Feature and Support Requests

[saadrahim]

No promises but let us know what you want here. New OS support, features, and anything else is helpful.

[Yanislar]

Please support the SteamOS on the Steam Deck and it's GPU: AMD Custom GPU 0405

[TheAkashicTraveller]

Does anyone know if stable diffusion works on another os that normally supports it? Or just base stable diffusion without webui?

[aemonge]

Isn't the thread getting off-topic?
It started for supporting SteamOS, and I can't understand why dockerized ubuntu is mentioned? Is the dockerized version of ubuntu working under SteamOS, working?

I got a bit lost, on the conversation and can't understand if the version has been now supported or we are still waiting :)

[deftdawg]

Isn't the thread getting off-topic?
It started for supporting SteamOS, and I can't understand why dockerized ubuntu is mentioned? Is the dockerized version of ubuntu working under SteamOS, working?

Nope, it's on the topic of finding a supportable solution.

Docker/Podman containerization are one way to support ROCm on devices that have the kernel driver stack but not the userspace stack on host systems where modifying the userspace stack is challenging/problematic (i.e on SteamDeck where the OS is read-only and subject to being flashed over by updates)

[aemonge]

It is helps, I'm using https://github.com/ValShaped/rwfus to “overwrite” the permission, and be able to handle pacman/pikaur without being frightened about the flash updates, nor forced to use steamos-readonly disable. I've been happily with all my packages, customizations, and updates for 4 months now.

Though, this might be off-topic :)

Anyhow, thanks @deftdawg I'll now try the dockerization solution. I didn't understand it, but now is obvious!

[xiaobo1025]

Dear friends, recently, according to the ROCm deployment guide on the windows 10 system provided by the AMD official website, I have not been successful in my own windows 10 system. I hope you can give me some help so that I can get started with ROCm. I really appreciate your help!

[workspace-public]

Gentoo compatibility, If it can be built on gentoo then it will work well for anyone who thinks they know what ROCm is but actually they don't. Rust for example, does not easily build on Gentoo but it does build.

It would also be a cool feature for the release cycle to be cutting edge exclusive to the intended platform, INSTINCT, etc., with a second LTS for a "tock" cycle of consumer GPUs. People are practically begging for this use case, and it can be used to purge held stock.

[tedliosu]

Umm here's two:

1. Please for the love of all that's good support WMMA-accelerated tensorflow and pytorch mixed precision training and inferencing through something like rocWMMA (which btw I saw has already been released for cards like the MI200, now pretty please just do it for the prosumer/consumer cards as well); Nvidia already has had at least 3 generations head start and I'd hate for the tech gap to widen even further.

2. Per here on a related note Nvidia also offers consumers the CUDA framework which works for consumers FROM DAY ONE OF HARDWARE RELEASE whereas ROCm has had a pretty poor track record thus far with catching up with RDNA support; I mean I get it datacenters are where all the $$$$$$ is at but won't it be wonderful to be able to offer a timely delivered viable alternative to the big green giant so that AMD can get more market share at the consumer end? I mean great job on the RX 7000 series pricing it looks fantastic imo at least for the high end imo, but please like Frank said on stage have the software to back it up as well :)

OK rant over; keep innovating AMD 👍

[Odin234]

@saadrahim Any update on that forward looking statement that you talked about previously?

rocm 5.5.0 has been late for few months. The last released rocm is February. No communications and no support. It's our mistake chose AMD. Switch to team green instead. It has cuda support from day 0 release. Pytorch just work out of the box.
And you still wish AMD is not ignoring this?

[saadrahim]

Apologies, but I have no firm update on when ROCm 5.5.0 will be out. I hope it will be out in the coming weeks.

[MatthiasGrandl]

@saadrahim appreciate the follow up, even though it's not what I would have hoped for...

[ReginaVistine]

Apologies, but I have no firm update on when ROCm 5.5.0 will be out. I hope it will be out in the coming weeks.

Don't be sorry, just be better. We, RDNA 3 users need ML/AI to work. That silence and unclear communication from AMD will make bad user experience.
Why don't you take this to higher-ups, in more serious manner, so your higher-ups can do something decent about this?

https://github.com/ROCmSoftwarePlatform/MIOpen/milestones

Miopen for 5.5 and 5.6 are done, even those are too far behind schedule.
Why don't you give a working release version for RDNA 3?

[saadrahim]

RDNA 3 support will be available in the fall of this year.

https://community.amd.com/t5/rocm/new-rocm-5-6-release-brings-enhancements-and-optimizations-for/ba-p/614745/jump-to/first-unread-message

[billy2618]

Hey AMD,
Watch this thread.
https://www.reddit.com/r/Amd/comments/yw7chy/will_rocm_finally_get_some_love/

There are so many people eagering to have ROCm work on their card and finally find that the only reasonable solution is to turn to the team green.

Don't fail us again.

[saadrahim]

Thanks for showing me the sentiment on reddit. Most of us are super busy and don't respond to these threads as regularly as we should. However, don't assume we are not paying attention.

[lsmine0]

Will there be a Windows 10 release any time? i understand that it is a whole new OS to make for but is it even in mind? i can't find any note or mention of windows os for ROCm support.

