setup.py

# Copyright (c) 2023, Tri Dao.

import sys
import functools
import warnings
import os
import re
import ast
import glob
import shutil
import tempfile
from pathlib import Path
from typing import Literal, Optional
from packaging.version import parse, Version
import platform

from setuptools import setup, find_packages
import subprocess

import urllib.request
import urllib.error
from wheel.bdist_wheel import bdist_wheel as _bdist_wheel

import torch
from torch.utils.cpp_extension import (
    BuildExtension,
    CppExtension,
    CUDAExtension,
    CUDA_HOME,
    ROCM_HOME,
    IS_HIP_EXTENSION,
)


with open("README.md", "r", encoding="utf-8") as fh:
    long_description = fh.read()


# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))

BUILD_TARGET = os.environ.get("BUILD_TARGET", "auto")

if BUILD_TARGET == "auto":
    if IS_HIP_EXTENSION:
        IS_ROCM = True
    else:
        IS_ROCM = False
else:
    if BUILD_TARGET == "cuda":
        IS_ROCM = False
    elif BUILD_TARGET == "rocm":
        IS_ROCM = True

PACKAGE_NAME = "flash_attn"

BASE_WHEEL_URL = (
    "https://github.com/Dao-AILab/flash-attention/releases/download/{tag_name}/{wheel_name}"
)

# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels
# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation
FORCE_BUILD = os.getenv("FLASH_ATTENTION_FORCE_BUILD", "FALSE") == "TRUE"
SKIP_CUDA_BUILD = os.getenv("FLASH_ATTENTION_SKIP_CUDA_BUILD", "FALSE") == "TRUE"
# For CI, we want the option to build with C++11 ABI since the nvcr images use C++11 ABI
FORCE_CXX11_ABI = os.getenv("FLASH_ATTENTION_FORCE_CXX11_ABI", "FALSE") == "TRUE"
ROCM_BACKEND: Optional[Literal["triton", "ck"]] = None
if IS_ROCM:
    ROCM_BACKEND = "triton" if os.getenv("FLASH_ATTENTION_TRITON_AMD_ENABLE", "FALSE") == "TRUE" else "ck"
NVCC_THREADS = os.getenv("NVCC_THREADS") or "4"

@functools.lru_cache(maxsize=None)
def cuda_archs() -> str:
    return os.getenv("FLASH_ATTN_CUDA_ARCHS", "80;90;100;110;120").split(";")


def get_platform():
    """
    Returns the platform name as used in wheel filenames.
    """
    if sys.platform.startswith("linux"):
        return f'linux_{platform.uname().machine}'
    elif sys.platform == "darwin":
        mac_version = ".".join(platform.mac_ver()[0].split(".")[:2])
        return f"macosx_{mac_version}_x86_64"
    elif sys.platform == "win32":
        return "win_amd64"
    else:
        raise ValueError("Unsupported platform: {}".format(sys.platform))


def get_cuda_bare_metal_version(cuda_dir):
    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    output = raw_output.split()
    release_idx = output.index("release") + 1
    bare_metal_version = parse(output[release_idx].split(",")[0])

    return raw_output, bare_metal_version


def add_cuda_gencodes(cc_flag, archs, bare_metal_version):
    """
    Adds -gencode flags based on nvcc capabilities:
      - sm_80/90 (regular)
      - sm_100/120 on CUDA >= 12.8
      - Use 100f on CUDA >= 12.9 (Blackwell family-specific)
      - Map requested 110 -> 101 if CUDA < 13.0 (Thor rename)
      - Embed PTX for newest arch for forward compatibility
    """
    # Always-regular 80
    if "80" in archs:
        cc_flag += ["-gencode", "arch=compute_80,code=sm_80"]

    # Hopper 9.0 needs >= 11.8
    if bare_metal_version >= Version("11.8") and "90" in archs:
        cc_flag += ["-gencode", "arch=compute_90,code=sm_90"]

