，奇数索引是sin(角度)。
"""
# 确保dim是偶数
if dim % 2 != 0:
raise ValueError(f"维度必须是偶数，但得到了 {dim}")
# 复制原始函数的频率计算逻辑
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
t = torch.arange(end, device=freqs.device)
freqs = torch.outer(t, freqs).float()
# 计算cos和sin值
# 在复数版本中，pos_cis = torch.polar(torch.ones_like(freqs), freqs)
# 等价于 cos(freqs) + i*sin(freqs)
cos = torch.cos(freqs)
sin = torch.sin(freqs)
# 创建实数张量，交错排列cos和sin
pos_emb = torch.zeros((end, dim), device=freqs.device)
pos_emb[:, 0::2] = cos # 偶数索引放cos
pos_emb[:, 1::2] = sin # 奇数索引放sin
return pos_emb
# apply_rotary_emb_real 函数用于应用旋转位置编码（实数版本）。
def apply_rotary_emb_real(xq, xk, pos_emb):
"""使用实数张量实现旋转位置编码，避免使用复数张量
这个函数与apply_rotary_emb完全等价，但使用实数张量而非复数张量。
原始函数将输入张量转换为复数形式，与位置编码相乘，然后再转回实数形式。
这个函数直接使用实数运算实现相同的旋转操作。
"""
# 获取形状信息
bsz, seq_len, n_heads, head_dim = xq.shape
# 确保pos_emb形状正确
assert pos_emb.shape[0] >= seq_len, f"位置编码长度 {pos_emb.shape[0]} 小于序列长度 {seq_len}"
assert pos_emb.shape[1] == head_dim, f"位置编码维度 {pos_emb.shape[1]} 与头维度 {head_dim} 不匹配"
# 截取需要的位置编码长度
pos_emb = pos_emb[:seq_len]
# 将pos_emb调整为广播形状 [1, seq_len, 1, head_dim]
pos_emb = pos_emb.unsqueeze(0).unsqueeze(2)
# 将head_dim分成两半
half_head_dim = head_dim // 2
# 提取cos和sin值（偶数索引是cos，奇数索引是sin）
cos = pos_emb[..., 0::2]
sin = pos_emb[..., 1::2]
# 将xq和xk重新排列，以便进行旋转操作
# 原始复数版本中，xq和xk被重塑为复数张量，其中实部和虚部交错排列
# 在实数版本中，我们需要将偶数索引和奇数索引分开处理
# 分离偶数和奇数索引
xq_even = xq[..., 0::2] # 偶数索引，对应复数的实部
xq_odd = xq[..., 1::2] # 奇数索引，对应复数的虚部
xk_even = xk[..., 0::2]
xk_odd = xk[..., 1::2]
# 应用旋转（等价于复数乘法）
# (a + bi)(cos + sin*i) = (a*cos - b*sin) + (a*sin + b*cos)i
# 其中a是偶数索引，b是奇数索引
xq_out_even = xq_even * cos - xq_odd * sin # 新的偶数索引（实部）
xq_out_odd = xq_even * sin + xq_odd * cos # 新的奇数索引（虚部）
xk_out_even = xk_even * cos - xk_odd * sin
xk_out_odd = xk_even * sin + xk_odd * cos
# 重新组合偶数和奇数索引
xq_out = torch.zeros_like(xq)
xk_out = torch.zeros_like(xk)
xq_out[..., 0::2] = xq_out_even
xq_out[..., 1::2] = xq_out_odd
xk_out[..., 0::2] = xk_out_even
xk_out[..., 1::2] = xk_out_odd
return xq_out.type_as(xq), xk_out.type_as(xk)
# repeat_kv 函数用于重复键值对。
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.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, persistent=False)
def forward(self,
x: torch.Tensor,
pos_cis: torch.Tensor,
db_value=None):
bsz, seq_len, _ = x.shape #bsz: 批量大小, seq_len: 序列长度, _: 隐藏维度
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) #将输入张量x分别通过线性层wq, wk, wv进行变换，得到查询、键和值。
xq = xq.view(bsz, seq_len, self.n_local_heads, self.head_dim) #将变换后的张量xq重塑为形状为(bsz, seq_len, n_local_heads, head_dim)的形状。
xk = xk.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim) #将变换后的张量xk重塑为形状为(bsz, seq_len, n_local_kv_heads, head_dim)的形状。
xv = xv.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim) #将变换后的张量xv重塑为形状为(bsz, seq_len, n_local_kv_heads, head_dim)的形状。
# 应用旋转位置编码（使用实数版本）
xq, xk = apply_rotary_emb_real(xq, xk, pos_cis)
# 重复键值对
xq, xk, xv = (
xq.transpose(1, 2),
repeat_kv(xk, self.n_rep).transpose(1, 2),
repeat_kv(xv, self.n_rep).transpose(1, 2)
)
# 如果提供了db_value，根据头的数量调整它的形状并与xv合并
if db_value is not None:
# 确保db_value的形状与xv兼容，假设db_value形状为[B, N, H, D]
if db_value.ndim == 4: # [B, N, H, D]
db_value = db_value.transpose(1, 2) # -> [B, H, N, D]
