MagpieTTS decoder model on top of NeMo main branch#15277

Open

paarthneekhara wants to merge 102 commits intoNVIDIA-NeMo:mainfrom

paarthneekhara:magpietts_decoderonly_2601

Commits on Mar 10, 2026

MagpieTTS decoder model working on top of NeMo main branch

paarthneekhara
authored and
shehzeen
committed
merge wit main again

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
handling changes in dataloader

paarthneekhara
authored and
shehzeen
committed
hack to avoid HF error

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
remove discriminatory temporarily

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
fix errors

paarthneekhara
authored and
shehzeen
committed
bug fix

paarthneekhara
authored and
shehzeen
committed
add moe

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
20 layer moe

paarthneekhara
authored and
shehzeen
committed
some refactoring and clean up

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
bug fix related to spectral codec

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
some clean up

paarthneekhara
authored and
shehzeen
committed
add docstrings and data classes

paarthneekhara
authored and
shehzeen
committed
more doc strings

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
support multiple training modes

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
default mode for backward compatibility

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
default config changes
shehzeen
committed
Magpietts decoderonly 2601 bpe ipa tokenizer (#57)
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
nemotron mamba model (#58)

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
inference function refactoring

paarthneekhara
authored and
shehzeen
committed
revert some changes and remove scripts

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
Magpietts decoderonly 2601 simplify code (#60)

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
include vocab file

paarthneekhara
authored and
shehzeen
committed
Magpietts decoderonly 2601 valinfer (#61)

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
bug fixes in inference and logging
shehzeen
committed
more tests
shehzeen
committed
tested and verified that infer batch works correctly with teacher forcing and matches process batch output
shehzeen
committed
added legacy option to still work with 21fps F2F model
shehzeen
committed
remove streaming decode because it not being used
shehzeen
committed
pass phoneme EOS to next step
shehzeen
committed
exlcude codec model from optimizer params
shehzeen
committed
reduce dropout prob, change default delays to 0,1
shehzeen
committed
bug fix
shehzeen
committed
phoneme EOS handling bug fix
shehzeen
committed
phoneme corruption methodology implemented
shehzeen
committed
revisit defaults and update
shehzeen
committed
bug fix phoneme loss
shehzeen
committed
another inference bug fix
shehzeen
committed
phoneme vocab size fix
shehzeen
committed
bug fix
shehzeen
committed
handle legacy model phoneme vocab size
shehzeen
committed
context duration handling - stop repeating excessively
shehzeen
committed
clamp cer and wer to 1
shehzeen
committed
Preference Optimization for EasyMagpieTTS (#64)
shehzeen
committed
po stabilize
shehzeen
committed
mamba config update

paarthneekhara
authored and
shehzeen
committed
fix weight initialization bugs in mamba

paarthneekhara
authored and
shehzeen
committed
Magpietts decoderonly 2601 flash (#65)

paarthneekhara
authored and
shehzeen
committed
add do tts method
shehzeen
committed
bug fix

paarthneekhara
authored and
shehzeen
committed
Magpietts decoderonly 2601 utmos po (#67)
shehzeen
committed
full phoneme channel dropout option

paarthneekhara
authored and
shehzeen
committed
gt phoneme option in do_tts
shehzeen
committed
bug fix
shehzeen
committed
ignore phoneme channel for some languages

paarthneekhara
authored and
shehzeen
committed
PO updates, cross lingual dataset creation
shehzeen
committed
add language to text contexts.
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
tokenizer import change

paarthneekhara
authored and
shehzeen
committed
remove unnecessary imports

paarthneekhara
authored and
shehzeen
committed
cleanup

paarthneekhara
authored and
shehzeen
committed
remove name from training modes

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
removing some debugging statements
shehzeen
committed
new base class (#68)

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
cleanup

paarthneekhara
authored and
shehzeen
committed
sanitize logits only for easy magpie to preserve magpietts functionality

paarthneekhara
authored and
shehzeen
committed
remove custom phoneme tokenizer instantiation and handle it in the tokenizer class

paarthneekhara
authored and
shehzeen
committed
Apply isort and black reformatting

paarthneekhara
authored and
shehzeen
committed
remove streaming inference script
shehzeen
committed

Commits on Mar 11, 2026

Commits on Mar 14, 2026

Commits on Mar 16, 2026

Commits on Mar 17, 2026