Piotr Żelasko comments

Results 523 comments of


                                            Piotr Żelasko

Use torch sdpa implementation in ASR mha

I'm on it - launched Canary-1B training with this, will keep you posted.

Use torch sdpa implementation in ASR mha

Unfortunately the Canary-1B training on 32 GPUs with bf16 AMP is diverging after ~10 hours with this SDPA implementation. I launched a second run with a different seed to confirm...

Use torch sdpa implementation in ASR mha

Uh, sorry for the lack of responsiveness lately. I used bf16-mixed and PyTorch 2.4.0 with CUDA 12.5. I agree we should merge it disabled by default. Let's take a look...

Add CI Tests for Canary/AEDMultitask "lang_field"

I'll look into that.

Question about `validate_cut_set` in qa.py

That used to be a constraint but at some point we dropped it. I may have missed that validation still checks for this. If you could make a PR to...

Zipformer performs well on short audios but terribly on long audios with silences

Roughly speaking, scalability in batch dimension is ~linear, but scalability in sequence length dimension is ~quadratic.

Zipformer performs well on short audios but terribly on long audios with silences

I never trained on utterances longer than ~40s so I can only share my suspicions. I’d expect with longer examples it may be more difficult for the model to find...

Zipformer performs well on short audios but terribly on long audios with silences

Would probably go with your strategy for fixed 30s chunks in both training and inference - consistent and simple..

UnicodeEncodeError: 'ascii' codec can't encode characters in position 505-506: ordinal not in range(128)

The only thing I can think of is that smart_open is now not used for local paths. I restored this functionality in PR https://github.com/lhotse-speech/lhotse/pull/1360 - can you try with that?

Issue using Lhotse in NeMo - RuntimeError: cannot join current thread

One other person has reported this error but I don't know yet where it's coming from. I have a suspicion though. It's going to be a long-shot, but can you...