给定一个问题(question)及其对应的选项内容(options)，大模型无法把选项内容(option content)和对应的选项标识符(symbol，特指选项标识A/B/C/D)关联到一起。例如，当把正确答案"the president"放到选项B时，模型能够正确选择出答案；当我们把正确答案放到C时，模型依然选择"B"，即模型偏向于选"B"或者第二个答案，而忽略了正确答案的内容。

lost in the middle

Lost in the Middle: How Language Models Use Long Contexts

reasoning boundary

NeurIPS 2024 (Oral) | 如何量化与提升思维链的推理能力边界？

Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought

https://github.com/LightChen233/reasoning-boundary

语言≠思维

语言≠思维，大模型学不了推理：一篇Nature让AI社区炸锅了

https://www.nature.com/articles/s41586-024-07522-w

一些其他比较重要的工作

几篇出现频率比较高的论文

Scaling instruction-finetuned language models 引用数800+

How can we know what language models know? 引用数800+

Chain of thought prompting elicits reasoning in large language models引用1800+

Anthropic的一些工作

Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

Studying Large Language Model Generalization with Influence Functions

Measuring Faithfulness in Chain-of-Thought Reasoning

从Claude 3中提取数百万特征，首次详细理解大模型的「思维」

Scaling Dictionary Learning to Claude 3 Sonnet

LLM惊现篡改代码获得奖励，欺骗人类无法根除逆转！Anthropic新作揭露惊人真相

SYCOPHANCY TO SUBTERFUGE: INVESTIGATING REWARD TAMPERING IN LANGUAGE MODELS

个性化搜索

随便找一篇Denoising Attention for Query-aware User Modeling in Personalized Search，来看看它的参考文献：

学术界：

工业界：

A Zero Attention Model for Personalized Product Search，CIKM19，亚马逊
Real-time Personalization using Embeddings for Search Ranking at Airbnb，KDD18，airbnb
End-to-End Deep Attentive Personalized Item Retrieval for Online Content-sharing Platforms，www20，Google
Towards Personalized and Semantic Retrieval: An End-to-End Solution for E-commerce Search via Embedding Learning，sigir20，京东
Personalized Query Suggestions，sigir20，LinkedIn
A GNN-based Multi-task Learning Framework for Personalized Video Search，WSDM22，百度

q-i双塔的改进

引入location+social

Embedding-based Retrieval in Facebook Search，KDD20

q-d双塔结构，在两个塔的最底层均加入：

location：用户所处地理位置，如城市
social：facebook是社交网络，通过另一个基于graph的模型训练得到的user和item emb，直接加进来

引入用户行为序列

Encoding History with Context-aware Representation Learning for Personalized Search，sigir20，人大，提出HTPS

把用户历史的q-d pair对和当前query一起，过短期transformer和长期transformer得到输出 $q^l$ 。

把 $q^l$ 加上[mask]，过transoformer得到预估的intent $q^p$

然后将 $q^l$ 和 $q^p$ 通过gate nn融合得到最终的context-aware的query表示 $q^f$

最终doc和query的打分包括两部分，通过 $\phi$ （一个MLP，激活是tanh）进行融合：

p(d \mid q, H)=\phi\left(p(d, q), p\left(d, q^H\right)\right)

$p(d, q)$ ：q和d的语义相似度，可以用正常的nlp模型得到
$p\left(d, q^H\right)$ ：q和d的个性化得分，公式如下，其中 $s^R$ 是cos：

p\left(d, q^H\right)=\phi\left(s^R\left(q^s, d^w\right), s^R\left(q^l, d^w\right), s^R\left(q^p, d^w\right), s^R\left(q^f, d^w\right)\right)

有两个loss：

pred loss：预估intent，即下一个query，拿 $q^p$ 与下一个query中各个词向量的avg算cos
rank loss：依据 $p(d \mid q, H)$ 算lambda rank的pairwise loss