    # Blackwell 10.x requires >= 12.8
    if bare_metal_version >= Version("12.8"):
        if "100" in archs:
            # CUDA 12.9 introduced "family-specific" for Blackwell (100f)
            if bare_metal_version >= Version("12.9"):
                cc_flag += ["-gencode", "arch=compute_100f,code=sm_100"]
            else:
                cc_flag += ["-gencode", "arch=compute_100,code=sm_100"]

        if "120" in archs:
            # sm_120 is supported in CUDA 12.8/12.9+ toolkits
            if bare_metal_version >= Version("12.9"):
                cc_flag += ["-gencode", "arch=compute_120f,code=sm_120"]
            else:
                cc_flag += ["-gencode", "arch=compute_120,code=sm_120"]


        # Thor rename: 12.9 uses sm_101; 13.0+ uses sm_110
        if "110" in archs:
            if bare_metal_version >= Version("13.0"):
                cc_flag += ["-gencode", "arch=compute_110f,code=sm_110"]
            else:
                # Provide Thor support for CUDA 12.9 via sm_101
                if bare_metal_version >= Version("12.8"):
                    cc_flag += ["-gencode", "arch=compute_101,code=sm_101"]
                # else: no Thor support in older toolkits

    # PTX for newest requested arch (forward-compat)
    numeric = [a for a in archs if a.isdigit()]
    if numeric:
        newest = max(numeric, key=int)
        cc_flag += ["-gencode", f"arch=compute_{newest},code=compute_{newest}"]

    return cc_flag


def get_hip_version():
    return parse(torch.version.hip.split()[-1].rstrip('-').replace('-', '+'))


def check_if_cuda_home_none(global_option: str) -> None:
    if CUDA_HOME is not None:
        return
    # warn instead of error because user could be downloading prebuilt wheels, so nvcc won't be necessary
    # in that case.
    warnings.warn(
        f"{global_option} was requested, but nvcc was not found.  Are you sure your environment has nvcc available?  "
        "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, "
        "only images whose names contain 'devel' will provide nvcc."
    )


def check_if_rocm_home_none(global_option: str) -> None:
    if ROCM_HOME is not None:
        return
    # warn instead of error because user could be downloading prebuilt wheels, so hipcc won't be necessary
    # in that case.
    warnings.warn(
        f"{global_option} was requested, but hipcc was not found."
    )


def detect_hipify_v2():
    try:
        from torch.utils.hipify import __version__
        from packaging.version import Version
        if Version(__version__) >= Version("2.0.0"):
            return True
    except Exception as e:
        print("failed to detect pytorch hipify version, defaulting to version 1.0.0 behavior")
        print(e)
    return False


def append_nvcc_threads(nvcc_extra_args):
    return nvcc_extra_args + ["--threads", NVCC_THREADS]


def rename_cpp_to_cu(cpp_files):
    for entry in cpp_files:
        shutil.copy(entry, os.path.splitext(entry)[0] + ".cu")


def validate_and_update_archs(archs):
    # List of allowed architectures
    allowed_archs = ["native", "gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1102", "gfx1150", "gfx1151", "gfx1200", "gfx1201"]

    # Validate if each element in archs is in allowed_archs
    assert all(
        arch in allowed_archs for arch in archs
    ), f"Invalid archs: {archs}. Allowed: {allowed_archs}"

    if "native" in archs and len(archs) > 1:
        raise ValueError(
            f"'native' cannot be combined with explicit archs: {archs}. "
            "Use either GPU_ARCHS='native' or GPU_ARCHS='gfx942;gfx950'."
        )


cmdclass = {}
ext_modules = []

