2025-09-21 nightly release (3cefe05)

Author: pytorchbot

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

@@ -19,6 +19,4 @@ constexpr int64_t nextPowerOf2(int64_t num) {
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
 
-int getDeviceArch();
-
 } // namespace fbgemm_gpu

fbgemm_gpu/experimental/gen_ai/src/quantize/common/utils.cpp renamed to fbgemm_gpu/experimental/gen_ai/src/quantize/common/include/fbgemm_gpu/quantize/utils_gpu.h

@@ -6,29 +6,25 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include "fbgemm_gpu/quantize/utils.h" // @manual
+#pragma once
 
-#include <ATen/ATen.h>
-#include <c10/cuda/CUDAException.h>
-#include <cuda_runtime.h>
+#include <ATen/cuda/CUDAContext.h>
 
 namespace fbgemm_gpu {
 
-int getDeviceArch() {
+inline int getDeviceArch() {
   static int arch = []() {
-    // Avoid expensive cudaGetDeviceProperties call.
-    cudaDeviceProp prop;
-    cudaGetDeviceProperties(&prop, 0);
-
-    if (prop.major >= 10) {
+    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 prop.major;
+    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 {