tinygrad copied to clipboard
You like pytorch? You like micrograd? You love tinygrad! ❤️
Tile CL images when needed, to fit into the 16384 max Metal image size; gets me to ~4.8s/iteration for SD on M1 Pro with `OPENCL=1 FLOAT16=1 LAZY=1`. ![rendered](https://user-images.githubusercontent.com/8733615/189498322-88e1a29a-bd1f-4da9-b8a0-b425dec3fa79.png)
Closes [Write arbitrary striding for maxpool](https://github.com/geohot/tinygrad/issues/366) TODO: Work out the backwards pass
- [x] Support asymmetric padding - [ ] Support caching of only unresolved LazyBuffers `See CACHE_LAZYBUFFERS` - [ ] Support not running the convs multiple times, aka late split. This...
Anyone want a challenge? Check out TestComplexShapeTracker, those tests are failing but they shouldn't be. Can you fix them in a few lines?
This will make backward convs fast.
This is doable as a day project if someone wants to do it!
Fixing issue 139 to be able to run all tests without Torch Acceptable amount of lines?
With llops, we should be able to write something fast in C to replace them.