[Core] Reduce unnecessary compute when logprobs=None #6532
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well. To run full CI, you can do one of these:
🚀 |
peng1999
changed the title
| if num_logprobs is None and not use_beam_search: | ||
| for next_token_id in next_token_ids: | ||
| # Use a dummy logprob | ||
| sampled_logprobs.append({next_token_id: Logprob(0.0)}) |
There was a problem hiding this comment.
Is 0 a value that makes sense for a logprob? We are still using real token ids here, so maybe 1.0 would be more representative?
There was a problem hiding this comment.
What do you think using NaN here? The value will propagate to cumulative_logprobs. I think it's better to show a NaN than an arbitrary positive number here.
The logprobs are already omitted if logprobs=None. So this API is not changing. Line 124 in 1689219 The changed part is |
| if num_logprobs is None and not use_beam_search: | ||
| for next_token_id in next_token_ids: | ||
| # Use a dummy logprob | ||
| sampled_logprobs.append({next_token_id: Logprob(0.0)}) |
|
It seems that some tests failed when I set the cumulative_logprobs to NaN because NaN != NaN. How should I deal with this? |
It would be a bit tricky: https://stackoverflow.com/questions/13003202/python-nan-nan I'd suggest one of the following:
|
|
1 sounds like a good solution to me! can you actually update the sampling parameter docstring? |
github-actions
bot
added
the
ready
Head branch was pushed to by a user without write access
|
I changed the dummy value from nan to inf for better comparability. The value will be discard eventually and end users will get both logprobs and cumulative_logprobs=None. I think the inf value is reasonable as an internal representation. Not using The failed test seems to be irrelevant to this PR. |
* upstream/main: (66 commits) [Bugfix] Fix PaliGemma MMP (vllm-project#6930) [TPU] Fix greedy decoding (vllm-project#6933) [Kernel] Tuned int8 kernels for Ada Lovelace (vllm-project#6848) [Kernel] Fix marlin divide-by-zero warnings (vllm-project#6904) [ci] GHA workflow to remove ready label upon "/notready" comment (vllm-project#6921) [Kernel] Remove unused variables in awq/gemm_kernels.cu (vllm-project#6908) [Frontend] New `allowed_token_ids` decoding request parameter (vllm-project#6753) [Bugfix] Allow vllm to still work if triton is not installed. (vllm-project#6786) [TPU] Support tensor parallelism in async llm engine (vllm-project#6891) [Kernel] Fix deprecation function warnings squeezellm quant_cuda_kernel (vllm-project#6901) [Core] Reduce unnecessary compute when logprobs=None (vllm-project#6532) [Kernel] Tuned FP8 Kernels for Ada Lovelace (vllm-project#6677) [Model] Initialize support for InternVL2 series models (vllm-project#6514) [Misc] Pass cutlass_fp8_supported correctly in fbgemm_fp8 (vllm-project#6871) Add Nemotron to PP_SUPPORTED_MODELS (vllm-project#6863) [Kernel] Increase precision of GPTQ/AWQ Marlin kernel (vllm-project#6795) [TPU] Reduce compilation time & Upgrade PyTorch XLA version (vllm-project#6856) [Docs] Add RunLLM chat widget (vllm-project#6857) [Model] Initial support for BLIP-2 (vllm-project#5920) [CI/Build][Doc] Update CI and Doc for VLM example changes (vllm-project#6860) ...
Signed-off-by: Alvant <alvasian@yandex.ru>
This PR refactors
_get_logprobsfunction in sampler module, to get rid of some unnecessary CPU and GPU works whenlogprobs=NoneinSamplerParams.Tested using A100 80G GPU, running Qwen2-0.5B-Instruct and batch_size=2048. The latency of
_get_logprobsreduced from ~30ms to ~5ms.Before:
After:
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,LLMclass, 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:
format.shto format your code.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-requiredand 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:
action-requiredlabel on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.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!