wsd12345 comments

Results 4 comments of


                                            wsd12345

In generate.py , prompt_length > 4096(generate_long), lead to max_new_tokens< 0.

> Your reference audio length should not exceed 90 seconds. Thanks for your response. I use 23s of reference audio。In decode_n_tokens，I find that stop sign is not predicted. The dimensions...

In generate.py , prompt_length > 4096(generate_long), lead to max_new_tokens< 0.

> You 100% need multinomial sampling, argmax will cause repetition pattern. The generated audio waveform is indeed repetitive noise at the back. But I don't know why it keeps repeating,...

In generate.py , prompt_length > 4096(generate_long), lead to max_new_tokens< 0.

> Using greedy for any LLM can also meet same issue. Thanks。

The model inference time is inconsistent

[Model and test code](https://github.com/wsd12345/sense_triton_question.git) The above problem uses test/1.py. The deployment server is in the cloud, Tesla V100-SXM2-16GB. And, I found the following problems: I run test/2.py, that turned out...