layer_norm / ln_fwd_256.cu
Ubuntu
feat: torch 2.5.1+cu124
a3d4355
raw
history blame
925 Bytes
#include "ln_fwd_kernels.cuh"
// Create forward launch function and register. Macro signature:
// HIDDEN_SIZE, WTYPE, ITYPE, RYTPE, OTYPE, CTYPE, CTAS_PER_ROW, WARPS_M, WARPS_N, BYTES_PER_LDG
REGISTER_FWD_LAUNCHER( 256, fp32, fp32, fp32, fp32, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp16, fp32, fp32, fp32, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp32, fp16, fp32, fp16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp16, fp16, fp32, fp16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp32, fp16, fp16, fp16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp32, bf16, fp32, bf16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, bf16, bf16, fp32, bf16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp32, bf16, bf16, bf16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, fp16, fp16, fp16, fp16, fp32, 1, 4, 1, 16);
REGISTER_FWD_LAUNCHER( 256, bf16, bf16, bf16, bf16, fp32, 1, 4, 1, 16);