MichaelJKlaiber commented on PR #12087: URL: https://github.com/apache/tvm/pull/12087#issuecomment-1190262638
> @cbalint13, appreciate sharing your thoughts! I also think this is a great out-of-tree approach and look forward to its actual usage. > > > Could you add more documentation in general (e.g., explain what parameters are)? > > > Current sample (_template) & doc are fine for a contrib module. > > > It greatly enables "out-of-tree" maintenance towards custom HW accelarator design. > > > Example parts Vanilla, Strawberry & Chocolate are nice levels of introductory & guidance. > > I agree that the overall docstring structure (e.g., Vanilla, Strawberry & Chocolate) would be a great introduction/tutorial (Truly appreciate the authors!). My issue was more like a code-level docstring since some of them are missing and I wasn't sure what each function or parameter exactly means. https://github.com/apache/tvm/blob/bc211fdc91932ef9b77f7edd2ff65151ee5faaff/python/tvm/relay/backend/contrib/uma/_template/codegen.py https://github.com/apache/tvm/blob/bc211fdc91932ef9b77f7edd2ff65151ee5faaff/python/tvm/relay/backend/contrib/uma/_template/codegen.py > > > It would be great to have a testing for end-to-end model-level testing. (e.g., run BERT with UMA on ARM EthosU) > > > Sounds beyond of scope, same way we could "dream" a fully verilated BERT running verilator-contrib module (!) @tmoreau89 , @vegaluisjose > > > > > > > As for ARM-Ethos ⟨™⟩, my understanding is a quite different (not so "out-of-tree" principled) way to connect things inside, but @manupa-arm can comment more on this. > > > > > > > Would be possible in the future to decouple ARM-Ethos⟨™⟩ (or Vitis-AI) to be more "out-of-tree" by porting it over UMA, but this would be quite bit of another task. > > I see. Thank you for the clarification! Maybe my example was not relevant. I just thought it would be nice to have some model-level test cases since we seem to have all necessary components (UMA Partitioner, UMA Pipeline, UMA Lower, UMA Codegen) in this PR. I just wanted to see if current PR also covers a case where we lower some parts with UMA while handling other unsupported parts via the conventional pipeline. We have an example testcase for running MobileNet (created in Relay) with uma and the Vanilla Accelerator in place. Not sure if it should be added to this PR or a later one, as this PR already is a lot to discuss. I'd be in favor of moving this to the next UMA PR. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
