[Kernel] Support sliding window in flash attention backend

[heheda12345]

Flash attention backend provides sliding window support now. We can use it instead of fall back to xformer.

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!

[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.

🚀

[comaniac]

LGTM

[comaniac]

The CI failure seems to be a true error. Please fix it and ping me again and I'll unblock the rest CI tests

[heheda12345]

@comaniac I can pass this failed test in my local env. Can you take a look?
pytest -vs entrypoints/openai/test_chat.py::test_response_format_json_schema

[comaniac]

@comaniac I can pass this failed test in my local env. Can you take a look? pytest -vs entrypoints/openai/test_chat.py::test_response_format_json_schema

I'm not sure if this is a flaky test. Let me just retry first.

[heheda12345]

The failure may be related to this pr because the model it used contains sliding window. But the failure is that the model fails to generate a valid json string, though it can generate a reasonable response.

s = 'Here\'s the JSON object you requested:\n\n```json\n{\n "result": 2\n}\n```\n\nAlternatively, you can also write it o...o enclosed in double quotes unless it\'s a number, boolean, or null. The entire object is enclosed in curly braces {}.'

Not sure why the sampler fails to control the output format. Any suggestions for debugging?

[2024-10-16T18:19:26Z]     @pytest.mark.asyncio
[2024-10-16T18:19:26Z]     async def test_response_format_json_schema(client: openai.AsyncOpenAI):
[2024-10-16T18:19:26Z]         for _ in range(2):
[2024-10-16T18:19:26Z]             resp = await client.chat.completions.create(
[2024-10-16T18:19:26Z]                 model=MODEL_NAME,
[2024-10-16T18:19:26Z]                 messages=[{
[2024-10-16T18:19:26Z]                     "role":
[2024-10-16T18:19:26Z]                     "user",
[2024-10-16T18:19:26Z]                     "content": ('what is 1+1? please respond with a JSON object, '
[2024-10-16T18:19:26Z]                                 'the format is {"result": 2}')
[2024-10-16T18:19:26Z]                 }],
[2024-10-16T18:19:26Z]                 response_format={
[2024-10-16T18:19:26Z]                     "type": "json_schema",
[2024-10-16T18:19:26Z]                     "json_schema": {
[2024-10-16T18:19:26Z]                         "name": "foo_test",
[2024-10-16T18:19:26Z]                         "schema": {
[2024-10-16T18:19:26Z]                             "type": "object",
[2024-10-16T18:19:26Z]                             "properties": {
[2024-10-16T18:19:26Z]                                 "result": {
[2024-10-16T18:19:26Z]                                     "type": "integer"
[2024-10-16T18:19:26Z]                                 },
[2024-10-16T18:19:26Z]                             },
[2024-10-16T18:19:26Z]                         },
[2024-10-16T18:19:26Z]                     }
[2024-10-16T18:19:26Z]                 })
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z]             content = resp.choices[0].message.content
[2024-10-16T18:19:26Z]             assert content is not None
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z] >           loaded = json.loads(content)
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z] entrypoints/openai/test_chat.py:881:
[2024-10-16T18:19:26Z] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
[2024-10-16T18:19:26Z] /usr/lib/python3.12/json/__init__.py:346: in loads
[2024-10-16T18:19:26Z]     return _default_decoder.decode(s)
[2024-10-16T18:19:26Z] /usr/lib/python3.12/json/decoder.py:337: in decode
[2024-10-16T18:19:26Z]     obj, end = self.raw_decode(s, idx=_w(s, 0).end())
[2024-10-16T18:19:26Z] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z] self = <json.decoder.JSONDecoder object at 0x7fb04bfe6450>
[2024-10-16T18:19:26Z] s = 'Here\'s the JSON object you requested:\n\n```json\n{\n  "result": 2\n}\n```\n\nAlternatively, you can also write it o...o enclosed in double quotes unless it\'s a number, boolean, or null. The entire object is enclosed in curly braces {}.'
[2024-10-16T18:19:26Z] idx = 0
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z]     def raw_decode(self, s, idx=0):
[2024-10-16T18:19:26Z]         """Decode a JSON document from ``s`` (a ``str`` beginning with
[2024-10-16T18:19:26Z]         a JSON document) and return a 2-tuple of the Python
[2024-10-16T18:19:26Z]         representation and the index in ``s`` where the document ended.
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z]         This can be used to decode a JSON document from a string that may
[2024-10-16T18:19:26Z]         have extraneous data at the end.
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z]         """
[2024-10-16T18:19:26Z]         try:
[2024-10-16T18:19:26Z]             obj, end = self.scan_once(s, idx)
[2024-10-16T18:19:26Z]         except StopIteration as err:
[2024-10-16T18:19:26Z] >           raise JSONDecodeError("Expecting value", s, err.value) from None
[2024-10-16T18:19:26Z] E           json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)
[2024-10-16T18:19:26Z]
[2024-10-16T18:19:26Z] /usr/lib/python3.12/json/decoder.py:355: JSONDecodeError

[comaniac]

Looks like the JSON mode and constraint decoding is not applied. Since FlashAttention doesn't support sliding window before this PR, that test should use other attention backends such as xFormers, so the general question becomes why JSON mode doesn't work with FlashAttention with sliding window. You can check

1. For models w/o sliding window, does JSON mode work with FlashAttention?
2. If (1) is true, then you may need to dive in JSON mode implementation to debug this failure test.

[heheda12345]

The CI fail is caused by some bug in json_schema support of vllm (fixing it in #9530) and is triggered by the numerical error between xformer and flash attention.

61c56d4 This commit uses flash attention without sliding window, and pytest -vs entrypoints/openai/test_chat.py also fails on L4 GPU. It can prove that the failure is not related to sliding window but is caused by the difference between xformer and flash attention.

Hope we can pass the CI in this pr after #9530 . I'll update this branch after that pr is approved.

[heheda12345]

@comaniac Can you help me to enable the remaining tests?

[heheda12345]

@comaniac Tests are passed now.

[comaniac]

Thanks for the great work!