[V1] Prototype Fully Async Detokenizer

[robertgshaw2-neuralmagic]

SUMMARY:

• In current v1 prototype, we create RequestOutput objects in the LLMEngine based on the value of step() and send them back to the Client. This has implications:
  • The Detokenizer runs in a separate process, step() may return an empty list as the Detokenizer can fall behind. This means that we need to maintain state about how many steps behind each request is (num_lagged_steps object), which complicates the code
  • Creating RequestOutputs and sending the RequestOutputs back to the MQLLMEngine run in the main LLMEngine process, increasing contention.
• PR makes the Detokenizer fully async by using the Detokenizer as the message sender (sends the RequestOutputs to the MQLLMEngineClient --> this avoids needing to send the outputs of the Detokenizer back to the LLMEngine (reducing the complexity of the LLMEngine and reducing overhead).

ARCH SUMMARY:

TODO:

• Get agreement from group that this strategy makes sense
• Figure out how to make it work with LLM
• Guarantee there is no race condition for sending a new request vs sending new tokens
• Figure out if it is okay for the MQLLMEngine and Detokenizer to connect and PUSH to the same socket.
• Propogate abort to the Detokenizer
• Shutdown for Detokenizer
• Error handling for Detokenizer
• Decide what to do with AsyncLLMEngine
• Optimize the message passing copy() like we did for the MQLLMEngine
• Performance benchmark

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.

🚀

[robertgshaw2-neuralmagic]

We reduce overhead because:

• We now create the RequestOutputs in the Detokenizer
• We now send the RequestOutputs to the MQLLMEngineClient in the Detokenizer, rather than from the MQLLMEngine (which is in the same process as the LLMEngine)

[robertgshaw2-neuralmagic]

This breaks the API, as step no longer returns RequestOutput.

However, I will note that this is already how we use the LLMEngine in the MQLLMEngine with async_process_outputs (the return value of step is [ ])

[robertgshaw2-neuralmagic]

The key benefit of this PR in terms of code complexity is that we simplify the LLMEngine by no longer needing to keep track of these data.

[robertgshaw2-neuralmagic]

I think this should be okay, but I want to double check. If this is a problem, we can convert this to a separate "error" socket and poll on the error and output socket from the MQLLMEngineClient

[njhill]

Thanks @robertgshaw2-neuralmagic! Actually @WoosukKwon and I were discussing something quite similar to this last week and I had started writing up a similar proposal ... I'd just been thinking some more about the best way to structure it to retain backwards compatibility with the various existing interfaces.

Here is what I propose:

• We have the core step logic in an LLMEngine._step method which returns only token ids in a compact form
• The LLMEngine.step method can wrap this and perform the detokenization and stop string evaluation inline, returning the same List[RequestOutput] for backwards compatibility in case folks are using this directly (though we would discourage it)
• We rework the LLM class to actually start the LLMEngine in a separate process which runs a loop invoking the _step method, connected with zmq, basically the same thing as MQLLMEngine. In this case tokenization and detokenization are done on the client side.
• We add a new AsyncLLM class which is the same thing just with an async interface. Again pretty much the same as MQLLMEngine except that we do the tokenization/detokenization on the client size, and the messages that we send internally are more compact (not RequestOutput objects).

So the main difference is that we don't have a dedicated process for detokenization, we just do it on the client side. The important thing is that we isolate the critical loop in its own process. And we don't have an explicit MQLLMEngine, this just becomes the standard behaviour when you use the LLM or AsyncLLM interfaces.

WDYT?

[robertgshaw2-neuralmagic]

Thanks @robertgshaw2-neuralmagic! Actually @WoosukKwon and I were discussing something quite similar to this last week and I had started writing up a similar proposal ... I'd just been thinking some more about the best way to structure it to retain backwards compatibility with the various existing interfaces.

Here is what I propose:

• We have the core step logic in an LLMEngine._step method which returns only token ids in a compact form
• The LLMEngine.step method can wrap this and perform the detokenization and stop string evaluation inline, returning the same List[RequestOutput] for backwards compatibility in case folks are using this directly (though we would discourage it)
• We rework the LLM class to actually start the LLMEngine in a separate process which runs a loop invoking the _step method, connected with zmq, basically the same thing as MQLLMEngine. In this case tokenization and detokenization are done on the client side.
• We add a new AsyncLLM class which is the same thing just with an async interface. Again pretty much the same as MQLLMEngine except that we do the tokenization/detokenization on the client size, and the messages that we send internally are more compact (not RequestOutput objects).

So the main difference is that we don't have a dedicated process for detokenization, we just do it on the client side. The important thing is that we isolate the critical loop in its own process. And we don't have an explicit MQLLMEngine, this just becomes the standard behaviour when you use the LLM or AsyncLLM interfaces.

WDYT?

I agree with the direction you described

[robertgshaw2-neuralmagic]

Closing in favor of #9741