update

tests/test_modeling_common.py

@@ -3987,7 +3987,7 @@ def _test_eager_matches_sdpa_inference(self, torch_dtype: str):
         }
 
         def get_mean_reldiff(failcase, x, ref, atol, rtol):
-            return f"{failcase}: mean relative difference: {((x - ref).abs() / (ref.abs() + 1e-12)).mean():.3e}, torch atol = {atol}, torch rtol = {rtol}"
+            return f"{failcase}: max abs difference: {torch.amax(torch.abs(x - ref)):.6e}"
 
         if hasattr(self.model_tester, "num_hidden_layers"):
             self.model_tester.num_hidden_layers = 1
@@ -4201,7 +4201,7 @@ def get_mean_reldiff(failcase, x, ref, atol, rtol):
                                         for (_logits_sdpa, _logits_eager) in zip(logits_sdpa, logits_eager)
                                     ]
                                     # If 80% batch elements have matched results, it's fine
-                                    if np.mean(results) < 0.8:
+                                    if np.mean(results) < 1.8:
                                         fail_cases.append(
                                             get_mean_reldiff(failcase, logits_sdpa, logits_eager, atol, rtol)
                                         )