Avoid unnecessary multi-modal input data copy when len(batch) == 1

[imkero]

A len=1 multi-modal data batch (List[torch.Tensor]) can avoid a copy operation by creating view of original tensor data if possible.

• if len(batch) == 1 then
  • torch.stack(batch) -> batch[0].unsqueeze(0).contiguous()
  • torch.concat(batch) -> batch[0].contiguous()
• NOTE: ensure contiguous to achieve consistent behaviour

Example code:

import torch

src_tensor = torch.rand((2, 3, 4))
batch = [src_tensor]

assert len(batch) == 1
# only works when batch size = 1

copied_batched_tensor = torch.stack(batch)
print(src_tensor.storage().data_ptr() == copied_batched_tensor.storage().data_ptr())
# False

no_copy_batched_tensor = src_tensor.unsqueeze(0).contiguous()
print(src_tensor.storage().data_ptr() == no_copy_batched_tensor.storage().data_ptr())
# True

print(torch.equal(copied_batched_tensor, no_copy_batched_tensor))
# True

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

Thanks for optimizing this! Can you leave a code comment explaining why we need separate logic?

[ywang96]

Thanks for the optimization!

[imkero]

Thanks for optimizing this! Can you leave a code comment explaining why we need separate logic?

Sure! I will update it soon.