---
title: "Discussion on \"Backprop 01: Where Did the Memory Go?\" | Hashnode"
description: "Discussion on \"Backprop 01: Where Did the Memory Go?\". Why the same Gemma 4 12B fit 1.83 users on vLLM and about 13 on llama.cpp"
url: "https://hashnode.com/posts/backprop-01-where-did-the-memory-go/6afd92054f515efa31a3329a"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Backprop 01: Where Did the Memory Go?](https://gradientdecent.hashnode.dev/backprop-01-where-did-the-memory-go?utm_source=hashnode&utm_medium=feed)

Why the same Gemma 4 12B fit 1.83 users on vLLM and about 13 on llama.cpp

16h ago · 24 min read · gradientdecent.hashnode.dev  
Read the article

-   [#llm](https://hashnode.com/tag/llm)
-   [#kv-cache](https://hashnode.com/tag/kv-cache)
-   [#vllm](https://hashnode.com/tag/vllm)
-   [#llamacpp](https://hashnode.com/tag/llamacpp)
-   [#amd](https://hashnode.com/tag/amd)

## Written by

[Gayatri Panse](https://hashnode.com/@gradientnotes)

@gradientnotes · Pune, India

  
[View profile](https://hashnode.com/@gradientnotes)

Building & Prototyping AI systems @Nuvolink

Hi, I’m Gayatri, co-founder of Nuvolink and a builder at heart. I’m part of a small founding team building AI products, working on agentic systems, ML and inference for real-world applications in legal tech and defence. Gradient Decent is where I learn in public: hands-on experiments with real numbers, literature surveys and deep dives, and short notes on what I’ve built or read. Topics include agentic systems, evals, search and retrieval, LLMs, and ML/DL. Every post is a small step forward, like a gradient update: not always optimal, but always slightly less wrong.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.