三塔+gnn邻居+mtl

A GNN-based Multi-task Learning Framework for Personalized Video Search，WSDM22，百度，提出MGNN-PVS

现有的PSM(g personalized search methods)大多使用用户反馈（如点击）进行训练，缺点：

反馈信号大部分表达的是吸引力而非相关性
用户的历史信号比较稀疏，很难学好PSM

两张二部图：u-q和q-d

3个塔：

user：
- user自己
- 一跳邻居（u->q）的q
- 二跳邻居（u->q->u）的u
query：
- query自己
- 一跳邻居（q->d）的doc
- 二跳邻居（q->d->q）的query
doc：
- doc自己的title向量（训练query-正title-负title的triplet loss）和video向量（训练video-正query-负query的triplet loss）
- 二跳邻居（d->q->d）的doc

两个task：

ctr预估：u和q拼一起过nn得到个性化的q，再和d过nn得到的向量算内积，得到预估值，用交叉熵
相关性预估：q过另一个nn，d过另一个nn，内积，用mse

LLM模型融合

https://github.com/arcee-ai/mergekit

SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling

LLM 合并新思路：进化算法+零训练->新任务

Evolutionary Optimization of Model Merging Recipes

https://github.com/SakanaAI/evolutionary-model-merge

LLM auto-ml

LLaMA-NAS

用神经架构搜索给LLM瘦身，模型变小，准确度有时反而更高

LLaMA-NAS: Efficient Neural Architecture Search for Large Language Models

SELA

MetaGPT开源SELA，用AI设计AI，效果超越OpenAI使用的AIDE

SELA: Tree-Search Enhanced LLM Agents for Automated Machine Learning

https://github.com/geekan/MetaGPT/tree/main/metagpt/ext/sela

可解释AI

XAI有什么用？探索LLM时代利用可解释性的10种策略

Usable XAI: 10 Strategies Towards Exploiting Explainability in the LLM Era

https://github.com/JacksonWuxs/UsableXAI_LLM

综述

可解释性终极追问，什么才是第一性解释？20篇CCF-A+ICLR论文给你答案

TransformerLens

Neel Nanda（deepmind）的项目

https://transformerlensorg.github.io/TransformerLens/

ecco

https://www.eccox.io/

https://jalammar.github.io/explaining-transformers/

https://jalammar.github.io/hidden-states/

interpretability in the wild

Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

https://github.com/redwoodresearch/Easy-Transformer

activation engineering

Activation Addition: Steering Language Models Without Optimization

representation engineering

Representation Engineering: A Top-Down Approach to AI Transparency

transformer-debugger

https://github.com/openai/transformer-debugger/tree/main

painter

八问八答搞懂Transformer内部运作原理

Transformer Layers as Painters

transformer explainer

黑匣子被打开了！能玩的Transformer可视化解释工具，本地运行GPT-2、还可实时推理

TRANSFORMER EXPLAINER: Interactive Learning of Text-Generative Models

http://poloclub.github.io/transformer-explainer/

https://bbycroft.net/llm

superposition

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

3Blue1Brown

用最直观的动画，讲解LLM如何存储事实，3Blue1Brown的这个视频又火了

https://www.youtube.com/watch?v=9-Jl0dxWQs8

Monitor

他们掰开神经元，终于让大模型9.8大于9.11了：神秘创业公司，开源AI「洗脑」工具

https://transluce.org/observability-interface

https://monitor.transluce.org/dashboard/chat

llm反思

ACL 2025｜自我怀疑还是自我纠正？清华团队揭示LLMs反思技术的暗面

Understanding the Dark Side of LLMs’ Intrinsic Self-Correction

反思失败的原因：

内部答案波动：在多轮问答任务上，「你确定吗？请思考后再回答」的提示语会让LLMs反复更改答案。说明反思技术会造成LLMs内部答案的波动，表现出「自我怀疑」的倾向，最终可能导致回答出错
prompt偏差：LLMs在反思失败时会过度关注提示语「你确定吗？想一想再回答。」，而忽略问题本身；当反思失败时，LLMs在76.1%的情况下会更关注反思指令，而当坚持正确答案时，LLMs对反思指令和问题本身的关注度非常相近，分别为50.8%和49.2%。
认知偏差：
- 过度思考：过度制定策略而不采取行动
- 认知过载：在长文本的反思中忽略关键信息
- 完美主义偏差：为了追求高效性而忽略环境限制

