[CI Failure] Fix backend selection for encoder-only models

[hl475]

Purpose

After #24794, encoder-only models (e.g., BERT) fail to initialize because the TRITON_ATTN backend is selected by default, but it doesn't support encoder self-attention, causing:

NotImplementedError: Encoder self-attention and encoder/decoder cross-attention are not implemented for TritonAttentionImpl

This PR implemented an opt-in approach for attention type support:

1. Added supports_attn_type() method to AttentionBackend:
   - Default behavior: Only supports DECODER attention
   - Backends must explicitly override to support ENCODER_ONLY or other attention types
   - This makes the system safe by default - new backends won't accidentally support encoder-only models
2. Propagated attn_type through the backend selection pipeline:
   - Added attn_type parameter to get_attn_backend() and validate_configuration()
   - Modified EncoderOnlyAttention to pass attn_type=AttentionType.ENCODER_ONLY
   - Platform classes now validate attention type compatibility during backend selection
3. Explicitly marked the 3 backends that support encoder-only models:
   - FlexAttention: Supports DECODER + ENCODER_ONLY
   - FlashAttention: Supports DECODER + ENCODER_ONLY
   - CPU/TorchSDPA: Supports all attention types

Test Plan

pytest -s -v tests/models/language/pooling/test_token_classification.py::test_bert_models[float-boltuix/NeuroBERT-NER]

Test Result

1 passed, 4 warnings in 18.16s

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[DarkLight1337]

cc @MatthewBonanni @LucasWilkinson

[MatthewBonanni]

Down the road I'd like to make AttentionType an enum, but this LGTM!

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @hl475.

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

[MatthewBonanni]

Which backends actually do support encoder self-attention? Want to make sure this doesn't just kick over to another backend that doesn't support it and continue failing the tests. Please also make sure to run the previously-failing CI tests if they aren't triggered automatically

[LucasWilkinson]

Overall looks good; thanks for fixing this! please rebase

[hl475]

rebase

[hl475]

Thanks @LucasWilkinson and @MatthewBonanni for reviewing!

Just rebased my PR. Could you folks please help add the ready label (I don't have the permission) so I can run all previous failing CIs, thanks!

[hl475]

Which backends actually do support encoder self-attention? Want to make sure this doesn't just kick over to another backend that doesn't support it and continue failing the tests.

DONE!I will check this and maybe come up with some additional changes!

[mgoin]

I think we can probably ignore my comments for now, but we should consider them in followup. Probably Matt can tackle that if you don't have time

[hl475]

I think we can probably ignore my comments for now, but we should consider them in followup. Probably Matt can tackle that if you don't have time

oops, sorry just saw your comment @mgoin - but I changed PR based on your comments (thanks)!

I am OK either way. Please let me know and then I can start to run previous failing CIs!

[mgoin]

What are the attention backends that support running ENCODER and ENCODER_DECODER? I don't see them mentioned anywhere cc @russellb @NickLucche

[hl475]

What are the attention backends that support running ENCODER and ENCODER_DECODER?

Regarding this, not sure if I get you correctly but this PR focuses on fixing ENCODER_ONLY model support. I will defer this questions to others. From my understanding,

• AttentionType.ENCODER - CPU supports it; FlashAttention supported it via _forward_encoder_attention; FlexAttention does not support it
• AttentionType.ENCODER_DECODER - None of the v1 backends currently support this ?

[russellb]

• AttentionType.ENCODER_DECODER - None of the v1 backends currently support this ?

flash_attn supports ENCODER_DECODER.

flashinfer would support it with this change: #25098

[hl475]

I manually triggered all 3 previously failing CIs due to NotImplementedError: Encoder self-attention and encoder/decoder cross-attention are not implemented for TritonAttentionImpl in this PR

Language Models Test (Extended Pooling) - https://buildkite.com/vllm/ci/builds/38812/steps/canvas?jid=019a7c6c-315c-418a-8273-e5b946fbac0f
Language Models Test (MTEB) - https://buildkite.com/vllm/ci/builds/38812/steps/canvas?jid=019a7c6c-315f-4f59-8fb6-bbc83f129594
Multi-Modal Models Test (Extended) 1 - https://buildkite.com/vllm/ci/builds/38812/steps/canvas?jid=019a7c6c-3166-424b-a9af-aa70e5fadf08