iomgaa/Minimind
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

experiment_1.3.0-1.3.2

Browse Source
This commit is contained in:
Yu Chengzhang 2025-07-13 21:28:46 +08:00
parent 5e464bbd3f
commit 2797b76939
10 changed files with 916 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
model/dataset.py
View File

@ -98,9 +98,14 @@ class PretrainDataset(Dataset):
def __getitem__(self, index):
sample = self.samples[index]
# 构建输入文本
text = f"{self.tokenizer.bos_token}{str(sample['text'])}{self.tokenizer.eos_token}"
text = str(sample['text'])
# 检查并添加<|im_start|>和<|im_end|>如果不存在
if not text.startswith(self.tokenizer.bos_token):
text = f"{self.tokenizer.bos_token}{text}"
if not text.endswith(self.tokenizer.eos_token):
text = f"{text}{self.tokenizer.eos_token}"
encoding = self.tokenizer(
text,
max_length=self.max_length,
@ -123,8 +128,8 @@ class SFTDataset(Dataset):
self.tokenizer = tokenizer
self.max_length = max_length
self.samples = self.load_data(jsonl_path)
self.bos_id = tokenizer('<s>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('</s>', add_special_tokens=False).input_ids
self.bos_id = tokenizer('<|im_start|>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('<|im_end|>', add_special_tokens=False).input_ids
def __len__(self):
return len(self.samples)
@ -191,8 +196,8 @@ class DPODataset(Dataset):
self.tokenizer = tokenizer
self.max_length = max_length
self.padding = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0
self.bos_id = tokenizer('<s>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('</s>', add_special_tokens=False).input_ids
self.bos_id = tokenizer('<|im_start|>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('<|im_end|>', add_special_tokens=False).input_ids
with open(file_path, 'r', encoding='utf-8') as f:
self.data = []
for line in f:
@ -502,8 +507,8 @@ class RLAIFDataset(Dataset):
self.tokenizer = tokenizer
self.max_length = max_length
self.samples = self.load_data(jsonl_path)
self.bos_id = tokenizer('<s>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('</s>', add_special_tokens=False).input_ids
self.bos_id = tokenizer('<|im_start|>assistant', add_special_tokens=False).input_ids
self.eos_id = tokenizer('<|im_end|>', add_special_tokens=False).input_ids
def __len__(self):
return len(self.samples)

4
model/model.py
View File

@ -540,7 +540,8 @@ class MiniMindLM(PreTrainedModel):
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))
# 统一不使用 aux_loss
aux_loss = 0
# 进一步简化，只保留必要的参数
output = CausalLMOutputWithPast(
@ -606,4 +607,3 @@ class MiniMindLM(PreTrainedModel):
yield input_ids[:, start:]
if input_ids_next.item() == eos_token_id:
break