反思失败的缓解方法：

问题重复：在反思prompt的最后附上初始问题以引导LLMs维持对初始问题的关注。
少样本微调：不引入知识的少样本（4-10 个样本）微调可纠正反思失败的异常行为。

自省

LLM 比之前预想的更像人类，竟也能「三省吾身」

Looking Inward: Language Models Can Learn About Themselves by Introspection

具身智能

大模型走向物理世界，TeleAI 发布大模型驱动的具身智能综述，覆盖300篇文献

Embodied-AI with large models: research and challenges

ReKep

李飞飞团队提出ReKep，让机器人具备空间智能，还能整合GPT-4o

ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation

https://github.com/huangwl18/ReKep

GR-2

GR-2登场！ByteDance Research提出机器人大模型，具备世界建模和强大泛化能力

GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation

https://gr2-manipulation.github.io/

RDT-1B

清华开源全球最大双臂机器人扩散大模型RDT，懂调酒能遛狗，登顶HF具身热榜

RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation

HIL-SERL

强化学习训练一两个小时，100%自主完成任务：机器人ChatGPT时刻真来了？

Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning

Genesis

历时2年，华人团队力作，震撼开源生成式物理引擎Genesis，可模拟世界万物

https://github.com/Genesis-Embodied-AI/Genesis

language of motion

李飞飞团队统一动作与语言，新的多模态模型不仅超懂指令，还能读懂隐含情绪

The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion

视觉空间智能

李飞飞、谢赛宁等探索MLLM「视觉空间智能」，网友：2025有盼头了

Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces

具身多模态推理

统一框架下的具身多模态推理：自变量机器人让AI放下海德格尔的锤子

PEVA

伯克利&Meta面向具身智能的世界模型：让AI通过全身动作「看见」未来

Whole-Body Conditioned Egocentric Video Prediction

https://dannytran123.github.io/PEVA/

MTU3D

ICCV 2025满分论文：一个模型实现空间理解与主动探索大统一

Move to Understand a 3D Scene: Bridging Visual Grounding and Exploration for Efficient and Versatile Embodied Navigation

https://github.com/MTU3D/MTU3D

CoA

模仿学习新范式，Chain-of-Action：轨迹自回归实现动作推理

Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation

https://github.com/ByteDance-Seed/Chain-of-Action

LLM+芯片设计

登上Nature的AI芯片设计屡遭质疑，谷歌发文反击，Jeff Dean：质疑者连预训练都没做

That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design

其他

安全性

Anthropic安全负责人：在超级AI「毁灭」人类之前，我们可以做这些准备

First-Person Fairness in Chatbots

翁荔B站分享原文：AI安全与“培养”之道

time-LLM

谁说大象不能起舞! 重编程大语言模型实现跨模态交互的时序预测 | ICLR 2024

Time-LLM: Time Series Forecasting by Reprogramming Large Language Models

https://github.com/KimMeen/Time-LLM

将时序预测任务转换成一个可以由 LLMs 有效解决的语言任务，成功激活了llm做高精度时序推理的能力。

时序输入重编程
提示做前缀

长尾

A Systematic Review on Long-Tailed Learning

文本匹配效果还行的模型

大多是基于sentence-bert的，m3e-base在电商语料上试过，效果不错

https://huggingface.co/moka-ai/m3e-base

https://huggingface.co/shibing624/text2vec-base-chinese

本地知识库

https://github.com/chatchat-space/Langchain-Chatchat

llm应用合辑

ChatGPT聚合站：https://hokex.com
游戏生成站：https://latitude.io/
家庭作业辅助站：https://ontimeai.com/
文字转语音站：https://www.resemble.ai/
艺术作画站：https://starryai.com/
logo制作站：https://www.logoai.com/
ai写作站：https://www.getconch.ai/
音乐制作站：https://soundraw.io/
声音模拟站：https://fakeyou.com/
一句话生成一段视频：https://runwayml.com/
文字转语音：https://murf.ai/

