search

tenstorrent / tt-mlir

Tenstorrent MLIR compiler
https://tenstorrent.github.io/tt-mlir/
Apache License 2.0
52 stars 7 forks source link

T5 Model Bringup #75

Open nsmithtt opened 2 months ago

nsmithtt commented 2 months ago

Unique ops in T5

nvukobratTT commented 1 month ago

What do you think about creating milestones for each model that we're bringing up?

As we're at the early stage of this compiler, there will be more jobs for each model to work e2e. Therefore, it seems reasonable to have milestones for each one until we see that the new model bringup is becoming less complex.

I'm asking this as a new op bringup can be a good ramp-up task for new folks on Forge and MLIR projects. Therefore, having some of them in the backlog is beneficial.

nvukobratTT commented 1 month ago

Besides, the proposed set of ops, these are related to the inference? I assume training for T5 is still TBD until we get few more PoC for the basic MNIST models working e2e.

nsmithtt commented 1 month ago

What do you think about creating milestones for each model that we're bringing up?

As we're at the early stage of this compiler, there will be more jobs for each model to work e2e. Therefore, it seems reasonable to have milestones for each one until we see that the new model bringup is becoming less complex.

I'm asking this as a new op bringup can be a good ramp-up task for new folks on Forge and MLIR projects. Therefore, having some of them in the backlog is beneficial.

I agree completely and I really like this idea of having milestones for each model. There will be a lot of overlapping work between the models so it could help to have individuals run ahead, try pushing on a model, file issues for the ops/issues they're running into and then we can cross-reference which models are blocked on what which can help us prioritize what the best bang for buck is.

And yes, above was just inference.