Support BERTModel (first encoder-only embedding model)

[robertgshaw2-neuralmagic]

SUMMARY:

• built on top of [Model] Bert Embedding Model #5447 (co-author: laishzh)
• introduces Encoder-Only models (BERT - tested with https://huggingface.co/BAAI/bge-base-en-v1.5 to match sentence-transformers implementation)
• introduces Encoder-Only attention for XFORMERS backend.

Note: this PR requires setting the VLLM_ATTENTION_BACKEND=XFORMERS variable to run. We throw a loud error if it is not set.

TODO:

• add end-to-end testing with MTEB
• test tp>1

FOLLOW UPS:

• refactor to extract some of logic out of XFORMERS. The logic is very complex right now.
• support other attention backend

FIX #5179

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Adding or changing kernels

Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.

• Make sure custom ops are registered following PyTorch guidelines: Custom C++ and CUDA Operators and The Custom Operators Manual
• Custom operations that return Tensors require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.
• Use torch.libary.opcheck() to test the function registration and meta-function for any registered ops. See tests/kernels for examples.
• When changing the C++ signature of an existing op, the schema must be updated to reflect the changes.
• If a new custom type is needed, see the following document: Custom Class Support in PT2.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[robertgshaw2-neuralmagic]

We need a follow up to clean up the XFormersAttention

[laishzh]

I see both #5447 and this PR are being worked on. Which one should we review?

@DarkLight1337 #5547 has some issues. This PR is built on top of #5547. I do not want to use the encoder-decoder runners for embedding models, so I believe this design is better + matches what sentence-transformers gives.

@laishzh - Im happy to move this PR under your github account or at minimum add you as a co-author.

Totally agree with you that it's better design! Also co-author works for me. Thanks for the reply.

[DarkLight1337]

I can't say to the implementation of the model (don't have time to compare it against HF), but nothing really jumps out at me from a brief look. If the tests pass then it's probably fine.

[robertgshaw2-neuralmagic]

I can't say to the implementation of the model (don't have time to compare it against HF), but nothing really jumps out at me from a brief look. If the tests pass then it's probably fine.

I have a test that compares against the hf implementation (via sentence-transformers)

[pooyadavoodi]

nit: first_token_flat_indices -> cls_token_flat_indices

[maxdebayser]

I've tested it with BAAI/bge-base-en-v1.5 and get results close to the MTEB benchmark.

[maxdebayser]

In the transformers code for this model I don't see normalization being applied: https://github.com/huggingface/transformers/blob/d00f1ca860f19f4c0962882e56044bb6ef7b5626/src/transformers/models/bert/modeling_bert.py#L743
How do we know if it has to be applied or not?

[robertgshaw2-neuralmagic]

I followed the sentence-transfomers implementation, which is used in MTEB. The baseline transformers model uses a different pooling strategy completely

[maxdebayser]

In transformers, if position_ids is None, self.position_ids is used. Should we do the same?

https://github.com/huggingface/transformers/blob/d00f1ca860f19f4c0962882e56044bb6ef7b5626/src/transformers/models/bert/modeling_bert.py#L196

[robertgshaw2-neuralmagic]

I think position_ids should never be None for us. But I should double check this

[DarkLight1337]

Overall LGTM - remember to add this model to the Supported Models page before you merge.

[DarkLight1337]

Is it intentional to not include this on the docs yet?

[robertgshaw2-neuralmagic]

Is it intentional to not include this on the docs yet?

Yes I think we should get #9388 in before we advertise this