Fix pre-existing CI failures on macOS Intel and model download tests

.github/workflows/ci.yml

@@ -10,11 +10,22 @@ on:
   # Allows you to run this workflow manually from the Actions tab
   workflow_dispatch:
 
+# Cancel a running PR check when new commits are pushed, so a wedged run does not
+# keep burning runner hours after it has been superseded. Pushes to main still run
+# to completion.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 jobs:
   run-tests:
     runs-on: ${{ matrix.os }}
     continue-on-error: false
-    timeout-minutes: 360
+    # Backstop for a wedged run: the fork/TensorFlow deadlock hangs in native code
+    # holding the GIL, which pytest-timeout cannot interrupt in either method, so the
+    # job-level limit is what stops it. Slowest healthy job is ~32 min; 50 leaves
+    # headroom while capping a hang's cost.
+    timeout-minutes: 50
 
     strategy:
       fail-fast: false

pyproject.toml

@@ -174,6 +174,11 @@ namespaces = true
 log_cli = true
 log_level = "DEBUG"
 timeout = 300
+# The signal method cannot interrupt a thread blocked in a C call, so a deadlocked
+# fork child let the run hang until the job limit killed it hours later. The thread
+# method kills the process instead, turning such hangs into a failure with a stack
+# dump. Windows already uses it, since it has no SIGALRM.
+timeout_method = "thread"
 addopts = ["--max-worker-restart=0"]
 testpaths = ["src/birdnet_tests"]
 norecursedirs = ["src/birdnet_v1_tests"]

src/birdnet_tests/acoustic_models/inference/encoding/encoding_result_py/test_encoding_to_structured_array.py

@@ -17,6 +17,12 @@
 
 DEFAULT_EMBEDDING_DIM = 6
 
+# start_time/end_time are stored as float32 and, on numpy 1.x, the intermediate
+# products are computed in float32 as well (numpy 2 keeps them float64 via NEP 50).
+# Speeds that are not exactly representable therefore drift by ~1 ULP, which exceeds
+# the assert_allclose default rtol of 1e-7 - that is below float32 eps (~1.19e-7).
+_FLOAT32_TIME_RTOL = 1e-6
+
 
 def create_mock_encoding_tensor(
   n_files: int,
@@ -268,8 +274,12 @@ def test_time_calculations_speedup_one_tenth_no_overlap() -> None:
   expected_starts = np.arange(len(structured)) * hop
   expected_ends = np.minimum(expected_starts + 3.0 * 0.1, result.input_durations[0])
 
-  np.testing.assert_allclose(structured["start_time"], expected_starts)
-  np.testing.assert_allclose(structured["end_time"], expected_ends)
+  np.testing.assert_allclose(
+    structured["start_time"], expected_starts, rtol=_FLOAT32_TIME_RTOL
+  )
+  np.testing.assert_allclose(
+    structured["end_time"], expected_ends, rtol=_FLOAT32_TIME_RTOL
+  )
 
 
 def test_time_calculations_speedup_decimal_no_overlap() -> None:
@@ -289,8 +299,12 @@ def test_time_calculations_speedup_decimal_no_overlap() -> None:
     expected_starts + 3.0 * 0.1387434856, result.input_durations[0]
   )
 
-  np.testing.assert_allclose(structured["start_time"], expected_starts)
-  np.testing.assert_allclose(structured["end_time"], expected_ends)
+  np.testing.assert_allclose(
+    structured["start_time"], expected_starts, rtol=_FLOAT32_TIME_RTOL
+  )
+  np.testing.assert_allclose(
+    structured["end_time"], expected_ends, rtol=_FLOAT32_TIME_RTOL
+  )
 
 
 def _test_end_time_clipping_multiple_segments(

src/birdnet_tests/acoustic_models/v3_0/model_py/test_encode/test_acoustic_encode_model_v3_0.py

@@ -10,6 +10,7 @@
   assert_encoding_result_is_close,
   ensure_onnxruntime_or_skip,
   ensure_torch_or_skip,
+  ensure_v3_0_torch_backend_or_skip,
 )
 from birdnet_tests.test_files import TEST_FILE_SHORT
 
