Update on "[PG NCCL] Add TDD, NCCL_DEBUG log"

Prints these env var setting during setup for easier debug. Differential Revision: [D44430875](https://our.internmc.facebook.com/intern/diff/D44430875/) [ghstack-poisoned]
pytorch · Apr 6, 2023 · f022842 · f022842
2 parents 3a66807 + 5715338
commit f022842
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 6 deletions.
diff --git a/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp b/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp
@@ -659,8 +659,10 @@ ProcessGroupNCCL::ProcessGroupNCCL(
 #endif
 
   init();
-  char * torch_distributed_debug = parseEnvVarString("TORCH_DISTRIBUTED_DEBUG", "OFF");
-  char * nccl_debug = parseEnvVarString("NCCL_DEBUG", "OFF");
+  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_
@@ -671,8 +673,7 @@ ProcessGroupNCCL::ProcessGroupNCCL(
             << options_->is_high_priority_stream
             << "\n TORCH_DISTRIBUTED_DEBUG: "
             << std::string(torch_distributed_debug)
-            << "\n NCCL_DEBUG: "
-            << std::string(nccl_debug);
+            << "\n NCCL_DEBUG: " << std::string(nccl_debug);
 
   RECORD_PARAM_COMMS(
       0, // seq

diff --git a/torch/csrc/distributed/c10d/Utils.hpp b/torch/csrc/distributed/c10d/Utils.hpp
@@ -98,8 +98,8 @@ inline int parseEnvVarInt(const char* envVarName) {
   return C10D_ENV_NOT_SET;
 }
 
-inline char* parseEnvVarString(const char* envVarName, char* default_val) {
-  char* val = std::getenv(envVarName);
+inline const char* parseEnvVarString(const char* envVarName, const char* default_val) {
+  const char* val = std::getenv(envVarName);
   if (val == nullptr) {
     val = default_val;
   }