479
model/model_no_feed.py Normal file
View File

@ -0,0 +1,479 @@
import math
import struct
import inspect
import time
import gc
#子空间二维分解+梯度更新
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
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)
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
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)
class KnowledgeDataset(nn.Module):
def __init__(self, params, tok_embeddings, is_train=True):
super().__init__()
self.is_train = is_train
self.params = params
self.tok_embeddings = tok_embeddings
# 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),
)
## 数据库参数
self.knowledge_num = params.knowledge_num
self.knowledge_length = params.knowledge_length
# 修改键存储为二维分解空间，设置为可训练参数
self.num_keys = int(math.sqrt(self.knowledge_num))
# 确保keys是可训练参数
self.keys = nn.Parameter(torch.randn(self.num_keys, 2, self.key_dim) * 0.02, requires_grad=True)
self.product_key_topk = min(16, self.num_keys)
# 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))
# 计算step数目，用于动态调整权重
self.step_counter = 0
# 移除批次计数器和更新频率相关代码
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
# 记录进入智能选择前的内存状态
if hasattr(self, 'step_counter'):
self.step_counter += 1
# 禁用GPU内存监控记录以提高性能
# if self.step_counter % 50 == 0: # 每50次调用记录一次
# if torch.cuda.is_available():
# allocated_before = torch.cuda.memory_allocated() / (1024**3)
# print(f"[INTEL_SELECT_ENTER] Step {self.step_counter}: GPU Memory: {allocated_before:.2f}GB")
# 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（保留这些变量以便其他地方使用）
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)
# 清理中间张量以防止内存泄漏
del all_candidate_indices, unique_indices, inverse_indices
del unique_candidate_features, normalized_candidates, normalized_queries
del batch_best_tokens, batch_best_tokens_embeddings
del flat_tokens, flat_embeddings, pre_update_embeddings
# 记录退出智能选择后的内存状态（已禁用以提高性能）
# if hasattr(self, 'step_counter') and self.step_counter % 50 == 0:
# if torch.cuda.is_available():
# allocated_after = torch.cuda.memory_allocated() / (1024**3)
# print(f"[INTEL_SELECT_EXIT] Step {self.step_counter}: GPU Memory: {allocated_after:.2f}GB")
# 强制垃圾回收（仅在监控步骤）
if hasattr(self, 'step_counter') and self.step_counter % 100 == 0:
gc.collect()
# if torch.cuda.is_available():
# torch.cuda.empty_cache()
return all_best_tokens, all_best_tokens_embeddings
def search_index(self, x):
batch_size, seq_len, dim = x.shape
# 1. 序列维度平均
x_flat = x.mean(dim=1) # [batch_size, dim]
# 2. 生成查询向量并重塑为两个子查询
queries = self.to_queries(x_flat) # [batch_size, knowledge_dim]
queries = queries.reshape(batch_size, 2, self.key_dim) # [batch_size, 2, key_dim]
# 调整维度顺序，使子空间维度位于首位
queries = queries.permute(1, 0, 2) # [2, batch_size, key_dim]
# 3. 计算每个子空间的相似度
sim = torch.einsum('p b d, k p d -> p b k', queries, self.keys)
# 4. 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]
# 5. 组合两个子空间的结果
all_scores = scores_x.unsqueeze(-1) + scores_y.unsqueeze(-2) # [batch_size, topk, topk]
all_indices = (indices_x.unsqueeze(-1) * self.num_keys) + indices_y.unsqueeze(-2) # [batch_size, topk, topk]
# 6. 将结果重塑为二维
all_scores = all_scores.reshape(batch_size, -1) # [batch_size, topk*topk]
all_indices = all_indices.reshape(batch_size, -1) # [batch_size, topk*topk]
# 7. 选择最终的top-k结果
scores, indices_of_indices = all_scores.topk(self.product_key_topk, dim=-1)
indices = torch.gather(all_indices, 1, indices_of_indices)
# 8. 应用智能分层选择策略
best_tokens, best_tokens_embeddings = self.intelligent_selection(x, scores, indices)
return best_tokens, best_tokens_embeddings
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)
# 监控交叉注意力开始时的内存（已禁用以提高性能）
if not hasattr(self, 'call_counter'):
self.call_counter = 0
self.call_counter += 1
# 禁用GPU内存监控记录以提高性能
# if self.call_counter % 100 == 0 and torch.cuda.is_available():
# allocated_before = torch.cuda.memory_allocated() / (1024**3)
# print(f"[CROSS_ATTN_ENTER] Call {self.call_counter}: GPU Memory: {allocated_before:.2f}GB")
# 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)
# 清理中间张量
del q, k, v, attn_scores, attn_weights
# 监控交叉注意力结束时的内存（已禁用以提高性能）
# if self.call_counter % 100 == 0 and torch.cuda.is_available():
# allocated_after = torch.cuda.memory_allocated() / (1024**3)
# print(f"[CROSS_ATTN_EXIT] Call {self.call_counter}: GPU Memory: {allocated_after:.2f}GB")
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 MiniMindBlock(nn.Module):
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.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)
# 移除 ffn_norm 和 feed_forward，因为不再使用 FeedForward 层
def forward(self, x, pos_cis):
h_attn = self.self_attention(
self.attention_norm(x),
pos_cis
)
db, db_embeddings = self.knowledge_dataset.search_index(h_attn)
h_attn = self.cross_attention(h_attn, db_embeddings)
h = x + h_attn
# 移除 FeedForward 层，直接返回注意力输出
return h
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
self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
self.dropout = nn.Dropout(params.dropout)
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.tok_embeddings.weight = self.output.weight
self.register_buffer("pos_cis",
precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta),
persistent=False)
self.OUT = CausalLMOutputWithPast()
self.freeze_embedding = False
def forward(self,
input_ids: Optional[torch.Tensor] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
step: int = 0,
**args):
start_pos = args.get('start_pos', 0)
# if self.freeze_embedding and step == 0:
# self.tok_embeddings.weight.requires_grad = False
# # 移除对knowledge_dataset.freeze_embedding的设置，让键更新由batch_counter控制
# # self.knowledge_dataset.freeze_embedding = True
# print("tok_embeddings.weight.requires_grad: ", self.tok_embeddings.weight.requires_grad)
h = self.dropout(self.tok_embeddings(input_ids))
pos_cis = self.pos_cis[start_pos:start_pos + input_ids.size(1)]
for l, layer in enumerate(self.layers):
h = layer(
h, pos_cis
)
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, :])
# 移除 aux_loss 计算，因为不再使用 FeedForward 层
aux_loss = 0
# 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
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, 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, 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:
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

