Tag

#LLM Training

101 posts tagged with this label. Back to all tags or the main feed.

2026

09-29 EN

ProRL: What Prolonged Reinforcement Learning Changes, and What Its Evidence Can Prove
09-29 中

ProRL 阅读笔记：长期强化学习与推理边界
09-24 EN

Deep Delta Learning: Editing the Residual Stream, and Accounting for the Cost
09-24 中

Deep Delta Learning 阅读笔记：怎样改写残差流，以及这次改写要付出什么代价
09-23 EN

Gated Attention: Controlling What a Softmax Head Writes
09-23 中

Gated Attention：让注意力头决定写入多少信息
09-22 EN

Progressive Point Matching: Dense Credit for Long-Horizon Language Model RL
09-22 中

PPM 阅读笔记：长程强化学习，怎样给尚未成功的推理记分
09-20 EN

mHC: Stable Residual Mixing, Its Guarantees, and Its Costs
09-20 中

mHC 阅读笔记：残差混合如何稳定，保证又止于哪里
09-19 EN

Attention Residuals: Learning What to Retrieve Across Network Depth
09-19 中

Attention Residuals 阅读笔记：让每一层选择自己需要的历史表示
09-15 EN

SAPO: Smooth Policy Updates, Their Mathematics, and Their Limits
09-15 中

SAPO 阅读笔记：平滑策略更新的推导、实现与边界
09-01 EN

RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction
09-01 中

RRC 阅读笔记：如何真正把生成式奖励模型的判卷能力用到 RL 里
08-27 EN

SCOUT: Localizing Latent Training Failures via Strict-Majority Consensus Among Equivalent Replicas
08-27 中

SCOUT 阅读笔记：用「对等副本多数一致」定位大模型预训练中的隐性故障
08-25 EN

Best-Practice Critic Optimization: Making a Learned Critic Beat Group Sampling in LLM RL
08-25 中

Best-Practice Critic Optimization：让训练出来的 Critic 打赢分组采样
08-21 EN

CARE-LoRA: Fixing LoRA-FA's Frozen-Subspace Problem Without Paying for Full Activations
08-21 中

CARE-LoRA：不用冻结子空间，也能省下 LoRA 的激活内存
08-18 EN

GCPO: Why Some GRPO Updates Quietly Break Your Model, and How to Structurally Forbid Them
08-18 中

GCPO 阅读笔记：为什么某些 GRPO 更新会悄悄搞垮模型，以及如何从结构上禁止它们
08-17 中

0 train-infer mismatch
08-13 EN

ZeroLock: Breaking Update Locking in Pipeline-Parallel LLM Fine-Tuning
08-13 中

ZeroLock 阅读笔记：用模块化解耦打破流水线并行训练的更新锁定
08-11 EN

CVPO: Curriculum-Guided Value-Variance Policy Optimization for LLM Reasoning
08-11 中

CVPO 阅读笔记：用价值方差驱动的课程式策略优化
08-02 EN

FEPLB: Exploiting Copy Engines for Nearly Free MoE Load Balancing in Distributed Training
08-02 中

FEPLB 阅读笔记：用 Copy Engine 几乎零成本地做 MoE 负载均衡
07-30 EN

Libra: Taming Attention Workload Skew in Long-Context LLM Training with Bounded Sequence Pool
07-30 中

Libra 阅读笔记:用有界序列池驯服长上下文训练中的注意力负载偏斜
07-23 EN

PHOENIX: Recovering LLM Training in 40 Seconds Instead of Restarting the Whole Job
07-23 中

PHOENIX：让大模型训练在故障后 40 秒内复活，而不是整个作业重启
07-21 EN

PS-PPO: Skipping the Boring Parts of a Trajectory Without Breaking PPO's Math
07-21 中

PS-PPO 阅读笔记:跳过轨迹里「已经确定」的部分,但不破坏 PPO 的数学
07-16 EN

GIFT: Why the Coordinate System You Quantize In Matters More Than the Quantizer
07-16 中

GIFT 阅读笔记:量化用的坐标系,比量化器本身更重要
07-14 EN

RIPO: Why PPO's Ratio Clipping Is the Wrong Metric, and What Riemannian Geometry Fixes
07-14 中

RIPO 阅读笔记：PPO 的比率裁剪为什么量错了尺度,黎曼几何给出的修正
07-11 EN

DPPO: Fixing PPO's Ratio Clipping by Constraining Divergence Instead of a Single Token's Probability
07-11 中

DPPO 阅读笔记：与其信任一个采样 token 的比率，不如直接约束策略散度
07-09 EN

RATrain: Training-State Lifecycle Scheduling for Dense LLM Training on Bandwidth-Constrained Heterogeneous Supercomputers
07-09 中

RATrain：面向带宽受限异构超算的稠密大模型训练状态生命周期调度
07-07 EN

RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents
07-07 中

RSPO：多轮LLM智能体的奖励交换策略优化
07-06 EN

The Mirage of Optimizing Training Policies: Monotonic Inference Policy Improvement for LLM RL
07-06 中

