Problem Statement
There should be E2E tests that function as benchmarks that spin up queue implementation and async-processor instance and generate load and record how long it takes the requests to be serviced and the saturation metrics. This can be used to show and prevent performance regressions and ensure flow control is working.
Proposed Solution
For the initial benchmark we assume homogenous inference requests:
- Generate homogenous inference requests and add to queue using producer.
- Start mock inference backend
- Start llm-d-async with a given config
- Record saturation metric and time how long it takes for all the inference requests to be completed.
See the results posted in #304
Alternatives Considered
No response
Willingness to Contribute
Yes, I can submit a PR
Additional Context
No response
Problem Statement
There should be E2E tests that function as benchmarks that spin up queue implementation and async-processor instance and generate load and record how long it takes the requests to be serviced and the saturation metrics. This can be used to show and prevent performance regressions and ensure flow control is working.
Proposed Solution
For the initial benchmark we assume homogenous inference requests:
See the results posted in #304
Alternatives Considered
No response
Willingness to Contribute
Yes, I can submit a PR
Additional Context
No response