diff --git a/.vscode/launch.json b/.vscode/launch.json new file mode 100644 index 0000000..cac702c --- /dev/null +++ b/.vscode/launch.json @@ -0,0 +1,102 @@ +{ + "version": "0.2.0", + "configurations": [ + { + "name": "Debug Train Pretrain Accelerate", + "type": "python", + "request": "launch", + "program": "${workspaceFolder}/train_pretrain_accelerate.py", + "console": "integratedTerminal", + "python": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "cwd": "${workspaceFolder}", + "env": { + "PYTHONPATH": "${workspaceFolder}", + "CUDA_VISIBLE_DEVICES": "0" + }, + "justMyCode": false, + "stopOnEntry": false, + "redirectOutput": true + }, + { + "name": "Debug Train Pretrain Accelerate (Multi-GPU)", + "type": "python", + "request": "launch", + "program": "${workspaceFolder}/train_pretrain_accelerate.py", + "console": "integratedTerminal", + "python": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "args": [ + "--hidden_size", "512", + "--max_seq_len", "512", + "--n_layers", "8", + "--batch_size", "8", + "--epochs", "1" + ], + "cwd": "${workspaceFolder}", + "env": { + "PYTHONPATH": "${workspaceFolder}", + "CUDA_VISIBLE_DEVICES": "0,1" + }, + "justMyCode": false, + "stopOnEntry": false, + "redirectOutput": true + }, + { + "name": "Debug Train Pretrain Accelerate (Small Test)", + "type": "python", + "request": "launch", + "program": "${workspaceFolder}/train_pretrain_accelerate.py", + "console": "integratedTerminal", + "python": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "args": [ + "--hidden_size", "512", + "--max_seq_len", "512", + "--n_layers", "8", + "--batch_size", "2", + "--epochs", "1", + "--log_interval", "10", + "--save_interval", "100", + "--accumulation_steps", "4" + ], + "cwd": "${workspaceFolder}", + "env": { + "PYTHONPATH": "${workspaceFolder}", + "CUDA_VISIBLE_DEVICES": "0", + "WANDB_MODE": "offline" + }, + "justMyCode": false, + "stopOnEntry": false, + "redirectOutput": true + }, + { + "name": "Debug ExtractDB Comparison", + "type": "python", + "request": "launch", + "program": "${workspaceFolder}/train_pretrain_accelerate.py", + "console": "integratedTerminal", + "python": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "args": [ + "--hidden_size", "512", + "--max_seq_len", "256", + "--n_layers", "4", + "--batch_size", "2", + "--epochs", "1", + "--log_interval", "10", + "--save_interval", "200", + "--accumulation_steps", "2", + "--comparison_mode", + "--knowledge_num", "256", + "--knowledge_length", "64", + "--comparison_mode" + ], + "cwd": "${workspaceFolder}", + "env": { + "PYTHONPATH": "${workspaceFolder}", + "CUDA_VISIBLE_DEVICES": "0", + "WANDB_MODE": "offline" + }, + "justMyCode": false, + "stopOnEntry": false, + "redirectOutput": true + } + ] +} \ No newline at end of file diff --git a/.vscode/settings.json b/.vscode/settings.json new file mode 100644 index 0000000..d5a2ff8 --- /dev/null +++ b/.vscode/settings.json @@ -0,0 +1,18 @@ +{ + "python.pythonPath": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "python.defaultInterpreterPath": "/home/iomgaa/miniconda3/envs/accelerate/bin/python", + "python.terminal.activateEnvironment": true, + "python.terminal.activateEnvInCurrentTerminal": true, + "python.linting.enabled": true, + "python.linting.pylintEnabled": false, + "python.linting.flake8Enabled": true, + "python.formatting.provider": "black", + "python.analysis.autoImportCompletions": true, + "python.analysis.typeCheckingMode": "off", + "files.exclude": { + "**/__pycache__": true, + "**/*.pyc": true, + "**/.git": false, + "**/wandb": false + } +} \ No newline at end of file diff --git a/model/LMConfig.py b/model/LMConfig.py index 1312e97..e9cc3f5 100644 --- a/model/LMConfig.py +++ b/model/LMConfig.py @@ -39,6 +39,7 @@ class LMConfig(PretrainedConfig): #################################################### knowledge_num: int = 64*64, knowledge_length: int = 8, + knowledge_dim: int = 128, **kwargs, ): self.dim = dim @@ -72,4 +73,5 @@ class LMConfig(PretrainedConfig): #################################################### self.knowledge_num = knowledge_num self.knowledge_length = knowledge_length + self.knowledge_dim = knowledge_dim super().