Nightwing-77
Nightwing-77
> Who will be doing it? u can work on it and open a pr i think
> # Intermittent CUDA Out of Memory Errors during Parakeet-TDT-0.6B Fine-tuning > ## Describe the bug > Intermittent and unpredictable CUDA out-of-memory errors occur during fine-tuning of the Parakeet-TDT-0.6B model....
Hey so I had a similar setup based on trial and error and I'm facing issues related to convergence of parakeet!! Any specific reason for it!! Or it's just the...
Instead of fully randomized I had sorting based batching in which we had 5 classes based on size of it and created varying batch sizes based on them(I thought that...
@jeremy110 hey , i've been stuck in one issue on how to train nemo model on parquet files !! i added https://huggingface.co/datasets/ai4bharat/IndicVoices to my training data ! but the issue...
@jeremy110 hey so actually i don't have any tensorboard graphs as i switched logging off cause i was getting many issues in it (kaggle) so i trained it for around...
> [@Nightwing-77](https://github.com/Nightwing-77) This is normal. It's only when the 10k steps are ready to start converging, and it's not until the 20k to 30k steps that the output will appear....
@jeremy110 does nemo not support control tokens !?! cause control tokens are not meant to be predicted , or computed in rnnt loss