[PG NCCL] Add TDD, NCCL_DEBUG log (#97692)

Prints these env var setting during setup for easier debug.

Differential Revision: [D44430875](https://our.internmc.facebook.com/intern/diff/D44430875/)
Pull Request resolved: #97692
Approved by: https://github.com/kumpera

torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp

@@ -659,14 +659,21 @@ ProcessGroupNCCL::ProcessGroupNCCL(
 #endif
 
   init();
+  const std::string OFF = "OFF";
+  const char* torch_distributed_debug =
+      parseEnvVarString("TORCH_DISTRIBUTED_DEBUG", OFF.c_str());
+  const char* nccl_debug = parseEnvVarString("NCCL_DEBUG", OFF.c_str());
   LOG(INFO) << "[Rank " << rank_
             << "] ProcessGroupNCCL initialized with following options:"
             << "\nNCCL_ASYNC_ERROR_HANDLING: " << asyncErrorHandling_
             << "\nNCCL_DESYNC_DEBUG: " << desyncDebug_
             << "\nNCCL_BLOCKING_WAIT: " << blockingWait_
             << "\nTIMEOUT(ms): " << options_->timeout.count()
             << "\nUSE_HIGH_PRIORITY_STREAM: "
-            << options_->is_high_priority_stream;
+            << options_->is_high_priority_stream
+            << "\n TORCH_DISTRIBUTED_DEBUG: "
+            << std::string(torch_distributed_debug)
+            << "\n NCCL_DEBUG: " << std::string(nccl_debug);
 
   RECORD_PARAM_COMMS(
       0, // seq

torch/csrc/distributed/c10d/Utils.hpp

@@ -98,6 +98,14 @@ inline int parseEnvVarInt(const char* envVarName) {
   return C10D_ENV_NOT_SET;
 }
 
+inline const char* parseEnvVarString(const char* envVarName, const char* default_val) {
+  const char* val = std::getenv(envVarName);
+  if (val == nullptr) {
+    val = default_val;
+  }
+  return val;
+}
+
 inline int parseEnvVarIntDefault(const char* envVarName, int defaultVal) {
     int val = parseEnvVarInt(envVarName);
     if (val == C10D_ENV_NOT_SET)