test: adjust is_timing_close (#17178)

Author: Borda

tests/tests_fabric/parity/test_parity_ddp.py

@@ -130,8 +130,8 @@ def train_fabric_ddp(fabric):
 @pytest.mark.parametrize(
     "accelerator, devices, tolerance",
     [
-        ("cpu", 2, 0.01),
-        pytest.param("cuda", 2, 0.005, marks=RunIf(min_cuda_gpus=2)),
+        ("cpu", 2, 0.02),
+        pytest.param("cuda", 2, 0.01, marks=RunIf(min_cuda_gpus=2)),
     ],
 )
 def test_parity_ddp(accelerator, devices, tolerance):

tests/tests_fabric/parity/test_parity_simple.py

@@ -145,7 +145,7 @@ def test_parity_single_device(precision, accelerator):
     assert is_state_dict_equal(state_dict_torch, state_dict_fabric)
 
     # Compare the time per iteration
-    assert is_timing_close(timings_torch, timings_fabric, rtol=1e-3, atol=1e-3)
+    assert is_timing_close(timings_torch, timings_fabric, rtol=1e-2, atol=0.1)
 
     # Compare memory usage
     if accelerator == "cuda":

tests/tests_fabric/parity/utils.py

@@ -24,7 +24,7 @@ def is_state_dict_equal(state0, state1):
     return all(eq_fn(w0.cpu(), w1.cpu()) for w0, w1 in zip(state0.values(), state1.values()))
 
 
-def is_timing_close(timings_torch, timings_fabric, rtol=1e-3, atol=1e-3):
+def is_timing_close(timings_torch, timings_fabric, rtol=1e-2, atol=0.1):
     # Drop measurements of the first iterations, as they may be slower than others
     # The median is more robust to outliers than the mean
     # Given relative and absolute tolerances, we want to satisfy: |torch – fabric| < RTOL * torch + ATOL