__init__(**kwargs) diff --git a/model/model.py b/model/model.py index 814674d..625be9a 100644 --- a/model/model.py +++ b/model/model.py @@ -11,14 +11,9 @@ 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__() @@ -31,7 +26,7 @@ class RMSNorm(torch.nn.Module): 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 @@ -39,7 +34,7 @@ def precompute_pos_cis(dim: int, end: int = int(32 * 1024), theta: float = 1e6): 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 @@ -55,200 +50,166 @@ def apply_rotary_emb(xq, xk, pos_cis): 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): +class KnowledgeDataset(nn.Module): + def __init__(self, params, tok_embeddings, is_train=True): 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) + self.is_train = is_train + self.params = params + self.tok_embeddings = tok_embeddings - 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) - # kv_cache实现 REMOVED - # if past_key_value is not None: - # xk = torch.cat([past_key_value[0], xk], dim=1) - # xv = torch.cat([past_key_value[1], xv], dim=1) - # past_kv = (xk, xv) if use_cache else None - - # 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) + # 嵌入参数 + self.knowledge_dim = params.knowledge_dim + self.key_dim = self.knowledge_dim // 2 + self.to_queries = nn.Sequential( + nn.Linear(params.dim, self.knowledge_dim, bias=False), ) - # 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] + ## 数据库参数 + self.knowledge_num = params.knowledge_num + self.knowledge_length = params.knowledge_length + self.keys = nn.Parameter(torch.randn(self.knowledge_num, self.knowledge_dim) * 0.02, requires_grad=True) + self.product_key_topk = min(16, self.knowledge_num) + + # 使用频率统计 - 使用register_buffer以便在GPU/CPU间正确移动 + self.register_buffer('has_update_keys', torch.zeros(self.knowledge_num)) - # 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]) + # 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('knowledge_dataset', + torch.randint(low=0, high=params.vocab_size, size=(self.knowledge_num, self.knowledge_length), dtype=torch.long) + ) - # 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 + # 计算step数目，用于动态调整权重 + self.step_counter = 0 - # 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 + def intelligent_selection(self, query, all_scores, all_indices): + """智能分层选择策略""" + if self.is_train == False: + return all_scores, all_indices + + batch_size = all_scores.size(0) + device = all_scores.device + dtype = all_scores.dtype + # 对每个batch进行分层选择 + enhanced_scores = all_scores.clone() + query_features = query.mean(dim=1) # [batch_size, dim] + # 预先计算所有候选条目的嵌入（批量优化） + all_candidate_indices = torch.cat([all_indices[i] for i in range(batch_size)], dim=0) + unique_indices, inverse_indices = torch.unique(all_candidate_indices, return_inverse=True) + # 批量计算唯一候选条目的嵌入 + candidate_tokens = self.knowledge_dataset[unique_indices] + flat_tokens = candidate_tokens.view(-1) + flat_embeddings = self.tok_embeddings(flat_tokens) + #获取flat_tokens对应的index + pre_update_indices = unique_indices.view(-1) + pre_update_embeddings = flat_embeddings.view( + len(unique_indices), self.knowledge_length, -1 + ) + + unique_candidate_features = flat_embeddings.view( + len(unique_indices), self.knowledge_length, -1 + ).mean(dim=1) # [num_unique_candidates, dim] + + # 归一化候选特征（优化相似度计算） + normalized_candidates = F.normalize(unique_candidate_features, dim=-1) + normalized_queries = F.normalize(query_features, dim=-1) + + # 收集所有batch的best_tokens + batch_best_tokens = [] + batch_best_tokens_embeddings = [] + + for batch_idx in range(batch_size): + indices = all_indices[batch_idx] + + # 获取当前batch候选条目对应的特征索引 + start_idx = batch_idx * len(indices) + end_idx = start_idx + len(indices) + batch_inverse_indices = inverse_indices[start_idx:end_idx] + + # 使用预计算的归一化特征进行优化相似度计算 + batch_candidate_features = normalized_candidates[batch_inverse_indices] + query_feature = normalized_queries[batch_idx] + + # 使用矩阵乘法计算余弦相似度 + similarity_scores = torch.mv(batch_candidate_features, query_feature) + + # 找到最大相似度分数的索引 + max_similarity_idx = torch.argmax(similarity_scores) + + # 获取最大相似度对应的候选条目索引 + best_candidate_idx = indices[max_similarity_idx] + + # 获取对应的tokens + best_tokens = self.knowledge_dataset[best_candidate_idx] + best_tokens_embeddings = self.tok_embeddings(best_tokens) + + # 将当前batch的best_tokens添加到列表中 + batch_best_tokens.append(best_tokens) + batch_best_tokens_embeddings.append(best_tokens_embeddings) + + # 将所有batch的best_tokens堆叠成一个张量 + # [batch_size, knowledge_length] + all_best_tokens = torch.stack(batch_best_tokens, dim=0) + all_best_tokens_embeddings = torch.stack(batch_best_tokens_embeddings, dim=0) + + # 获取 + + # 使用重新计算的embeddings更新self.keys + if self.is_train: + self._update_keys_with_embeddings(pre_update_indices, pre_update_embeddings) + + # 更新被修改过的key + with torch.no_grad(): + self.has_update_keys[pre_update_indices] = 1 + + return all_best_tokens, all_best_tokens_embeddings + + def _update_keys_with_embeddings(self, pre_update_indices, pre_update_embeddings): + # 使用pre_update_embeddings更新self.keys + with torch.no_grad(): + pre_update_embeddings = pre_update_embeddings.mean(dim=1) # [337, 512] + pre_update_embeddings = self.to_queries(pre_update_embeddings) + self.keys[pre_update_indices] = pre_update_embeddings + + def search_index(self,x): + 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(batch_size, 2, self.key_dim) + # queries = queries.permute(1, 0, 2) + + # 2. 计算queries与keys的相似度 + sim = torch.einsum('b d, k d -> b k', queries, self.keys) + + # 3. 在两个子空间分别做top-k + scores_and_indices = sim.topk(self.product_key_topk, dim=-1) + scores, indices = scores_and_indices[0], scores_and_indices[1] + + # 5. 应用智能分层选择策略 + best_tokens, best_tokens_embeddings = self.intelligent_selection(x, scores, indices) + + # 6. 更新1%的keys + if self.is_train: + # 获取未更新过的keys的索引 + not_updated_indices = torch.where(self.has_update_keys == 0)[0] + # 如果有未更新的keys，随机选择num_update_keys个进行更新 + if len(not_updated_indices) > 0: + num_update_keys = int(self.knowledge_num * 0.01) + perm = torch.randperm(len(not_updated_indices))[:num_update_keys] + pre_update_indices = not_updated_indices[perm] + pre_update_tokens = self.knowledge_dataset[pre_update_indices] + pre_update_embeddings = self.tok_embeddings(pre_update_tokens.view(-1)) + pre_update_embeddings = pre_update_embeddings.view(num_update_keys, self.knowledge_length, -1) + self._update_keys_with_embeddings(pre_update_indices, pre_update_embeddings) + + return best_tokens, best_tokens_embeddings class CrossAttention(nn.Module): def __init__( @@ -295,6 +256,58 @@ class CrossAttention(nn.Module): return context +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): + bsz, seq_len, _ = x.shape + xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) + xq = xq.view(bsz, seq_len, self.n_local_heads, self.head_dim) + xk = xk.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim) + xv = xv.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim) + + xq, xk = apply_rotary_emb(xq, xk, pos_cis) + 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 FeedForward(nn.Module): def __init__(self, config: LMConfig): super().__init__() @@ -427,169 +440,30 @@ class MOEFeedForward(nn.Module): class MiniMindBlock(nn.Module): - def __init__(self, layer_id: int, config: LMConfig): + def __init__(self, layer_id: int, config: LMConfig, knowledge_dataset: KnowledgeDataset): 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.self_attention = Attention(config) + self.cross_attention = CrossAttention(config) + self.knowledge_dataset = knowledge_dataset 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) - # 假设num_experts是已定义的总专家数量的平方根 - - - # 查询生成的参数 - - - # 创建查询生成模块 - # if weight_down_embed is not None: - # self.to_queries = nn.Sequential( - # nn.Linear(config.dim, self.dim_key * 2, bias=False), - # # nn.Unflatten(2, (2, self.n_heads, self.dim_key)) # 替代Rearrange - # ) - - # # 超参数 - # self.product_key_topk = min(16, self.num_keys) # 确保不超过num_keys - # self.num_experts_per_head_topk = 1 # 最终每个头选取的专家数 - - def forward(self, x, db_value, pos_cis): - # import pdb;pdb.set_trace() - # db_value = None - - # # 如果有weight_down_embed，使用Product Key机制 - # if self.weight_down_embed is not None: - # # 1. 