build(cuda): NVCC_ALLOW_UNSUPPORTED, so a host compiler newer than VS2022 is reachable

c/Makefile

@@ -296,6 +296,23 @@ ifneq ($(LINUX),)
 NVCCFLAGS += -Xcompiler=-include,$(CURDIR)/glibc_c23_math_compat.h
 endif
 endif
+# CUDA 13.1 host_config.h accepts only Visual Studio 2019-2022 and refuses
+# anything newer outright, so on a box with VS Build Tools 18 (MSVC 19.50)
+# EVERY nvcc target fails at the first #include with C1189 -- cuda-test,
+# cuda-dll, all of it. nvcc offers -allow-unsupported-compiler to override the
+# check, but there was no way to get one flag in: NVCCFLAGS is the only handle
+# and replacing it wholesale discards $(CUDA_GENCODE), the -ccbin set above and
+# the -Xcompiler warning form that Windows specifically needs -- the same
+# argument that made NVCC_STD a variable rather than a rewrite.
+#
+# Defaults to 0: nvcc's own wording is "may cause compilation failure or
+# incorrect run time execution", so an unsupported host compiler is the
+# builder's decision to make explicitly, not a default this Makefile makes for
+# them. Any result produced with it set should say so.
+NVCC_ALLOW_UNSUPPORTED ?= 0
+ifeq ($(NVCC_ALLOW_UNSUPPORTED),1)
+NVCCFLAGS += -allow-unsupported-compiler
+endif
 # HIP=1 builds the SAME backend for AMD GPUs via ROCm: backend_cuda.cu is
 # compiled unchanged through backend_gpu_compat.h (one source, two vendors,
 # like compat.h does for Windows). HIP_ARCH=native targets the GPU in this

c/tests/test_cuda_test_makefile.py

@@ -15,6 +15,15 @@ def cuda_test_recipe(*variables):
     return result.stdout + result.stderr
 
 
+def nvcc_compile_line(recipe):
+    """First real nvcc invocation, skipping the `command -v "nvcc"` guard line
+    that a naive grep for "nvcc" matches first."""
+    return next(
+        (line for line in recipe.splitlines()
+         if "backend_cuda.cu" in line and " -o " in line and "command -v" not in line),
+        "")
+
+
 def mxfp4_link_line(recipe):
     return next(
         (line for line in recipe.splitlines()
@@ -53,6 +62,36 @@ def test_hipcc_does_not_link_an_openmp_runtime_for_the_oracle(self):
         self.assertTrue(line, "no MXFP4 HIP link command in dry-run recipe")
         self.assertNotIn("-fopenmp", line)
 
+    def test_allow_unsupported_compiler_is_off_by_default(self):
+        """nvcc's own wording for this flag is "may cause compilation failure or
+        incorrect run time execution", so it is never a default."""
+        line = nvcc_compile_line(cuda_test_recipe("CUDA=1", "NVCC=nvcc"))
+        self.assertTrue(line, "no nvcc compile command in dry-run recipe")
+        self.assertNotIn("-allow-unsupported-compiler", line)
+
+    def test_allow_unsupported_compiler_is_reachable(self):
+        """CUDA 13.1's host_config.h rejects any MSVC newer than VS2022, which
+        fails every nvcc target at the first #include with C1189. Before this
+        variable the only handle was NVCCFLAGS, and overriding that wholesale
+        drops the gencode, the -ccbin and the platform warning form."""
+        line = nvcc_compile_line(
+            cuda_test_recipe("CUDA=1", "NVCC=nvcc", "NVCC_ALLOW_UNSUPPORTED=1"))
+        self.assertTrue(line, "no nvcc compile command in dry-run recipe")
+        self.assertIn("-allow-unsupported-compiler", line)
+
+    def test_allow_unsupported_compiler_preserves_the_other_flags(self):
+        """The load-bearing one. Appending must not do what a wholesale
+        NVCCFLAGS override does, which is silently drop the arch and lose
+        -ftz=false -- and -ftz is a correctness flag here, not a tuning knob:
+        the fmt=8 kernels are cross-tier parity instruments and a flushed
+        subnormal diverges from the CPU reference."""
+        line = nvcc_compile_line(
+            cuda_test_recipe("CUDA=1", "NVCC=nvcc", "CUDA_ARCH=sm_86",
+                             "NVCC_ALLOW_UNSUPPORTED=1"))
+        self.assertTrue(line, "no nvcc compile command in dry-run recipe")
+        for flag in ("-ftz=false", "-std=c++17", "sm_86"):
+            self.assertIn(flag, line, f"{flag} lost when the override is set")
+
     def test_setup_openmp_probe_does_not_require_tmp(self):
         setup = (HERE / "setup.sh").read_text(encoding="utf-8")
         self.assertNotIn("/tmp/_omp", setup)