# We want this even if SKIP_CUDA_BUILD because when we run python setup.py sdist we want the .hpp
# files included in the source distribution, in case the user compiles from source.
if IS_ROCM:
    if ROCM_BACKEND == "triton":
        if os.path.isdir(".git"):
            subprocess.run(["git", "submodule", "update", "--init", "third_party/aiter"], check=True)
        else:
            assert os.path.isdir("third_party/aiter"), (
                "third_party/aiter is missing, please use source distribution or git clone"
            )
        subprocess.run(
            [sys.executable, "-m", "pip", "install", "--no-build-isolation", "third_party/aiter"],
            check=True,
        )
    elif ROCM_BACKEND == "ck":
        if os.path.isdir(".git"):
            subprocess.run(["git", "submodule", "update", "--init", "csrc/composable_kernel"], check=True)
        else:
            assert os.path.exists("csrc/composable_kernel/example/ck_tile/01_fmha/generate.py"), (
                "csrc/composable_kernel is missing, please use source distribution or git clone"
            )
else:
    # CUDA: cutlass submodule
    if os.path.isdir(".git"):
        subprocess.run(["git", "submodule", "update", "--init", "csrc/cutlass"], check=True)
    else:
        assert os.path.exists("csrc/cutlass/include/cutlass/cutlass.h"), (
            "csrc/cutlass is missing, please use source distribution or git clone"
        )

if not SKIP_CUDA_BUILD and not IS_ROCM:
    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
    TORCH_MAJOR = int(torch.__version__.split(".")[0])
    TORCH_MINOR = int(torch.__version__.split(".")[1])

    check_if_cuda_home_none("flash_attn")
    # Check, if CUDA11 is installed for compute capability 8.0
    cc_flag = []
    if CUDA_HOME is not None:
        _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
        if bare_metal_version < Version("11.7"):
            raise RuntimeError(
                "FlashAttention is only supported on CUDA 11.7 and above.  "
                "Note: make sure nvcc has a supported version by running nvcc -V."
            )
        # Build -gencode (regular + PTX + family-specific 'f' when available)
        add_cuda_gencodes(cc_flag, set(cuda_archs()), bare_metal_version)
    else:
        # No nvcc present; warnings already emitted above
        pass

    # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
    # torch._C._GLIBCXX_USE_CXX11_ABI
    # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920
    if FORCE_CXX11_ABI:
        torch._C._GLIBCXX_USE_CXX11_ABI = True

    nvcc_flags = [
    "-O3",
    "-std=c++17",
    "-U__CUDA_NO_HALF_OPERATORS__",
    "-U__CUDA_NO_HALF_CONVERSIONS__",
    "-U__CUDA_NO_HALF2_OPERATORS__",
    "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
    "--expt-relaxed-constexpr",
    "--expt-extended-lambda",
    "--use_fast_math",
    # "--ptxas-options=-v",
    # "--ptxas-options=-O2",
    # "-lineinfo",
    # "-DFLASHATTENTION_DISABLE_BACKWARD",
    # "-DFLASHATTENTION_DISABLE_DROPOUT",
    # "-DFLASHATTENTION_DISABLE_ALIBI",
    # "-DFLASHATTENTION_DISABLE_SOFTCAP",
    # "-DFLASHATTENTION_DISABLE_UNEVEN_K",
    # "-DFLASHATTENTION_DISABLE_LOCAL",
    ]

    compiler_c17_flag=["-O3", "-std=c++17"]
    # Add Windows-specific flags
    if sys.platform == "win32" and os.getenv('DISTUTILS_USE_SDK') == '1':
        nvcc_flags.extend(["-Xcompiler", "/Zc:__cplusplus"])
        compiler_c17_flag=["-O2", "/std:c++17", "/Zc:__cplusplus"]

    # Opt-in: skip the num_splits==1 blocksize-alignment instantiation in the
    # splitkv dispatch. That alignment (PR #2448) compiles a second splitkv kernel
    # tree per head dim, roughly doubling ptxas time for hd32/64/96/128 (hd64 can
    # stall ptxas for hours). Disabling it keeps num_splits==1 correct but no
    # longer bitwise-identical to the standard kernel. nvcc-only (header in .cu).
    if os.getenv("FLASH_ATTENTION_DISABLE_SPLIT_ALIGNMENT", "FALSE") == "TRUE":
        nvcc_flags.append("-DFLASHATTENTION_DISABLE_SPLIT_ALIGNMENT")

