Welcome to the Apsi community ๐ Apsi 2.4.4 is here
Hi everyone, and welcome!
I'm the developer of Apsi, a programming language for AI and physics. This is the place to share what you build, ask questions, report what breaks, and help shape where Apsi goes next.
What Apsi is
- Reads like Python, with physical units and tensor shapes checked before the program runs
- One installer: no Python, PyTorch or CUDA toolkit to set up
- Trains AI on your own NVIDIA GPU with kernels Apsi generates itself from the display driver
- Runs language models from Hugging Face straight on your card
- Converts Python and PyTorch code to Apsi with the same behaviour
What's new in 2.4.4
On a Ryzen 5 5500 + RTX 3060 12GB, same model and same data:
| Training a 10.9M-param GPT (bf16) | Speed |
|---|---|
| Apsi | 151,744 tokens/s |
| PyTorch 2.14 | 126,993 tokens/s |
- The card stays at a steady ~77 ยฐC: Apsi moves the power limit instead of pausing
- Several GPUs at once with
device="cuda:all" - Qwen 0.5B / 3B / 14B run fully on a 12 GB card
- The Apsi GPU Tuner on the download page sets the best power limit for any NVIDIA card in one click
Every number comes with its raw log on the Benchmarks page.
Get started in two minutes
- Download Apsi 2.4.4 for Windows or Linux
- Run
apsi new helloto create a project - Read Start in 2 minutes and the GPU guide
How you can help
Apsi is young and still has rough edges, and testing on more machines is the most valuable thing right now:
- Show & tell: share what you built, and your training speed on your own GPU
- Questions: stuck on something? Ask, there are no silly questions
- Bugs: tell me what broke, on which OS and GPU, and the error you saw
- Libraries: publish a package with
apsi publishand post about it here
Please keep it friendly and on topic. Thanks for being here early. I read every post! ๐
0 comments
Sign in to comment.