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…
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
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% 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 LoRA finetuning. 8. Train Slim Orca fully locally in 260 hours from 1301 hours (5x faster). 9. Open source version trains 5x faster or you can check out Unsloth Pro and Max codepaths for 30x faster training! https://www.reddit.com/r/LocalLLaMA/comments/188197j/80_fast... has more info about Unsloth! Hopefully you can try it out! Wrote a blog post at https://unsloth.ai/introducing if you want to learn more about our manual hand derived backprop or Triton kernels and stuff! Thanks once again!
Does the same job
all alternatives →- FLFinetune LLaMA-7B on commodity GPUs using your own text2023 · github.com · ▲449
I'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!
- TLTune LLaMa3.1 on Google Cloud TPUs2024 · github.com · ▲189
Hey 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…
- SUSpeeding up LLM inference 2x times (possibly)2024 · asciinema.org · ▲419
Here'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…
- L3Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPUFeb 2026 · github.com · ▲395
Hi 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 →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, December 2023
the whole month →- BM
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…
Commerce · 2023 · beeper.com
- HA
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▲941Client 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
