ROCm 6.1 Released With Ubuntu 22.04.4 Support, rocDecode For AMD Video Decode

[Jabberwocky]

The people dual booting have been waiting for this update for a very long time, IIRC new MIOpen was the thing preventing Pytorch from running on Windows. I think this is the first time we can do crossplatform benchmarking of native ROCm & pytorch. I'm excited to do some tests around this.

Grabbing me some popcorn and enjoying my 7900 XTX with Stable Diffusion and ollama while watching the AMD haters.

What's different for you with 6.1, weren't you have to enjoy it with 6.0? Is it just that morons like Tomshardware still think that highend AMD GPUs can't do SD/LLM at all?

Grabbing my AMD iGPU and struggling with hacks for any ROCm tasks to work on it.

To be fair which iGPUs from other vendors (without dedicated AI hardware) has good support?

I have a 5700G with 128GB ram (used as VM host) on paper it has potential for compute. 16GB UMA Frame Buffer and IIRC max of 64GB SVM. It's a low power system which feels like it as a lot to offer for SD/LLMs, but in practice I think it's a waste of time trying to do anything meaningful on it because everyone at AMD is focused on the highend chips. The whole world is looking at Nvidia right now and everyone else wants a piece of the hype-pie. I read most people doing ROCm compute straight up disable iGPUs. 🙈

[CochainComplex]

Just recently tried to use rocm 6.0.6/6.0.2 with tensorflow on my 5700Xt ...didn't work. Did a dedicated ubuntu22.04 vanilla installation just for using tensorflow with my navi card. But nothing has worked yet. There are some hacks but wasn't able to make it run.

This is really annoying. I want to use AMD for AI but its diffcult and frusttrating.

[Mathias]

Just recently tried to use rocm 6.0.6/6.0.2 with tensorflow on my 5700Xt ...didn't work. Did a dedicated ubuntu22.04 vanilla installation just for using tensorflow with my navi card. But nothing has worked yet. There are some hacks but wasn't able to make it run.

This is really annoying. I want to use AMD for AI but its diffcult and frusttrating.

AFAIK Rocm inofficially supported NAVI1 (5700) with earlier versions when using `HSA_OVERRIDE_GFX_VERSION=10.3.0`. But since a few versions (at least 5.7.0) rocm seems to use some NAVI2 instructions, so that fails.

But there has been progress, as stated here, that gives hope it will work in the near future.

I can't test it right now. But I checked the files installed on Fedora with the RHEL ROCm repository and I don't think gfx1010 support is fully there

$ ls -1 /opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx*
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1030.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1100.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1101.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1102.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx900.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx906.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx908.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx90a.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx940.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx941.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx942.dat​​

I guess there should be a gfx1010.dat, if NAVI1 would be supported.

I don't know much about tensile, but there are just as many gfx1010 hsaco files then there are for other platforms... So maybe there is some hope that you can somehow easily generate (or download) the lazy_gfx1010.dat...

[CochainComplex]

AFAIK Rocm inofficially supported NAVI1 (5700) with earlier versions when using `HSA_OVERRIDE_GFX_VERSION=10.3.0`. But since a few versions (at least 5.7.0) rocm seems to use some NAVI2 instructions, so that fails.

But there has been progress, as stated here, that gives hope it will work in the near future.

I can't test it right now. But I checked the files installed on Fedora with the RHEL ROCm repository and I don't think gfx1010 support is fully there

$ ls -1 /opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx*
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1030.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1100.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1101.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx1102.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx900.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx906.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx908.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx90a.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx940.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx941.dat
/opt/rocm-6.1.0/lib/rocblas/library/TensileLibrary_lazy_gfx942.dat​​

I guess there should be a gfx1010.dat, if NAVI1 would be supported.

I don't know much about tensile, but there are just as many gfx1010 hsaco files then there are for other platforms... So maybe there is some hope that you can somehow easily generate (or download) the lazy_gfx1010.dat...

thx for the input I will check it out.

[nyanmisaka]

Considering the existence of AMF, why do they reinvent the wheel instead of having AMF support HIP backend???

[oleid]

It's a low power system which feels like it as a lot to offer for SD/LLMs, but in practice I think it's a waste of time trying to do anything meaningful on it because everyone at AMD is focused on the highend chips.

A 7600 works fine for llms.

https://m.youtube.com/watch?v=VXHryjPu52k

[pWe00Iri3e7Z9lHOX2Qx]

Grabbing me some popcorn and enjoying my 7900 XTX with Stable Diffusion and ollama while watching the AMD haters.

Calling everyone a hater that bought recent AMD GPU hardware and wants GPU compute to work is a strange stance.

[Panix]

Calling everyone a hater that bought recent AMD GPU hardware and wants GPU compute to work is a strange stance.

What else do you expect from the AMD fanboys on this site?

[duby229]

Just recently tried to use rocm 6.0.6/6.0.2 with tensorflow on my 5700Xt ...didn't work. Did a dedicated ubuntu22.04 vanilla installation just for using tensorflow with my navi card. But nothing has worked yet. There are some hacks but wasn't able to make it run.

This is really annoying. I want to use AMD for AI but its diffcult and frusttrating.

In a similar boat. Although I haven't actually tried the newest ROCm versions yet, Actually I gave up on it a long time ago. I think I'll put my current bias to the side for a bit and try it again.

[Mathias]

I just tried to run a very basic pytorch script on my 5700 with ROCm 6.1, but I didn't manage to get it running. Another few hours I wasted on this again... Like I said, I think the gfx1010 tensile library is not shipped. I tried it with and without the HSA_Override to no avail.

I tried to compile the tensile library myself. I got to the point, where it was benchmarking on the GPU (I guess). But after 20 minutes, it never finished (was supposed to be 1 Minute)... I'll try to compile it some more tomorrow...