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.

🚀

[lihaoyang-amd]

tp2
custom when 4k-128k
qr2stagef16 when 256k-512k
qr+quant when 1MB-64MB

<style> </style>

kenrel		vllm	qr	qr	qr	qr	qr
time(us)		custom allreduce	1stage	2stage f16	2stage f8	2stage Q6	2stage Q4
tp2	4k	9.7	15	15.42	25.79	34.98	27.59
	8k	11.25	16.98	15.87	27.07	35.91	29.32
	16k	12.3	14.48	18.98	28.06	31.31	28.67
	32k	13.27	18.3	16.4	26.9	33.46	28.37
	64k	9.6	20.97	18.03	25.33	31.38	42.63
	128k	13.44	15.77	18.17	27.24	36.72	30.7
	256k	38.84	19.39	18.86	30.02	32.33	30.96
	512k	30.6	25.64	24.46	33.3	32.05	29.04
	1M	33.46	36.05	39.58	35.57	37.65	36.32
	2M	57.96	60.01	62.22	66.15	47.17	44.97
	4M	108.15	106.8	105.33	75.19	66.4	55.14
	8M	204.3	195.46	199	134.91	113.08	90.14
	16M	394.06	369.52	371.06	241.09	203.82	154.48
	32M	781.73	732.85	735.03	446.94	367.06	290.67
	64M	1552.63	1443.03	1447.09	833.53	674.4	529.25

I'll add tp4, tp8 and e2e experimental data later, and cleaning up.

[youkaichao]

what's the relationship between this and #16804 ?

[lihaoyang-amd]

what's the relationship between this and #16804 ?

Aha, maybe we are competitive. I am 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