Integrate quick allreduce and select the best allreduce implementation

[lihaoyang-amd]

1. With its low-granularity quantization, https://github.com/mk1-project/quickreduce brings huge performance gains to allreduce on tp2 and tp4 on rocm, and does not significantly degrade the model's performance.
2. We integrated quick allreduce into vllm to support 1stage(f16 ), and 2stage(f16, fp8, Q8, Q6, Q4).
3. It is worth mentioning that the speedup of qr is brought about by sacrificing a certain amount of precision, and custom_qr is significantly better than qr's 1stage and 2stage methods at low data volumes (less than 128kb), so we need to judge whether to choose qr or cr or rccl by some conditions.

[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.

🚀

[youkaichao]

what's the relationship between this and #16804 ?

[lihaoyang-amd]

what's the relationship between this and #16804 ?

Aha, maybe we are in competition. We're from amd. We recently spent some time trying to integrate qr into vllm (because qr is very suitable for rocm)

Integrating qr makes the two pr have many similarities, but it seems that the pr you mentioned #16804 only supports Q8 and Q 4. There are no obvious boundary conditions, quantization seems to have some problems, and lack of experimental data.

Maybe we can work together to finish the work.

[mergify]

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

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

[mergify]

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

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

[mergify]

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

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

[mergify]

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

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

[lihaoyang-amd]

@youkaichao hi, kaichao,
from the experimental results, qr has no acceleration for bfloat16 with tp4 and tp8 due to the lack of bfloat16 instruction in mi300, also qr supports rocm for the time being, so we still need custom allreduce.
I noticed custom supports graph capture, does quick allreduce need to support graph mode as well？This 35 times captured graph will be used in formal reasoning and reasoning that doesn't match the shape of this 35-capture graph then uses eager_mode, right?

[mergify]

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

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

[youkaichao]

qr has no acceleration for bfloat16 with tp4 and tp8 due to the lack of bfloat16 instruction in mi300, also qr supports rocm for the time being, so we still need custom allreduce.

I wish qr itself contains the logic of selecting qr or custom allreduce, since their interface is quite the same. My request is that we don't touch the cuda code path, so that people reading the code will not need to think about quick reduce.

graph mode allreduce is necessary for some low-latency workloads where the batchsize is small.

[lihaoyang-amd]

qr has no acceleration for bfloat16 with tp4 and tp8 due to the lack of bfloat16 instruction in mi300, also qr supports rocm for the time being, so we still need custom allreduce.

I wish qr itself contains the logic of selecting qr or custom allreduce, since their interface is quite the same. My request is that we don't touch the cuda code path, so that people reading the code will not need to think about quick reduce.

graph mode allreduce is necessary for some low-latency workloads where the batchsize is small.

Hi, @youkaichao
I made the initial changes you suggested and the code is now working.
I'm not sure if this meets your expectations, please leave me some comments so I can move on to the next step, thank you very much!
#19094

[tlrmchlsmth]

closing in favor of #19744