Skip to content

Graduate debug handle in torchao #2452

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 27, 2025
Merged

Graduate debug handle in torchao #2452

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
123 changes: 70 additions & 53 deletions test/quantization/pt2e/test_numeric_debugger.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,10 +37,12 @@ def test_simple(self):
example_inputs = m.example_inputs()
ep = export_for_training(m, example_inputs, strict=True)
m = ep.module()
self._assert_each_node_has_debug_handle(m)
debug_handle_map = self._extract_debug_handles(m)
self._assert_each_node_has_from_node_source(m)
from_node_source_map = self._extract_from_node_source(m)

self.assertEqual(len(set(debug_handle_map.values())), len(debug_handle_map))
self.assertEqual(
len(set(from_node_source_map.values())), len(from_node_source_map)
)

@unittest.skip("debug flow not working on model with conditional control flow")
def test_control_flow(self):
Expand All @@ -49,37 +51,42 @@ def test_control_flow(self):
ep = export_for_training(m, example_inputs, strict=True)
m = ep.module()

self._assert_each_node_has_debug_handle(m)
debug_handle_map = self._extract_debug_handles(m)
self._assert_each_node_has_from_node_source(m)
from_node_source_map = self._extract_from_node_source(m)

self.assertEqual(len(set(debug_handle_map.values())), len(debug_handle_map))
self.assertEqual(
len(set(from_node_source_map.values())), len(from_node_source_map)
)

def test_copy_preserve_handle(self):
m = TestHelperModules.Conv2dThenConv1d()
example_inputs = m.example_inputs()
ep = torch.export.export(m, example_inputs, strict=True)
m = ep.module()

self._assert_each_node_has_debug_handle(m)
debug_handle_map_ref = self._extract_debug_handles(m)
self._assert_each_node_has_from_node_source(m)
from_node_source_map_ref = self._extract_from_node_source(m)

ep_copy = copy.copy(ep)
debug_handle_map = self._extract_debug_handles(ep_copy.module())
from_node_source_map = self._extract_from_node_source(ep_copy.module())

self._assert_each_node_has_debug_handle(ep)
self.assertEqual(debug_handle_map, debug_handle_map_ref)
self._assert_each_node_has_from_node_source(ep)
self.assertEqual(from_node_source_map, from_node_source_map_ref)

def test_deepcopy_preserve_handle(self):
m = TestHelperModules.Conv2dThenConv1d()
example_inputs = m.example_inputs()
ep = torch.export.export(m, example_inputs, strict=True)

debug_handle_map_ref = self._extract_debug_handles(ep.module())
from_node_source_map_ref = self._extract_from_node_source(ep.module())
ep_copy = copy.deepcopy(ep)
debug_handle_map = self._extract_debug_handles(ep_copy.module())
from_node_source_map = self._extract_from_node_source(ep_copy.module())

self._assert_each_node_has_debug_handle(ep.module())
self.assertEqual(debug_handle_map, debug_handle_map_ref)
self._assert_each_node_has_from_node_source(ep.module())
self.assertEqual(from_node_source_map, from_node_source_map_ref)
self.assertEqual(
set(from_node_source_map.values()), set(from_node_source_map_ref.values())
)

@unittest.skip(
"torch._dynamo.exc.FailOnRecompileLimitHit: recompile_limit reached with one_graph=True. Excessive recompilations can degrade performance due to the compilation overhead of each recompilation. To monitor recom..."
Expand All @@ -90,16 +97,16 @@ def test_re_export_preserve_handle(self):
ep = export_for_training(m, example_inputs, strict=True)
m = ep.module()

self._assert_each_node_has_debug_handle(m)
debug_handle_map_ref = self._extract_debug_handles(m)
self._assert_each_node_has_from_node_source(m)
from_node_source_map_ref = self._extract_from_node_source(m)

ep_reexport = export_for_training(m, example_inputs, strict=True)
m_reexport = ep_reexport.module()

self._assert_each_node_has_debug_handle(m_reexport)
debug_handle_map = self._extract_debug_handles(m_reexport)
self._assert_each_node_has_from_node_source(m_reexport)
from_node_source_map = self._extract_from_node_source(m_reexport)

self.assertEqual(debug_handle_map, debug_handle_map_ref)
self.assertEqual(from_node_source_map, from_node_source_map_ref)

@unittest.skip(
"torch._dynamo.exc.FailOnRecompileLimitHit: recompile_limit reached with one_graph=True. Excessive recompilations can degrade performance due to the compilation overhead of each recompilation. To monitor recom..."
Expand All @@ -110,19 +117,19 @@ def test_run_decompositions_same_handle_id(self):
ep = export_for_training(m, example_inputs, strict=True)
m = ep.module()

self._assert_each_node_has_debug_handle(m)
debug_handle_map_ref = self._extract_debug_handles(m)
self._assert_each_node_has_from_node_source(m)
from_node_source_map_ref = self._extract_from_node_source(m)

ep_copy = copy.copy(ep)
ep_copy = ep_copy.run_decompositions()
m_decomposed = ep_copy.module()

self._assert_each_node_has_debug_handle(m_decomposed)
debug_handle_map = self._extract_debug_handles(m_decomposed)
self._assert_each_node_has_from_node_source(m_decomposed)
from_node_source_map = self._extract_from_node_source(m_decomposed)

