triton VS dfdx

There's a ton of cool opportunity in the runtime layer. I've been keeping my eye on the compiler-based approaches. From what I've gathered many of the larger "production" inference tools use compilers:

- https://github.com/openai/triton

Does anyone have more inside knowledge from OpenAI or AMD on AMDGPU support for Triton?

I see this:

https://github.com/openai/triton/issues/1073

But it's not clear to me if we will see AMD GPUs as first class citizens for pytorch in the future?

I thought this was a good overview of the idea Triton can circumvent the CUDA moat: https://www.semianalysis.com/p/nvidiaopenaitritonpytorch

It also looks like they added MLIR backend to Triton though I wonder if Mojo has advantages since it was built on MLIR? https://github.com/openai/triton/pull/1004

I'm pretty happy with the setup, because it allows me to keep all the AI stuff and its dozens of conda envs and repos etc. seperate from my normal setup and "portable". It may have some performance impact (although I don't personally notice any significant difference to running it "natively" on windows), and it may enable some extra functionality, such as access to OpenAi's Triton etc., but that's currently neither here nor there.

Very cool development. There is too much busy work going from development to test to production. This will help to unify everything. OpenAI Triton https://github.com/openai/triton/ is going for a similar goal. But this is a more fundamental approach.

I keep checking the progress on dfdx for this reason. It does what I (and, I assume from context, you) want: Provides static checking of tensor shapes. Which is fantastic. Not quite as much inference as I'd like but I love getting compile-time errors that I forgot to transpose before a matmul.

It depends on the generic_const_exprs feature which is still, to quote, "highly experimental":

https://github.com/rust-lang/rust/issues/76560

Definitely not for production use, but it gives a flavor for where things can head in the medium term, and it's .. it's nice. You could imagine future type support allowing even more inference for some intermediate shapes, of course, but even what it has now is really nice. Like this cute little convnet example:

https://github.com/coreylowman/dfdx/blob/main/examples/night...

And for transformers, it's really early days for dfdx, but it's a library that aims to sit basically at the Pytorch level of abstraction, that the difference is it's not just coded in Rust, but it follows the Rust-y/functional-y philosophy of "if it compiles it runs".

Wow that is super interesting. I actually tried to use GATs at first to be generic over shapes, but I couldn't do it, I'm sure it would be possible in the future though. There is this library dfdx that does something similar to what you mentioned, but it feels a little clumsy to me.

Awesome, I added an issue here https://github.com/coreylowman/dfdx/issues/597. We can discuss more there! The first step will just be adding the device and implementing tensor creation methods for it.

The next release of dfdx includes a CUDA device and implements many ops. The same dev created a new crate, cudarc, for a wrapper around CUDA toolkit.

A question I have is: what are the philosophical/design differences with dfdx? As someone who's played around with dfdx and only skimmed the README of burn, it seems like dfdx leans into Rust's type system/type inference for compile time checking of as much as is possible to check at compile time. I wonder if you've gotten a chance to look at dfdx and would like to outline what you think the differences are. Thanks!