Stories by limoce

FP8 is ~100 tflops faster when the kernel name has "cutlass" in it

Polaris: A Post-training recipe for scaling RL on Advanced Reasoning models

Overclocking LLM Reasoning: Monitoring and Controlling LLM Thinking Path Lengths

Neutrino: Probing-Based eBPF-Like GPU Kernel Profiling

Machine Learning Conferences Should Establish "Refutations and Critiques" Track

SuperGPQA: Scaling LLM Evaluation Across 285 Graduate Disciplines

SepLLM: Accelerate LLMs by Compressing One Segment into One Separator

Step-Video-T2V: The Practice, Challenges, and Future of Video Foundation Model

Logic R1: Reproduce DeepSeek R1 Zero on 2K Logic Puzzle Dataset

Libnginx: Nginx as a Shared Library

DeepSeek-VL2: Moe Vision-Language Models for Advanced Multimodal Understanding [pdf]

Fast vectorizable algorithms of binary searching for floating point numbers

New OpenAI Feature: Predicted Outputs

Collaborative Filtering Is Wrong and Here Is Why

REST: A Plug-and-Play Method for Accelerating LLM Without Additional Training

Smoke 'em if you got 'em: Hacker gains root access using cigarette lighter

O1 Replication Journey: A Strategic Progress Report

Failures of Gradient-Based Deep Learning (2017) [pdf]

Qwen2-VL

Qwen2-Math

1
2
3
4
5
6