Error while get_wer_feat_logit with stt_en_fastconformer_hybrid_large_pc model

[Khimer]

Hello!
I'm trying to use FrameBatchASRLogits to recognize large audio with a ctc type decoder with timestamps. (to change the type of decoder I execute:

new_decoding_config = copy.deepcopy(asr_model.decoding.cfg)
         with open_dict(new_decoding_config):
             new_decoding_config.preserve_alignments = True
             new_decoding_config.strategy = 'greedy'
         self.asr_model.change_decoding_strategy(new_decoding_config, decoder_type='ctc')

But when I run:

hyp, greedy_predictions_list, log_prob = get_wer_feat_logit(
                 audio_file_path=audio_file_path,
                 asr=frame_asr,
                 frame_len=chunk_len,
                 tokens_per_chunk=tokens_per_chunk,
                 delay=mid_delay,
           model_stride_in_secs=transcription_config.model_stride_in_secs,
             )

I get an error:

CondaEnvs/nemo_1.19/lib/python3.8/site-packages/nemo/collections/asr/parts/utils/decoder_timestamps_utils.py:269, in FrameBatchASRLogits._get_batch_preds(self, keep_logits)
     267 feat_signal, feat_signal_len = feat_signal.to(device), feat_signal_len.to(device)
     268 print(self.asr_model(processed_signal=feat_signal, processed_signal_length=feat_signal_len))
--> 269 log_probs, encoded_len, predictions = self.asr_model(
     270 processed_signal=feat_signal, processed_signal_length=feat_signal_len
     271)
     272 preds = torch.unbind(predictions)
     273 for pred in preds:
ValueError: not enough values to unpack (expected 3, got 2)

I looked at the output of self.asr_model(processed_signal=feat_signal, processed_signal_length=feat_signal_len) and see that it lacks predictions tensor.
The nvidia/stt_en_conformer_ctc_large model does not have this problem.
Do I need to configure any additional options for the fastconformer_hybrid model?

[titu1994]

The hybrid model is not currently supported with buffered inference. We can try adding it in the next few versions. In the mean time you can try using the fast conformer RNNT model instead

[titu1994]

@fayejf has submitted this support here - #6570

Could you try using the main Branch if NeMo ?