Le thème : les modèles de language et la robotique open hardware. Si ça vous intéresse de découvrir une autre facette que Skynet et la machine à billet,
Andrej Karpathy just released a new repo with an implementation of training LLM with pure C/Cude with a few lines of code 🚀. This repo, according to Andrej Karpathy, is still WIP, and the first working example is of GPT-2 (or the grand-daddy of LLMS 😅) 👇🏼
This weekend, on Sunday afternoon non-important facts - the PyTorch name👇🏼
The #PyTorch library was built as a wrapper for the Torch open-source project. Therefore, it was named #Python Torch and, in short, PyTorch.
Torch is a machine-learning library and scientific computing framework based on Lua (I just learned about this language). The Torch library has a deep learning implementation in C, which PyTorch was primarily based on.
Hey, any Fedora folks out there ( also could include future Fedora people!) with involvement in upstream Pytorch and related development communities? #pytorch
We recently upgraded from #PyTorch 1.8.1 to 2.0.0 and all of a sudden our ONNX exports for one model seem to be broken, we get too many unspecific dimensions if we set dynamic_axes at all on the outputs (i.e. for the batch size)
I have a terrible feeling that this is some kind of change in the tracing JIT, but haven't been able to find a tidy reproducer yet.
Now again #LLMs: if you do not want to train your own #ai foundation model, you can patch it with so-called #adapters. Benjamin Trim talked about their own open-source adapter micro framework: #Refiners work on top of #PyTorch and use declarative layers to patch models, context API to store state. #fosdem2024
A new crash course for getting started with #CUDA with #Python by Jeremy Howard 🚀. CUDA is NVIDIA's programming model for parallel computing on GPUs. CUDE is being used by tools such as #PyTorch#tensorflow and other #deeplearning and LLMs frameworks to speed up calculations. The course covers the following topics:
✅ Setting up CUDA
✅ CUDA foundation
✅ Working with Kernel
✅ CUDA with PyTorch
Your submission in “GPU-based ODE solvers which are 20x-100x faster than those in #jax and #pytorch” was removed for Testing functions using new Proton UI.
During the holidays, i.e. from the 23rd to the 7th, I will not work.
I'd like to continue playing with the #julialang and some #pytorch but it's really for pleasure.
I will read some stuff about brains and perception but it'll be Ed Yong's "An Immense World", and perhaps McCulloch and Ashby.
Also, I'm finishing Dune.
If you too have the "privilege" of taking a break, what are you up to?
JOSS publishes articles about open source research software. It is a free, open-source, community driven and developer-friendly online journal. JOSS reviews involve downloading and installing the software, and inspecting the repository and submitted paper for key elements
Please reach out if you are interested in reviewing this paper or know one who could review this paper.
Trying to live the single-(usbc)-cable-dream at work as well by going through my old thinkpad dock: mac to dock via usbc, dock to Delly U2713HM DisplayPort. However, DP to DP connection from dock to old Dell U2713HM display only sometimes flickers on and often not. usbc to DP from dock is solid. DP-DP cable is extremely sus. Work has ordered new DP-DP cable and I'm crossing my fingers. Singel cable life is fun.
Again discovered how unbelievably badly macOS renders fonts on resolutions that Apple believes to be too low, in this case my Dell Ultrasharp 27" at 2560x1440 aka QHD at work. Microsoft and Windows do an absolutely great job on exactly the same hardware, and fonts look great.
BTW, although macOS does marginally better on my 32" 4K Dell IPS display at home, here Windows even further increases its font rendering dominance with fractional scaling and cleartype.
As is often the case with Apple, there is a (paid) third-party software tool that works around their attempts to improve matters, namely BetterDisplay: https://github.com/waydabber/BetterDisplay
As I wrote in my notes some years ago when this became apparent, this is just the company's philosophy. They want to control all the hardware. They will begrudgingly let you use some third party displays, but they pick their battles to look good. In this case, it does feel quite user-hostile.
Ran into a M1-specific bug in the ruff vscode extension, where the arm64 extension build bundles the x86_64 ruff binary. Worked-around, and then reported at https://github.com/astral-sh/ruff-vscode/issues/364
I've been looking at Apple's MLX machine learning / array framework for Apple Silicon https://github.com/ml-explore/mlx as well as at CoreML because I'm curious whether this will give me faster Jina AI embeddings inference on the M1 than I'm currently getting with the PyTorch MPS backend, which is muuuuuch slower on this M1 Pro 10C / 16C GPU / 16C neural than PyTorch CUDA on my oooold GeForce RTX2070 with 8GB.
@gabmus
I use #rocm 5.7 to run #opencl, google's #jax (for pymc), and #pytorch on two vega cards (Vega 64 and Radeon pro WX9100) on arch and ubuntu. They all run Ok, but correct setup needs some googling around, and jax beeds exporting some #xla flags. Situation is much, much better than 2 years ago, though. @oblomov
TorchServe Potentially Exposed to Remote Code Execution
Israeli-based security company Oligo has uncovered multiple vulnerabilities in TorchServe, the tool used to serve PyTorch models, that could allow an attacker to run arbitrary code on vulnerable systems. The vulnerabilities have been promptly fixed in TorchServe version 0.82.
Je bosse au 4/5 sur les modèles de langage (LLM, parfois appelées IAs) et à 2/5 sur la robotique open hardware AMA (jlai.lu) French
Hello!...
GPU-based ODE solvers which are 20x-100x faster than those in #jax and #pytorch (programming.dev)
cross-posted from: programming.dev/post/8391233...