Add quickreduce as alternative to custom allreduce

[ilmarkov]

Add quickreduce alternative to custom allreduce.
The collective is only enabled on AMD. The kernels support quantized all reduce collective int4/int8 symmetric and asymmetric quantization algorithms.
The kernels can be enabled by quick_reduce_allreduce_algo config parameter. On AMD we first check if it is profitable to use quickreduce, otherwise we fallback to custom allreduce.

[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.

🚀

[lixixicommute]

hi, @ilmarkov ,
Thank you for your great work.
Does this program have any test data and how well does it work?
It looks like it's still a draft at the moment, will it be refined afterward?

[mergify]

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

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

[youkaichao]

On AMD we first check if it is profitable to use quickreduce, otherwise we fallback to custom allreduce.

what are the cases when custom allreduce performs better than quickreduce? It would better if quickreduce can surpass custom allreduce in all cases, then we can use quickreduce as a drop-in replacement of custom allreduce without a new user-facing flag.

[ilmarkov]

@youkaichao It is slower for smaller input sizes. We could do the similar approach as custom allreduce has - use one shot for small buffers and two shot for larger ones.

[youkaichao]

@youkaichao It is slower for smaller input sizes. We could do the similar approach as custom allreduce has - use one shot for small buffers and two shot for larger ones.

that would be great, can you implement it? we can use either quickreduce or custom allreduce at the engine level, instead of dynamically switching based on the input size.

[ilmarkov]

Yes, we can try to implement this approach.
Although, custom allreduce setup and implementation is more suitable for low latency small input sizes, whereas quick reduce performs well for bandwidth bottlenecked workloads. At the moment, we use custom allreduce or nccl based on input size.
Also, we will still need the new uder-facing flag as we need to provide switch between quantization regimes allowing user to find a trade-off between accuracy and performance.

[youkaichao]

Also, we will still need the new uder-facing flag as we need to provide switch between quantization regimes allowing user to find a trade-off between accuracy and performance.

you can use an environment variable, like VLLM_ROCM_CA_BACKEND.

[mergify]

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

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