triton VS automatic

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.

> it won't ever be a viable option

For production workloads, I generally agree. It's an unsupported hack with a questionable future, I wouldn't do anything money-making with it.

However, for tinkering and consumer workloads, it already works pretty well. Enough of cuDNN and cuBLAS work to run PyTorch and in turn, Stable Diffusion with https://github.com/lshqqytiger/ZLUDA - there's even a fairly user-friendly setup process already in https://github.com/vladmandic/automatic .

I was able to get a personal non-ML related project working on my AMD card in just a few minutes, which saved me a lot of development time before I then deployed the production workload on NV hardware (this is probably why AMD pulled the plug on the project - it's almost more of a boost to NV than anything else, AMD really need people to be writing code on ROCm to deploy on AMD datacenter hardware).

While I currently use SD.Next[1], I have tested ComfyUI locally with my AMD card. The UI can be daunting, but you learn quite a great deal about how a Stable Diffusion pipeline works. In addition some innovations and advances find their way into ComfyUI first.

[1] https://github.com/vladmandic/automatic

I used SD.NEXT ( https://github.com/vladmandic/automatic ) and https://civitai.com/models/82098/add-more-details-detail-enhancer-tweaker-lora and epicphotogasm_lastUnicorn

I found this thread on GitHub talking about how this was fixed in the latest version with an optional setting. I tried enabling it, as they mentioned, but it just resulted in an immediate CUDA out of memory error when starting generation. So it seems I'm actually needing the shared memory, which I assume is my issue.

As always, do check out our more detailed changelog, give us a quick install from our Repo, and stop by our Discord Server for any questions or help you may need.

SD.Next

Today I finally wanted to give SDXL a chance, so I set everythin up according to Vladmandic's Wiki https://github.com/vladmandic/automatic/wiki/SD-XL

I'm trying to get SDXL working on Vlad's SDNext, but I keep getting the error in the title when trying to run basic operations. I'm not sure what's going on, I followed his guide for it to a T.

We used an inference container based on SDNext, along with a custom worker written in Typescript that implemented the job processing pipeline. The worker used HTTP to communicate with both the SDNext container and with our batch framework.