File size: 838 Bytes
58b76f1 81fff32 f2afc26 58b76f1 81fff32 58b76f1 05bebc1 58b76f1 9b13459 58b76f1 05bebc1 9b13459 05bebc1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 |
# /// script
# requires-python = ">=3.10"
# dependencies = [
# "numpy",
# "torch==2.8.0",
# "kernels-benchmark-tools",
# ]
#
# [tool.uv.sources]
# kernels-benchmark-tools = { path = "../../../../../tools", editable = true }
# ///
import torch
import sys
from kernels_benchmark_tools import KernelTypeEnum, run_benchmark
def torch_flash(q, k, v):
qt, kt, vt = (x.transpose(1, 2).contiguous() for x in (q, k, v))
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.FLASH_ATTENTION):
o = torch.nn.functional.scaled_dot_product_attention(qt, kt, vt)
return o.transpose(1, 2).contiguous()
run_benchmark(
kernel_type=KernelTypeEnum.ATTENTION,
impl_name="torch_flash_ma",
impl_tags={"family": "torch-sdpa", "backend": "FLASH", "compile": "max-autotune"},
impl_func=torch_flash,
) |