Daniel Galvez
Daniel Galvez
jenkins Previous failure seems to be a spurious failure caused by git clone failing. @artbataev I incorporated your change after verifying it on a multi-GPU machine. Thank you again. I...
jenkins Sorry for missing the guard in that test. Hopefully things go through now.
FYI, CI is faling with: ``` extras/check_dependencies.sh: python2.7 is not installed extras/check_dependencies.sh: Some prerequisites are missing; install them using the command: sudo apt-get install python2.7 make: *** [Makefile:39: check_required_programs] Error...
Fixed CI (so far)
FYI @danpovey you might find these very low latencies exciting. I'm going to be incorporating this into https://github.com/nvidia-riva/riva-asrlib-decoder (via the kaldi submodule within that project) so that CTC models (and...
@ravi-shanker-m that's been deprecated for a few years now: https://github.com/kaldi-asr/kaldi/blob/be22248e3a166d9ec52c78dac945f471e7c3a8aa/src/cudadecoder/batched-threaded-nnet3-cuda-pipeline.h#L35-L36 I'm happy to go ahead and remove that code. Were you using it for some reason?
@trunglebka, I'm happy to provide advice if you give more detail. I would sincerely doubt that you reach anywhere near 8000 RTFx on the v1 cuda decoder on an A100...
@trunglebka Okay. I found several performance problems with the v2 decoder during my work on making this PR and this is very close to the "speed of light", so I'm...
My only comment is that I'm not a big fan of polluting the global namespace with int32, etc. It can cause downstream pain for those who use kaldi as a...