Build the CUDA extension with C++20 on PyTorch 2.13 and newer, whose ATen
headers require it; setup.py pins -std=c++17, which PyTorch cannot
override. Backport of upstream commits f87af617f and edb5c76ee
(Dao-AILab/flash-attention#2879, #2899), not in any 2.8.3 release.

--- a/setup.py
+++ b/setup.py
@@ -196,9 +196,17 @@
     if FORCE_CXX11_ABI:
         torch._C._GLIBCXX_USE_CXX11_ABI = True
 
+    # PyTorch 2.13+ requires C++20 for extensions that include ATen headers
+    # (ATen raises "bit-field default initializer error" for 2.13,
+    # "#error C++20 or later ... required" for 2.14+). Because we pass an
+    # explicit -std flag below, PyTorch's build machinery cannot upgrade the
+    # standard for us, so select it from the installed torch version. Older
+    # torch keeps C++17 to avoid requiring a newer toolchain unnecessarily.
+    cxx_standard = "c++20" if (TORCH_MAJOR, TORCH_MINOR) >= (2, 13) else "c++17"
+
     nvcc_flags = [
     "-O3",
-    "-std=c++17",
+    f"-std={cxx_standard}",
     "-U__CUDA_NO_HALF_OPERATORS__",
     "-U__CUDA_NO_HALF_CONVERSIONS__",
     "-U__CUDA_NO_HALF2_OPERATORS__",
@@ -217,11 +225,11 @@
     # "-DFLASHATTENTION_DISABLE_LOCAL",
     ]
 
-    compiler_c17_flag=["-O3", "-std=c++17"]
+    compiler_cxx_flag = ["-O3", f"-std={cxx_standard}"]
     # 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"]
+        compiler_cxx_flag = ["-O2", f"/std:{cxx_standard}", "/Zc:__cplusplus"]
 
     ext_modules.append(
         CUDAExtension(
@@ -302,7 +310,7 @@
                 "csrc/flash_attn/src/flash_fwd_split_hdim256_bf16_causal_sm80.cu",
             ],
             extra_compile_args={
-                "cxx": compiler_c17_flag,
+                "cxx": compiler_cxx_flag,
                 "nvcc": append_nvcc_threads(nvcc_flags + cc_flag),
             },
             include_dirs=[