    ext_modules.append(
        CUDAExtension(
            name="flash_attn_2_cuda",
            sources=[
                "csrc/flash_attn/flash_api.cpp",
                "csrc/flash_attn/src/flash_fwd_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim32_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim32_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim64_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim96_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim128_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim192_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_hdim256_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim32_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim64_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim96_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim128_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim192_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_bwd_hdim256_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_fp16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim32_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim32_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim64_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim96_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim128_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim192_bf16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_fp16_causal_sm80.cu",
                "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_causal_sm80.cu",
            ],
            extra_compile_args={
                "cxx": compiler_c17_flag,
                "nvcc": append_nvcc_threads(nvcc_flags + cc_flag),
            },
            include_dirs=[
                Path(this_dir) / "csrc" / "flash_attn",
                Path(this_dir) / "csrc" / "flash_attn" / "src",
                Path(this_dir) / "csrc" / "cutlass" / "include",
            ],
        )
    )
elif not SKIP_CUDA_BUILD and IS_ROCM:
    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
    TORCH_MAJOR = int(torch.__version__.split(".")[0])
    TORCH_MINOR = int(torch.__version__.split(".")[1])

    # Skips CK C++ extension compilation if using Triton Backend
    if ROCM_BACKEND == "ck":
        ck_dir = "csrc/composable_kernel"

        #use codegen get code dispatch
        if not os.path.exists("./build"):
            os.makedirs("build")

        optdim = os.getenv("OPT_DIM", "32,64,128,256")
        archs = [arch.lower() for arch in os.getenv("GPU_ARCHS", "native").split(";")]
        validate_and_update_archs(archs)

        if archs != ["native"]:
            kernel_targets = archs
        else:
            if not torch.cuda.is_available():
                raise RuntimeError(
                    "GPU_ARCHS not set and no GPU detected. "
                    "Please set GPU_ARCHS (e.g. GPU_ARCHS='gfx942') to cross-compile."
                )
            props = torch.cuda.get_device_properties(torch.cuda.current_device())
            gcn_arch = getattr(props, "gcnArchName", None)
            if not gcn_arch:
                raise RuntimeError(
                    "GPU_ARCHS not set and current device does not expose gcnArchName. "
                    "This usually means the active PyTorch build is not ROCm. "
                    "Please set GPU_ARCHS explicitly."
                )
            detected_arch = gcn_arch.split(":")[0]
            kernel_targets = [detected_arch.lower()]
            validate_and_update_archs(kernel_targets)

        # NOTE: --targets requires CK >= 859acb5 (the submodule version pinned in this repo).
        # If generate.py fails with an unknown argument error, ensure the
        # composable_kernel submodule is up to date.
        targets_arg = ",".join(kernel_targets)
        for direction in ["fwd", "fwd_appendkv", "fwd_splitkv", "bwd"]:
            subprocess.run([sys.executable, f"{ck_dir}/example/ck_tile/01_fmha/generate.py", "-d", direction, "--output_dir", "build", "--receipt", "2", "--optdim", optdim, "--targets", targets_arg], check=True)

        # Check, if ATen/CUDAGeneratorImpl.h is found, otherwise use ATen/cuda/CUDAGeneratorImpl.h
        # See https://github.com/pytorch/pytorch/pull/70650
        generator_flag = []
        torch_dir = torch.__path__[0]
        if os.path.exists(os.path.join(torch_dir, "include", "ATen", "CUDAGeneratorImpl.h")):
            generator_flag = ["-DOLD_GENERATOR_PATH"]

        check_if_rocm_home_none("flash_attn")
        cc_flag = [f"--offload-arch={arch}" for arch in kernel_targets]

        # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
        # torch._C._GLIBCXX_USE_CXX11_ABI
        # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920
        if FORCE_CXX11_ABI:
            torch._C._GLIBCXX_USE_CXX11_ABI = True

        sources = ["csrc/flash_attn_ck/flash_api.cpp",
                "csrc/flash_attn_ck/flash_common.cpp",
                "csrc/flash_attn_ck/mha_bwd.cpp",
                "csrc/flash_attn_ck/mha_fwd_kvcache.cpp",
                "csrc/flash_attn_ck/mha_fwd.cpp",
                "csrc/flash_attn_ck/mha_varlen_bwd.cpp",
                "csrc/flash_attn_ck/mha_varlen_fwd.cpp"] + glob.glob(
            f"build/fmha_*wd*.cpp"
        )

