[ROCm] Integrate AITER PagedAttention

[SageMoore]

This PR is largely cut out from #14007 with a few modifications.

1. Added the AiterPagedAttention class that's a subclass of PagedAttention. This removes a significant amount of code duplication.
2. Moved the AITER environment variable checks from the platform class down to the rocm attention backend. We could likely clean this up even further, but I think this is fine for now.
3. Disabled AITER paged attention on non 8-bit data types. It was previously only enabled for 16 bit models with one kv head. I tried running lm_eval with falcon 7b and got a score of 0 so it seems that there are some issues here. The set of supported models is so small that I think it's safe just disable the kernel for bf16/fp16.
4. Added a fallback mechanism that will just go through the original paged attention kernel when we detect at runtime that the AITER kernel is unsupported.

[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.

🚀

[hongxiayang]

cc @tjtanaa

[tjtanaa]

@SageMoore
Can you enable module level and model level unit tests for AITER kernel as well?
Module level is like checking if when AITER is invoked, is it calling AITER kernel as expected and fallback is working as expected, using something like a dispatcher function?
Model level unit tests would be referring to
tests/quantization/test_fp8.py . Refer to https://github.com/vllm-project/vllm/pull/14007/files#diff-1752a3edce0e5a3cce7225c9a1d4b06738ef5b624c24f57096882693134888e2 (to ensure FP8 Paged Attention works end-to-end)
and
decoder_only/language/test_models.py https://github.com/vllm-project/vllm/pull/14007/files#diff-99ded1cb38946be58b246c0fd8766bffb39b8c17154f43d3de4592c2cebb1232 (to ensure fallback works end-to-end)

I am working on enabling other AITER kernels AITER Linear, AITER Fused MoE, AITER RMSNorm, and AITER Block Scaled GEMM in other PRs.

In the process, I will be enabling model level unit tests would be referring to
tests/quantization/test_fp8.py . Refer to https://github.com/vllm-project/vllm/pull/14007/files#diff-1752a3edce0e5a3cce7225c9a1d4b06738ef5b624c24f57096882693134888e2 (to ensure FP8 Paged Attention works end-to-end)
and
decoder_only/language/test_models.py https://github.com/vllm-project/vllm/pull/14007/files#diff-99ded1cb38946be58b246c0fd8766bffb39b8c17154f43d3de4592c2cebb1232 (to ensure fallback works end-to-end)

I will also be updating the Dockerfile.rocm_base with AITER setup and proper tested commit.

It would be preferrable to merge this PR after we at least get our first AITER PR merged with AITER model level unit test enabled, and proper Dockerfile.rocm_base that creates Docker environment that has AITER for the unit test.

[houseroad]

do we really fallback here? Only see the warning here.

[mergify]

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

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