Inspiration
The inspiration for this project came from Huawei's Custom Challenge #2: Automatic Decision for Re-computation in AI model optimization. We were highly interested in the research aspects of this project looking into new algorithms implemented in the recent years
What it does
Our scheduler implements multiple algorithms to solve the AI model computation scheduling problem based on memory constraints. We took a multi-algorithmic approach implementing naive, greedy, heuristic, beam search, branch bounding scheduling algorithms to produce our results. We have also implemented aggressive memory management to remove intermediate results when memory is tight and minimizes the memory peak through node recomputation.
Challenges we ran into
Our main issue was our memory peak hitting the ceiling due to our focus on finding the optimal time.
Accomplishments that we're proud of
What we are proud of is our strong understanding of theory along with the implementation of several recent (1-2years) research papers work on re-computation algorithms
What's next for Racoon Works Hua Wei CC 2
We plan on bettering our implementation of our algorithm, making our re computation logic smarter and fall backs more consistent as currently we have a strong theory understanding but our implementation could be improved.
Log in or sign up for Devpost to join the conversation.