        # Check if torch is using hipify v2. Until CK is updated with HIPIFY_V2 macro,
        # we must replace the incorrect APIs.
        maybe_hipify_v2_flag = []
        if detect_hipify_v2():
            maybe_hipify_v2_flag = ["-DHIPIFY_V2"]

        rename_cpp_to_cu(sources)

        renamed_sources = ["csrc/flash_attn_ck/flash_api.cu",
                        "csrc/flash_attn_ck/flash_common.cu",
                        "csrc/flash_attn_ck/mha_bwd.cu",
                        "csrc/flash_attn_ck/mha_fwd_kvcache.cu",
                        "csrc/flash_attn_ck/mha_fwd.cu",
                        "csrc/flash_attn_ck/mha_varlen_bwd.cu",
                        "csrc/flash_attn_ck/mha_varlen_fwd.cu"] + glob.glob(f"build/fmha_*wd*.cu")

        cc_flag += ["-O3","-std=c++20",
                    "-Wno-unknown-warning-option",
                    "-fbracket-depth=1024",
                    "-DCK_TILE_FMHA_FWD_FAST_EXP2=1",
                    "-fgpu-flush-denormals-to-zero",
                    "-DCK_ENABLE_BF16",
                    "-DCK_ENABLE_BF8",
                    "-DCK_ENABLE_FP16",
                    "-DCK_ENABLE_FP32",
                    "-DCK_ENABLE_FP64",
                    "-DCK_ENABLE_FP8",
                    "-DCK_ENABLE_INT8",
                    "-DCK_USE_XDL",
                    "-DUSE_PROF_API=1",
                    # "-DFLASHATTENTION_DISABLE_BACKWARD",
                    "-D__HIP_PLATFORM_HCC__=1"]

        ck_tile_float_to_bfloat16_default = os.environ.get("CK_TILE_FLOAT_TO_BFLOAT16_DEFAULT")
        if ck_tile_float_to_bfloat16_default is None:
            has_gfx11_target = any(arch.startswith("gfx11") for arch in kernel_targets)
            ck_tile_float_to_bfloat16_default = "0" if has_gfx11_target else "3"
        cc_flag += [f"-DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT={ck_tile_float_to_bfloat16_default}"]

        # Imitate https://github.com/ROCm/composable_kernel/blob/c8b6b64240e840a7decf76dfaa13c37da5294c4a/CMakeLists.txt#L190-L214
        hip_version = get_hip_version()
        if hip_version > Version('5.5.00000'):
            cc_flag += ["-mllvm", "--lsr-drop-solution=1"]
        if hip_version > Version('5.7.23302'):
            cc_flag += ["-fno-offload-uniform-block"]
        if hip_version > Version('6.1.40090'):
            cc_flag += ["-mllvm", "-enable-post-misched=0"]
        if hip_version > Version('6.2.41132'):
            cc_flag += ["-mllvm", "-amdgpu-early-inline-all=true",
                        "-mllvm", "-amdgpu-function-calls=false"]
        if hip_version > Version('6.2.41133') and hip_version < Version('6.3.00000'):
            cc_flag += ["-mllvm", "-amdgpu-coerce-illegal-types=1"]

        extra_compile_args = {
            "cxx": ["-O3", "-std=c++20"] + generator_flag + maybe_hipify_v2_flag,
            "nvcc": cc_flag + generator_flag + maybe_hipify_v2_flag,
        }

        include_dirs = [
            Path(this_dir) / "csrc" / "composable_kernel" / "include",
            Path(this_dir) / "csrc" / "composable_kernel" / "library" / "include",
            Path(this_dir) / "csrc" / "composable_kernel" / "example" / "ck_tile" / "01_fmha",
        ]

        ext_modules.append(
            CUDAExtension(
                name="flash_attn_2_cuda",
                sources=renamed_sources,
                extra_compile_args=extra_compile_args,
                include_dirs=include_dirs,
            )
        )


def get_package_version():
    with open(Path(this_dir) / "flash_attn" / "__init__.py", "r") as f:
        version_match = re.search(r"^__version__\s*=\s*(.*)$", f.read(), re.MULTILINE)
    public_version = ast.literal_eval(version_match.group(1))
    local_version = os.environ.get("FLASH_ATTN_LOCAL_VERSION")
    if local_version:
        return f"{public_version}+{local_version}"
    else:
        return str(public_version)


def get_wheel_url():
    torch_version_raw = parse(torch.__version__)
    python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
    platform_name = get_platform()
    flash_version = get_package_version()
    torch_version = f"{torch_version_raw.major}.{torch_version_raw.minor}"
    cxx11_abi = str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()

    if IS_ROCM:
        torch_hip_version = get_hip_version()
        hip_version = f"{torch_hip_version.major}{torch_hip_version.minor}"
        wheel_filename = f"{PACKAGE_NAME}-{flash_version}+rocm{hip_version}torch{torch_version}cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"
    else:
        # Determine the version numbers that will be used to determine the correct wheel
        # We're using the CUDA version used to build torch, not the one currently installed
        # _, cuda_version_raw = get_cuda_bare_metal_version(CUDA_HOME)
        torch_cuda_version = parse(torch.version.cuda)
        # For CUDA 11 we compile for 11.8, for CUDA 12 for 12.3, and for CUDA 13 for 13.0
        # to save CI time. Minor versions should be compatible.
        if torch_cuda_version.major == 11:
            torch_cuda_version = parse("11.8")
        elif torch_cuda_version.major == 12:
            torch_cuda_version = parse("12.3")
        else:
            torch_cuda_version = parse("13.0")
        # cuda_version = f"{cuda_version_raw.major}{cuda_version_raw.minor}"
        cuda_version = f"{torch_cuda_version.major}"

        # Determine wheel URL based on CUDA version, torch version, python version and OS
        wheel_filename = f"{PACKAGE_NAME}-{flash_version}+cu{cuda_version}torch{torch_version}cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"

    wheel_url = BASE_WHEEL_URL.format(tag_name=f"v{flash_version}", wheel_name=wheel_filename)

    return wheel_url, wheel_filename


class CachedWheelsCommand(_bdist_wheel):
    """
    The CachedWheelsCommand plugs into the default bdist wheel, which is ran by pip when it cannot
    find an existing wheel (which is currently the case for all flash attention installs). We use
    the environment parameters to detect whether there is already a pre-built version of a compatible
    wheel available and short-circuits the standard full build pipeline.
    """

    def run(self):
        if FORCE_BUILD:
            return super().run()

        wheel_url, wheel_filename = get_wheel_url()
        print("Guessing wheel URL: ", wheel_url)
        try:
            urllib.request.urlretrieve(wheel_url, wheel_filename)

            # Make the archive
            # Lifted from the root wheel processing command
            # https://github.com/pypa/wheel/blob/cf71108ff9f6ffc36978069acb28824b44ae028e/src/wheel/bdist_wheel.py#LL381C9-L381C85
            if not os.path.exists(self.dist_dir):
                os.makedirs(self.dist_dir)

            impl_tag, abi_tag, plat_tag = self.get_tag()
            archive_basename = f"{self.wheel_dist_name}-{impl_tag}-{abi_tag}-{plat_tag}"

            wheel_path = os.path.join(self.dist_dir, archive_basename + ".whl")
            print("Raw wheel path", wheel_path)
            os.rename(wheel_filename, wheel_path)
        except (urllib.error.HTTPError, urllib.error.URLError):
            print("Precompiled wheel not found. Building from source...")
            # If the wheel could not be downloaded, build from source
            super().run()


class NinjaBuildExtension(BuildExtension):
    def __init__(self, *args, **kwargs) -> None:
        # do not override env MAX_JOBS if already exists
        if not os.environ.get("MAX_JOBS"):
            import psutil

            nvcc_threads = max(1, int(NVCC_THREADS))

