Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[V1] Refactor LLMEngine To Use Multiprocessing #9741

Closed
wants to merge 26 commits into from
Closed

[V1] Refactor LLMEngine To Use Multiprocessing #9741

wants to merge 26 commits into from

Conversation

robertgshaw2-redhat
Copy link
Collaborator

@robertgshaw2-redhat robertgshaw2-redhat commented Oct 27, 2024 •
edited
Loading

SUMMARY:

Run with:

from vllm import LLM

model = LLM("Qwen/Qwen2-0.5B-Instruct", enforce_eager=True)
output = model.generate("Hello my name is")
print(output)

# Need to call this, else hangs. TODO: fix
del model

AsyncLLMEngine Plan

  • Per request asyncio queues

  • On generate():

    • Create new queue for this request
    • Add new request to the LLMEngine
    • Stream from the queue

  • Output Handling Option A) Works like MQLLMEngineClient

    • output_handler which calls the LLMEngine.step() in a loop
    • Get back gets back list of RequestOutputs from Detokenizer.step()
    • Loop through RequestOuput, pushes into the per request generator queue

  • Output Handling Option B) Avoid a loop << leaning towards this option

    • Avoid an output_handler loop by pushing the Queue into the Detokenizer. When we call Detokenizer.add_request() (which is called by AsyncLLMEngine.generate()), pass the per request queue into the Detokenizer
    • Final action in the inner loop in the Detoknizer.step() puts the RequestOutput into the queue
    • Challenge → how does this work with LLM? Might need two Detokenizer modes.

cc @njhill @WoosukKwon

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 GitHub Actions
Copy link

👋 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-redhat robertgshaw2-redhat mentioned this pull request Oct 28, 2024
Closed
10 tasks
robertgshaw2-redhat
robertgshaw2-redhat commented Oct 28, 2024
View reviewed changes
vllm/v1/engine/detokenizer.py
output_kind=request.output_kind,
request_output=request_output)

@staticmethod
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Make this a @classmethod for RequestOutput

@robertgshaw2-redhat
Copy link
Collaborator Author

moving to #9782

@robertgshaw2-redhat robertgshaw2-redhat deleted the rs-prototype-2 branch October 29, 2024 00:41
@njhill
Copy link
Member

njhill commented Oct 29, 2024

@robertgshaw2-neuralmagic some high level thoughts (not reviewed in detail yet):

My thinking was that for performance-sensitive usage, we need to own/isolate the critical loop and as such it doesn't really make sense for folks to integrate using the current LLMEngine.add_request/step API because this implies they own that loop.

I did think about something like you've implemented here, i.e. have add_request dispatch the request so that it will run immediately and step is then more like a synchronous iterator to get back the results of all in-progress requests (although I had assumed it would be blocking rather than non-blocking as you have it rn). However this does change the semantics because folks might be doing things where they make decisions about requests to add/remove between each step, and/or they may be wanting to create a batch of requests up-front to run all together.

So if we want to keep this interface for backwards compatibility (maybe we don't? but I think @WoosukKwon wants this) then my thinking was to just to not have it do the detokenization asynchronously (since again we will say that this is not the API to use for max performance anyhow). When I said "wrap" I meant just wrapping in a method in the same process.

So then what is currently LLMEngine we would effectively rename to LLMEngineCore (or whatever) and then LLMEngine can subsume it either as a subclass or via an instance field. Just the public add_remove() and step() methods would be implemented in LLMEngine and these would do the tok/detok inline (synchronously), i.e. after calling the _step() method on the engine core.

It would be just the LLM and AsyncLLM classes where we start LLMEngineCore as a separate process running the _step() loop itself along with the zmq queues. These would essentially just expose the a generate method (either sync/async). But we could extend them to be a bit more flexible by:

  • Allowing multiple inputs in a single generate call so that they'll be guaranteed to all get added to the batch together (assuming it has space)
  • For LLM generate we could have a synchronous streaming option where an iterator is returned
  • Possibly even methods providing semantics similar to the "breaking" ones discussed above where you can call an add_request method at any time to add new requests but there's a single-consumer method for iteratively receiving results of all running requests

And then would suggest anyone implementing a custom front-end to just use either LLM or AsyncLLM. AsyncLLMEngine I suggest we deprecate, it could effectively be an alias for AsyncLLM in the meantime (though we need to decide about the PP case).

The same Detokenizer class can be used across all of these of course (including the inline LLMEngine.step() case)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@WoosukKwon WoosukKwon Awaiting requested review from WoosukKwon WoosukKwon is a code owner

@njhill njhill Awaiting requested review from njhill njhill is a code owner

@ywang96 ywang96 Awaiting requested review from ywang96 ywang96 will be requested when the pull request is marked ready for review ywang96 is a code owner

@comaniac comaniac Awaiting requested review from comaniac comaniac will be requested when the pull request is marked ready for review comaniac is a code owner

@alexm-redhat alexm-redhat Awaiting requested review from alexm-redhat alexm-redhat will be requested when the pull request is marked ready for review alexm-redhat is a code owner

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@robertgshaw2-redhat @njhill