[Usage]: Profiling Prefill and Decode Phases Separately

[Msiavashi]

Your current environment

I'm attempting to independently measure the performance (e.g., latency, throughput, etc.) of the prefill and decode phases. Is there a way to achieve this? I have noticed a few benchmarks that measure end-to-end throughput and latency but do not provide separate metrics for each phase.

I would greatly appreciate any guidance on profiling these two phases separately.

How would you like to use vllm

No response

[leiwen83]

stream mode shall get each token's latency, and thus prefill and decode phase could be measured.
While current benchmark using sync mode, another workaround may be considered is:

1. measure latency for input_len=1000, output_len=1, thus get prefill latency for input_len=1000
2. measure latency for input_len=1, output_len=1, get average latency A, and then input_len=1, output_len=1000, get average latency B. (B-A)/999 to get the decode latency...

[Msiavashi]

So, there is still no embedded mechanism for these measurements/profiling, right?

[KevinZeng08]

stream mode shall get each token's latency, and thus prefill and decode phase could be measured. While current benchmark using sync mode, another workaround may be considered is:

1. measure latency for input_len=1000, output_len=1, thus get prefill latency for input_len=1000
2. measure latency for input_len=1, output_len=1, get average latency A, and then input_len=1, output_len=1000, get average latency B. (B-A)/999 to get the decode latency...

Hi, now do you have some other ways for profiling prefill and decode phase separately?

[kerthcet]

There's a ongoing PR related. #2809

[dshm]

same question. now do we have some other ways for profiling prefill and decode phase separately?

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!