[Hardware][AMD] integrate aiter chunked prefill into vllm

[Zzz9990]

CMD: VLLM_TORCH_PROFILER_DIR=/mnt/raid0/sixifang/vllm/vllm_profile HIP_VISIBLE_DEVICES=4,5,6,7 VLLM_ROCM_USE_AITER=1 VLLM_USE_V1=1 vllm serve /models/models--amd--Meta-Llama-3.1-8B-Instruct-FP8-KV/snapshots/fa42f9a9105c545755fea25cf69f49ac8c8b40e1/ --tensor-parallel-size 4 --gpu-memory-utilization 0.9 --trust-remote-code --disable-log-requests --block-size 16 --max-model-len 32768 --dtype float16 --quantization fp8 --no-enable-prefix-caching --max-num-batched-tokens=8192

Performance with aiter:

prompts	isl	osl	con	req_throughput	median_e2e	median_ttft	median_tpot	median_itl	output_tps	total_tps
1280	2048	1024	128	2.18	58777.36	1041.87	56.22	42.28	2236.56	6707.49

Performance without aiter:

prompts	isl	osl	con	req_throughput	median_e2e	median_ttft	median_tpot	median_itl	output_tps	total_tps
1280	2048	1024	128	0.94	135871.09	2621.44	129.89	94.26	964.17	2891.56

[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.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Zzz9990.

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

[tjtanaa]

@Zzz9990 What is the difference with this PR? #17710
Could you rename the PR to be semantically related to the feature you are adding? "integrate aiter into vllm" is too broad and generic

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Zzz9990.

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

[HAIAI]

Let's get this merged soon. cc @houseroad @WoosukKwon @simon-mo

[houseroad]

Accept to unblock. Since only touch AMD related logic, should be safe on other platform.

Could @HAIAI or someone from AMD approve this PR?

[HAIAI]

LGTM

[HAIAI]

@Zzz9990 would you take a look at failed checks?

[houseroad]

https://buildkite.com/vllm/ci/builds/22190#01977cbb-a80d-4f96-93d8-a821ca8095cf failure seems related? cc: @Zzz9990

[houseroad]

ImportError: cannot import name 'rocm_aiter_fused_add_rms_norm' from 'vllm.model_executor.layers.layernorm' (/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/layernorm.py)

[Zzz9990]

ImportError: cannot import name 'rocm_aiter_fused_add_rms_norm' from 'vllm.model_executor.layers.layernorm' (/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/layernorm.py)
Done

[gshtras]

Please note all the failing AMD checks. Updating from the latest should help fix the build to get started.

[HAIAI]

@Zzz9990 @fsx950223 can you update/rebase from main?

[SageMoore]

Looks reasonable. If you can delete some more of the cascade attention code in the builder/metadata that would be great. If not, it's no big deal.

[SageMoore]

AFAICT you need the use_cascade_attention method for API compatibility but everything else should be able to be removed. If not it's fine. No worries.