diff --git a/setup.py b/setup.py
index 81159acbd313f3..9b460e32a307c2 100644
--- a/setup.py
+++ b/setup.py
@@ -1273,6 +1273,7 @@ def main():
         "include/torch/csrc/profiler/orchestration/*.h",
         "include/torch/csrc/profiler/stubs/*.h",
         "include/torch/csrc/profiler/unwind/*.h",
+        "include/torch/csrc/profiler/python/*.h",
         "include/torch/csrc/utils/*.h",
         "include/torch/csrc/tensor/*.h",
         "include/torch/csrc/lazy/backend/*.h",
diff --git a/torch/csrc/profiler/python/combined_traceback.h b/torch/csrc/profiler/python/combined_traceback.h
index 0dcb9a11c04acc..f71033fd88e708 100644
--- a/torch/csrc/profiler/python/combined_traceback.h
+++ b/torch/csrc/profiler/python/combined_traceback.h
@@ -11,7 +11,7 @@ namespace torch {
 // returns std::vector because one use is to call it with a batch of
 // tracebacks that come from a larger datastructure (e.g. a memory snapshot)
 // and then have more c++ code to put those objects in the right place.
-std::vector<pybind11::object> py_symbolize(
+TORCH_API std::vector<pybind11::object> py_symbolize(
     std::vector<CapturedTraceback*>& to_symbolize);
 
 // requires GIL to be held, frees any pending free frames