-
Open Source Model Quantization Strategies
-
MoE Parallel Folding
-
Engram: Scaling Large Language Models via Conditional Memory Lookup
-
Manifold-Constrained Hyper-Connections
-
High-Performance Matmul Kernels on NVIDIA Hopper
-
Rollout Routing Replay: Stabilizing MoE Reinforcement Learning
-
LLM Agent Memory
-
Self-Play SWE-RL: Superintelligent Agents via Autonomous Bug Discovery
-
LLMs as Improvement Operators & Parallel-Distill-Refine (PDR)
-
Reading Note on Performance Hints
-
Reading Note on SonicMoE
-
Reading Note on Nvidia Nemotron 3
-
Interplay of Training Stages
-
Linear Attention: Kimi Delta Attention
-
Pure BF16 Training via Stochastic Rounding
-
Adaptive NVFP4 Quantization
-
LongCat Flash
-
MXFP8 Training
-
Tokenizer Learning
-
LLM Architecture Evolution
-
First-Order Approximation for Stable LLM-RL Training
-
DeepSeek-V3.2 Reading Note
-
vLLM V1 Understanding
-
Smol Training Playbook Reading Note
-
Infra Math for LLM Training
-
Predictable Scaling of Reinforcement Learning for LLMs
-
Reasoning Limits of LLMs Under RLVR
-
LoRA, Manifolds, and OPD
-
NVFP4: Stable 4-Bit Training at 10 Trillion Tokens
-
Kimi-K2 Reading Note
-
Recent RL Infra Related Papers
-
High Precision Used for Reasoning Recipes
-
DeepSeek-V3's Hardware-Aware Design
-
Summary on Llama-Nemotron
-
Summary on StreamRL
-
DAPO Reading Note
-
Disaggregate Prefill and Decoding
-
Summary on DeepSeek R1 and Kimi k1.5
-
Summary on MiniMax-01
-
Summary on Zero Bubble
-
CUDA H100 GEMM Optimization
-
Summary of SemiAnalysis o1 Reasoning Report
-
People Retrospective: Communication, Growth, Collaboration, and Challenges
-
Summary on DeepSeek V3
-
Scaling Law
-
Understand Speculative Decoding for LLM Inference
-
Notes on Reading Hunyuan Model
-
Educational Materials for GEMM Optimizations on CPUs and GPUs