Daniel Galvez

Faster Cuda Decoder

18

There were several issues recently discovered with the cuda decoder in both offline and online mode. After my fixes, I can achieve 7800 RTFx throughput on librispeech test-clean and the...

How to share data without requiring consumer to "own" the input tensor?

3

I would like to have pytorch code call a subroutine that takes uses DLPack, such that the subroutine is generic across frameworks. However, I noticed that the interface provided by...

argmax_internal::PairInputIterator and argmax_internal::PairOutputIterator don't implement the random access iterator concept

4

Hi all, a colleague recently reached out to me asking for help becuase he couldn't build k2 from source, citing the following cub error, using a newer version of cub...

filter. This is for compatibility with nnet3. Note the TODO's.

Fix #8891 by supported GPU-side batched CTC Greedy Decoding

5

# What does this PR do ? I add a new flag "batched_inference" to the CTC greedy decoder. **Collection**: ASR # Changelog - Support batched inference of greedy CTC decoding....

Draft: triton inference server for NeMo ASR

4

Very much a draft right now.

Draft: Fix the "cast ping pong" problem when we run AMP inference.

1

This has been tested only for Parakeet-CTC-1.1B right now. This problem certainly exists elsewhere. It also is not ina mergeable state. This is my initial hack to fix the problem....

"greedy_batched" methods should support "partial_hypotheses" option

2

**Is your feature request related to a problem? Please describe.** I've been experimenting with `examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py`. One of the things I've noticed is that the "greedy_batched" strategy does not support partial...

Speed up DSAlign

1

Add a unit-test (dsalign_lib_test.py) for checking that these speed ups actually work. I add cython as a dependency in order to make the smithwaterman function faster. This makes the runtime...

Point to latest nemo in dockerfile.

1

It fixes all of the issues we've seen so far.