---
title: "Aman Karki (@amankarki) | Hashnode"
description: "C++/CUDA Engineer · LLM Inference & Systems. Read the latest articles by Aman Karki on Hashnode."
url: "https://hashnode.com/@amankarki"
---

@amankarki

# Aman Karki

@amankarki · India · Joined August 2026

C++/CUDA Engineer · LLM Inference & Systems

[Website](https://aman-portfolio-rho-six.vercel.app/)

### [My GPU fix was 13% faster on one card and 81% slower on another](https://amankarki.hashnode.dev/gpu-fix-faster-on-one-card-slower-on-another-waves?utm_source=hashnode&utm_medium=feed)

A vLLM int8 GEMM change, a reviewer with a smaller Ada GPU, and the one idea that explained it: waves

7 min read · [amankarki.hashnode.dev](https://amankarki.hashnode.dev/gpu-fix-faster-on-one-card-slower-on-another-waves?utm_source=hashnode&utm_medium=feed)

[Oct 26](https://hashnode.com/posts/gpu-fix-faster-on-one-card-slower-on-another-waves/6adf75af7a752cdd73090b42)

### [Blockwise FP8 on Ada GPUs, part 2: my kernel was correct and 4x too slow](https://amankarki.hashnode.dev/blockwise-fp8-ada-kernel-4x-too-slow?utm_source=hashnode&utm_medium=feed)

How I found register spills in the main loop, fixed them, and got a 1.42x end-to-end speedup on an L4

8 min read · [amankarki.hashnode.dev](https://amankarki.hashnode.dev/blockwise-fp8-ada-kernel-4x-too-slow?utm_source=hashnode&utm_medium=feed)

[Oct 17](https://hashnode.com/posts/blockwise-fp8-ada-kernel-4x-too-slow/6ad2ef38195837f31546215e)

### [Blockwise FP8 on Ada GPUs: why your L4 falls back to Marlin, measured](https://amankarki.hashnode.dev/blockwise-fp8-ada-gpus-marlin-fallback?utm_source=hashnode&utm_medium=feed)

What 128×128 block scaling is, why it needs its own kernel, and what the missing kernel costs on a real L4

8 min read · [amankarki.hashnode.dev](https://amankarki.hashnode.dev/blockwise-fp8-ada-gpus-marlin-fallback?utm_source=hashnode&utm_medium=feed)

[Oct 2](https://hashnode.com/posts/blockwise-fp8-ada-gpus-marlin-fallback/6abfad8204a9ca58d66a76ac)

### [Two CUDA PRs into llama.cpp, and what they taught me about a 100K-star codebase](https://amankarki.hashnode.dev/llama-cpp-two-cuda-prs?utm_source=hashnode&utm_medium=feed)

A 95-line pooling kernel, a 9-line fix for a silent CPU fallback, and how ggml decides where your ops run

11 min read · [amankarki.hashnode.dev](https://amankarki.hashnode.dev/llama-cpp-two-cuda-prs?utm_source=hashnode&utm_medium=feed)

[Sep 27](https://hashnode.com/posts/llama-cpp-two-cuda-prs/6ab9471561ca1c18dd7082e6)

### [Two From-Scratch Systems, and the Day They Talked](https://amankarki.hashnode.dev/two-from-scratch-systems-and-the-day-they-talked?utm_source=hashnode&utm_medium=feed)

I built a vector database and an LLM inference engine, each without leaning on someone else's runtime. This is the demo where they finally talked to each other — and the two real bugs that showed up doing it.

5 min read · [amankarki.hashnode.dev](https://amankarki.hashnode.dev/two-from-scratch-systems-and-the-day-they-talked?utm_source=hashnode&utm_medium=feed)

[Sep 5](https://hashnode.com/posts/two-from-scratch-systems-and-the-day-they-talked/6a9c5c43ad3755e014f0cf68)