| This CUDA extension implements fused matmul + bias (forward and backward), and fused matmul + bias + gelu | |
| (forward and backward), adapted from Apex's | |
| [FusedDense](https://github.com/NVIDIA/apex/tree/master/apex/fused_dense). | |
| 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. | |
| ```sh | |
| cd csrc/fused_dense_lib && pip install . | |
| ``` | |