# checking the map still has the same ids, the node may change
self.assertEqual(
set(debug_handle_map.values()), set(debug_handle_map_ref.values())
set(from_node_source_map.values()), set(from_node_source_map_ref.values())
)

@unittest.skip(
Expand All @@ -139,22 +146,23 @@ def test_run_decompositions_map_handle_to_new_nodes(self):
ep = export_for_training(m, example_inputs, strict=True)
m = ep.module()

self._assert_each_node_has_debug_handle(m)
pre_decomp_to_debug_handle_map_ref = (
self._extract_debug_handles_with_prev_decomp_op(m)
self._assert_each_node_has_from_node_source(m)
pre_decomp_to_from_node_source_map_ref = (
self._extract_from_node_source_with_prev_decomp_op(m)
)

ep_copy = copy.copy(ep)
ep_copy = ep_copy.run_decompositions()
m_decomposed = ep_copy.module()
self._assert_each_node_has_debug_handle(m_decomposed)
pre_decomp_to_debug_handle_map = (
self._extract_debug_handles_with_prev_decomp_op(m_decomposed)
self._assert_each_node_has_from_node_source(m_decomposed)
pre_decomp_to_from_node_source_map = (
self._extract_from_node_source_with_prev_decomp_op(m_decomposed)
)

# checking the map still has the same ids, the node may change
# checking the map still has the same infos, the node may change
self.assertEqual(
pre_decomp_to_debug_handle_map, pre_decomp_to_debug_handle_map_ref
pre_decomp_to_from_node_source_map,
pre_decomp_to_from_node_source_map_ref,
)

def test_prepare_for_propagation_comparison(self):
Expand All @@ -178,18 +186,18 @@ def test_added_node_gets_unique_id(self) -> None:
example_inputs = m.example_inputs()
ep = export_for_training(m, example_inputs, strict=True)

ref_handles = self._extract_debug_handles(ep.module())
ref_counter = Counter(ref_handles.values())
ref_from_node_source = self._extract_from_node_source(ep.module())
ref_counter = Counter(ref_from_node_source.values())

for k, v in ref_counter.items():
self.assertEqual(
v,
1,
msg=f"For handle {k}, there were {v} nodes with that handle, but expected only 1",
msg=f"For from_node info {k}, there were {v} nodes with that info, but expected only 1",
)

# Now that we have unique ids, add a new node into the graph and re-generate
# to make sure that the new node gets a unique id.
# Now that we have unique infos, add a new node into the graph and re-generate
# to make sure that the new node gets a unique info.
last_node = next(iter(reversed(ep.graph.nodes)))
with ep.graph.inserting_before(last_node):
arg = last_node.args[0]
Expand All @@ -200,30 +208,39 @@ def test_added_node_gets_unique_id(self) -> None:
arg.replace_all_uses_with(n, lambda x: x != n)
ep.graph_module.recompile()

# Regenerate handles, make sure only the new relu node has a new id, and
# it doesn't clash with any of the existing ids.
# Regenerate from_node info, make sure only the new relu node has a new info, and
# it doesn't clash with any of the existing infos.

m = ep.module()
self._assert_each_node_has_debug_handle(m)
handles_after_modification = self._extract_debug_handles(m)
handles_counter = Counter(handles_after_modification.values())
for name, handle in ref_handles.items():
self.assertIn(name, handles_after_modification)
# Check that handle was unchanged.
self.assertEqual(handles_after_modification[name], handle)
self._assert_each_node_has_from_node_source(m)
from_node_source_after_modification = self._extract_from_node_source(m)
from_node_source_counter = Counter(from_node_source_after_modification.values())
for name, from_node_source in ref_from_node_source.items():
self.assertIn(name, from_node_source_after_modification)
# Check that from_node info was unchanged.
self.assertEqual(
from_node_source_after_modification[name], from_node_source
)
# Check that total count was unchanged.
ref_count = ref_counter[handle]
after_count = handles_counter[handle]
ref_count = ref_counter[from_node_source]
after_count = from_node_source_counter[from_node_source]
self.assertEqual(
after_count,
ref_count,
msg=f"For handle {handle}, there were {after_count} nodes with that handle, but expected only {ref_count}",
msg=f"For from_node info {from_node_source}, there were {after_count} nodes with that info, but expected only {ref_count}",
)

# Check for relu specifically. Avoid hardcoding the handle id since it
# Check for relu specifically. Avoid hardcoding the from_node info since it
# may change with future node ordering changes.
self.assertNotIn(handles_after_modification["relu_default"], ref_counter)
self.assertEqual(handles_counter[handles_after_modification["relu_default"]], 1)
self.assertNotIn(
from_node_source_after_modification["relu_default"], ref_counter
)
self.assertEqual(
from_node_source_counter[
from_node_source_after_modification["relu_default"]
],
1,
)


if __name__ == "__main__":
Expand Down
Loading
Loading