Dilshod Tadjibaev
Dilshod Tadjibaev
@laggui I hope you'd be able to review it soon so we can merge it before other deviations appear on the main.
I think for uninitiated developers, before we jump to a solution, we need to explain a couple of things: why the current design is "less flexible" and why we still...
> Note that most, if not all, of our kernels have a static rank, so even if you fuse 10 tensor operations with 5 actual tensors of different ranks, we'll...
> > Would it be fair to say tracking rank statically helps with graph optimizations? > > I don't think it does. At some point, we had the rank being...
> So with your new design, tracking rank becomes optional? Can you show us how to use dynamic rank? What happens when a tensor with dynamic rank interacts with a...
Just to be clear these feature flags are part of the meta package `burn`?
> [c2b6318](https://github.com/tracel-ai/burn/commit/c2b6318fc3232536c91910923e09ad0f9f30763a) switched to using `ArgType::Shape` for the output of the shapes-Operator instead of just using a 1D `Tensor`. > > As far as I can tell, the ONNX Spec...
Tested the model. it seems to be fixed by https://github.com/tracel-ai/burn/pull/3381
This requires enhancement to Burn's average pool implementation before we can take on ONNX
Working on this issue