生成queries - # 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(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. 最终top-k选择 - # scores, pk_indices = all_scores.topk(self.num_experts_per_head_topk, dim=-1) - # indices = all_indices.gather(-1, pk_indices) - - # # 6. 从embedding中获取专家值 - - # # 从embedding中获取值 - # flat_indices = indices.view(-1) # 将索引展平为一维张量 - # db_values = self.weight_down_embed(flat_indices) - - # # 重塑回原始形状 - # db_value = db_values.view(batch_size, -1, dim) - - - # 注意力计算 - h_attn = self.attention( + def forward(self, x, pos_cis): + h_attn = self.self_attention( self.attention_norm(x), - pos_cis, - db_value=db_value + pos_cis ) - - h_attn = self.cross_att(h_attn, db_value) - - # 残差连接 + db, db_embeddings = self.knowledge_dataset.search_index(h_attn) + h_attn = self.cross_attention(h_attn, db_embeddings) 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): - # 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 - - # 使用register_buffer代替nn.Parameter，避免梯度问题 - # self.register_buffer('weight_down_embed', torch.randn(self.knowledge_num, self.knowledge_length) * 0.02) - 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), - ) - - 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. 最终top-k选择 - scores, pk_indices = all_scores.topk(self.num_experts_per_head_topk, dim=-1) - indices = all_indices.gather(-1, pk_indices) - flat_indices = indices.view(-1) - 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 - + return out class MiniMindLM(PreTrainedModel): @@ -601,46 +475,15 @@ class MiniMindLM(PreTrainedModel): self.vocab_size, self.n_layers = params.vocab_size, params.n_layers self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim) self.dropout = nn.Dropout(params.dropout) - # 移除旧的weight_down_embed声明 - self.extract_db = ExtractDB(self.params) - - # 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.knowledge_dataset = KnowledgeDataset(params, self.tok_embeddings) + self.layers = nn.ModuleList([MiniMindBlock(l, params, self.knowledge_dataset) 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), + self.register_buffer("pos_cis", + precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta), persistent=False) - self.params = params + self.OUT = CausalLMOutputWithPast() def forward(self, input_ids: Optional[torch.Tensor] = None, @@ -648,63 +491,12 @@ class MiniMindLM(PreTrainedModel): **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 = [] - + pos_cis = self.pos_cis[start_pos:start_pos + input_ids.size(1)] for l, layer in enumerate(self.layers): - # 禁用数据库模式，使用固定值替代数据库查询 - if self.params.disable_db: - # 创建一个形状为[batch_size, n_layers, dim]的tensor，所有元素值为1e-4 - batch_size = h.size(0) - db_value = torch.full((batch_size, self.n_layers, self.params.dim), 1e-4, - dtype=h.dtype, device=h.device) - else: - # 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, pos_cis ) - 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)) @@ -717,12 +509,6 @@ class MiniMindLM(PreTrainedModel): 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() @@ -755,13 +541,14 @@ class MiniMindLM(PreTrainedModel): 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 + start, first_seq, past_kvs = input_ids.shape[1], True, None 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, :] + out = self(input_ids[:, -1:], + start_pos=input_ids.shape[1] - 1, **args) + logits, past_kvs = out.logits[:, -1, :], out.past_key_values logits[:, list(set(input_ids.tolist()[0]))] /= rp logits /= (temperature + 1e-9) if top_p is not None and top_p < 1.0: @@ -777,4 +564,4 @@ class MiniMindLM(PreTrainedModel): input_ids = torch.cat((input_ids, input_ids_next), dim=1) yield input_ids[:, start:] if input_ids_next.item() == eos_token_id: - break + break \ No newline at end of file diff --git a/train_pretrain_accelerate.py b/train_pretrain_accelerate.py index ab1cc81..d310c6c 100644 --- a/train_pretrain_accelerate.py +++ b/train_pretrain_accelerate.py @@ -74,8 +74,8 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non nn.init.ones_(module.weight) # 初始化位置编码相关参数 - if hasattr(model.extract_db, 'keys'): - nn.init.normal_(model.extract_db.keys, mean=0.0, std=0.02) + if hasattr(model.knowledge_dataset, 'keys'): + nn.init.normal_(model.knowledge_dataset.keys, mean=0.0, std=0.02) Logger("Default model initialization completed") @@ -434,11 +434,11 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non Logger(f"Failed to save cluster results: {e}") # 3. 初始化模型的weight_down_embed - if hasattr(model, 'extract_db') and hasattr(model.extract_db, 'weight_down_embed'): - model.extract_db.weight_down_embed.data.copy_(clustered_tensor) - Logger("Successfully initialized model.extract_db.weight_down_embed with clustered data") + if hasattr(model, 'knowledge_dataset') and hasattr(model.knowledge_dataset, 'knowledge_dataset'): + model.knowledge_dataset.knowledge_dataset.data.copy_(clustered_tensor) + Logger("Successfully initialized model.knowledge_dataset.knowledge_dataset with clustered data") else: - Logger("Warning: Could not find model.extract_db.weight_down_embed to initialize") + Logger("Warning: Could not find model.knowledge_dataset.knowledge_dataset to initialize") # 存储为全局变量作为备选 globals()['clustered_database'] = clustered_tensor @@ -659,21 +659,21 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a def main(): parser = argparse.ArgumentParser(description="MiniMind Pretraining with Accelerate") parser.add_argument("--out_dir", type=str, default="out") - parser.add_argument("--epochs", type=int, default=3) - parser.add_argument("--batch_size", type=int, default=24) + parser.add_argument("--epochs", type=int, default=4) + parser.add_argument("--batch_size", type=int, default=48) parser.add_argument("--learning_rate", type=float, default=2e-4) parser.add_argument("--dtype", type=str, default="bfloat16") parser.add_argument("--use_wandb", default=True, action="store_true") parser.add_argument("--wandb_project", type=str, default="MiniMind-Pretrain") - parser.add_argument("--num_workers", type=int, default=48) + parser.add_argument("--num_workers", type=int, default=8) parser.add_argument("--accumulation_steps", type=int, default=32) parser.add_argument("--grad_clip", type=float, default=1.0) parser.add_argument("--warmup_iters", type=int, default=0) parser.add_argument("--log_interval", type=int, default=100) parser.add_argument("--save_interval", type=int, default=10000) - parser.add_argument('--dim', default=1024, type=int) - parser.add_argument('--n_layers', default=32, type=int) - parser.add_argument('--max_seq_len', default=1024, type=int) + parser.add_argument('--dim', default=512, type=int) + parser.add_argument('--n_layers', default=8, type=int) + parser.add_argument('--max_seq_len', default=512, type=int) parser.add_argument('--use_moe', default=False, type=bool) parser.add_argument('--disable_db', action='store_true', help="禁用数据库功能，使用固定值1e-4替代") parser.add_argument("--data_path", type=str, default="./dataset/pretrain_hq.jsonl") @@ -681,11 +681,11 @@ def main(): parser.add_argument("--profile", action="store_true", default=True, help="启用性能分析") parser.add_argument("--profile_interval", type=int, default=10, help="性能分析打印间隔（步数）") parser.add_argument("--use_flash_attn", action="store_true", default=True, help="启用FlashAttention") - parser.add_argument("--knowledge_num", type=int, default=65536,help="知识库的数据数目") - parser.add_argument("--knowledge_length", type=int, default=64,help="知识库的句子长度") + parser.add_argument("--knowledge_num", type=int, default=4096,help="知识库的数据数目") + parser.add_argument("--knowledge_length", type=int, default=16,help="知识库的句子长度") parser.add_argument("--database_init_path", type=str, default="./dataset/database_init.json", help="数据库初始化路径") parser.add_argument("--fast_clustering", action="store_true", default=True, help="使用快速近似聚类算法（适用于大数据集）") - parser.add_argument("--cluster_cache_path", type=str, default="./cache/cluster_tokens.pt", help="聚类结果缓存文件路径") + parser.add_argument("--cluster_cache_path", type=str, default="./cache/cluster_tokens_single.pt", help="聚类结果缓存文件路径") parser.add_argument("--recompute_clusters", action="store_true", default=False, help="强制重新计算聚类，忽略缓存文件") args = parser.parse_args()