Life is like riding a bicycle. To keep your balance, you must keep moving.
Reading Time: 21 minutes TL;DR: A vibe coded Next.js 15 app (https://visualneurons.com/) that lets you create images, edit them with precision masks, and generate short videos. It runs on a small t3.xlarge EC2 box behind nginx with TLS, stores files on disk, tracks usage and costs, and authenticates with AWS Cognito. Models: Imagen 4 Ultra, Nova Canvas, Nano Banana, Read More Building Visual…
Reading Time: 15 minutes 👉 Useful links: Youtube lecture, minbpe GitHub repo, Colab notebook What is a tokenizer and why do we need one The purpose of this post is to put in writing (part of) the latest Andrej Karpathy s lecture on LLM Tokenization. I ll specifically try to cover the Byte Pair Encoding (BPE) algorithm, which is at the Read More Byte Pair Encoding: building the GPT tokenizer with…
Reading Time: 18 minutes Introduction In this post, I go through the content covered by Jeremy Howard in his lecture CUDA for Python programmers. This is a talk he gave as part of the Programming Massively Parallel Processors (PMPP) study group, currently happening on the CUDA MODE Discord server (with videos uploaded to the homonymous YouTube channel). My initial Read More CUDA for Python…
Reading Time: 9 minutes Context In this post, Lucas and I describe the end2end cloud solution we built to deploy LogoNet. That s the name we gave to a web application accepting an image as user input, and returning 30 variations of it, based on curated prompts to a ControlNet model. Below you can find the high-level workflow, starting with Read More LogoNet: the journey to an AWS-powered cloud…
Reading Time: 5 minutes Useful links The what and the why If you haven t yet watched Karpathy s lecture Let s build GPT: from scratch, in code, spelled out. you are missing out on an absolute masterclass. From scratch, in two hours, Andrej builds and trains a character-level Generatively Pretrained Transformer (GPT) on the entire Shakespeare s corpus, following the paper Attention Read More A…
Reading Time: 9 minutes Context In October 2022 I started attending the new version of part 2 of the fastai course. In this one, Jeremy builds up a Deep Learning training and evaluation framework from scratch. Literally starting out from matrix multiplication and climbing all the way up to torch. It s a somewhat simplified version of the fastai library Read More Fastai Course Part 2 2022:…
Reading Time: 6 minutes Note: you can follow along with the post in this Jupyter notebook Context The ML announcement I liked the most at AWS re:Invent 2022 was SageMaker support for shadow deployment. What is it? The idea is the following. You have model A deployed on an endpoint. You have developed model B (say, a supposedly more Read More Amazon SageMaker Shadow Deployment: Semantic…
Reading Time: 9 minutes Disclaimers TLDR All scenarios running on NVIDIA GPU: Context This post is a continuation of the experiments I conducted and summarized here around deploying and benchmarking ResNet18 to SageMaker endpoints on GPU and AWS Inferentia. One of the comments I received on that work was: It seems the combo SageMaker Neo + Inferentia wins hands Read More Benchmarking TorchVision…
Reading Time: 8 minutes Code and disclaimer: you can find the notebook with the relevant code here. Most of it was copied from the official Amazon SageMaker repo, without much change. TLDR 🤷♂️ The what and the why💡 The goal of this short post is to benchmark, both cost and latency-wise, a couple of ML inference options we have Read More Benchmarking TorchVision ResNet18 on Amazon SageMaker CPU,…
Reading Time: 9 minutes Note: You can find the Jupyter notebook with all the steps I followed here, and the folder with the relevant accompanying files here. Introduction The goal of this post is to experiment with the Neural Magic (NM) suite of open-source libraries by training a face detector running at GPU speed on CPU (yes, you heard Read More Neural Magic: Training YoloV5 with Sparse Transfer…