[Kernel][Model] PagedAttention: Support custom attention bias for T5 model (1/2)

[NickLucche]

This contribution is part of 2 PRs attempting to add support for the T5 model.
Based on the great work done in #7366 (comment) and #3117.

It implements:

$$ Attention(Q, K, V) = \text{softmax}\left(\frac{QK^\top + A}{\sqrt{d_k}}\right)V $$

In particular, the following aims to be a more generalized addition (not tied to T5), enabling any
custom attention bias in the PagedAttention kernel.

While I am aware this introduces a fully materialized matrix (akin to what's done in xformers),
currently padded to max_seq_len, I believe the flexibility brought by allowing any custom bias
would still pay for itself when providing initial compatibility for yet unsupported models (such as T5).
I'd be happy to focus on performance optimization later (eg variable len bias matrix or flashattn-like IO optimization)
or even have yet another model-specific feature in the kernel for T5 to compute relative positional embeddings.

TODO work left for future contribs (ie extend to all platforms):

• ROCm attn bias support
• CPU attn bias support
• ..?

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[sfc-gh-phalama]

T5 uses relative bias, so I believe attn_bias_vec should be instead indexed with seq_len - token_idx - 1.

Suggested change

	qk += (attn_bias_vec != nullptr) ? attn_bias_vec[token_idx] : 0;
	const int custom_bias_idx = max(min(seq_len - token_idx - 1, padded_max_seq_len), 0);
	qk += (attn_bias_vec != nullptr) ? attn_bias_vec[custom_bias_idx] : 0;

I have tested it with T5 using similar implementation.

[sfc-gh-phalama]

While I am aware this introduces a fully materialized matrix (akin to what's done in xformers),
currently padded to max_seq_len

@NickLucche In case of T5, it can be limited to the maximum relative distance, which for T5 Large is 128. Anything beyond 128 just uses the last value from the bias tensor.

[NickLucche]

Thanks for the quick input @sfc-gh-phalama, personally I'd rather not have the attn bias implementation depend or be tied to T5/relative positional encoding mechanisms, otherwise we could probably implement the inline formula.

[NickLucche]

Fixed the kernel tests failing and rebased, not sure how related the rest of errors may be tbh..
https://buildkite.com/vllm/ci/builds/12022#01946b42-da66-4512-919d-c2115515eb6d/147-444

AssertionError: please set tensor_parallel_size (4) to less than max local gpu count (1)
looks more like a CI/node issue

[joerunde]

@NickLucche the gpu count issue might be fixed by this PR: #12111

[NickLucche]

CI failures still look unrelated to me :/

[NickLucche]

I believe this PR could also be useful in places like https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/mllama.py#L863, where we can pass in the custom attention mask as bias term A, while getting rid of the logic to handle kv cache in the model.