# /// script
# requires-python = ">=3.10"
# dependencies = [
# "numpy",
# "torch==2.8.0",
# "kernels-benchmark-tools",
# "kernels",
# ]
#
# [tool.uv.sources]
# kernels-benchmark-tools = { path = "../../../../../tools", editable = true }
# ///
import torch
import sys
from kernels_benchmark_tools import KernelTypeEnum, run_benchmark
from kernels import get_kernel
# Load the deformable DETR kernel
deformable_detr = get_kernel("kernels-community/deformable-detr")
def hf_kernels_deformable_detr(
value, spatial_shapes, level_start_index, sampling_locations, attention_weights, im2col_step=64
):
"""HuggingFace Kernels Deformable DETR Multi-Scale Deformable Attention"""
return deformable_detr.ms_deform_attn_forward(
value=value,
spatial_shapes=spatial_shapes,
level_start_index=level_start_index,
sampling_loc=sampling_locations,
attn_weight=attention_weights,
im2col_step=im2col_step
)
run_benchmark(
kernel_type=KernelTypeEnum.DEFORMABLE_DETR,
impl_name="hf_kernels_deformable_detr",
impl_tags={"family": "hf-kernels", "backend": "cuda"},
impl_func=hf_kernels_deformable_detr,
dtype="float32",
)