HENT-SRT recipe training problem

[redsheep913]

I've use hent-srt recipe to train a model, the default setting like using RNN-T loss is True, other loss like ctc-loss, cr-ctc-loss are False, However in HENT-SRT paper cr-ctc-loss for ASR and ST are enabled, ctc-loss is also required when cr-ctc-loss is enabled. Therefore I train a model with RNN-T loss, ctc-loss, cr-ctc-loss are enabled. It turned out to be the grad_scale is too small and the training had failed.
I am wondering how to train HENT-SRT model correctly.
log-train-2025-11-03-18-21-05.log