Grad overflow on iteration occurs frequently?

Hi, Grad overflow on iteration occurs almost every step in my experiments, and the result is bad, what's going on? Thanks for your reply.

Epoch 1/20
  0% 0/184 [00:00<?, ?it/s]
Grad overflow on iteration 0
Using dynamic loss scale of 65536
  1% 1/184 [00:00<01:34,  1.93it/s, loss=4.215]
Grad overflow on iteration 1
Using dynamic loss scale of 32768.0
  1% 2/184 [00:00<01:24,  2.15it/s, loss=4.174]
Grad overflow on iteration 2
Using dynamic loss scale of 16384.0
  4% 8/184 [00:03<01:14,  2.35it/s, loss=4.251]
Grad overflow on iteration 8
Using dynamic loss scale of 8192.0
  8% 14/184 [00:05<01:02,  2.72it/s, loss=3.725]
Grad overflow on iteration 14
Using dynamic loss scale of 4096.0
  8% 15/184 [00:05<00:58,  2.87it/s, loss=3.562]
Grad overflow on iteration 15
Using dynamic loss scale of 2048.0
  9% 16/184 [00:06<01:10,  2.38it/s, loss=3.418]
Grad overflow on iteration 16
Using dynamic loss scale of 1024.0
  9% 17/184 [00:06<01:09,  2.41it/s, loss=3.292]
Grad overflow on iteration 17
Using dynamic loss scale of 512.0
 10% 18/184 [00:07<01:06,  2.50it/s, loss=3.187]
Grad overflow on iteration 18
Using dynamic loss scale of 256.0
 10% 19/184 [00:07<01:08,  2.40it/s, loss=3.098]
Grad overflow on iteration 19
Using dynamic loss scale of 128.0
 11% 20/184 [00:08<01:07,  2.44it/s, loss=3.017]
Grad overflow on iteration 20
Using dynamic loss scale of 64.0
 12% 22/184 [00:08<01:06,  2.45it/s, loss=2.866]
Grad overflow on iteration 22
Using dynamic loss scale of 32.0
 12% 23/184 [00:09<01:11,  2.26it/s, loss=2.791]
Grad overflow on iteration 23
Using dynamic loss scale of 16.0
 13% 24/184 [00:09<01:07,  2.37it/s, loss=2.718]
Grad overflow on iteration 24
Using dynamic loss scale of 8.0
 17% 32/184 [00:13<00:59,  2.55it/s, loss=2.405]
Grad overflow on iteration 32
Using dynamic loss scale of 4.0
 20% 37/184 [00:15<01:12,  2.02it/s, loss=2.350]
Grad overflow on iteration 37
Using dynamic loss scale of 2.0
 21% 38/184 [00:16<01:24,  1.72it/s, loss=2.312]
Grad overflow on iteration 38
Using dynamic loss scale of 1.0
 22% 40/184 [00:17<01:10,  2.05it/s, loss=2.245]
Grad overflow on iteration 40
Using dynamic loss scale of 1
 22% 41/184 [00:17<01:13,  1.96it/s, loss=2.211]
Grad overflow on iteration 41
Using dynamic loss scale of 1
 23% 42/184 [00:18<01:08,  2.06it/s, loss=2.175]
Grad overflow on iteration 42
Using dynamic loss scale of 1
 25% 46/184 [00:19<00:57,  2.41it/s, loss=2.082]
Grad overflow on iteration 46
Using dynamic loss scale of 1
 26% 48/184 [00:20<00:56,  2.41it/s, loss=2.040]
Grad overflow on iteration 48
Using dynamic loss scale of 1
 27% 49/184 [00:21<00:58,  2.30it/s, loss=2.022]
Grad overflow on iteration 49
Using dynamic loss scale of 1
 27% 50/184 [00:21<00:58,  2.29it/s, loss=2.002]
Grad overflow on iteration 50
Using dynamic loss scale of 1

No file, test, or entry point is named. Start by reproducing the reported mixed-precision training run and inspect the dynamic loss-scale behavior around the logged gradient overflows; done means explaining whether the repeated overflows are expected and why they produce a bad result.