[eazy] add more log when accuracy check fail (#121656)

shunting314 · pytorchmergebot · commit 522d9729247d · 2024-03-12T20:58:20.000Z
Add these log to debug the regress of accuracy test for dm_nfnet_f0 model for training. With these extra log when the accuracy check fail, we can verify if it's close to succeed or not. If yes that indicates there is no real issue but just flaky and we probably can tune the tolerance to fix. Pull Request resolved: #121656 Approved by: https://github.com/jansel, https://github.com/Skylion007
diff --git a/torch/_dynamo/utils.py b/torch/_dynamo/utils.py
@@ -1350,10 +1350,13 @@ def to_tensor(t):
                 passes_test = res_error <= (multiplier * ref_error + tol / 10.0)
                 if not passes_test:
                     log_error(
-                        "RMSE (res-fp64): %.5f, (ref-fp64): %.5f and shape=%s",
+                        "RMSE (res-fp64): %.5f, (ref-fp64): %.5f and shape=%s. res.dtype: %s, multiplier: %f, tol: %f",
                         res_error,
                         ref_error,
                         res.size(),
+                        res.dtype,
+                        multiplier,
+                        tol,
                     )
                     # import pdb; pdb.set_trace()
                 return passes_test
