Extremely slow generation time (~15 hours for 5 samples) — is this expected?

[Yohansenanayake]

Hi,

I'm a final-year undergraduate majoring in Computer Engineering, and for my research project I selected Museformer as one of the baseline models.

I successfully trained a model following the instructions in the repository and then attempted to generate 5 samples according to the generation instructions. However, generating the token sequences took nearly 15 hours on a server with 4× NVIDIA RTX Ada 6000 GPUs (48 GB each).

I noticed that the README mentions generation may take a while, but it does not quantify the expected time. Therefore I wanted to clarify whether this runtime is expected behavior, or if there might be something incorrect in my setup.

For reference, the final line of the generation log was:

2026-03-08 06:06:30 | INFO | fairseq_cli.interactive | Total time: 54132.240 seconds; translation time: 54126.992

Any clarification on whether this runtime is typical (or suggestions for improving generation speed) would be greatly appreciated.

Thanks!