[LoRA] Change lora_tokenizers capacity

[xyang16]

Currently lora_tokenizers capacity is max_num_seqs. This is fine if max_loras is smaller than max_num_seqs.

But if max_loras is greater than max_num_seqs, lora_tokenizers should have capacity=max_loras, just like active_adapters LoRALRUCache have capacity=max_loras. If not, items in active_adapters are not evicted, but items in lora_tokenizers are evicted. I think this is unnecessarily evicted, causing performance degradation.

In our benchmark, we test max_loras=50, max_num_seqs=32, 50 adapters are invoked evenly distributed. We found this causing 50% performance degradation in P90 latency, comparing to tokenizers are not evicted.

So raising this PR to make lora_tokenizers capacity to be max(max_loras, max_num_seqs).

Another option is to add max_lora_tokenizers, just like max_loras and max_cpu_loras which controls GPU cache and CPU cache. But this might be an overkill.

[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 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 do one of these:

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

🚀

[jeejeelee]

Thanks for your contribution, I will look at this PR tomorrow or the day after tomorrow

[jeejeelee]

Thank you very much for your contribution. Could we pass max_loras as a keyword argument? This way the overall changes would be minimal.

[xyang16]

Thanks for your contribution, I will look at this PR tomorrow or the day after tomorrow

@jeejeelee Thanks for reviewing. There's already **tokenizer_config https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/tokenizer_group/tokenizer_group.py#L18, you mean add to **tokenizer_config?

[jeejeelee]

Thanks for your contribution, I will look at this PR tomorrow or the day after tomorrow

@jeejeelee Thanks for reviewing. There's already **tokenizer_config https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/tokenizer_group/tokenizer_group.py#L18, you mean add to **tokenizer_config?

Yeah

[xyang16]

Thanks for your contribution, I will look at this PR tomorrow or the day after tomorrow

@jeejeelee Thanks for reviewing. There's already **tokenizer_config https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/tokenizer_group/tokenizer_group.py#L18, you mean add to **tokenizer_config?

Yeah

OK

[xyang16]

@jeejeelee I have updated the code, please review.

[jeejeelee]

overall LGTM, thanks for your contribution and patience