| .. | ||
| fused_dense_cuda.cu | ||
| fused_dense.cpp | ||
| README.md | ||
| setup.py | ||
This CUDA extension implements fused matmul + bias (forward and backward), and fused matmul + bias + gelu (forward and backward), adapted from Apex's FusedDense. We make it work for bfloat16.
For best performance, you should use CUDA >= 11.8. CuBLAS versions before this doesn't have the best matmul + bias + gelu performance for bfloat16.
It has only been tested on A100s.
cd csrc/fused_dense_lib && pip install .