nvukobratTT
commented
2 weeks ago @nsmithtt @tapspatel During this sprint, we plan to focus on E2E tests mostly for higher robustness. Therefore, we'll require more silicon CI machines than one.
What are the plans for MLIR testing? Do you have some estimates of the required workflow?
From the FFE perspective, it'll be good to have:
- 2 build runners - in the next few weeks 4 more folks will join FFE, so it'll be good to have those faster builders
- 2 more n150 runners - in the upcoming week or two, we're starting with robust op testing (e.g. testing added with many different variations and attributes e2e). This is mostly to error out FFE, MLIR, and TTNN/Metal bugs sooner rather than later, and plan accordingly for patching (especially focusing on the tt-nn side here).
- Regarding n300, as multi-chip is still WIP, there is no need to have them on FFE as well
By the above points, it'll be good to balance things out a bit. Do you think that is possible?
vmilosevic
Right now Forge Fe only has 1 machine allocated to it. This is causing a bottle neck on their development.
@nvukobratTT How many additional machines will ease the constrain?
@tapspatel How many machines can we afford to allocate to the FE side of things?
I know we looked into pooling which seemed non-trivial, however I wonder if there is a way to make this easier in the future / understand how our infra utilization is going.