| #include "ln_parallel_residual_bwd_kernels.cuh" |
|
|
| |
| |
|
|
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp32, fp32, fp32, fp32, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp16, fp32, fp32, fp32, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp32, fp16, fp32, fp16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp16, fp16, fp32, fp16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp32, fp16, fp16, fp16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp32, bf16, fp32, bf16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, bf16, bf16, fp32, bf16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp32, bf16, bf16, bf16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, fp16, fp16, fp16, fp16, fp32, 1, 1, 4, 16, 4); |
| REGISTER_PARALLEL_BWD_LAUNCHER( 2048, bf16, bf16, bf16, bf16, fp32, 1, 1, 4, 16, 4); |