# 检查是否需要调整D维度
if db_value.shape[-1] != xv.shape[-1]:
# 如果db_value的维度与xv不同，可以添加一个投影层
# 或者在这里使用简单的调整方法
# 这里我们简单地通过均值池化或重复来调整维度
if db_value.shape[-1] > xv.shape[-1]:
# 降维
factor = db_value.shape[-1] // xv.shape[-1]
db_value = db_value.view(bsz, self.n_local_heads, seq_len, factor, xv.shape[-1])
db_value = db_value.mean(dim=3)
else:
# 升维
factor = xv.shape[-1] // db_value.shape[-1]
db_value = db_value.unsqueeze(-1).repeat(1, 1, 1, 1, factor)
db_value = db_value.view(bsz, self.n_local_heads, seq_len, xv.shape[-1])
# 将db_value与xv相加或融合
# 这里我们简单地将它们相加，但你也可以使用其他融合方法
xv = xv + db_value
# 使用Flash Attention
if self.flash and seq_len != 1:
dropout_p = self.dropout if self.training else 0.0
output = F.scaled_dot_product_attention(
xq, xk, xv,
attn_mask=None,
dropout_p=dropout_p,
is_causal=True
)
else:
scores = (xq @ xk.transpose(-2, -1)) / math.sqrt(self.head_dim)
scores += self.mask[:, :, :seq_len, :seq_len]
scores = F.softmax(scores.float(), dim=-1).type_as(xq)
scores = self.attn_dropout(scores)
output = scores @ xv
output = output.transpose(1, 2).reshape(bsz, seq_len, -1)
output = self.resid_dropout(self.wo(output))
return output
class CrossAttention(nn.Module):
def __init__(
self,
config
):
super().__init__()
self.config = config
self.num_heads = 8
self.head_dim = self.config.dim // self.num_heads
self.to_q = nn.Linear(self.config.dim, self.config.dim, bias=False)
self.to_k = nn.Linear(self.config.dim, self.config.dim, bias=False)
self.to_v = nn.Linear(self.config.dim, self.config.dim, bias=False)
self.to_out = nn.Linear(self.config.dim, self.config.dim, bias=False)
def forward(self, x, db, context_mask=None, pos_emb=None):
batch_size = x.size(0)
# 分离多头
q = self.to_q(x).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
k = self.to_k(db).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
v = self.to_v(db).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
if pos_emb is not None:
pos_emb = pos_emb.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
q = q + pos_emb
k = k + pos_emb
v = v + pos_emb
attn_scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
if context_mask is not None:
expanded_mask = context_mask.unsqueeze(1).expand(-1, self.num_heads, -1, -1)
attn_scores = attn_scores.masked_fill(expanded_mask == 0, -1e10)
attn_weights = F.softmax(attn_scores, dim=-1)
context = torch.matmul(attn_weights, v)
context = context.transpose(1, 2).contiguous().view(batch_size, -1, self.config.dim)
context = self.to_out(context)
return context
class FeedForward(nn.Module):
def __init__(self, config: LMConfig):
super().__init__()
if config.hidden_dim is None:
hidden_dim = 4 * config.dim
hidden_dim = int(2 * hidden_dim / 3)
config.hidden_dim = config.multiple_of * ((hidden_dim + config.multiple_of - 1) // config.multiple_of)
self.w1 = nn.Linear(config.dim, config.hidden_dim, bias=False)
self.w2 = nn.Linear(config.hidden_dim, config.dim, bias=False)
self.w3 = nn.Linear(config.dim, config.hidden_dim, bias=False)
self.dropout = nn.Dropout(config.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 = 0
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(config)
for _ in range(config.n_routed_experts)
])
self.gate = MoEGate(config)
