AMD ROCm 5.6.1 Compute Stack Released With A Few Fixes

[phoronix]

Phoronix: AMD ROCm 5.6.1 Compute Stack Released With A Few Fixes

While we are eagerly awaiting ROCm support for more RDNA3 GPUs said to be coming later this calendar year, shipping Tuesday night was ROCm 5.6.1 as the newest point release for this open-source GPU compute stack...

https://www.phoronix.com/news/AMD-ROCm-5.6.1

[fraz0815]

What's missing in ROCm 5.6 for RDNA 3?

I am using:
AI Image Stuff: AUTOMATIC111's stable-diffusion-webui fork (use dev branch, torch dependencies were outdated, made a PR which is in dev), SadTalker (did not merge my PR yet)

And for LLM's
llama.cpp compiled with 'CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ make LLAMA_HIPBLAS=1'. 13B Models run fine, did not test bigger models yet.

[NeoMorpheus]

Grabs popcorn for the barrage of negative posts from Ngreedia fans.

[piotrj3]

Grabs popcorn for the barrage of negative posts from Ngreedia fans.

I don't have to say anything, post is embarrasment for AMD in itself. I would expect day 1 support for RDNA3 GPUs with rocm for HIP and opencl. 9 months no support, and we don't know when it is even comming. At this point Micheal should compare 5000 series Nvidia vs 7000 series AMD if they gonna be so late with software :P

[Grinness]

I don't have to say anything, post is embarrasment for AMD in itself. I would expect day 1 support for RDNA3 GPUs with rocm for HIP and opencl. 9 months no support, and we don't know when it is even comming. At this point Micheal should compare 5000 series Nvidia vs 7000 series AMD if they gonna be so late with software :P

According to phoronix, RDNA3 are already supported (unofficially) since ROCm 5.5:

While ROCm 5.5 has been tested to have better RDNA3 support, somewhat surprisingly it's not mentioned at all in the now-published v5.5 release notes... But then again AMD tends to "officially" just focus on their professional/workstation graphics card / accelerator support. The GPU support matrix for ROCm 5.5 remains quite sad with just a handful of GFX9 / CDNA / RDNA GPUs. But all indications so far are that ROCm 5.5 should be working out better now for Radeon RX 7000 series Linux users.

https://www.phoronix.com/news/ROCm-5.5-Released

Furthermore the ROCm windows support includes a long list of GPUs, including _official_ support for RDN3 :

https://rocm.docs.amd.com/en/latest/release/windows_support.html

I personally use ROCm on rx 6800 and rx 480 with pytorch no problem in Linux (Arch linux repos, version 5.6.0)

Surely AMD has no great communication strategy (EDIT: but the cards just work)

(opening my popcorn bag)

[david-nk]

I personally use ROCm on rx 6800 and rx 480 with pytorch no problem in Linux (Arch linux repos, version 5.6.0)

Surely AMD has no great communication strategy (EDIT: but the cards just work)

(opening my popcorn bag)

Okay, but does it actually work? When I tried to get anything done with ROCm, the were a never-ending stream of issues from random internal errors to random data corruption, getting NaNs, straight crashes, internal compiler failures and whatnot. Basically only training a standard ResNet worked fine and a few other things.

So can you now successfully train networks of various different architectures, like let's say BigGAN, ProGAN, StyleGAN2/3, EfficientNet, NASNet, GPT-like models, Stable Diffusion, Insgen... especially most of the GAN-type architectures failed with ROCm, even relatively simple ones like CycleGAN were unstable.

If most of those work, I'm convinced.

[Grinness]

Okay, but does it actually work? When I tried to get anything done with ROCm, the were a never-ending stream of issues from random internal errors to random data corruption, getting NaNs, straight crashes, internal compiler failures and whatnot. Basically only training a standard ResNet worked fine and a few other things.

So can you now successfully train networks of various different architectures, like let's say BigGAN, ProGAN, StyleGAN2/3, EfficientNet, NASNet, GPT-like models, Stable Diffusion, Insgen... especially most of the GAN-type architectures failed with ROCm, even relatively simple ones like CycleGAN were unstable.

If most of those work, I'm convinced.

In terms of AI, I use ROCm 5.x to run the following reference algorithms via pytorch-rocm:
* OpenAI wisper
* Huggingface BERT for text classification (EDIT: including training of the model)

I also built my own algorithms, mostly for testing and verification.
For example, on the MNIST optical character database I tested (and run on ROCm) bespoke implementations of (EDIT: including training):
* standard feed forward neural networks
* multi-layer convolutional neural networks ('traditional' computer vision deep learning)

During the ROCm 4.x cycle I had errors from the stack when running (EDIT: including training) convolutional nn, these have been resolved since rocm 5.x (possibly earlier, I do not remember)

I am currently planning to test on the same stack OpenAI CLIP

Outside AI, I use the rx 6800 and the ROCm stack to run blender cycles with HIP -- the rx 480 is not picked up by blender despite rocm is installed and working with pytorch.

If you have a specific reference implementation to test, I am happy to do so.

Cheers.

[dreamcat4]

i have 5700xt on 23.04 and cannot install this software... the problem isn't a question of what is considered to be 'official support'. the problem is that this is

a) generally just way too convoluted and difficult to install
b) they wont rebuild on newer lib dependancies for 23.04

and then maybe c) that it has other extra issues too...

owning an amd card does not make my some nvidia shill for saying this: it just plain isn't good enough to meet some minimum levels of expectations. my distro is brand new. and the generation of this gpu is 3 behind. which was well enough plenty of time for amd to figure it all out.

i am not even saying support has to be at some levels approaching nvidia. it just basically has to be installable, to run, execute. and perform at whatever level(s) that the developers have currently gotten to implementing in the upstream repository (but on my own system). nothing more, nothing less.