swiftsage

大语言模型在开放世界中的推理能力探索实践

SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks

达摩院大模型技术交流

https://developer.aliyun.com/live/248332

ppt：链接密码：5yyf

回译

通过单语数据提升 NMT 模型最高效的方法之一是回译（back-translation）。如果我们的目标是训练一个英语到德语的翻译模型，那么可以首先训练一个从德语到英语的翻译模型，并利用该模型翻译所有的单语德语数据。然后基于原始的英语到德语数据，再加上新生成的数据，我们就能训练一个英语到德语的最终模型。

Understanding Back-Translation at Scale

nan问题

解决pytorch半精度amp训练nan问题

时间序列

LLM用于时序预测真的不行，连推理能力都没用到

Are Language Models Actually Useful for Time Series Forecasting?

人脑

MIT大牛新作震惊学界！AI「长脑子」了？LLM惊现「人类脑叶」结构并有数学代码分区

The Geometry of Concepts: Sparse Autoencoder Feature Structure

LLM for 算法设计

调研180多篇论文，这篇综述终于把大模型做算法设计理清了

A Systematic Survey on Large Language Models for Algorithm Design

深度生成模型课程

教授何恺明在MIT的第二门课——《深度生成模型》，讲座PPT陆续已出

https://mit-6s978.github.io/schedule.html

https://mit-6s978.github.io/assets/pdfs/lec1_intro.pdf

https://mit-6s978.github.io/assets/pdfs/lec2_vae.pdf

https://mit-6s978.github.io/assets/pdfs/lec3_ar.pdf

https://mit-6s978.github.io/assets/pdfs/lec4_gan.pdf

https://mit-6s978.github.io/assets/pdfs/lec5_diffusion.pdf

时序db

influxdb

https://github.com/influxdata/influxdb

https://jasper-zhang1.gitbooks.io/influxdb/content/Introduction/getting_start.html

其他

原来，这些顶级大模型都是蒸馏的

Distillation Quantification for Large Language Models

小模型性能饱和、表现不佳，根源是因为Softmax?

Why do small language models underperform? Studying LM Saturation via the Softmax Bottleneck

Ilya Sutskever的推荐清单

传说中Ilya Sutskever精选论文清单：AI领域40大论文完整版「破解」完成

2024年大模型LLM还有哪些可研究的方向？

Hinton万字访谈：用更大模型「预测下一个词」值得全力以赴

ChatGPT如何「思考」？心理学和神经科学破解AI大模型，Nature发文

适应多形态多任务，最强开源机器人学习系统「八爪鱼」诞生

Octo: An Open-Source Generalist Robot Policy

LeCun新作：神经网络在实践中的灵活性到底有多大？

Just How Flexible are Neural Networks in Practice?

清华包揽最佳论文+时间检验奖，山大获荣誉提名，SIGIR 2024奖项出炉

https://zhuanlan.zhihu.com/p/654910335

一些记录

打印模型参数量

https://stackoverflow.com/questions/49201236/check-the-total-number-of-parameters-in-a-pytorch-model


pytorch_total_params = sum(p.numel() for p in model.parameters())

pytorch_total_trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

# Load the model
from transformers import BartForConditionalGeneration
from transformers import T5ForConditionalGeneration
def cal(model):
  pytorch_total_trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
  return pytorch_total_trainable_params

model = BartForConditionalGeneration.from_pretrained("facebook/bart-base")
print("bart-base")
print(cal(model)) # 6L 139420416 139M

model = BartForConditionalGeneration.from_pretrained("facebook/bart-base")
print("bart-base")
print(cal(model)) # 12L 406291456 406M

model = T5ForConditionalGeneration.from_pretrained("t5-small")
print("t5-small")
print(cal(model)) # 6L 60506624 65M

model = T5ForConditionalGeneration.from_pretrained("t5-base")
print("t5-base")
print(cal(model)) # 12L 222903552 223M


model = T5ForConditionalGeneration.from_pretrained("t5-large")
print("t5-large")
print(cal(model)) # 24L 737668096 738M

