[V1] Decouple GPU and TPU InputBatch

[afeldman-nm]

Purpose

Decouple TPU and GPU InputBatch implementations. This PR adds a new tpu_input_batch.py, which is identical to gpu_input_batch.py but with all but TPU-oriented logic stripped out. Rationale: (1) TPU code lags GPU code in terms of feature support, so it makes sense to allow GPU to have a separate input batch from TPU to support more advanced features; (2) This PR is a prerequisite for the new logitsprocs programming model introduced in #16728 , which makes InputBatch modifications that are incompatible with TPU.

Test Plan

Covered by existing GPU and TPU unit tests.

Test Result

N/A

(Optional) Documentation Update

N/A

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @afeldman-nm, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a foundational architectural change by decoupling the InputBatch data structure implementations for TPU and GPU. This separation is a prerequisite for integrating a new logitsprocs programming model, which necessitates InputBatch modifications that are incompatible with TPU. The changes involve creating a dedicated TPU-specific InputBatch class and updating relevant modules to conditionally use the appropriate batch implementation based on the hardware platform.

Highlights

• Dedicated TPU InputBatch: A new file, vllm/v1/worker/tpu_input_batch.py, has been introduced. This file contains a distinct InputBatch class implementation specifically tailored for operations on Tensor Processing Units (TPUs).
• GPU-Specific InputBatch Clarification: The existing vllm/v1/worker/gpu_input_batch.py has been updated with a clarifying comment, explicitly stating that it defines the InputBatch for GPU, thereby formalizing the separation of concerns between GPU and TPU batch implementations.
• Dynamic InputBatch Selection: The vllm/v1/worker/lora_model_runner_mixin.py module now includes conditional import logic. It dynamically imports either the TPUInputBatch or GPUInputBatch based on the detected current_platform.device_name, ensuring that the correct batch structure is utilized at runtime depending on the hardware platform.
• TPU Component Alignment: Files within the TPU-specific paths, namely vllm/v1/sample/tpu/metadata.py and vllm/v1/worker/tpu_model_runner.py, have been updated to exclusively import and utilize the newly defined tpu_input_batch.InputBatch. This change aligns TPU-specific components with their dedicated batch implementation, reinforcing the decoupling.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a dedicated InputBatch implementation for TPU, decoupling it from the GPU version. It involves creating a new tpu_input_batch.py file and updating imports and logic in related files (metadata.py, lora_model_runner_mixin.py, tpu_model_runner.py). The changes are a necessary step towards a TPU-specific programming model. The review identified several high-severity issues related to potential correctness problems in sampling metadata handling, generator state management, and logprobs calculation, as well as a duplicate code block. A medium-severity issue regarding an undocumented assumption in the condense method was also found.

[njhill]

Thanks @afeldman-nm LGTM.

This PR adds a new tpu_input_batch.py, which is identical to gpu_input_batch.py but with all but TPU-oriented logic stripped out.

presumably you meant something like "but with non-TPU-supported logic stripped out"?

[njhill]

Thanks @afeldman-nm.

As discussed we can look at how to reconcile this again once the LogitsProcessor rework is done (and hopefully support LPs with TPU somehow).

[afeldman-nm]

Thanks @afeldman-nm.

As discussed we can look at how to reconcile this again once the LogitsProcessor rework is done (and hopefully support LPs with TPU somehow).

Certainly. To be clear, vLLM's TPU logic currently supports certain hard-coded logits processors; the gap which needs to be filled is specifically to (1) support the new logitsproc programming model introduced in #16728 , and (2) support logitsproc extensibility. Basically, to have feature parity with GPU.

[yaochengji]

LGTM, thanks for separating TPU and GPU on InputBatch!

[afeldman-nm]

Thanks for the reviews and merge @njhill @yaochengji @aarnphm