[Profiler][Minor] Separate standalone profilers from the main PyTorch profiler. (#85511)

There are a number of instrumentation utils which have been added to the profiler toolkit. They are generally small and self contained, often wrapping vendor APIs. (NVTX, ITT)

They don't really interact with the much more expansive machinery of the PyTorch profiler beyond registration / unregistration, minor util sharing, and reusing the profiler base class. Just as in the case of stubs, it makes sense to group them in a dedicated subfolder.

Differential Revision: [D39108649](https://our.internmc.facebook.com/intern/diff/D39108649/)

**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D39108649/)!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85511
Approved by: https://github.com/albanD

build_variables.bzl

@@ -134,13 +134,13 @@ libtorch_profiler_sources = [
     "torch/csrc/autograd/profiler_legacy.cpp",
     "torch/csrc/autograd/profiler_kineto.cpp",
     "torch/csrc/profiler/collection.cpp",
-    "torch/csrc/profiler/execution_graph_observer.cpp",
     "torch/csrc/profiler/kineto_shim.cpp",
-    "torch/csrc/profiler/nvtx_observer.cpp",
     "torch/csrc/profiler/kineto_client_interface.cpp",
-    "torch/csrc/profiler/itt_observer.cpp",
     "torch/csrc/profiler/orchestration/observer.cpp",
     "torch/csrc/profiler/orchestration/python_tracer.cpp",
+    "torch/csrc/profiler/standalone/execution_graph_observer.cpp",
+    "torch/csrc/profiler/standalone/itt_observer.cpp",
+    "torch/csrc/profiler/standalone/nvtx_observer.cpp",
     "torch/csrc/profiler/stubs/base.cpp",
     "torch/csrc/monitor/counters.cpp",
     "torch/csrc/monitor/events.cpp",

torch/_C/_autograd.pyi

@@ -68,10 +68,6 @@ def _profiler_enabled() -> bool: ...
 def _add_metadata_json(key: str, value: str) -> None: ...
 def _kineto_step() -> None: ...
 def kineto_available() -> bool: ...
-def _add_execution_graph_observer(output_file_path: str) -> bool: ...
-def _remove_execution_graph_observer() -> None: ...
-def _enable_execution_graph_observer() -> None: ...
-def _disable_execution_graph_observer() -> None: ...
 def _record_function_with_args_enter(name: str, args: List[Any]) -> torch.Tensor: ...
 def _record_function_with_args_exit(handle: torch.Tensor) -> None: ...
 def _supported_activities() -> Set[ProfilerActivity]: ...

torch/_C/_profiler.pyi

@@ -159,3 +159,8 @@ class _ExtraFields_PyCall:
     caller: _PyFrameState
 
 class _ExtraFields_Kineto: ...
+
+def _add_execution_graph_observer(output_file_path: str) -> bool: ...
+def _remove_execution_graph_observer() -> None: ...
+def _enable_execution_graph_observer() -> None: ...
+def _disable_execution_graph_observer() -> None: ...

torch/csrc/autograd/init.cpp

@@ -24,7 +24,6 @@
 #include <torch/csrc/autograd/utils/wrap_outputs.h>
 #include <torch/csrc/jit/python/pybind_utils.h>
 #include <torch/csrc/profiler/collection.h>
-#include <torch/csrc/profiler/execution_graph_observer.h>
 #include <torch/csrc/profiler/kineto_shim.h>
 #include <torch/csrc/utils/disable_torch_function.h>
 #include <torch/csrc/utils/pybind.h>
@@ -238,21 +237,6 @@ PyObject* THPAutograd_initExtension(PyObject* _unused, PyObject* unused) {
   m.def("_kineto_step", profilerStep); // Only if `USE_KINETO` is set
   m.def("kineto_available", []() { return torch::profiler::kKinetoAvailable; });
 
-  // PyTorch profiler execution graph internal interface.
-  m.def(
-      "_add_execution_graph_observer",
-      &torch::profiler::impl::addExecutionGraphObserver,
-      py::arg("output_file_name"));
-  m.def(
-      "_remove_execution_graph_observer",
-      &torch::profiler::impl::removeExecutionGraphObserver);
-  m.def(
-      "_enable_execution_graph_observer",
-      &torch::profiler::impl::enableExecutionGraphObserver);
-  m.def(
-      "_disable_execution_graph_observer",
-      &torch::profiler::impl::disableExecutionGraphObserver);
-
   // NOTICE: These record functions are not torch operators and may not show up
   // in TorchScript tracing, FX transforms, or operator serialization. For these
   // use cases, please use `torch.profiler.record_function`.

