Batched PostXferReq

[robertgshaw2-redhat]

• install https://github.com/Pyngon/nixl/pull/1/files

Key implementation:

• enable multiple ucp workers
• on nixl side: create batched api which posts the Xfers for each worker in a separate thread
• on vllm side: chunk xfer up into groups of 100 descs --- each gets a separate handle (which is done by each worker). Pass these chunks to the batched api
• on vllm side: wait until each one of these handles is done. send a notif once done

This seems to get ~60ms of latency in the launch for 30k tokens on llama-8b

[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.

🚀