search

triton-inference-server / client

Triton Python, C++ and Java client libraries, and GRPC-generated client examples for go, java and scala.
BSD 3-Clause "New" or "Revised" License
520 stars 225 forks source link

Update scaling from ns to ms #639

Open debermudez opened 2 months ago

debermudez commented 2 months ago

The default scaling for timestamps is now ms from ns. Tests are updated. Console output is updated Plot labels are updated.

Screenshot 2024-05-08 at 20 16 00

nv-hwoo commented 2 months ago

Before

image

After

image

Should we keep some precisions for both accuracy as well as aesthetic? (columns seem a bit too narrow 😅) I know @dyastremsky argued for integral number for nanosecond time, which I think made sense since fractional part of nanosecond is way too small to be significant. But for millisecond, I think we may want to keep the precision (up to 3 decimals?). What do you think @dyastremsky ?

dyastremsky commented 2 months ago

Before image

After image

Should we keep some precisions for both accuracy as well as aesthetic? (columns seem a bit too narrow 😅) I know @dyastremsky argued for integral number for nanosecond time, which I think made sense since fractional part of nanosecond is way too small to be significant. But for millisecond, I think we may want to keep the precision (up to 3 decimals?). What do you think @dyastremsky ?

Those are good points! I would separate the accuracy concerns from the aesthetic ones. You should be able to make the columns artificially wider, though having them be narrow is better for small screens IMO. There is hopefully a setting in the library that can help with adding a buffer for larger screens.

As far as the accuracy concerns, I think changes in accuracy should be fully based on customer needs. I would confirm with a few users on what is useful for them.