fix pynccl reduce_scatter

[youzhedian]

This PR fix the reduce_scatter interface in pynccl-wrapper. The old implementation fails when scatter_dim != 0 because the input_tensor is incorrectly supplied, it doesn’t use the tensor rearranged by movedim.

[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 ask your reviewers to trigger select CI tests on top of fastcheck CI.

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.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[gemini-code-assist]

Code Review

This pull request correctly fixes a bug in the reduce_scatter function where the wrong tensor was being passed to the underlying pynccl communication primitive. The fix ensures that the tensor with the rearranged dimensions is used. However, a similar bug exists in the reduce_scatterv function within the same file, which has been overlooked. I've added a comment to highlight this critical issue, as fixing it would ensure correctness and consistency across related communication operations.

[youkaichao]

cc @tlrmchlsmth