Fix ROCm build I broke (#4902)

Summary:
X-link: facebookresearch/FBGEMM#1929

Pull Request resolved: #4902

Broke the ROCm build by accident in D82855103 :(.

Since the function is cuda specific, we can just only include it when being used for cuda for now. For some reason it cannot build properly standalone still in fbcode, which Is why I originally tried to move it to inline header, but that broke OSS since hipify would not get run.

Reviewed By: q10

Differential Revision: D82895451

fbshipit-source-id: 41553cf51c8a93f72f891e206de97ff00fd108dc

Author: cthi

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

@@ -11,8 +11,6 @@
 #include <climits>
 #include <cstdint>
 
-#include <ATen/cuda/CUDAContext.h>
-
 namespace fbgemm_gpu {
 
 constexpr int64_t nextPowerOf2(int64_t num) {
@@ -21,19 +19,4 @@ constexpr int64_t nextPowerOf2(int64_t num) {
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
 
-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

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

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <ATen/cuda/CUDAContext.h>
+
+namespace fbgemm_gpu {
+
+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

fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/bf16bf16bf16_grouped.cu

@@ -12,6 +12,7 @@
 #include "bf16bf16bf16_grouped/bf16bf16bf16_grouped_manifest.cuh"
 #include "fbgemm_gpu/quantize/tuning_cache.hpp"
 #include "fbgemm_gpu/quantize/utils.h"
+#include "fbgemm_gpu/quantize/utils_gpu.h"
 
 namespace fbgemm_gpu {
 

fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/f8f8bf16_groupwise.cu

@@ -13,6 +13,7 @@
 #include "f8f8bf16_groupwise/f8f8bf16_groupwise_manifest.cuh"
 #include "fbgemm_gpu/quantize/tuning_cache.hpp"
 #include "fbgemm_gpu/quantize/utils.h"
+#include "fbgemm_gpu/quantize/utils_gpu.h"
 
 namespace fbgemm_gpu {
 

fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/f8f8bf16_rowwise_batched.cu

@@ -11,6 +11,7 @@
 #include "f8f8bf16_rowwise_batched/f8f8bf16_rowwise_batched_manifest.cuh"
 
 #include "fbgemm_gpu/quantize/utils.h"
+#include "fbgemm_gpu/quantize/utils_gpu.h"
 
 namespace fbgemm_gpu {
 

fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/f8f8bf16_rowwise_grouped.cu

@@ -14,6 +14,7 @@
 #include "f8f8bf16_rowwise_grouped_sm100/f8f8bf16_rowwise_grouped_manifest.cuh"
 #include "fbgemm_gpu/quantize/tuning_cache.hpp"
 #include "fbgemm_gpu/quantize/utils.h"
+#include "fbgemm_gpu/quantize/utils_gpu.h"
 
 namespace fbgemm_gpu {