[no_wizard]

Can't speak to how it compares to CUDA however they are developing ROCm[0]

[0]: https://www.amd.com/en/products/software/rocm.html

[zamalek]

I can attest that it works really well on my 6900xt. Compiling CUDA kernels is merely a matter of a using a #define shim. Also, provided you download the ROCm pytorch (and force compatibility with the HSA_OVERRIDE_GFX_VERSION) everything just works.

[echelon]

Next to nothing is written that targets this. Not Stable Diffusion, not RVC, not Vall-E, not Tacotron, not Tortoise, ...

Maybe the LLM space is better about this, but the generative media side definitely isn't.

AMD has a market share of 0% here, and nobody publishes models with AMD support.

[Symmetry]

The things is you can actually run Stable Diffusion.

And I got PyTorch working on my AMD 7900 XT graphics card recently, though it was a bit of a hassle to do so.

[godshatter]

You can also run Stable Diffusion in cpu mode, if you don't mind it being slower. I have an NVIDIA card but it's not powerful enough to run it. I'm on Ubuntu.

[echelon]

> though it was a bit of a hassle to do so.

Incredible understatement. And the diverse set of community tools also breaks down.

We're still a year or more out from proper AMD support in the ecosystem.

[coffeebeqn]

Could there be a compiler/transpiler from CUSA to whatever AMD is pushing ?

[puzzlingcaptcha]

That's what https://github.com/ROCm/HIPIFY is (as a part of ROCm)

[zamalek]

I think it's llama.cpp that simply #defines all cuda_ functions to rocm_ (99% name-name). Porting seems to be that trivial.