5
model/model_original.py
View File

@ -322,7 +322,8 @@ class MiniMindLM(PreTrainedModel):
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))
# 统一不使用 aux_loss
aux_loss = 0
self.OUT.__setitem__('last_hidden_state', h)
self.OUT.__setitem__('logits', logits)
self.OUT.__setitem__('aux_loss', aux_loss)
@ -382,4 +383,4 @@ 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

47
run_file/experiment_1.3.0.sh Normal file
View File

@ -0,0 +1,47 @@
#!/bin/bash
# 激活conda环境
source $(conda info --base)/etc/profile.d/conda.sh
conda activate ycz_accelerate
# 设置环境变量以帮助调试
export NCCL_DEBUG=INFO
export PYTHONFAULTHANDLER=1
# 实验1.3.0 - 使用命令行参数直接配置accelerate
CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
--multi_gpu \
--num_processes=4 \
--mixed_precision=bf16 \
--main_process_port=29500 \
train_pretrain_accelerate.py \
--out_dir "out" \
--epochs 3 \
--embedding_epoch 2 \
--batch_size 48 \
--learning_rate 2e-4 \
--dtype bfloat16 \
--use_swanlab \
--swanlab_project "MiniMind-Pretrain" \
--num_workers 1 \
--accumulation_steps 32 \
--grad_clip 1.0 \
--warmup_iters 0 \
--log_interval 100 \
--save_interval 10000 \
--dim 1024 \
--n_layers 18 \
--max_seq_len 512 \
--use_moe False \
--data_path "./dataset/stable/merged_pretrain.jsonl" \
--profile \
--profile_interval 10 \
--use_flash_attn \
--knowledge_num 1048576 \
--knowledge_length 32 \
--database_init_path "./dataset/stable/sentence_trex_data.json" \
--fast_clustering \
--cluster_cache_path "./cache/cluster_tokens_single.pt" \
--memory_monitor_interval 10 \
--model_type "model_original" \
--model_size 814.724

47
run_file/experiment_1.3.1.sh Normal file
View File

@ -0,0 +1,47 @@
#!/bin/bash
# 激活conda环境
source $(conda info --base)/etc/profile.d/conda.sh
conda activate ycz_accelerate
# 设置环境变量以帮助调试
export NCCL_DEBUG=INFO
export PYTHONFAULTHANDLER=1
# 实验1.3.0 - 使用命令行参数直接配置accelerate
CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
--multi_gpu \
--num_processes=4 \
--mixed_precision=bf16 \
--main_process_port=29500 \
train_pretrain_accelerate.py \
--out_dir "out" \
--epochs 3 \
--embedding_epoch 2 \
--batch_size 48 \
--learning_rate 2e-4 \
--dtype bfloat16 \
--use_swanlab \
--swanlab_project "MiniMind-Pretrain" \
--num_workers 1 \
--accumulation_steps 32 \
--grad_clip 1.0 \
--warmup_iters 0 \
--log_interval 100 \
--save_interval 10000 \
--dim 1024 \
--n_layers 18 \
--max_seq_len 512 \
--use_moe False \
--data_path "./dataset/stable/merged_pretrain.jsonl" \
--profile \
--profile_interval 10 \
--use_flash_attn \
--knowledge_num 1048576 \
--knowledge_length 32 \
--database_init_path "./dataset/stable/sentence_trex_data.json" \
--fast_clustering \
--cluster_cache_path "./cache/cluster_tokens_single.pt" \
--memory_monitor_interval 10 \
--model_type "model_no_feed" \
--model_size 814.724