torch/csrc/autograd/profiler_kineto.cpp

@@ -12,10 +12,10 @@
 #include <torch/csrc/profiler/api.h>
 #include <torch/csrc/profiler/collection.h>
 #include <torch/csrc/profiler/containers.h>
-#include <torch/csrc/profiler/itt_observer.h>
 #include <torch/csrc/profiler/kineto_shim.h>
-#include <torch/csrc/profiler/nvtx_observer.h>
 #include <torch/csrc/profiler/orchestration/observer.h>
+#include <torch/csrc/profiler/standalone/itt_observer.h>
+#include <torch/csrc/profiler/standalone/nvtx_observer.h>
 #include <torch/csrc/profiler/util.h>
 
 #include <ATen/Context.h>

torch/csrc/profiler/python/init.cpp

@@ -5,6 +5,7 @@
 #include <torch/csrc/autograd/utils/wrap_outputs.h>
 #include <torch/csrc/jit/python/pybind_utils.h>
 #include <torch/csrc/profiler/collection.h>
+#include <torch/csrc/profiler/standalone/execution_graph_observer.h>
 #include <torch/csrc/utils/pybind.h>
 
 namespace torch {
@@ -251,6 +252,21 @@ void initPythonBindings(PyObject* module) {
       .def_property_readonly("duration_time_ns", [](const Result& r) {
         return r.endTimeNS() - r.start_time_ns_;
       });
+
+  // PyTorch profiler execution graph internal interface.
+  m.def(
+      "_add_execution_graph_observer",
+      &torch::profiler::impl::addExecutionGraphObserver,
+      py::arg("output_file_name"));
+  m.def(
+      "_remove_execution_graph_observer",
+      &torch::profiler::impl::removeExecutionGraphObserver);
+  m.def(
+      "_enable_execution_graph_observer",
+      &torch::profiler::impl::enableExecutionGraphObserver);
+  m.def(
+      "_disable_execution_graph_observer",
+      &torch::profiler::impl::disableExecutionGraphObserver);
 }
 
 } // namespace profiler

torch/csrc/profiler/standalone/execution_graph_observer.cpp

@@ -27,7 +27,7 @@
 #include <ATen/core/stack.h>
 #include <ATen/record_function.h>
 #include <c10/util/irange.h>
-#include <torch/csrc/profiler/execution_graph_observer.h>
+#include <torch/csrc/profiler/standalone/execution_graph_observer.h>
 #include <torch/csrc/profiler/util.h>
 
 using namespace at;

torch/csrc/profiler/standalone/itt_observer.cpp

@@ -1,4 +1,4 @@
-#include <torch/csrc/profiler/itt_observer.h>
+#include <torch/csrc/profiler/standalone/itt_observer.h>
 
 #include <torch/csrc/profiler/stubs/base.h>
 #include <torch/csrc/profiler/util.h>

torch/csrc/profiler/standalone/nvtx_observer.cpp

@@ -1,4 +1,4 @@
-#include <torch/csrc/profiler/nvtx_observer.h>
+#include <torch/csrc/profiler/standalone/nvtx_observer.h>
 
 #include <torch/csrc/profiler/stubs/base.h>
 #include <torch/csrc/profiler/util.h>

torch/profiler/profiler.py

@@ -9,13 +9,13 @@ from warnings import warn
 
 import torch
 import torch.autograd.profiler as prof
-from torch._C._autograd import (
+from torch._C._profiler import (
     _add_execution_graph_observer,
-    _remove_execution_graph_observer,
-    _enable_execution_graph_observer,
     _disable_execution_graph_observer,
+    _enable_execution_graph_observer,
+    _ExperimentalConfig,
+    _remove_execution_graph_observer,
 )
-from torch._C._profiler import _ExperimentalConfig
 from torch.autograd import ProfilerActivity, kineto_available
 
 __all__ = ['supported_activities', 'ProfilerAction', 'schedule', 'tensorboard_trace_handler', 'profile',