| | #include "ln_parallel_residual_bwd_kernels.cuh" |
| |
|
| | |
| | |
| |
|
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp32, fp32, fp32, fp32, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp16, fp32, fp32, fp32, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp32, fp16, fp32, fp16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp16, fp16, fp32, fp16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp32, fp16, fp16, fp16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp32, bf16, fp32, bf16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, bf16, bf16, fp32, bf16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp32, bf16, bf16, bf16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, fp16, fp16, fp16, fp16, fp32, 1, 4, 1, 16, 4); |
| | REGISTER_PARALLEL_BWD_LAUNCHER( 1024, bf16, bf16, bf16, bf16, fp32, 1, 4, 1, 16, 4); |
| |
|