Collection of toy implementations of memory efficient attention.

All numpy implementations are in attn.py, and there's a cuda implementation in attn_chunk_q_chunk_kv_cuda/attn_chunk_q_chunk_kv_kernel.cu

Building the cuda implementation

cd attn_chunk_q_chunk_kv_cuda
python setup.py install

Running the tests

No formal testing framework, just run

python test.py

If no assertions are thrown, tests pass :)