if config.n_shared_experts is not None:
self.shared_experts = FeedForward(config)
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:
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]).to(y.dtype) # 确保类型一致
y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
y = y.view(*orig_shape)
else:
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)
self.aux_loss = aux_loss
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
# 当tokens_per_expert = [6, 15, 20, 26]，tokens_per_expert.shape[0]即为专家数量（此时为4）
# 且token_idxs = [3, 7, 19, 21, 24, 25, 4, 5, 6, 10, 11, 12...] 时
# 意味token_idxs[:6] -> [3, 7, 19, 21, 24, 25]这6个位置属于专家0处理的token（每个token有可能被多个专家处理，这取决于num_experts_per_tok）
# 接下来9个位置token_idxs[6:15] -> [4, 5, 6, 10, 11, 12...]属于专家1处理的token...依此类推
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).to(expert_cache.dtype)
expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
expert_cache.scatter_add_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out)
return expert_cache
class MiniMindBlock(nn.Module):
def __init__(self, layer_id: int, config: LMConfig):
super().__init__()
self.n_heads = config.n_heads
self.dim = config.dim
self.head_dim = config.dim // config.n_heads
self.attention = Attention(config)
self.cross_att = CrossAttention(config)
self.layer_id = layer_id
self.attention_norm = RMSNorm(config.dim, eps=config.norm_eps)
self.ffn_norm = RMSNorm(config.dim, eps=config.norm_eps)
self.feed_forward = FeedForward(config) if not config.use_moe else MOEFeedForward(config)
def forward(self, x, db_value, pos_cis):
# 注意力计算
h_attn = self.attention(
self.attention_norm(x),
pos_cis,
db_value=db_value
)
h_attn = self.cross_att(h_attn, db_value)
# 残差连接
h = x + h_attn
# 前馈神经网络
out = h + self.feed_forward(self.ffn_norm(h))
return out
class ExtractDB(nn.Module):
def __init__(self, params, tok_embeddings=None):
# 修改专家数量和知识维度，确保能开方
super().__init__()
self.batch_size = None
self.dim = params.dim
self.dim_key = self.dim // 2
self.knowledge_num = params.knowledge_num # 100专家，确保是完全平方数
# 将knowledge_dim设置为与head_dim相同，以便在attention中直接使用
self.head_dim = params.dim // params.n_heads
self.knowledge_length = params.knowledge_length
# 智能负载均衡相关参数
self.enable_intelligent_balance = getattr(params, 'db_intelligent_balance', True)
self.relevance_threshold = getattr(params, 'db_relevance_threshold', 0.7)
self.base_balance_strength = getattr(params, 'db_balance_strength', 0.3)
self.momentum = getattr(params, 'db_momentum', 0.9)
self.adaptive_weights = getattr(params, 'db_adaptive_weights', True)
# 动态权重调整参数
self.current_relevance_weight = 0.8 # 开始时更重视相关性
self.current_balance_weight = 0.2
self.weight_update_frequency = 100 # 每100步调整一次权重
self.step_counter = 0
# 使用频率统计 - 使用register_buffer以便在GPU/CPU间正确移动
self.register_buffer('usage_counts', torch.zeros(self.knowledge_num))
self.register_buffer('total_queries', torch.tensor(0.0))
# 知识库存储 - 使用register_buffer因为这是整数索引，不需要梯度
self.register_buffer('weight_down_embed',
torch.randint(low=0, high=6400, size=(self.knowledge_num, self.knowledge_length), dtype=torch.long)
)
self.num_keys = int(math.sqrt(self.knowledge_num)) if self.knowledge_num > 0 else 0
self.product_key_topk = min(16, self.num_keys)