往现有tokenizer里加一些特殊token

https://stackoverflow.com/questions/69191305/how-to-add-new-special-token-to-the-tokenizer

num_added_toks = tokenizer.add_tokens(['[EOT]'], special_tokens=True) ##This line is updated
model.resize_token_embeddings(len(tokenizer))

###The tokenizer has to be saved if it has to be reused
tokenizer.save_pretrained(<output_dir>)

示例

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

print("Before")
print(tokenizer.all_special_tokens) # --> ['[UNK]', '[SEP]', '[PAD]', '[CLS]', '[MASK]']
print(tokenizer.all_special_ids)    # --> [100, 102, 0, 101, 103]


special_tokens_dict = {'additional_special_tokens': ['[EOT]']}
num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)
# model.resize_token_embeddings(len(tokenizer))  # --> Embedding(30523, 768)

tok_id = tokenizer.convert_tokens_to_ids('[EOT]')  # --> 30522

print("After")
print(tokenizer.all_special_tokens) # --> ['[UNK]', '[SEP]', '[PAD]', '[CLS]', '[MASK]']
print(tokenizer.all_special_ids)    # --> [100, 102, 0, 101, 103]

python的读写锁

一个写，多个并行读：https://pypi.org/project/readerwriterlock/

pytorch的显存泄露

https://github.com/pytorch/pytorch/issues/13246#issuecomment-445770039

torch profiling

https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html

可以拿这个来可视化：https://ui.perfetto.dev/

点击open trace file上传json文件
timeline中有两个python进程，点击cuda kernel会出现箭头，方便找到是哪个op调用了该kernel
- 靠上的python进程是host侧进程（主要是用户代码中调用的一些API/pytorch op，能比较方便能和训练代码对应上）
- 靠下的python进程是device（gpu）侧进程（记录实际cuda kernel 的执行和一些性能相关的数据）

device timeline比较稀疏的情况下训练性能较差，GPU利用率较低，可能需要排查下训练代码是否有问题

显存泄露排查

https://pytorch.ac.cn/docs/stable/torch_cuda_memory.html

https://pytorch.org/blog/understanding-gpu-memory-1/

https://pytorch.org/blog/understanding-gpu-memory-2/

检查显存

# (c) Meta Platforms, Inc. and affiliates. 
# https://pytorch.org/blog/understanding-gpu-memory-1/
import logging
import socket
from datetime import datetime, timedelta

import torch

from torchvision import models

logging.basicConfig(
   format="%(levelname)s:%(asctime)s %(message)s",
   level=logging.INFO,
   datefmt="%Y-%m-%d %H:%M:%S",
)
logger: logging.Logger = logging.getLogger(__name__)
logger.setLevel(level=logging.INFO)

TIME_FORMAT_STR: str = "%b_%d_%H_%M_%S"

# Keep a max of 100,000 alloc/free events in the recorded history
# leading up to the snapshot.
MAX_NUM_OF_MEM_EVENTS_PER_SNAPSHOT: int = 100000

def start_record_memory_history() -> None:
   if not torch.cuda.is_available():
       logger.info("CUDA unavailable. Not recording memory history")
       return

   logger.info("Starting snapshot record_memory_history")
   torch.cuda.memory._record_memory_history(
       max_entries=MAX_NUM_OF_MEM_EVENTS_PER_SNAPSHOT
   )

def stop_record_memory_history() -> None:
   if not torch.cuda.is_available():
       logger.info("CUDA unavailable. Not recording memory history")
       return

   logger.info("Stopping snapshot record_memory_history")
   torch.cuda.memory._record_memory_history(enabled=None)

def export_memory_snapshot() -> None:
   if not torch.cuda.is_available():
       logger.info("CUDA unavailable. Not exporting memory snapshot")
       return

