GitXplorerGitXplorer
c

torchcsr

public
1 stars
0 forks
0 issues

Commits

List of commits on branch main.
Unverified
03cfe69e428a78eef791d8327d7f315c3ca3ef50

Padded+masked softmax and better README

ccpuhrsch committed 4 years ago
Unverified
442b092f103144ff3c3af170dff2c683081ea1ea

Latest changes

ccpuhrsch committed 4 years ago
Unverified
6908bb76060690d252f08974682ee003eab1bdf3

build_tools

committed 4 years ago
Unverified
3a58d3fa0275301bffed10e1b330908f1278cb57

And some more CMake

committed 4 years ago
Unverified
22f6c3748647e9ad68899e06700a96675c898ec8

Some more CMake

committed 4 years ago
Unverified
f9a9a9f7991474ed4be956f4991aa6cc3f2e992d

Some more CMake

committed 4 years ago

README

The README file for this repository.

Build

Run on Linux, Volta GPU, recent PyTorch nightly

python setup.py clean && TORCH_CUDA_ARCH_LIST=Volta python setup.py develop

Run softmax

Run

python run.py

which contains a padded+masked implementation of softmax and a call into sparse_softmax.

TODO

  • Sparse CUDA kernel is registered at CPU with to/from device conversion. SparseCsrCUDA has no constructor, but we can add coverage out of tree here.

  • Need to setup row_offsets correctly in call to sputnik sparse_softmax