self.keys = nn.Parameter(torch.randn(self.num_keys, 2, self.dim_key) * 0.02)
self.num_experts_per_head_topk = 1
self.to_queries = nn.Sequential(
nn.Linear(params.dim, self.dim_key * 2, bias=False),
)
# 存储token embeddings的引用，用于计算真实的语义相关性
self.tok_embeddings = tok_embeddings
def update_usage_statistics(self, selected_indices):
"""更新数据库条目的使用统计"""
if not self.training or not self.enable_intelligent_balance:
return
with torch.no_grad():
# 统计当前batch中每个条目的使用次数
batch_usage = torch.zeros(self.knowledge_num, device=selected_indices.device)
unique_indices, counts = torch.unique(selected_indices, return_counts=True)
batch_usage[unique_indices] = counts.float()
# 使用简单的tensor操作来更新统计
current_usage = self.usage_counts.clone()
current_total = self.total_queries.clone()
new_usage = self.momentum * current_usage + (1 - self.momentum) * batch_usage
new_total = current_total + selected_indices.numel()
# 直接替换buffer内容
self.usage_counts.copy_(new_usage)
self.total_queries.copy_(new_total)
def update_dynamic_weights(self):
"""动态调整相关性和平衡权重"""
if not self.adaptive_weights or not self.training:
return
self.step_counter += 1
# 每隔一定步数调整权重
if self.step_counter % self.weight_update_frequency == 0:
with torch.no_grad():
if self.total_queries > 0:
# 计算使用分布的方差（不平衡程度）
usage_rates = self.usage_counts / self.total_queries
usage_variance = usage_rates.var().item()
# 根据不平衡程度调整权重
if usage_variance > 0.01: # 高度不平衡
self.current_relevance_weight = max(0.5, self.current_relevance_weight - 0.1)
self.current_balance_weight = min(0.5, self.current_balance_weight + 0.1)
elif usage_variance < 0.001: # 已经很平衡
self.current_relevance_weight = min(0.9, self.current_relevance_weight + 0.1)
self.current_balance_weight = max(0.1, self.current_balance_weight - 0.1)
# 确保权重和为1
total_weight = self.current_relevance_weight + self.current_balance_weight
self.current_relevance_weight /= total_weight
self.current_balance_weight /= total_weight
def intelligent_selection(self, query, all_scores, all_indices):
"""智能分层选择策略"""
if not self.enable_intelligent_balance or not self.training:
# 如果禁用智能平衡或在推理模式，使用原始分数
return all_scores
with torch.no_grad():
batch_size = all_scores.size(0)
device = all_scores.device
dtype = all_scores.dtype
# 更新动态权重
self.update_dynamic_weights()
# 对每个batch进行分层选择
enhanced_scores = all_scores.clone()
# 预先计算query的特征表示（取平均）
query_features = query.mean(dim=1) # [batch_size, dim]
for batch_idx in range(batch_size):
indices = all_indices[batch_idx] # 当前batch的候选条目
scores = all_scores[batch_idx] # 当前batch的原始分数
# 第一层：基于value内容计算真正的相关性
# 1. 获取候选条目的value tokens（只获取当前需要的）
candidate_tokens = self.weight_down_embed[indices] # [num_candidates, knowledge_length]
# 2. 高效计算：直接使用embedding层，避免中间变量
# 将tokens reshape为一维，批量计算embeddings，然后reshape回来
num_candidates, knowledge_length = candidate_tokens.shape
flat_tokens = candidate_tokens.view(-1) # [num_candidates * knowledge_length]
# 批量计算所有token的embeddings
flat_embeddings = self.tok_embeddings(flat_tokens) # [num_candidates * knowledge_length, dim]
# Reshape回原始形状并进行mean pooling
candidate_embeddings = flat_embeddings.view(num_candidates, knowledge_length, -1)
candidate_features = candidate_embeddings.mean(dim=1) # [num_candidates, dim]