   # Prefix for file names.
   host_name = socket.gethostname()
   timestamp = datetime.now().strftime(TIME_FORMAT_STR)
   file_prefix = f"{host_name}_{timestamp}"

   try:
       logger.info(f"Saving snapshot to local file: {file_prefix}.pickle")
       torch.cuda.memory._dump_snapshot(f"{file_prefix}.pickle")
   except Exception as e:
       logger.error(f"Failed to capture memory snapshot {e}")
       return

# Simple Resnet50 example to demonstrate how to capture memory visuals.
def run_resnet50(num_iters=5, device="cuda:0"):
   model = models.resnet50().to(device=device)
   inputs = torch.randn(1, 3, 224, 224, device=device)
   labels = torch.rand_like(model(inputs))
   optimizer = torch.optim.SGD(model.parameters(), lr=1e-2, momentum=0.9)
   loss_fn = torch.nn.CrossEntropyLoss()

   # Start recording memory snapshot history
   start_record_memory_history()

   for _ in range(num_iters):
       pred = model(inputs)
       loss_fn(pred, labels).backward()
       optimizer.step()
       optimizer.zero_grad(set_to_none=True)

   # Create the memory snapshot file
   export_memory_snapshot()

   # Stop recording memory snapshot history
   stop_record_memory_history()

if __name__ == "__main__":
    # Run the resnet50 model
    run_resnet50()

同时profile cpu和显存

# (c) Meta Platforms, Inc. and affiliates. 
# https://pytorch.org/blog/understanding-gpu-memory-1/
import logging
import socket
from datetime import datetime, timedelta

import torch

from torch.autograd.profiler import record_function
from torchvision import models

logging.basicConfig(
   format="%(levelname)s:%(asctime)s %(message)s",
   level=logging.INFO,
   datefmt="%Y-%m-%d %H:%M:%S",
)
logger: logging.Logger = logging.getLogger(__name__)
logger.setLevel(level=logging.INFO)

TIME_FORMAT_STR: str = "%b_%d_%H_%M_%S"

def trace_handler(prof: torch.profiler.profile):
   # Prefix for file names.
   host_name = socket.gethostname()
   timestamp = datetime.now().strftime(TIME_FORMAT_STR)
   file_prefix = f"{host_name}_{timestamp}"

   # Construct the trace file.
   prof.export_chrome_trace(f"{file_prefix}.json.gz")

   # Construct the memory timeline file.
   prof.export_memory_timeline(f"{file_prefix}.html", device="cuda:0")

def run_resnet50(num_iters=5, device="cuda:0"):
   model = models.resnet50().to(device=device)
   inputs = torch.randn(1, 3, 224, 224, device=device)
   labels = torch.rand_like(model(inputs))
   optimizer = torch.optim.SGD(model.parameters(), lr=1e-2, momentum=0.9)
   loss_fn = torch.nn.CrossEntropyLoss()

   with torch.profiler.profile(
       activities=[
           torch.profiler.ProfilerActivity.CPU,
           torch.profiler.ProfilerActivity.CUDA,
       ],
       schedule=torch.profiler.schedule(wait=0, warmup=0, active=6, repeat=1),
       record_shapes=True,
       profile_memory=True,
       with_stack=True,
       on_trace_ready=trace_handler,
   ) as prof:
       for _ in range(num_iters):
           prof.step()
           with record_function("## forward ##"):
               pred = model(inputs)

           with record_function("## backward ##"):
               loss_fn(pred, labels).backward()

           with record_function("## optimizer ##"):
               optimizer.step()
               optimizer.zero_grad(set_to_none=True)

if __name__ == "__main__":
    # Warm up
    run_resnet50()
    # Run the resnet50 model
    run_resnet50()

