[Paged KV] Add ShareGPT avg generation length tool#127
Merged
LxYuan0420 merged 3 commits intovllm-project:mainfrom Mar 2, 2026
Merged
[Paged KV] Add ShareGPT avg generation length tool#127LxYuan0420 merged 3 commits intovllm-project:mainfrom
LxYuan0420 merged 3 commits intovllm-project:mainfrom
Conversation
Signed-off-by: ran <hzz5361@psu.edu>
This was referenced Mar 1, 2026
Signed-off-by: ran <hzz5361@psu.edu>
WindChimeRan
changed the title
LxYuan0420
requested changes
Mar 2, 2026
Collaborator
LxYuan0420
left a comment
LxYuan0420
left a comment
There was a problem hiding this comment.
Few minor changes and I think we are good to merge; I like the direction here. Having small, repeatable benchmark/smoke scripts will really help us validate end-to-end behavior as paged attention evolves.
Nit (optional): could we put this under benchmarks/ instead of tools/ to match the repo layout?
Comment on lines
+7
to
+8
| huggingface-cli download anon8231489123/ShareGPT_Vicuna_unfiltered \ | ||
| --repo-type dataset --local-dir . ShareGPT_V3_unfiltered_cleaned_split.json |
Collaborator
There was a problem hiding this comment.
Can we update the example command to use hf download ... because hugging face-cli download is deprecated
| max_model_len=max_model_len, | ||
| max_num_seqs=max_num_seqs, | ||
| ) | ||
| sampling_params = SamplingParams(temperature=0, seed=42, max_tokens=max_tokens) |
Collaborator
There was a problem hiding this comment.
the 42 is hardcoded value ; should be wired to use --seed
LxYuan0420
approved these changes
Mar 2, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add a diagnostic tool for paged KV cache development. Runs offline vLLM inference on ShareGPT prompts and reports response length statistics (mean/std tokens).
The intended workflow is to compare the non-paged (standard MLX cache) and paged (Metal kernel) paths: if a KV cache bugfix improves alignment between the two distributions, that's a strong signal the fix is correct.
Also useful for comparing across batch sizes (--max-num-seqs 1 vs 8) to verify batched decode consistency. Related: #119
I learnt this from https://arxiv.org/pdf/2601.11580 table 1.
relevant post from thinking machine lab: https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
Usage
VLLM_METAL_USE_PAGED_ATTENTION=1 VLLM_METAL_MEMORY_FRACTION=0.7 \ python tools/avg_gen_length.py --max-num-seqs 1 8