# 3. 计算query与候选条目的相似度
query_feature = query_features[batch_idx] # [dim]
similarity_scores = F.cosine_similarity(
query_feature.unsqueeze(0), candidate_features, dim=1
) # [num_candidates]
# 4. 将相似度分数归一化为概率分布
relevance_probs = F.softmax(similarity_scores.float(), dim=-1).to(dtype)
# 相关性阈值：选择概率大于某个阈值的候选项
# 动态阈值：如果所有候选项的相似度都很平均，降低阈值
mean_prob = relevance_probs.mean()
adaptive_threshold = max(self.relevance_threshold * mean_prob, mean_prob * 0.5)
relevant_mask = relevance_probs > adaptive_threshold
if relevant_mask.sum() == 0:
# 如果没有足够相关的，选择相似度最高的top-k
top_k = min(5, len(indices))
_, top_indices = similarity_scores.topk(top_k)
relevant_mask = torch.zeros_like(relevant_mask, dtype=torch.bool)
relevant_mask[top_indices] = True
# 第二层：在相关候选中应用平衡策略
if relevant_mask.sum() > 1:
# 计算平衡分数（使用频率低的分数高）
relevant_indices = indices[relevant_mask]
relevant_usage = self.usage_counts[relevant_indices]
# 平衡分数：使用频率的倒数（加1避免除零）
balance_scores = 1.0 / (relevant_usage + 1.0)
balance_scores = balance_scores / (balance_scores.sum() + 1e-8)
# 相关性分数（基于真实的语义相似度）
relevant_rel_scores = relevance_probs[relevant_mask]
relevant_rel_scores = relevant_rel_scores / (relevant_rel_scores.sum() + 1e-8)
# 综合分数：动态权重组合
combined_scores = (self.current_relevance_weight * relevant_rel_scores +
self.current_balance_weight * balance_scores.to(dtype))
# 确保数据类型一致
adjustment = self.base_balance_strength * combined_scores.to(dtype)
# 将综合分数应用到enhanced_scores
enhanced_scores[batch_idx, relevant_mask] = (
scores[relevant_mask] + adjustment
)
# 清理中间变量，释放显存
del candidate_tokens, flat_tokens, flat_embeddings, candidate_embeddings, candidate_features
return enhanced_scores.to(device)
def q_to_k(self,x):
# 1. 生成queries
self.batch_size, seq_len, dim = x.shape
# collapse sequence dimension by averaging
x_flat = x.mean(dim=1) # [batch_size, dim]
queries = self.to_queries(x_flat) # [batch_size, 2*dim_key]
queries = queries.reshape(self.batch_size, 2, self.dim_key) # [batch_size, 2, dim_key]
queries = queries.permute(1, 0, 2) # [2, batch_size, dim_key]
# 2. 计算queries与keys的相似度
sim = torch.einsum('p b d, k p d -> p b k', queries, self.keys)
# 3. 在两个子空间分别做top-k
scores_and_indices = [sim[p].topk(self.product_key_topk, dim=-1) for p in range(2)]
scores_x, scores_y = scores_and_indices[0][0], scores_and_indices[1][0]
indices_x, indices_y = scores_and_indices[0][1], scores_and_indices[1][1]
# 4. 组合两个子空间的分数和索引
all_scores = scores_x.unsqueeze(-1) + scores_y.unsqueeze(-2)
all_scores = all_scores.view(*all_scores.shape[:-2], -1)
all_indices = (indices_x.unsqueeze(-1) * self.num_keys) + indices_y.unsqueeze(-2)
all_indices = all_indices.view(*all_indices.shape[:-2], -1)
# 5. 应用智能分层选择策略
enhanced_scores = self.intelligent_selection(x, all_scores, all_indices)
# 6. 基于增强后的分数进行最终top-k选择
scores, pk_indices = enhanced_scores.topk(self.num_experts_per_head_topk, dim=-1)
indices = all_indices.gather(-1, pk_indices)
flat_indices = indices.view(-1)
# 7. 更新使用统计
self.update_usage_statistics(flat_indices)
return flat_indices
def get_data(self, index):
# 直接从GPU获取embedding
db_values = self.weight_down_embed[index]#变成token了所以是1,后续再过emb
# db_value = db_values.view(self.batch_size,-1)
return db_values
@torch.no_grad()
