Improve error message for unsupported NCCL mem allocator

Summary:
Updated error messages in TorchCommNCCL and TorchCommNCCLX to include
the CUDART version number when the NCCL memory allocator is not
supported. This provides better diagnostic information to users,
making it clearer which CUDA runtime version is causing the
incompatibility. The error message now displays "NCCL mem allocator is
not supported in CUDART version %d" instead of the generic "not
supported in this NCCL version" message.

Reviewed By: siyengar

Differential Revision: D88564927

fbshipit-source-id: 2daf1f8bc6435ba0612af8c25ee5e17ad7a5f65f

Author: pavanbalaji

comms/torchcomms/nccl/TorchCommNCCL.cpp

@@ -1435,7 +1435,9 @@ std::shared_ptr<c10::Allocator> TorchCommNCCL::getMemAllocator() {
   c10::DeviceIndex deviceIdx = device_.index();
   if (!deviceSupportsMulticast(deviceIdx)) {
     TORCH_CHECK(
-        false, "NCCL mem allocator is not supported in this NCCL version");
+        false,
+        "NCCL mem allocator is not supported in CUDART version %d",
+        CUDART_VERSION);
   }
   static std::shared_ptr<c10::cuda::CUDACachingAllocator::CUDAAllocator>
       ncclMemAllocator =

comms/torchcomms/ncclx/TorchCommNCCLX.cpp

@@ -1776,7 +1776,9 @@ std::shared_ptr<c10::Allocator> TorchCommNCCLX::getMemAllocator() {
   c10::DeviceIndex deviceIdx = device_.index();
   if (!deviceSupportsMulticast(deviceIdx)) {
     TORCH_CHECK(
-        false, "NCCL mem allocator is not supported in this NCCL version");
+        false,
+        "NCCLX mem allocator is not supported in CUDART version %d",
+        CUDART_VERSION);
   }
   static std::shared_ptr<c10::cuda::CUDACachingAllocator::CUDAAllocator>
       ncclMemAllocator =