训练策略的幻觉：为什么LLM强化学习的真正目标是推理策略单调改进
07-02 EN

Tangram: Hiding GPU Heterogeneity for Efficient LLM Parallelization
07-02 中

Tangram：为异构GPU集群隐藏硬件差异的高效LLM并行化系统
06-30 EN

DAPO: An Open-Source LLM Reinforcement Learning System at Scale — Technical Review
06-30 中

DAPO：大规模 LLM 强化学习系统阅读笔记
06-16 EN

Back to Basics: Revisiting REINFORCE Style Optimization for RLHF (RLOO)
06-16 中

回归基础：用 RLOO 重新思考 RLHF 中的策略梯度优化
06-13 EN

ForeMoE: Micro-step-level MoE Load Balancing for RL Post-training via Routing Foresight
06-13 中

ForeMoE：利用路由预见性实现 RL 后训练中 MoE 微步级负载均衡
06-11 EN

MegaScale: Engineering 55% MFU at 12,288 GPUs for LLM Training
06-11 中

MegaScale：ByteDance 如何在 12,288 块 GPU 上实现 55% MFU 的大规模 LLM 训练
06-09 EN

VAPO: Value-Augmented Proximal Policy Optimization for Long-CoT Reasoning
06-09 中

VAPO：面向长链推理的价值增强近端策略优化
06-06 EN

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
06-06 中

DeepSeek-R1：用强化学习激发大语言模型的推理能力
06-02 EN

REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization
06-02 中

REINFORCE++：用全局优势归一化稳定免批评家策略优化
05-31 EN

Group Sequence Policy Optimization: A Sequence-Level RL Algorithm for Training Large Language Models
05-31 中

Group Sequence Policy Optimization：序列级重要性采样修正 GRPO 的 RL 训练方法
05-26 EN

SimPO: Simple Preference Optimization with a Reference-Free Reward
05-26 中

SimPO：无需参考模型的简洁偏好优化
05-25 EN

CodeAct: Executable Code Actions Elicit Better LLM Agents
05-25 中

CodeAct：用可执行代码驱动更强的 LLM Agent
05-24 EN

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
05-24 中

FlashAttention-2：更好的并行策略与线程块工作划分
05-22 EN

DoRA: Weight-Decomposed Low-Rank Adaptation — Technical Review
05-22 中

DoRA：权重分解低秩自适应——用幅度与方向解耦提升 LoRA 学习能力 | 阅读笔记
05-19 EN

KTO: Model Alignment as Prospect Theoretic Optimization — Technical Blog Review
05-19 中

KTO：把模型对齐看成「前景理论」优化 —— 阅读笔记
05-15 EN

Zero Sum SVD: A Global, Loss-Aware Rank Budget for LLM Compression
05-15 中

Zero Sum SVD：用「损失零和」做全局奇异值预算分配的 LLM 压缩方法
05-14 EN

DisagMoE: Disaggregating Attention and FFN to Beat the MoE All-to-All Bottleneck
05-14 中

DisagMoE：用解耦 Attention 和 FFN 打通 MoE 训练的 all-to-all 瓶颈
05-09 EN

Queueing Stability for LLM Inference with KV Cache Memory Constraints
05-07 EN

Piper: Efficient Large-Scale MoE Training via Resource Modeling and Pipelined Hybrid Parallelism
05-01 EN

Low-Rank Optimization Trajectories for LLM RLVR Acceleration: A Technical Review of NExt
04-24 EN

Generalization at the Edge of Stability: A Random Dynamical Systems Perspective
04-16 EN

PipeDream: Turning Pipeline Parallelism into a Practical Training System — Deep Technical Review
04-16 中

PipeDream：把 Pipeline Parallelism 做成真正可训练系统——深度阅读笔记
04-09 EN

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving — Deep Technical Review
04-09 中

DistServe：通过 Prefill/Decoding 解耦实现面向 Goodput 的大模型服务优化 — 深度阅读笔记
04-07 EN

ORPO: Monolithic Preference Optimization without Reference Model — In-Depth Technical Review
04-07 中

ORPO：不用参考模型的一体化偏好优化 — 深度阅读笔记
03-27 EN

GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection — In-Depth Technical Review
03-26 EN

Alpa: Automating Inter- and Intra-Operator Parallelism — In-Depth Technical Review
03-19 EN

ZeRO: Shattering the Memory Wall — How DeepSpeed Trains Trillion-Parameter Models
03-12 EN

Megatron-LM: NVIDIA's Blueprint for Training Billion-Parameter Models at Scale
03-12 EN

PaRO: Smarter Partitioning for Distributed Training — Beyond ZeRO's One-Size-Fits-All
02-19 EN

vLLM and PagedAttention: Efficient Memory Management for Large Language Model Serving — Technical Review
02-18 EN

GLM-5 Technical Review: From Vibe Coding to Agentic Engineering

2023

04-29 EN

ComputerArchitecture-Day1

2019

11-26 EN

Tensorflow-Day1-DNN Explain
11-24 EN

Reinforcement Learning\_WatermelonBook\_Summary