def updata_value(self, k, v):#要加一个从向量返回index的过程
# 直接更新buffer上的值 (不需要梯度)
v_reshaped = v.view(v.size(0), -1)
# 确保数据类型匹配
v_reshaped = v_reshaped.to(dtype=self.weight_down_embed.dtype)
self.weight_down_embed[k] = v_reshaped
class MiniMindLM(PreTrainedModel):
config_class = LMConfig
def __init__(self, params: LMConfig = None):
self.params = params or LMConfig()
super().__init__(self.params)
self.vocab_size, self.n_layers = params.vocab_size, params.n_layers
# 先创建token embeddings
self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
self.dropout = nn.Dropout(params.dropout)
# 创建ExtractDB，传入tok_embeddings引用
self.extract_db = ExtractDB(self.params, self.tok_embeddings)
# 将self.weight_down_embed传递给每个MiniMindBlock
self.layers = nn.ModuleList([MiniMindBlock(l, params) for l in range(self.n_layers)])
self.norm = RMSNorm(params.dim, eps=params.norm_eps)
self.output = nn.Linear(params.dim, params.vocab_size, bias=False)
self.database_output = nn.Linear(params.dim, params.knowledge_length, bias=False)
self.tok_embeddings.weight = self.output.weight
self.database_output.weight = self.output.weight
# Calculate input dimension
input_dim = (self.params.max_seq_len-1)*self.params.n_layers
# Use a bottleneck architecture to reduce parameters
bottleneck_dim = 256 # Significantly smaller bottleneck dimension
# Factorized shared downsampling using two smaller convolutions
self.shared_downsample = nn.Sequential(
# First reduce input dimension to bottleneck
nn.Conv1d(input_dim, bottleneck_dim, kernel_size=1, padding='same'),
nn.ReLU(), # Non-linearity to improve representation capacity
# Then expand to target dimension
nn.Conv1d(bottleneck_dim, 128*8, kernel_size=1, padding='same')
)
# Specific layers for v path
self.downsample_v_specific = nn.Sequential(
nn.Conv1d(128*8, 128, kernel_size=1, padding='same'),
nn.Conv1d(128, self.params.knowledge_length, kernel_size=1, padding='same')
)
# Specific layers for q path
self.downsample_q_specific = nn.Sequential(
nn.Conv1d(128*8, 512, kernel_size=1, padding='same')
)
# 使用实数版本的位置编码，避免复数张量可能导致的段错误
self.register_buffer("pos_cis_real",
precompute_pos_cis_real(dim=params.dim // params.n_heads, theta=params.rope_theta),
persistent=False)
self.params = params
def forward(self,
input_ids: Optional[torch.Tensor] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
**args):
start_pos = args.get('start_pos', 0)
h = self.dropout(self.tok_embeddings(input_ids))
pos_cis_real = self.pos_cis_real[start_pos:start_pos + input_ids.size(1)]
h_list = []
for l, layer in enumerate(self.layers):
# 正常模式，使用数据库查询
# import pdb;pdb.set_trace()
index = self.extract_db.q_to_k(h)
token_idx = self.extract_db.get_data(index) #这里是index
db_value =self.tok_embeddings(token_idx)
h = layer(
h, db_value, pos_cis_real
)
h_list.append(h.unsqueeze(0))
h_tensor = torch.cat(h_list, dim=0).permute(1, 0, 2, 3)
# 只在非禁用数据库模式下执行数据库更新逻辑
if not self.params.disable_db:
# 使用detach()分离计算图，避免多次反向传播
h_tensor_detached = h_tensor.detach()
h_tensor_detached = h_tensor_detached.reshape(h_tensor_detached.shape[0], -1, self.params.dim)
# 数据库更新逻辑与主计算图分离
with torch.no_grad():
# Compute shared downsampling layer once
shared_features = self.shared_downsample(h_tensor_detached)
