[Paper Reading Club] Stop Wasting GPU VRAM — PagedAttention & vLLM Architecture

Name: [Paper Reading Club] Stop Wasting GPU VRAM — PagedAttention & vLLM Architecture
Start: 2026-08-11T22:00:00.000+05:30
End: 2026-08-11T23:00:00.000+05:30
Location: Online

Machine Learning Indore

[Paper Reading Club] Stop Wasting GPU VRAM — PagedAttention & vLLM Architecture

Details

Welcome to Session #01 of the ML Indore Paper Reading Club!

In this online session, we are opening up and reading through: 📄 Efficient Memory Management for Large Language Model Serving with PagedAttention Authors: Woosuk Kwon et al. (UC Berkeley SkyLab) — Read on arXiv

📖 How We Read Papers Together

You don't need to be a C++ kernel developer or math wizard to join. We will have the PDF on screen, walking through the core architectural diagrams, section by section:

The Context & Bottleneck (Sections 1–2): Why static KV cache pre-allocation wastes 60–80% of GPU VRAM.
The Core Intuition (Section 3): How the paper adapts OS Virtual Memory Paging to GPU tensors.
Architecture & Implementation (Section 4): Looking at logical-to-physical block tables and dynamic allocation.
Group Discussion & Critique: Where does this paper break down? What are the hardware trade-offs?

👤 Session Facilitator

Karan Mittal (CTO & Co-Founder at Dextar Intelligence) will lead the walkthrough, breaking down key diagrams and facilitating group discussion.

🌐 Portfolio & Background: karan-s-mittal.github.io/about

💡 How to Prepare

📄 Skim the PDF ahead of time: Don't worry about understanding every equation—just glance at Figure 1 and Figure 2.
❓ Bring 1 question or takeaway: Note down one thing that made you curious or confused.
💻 Have the PDF open: We encourage everyone to read along during the session.