Getting "Caught signal 11 (Segmentation fault: address not mapped to object" when run "eval_beamsearch_ngram_ctc.py"

[mehadi92]

For evaluating beamsearch I use NeMo image docker pull nvcr.io/nvidia/nemo:24.01.speech inside this container I run install_beamsearch_decoders.sh after I run eval script like bellow,

time python eval_beamsearch_ngram_ctc.py \
    'nemo_model_file="/models/my-asr-model.nemo"' \
    'input_manifest="nemo_manifest.json"' \
    'kenlm_model_file="/models/my-lm-model.arpa"' \
    'beam_width=[8,16]' \
    'beam_alpha=[1.0,2.0]' \
    'beam_beta=[0.5,1.0]' \
    'probs_cache_file=null' \
    'decoding_mode=beamsearch_ngram'

But I'm getting Caught signal 11 (Segmentation fault: address not mapped to object at address 0x10) Here is the error logs,

[NeMo I 2024-09-20 17:47:27 eval_beamsearch_ngram_ctc:401] ==============================Starting the beam search decoding===============================
[NeMo I 2024-09-20 17:47:27 eval_beamsearch_ngram_ctc:402] Grid search size: 8
[NeMo I 2024-09-20 17:47:27 eval_beamsearch_ngram_ctc:403] It may take some time...
[NeMo I 2024-09-20 17:47:27 eval_beamsearch_ngram_ctc:404] ==============================================================================================
Beam search decoding with width=8, alpha=1.0, beta=0.5:   0%|                                    | 0/11 [00:00<?, ?it/s][NeMo I 2024-09-20 17:47:27 ctc_decoding:309] Beam search requires that consecutive ctc tokens are not folded. 
    Overriding provided value of `fold_consecutive` = True to False
Loading the LM will be faster if you build a binary file.
Reading /models/my-lm-model.arpa
----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
****************************************************************************************************
[efba42f53a9a:375  :0:458] Caught signal 11 (Segmentation fault: address not mapped to object at address 0x10)
Segmentation fault (core dumped)

Is it supposed to happen? or I'm missing something?

Thanks

[mehadi92]

I also tried Dockerfile.ngramtools but get the same Segmentation fault issue

[eyalgenesys]

similar segmentation fault when running eval_greedy_decoding_with_context_biasing with batch size =1 and device=cpu