47
run_file/experiment_1.3.2.sh Normal file
View File

@ -0,0 +1,47 @@
#!/bin/bash
# 激活conda环境
source $(conda info --base)/etc/profile.d/conda.sh
conda activate ycz_accelerate
# 设置环境变量以帮助调试
export NCCL_DEBUG=INFO
export PYTHONFAULTHANDLER=1
# 实验1.3.0 - 使用命令行参数直接配置accelerate
CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
--multi_gpu \
--num_processes=4 \
--mixed_precision=bf16 \
--main_process_port=29500 \
train_pretrain_accelerate.py \
--out_dir "out" \
--epochs 3 \
--embedding_epoch 2 \
--batch_size 48 \
--learning_rate 2e-4 \
--dtype bfloat16 \
--use_swanlab \
--swanlab_project "MiniMind-Pretrain" \
--num_workers 1 \
--accumulation_steps 32 \
--grad_clip 1.0 \
--warmup_iters 0 \
--log_interval 100 \
--save_interval 10000 \
--dim 1024 \
--n_layers 18 \
--max_seq_len 512 \
--use_moe False \
--data_path "./dataset/stable/merged_pretrain.jsonl" \
--profile \
--profile_interval 10 \
--use_flash_attn \
--knowledge_num 1048576 \
--knowledge_length 32 \
--database_init_path "./dataset/stable/sentence_trex_data.json" \
--fast_clustering \
--cluster_cache_path "./cache/cluster_tokens_single.pt" \
--memory_monitor_interval 10 \
--model_type "model" \
--model_size 814.724

16
scripts/train_tokenizer.py
View File

@ -32,7 +32,7 @@ def train_tokenizer():
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
# 定义特殊token
special_tokens = ["<unk>", "<s>", "</s>"]
special_tokens = ["<unk>", "<|im_start|>", "<|im_end|>"]
# 设置训练器并添加特殊token
trainer = trainers.BpeTrainer(
@ -53,8 +53,8 @@ def train_tokenizer():
# 检查特殊token的索引
assert tokenizer.token_to_id("<unk>") == 0
assert tokenizer.token_to_id("<s>") == 1
assert tokenizer.token_to_id("</s>") == 2
assert tokenizer.token_to_id("<|im_start|>") == 1
assert tokenizer.token_to_id("<|im_end|>") == 2
# 保存tokenizer
tokenizer_dir = "../model/minimind_tokenizer"
@ -77,7 +77,7 @@ def train_tokenizer():
"special": True
},
"1": {
"content": "<s>",
"content": "<|im_start|>",
"lstrip": False,
"normalized": False,
"rstrip": False,
@ -85,7 +85,7 @@ def train_tokenizer():
"special": True
},
"2": {
"content": "</s>",
"content": "<|im_end|>",
"lstrip": False,
"normalized": False,
"rstrip": False,
@ -94,9 +94,9 @@ def train_tokenizer():
}
},
"additional_special_tokens": [],
"bos_token": "<s>",
"bos_token": "<|im_start|>",
"clean_up_tokenization_spaces": False,
"eos_token": "</s>",
"eos_token": "<|im_end|>",
"legacy": True,
"model_max_length": 32768,
"pad_token": "<unk>",
@ -104,7 +104,7 @@ def train_tokenizer():
"spaces_between_special_tokens": False,
"tokenizer_class": "PreTrainedTokenizerFast",
"unk_token": "<unk>",
"chat_template": "{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{{ '<s>system\\n' + system_message + '</s>\\n' }}{% else %}{{ '<s>system\\n你是 MiniMind，是一个有用的人工智能助手。</s>\\n' }}{% endif %}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<s>user\\n' + content + '</s>\\n<s>assistant\\n' }}{% elif message['role'] == 'assistant' %}{{ content + '</s>' + '\\n' }}{% endif %}{% endfor %}"
"chat_template": "{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{{ '<|im_start|>system\\n' + system_message + '<|im_end|>\\n' }}{% else %}{{ '<|im_start|>system\\n你是 MiniMind，是一个有用的人工智能助手。<|im_end|>\\n' }}{% endif %}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|im_start|>user\\n' + content + '<|im_end|>\\n<|im_start|>assistant\\n' }}{% elif message['role'] == 'assistant' %}{{ content + '<|im_end|>' + '\\n' }}{% endif %}{% endfor %}"
}
# 保存配置文件

272
train_pretrain_accelerate.py
View File

@ -200,14 +200,20 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non
with open(database_init_path, 'r', encoding='utf-8') as f:
database_data = json.load(f)
sentences_data = []
for data in database_data:
sentences_data.append(data['target'][0]['sentence'])
# 提取sentences列表
sentences_data = database_data.get('sentences', [])
# sentences_data = database_data.get('sentences', [])
Logger(f"Loaded {len(sentences_data)} sentences from database")
# 2. 按照importance_score进行排序（从高到低）
sorted_sentences = sorted(sentences_data, key=lambda x: x.get('importance_score', 0.0), reverse=True)
Logger(f"Sorted sentences by importance score (highest: {sorted_sentences[0].get('importance_score', 0.0)}, lowest: {sorted_sentences[-1].get('importance_score', 0.0)})")
try:
sorted_sentences = sorted(sentences_data, key=lambda x: x.get('importance_score', 0.0), reverse=True)
Logger(f"Sorted sentences by importance score (highest: {sorted_sentences[0].get('importance_score', 0.0)}, lowest: {sorted_sentences[-1].get('importance_score', 0.0)})")
except:
sorted_sentences = sentences_data
# 3. 处理每条数据，不进行聚类
Logger("Processing individual sentences...")
processed_rows = []
@ -218,16 +224,25 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non
# 处理所需数量的句子
num_to_process = min(knowledge_num, len(sorted_sentences))
# 添加截断统计变量
total_sentences = 0
truncated_sentences = 0
for i in range(num_to_process):
sentence_data = sorted_sentences[i]
sentence = sentence_data.get('corrected_sentence', '')
try:
sentence = sentence_data.get('corrected_sentence')
except:
sentence = sentence_data
# 将句子转换为tokens
sentence_tokens = tokenizer.encode(sentence, add_special_tokens=False)
# 截断或填充到knowledge_length
total_sentences += 1
if len(sentence_tokens) > knowledge_length:
# 如果超过长度，截断
truncated_sentences += 1
sentence_tokens = sentence_tokens[:knowledge_length]
Logger(f"Sentence {i+1} truncated from {len(tokenizer.encode(sentence, add_special_tokens=False))} to {knowledge_length} tokens")
else:
@ -254,6 +269,13 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non
# 转换为tensor
processed_tensor = torch.tensor(processed_rows, dtype=torch.long)
# 计算并打印截断句子的占比
truncation_ratio = truncated_sentences / total_sentences if total_sentences > 0 else 0.0
Logger(f"截断句子统计:")
Logger(f" - 总句子数: {total_sentences}")
Logger(f" - 截断句子数: {truncated_sentences}")
Logger(f" - 截断句子占比: {truncation_ratio:.4f} ({truncation_ratio*100:.2f}%)")
Logger(f"Data processing completed:")
Logger(f" - Processed {num_to_process} sentences")
Logger(f" - Added {knowledge_num - num_to_process} empty entries")
@ -285,6 +307,202 @@ def init_model(lm_config, pretrained_embedding_path=None, database_init_path=Non
tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
model = MiniMindLM(lm_config)
Logger(f'LLM总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} 百万')
elif args.model_type == "model_no_feed":
Logger(f"Using model type: {args.model_type}")
from model.model_no_feed import MiniMindLM, RMSNorm
tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
model = MiniMindLM(lm_config)
# 默认模型初始化
Logger("Performing default model initialization...")
# 初始化嵌入层权重
nn.init.normal_(model.tok_embeddings.weight, mean=0.0, std=0.02)
# 初始化输出层权重（如果不共享权重的话）
if not hasattr(model.tok_embeddings, 'weight') or model.output.weight is not model.tok_embeddings.weight:
nn.init.normal_(model.output.weight, mean=0.0, std=0.02)
# 初始化所有线性层
for name, module in model.named_modules():
if isinstance(module, nn.Linear):
# 使用Xavier/Glorot初始化
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
# 嵌入层使用正态分布初始化
nn.init.normal_(module.weight, mean=0.0, std=0.02)
elif isinstance(module, RMSNorm):
# RMSNorm的权重初始化为1
if hasattr(module, 'weight'):
nn.init.ones_(module.weight)
# 初始化位置编码相关参数
if hasattr(model.knowledge_dataset, 'keys'):
nn.init.normal_(model.knowledge_dataset.keys, mean=0.0, std=0.02)
Logger("Default model initialization completed")
# 如果提供了预训练的嵌入权重，加载它们
if pretrained_embedding_path:
Logger(f"Loading pretrained token embeddings from {pretrained_embedding_path}")
pretrained_embeddings = torch.load(pretrained_embedding_path)
model.tok_embeddings.weight.data.copy_(pretrained_embeddings)
model.output.weight.data.copy_(pretrained_embeddings) # 共享权重
if database_init_path:
import json
import os
# 数据库参数
knowledge_num = args.knowledge_num
knowledge_length = args.knowledge_length
# 检查是否使用缓存
cache_dir = os.path.dirname(args.cluster_cache_path)
if cache_dir:
os.makedirs(cache_dir, exist_ok=True)
processed_tensor = None
# 尝试加载缓存的处理结果
if not args.recompute_clusters and os.path.exists(args.cluster_cache_path):
try:
Logger(f"Loading cached processed results from {args.cluster_cache_path}")
processed_tensor = torch.load(args.cluster_cache_path)
# 验证缓存文件的形状是否可用
cached_knowledge_num, cached_knowledge_length = processed_tensor.shape
if cached_knowledge_length == knowledge_length:
if cached_knowledge_num >= knowledge_num:
# 缓存足够大，可以截取使用
processed_tensor = processed_tensor[:knowledge_num, :]
Logger(f"Successfully loaded cached data with shape {processed_tensor.shape}")
Logger(f"Truncated from cached shape ({cached_knowledge_num}, {cached_knowledge_length}) to required shape ({knowledge_num}, {knowledge_length})")
Logger("Skipping database initialization - using cached results")
else:
# 缓存太小，需要重新计算
Logger(f"Cached knowledge_num ({cached_knowledge_num}) < required knowledge_num ({knowledge_num}), recomputing...")
processed_tensor = None
else:
# knowledge_length不匹配，需要重新计算
Logger(f"Cached knowledge_length ({cached_knowledge_length}) != required knowledge_length ({knowledge_length}), recomputing...")
processed_tensor = None
except Exception as e:
Logger(f"Failed to load cached data: {e}, recomputing...")
processed_tensor = None
# 只有在没有有效缓存时才进行数据库初始化和处理
if processed_tensor is None:
Logger(f"Loading database initialization data from {database_init_path}")
# 1. 加载JSON文件
with open(database_init_path, 'r', encoding='utf-8') as f:
database_data = json.load(f)
sentences_data = []
for data in database_data:
sentences_data.append(data['target'][0]['sentence'])
# 提取sentences列表
# sentences_data = database_data.get('sentences', [])
Logger(f"Loaded {len(sentences_data)} sentences from database")
# 2. 按照importance_score进行排序（从高到低）
try:
sorted_sentences = sorted(sentences_data, key=lambda x: x.get('importance_score', 0.0), reverse=True)
Logger(f"Sorted sentences by importance score (highest: {sorted_sentences[0].get('importance_score', 0.0)}, lowest: {sorted_sentences[-1].get('importance_score', 0.0)})")
except:
sorted_sentences = sentences_data
# 3. 处理每条数据，不进行聚类
Logger("Processing individual sentences...")
processed_rows = []
# 获取空token的id（用于填充）
pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0
# 处理所需数量的句子
num_to_process = min(knowledge_num, len(sorted_sentences))
# 添加截断统计变量
total_sentences = 0
truncated_sentences = 0
for i in range(num_to_process):
sentence_data = sorted_sentences[i]
try:
sentence = sentence_data.get('corrected_sentence')
except:
sentence = sentence_data
# 将句子转换为tokens
sentence_tokens = tokenizer.encode(sentence, add_special_tokens=False)
# 截断或填充到knowledge_length
total_sentences += 1
if len(sentence_tokens) > knowledge_length:
# 如果超过长度，截断
truncated_sentences += 1
sentence_tokens = sentence_tokens[:knowledge_length]
Logger(f"Sentence {i+1} truncated from {len(tokenizer.encode(sentence, add_special_tokens=False))} to {knowledge_length} tokens")
else:
# 如果不足长度，用空token填充
original_length = len(sentence_tokens)
sentence_tokens.extend([pad_token_id] * (knowledge_length - len(sentence_tokens)))
if original_length < knowledge_length:
Logger(f"Sentence {i+1} padded from {original_length} to {knowledge_length} tokens")
processed_rows.append(sentence_tokens)
if (i + 1) % 1000 == 0:
Logger(f"Processed {i + 1}/{num_to_process} sentences")
# 如果句子数量不足，用空token填充剩余位置
while len(processed_rows) < knowledge_num:
empty_tokens = [pad_token_id] * knowledge_length
processed_rows.append(empty_tokens)
if len(processed_rows) % 1000 == 0:
Logger(f"Added empty entry {len(processed_rows)}/{knowledge_num}")
Logger(f"Finished adding empty entries. Total: {len(processed_rows)}/{knowledge_num}")
# 转换为tensor
processed_tensor = torch.tensor(processed_rows, dtype=torch.long)
# 计算并打印截断句子的占比
truncation_ratio = truncated_sentences / total_sentences if total_sentences > 0 else 0.0
