Fix RNN-T loss memory usage

[artbataev]

What does this PR do ?

Fixes memory usage for Numba-based implementation of RNN-T and Multi-blank Transducer losses.
The current implementation requires 3x memory compared to the size of logits (logits, gradient, extra memory = size of logits). This PR fixes memory usage to the minimal possible 2x (logits, gradient).

It looks like assigning tensors directly to ctx instead of saving them through save_for_backward breaks PyTorch logic, and it copies the gradient tensor (which results in extra memory usage).

TDT loss was not affected by this issue (I'm unsure why, but it requires contiguous tensor for labels-related logits).

Before (main):

• Peak memory before loss: 2.09 GB
• Peak memory after loss: 6.27 GB

After (this PR):

• Peak memory before loss: 2.09 GB
• Peak memory after loss: 4.18 GB

Code to check memory usage (size of tensors besides logits is negligible compared to logits):

import torch
from nemo.collections.asr.parts.numba.rnnt_loss import RNNTLossNumba
device = torch.device("cuda")

loss = RNNTLossNumba(blank=1023, reduction='none')
batch_size = 32
logits = torch.rand([batch_size, 188, 90 + 1, 1024], device=device, dtype=torch.float32, requires_grad=True)
encoder_lengths = torch.full([batch_size], fill_value=188, device=device, dtype=torch.long)
label_lengths = torch.full([batch_size], fill_value=90, device=device, dtype=torch.long)
labels = torch.randint(0, 1022, size=[batch_size, 90], dtype=torch.long, device=device)
print(f"Peak memory before loss: {torch.cuda.max_memory_allocated() / (2 ** 30):.2f} GB")

loss_value = loss(acts=logits, act_lens=encoder_lengths, labels=labels, label_lens=label_lengths)
loss_value.sum().backward()
print(f"Peak memory after loss: {torch.cuda.max_memory_allocated() / (2 ** 30):.2f} GB")

Collection: [ASR]

Changelog

• In transducer-related functions that extend autograd, save tensors using ctx.save_for_backward(...) instead of directly assigning tensors according to PyTorch documentation.

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[hainan-xv]

Great discovery of the memory usage issue and a very clean fix! Approved and thanks.

[github-actions]

[🤖]: Hi @artbataev 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully

So it might be time to merge this PR or get some approvals

I'm just a bot so I'll leave it you what to do next.

//cc @pablo-garay @ko3n1g

[titu1994]

Thanks for the fix !