11 lines
455 B
Markdown
11 lines
455 B
Markdown
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.
|
|
```sh
|
|
cd csrc/fused_dense_lib && pip install .
|
|
```
|