            # calculate the maximum allowed NUM_JOBS based on cores
            max_num_jobs_cores = max(1, os.cpu_count() // 2)

            # calculate the maximum allowed NUM_JOBS based on free memory
            free_memory_gb = psutil.virtual_memory().available / (1024 ** 3)  # free memory in GB
            # Assume worst-case peak observed memory usage of ~5GB per NVCC thread.
            # Limit: peak_threads = max_jobs * nvcc_threads and peak_threads * 5GB <= free_memory.
            max_num_jobs_memory = max(1, int(free_memory_gb / (5 * nvcc_threads)))

            # pick lower value of jobs based on cores vs memory metric to minimize oom and swap usage during compilation
            max_jobs = max(1, min(max_num_jobs_cores, max_num_jobs_memory))
            print(
                f"Auto set MAX_JOBS to `{max_jobs}`, NVCC_THREADS to `{nvcc_threads}`. "
                "If you see memory pressure, please use a lower `MAX_JOBS=N` or `NVCC_THREADS=N` value."
            )
            os.environ["MAX_JOBS"] = str(max_jobs)

        super().__init__(*args, **kwargs)

    def build_extensions(self) -> None:
        original_spawn = None
        if sys.platform == "win32" and self.compiler.compiler_type == "msvc":
            original_spawn = self.compiler.spawn

            def spawn(cmd):
                if not cmd or Path(str(cmd[0])).name.lower() != "link.exe":
                    return original_spawn(cmd)
                cmd = [str(arg) for arg in cmd]
                if len(subprocess.list2cmdline(cmd)) <= 32767:
                    return original_spawn(cmd)
                # Temporary workaround adapted from https://github.com/pypa/distutils/pull/406
                # until setuptools/distutils ships response-file handling for long MSVC links.
                with tempfile.TemporaryDirectory() as tmpdir:
                    rsp_path = Path(tmpdir) / "cmdline.txt"
                    rsp_path.write_text(
                        "\n".join(subprocess.list2cmdline([arg]) for arg in cmd[1:]) + "\n",
                        encoding="ascii",
                    )
                    return original_spawn([cmd[0], f"@{rsp_path}"])

            self.compiler.spawn = spawn

        try:
            super().build_extensions()
        finally:
            if original_spawn is not None:
                self.compiler.spawn = original_spawn


# Build install_requires based on platform
if ROCM_BACKEND == "triton":
    # Note: torch is excluded because pip resolves it to CUDA PyTorch from PyPI, overwriting any pre-installed ROCm PyTorch. Users must have torch installed.
    install_requires = [
        "einops",
        "triton>=3.6.0" if sys.platform != "win32" else "triton-windows>=3.6.0",
    ]
else:
    install_requires = [
        "torch",
        "einops",
    ]

setup(
    name=PACKAGE_NAME,
    version=get_package_version(),
    packages=find_packages(
        exclude=(
            "build",
            "csrc",
            "include",
            "tests",
            "dist",
            "docs",
            "benchmarks",
            "flash_attn.egg-info",
            "flash_attn.cute",
            "flash_attn.cute.*",
        )
    ),
    author="Tri Dao",
    author_email="tri@tridao.me",
    description="Flash Attention: Fast and Memory-Efficient Exact Attention",
    long_description=long_description,
    long_description_content_type="text/markdown",
    url="https://github.com/Dao-AILab/flash-attention",
    classifiers=[
        "Programming Language :: Python :: 3",
        "License :: OSI Approved :: BSD License",
        "Operating System :: Unix",
    ],
    ext_modules=ext_modules,
    cmdclass={"bdist_wheel": CachedWheelsCommand, "build_ext": NinjaBuildExtension}
    if ext_modules
    else {
        "bdist_wheel": CachedWheelsCommand,
    },
    python_requires=">=3.9",
    install_requires=install_requires,
    setup_requires=[
        "packaging",
        "psutil",
        "ninja",
    ],
)