Fix CUDA 12.9 OSS compilation for HSTU (#4360)

Summary:
X-link: facebookresearch/FBGEMM#1429

- Fix CUDA 12.9 OSS compilation for HSTU

Pull Request resolved: #4360

Reviewed By: spcyppt

Differential Revision: D76792311

Pulled By: q10

fbshipit-source-id: 82e0ab01f72e4e40cf5f0d4652c0c5aa3f9d2b61

Author: q10

.github/scripts/fbgemm_gpu_build.bash

@@ -272,7 +272,8 @@ __configure_fbgemm_gpu_build_cuda () {
       local arch_list="7.0"
     fi
 
-    if    [[ $cuda_version_nvcc == *"V12.8"* ]]; then
+    if    [[ $cuda_version_nvcc == *"V12.9"* ]] ||
+          [[ $cuda_version_nvcc == *"V12.8"* ]]; then
       local arch_list="${arch_list};8.0;9.0a;10.0a;12.0a"
 
     elif  [[ $cuda_version_nvcc == *"V12.6"* ]] ||
@@ -282,6 +283,7 @@ __configure_fbgemm_gpu_build_cuda () {
 
     else
       local arch_list="${arch_list};8.0;9.0"
+      echo "[BUILD] Unknown NVCC version $cuda_version_nvcc - setting TORCH_CUDA_ARCH_LIST to: ${arch_list}"
     fi
   fi
   echo "[BUILD] Setting the following CUDA targets: ${arch_list}"

cmake/modules/GpuCppLibrary.cmake

@@ -308,7 +308,7 @@ function(gpu_cpp_library)
     target_compile_options(${lib_name} PRIVATE
         ${args_CC_FLAGS}
         # Silence compiler warnings (in asmjit)
-        -Wno-deprecated-anon-enum-enum-conversion
+        -Wno-deprecated-enum-enum-conversion
         -Wno-deprecated-declarations)
 
     ############################################################################

fbgemm_gpu/codegen/training/optimizer/embedding_optimizer_split_device_kernel_template.cuh

@@ -117,7 +117,7 @@ DEVICE_INLINE void {{ mdesc }}_{{ optimizer }}_table_update_kernel(
     }
 
     {%- if not ssd %}
-    constexpr auto enable_optimizer_offloading = false;
+    [[maybe_unused]] constexpr auto enable_optimizer_offloading = false;
     {%- endif %}
 
     {{ split_precomputation }}

fbgemm_gpu/src/tbe/eeg/eeg_models.h

@@ -64,10 +64,6 @@ struct IndicesDistributionParameters {
   // Number of indices to generate
   int64_t numIndices;
 
-  // NOTE: Compiler-generated aggregate initialization constructors (P0960R3,
-  // P1975R0) did not exist prior to C++20, but FBGEMM_GPU OSS still uses C++17,
-  // namely when building against CUDA 11.8.  Remove this constructor once CUDA
-  // 11.8 is deprecated from FBGEMM_GPU support.
   IndicesDistributionParameters(
       const std::vector<double>& _1,
       const ZipfParameters& _2,