Thanks. Now it’s fixed. See comments in https://github.com/d2l-ai/d2l-en/issues/1448
the second traning (use_random_iter=True) uses the model trained by the first time:
(1)the ppl is from about 2.0
(2)the curve is bumpy
So misleading 
I argee with “uses the model trained by the first time”. Thank you for reminding us.
@zhangjiekui
I think use_random_iter=True maybe is a way to converge better when it is hard to converge our model.
There is a statement H, = state in function rnn.
I’ve never seen such a usage. What does it mean?
I asked some colleagues around me, but nobody knows.
Can somebody kindly explain it to me?
Many thanks!
The second training still uses the same data iterator which is sequential not random. Am I missing something?