[V1][Perf] Avoid mem duplication when aggregating MM tensors

[njhill]

When stacking or concatenating, replace the original tensors with slices into the larger one.

cc @ywang96 @DarkLight1337

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[DarkLight1337]

Do you know why the tensors couldn't be freed in the original code?

Ok I think I get it now. When _reduce_data is called, the original tensors are copied but they are still referenced via MultiModalFieldElem.data which prevents them from being garbage collected.

[DarkLight1337]

I think we may need to abstract this a bit further since #15130 will introduce dim argument to concat which complicates things. Can we reuse the logic of build_elems to assign the slices back into the original tensors?

[njhill]

Thanks @DarkLight1337. it should be simple to generalize this for other dims, however that will result it non-contiguous slices and so could be detrimental depending on how we end up serializing things in #16432. I think if we can defer doing the "reduce" until later (i.e. only after they have been transferred) it would also solve that though.

[DarkLight1337]

Thanks @DarkLight1337. it should be simple to generalize this for other dims, however that will result it non-contiguous slices and so could be detrimental depending on how we end up serializing things in #16432. I think if we can defer doing the "reduce" until later (i.e. only after they have been transferred) it would also solve that though.

This may take a bit of work. I would like to at least finish #15686 before attempting this as it helps simplifies the code (considering that multi-modal processor is used in both V0 and V1).

[github-actions]

This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @njhill.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork