Skip to content

How to independently measure the performance of the Prefill phase and the Decode phase? #50

@J1nLo

Description

@J1nLo

As shown in the Figure 1 of the paper, I want to know how to independently measure the performance(rps or latency) of the Prefill phase and the Decode phase?

Now, I have tried these ways:

  • Prefill: input_len = x, output_len = 1
  • Decode: input_len = 1, output_len = y

Is this method correct? Maybe this method doesn't take into account kvcache, but I don't know how to simulate this part.

And is there any other more precise way?

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions