wip

Author: q10

fbgemm_gpu/src/memory_utils/memory_utils.cu

@@ -150,7 +150,7 @@ using gpuMemLocation = hipMemLocation;
 using gpuMemLocation = cudaMemLocation;
 #endif
 
-#if ROCM_VERSION >= 70000
+#if ROCM_VERSION >= 70100
 // In contrast to CUDA 13, which overrides the API signature of cudaMemAdvise,
 // ROCm 7.0, adds hipMemAdvise_v2 to maintain backwards compatibiility.
 //
@@ -203,7 +203,7 @@ Tensor new_managed_tensor(
       ptr,
       size_bytes,
       cudaMemAdviseSetPreferredLocation,
-#if CUDART_VERSION >= 13000 || ROCM_VERSION >= 70000
+#if CUDART_VERSION >= 13000 || ROCM_VERSION >= 70100
       // Starting with CUDA 13, the deviceId arg (int) is replaced with
       // cudaMemLocation (struct)
       new_mem_location_cpu()
@@ -218,7 +218,7 @@ Tensor new_managed_tensor(
       ptr,
       size_bytes,
       cudaMemAdviseSetAccessedBy,
-#if CUDART_VERSION >= 13000 || ROCM_VERSION >= 70000
+#if CUDART_VERSION >= 13000 || ROCM_VERSION >= 70100
       new_mem_location_from_device(at::cuda::current_device())
 #else
       at::cuda::current_device()
@@ -415,7 +415,7 @@ void uvm_cuda_mem_advise(const Tensor& t, int64_t cuda_memory_advise) {
       ptr,
       size_bytes,
       static_cast<enum cudaMemoryAdvise>(cuda_memory_advise),
-#if CUDART_VERSION >= 13000
+#if CUDART_VERSION >= 13000 || ROCM_VERSION >= 70100
       new_mem_location_from_device(hint_device)
 #else
       hint_device