# Get features from v path - now we output embedding-dimension vectors
z_v_features = self.downsample_v_specific(shared_features)
batch_z, seq_len, dim_z = z_v_features.shape
# Reshape to batch_size * knowledge_length, dim
z_v_flat = z_v_features.reshape(-1, dim_z)
# Direct token prediction - like the main language model head
token_logits = self.database_output(z_v_flat) # [batch_z * seq_len, vocab_size]
# Get token indices directly from logits
token_indices_flat = torch.argmax(token_logits, dim=-1)
token_indices = token_indices_flat.reshape(batch_z, -1)
# Process query path as before
z_q = self.downsample_q_specific(shared_features)
z_k = self.extract_db.q_to_k(z_q)
# self.extract_db.updata_value(z_k, token_indices)
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
logits = self.output(self.norm(h)[:, slice_indices, :])
aux_loss = sum(l.feed_forward.aux_loss for l in self.layers if isinstance(l.feed_forward, MOEFeedForward))
# 进一步简化，只保留必要的参数
output = CausalLMOutputWithPast(
logits=logits,
)
output.hidden_states = h
output.aux_loss = aux_loss
# 尝试添加其他属性（如果支持的话）
# try:
# output.hidden_states = h
# except:
# pass
return output
@torch.inference_mode()
def generate(self, input_ids, eos_token_id=2, max_new_tokens=1024, temperature=0.75, top_p=0.90,
stream=False, rp=1., pad_token_id=0, num_return_sequences=1, **args):
# 流式生成
if stream:
return self._stream(input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, **args)
# 直接生成
generated = []
for i in range(input_ids.size(0)):
non_pad = input_ids[i][input_ids[i] != pad_token_id].unsqueeze(0)
for _ in range(num_return_sequences):
out = self._stream(non_pad, eos_token_id, max_new_tokens, temperature, top_p, rp, **args)
tokens_list = [tokens[:, -1:] for tokens in out]
gen = torch.cat(tokens_list, dim=-1) if tokens_list else non_pad
full_sequence = torch.cat([non_pad, gen], dim=-1)
generated.append(full_sequence)
max_length = max(seq.size(1) for seq in generated)
generated = [
torch.cat(
[seq, torch.full((1, max_length - seq.size(1)), pad_token_id, dtype=seq.dtype, device=seq.device)],
dim=-1)
for seq in generated
]
output = torch.cat(generated, dim=0)
res = output.view(input_ids.size(0) * num_return_sequences, -1)
return res
def _stream(self, input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, **args):
start, first_seq = input_ids.shape[1], True
while input_ids.shape[1] < max_new_tokens - 1:
if first_seq:
out, first_seq = self(input_ids, **args), False
else:
out = self(input_ids[:, -1:], start_pos=input_ids.shape[1] - 1, **args)
logits = out.logits[:, -1, :]
logits[:, list(set(input_ids.tolist()[0]))] /= rp
logits /= (temperature + 1e-9)
if top_p is not None and top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
sorted_probs = F.softmax(sorted_logits, dim=-1)
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
sorted_indices_to_remove[:, 0] = False
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
logits[indices_to_remove] = -float('Inf')
input_ids_next = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)
input_ids = torch.cat((input_ids, input_ids_next), dim=1)
yield input_ids[:, start:]
if input_ids_next.item() == eos_token_id:
break