[wip/spmd] Serialization Optimization

[rkooo567]

• Send delta to reduce the payload size to send to workers
• Use msgspec.msgpack. This requires to change existing data classes to msgspec compatible data structure. I found this has 10~20% e2e throughput improvement over raw pickle when running benchmark_throughput.py with opt-175m

---

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.

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 consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[njhill]

@rkooo567 does this need to be spmd-specific? It would be good to move to deltas in all cases

[rkooo567]

@rkooo567 does this need to be spmd-specific? It would be good to move to deltas in all cases

yeah right now, it is spmd specific. I think the goal is to migrate to spmd architecture, so maybe we should just focus on making it work on spmd?

[njhill]

I just mean there is nothing inherent about this optimization that means it couldn't just be made generally rater than just for spmd? I thought we were still evaluating spmd to some extent, imo there are some tradeoffs that would still be good to understand better. I just thought better to minimize the special case paths and do this across the board...

[rkooo567]

I just mean there is nothing inherent about this optimization that means it couldn't just be made generally rater than just for spmd? I thought we were still evaluating spmd to some extent, imo there are some tradeoffs that would still be good to understand better. I just thought better to minimize the special case paths and do this across the board...

So given we have 2 archs co-existing now (spmd and driver as worker that uses nccl to broadcast metadata), I think this optimization is only applied to spmd. the driver-as-worker and spmd sends different set of data to "workers". I can technically make it work with driver-as-worker, but this has 0 benefit with it (or we should cache different data that's prepared and broadcasted after prepare_input).

[youkaichao]

Thanks for the hard working.

I think there are still some dispute around how we deal with the serialization optimization.

One consensus is the current architecture is slow. And there are two possible ways to solve it:

• sending diff of the data every step
• just send the request from the begining, and the data lives in each worker, so we don't need to send the diff any more.

It is not determined yet. Personally I feel the second solution might be better, since we don't need to manually diff the object (which can be quite difficult to understand).

[rkooo567]

@youkaichao good point. I actually considered the second approach (sending seq group to workers as soon as it is accepted, and update states inside a worker), but there are some gaps I found;

• block tables are decided inside scheduler, so sending this is necessary unless we move this to workers.
• For new token update, the beam search logics are pretty complicated, and i didn't know how good idea it is to duplicate that logic in the worker layer.

[youkaichao]

You are correct. And there's also ongoing discussion, on whether we should move scheduler and block manager into every worker. As long as they are deterministic, their output should be the same in all workers, so we don't need to communicate.

This might be complicated when we consider pipeline parallel though.

@cadedaniel also has some thoughts about this.

[rkooo567]

Please see #7109