from setuptools import setup
from torch.utils.cpp_extension import BuildExtension, CUDAExtension
import torch
# Compiler flags.
CXX_FLAGS = ["-g", "-O2", "-std=c++17"]
# TODO(woosuk): Should we use -O3?
NVCC_FLAGS = ["-O2", "-std=c++17","--gpu-max-threads-per-block=1024"]

ABI = 1 if torch._C._GLIBCXX_USE_CXX11_ABI else 0
CXX_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]

extra_compile_args={
    "cxx": CXX_FLAGS,
    "nvcc": NVCC_FLAGS,
}

setup(
    name="exllama_kernels",
    ext_modules=[
        CUDAExtension(
            name="exllama_kernels",
            sources=[
                "exllama_kernels/exllama_ext.cpp",
                "exllama_kernels/cuda_buffers.cu",
                "exllama_kernels/cuda_func/column_remap.cu",
                "exllama_kernels/cuda_func/q4_matmul.cu",
                "exllama_kernels/cuda_func/q4_matrix.cu",
            ],
            extra_compile_args=extra_compile_args,
        )
    ],
    cmdclass={"build_ext": BuildExtension},
)