@@ -23,6 +24,7 @@
 def _load_model(backend: _Backend) -> AcousticModelV3_0:
   if backend == "pt":
     ensure_torch_or_skip()
+    ensure_v3_0_torch_backend_or_skip()
   else:
     ensure_onnxruntime_or_skip()
 
@@ -84,6 +86,7 @@ def test_v3_0_encode_respects_segment_size(
 
 def test_v3_0_encode_pt_and_onnx_are_close() -> None:
   ensure_torch_or_skip()
+  ensure_v3_0_torch_backend_or_skip()
   ensure_onnxruntime_or_skip()
 
   pt_model = load("acoustic", "3.0", "pt", precision="fp32")
@@ -108,6 +111,7 @@ def test_v3_0_encode_pt_and_onnx_are_close_with_custom_segment_size(
   segment_size_s: float,
 ) -> None:
   ensure_torch_or_skip()
+  ensure_v3_0_torch_backend_or_skip()
   ensure_onnxruntime_or_skip()
 
   pt_model = load("acoustic", "3.0", "pt", precision="fp32")

src/birdnet_tests/acoustic_models/v3_0/model_py/test_predict/test_acoustic_predict_model_v3_0.py

@@ -10,6 +10,7 @@
   assert_prediction_result_is_close,
   ensure_onnxruntime_or_skip,
   ensure_torch_or_skip,
+  ensure_v3_0_torch_backend_or_skip,
 )
 from birdnet_tests.test_files import TEST_FILE_SHORT
 
@@ -23,6 +24,7 @@
 def _load_model(backend: _Backend) -> AcousticModelV3_0:
   if backend == "pt":
     ensure_torch_or_skip()
+    ensure_v3_0_torch_backend_or_skip()
   else:
     ensure_onnxruntime_or_skip()
 
@@ -98,6 +100,7 @@ def test_v3_0_predict_respects_segment_size(
 
 def test_v3_0_predict_pt_and_onnx_are_close() -> None:
   ensure_torch_or_skip()
+  ensure_v3_0_torch_backend_or_skip()
   ensure_onnxruntime_or_skip()
 
   pt_model = load("acoustic", "3.0", "pt", precision="fp32")
@@ -130,6 +133,7 @@ def test_v3_0_predict_pt_and_onnx_are_close_with_custom_segment_size(
   segment_size_s: float,
 ) -> None:
   ensure_torch_or_skip()
+  ensure_v3_0_torch_backend_or_skip()
   ensure_onnxruntime_or_skip()
 
   pt_model = load("acoustic", "3.0", "pt", precision="fp32")

src/birdnet_tests/conftest.py

@@ -1,7 +1,20 @@
 import logging
 
+import pytest
+
 from birdnet.utils.logging_utils import get_package_logger
 
+# The v3.0 models are ~520 MiB and Zenodo serves them at roughly 2 MiB/s, so a single
+# download takes ~5-6 min. That sits right on the global 300s timeout, which cut the
+# downloads off just short of completion instead of letting them finish.
+LOAD_MODEL_TIMEOUT_S = 1800
+
+
+def pytest_collection_modifyitems(items: list[pytest.Item]) -> None:
+  for item in items:
+    if item.get_closest_marker("load_model") is not None:
+      item.add_marker(pytest.mark.timeout(LOAD_MODEL_TIMEOUT_S))
+
 
 def pytest_configure() -> None:
   loggers = {"tensorflow", "absl", "urllib3"}

src/birdnet_tests/helper.py

@@ -256,6 +256,14 @@ def ensure_not_intel_macos_or_skip() -> None:
     pytest.skip("Test not supported on Intel macOS systems")
 
 
+def ensure_v3_0_torch_backend_or_skip() -> None:
+  # torch 2.2.2 is the last release shipping x86 macOS wheels and is too old for the
+  # v3.0 TorchScript model: loading it succeeds, but the inference workers then die and
+  # the session is cancelled. The onnx backend works on the same machines.
+  if check_is_intel_macos():
+    pytest.skip("Acoustic model v3.0 needs a newer torch than Intel macOS provides")
+
+
 def ensure_gpu_or_skip_smi() -> None:
   gpu_available = False
   try: