[general]
name = "layer_norm"
universal = false

[torch]
src = [
  "torch-ext/torch_binding.cpp",
  "torch-ext/torch_binding.h",
]

[kernel.layer_norm]
depends = ["torch"]
backend = "cuda"
cuda-capabilities = [
    "8.0",
    "8.9",
    "9.0",
    "10.0",
    "12.0",
]
include = ["."]
src = [
    "layer_norm/ln.h",
    "layer_norm/ln_api.cpp",
    "layer_norm/ln_bwd_1024.cu",
    "layer_norm/ln_bwd_1280.cu",
    "layer_norm/ln_bwd_1536.cu",
    "layer_norm/ln_bwd_2048.cu",
    "layer_norm/ln_bwd_256.cu",
    "layer_norm/ln_bwd_2560.cu",
    "layer_norm/ln_bwd_3072.cu",
    "layer_norm/ln_bwd_4096.cu",
    "layer_norm/ln_bwd_512.cu",
    "layer_norm/ln_bwd_5120.cu",
    "layer_norm/ln_bwd_6144.cu",
    "layer_norm/ln_bwd_7168.cu",
    "layer_norm/ln_bwd_768.cu",
    "layer_norm/ln_bwd_8192.cu",
    "layer_norm/ln_bwd_kernels.cuh",
    "layer_norm/ln_fwd_1024.cu",
    "layer_norm/ln_fwd_1280.cu",
    "layer_norm/ln_fwd_1536.cu",
    "layer_norm/ln_fwd_2048.cu",
    "layer_norm/ln_fwd_256.cu",
    "layer_norm/ln_fwd_2560.cu",
    "layer_norm/ln_fwd_3072.cu",
    "layer_norm/ln_fwd_4096.cu",
    "layer_norm/ln_fwd_512.cu",
    "layer_norm/ln_fwd_5120.cu",
    "layer_norm/ln_fwd_6144.cu",
    "layer_norm/ln_fwd_7168.cu",
    "layer_norm/ln_fwd_768.cu",
    "layer_norm/ln_fwd_8192.cu",
    "layer_norm/ln_fwd_kernels.cuh",
    "layer_norm/ln_kernel_traits.h",
    "layer_norm/ln_parallel_bwd_1024.cu",
    "layer_norm/ln_parallel_bwd_1280.cu",
    "layer_norm/ln_parallel_bwd_1536.cu",
    "layer_norm/ln_parallel_bwd_2048.cu",
    "layer_norm/ln_parallel_bwd_256.cu",
    "layer_norm/ln_parallel_bwd_2560.cu",
    "layer_norm/ln_parallel_bwd_3072.cu",
    "layer_norm/ln_parallel_bwd_4096.cu",
    "layer_norm/ln_parallel_bwd_512.cu",
    "layer_norm/ln_parallel_bwd_5120.cu",
    "layer_norm/ln_parallel_bwd_6144.cu",
    "layer_norm/ln_parallel_bwd_7168.cu",
    "layer_norm/ln_parallel_bwd_768.cu",
    "layer_norm/ln_parallel_bwd_8192.cu",
    "layer_norm/ln_parallel_fwd_1024.cu",
    "layer_norm/ln_parallel_fwd_1280.cu",
    "layer_norm/ln_parallel_fwd_1536.cu",
    "layer_norm/ln_parallel_fwd_2048.cu",
    "layer_norm/ln_parallel_fwd_256.cu",
    "layer_norm/ln_parallel_fwd_2560.cu",
    "layer_norm/ln_parallel_fwd_3072.cu",
    "layer_norm/ln_parallel_fwd_4096.cu",
    "layer_norm/ln_parallel_fwd_512.cu",
    "layer_norm/ln_parallel_fwd_5120.cu",
    "layer_norm/ln_parallel_fwd_6144.cu",
    "layer_norm/ln_parallel_fwd_7168.cu",
    "layer_norm/ln_parallel_fwd_768.cu",
    "layer_norm/ln_parallel_fwd_8192.cu",
    "layer_norm/ln_parallel_residual_bwd_kernels.cuh",
    "layer_norm/ln_parallel_residual_fwd_kernels.cuh",
    "layer_norm/ln_utils.cuh",
    "layer_norm/static_switch.h"
]
cxx-flags = ["-DFLASHATTENTION_DISABLE_PYBIND", "-mcmodel=large"]
cuda-flags = [
    "-O3",
    "-U__CUDA_NO_HALF_OPERATORS__",
    "-U__CUDA_NO_HALF_CONVERSIONS__",
    "-U__CUDA_NO_BFLOAT16_OPERATORS__",
    "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
    "-U__CUDA_NO_BFLOAT162_OPERATORS__",
    "-U__CUDA_NO_BFLOAT162_CONVERSIONS__",
    "--expt-relaxed-constexpr",
    "--expt-extended-lambda",
    "--use_fast_math",
]