Show input and output length on vLLM dashboard

[huydhn]

Fixes #6974

Input and output lengths are new dimensions on the dashboard that needs to be displayed after pytorch/pytorch-integration-testing#42. This PR also cleans up some old TODO code path for vLLM dashboard.

Testing

Different input and output lengths are showing up correctly now with their benchmark results on the preview

[vercel]

@huydhn is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Project	Deployment	Preview	Updated (UTC)
torchci	✅ Ready	Preview	Aug 12, 2025 7:42pm

[BoyuanFeng]

Thanks for the pr!

Is it possible to add in/out to the name? E.g., "meta-llama/Llama-4-Scout-17B-16E-Instruct / tp4 / qps_10 / 200 / 200" -> "meta-llama/Llama-4-Scout-17B-16E-Instruct / tp4 / qps_10 / input-200 / output-200":

Also, curious what is the setting here?

[huydhn]

Thanks for the pr!

Is it possible to add in/out to the name? E.g., "meta-llama/Llama-4-Scout-17B-16E-Instruct / tp4 / qps_10 / 200 / 200" -> "meta-llama/Llama-4-Scout-17B-16E-Instruct / tp4 / qps_10 / input-200 / output-200":

Oh, they're already part of the name, you can see them there in the format / 200 / 200. But let me make them clearer with their prefix, i.e. / input_200 / output_200, otherwise, it's unclear which is which

[huydhn]

Also, curious what is the setting here?

For latency and thoughput benchmarks, input and output length aren't set directly, but from what I see input_len is 32 and output_len is 128, probably the default values somewhere, I could show them in the chart

[huydhn]

One last note here is that input and output lengths are set in latency and serving benchmarks, but not throughput. I try to look into the default that it uses, but there is none for the ShareGPT dataset https://github.com/vllm-project/vllm/blob/main/vllm/benchmarks/datasets.py#L402. Few other datasets like Random or Sonnet define input and output lengths, but not ShareGPT. I will keep these fields empty for throughput benchmark for now.