Hi! I recently got nerdsniped by Modular and would love to do a project in it. I also have a long-standing interest in ML for System (https://mlforsystems.org/). I would love to see whether there were past projects that people have done around the area, or whether people would have any pointers of things that I could potentially explore here.
We invite submission of up to 4-page extended abstracts in the broad area of using machine learning in the design and management of computer systems . We are especially interested in submissions that move beyond using machine learning to replace numerical heuristics.
This year, we additionally look for
Using LLMs for systems challenges, such as program synthesis for hardware and other specialized domains.
Applying ML to systems issues that emerge from large-scale training and serving, such as compiler partitioning schemes for training LLMs across thousands of GPU or TPU devices.
Applying ML for compute sustainability, including power/energy/carbon optimization. Examples include energy-aware job scheduling, dynamic power management based on workload and carbon predictions, and ML-driven carbon footprint assessment for cloud datacenters.
Regarding the second bullet point, I do think that Modular is already latched onto that problem and chewing away at it. It feels like they are thinking that ML is used for a control plane because who could possibly rethink the stack and create a solutions for different types of hardware at scale.
Back to the first bullet point, there was a LLM-SQL buzz about a year ago, originally direct to database and later using RAG. I don’t have any real nuggets to suggest for this one.
Regarding mojo, you might to take a peek at Nabla-ml
If you have a model in mind, have a peek at Modular builds
And if you have your own graph in mind or want to leverage a systems approach to a graph, I’d suggest