2025-02-25 22:52:41 +08:00
|
|
|
import torch
|
|
|
|
|
|
|
|
|
|
from . import jit
|
|
|
|
|
from .jit_kernels import (
|
|
|
|
|
gemm_fp8_fp8_bf16_nt,
|
|
|
|
|
m_grouped_gemm_fp8_fp8_bf16_nt_contiguous,
|
|
|
|
|
m_grouped_gemm_fp8_fp8_bf16_nt_masked,
|
2025-02-26 18:37:22 +08:00
|
|
|
ceil_div,
|
2025-02-25 22:52:41 +08:00
|
|
|
set_num_sms, get_num_sms,
|
|
|
|
|
get_col_major_tma_aligned_tensor,
|
|
|
|
|
get_m_alignment_for_contiguous_layout
|
|
|
|
|
)
|
|
|
|
|
from .utils import bench, bench_kineto, calc_diff
|