各型号gpu对比

https://zhuanlan.zhihu.com/p/441153412

查看python的栈

pip install py-spy
py-spy dump --pid 1199

打出来：

Process 1199: /usr/bin/python3.10 -u torch_main.py
Python v3.10.14 (/usr/bin/python3.10)

Thread 0x7F62A2C43740 (active): "MainThread"
    _wait_for_tstate_lock (threading.py:1116)
    join (threading.py:1096)
    main (torch_main.py:776)
    <module> (torch_main.py:785)
Thread 0xAABBBCC (idle): "Thread-1"
    wait (threading.py:324)
    wait (threading.py:607)
    run (threading.py:1376)
    _bootstrap_inner (threading.py:1016)
    _bootstrap (threading.py:973)
Thread 0xAAAAA (idle): "Thread-3 (process)"
    wait (threading.py:320)
    get (queue.py:171)
    process (abase_writer.py:73)
    run (threading.py:953)
    _bootstrap_inner (threading.py:1016)
    _bootstrap (threading.py:973)
Thread 0xA992ACDA (idle): "Thread-4 (process)"
    wait (threading.py:320)
    get (queue.py:171)
    process (abase_writer.py:73)
    run (threading.py:953)
    _bootstrap_inner (threading.py:1016)
    _bootstrap (threading.py:973)
Thread 0xAFF11AA (active): "Thread-5 (read_file)"
    get_seq (ecom_seq_reader.py:200)
    read_file (torch_main.py:494)
    run (threading.py:953)
    _bootstrap_inner (threading.py:1016)
    _bootstrap (threading.py:973)
Thread 0x9922BCDA (idle): "Thread-6"
    wait (threading.py:324)
    wait (threading.py:607)
    run (tqdm/_monitor.py:60)
    _bootstrap_inner (threading.py:1016)
    _bootstrap (threading.py:973)

国内的huggingface模型下载地址

https://hf-mirror.com/

一些报错的解法

flash-attention2

https://github.com/Dao-AILab/flash-attention/issues/451

FLASH_ATTENTION_FORCE_BUILD=TRUE pip install flash-attn

GPU机型对比

L20：https://www.techpowerup.com/gpu-specs/l20.c4206
- T flops：
  - tf32：59.8
  - fp32：59.8
  - bf16：119.5
  - fp16：119.5
A800-40G：https://www.techpowerup.com/gpu-specs/a800-pcie-40-gb.c3964
- T flops：
  - tf32：156
  - fp32：19.5
  - bf16：312
  - fp16：77.97

一些问题和经验

坍缩（稳定召回那k个item），attention score太集中了，低秩特征（泛化特征）容易导致这个问题

看attention score的分布，如果第一层偏向target item，但第二层可能就很平均了，这种可能就释放不出收益，应该是没学好

auc离线有收益，在线没收益：

reload正确性，warmup有没有报错
nn的分发效率，20min降到x min，压缩、解压耗时
学习充分性：发现ab开得更久的时候，就看到收益了。。
- 累积梯度太小的不充分
- 历史样本变多
- 多epoch(参考快手阿里的一些做法，例如reset emb等)
- 加一些辅助loss，例如生成式、蒸馏
出nan：bf16转化等问题，加一些grad clip，norm等

上一页1.8.llm_o1 下一页2.recommend

最后更新于18天前

这有帮助吗？

泄露的系统提示词

语言物理学

蒸馏

ABKD

LLM+math

mathscale

AlphaGeometry

AlphaProof & AlphaGeometry 2

WE-Math基准

case-based or rule-based

LLM常见难题

LLM as a judge

重复生成

幻觉

综述

语义熵

Zilliz

记忆能力

reasoning

memorizing

越狱

LLM compiler

ProLong

白化

蒸馏

证明者-验证者博弈

道德风险

选择性偏差