Empty responses for complex queries when running demo_multiturn_chat.py

[tanhuajie]

Hi team! Thanks for such an amazing work on this project!

I'm encountering an issue when running the provided demo_multiturn_chat.py script. The model seems to work fine with only simple inputs like "Hi" or "What's your name?".

But when I try more complex questions such as:

• "Can you recommend me some places to visit in Beijing?"
• "How can I learn to play the violin?"

After running through the default 256 steps of forward diffusion, the model returns empty responses. This behavior seems inconsistent since simpler queries work as expected.

Have you encountered this issue before? Any suggestions on how to resolve it or what might be causing this behavior? Thanks again!!

[zihaoli0629]

Same issue here. I also notice that when alg_temp tends to 0, the model behaves like an AR model and the accuracy improves significantly for math problems.

[jiacheng-ye]

Hi there, the empty response is mostly because the model generates a sentence with all tokens, which may be because of the lower temperature and randomness. You can consider trying a slightly higher temperature.

Hi team! Thanks for such an amazing work on this project!

I'm encountering an issue when running the provided demo_multiturn_chat.py script. The model seems to work fine with only simple inputs like "Hi" or "What's your name?".

But when I try more complex questions such as:

• "Can you recommend me some places to visit in Beijing?"
• "How can I learn to play the violin?"

After running through the default 256 steps of forward diffusion, the model returns empty responses. This behavior seems inconsistent since simpler queries work as expected.

Have you encountered this issue before? Any suggestions on how to resolve it or what might be causing this behavior? Thanks again!!

[jiacheng-ye]

Same issue here. I also notice that when alg_temp tends to 0, the model behaves like an AR model and the accuracy improves significantly for math problems.

Good observation!