Fix Windows builds with CUDA 13 and PyTorch 2.15 (#2872)

* Add opt-in SM121 and fix Windows CUDA 13 builds

Add explicit SM121 code generation while keeping it out of the default release targets. Select forward-compatible PTX from architectures supported by the installed toolkit and reject unsupported-only requests. Enable the conforming MSVC preprocessor and math constants on Windows.

For C++20 builds on Windows, pass /permissive to both cl and nvcc's host compiler. /std:c++20 implies /permissive-, and MSVC 2022 (<19.50) then rejects CuTe templates (C3545 in cute/stride.hpp) with CUDA 13.x. This relaxes conformance checking for host code in that configuration only.

* Update comments on C++20 and nvcc flags

Clarify comments regarding MSVC 2022 and CUDA 13.x compatibility.

setup.py

@@ -105,17 +105,22 @@ 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)
+      - sm_121 on CUDA >= 12.9
+      - Use family-specific virtual architectures on CUDA >= 12.9
       - Map requested 110 -> 101 if CUDA < 13.0 (Thor rename)
-      - Embed PTX for newest arch for forward compatibility
+      - Embed PTX for the newest supported arch for forward compatibility
     """
+    supported_archs = []
+
     # Always-regular 80
     if "80" in archs:
         cc_flag += ["-gencode", "arch=compute_80,code=sm_80"]
+        supported_archs.append("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"]
+        supported_archs.append("90")
 
     # Blackwell 10.x requires >= 12.8
     if bare_metal_version >= Version("12.8"):
@@ -125,6 +130,7 @@ def add_cuda_gencodes(cc_flag, archs, bare_metal_version):
                 cc_flag += ["-gencode", "arch=compute_100f,code=sm_100"]
             else:
                 cc_flag += ["-gencode", "arch=compute_100,code=sm_100"]
+            supported_archs.append("100")
 
         if "120" in archs:
             # sm_120 is supported in CUDA 12.8/12.9+ toolkits
@@ -132,23 +138,33 @@ def add_cuda_gencodes(cc_flag, archs, bare_metal_version):
                 cc_flag += ["-gencode", "arch=compute_120f,code=sm_120"]
             else:
                 cc_flag += ["-gencode", "arch=compute_120,code=sm_120"]
+            supported_archs.append("120")
 
+        if bare_metal_version >= Version("12.9") and "121" in archs:
+            cc_flag += ["-gencode", "arch=compute_121f,code=sm_121"]
+            supported_archs.append("121")
 
         # 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"]
+                supported_archs.append("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"]
+                    supported_archs.append("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}"]
+    if not supported_archs:
+        raise RuntimeError(
+            f"None of the requested CUDA architectures {sorted(archs)} "
+            f"are supported by CUDA {bare_metal_version}"
+        )
+
+    # PTX for newest supported requested arch (forward-compat)
+    newest = max(supported_archs, key=int)
+    cc_flag += ["-gencode", f"arch=compute_{newest},code=compute_{newest}"]
 
     return cc_flag
 
@@ -362,8 +378,24 @@ if not SKIP_CUDA_BUILD and not IS_ROCM:
     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_cxx_flag = ["-O2", f"/std:{cxx_standard}", "/Zc:__cplusplus"]
+        nvcc_flags.extend([
+            "-Xcompiler", "/Zc:__cplusplus",
+            "-Xcompiler", "/Zc:preprocessor",
+            "-D_USE_MATH_DEFINES",
+        ])
+        compiler_cxx_flag = [
+            "-O2",
+            f"/std:{cxx_standard}",
+            "/Zc:__cplusplus",
+            "/Zc:preprocessor",
+            "/D_USE_MATH_DEFINES",
+        ]
+        # /std:c++20 implies /permissive-. MSVC 2022 (<19.50) then rejects CuTe
+        # templates (C3545 in cute/stride.hpp) with CUDA 13.x, so relax conformance
+        # for nvcc's host compilation only. Do not add it to the cl flags: torch's
+        # C++20 headers fail to compile under /permissive (C2666/C2139).
+        if cxx_standard == "c++20":
+            nvcc_flags += ["-Xcompiler", "/permissive"]
 
     # Opt-in: disable building dropout and its dependent headers (ATen philox/RNG
     # headers) from the FA2 build. This flag must be shared across both cxx and nvcc