Logger(f"截断句子统计:")
Logger(f" - 总句子数: {total_sentences}")
Logger(f" - 截断句子数: {truncated_sentences}")
Logger(f" - 截断句子占比: {truncation_ratio:.4f} ({truncation_ratio*100:.2f}%)")
Logger(f"Data processing completed:")
Logger(f" - Processed {num_to_process} sentences")
Logger(f" - Added {knowledge_num - num_to_process} empty entries")
Logger(f" - Final shape: {processed_tensor.shape}")
Logger(f" - Expected shape: ({knowledge_num}, {knowledge_length})")
# 保存处理结果到缓存文件
try:
torch.save(processed_tensor, args.cluster_cache_path)
Logger(f"Processed results saved to {args.cluster_cache_path}")
except Exception as e:
Logger(f"Failed to save processed results: {e}")
# 4. 初始化模型的knowledge_dataset
if hasattr(model, 'knowledge_dataset') and hasattr(model.knowledge_dataset, 'knowledge_dataset'):
model.knowledge_dataset.knowledge_dataset.data.copy_(processed_tensor)
Logger("Successfully initialized model.knowledge_dataset.knowledge_dataset with processed data")
else:
Logger("Warning: Could not find model.knowledge_dataset.knowledge_dataset to initialize")
# 存储为全局变量作为备选
globals()['processed_database'] = processed_tensor
Logger(f"Database embeddings and sentences stored in model")
Logger(f'LLM总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} 百万')
return model, tokenizer
@ -312,7 +530,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
optimizer_end = torch.cuda.Event(enable_timing=True)
# 预取数据
prefetch_factor = 2 # 预取的批次数
prefetch_factor = 8 # 预取的批次数
data_iter = iter(train_loader)
prefetch_batches = []
@ -396,14 +614,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
Y.view(-1)
).view(Y.size())
loss = (loss * loss_mask).sum() / loss_mask.sum()
# 添加辅助损失，如果存在的话
try:
aux_loss = sum(l.feed_forward.aux_loss for l in model.module.layers
if hasattr(l, 'feed_forward') and hasattr(l.feed_forward, 'aux_loss'))
loss += aux_loss
except Exception as e:
Logger(f"Warning: Could not add auxiliary loss: {e}")
# 如果出错，不添加辅助损失
# 移除辅助损失计算，统一不使用 aux_loss
loss = loss / args.accumulation_steps
# 计时前向传播结束 (只在主进程进行)
@ -603,13 +814,15 @@ def main():
parser.add_argument("--use_flash_attn", action="store_true", default=True, help="启用FlashAttention")
parser.add_argument("--knowledge_num", type=int, default=960400,help="知识库的数据数目")
parser.add_argument("--knowledge_length", type=int, default=32,help="知识库的句子长度")
parser.add_argument("--database_init_path", type=str, default="./dataset/combined_prepare.json", help="数据库初始化路径")
parser.add_argument("--database_init_path", type=str, default="./dataset/stable/sentence_trex_data.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_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")
parser.add_argument("--memory_monitor", action="store_true", default=False, help="启用内存监控")
parser.add_argument("--memory_monitor_interval", type=int, default=10, help="内存监控间隔（步数）")
parser.add_argument("--model_type", type=str, default="model", help="使用什么模型训练") #model,model_original
parser.add_argument("--model_type", type=str, default="model", help="使用什么模型训练") #model,model_original,model_no_feed
parser.add_argument("--model_size", type=float, default=50.0, help="模型大小")
parser.add_argument("--swanlab_online", type=bool, default=False, help="是否使用在线SwanLab服务")
args = parser.parse_args()
#########################################################
@ -678,16 +891,23 @@ def main():
# 初始化SwanLab实验实例
swanlab_run = None
if args.use_swanlab and accelerator.is_main_process:
# 初始化SwanLab
swanlab_run = swanlab.init(
project=args.swanlab_project,
experiment_name=args.swanlab_run_name,
description="MiniMind预训练实验，使用本地部署的SwanLab进行可视化",
config=config_dict
# 设置SwanLab服务器地址和API Key
# host="http://100.123.118.114:11071",
# api_key="LesBT7HRq23HNBrOPKP8S"
)
if args.swanlab_online:
# 使用在线SwanLab服务
# 初始化SwanLab
swanlab_run = swanlab.init(
project=args.swanlab_project,
experiment_name=args.swanlab_run_name,
description="MiniMind预训练实验，使用本地部署的SwanLab进行可视化",
config=config_dict
)
else:
swanlab_run = swanlab.init(
project=args.swanlab_project,
experiment_name=args.swanlab_run_name,
description="MiniMind预训练实验，使用本地部署的SwanLab进行可视化",
config=config_dict,
mode="offline"
)
else:
swanlab_run = None

