Use ATen API to get the device major arch (#4901)

Summary:
Pull Request resolved: #4901

X-link: facebookresearch/FBGEMM#1928

There is some weird stuff going on, and CUDA13 build is broken in torch when we bump the pin.

I have a gut feeling there is some weirdness with the cuda headers. To fix this instead we can just try to use the more modern ATen API, as that seems ok.

Reviewed By: q10

Differential Revision: D82855103

fbshipit-source-id: 0cb5931de3f4997702da3c9b101e8cd3798ee51f

Author: cthi

fbgemm_gpu/experimental/gen_ai/src/quantize/common/include/fbgemm_gpu/quantize/utils.h

@@ -11,6 +11,8 @@
 #include <climits>
 #include <cstdint>
 
+#include <ATen/cuda/CUDAContext.h>
+
 namespace fbgemm_gpu {
 
 constexpr int64_t nextPowerOf2(int64_t num) {
@@ -19,6 +21,19 @@ constexpr int64_t nextPowerOf2(int64_t num) {
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
 
-int getDeviceArch();
+inline int getDeviceArch() {
+  static int arch = []() {
+    const int majorVersion =
+        at::cuda::getDeviceProperties(at::cuda::current_device())->major;
+    if (majorVersion >= 10) {
+      int runtimeVersion = 0;
+      C10_CUDA_CHECK(cudaRuntimeGetVersion(&runtimeVersion));
+      TORCH_CHECK(
+          runtimeVersion >= 12080, "SM100a+ kernels require cuda >= 12.8");
+    }
+    return majorVersion;
+  }();
+  return arch;
+}
 
 } // namespace fbgemm_gpu