Create your own GitHub profile
Sign up for your own profile on GitHub, the best place to host code, manage projects, and build software alongside 50 million developers.
Sign upPinned
2,124 contributions in the last year
Contribution activity
November 2020
Created 113 commits in 4 repositories
Created 2 repositories
- geohot/pyopencl Python
- geohot/ipykernel Python
Created a pull request in commaai/openpilot that received 4 comments
set cpubw perf governor
I don't know the power or thermal implications of this, but CPU <-> GPU transfers (which I think is just memcpy) are 4x faster with the change. To …
Opened 11 other pull requests in 4 repositories
commaai/openpilot
6
merged
1
closed
geohot/tinygrad
1
open
1
merged
inducer/pyopencl
1
merged
ipython/ipykernel
1
open
Reviewed 15 pull requests in 2 repositories
geohot/tinygrad 11 pull requests
- Define `ProfileOp` class once
- better integration of tensor static methods and gpu interface slightly simplified
- Topological sort, zero_grads
- Backwards ops supporting broadcasting
- General broadcasting of binary operations
- Conv2D backward on GPU
- :hammer: clean up
- Fix OOM in efficientnet on CPU
- strided Pool funcs
- implemented backward for Pad2D & test
- alternative to einsum which is slow in extreme cases
commaai/openpilot 4 pull requests
Created an issue in geohot/tinygrad that received 20 comments
GPU=1 examples/train_efficientnet.py doesn't work
Fix broadcasting. Munch 1's from left for mod, munch 1's from right for div.
Opened 16 other issues in 2 repositories
geohot/tinygrad
10
closed
3
open
- We have to lose 4 lines!
- GPU EfficientNet is weirdly slow
- RMSProp and Adam use .data
- PyOpenCL doesn't work in GitHub actions
- Currently at 989 lines. Reduce lines while improving readability!
- Padding is wrong in enet
- Broadcasting on GPU doesn't work for BS > 1
- b0c0c5d0d62e40df97999c8df3c1809ecb7842db breaks enet inference on CPU!
- README should shill harder for our ImageNet support
- TestMNIST.test_conv_gpu doesn't work
- Stride is wrong here
- tinygrad has grown too large!
- Many GPU ops are missing for backward pass