Repository navigation
collaboration from OjasX project #175
mkrishna793
started this conversation in
Show and tell
Replies: 2 comments 2 replies
|
Hello, your idea is interesting, will you be open for a discussion sometime? |
1 reply
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
OjasX: An Open, Math-Level Compiler for Efficient AI Training
Hi, I’m Mohana Krishna. I’ve been independently building OjasX for the past six months.
OjasX is not just an “open-source CUDA.” It is a math-level compiler designed to reduce the actual computational and memory cost of deep-learning workloads—before kernel generation. The goal is to make PyTorch training, fine-tuning, and inference practical on much smaller GPUs, unified-memory systems, and integrated GPUs.
The Problem
Large-model training is expensive because of activation memory, gradients, optimizer states, and repeated matrix arithmetic. Many workloads that currently require multiple high-end GPUs could become feasible on far smaller hardware if the underlying mathematics were represented and executed more efficiently.
OjasX focuses on transforming arithmetic—especially matrix reductions—into lower-cost equivalent forms, rather than only fusing kernels or porting code between GPU vendors.
What I’m Building
The project combines:
In early experiments, I have observed activation and gradient-memory reductions of up to 98.6% in some cases. These results are experimental and still need broader validation, but they show why I believe this direction is worth pursuing.
The long-term goal is ambitious: reduce the hardware needed for workloads that might otherwise require around 20 GPUs, while preserving near bit-level numerical accuracy where possible.
Current Status
The public repository still contains earlier compiler prototypes and does not reflect the latest architecture:
I plan to push a major update before November with clearer documentation, tests, benchmarks, and the new compiler design. Please treat the current code as an early work-in-progress.
I’m also building NexusSilicon, an open-source EDA tool for chip design:
Looking for Contributors
I can do much of the work alone, but OjasX needs a community to grow into something practical and maintainable. I’m looking for people interested in:
You do not need to be an expert. Help with documentation, tests, benchmarks, issue triage, or reproducing results would be very valuable.
This project is not against NVIDIA or CUDA. It is about giving researchers, students, startups, and builders more freedom to train and deploy AI without being limited by GPU memory, cloud costs, or proprietary tools.
If this interests you, please comment, open an issue, or reach out. Let’s build this openly.
All reactions