Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

SpeechLM Update #12430

Open
wants to merge 165 commits into
base: main
Choose a base branch
from
Open

SpeechLM Update #12430

wants to merge 165 commits into from

Conversation

stevehuang52
Copy link
Collaborator

@stevehuang52 stevehuang52 commented Feb 28, 2025

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

What does this PR do ?

A couple of important updates in SpeechLM

Collection: [speechlm]

Changelog

  • Unified various input format into multimodal conversation, where audio and text are interleaved. This is achieved by mapping single turn Cut samples into NeMoMultimodalConversation. Later should also convert text SFT data to NeMoMultimodalConversation in nemo/collections/speechlm/data/dataset/audio_text_lhotse_dataset.py:MultimodalConversationDataset as well.
  • Added Whisper encoder support.
  • Fixed issues with PEFT not saving batch norm stats.
  • Fixed loading PEFT ckpt for inference.
  • Context parallel support in LLM. Speech encoder still cannot.
  • Various improvements and refactoring.

stevehuang52 and others added 30 commits September 16, 2024 10:38
Signed-off-by: stevehuang52 <[email protected]>
add type hint

Signed-off-by: He Huang (Steve) <[email protected]>
Signed-off-by: He Huang (Steve) <[email protected]>
Signed-off-by: He Huang (Steve) <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: stevehuang52 <[email protected]>
@stevehuang52 stevehuang52 marked this pull request as ready for review March 3, 2025 20:59
@stevehuang52 stevehuang52 requested a review from huvunvidia March 3, 2025 21:03
@@ -330,7 +330,7 @@ def parse_and_combine_datasets(
if (w := item.get("weight")) is not None:
weights.append(w)

assert all(t == tarred_status[0] for t in tarred_status), "Mixing tarred and non-tarred datasets is not supported."
# assert all(t == tarred_status[0] for t in tarred_status), "Mixing tarred and non-tarred datasets is not supported."
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is commented out to allow mixing tarred and untarred datasets, then use force_map_dataset or force_iterable_dataset later to control whether they're wrapped into map or iterable dataset. @pzelasko Do we have a better solution now?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants