[Misc][Speculative Decoding] Improve top1_proposal output tensor initialization.

[ShangmingCai]

This PR improves top1_proposal output tensor initialization efficiency and code readability.

Details:
I have been following the integration progress of speculative decoding technology in vllm and reading the relevant source code recently. During the process, I found that the initialization of proposal output in Top1Proposer uses three torch.tensor().expand() ops. Considering the initialization cost, I wonder if this part can be implemented using operations such as torch.full() and torch.zero() for they have better efficiency.

After such modification, the initialization time cost of these output tensors can be reduced to about half of its origin value. Also, these three lines of code might have better readability.

cc: @cadedaniel

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!

[njhill]

@ShangmingCai the code was actually originally as you are proposing here, I recently changed it to use expand in #5368. Since these are empty "placeholder" tensors containing a single value, there is no need to allocate their full size. expand() doesn't involve any additional memory allocation, it's just a view.

I'm not even sure whether this code path would be followed very often.

After such modification, the initialization time cost of these output tensors can be reduced to about half of its origin value.

This is a bit surprising, do you have any (micro-) benchmark results?

[ShangmingCai]

@ShangmingCai the code was actually originally as you are proposing here, I recently changed it to use expand in #5368. Since these are empty "placeholder" tensors containing a single value, there is no need to allocate their full size. expand() doesn't involve any additional memory allocation, it's just a view.

I'm not even sure whether this code path would be followed very often.

After such modification, the initialization time cost of these output tensors can be reduced to about half of its origin value.

My bad. I think you are right that creating a view is more memory-efficient.

This is a bit surprising, do you have any (micro-) benchmark results?

I was testing and timing multiple steps of the proposal procedure forgetting to add some cuda synchronize in this function. Guess I was misled by the wrong results.
Thanks for your reply, I will close this PR.

[cadedaniel]

Thanks for looking at this!

by the way, this method and the other framework ops take some time that can be optimized. Cell R30 in this sheet has a table that shows the overhead. I estimate ~800µs - 1000µs can be removed by lowering these ops to the GPU.

[ShangmingCai]

@ShangmingCai the code was actually originally as you are proposing here, I recently changed it to use expand in #5368. Since these are empty "placeholder" tensors containing a single value, there is no need to allocate their full size. expand() doesn't involve any additional memory allocation, it's just a view.

I'm not even sure whether this code path would be followed very often.

After such modification, the initialization time cost of these output tensors can be reduced to about half of its origin value.

This is a bit surprising, do you have any (micro-) benchmark results?

Hello @njhill, after our last conversation, I fixed the timing synchronization and ran a micro benchmark. I found that although using expand() can reduce GPU memory cost, for small Tensors, the time overhead of torch.full() and torch.zeros() seems to be smaller than creating views through expand(). I am not sure about the reason behind it. Maybe Torch or Cuda has optimized enough for Tensor initialization. Yet when the batch size becomes large enough, the advantage of expand() starts to be reflected.

Under this circumstance, considering that the tensor sizes of proposal_tokens[batch_size, proposal_len] and proposal_lens_tensor[len(proposal_lens)] are very small (their initialization has nothing to do with self._vocab_size), I did another set of experiments, using expand() only for proposal_probs initialization, I found that this strategy can reduce the initialization time of proposal placeholder in all cases without hurting memory efficiency, although the reduction is certainly limited (~300µs - 1000µs for different batch sizes).

Anyway, just want to share the results and thank you for conveying the trick that creating a view for a placeholder.