layer_norm / ln_parallel_bwd_1024.cu
Ubuntu
feat: torch 2.5.1+cu124
a3d4355
raw
history blame
1.1 kB
#include "ln_parallel_residual_bwd_kernels.cuh"
// Create backward launch function and register. Macro signature:
// HIDDEN_SIZE, WTYPE, ITYPE, RTYPE, OTYPE, CTYPE, CTAS_PER_ROW, WARPS_M, WARPS_N, BYTES_PER_LDG, BYTES_PER_LDG_FINAL
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);