Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
2 changes: 2 additions & 0 deletions code/evaluation/failure_analysis.py
Original file line number Diff line number Diff line change
Expand Up @@ -698,6 +698,8 @@ def _setup_trt_for_ablation(model, cfg, dist, device, basis, D, half, stim_dets)
Returns ``(trt_context, onnx_workflow)`` where *trt_context* is either a
``(execution_context, engine)`` pair or ``None`` (PyTorch fallback).
"""
import time as _time

trt_context = None
onnx_workflow = OnnxWorkflow.TORCH_ONLY
try:
Expand Down
21 changes: 20 additions & 1 deletion code/tests/test_failure_analysis.py
Original file line number Diff line number Diff line change
Expand Up @@ -992,7 +992,8 @@ def get_tensor_shape(self, name):
return (B_holder[0], 1 + num_detectors)

def execute_v2(self, bindings):
pass # output tensor remains zeroed — valid binary values
# Output tensor remains zeroed — valid binary values.
mock_trt.execute_v2_calls += 1

@property
def _engine(self):
Expand All @@ -1014,6 +1015,9 @@ def deserialize_cuda_engine(self, data):
class _Logger:
WARNING = 1

def __init__(self, severity=None):
pass

class _BuilderFlag:
FP16 = 0

Expand Down Expand Up @@ -1060,6 +1064,7 @@ def build_serialized_network(self, network, config):
return b""

mock_trt = types.ModuleType("tensorrt")
mock_trt.execute_v2_calls = 0
mock_trt.Logger = _Logger
mock_trt.Runtime = lambda logger: _Runtime()
mock_trt.Builder = lambda logger: _Builder()
Expand Down Expand Up @@ -1384,8 +1389,16 @@ def _run_once(cls, basis="X"):
_patch_tensor_to_for_mock_cuda():
mf.create_datapipe_inference.return_value = real_ds
result = decoder_ablation_study(_ZeroModel(), mock_device, _DummyDist(), cfg)
cls._execute_v2_calls = getattr(cls, "_execute_v2_calls", {})
cls._execute_v2_calls[basis] = mock_trt.execute_v2_calls
return result

def test_trt_path_runs_the_engine(self):
# A failed engine load falls back to PyTorch and still returns a
# well-formed result, so check that the TRT context was used.
self.assertGreater(self._execute_v2_calls["X"], 0)
self.assertGreater(self._execute_v2_calls["Z"], 0)

def test_trt_path_returns_correct_sample_count(self):
self.assertEqual(self._result_x["total_samples"], self._N)

Expand Down Expand Up @@ -1473,8 +1486,14 @@ def _fake_onnx_export(module, *args, **kwargs):
_patch_tensor_to_for_mock_cuda():
mf.create_datapipe_inference.return_value = real_ds
result = decoder_ablation_study(_ZeroModel(), mock_device, _DummyDist(), cfg)
cls._execute_v2_calls = mock_trt.execute_v2_calls
return result

def test_export_and_build_runs_the_engine(self):
# A failed engine build falls back to PyTorch and still returns a
# well-formed result, so check that the TRT context was used.
self.assertGreater(self._execute_v2_calls, 0)

def test_export_and_build_returns_correct_sample_count(self):
self.assertEqual(self._result["total_samples"], self._N)

Expand Down