#!/usr/bin/env bash

source /usr/local/bin/fastpt -C
export MAX_JOBS=64
# export TORCH_CUDA_ARCH_LIST=70
export DCNv4_BUILD_SO=1

python setup.py bdist_wheel 2>&1 | tee debug_$(date +%Y%m%d_%H%M%S).log

# /usr/local/lib/python3.10/dist-packages/torch/utils/cpp_extension.py
# 234:239
# COMMON_NVCC_FLAGS = [
#     '-D__CUDA_NO_HALF_OPERATORS__',
#     '-D__CUDA_NO_HALF_CONVERSIONS__',
#     '-D__CUDA_NO_BFLOAT16_CONVERSIONS__',
#     '-D__CUDA_NO_HALF2_OPERATORS__',
#     '--expt-relaxed-constexpr'
# ]

# /opt/dtk/cuda/cuda/include/cuda_pipeline_helpers.h
# 22行注释