80% faster, 50% less memory, 0% loss of accuracy Llama finetuning
Hi HN! I'm just sharing a project I've been working on during the LLM Efficiency Challenge - you can now finetune Llama with QLoRA 5x faster than Huggingface's original implementation on your own local GPU. Some highlights: 1. Manual autograd engine - hand derived backprop steps. 2. QLoRA / LoRA 80% faster, 50% less memory. 3. All kernels written in OpenAI's Triton language. 4. 0% loss in accuracy - no approximation methods - all exact. 5. No change of hardware necessary. Supports NVIDIA GPUs since 2018+. CUDA 7.5+. 6. Flash Attention support via Xformers. 7. Supports 4bit and 16bit…
80% faster, 50% less memory, 0% loss of accuracy Llama finetuning launched on December 1, 2023 with 385 votes, #65 of 3,780 launches that month and more than 99% of that year's launches. The site was still online in September 2026, 2 years on.
In plain words
This project provides an optimized implementation for finetuning Llama models using QLoRA and LoRA techniques on local GPUs. It uses a manually derived autograd engine with kernels written in Triton to achieve 80% faster training and 50% less memory usage compared to standard implementations, while maintaining full accuracy. The tool supports NVIDIA GPUs from 2018 onwards and offers both open-source and commercial versions, enabling researchers and developers to finetune large language models efficiently on consumer hardware without hardware upgrades.
written from the facts on this page · September 2026
From the sources
Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, FLUX and more. - unslothai/unsloth
In the maker’s words, at launch
Hi HN! I'm just sharing a project I've been working on during the LLM Efficiency Challenge - you can now finetune Llama with QLoRA 5x faster than Huggingface's original implementation on your own local GPU. Some highlights: 1. Manual autograd engine - hand derived backprop steps. 2. QLoRA / LoRA 80% faster, 50%…
Does a similar job
all alternatives →
Finetune LLaMA-7B on commodity GPUs using your own text2023 · github.com · ▲449I've been playing around with https://github.com/zphang/minimal-llama/ and https://github.com/tloen/alpaca-lora/blob/main/finetune.py, and wanted to create a simple UI where you can just paste text, tweak the parameters, and finetune the model quickly using a modern GPU. To prepare the data, simply separate your text with two blank lines. There's an inference tab, so you can test how the tuned model behaves. This is my first foray into the world of LLM finetuning, Python, Torch, Transformers, LoRA, PEFT, and Gradio. Enjoy!
Tune LLaMa3.1 on Google Cloud TPUs2024 · github.com · ▲189Hey HN, we wanted to share our repo where we fine-tuned Llama 3.1 on Google TPUs. We’re building AI infra to fine-tune and serve LLMs on non-NVIDIA GPUs (TPUs, Trainium, AMD GPUs). The problem: Right now, 90% of LLM workloads run on NVIDIA GPUs, but there are equally powerful and more cost-effective alternatives out there. For example, training and serving Llama 3.1 on Google TPUs is about 30% cheaper than NVIDIA GPUs. But developer tooling for non-NVIDIA chipsets is lacking. We felt this pain ourselves. We initially tried using PyTorch XLA to train Llama 3.1 on TPUs, but it was rough: xla…
Speeding up LLM inference 2x times (possibly)2024 · asciinema.org · ▲419Here's a project I've been working on for the last few months. It's a new (I think) algorithm, that allows to adjust smoothly - and in real time - how many calculations you'd like to do during inference of an LLM model. It seems that it's possible to do just 20-25% of weight multiplications instead of all of them, and still get good inference results. I implemented it to run on M1/M2/M3 GPU. The mmul approximation itself can be pushed to run 2x fast before the quality of output collapses. The inference speed is just a bit faster than Llama.cpp's, because the rest of implementation…
Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPUFeb 2026 · github.com · ▲395Hi everyone, I'm kinda involved in some retrogaming and with some experiments I ran into the following question: "It would be possible to run transformer models bypassing the cpu/ram, connecting the gpu to the nvme?" This is the result of that question itself and some weekend vibecoding (it has the linked library repository in the readme as well), it seems to work, even on consumer gpus, it should work better on professional ones tho
More ai this month
the category →
Paintings by AI models: programs run through a simulation of oil paint.
AI · 22d ago · stillwet.art
FastRouter.ai▲364Route requests to the right LLM for cost, latency & quality
AI · 23d ago · fastrouter.ai
Source: Show HN post ↗ · launch text quoted in part
Launched alongside, December 2023
the whole month →
Hi HN! I’m proud to share that we have built a real 3rd party iMessage client for Android. We did it by reverse engineering the iMessage protocol and encryption system. It's available to download today (no waitlist): https://play.google.com/store/apps/details?id=com.beeper.ima and there's a technical writeup here: https://blog.beeper.com/p/how-beeper-mini-works. Unlike every other attempt to build an iMessage app for Android (including our first gen app), Beeper Mini does not use a Mac server relay in the cloud. The app connects directly to Apple…
Work · 2023 · beeper.com
Hey! I made Heynote entirely for my own use case. For many years, I always had an Emacs instance running with the scratch buffer open, even long after I had abandoned Emacs as my programming editor in favor of more recent IDE:s. The simplicity of having just one big scratch buffer appeals to me, but I still want to separate the different things I jot down somehow (without using tabs or similar). Previously, my solution was to insert a bunch of blank lines between the notes, but hitting C-A would still select the entire buffer. That's why I came up with the concept of "blocks", which turned…
Work · 2023 · heynote.com
FuseBase▲939Client collaboration platform for professional services
Dev tools · 2023 · thefusebase.com
Beep! 2.0▲871Fastest tool to assign and share visual tasks with your team
Work · 2023 · justbeepit.com
Inbox Zero▲821Clean Up Your Inbox In Minutes Newsletter management, AI automation, and email analytics. Inbox Zero is the open-source email app that puts you back in control of your inbox.
AI · 2023 · getinboxzero.com