New Show Hacker News story: Show HN: 80% faster, 50% less memory, 0% loss of accuracy Llama finetuning

Show HN: 80% faster, 50% less memory, 0% loss of accuracy Llama finetuning
55 by danielhanchen | 10 comments on Hacker News.
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://ift.tt/5asFgeP... has more info about Unsloth! Hopefully you can try it out! Wrote a blog post at https://ift.tt/FmnplJh if you want to learn more about our manual hand derived backprop or Triton kernels and stuff! Thanks once again!

Comments

Post a Comment

Popular posts from this blog

New Show Hacker News story: Show HN: Natural language Twitter search using Codex

Show HN: Natural language Twitter search using Codex 3 by andyk | 0 comments on Hacker News. We built a structured search engine for Twitter called Bird SQL, available at https://ift.tt/7BU0hHb . Our search interface uses OpenAI Codex to translate natural language to SQL. Our backend then verifies the SQL, executes it, and displays the results on the web app. This makes large structured datasets like a scrape of Twitter easy for anyone to explore. As background, while working on text-to-SQL as a general problem, we came to believe one of its most powerful applications is as a search tool because: - SQL is hard to write by hand and prone to errors - It allows you to iterate quickly if you’re exploring a new dataset - A lot of contextual information that you’d normally have to internalize (e.g. your data’s schema) can be automatically generated and offloaded to the language model Using large language models (LLMs) like Codex to write the SQL for you means you don’t have to worry about
Read more

Internet Download Manager Universal Crack is Here ! [IDM 6.25 Build 10 UPDATED]

Image
Internet Download Manager Universal Crack Internet Download Manager aka IDM developers release a new build of  their well-known Download  Accelerator; Internet Download Manager  regularly. So everybody who uses pirated versions get so  pissed off that they have to find a new version of crack every time when an IDM update comes. So  we made a really awesome fix for that An Universal Web Crack. A new era in scene world. Web crack is a new concept of cracks that download and update crack content as the target program gets updated.  So whenever IDM releases a new version. Update your crack using built-in updater and apply crack. Easy as that ! You can simply trust crack because it has a history of more than 3 year with regular updates, and also it has more than 10000K worldwide downloads. How to Crack any IDM Version ? Install the latest version of IDM / Or update via quick update (yes its possible now)  Turn off any virus guards Download and 
Read more

New Show Hacker News story: Show HN: Movis – A Video Editing Library in Python

Show HN: Movis – A Video Editing Library in Python 2 by rezoolab | 0 comments on Hacker News. Originally, this library was created due to my motivation to make videos in Python. Of course, I considered other Python libraries, but I decided to create a new library for video editing because I needed the following features and also for study purposes: - The existence of composition. Especially, the ability to perform advanced video editing by inserting compositions within compositions. - The ability to add animations based on keyframes. - Effects like chroma key and drop shadow. - A caching mechanism to render the same frame quickly. - Various blending modes. These features, despite being standard in proprietary video editing software, were not present in previous Python libraries. Currently, I am using this library to automatically generate videos from raw materials. I would be grateful for any comments, thoughts, or requests.
Read more