O(n) and Fast JIT Compiler Workload (Transformer > 70 layers)

hikettei / Caten

[wip] Deep Learning Compiler based on Polyhedral Compiler and Light-weight IRs based on Optimizing Pattern Matcher

https://hikettei.github.io/Caten/

Other

20 stars 3 forks source link

Open hikettei opened 1 week ago

hikettei commented 1 week ago

[ ] #192
- [ ] the idea is that caching the result of simplifies defined in the defmodule.
- [ ] 70 layers transformer simplification in zero cost
[x] #228
- [x] Maximize the use of the schedule caching
- [ ] 1 Layer Transformer JIT Time = 70 Layer Transformer JIT Time
- [ ] remove id->users

hikettei commented 1 week ago

Currently (with JIT=1, including gcc time)

what we should optimize is not an algorithm, but the number of N (input nodes) by caching the aasm per modules.