Add backtrace to NCCLXException message (#744)

Summary:
Pull Request resolved: #744

Embed a symbolized stack trace in the NCCLXException message using
folly::symbolizer::getStackTraceStr(). Since NCCLXException is never
caught within the torchcomms ncclx layer and propagates to upstream
callers, embedding the backtrace directly in the message ensures it is
always visible via e.what() without requiring the caller to use a
separate exception tracer.

Added NCCLXExceptionIncludesBacktrace unit test to verify the
backtrace is present in the exception message.

Reviewed By: mingrany

Differential Revision: D93783839

fbshipit-source-id: d8be63b2daf97503740538849654bb791c908b7d

Author: pavanbalaji

comms/torchcomms/ncclx/NcclxApi.cpp

@@ -2,6 +2,8 @@
 
 #include "comms/torchcomms/ncclx/NcclxApi.hpp"
 
+#include <folly/debugging/symbolizer/Symbolizer.h>
+
 // Check NCCL version at compile time
 #if NCCL_VERSION_CODE < NCCL_VERSION(2, 25, 0)
 #error \
@@ -19,7 +21,8 @@ NCCLXException::NCCLXException(
     ncclComm_t comm)
     : message_(
           message + ": " + nccl_api.getErrorString(result) +
-          " \nNCCL Last Error: " + nccl_api.getLastError(comm)),
+          " \nNCCL Last Error: " + nccl_api.getLastError(comm) +
+          " \nBacktrace:\n" + folly::symbolizer::getStackTraceStr()),
       result_(result) {}
 
 const char* NCCLXException::what() const noexcept {

comms/torchcomms/ncclx/tests/unit/cpp/TorchCommNCCLXTest.cpp

@@ -1484,6 +1484,28 @@ TEST_F(TorchCommNCCLXTest, NCCLXExceptionIncludesLastErrorString) {
   EXPECT_EQ(exception.getResult(), ncclInternalError);
 }
 
+TEST_F(TorchCommNCCLXTest, NCCLXExceptionIncludesBacktrace) {
+  // Test that NCCLXException message includes a backtrace
+
+  nccl_mock_->setupDefaultBehaviors();
+
+  EXPECT_CALL(*nccl_mock_, getErrorString(ncclInternalError))
+      .WillOnce(Return("internal error"));
+  EXPECT_CALL(*nccl_mock_, getLastError(_)).WillOnce(Return("some error"));
+
+  ncclComm_t mock_comm = reinterpret_cast<ncclComm_t>(0x3000);
+  NCCLXException exception(
+      *nccl_mock_, "Test operation failed", ncclInternalError, mock_comm);
+
+  std::string what_message = exception.what();
+  EXPECT_TRUE(what_message.find("Backtrace:") != std::string::npos)
+      << "Exception message should contain 'Backtrace:' label";
+  EXPECT_TRUE(
+      what_message.find("NCCLXException") != std::string::npos ||
+      what_message.find("getStackTraceStr") != std::string::npos)
+      << "Backtrace should contain a recognizable frame: " << what_message;
+}
+
 TEST_F(TorchCommNCCLXTest, NCCLXExceptionFromFailedSendIncludesLastError) {
   // Test that when send() fails, the thrown NCCLXException includes
   // the NCCL last error string