Test reproducibility

Author: cwindolf

tests/test_grab_and_featurize.py

@@ -253,7 +253,11 @@ def test_grab_and_featurize():
 
     # this is kind of a good test of reproducibility
     # totally reproducible on CPU, suprprisingly large diffs on GPU
-    if not torch.cuda.is_available():
+    # reproducibility is fine on some BLAS but not MKL?
+    repro = (not torch.cuda.is_available()) and (
+        "BLAS_INFO=mkl" not in torch.__config__.show()
+    )
+    if repro:
         assert np.array_equal(locs0, locs1)
     else:
         valid = np.clip(locs1[:, 2], geom[:, 1].min(), geom[:, 1].max())

tests/test_matching.py

@@ -92,11 +92,12 @@ def test_tiny(tmp_path):
     assert np.isclose(
         torch.square(res["residual"]).mean(),
         0.0,
+        atol=1e-4,
     )
     assert np.isclose(
         torch.square(res["conv"]).mean(),
         0.0,
-        atol=1e-5,
+        atol=1e-4,
     )
 
     matcher = main.ObjectiveUpdateTemplateMatchingPeeler.from_config(
@@ -124,11 +125,12 @@ def test_tiny(tmp_path):
     assert np.isclose(
         torch.square(res["residual"]).mean(),
         0.0,
+        atol=1e-4,
     )
     assert np.isclose(
         torch.square(res["conv"]).mean(),
         0.0,
-        atol=1e-5,
+        atol=1e-4,
     )
 
 
@@ -270,11 +272,12 @@ def test_tiny_up(tmp_path, up_factor=8):
     assert np.isclose(
         torch.square(res["residual"]).mean(),
         0.0,
+        atol=1e-4,
     )
     assert np.isclose(
         torch.square(res["conv"]).mean(),
         0.0,
-        atol=1e-5,
+        atol=1e-4,
     )
 
 
@@ -410,6 +413,7 @@ def static_tester(tmp_path, up_factor=1):
     assert np.isclose(
         torch.square(res["residual"]).mean(),
         0.0,
+        atol=1e-4,
     )
     assert np.isclose(
         torch.square(res["conv"]).mean(),