23
uv.lock generated
View File

@ -12,6 +12,15 @@ resolution-markers = [
"python_full_version < '3.11' and sys_platform != 'linux'",
]
[[package]]
name = "absl-py"
version = "2.3.1"
source = { registry = "https://pypi.org/simple" }
sdist = { url = "https://files.pythonhosted.org/packages/10/2a/c93173ffa1b39c1d0395b7e842bbdc62e556ca9d8d3b5572926f3e4ca752/absl_py-2.3.1.tar.gz", hash = "sha256:a97820526f7fbfd2ec1bce83f3f25e3a14840dac0d8e02a0b71cd75db3f77fc9", size = 116588, upload-time = "2025-07-03T09:31:44.05Z" }
wheels = [
{ url = "https://files.pythonhosted.org/packages/8f/aa/ba0014cc4659328dc818a28827be78e6d97312ab0cb98105a770924dc11e/absl_py-2.3.1-py3-none-any.whl", hash = "sha256:eeecf07f0c2a93ace0772c92e596ace6d3d3996c042b2128459aaae2a76de11d", size = 135811, upload-time = "2025-07-03T09:31:42.253Z" },
]
[[package]]
name = "accelerate"
version = "1.7.0"
@ -1794,6 +1803,7 @@ dependencies = [
{ name = "regex" },
{ name = "requests" },
{ name = "rich" },
{ name = "rouge-score" },
{ name = "rpds-py" },
{ name = "s3transfer" },
{ name = "safetensors" },
@ -1965,6 +1975,7 @@ requires-dist = [
{ name = "regex", specifier = "==2024.11.6" },
{ name = "requests", specifier = "==2.32.3" },
{ name = "rich", specifier = "==13.7.1" },
{ name = "rouge-score", specifier = ">=0.1.2" },
{ name = "rpds-py", specifier = "==0.24.0" },
{ name = "s3transfer", specifier = "==0.13.0" },
{ name = "safetensors", specifier = "==0.5.3" },
@ -3496,6 +3507,18 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/87/67/a37f6214d0e9fe57f6ae54b2956d550ca8365857f42a1ce0392bb21d9410/rich-13.7.1-py3-none-any.whl", hash = "sha256:4edbae314f59eb482f54e9e30bf00d33350aaa94f4bfcd4e9e3110e64d0d7222", size = 240681, upload-time = "2024-02-28T14:51:14.353Z" },
]
[[package]]
name = "rouge-score"
version = "0.1.2"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "absl-py" },
{ name = "nltk" },
{ name = "numpy" },
{ name = "six" },
]
sdist = { url = "https://files.pythonhosted.org/packages/e2/c5/9136736c37022a6ad27fea38f3111eb8f02fe75d067f9a985cc358653102/rouge_score-0.1.2.tar.gz", hash = "sha256:c7d4da2683e68c9abf0135ef915d63a46643666f848e558a1b9f7ead17ff0f04", size = 17400, upload-time = "2022-07-22T22:46:22.909Z" }
[[package]]
name = "rpds-py"
version = "0.24.0"