[DuckersMcQuack]

Meanwhile windows can with Nvidia GPU and cuda. Just GeForce cards can't split workloads like for a VM because buy Quadro lol. But i do computational stuff with my 3090 on windows for SD and blender. But and needs rocm-HIIP to do blender cycles

[lsmine0]

it doesn't matter that windows was mainly to busness needs. Ai and ML is getting main stream on many aplications and for rending do need CUDA like functionality. Pytorch or any platform or framework for neural networks suddenly get cut off. HIP functionality is good but for development on windows there has to be ROCm platform. Does that mean if i chose to use stable diffusion or GPT text model development my only option is cuda with no plans for AMD?

[workspace-public]

Not exactly. Vulkan is capable of implementing HPC workloads and that is natively capable on Windows. The ROCm system is a collection of devoted mathematical kernel components for advanced AMD GPGPU with narrow focus on dedicated hardware designed for massive scale. They can do this project because it is roughly equivalent to the IBM mainframe being made for the first time in its current lineage from the System 360. They are not denying you some prized capacity, they have the grand luck to be capable of providing this at all.

[saadrahim]

A Windows 10 and Windows 11 release is planned. Preparatory work is underway. Amongst the publicly visible activity, you can see the team is busy resolving Visual Studio solution file issues at ROCm/rocm-examples#22.

[xiaobo1025]

Dear friends, recently, according to the ROCm deployment guide on the windows 10 system provided by the AMD official website, I have not been successful in my own windows 10 system. I hope you can give me some help so that I can get started with ROCm. I really appreciate your help!

[james-lawrence]

highly interested out of the box support in ROCM support for arch linux for prosumers. I want to play and be productive at a reasonable price basically want rocm to work as seamlessly as mesa does these days for AMD on linux. (fine with an additional package install; but it just 'needs to work')

[cmlibo]

现在买7900xt是不合时宜的，买了这个卡，两个多星期了，Ubuntu都重装了7次了，试过各类方法，还是无法运行stable-diffusion，甚至Linux游戏也无法玩，显卡无法被游戏识别，等AMD更新还不知道猴年马月。至少目前来看还不如上代的6600xt

[Maetveis]

In #994 @abuccts requests a container runtime for docker/kubernetes to allow specifying accessible GPUs by environment variables.

To expand further with some personal points:

• Having this runtime would also allow for the same simple commands to be used for GPUs in docker containers as NVIDIA i.e. --gpus all instead of the much more intimidating --device /dev/kfd --device=/dev/dri.
• The broader software ecosystem also makes use of the container runtime: for example Visual Studio Code allows requiring GPU access for devcontainers, but this currently only works for NVIDIA GPUs, having a compatible runtime would probably allow implementation for AMD.

[billy2618]

If windows support is not on the plan, how about WSL2 ?

[xiaobo1025]

Dear friends, recently, according to the ROCm deployment guide on the windows 10 system provided by the AMD official website, I have not been successful in my own windows 10 system. I hope you can give me some help so that I can get started with ROCm. I really appreciate your help!

[burlen]

Please add support for OpenMP offload and HIP in the same translation unit. It is needed for templated C++ codes that could use both OpenMP offload and HIP, for instance depending on downstream users' preferences to enable interoperability of the two strategies. The use of templates makes it impossible/impractical to factor things into separately compiled translation units and I would suspect that the majority of the technological underpinnings to do this are already available, given that OpenMP offload has the capability to compile and launch HIP kernels and device functions. NVIDIA's HPC compiler currently supports this feature, would be great if AMD did as well.

Discussed in #2137

[johnnynunez]

RDNA3 Support: https://community.amd.com/t5/rocm/new-rocm-5-6-release-brings-enhancements-and-optimizations-for/ba-p/614745/jump-to/first-unread-message

[dbenedb]

I'm using ROCm/HIP mostly for rendering in Blender but I'm also experimenting with Stable Diffusion (inference). All under Linux.
My list of requests that will make AMD GPUs acceptable as a compute platform is pretty short:

1. DAY ONE support for ROCm on consumer/prosumer hardware (both Linux and Windows).
2. Shorter release cycle or more frequent corrective releases.
3. Developing for both platforms (Linux and Windows) at the same time.

AD1: I'm not going to buy any AMD GPU unless it has official ROCm support for the parts/libs that are required by the tools that I use. Period. I'm still on Radeon VII btw.
AD2: If there is any bug, that involves HIP in Blender (especially on Linux) not only we usually need to wait extremely long for it to be diagnosed and fixed, and if it is then we need to wait another couple months for the release with a fix to land. This is unacceptable. Same additional delay happens with shipping new features.
AD3: Again an example from ROCm/HIP in Blender - there is a trend of providing support or shipping new features to Windows first ten after another couple months to Linux. This is also unacceptable. You are making your GPUs useless for Linux users for practically a year after the release! And when you will drop support for that GPU you will do it for all OS' at once. This is an absolute joke. I'm not going to pay for a product that has it's lifespan shortened by a year.

[johnnynunez]

@dbenedb @saadrahim Hopefully they speed up to all official RDNA support and RDNA4 will be supported by ROCm from day one. At the hardware level, unless they also bring systolic arrays (matrix cores in CDNA / tensor cores in Nvidia) in RDNA4.

On the other hand I think that AMD should push to bring zen4/5 + rdna4 consoles with systolic arrays and infinity cache for 2024/25 (ps5 pro or similar), it would not only be beneficial for AI algorithms on consoles (like behavior of bots, FSR3 etc) because AMD is not associated much with AI, so at the marketing level where millions of players move it would be a masterpiece.

On the other hand I think native windows support is essential but I think right now it's optional, most of the people who work for AI use Linux.

Another important feature would be direct storage in deep learning or similar, or even compression and decompression of data in real time for DL in ROCm like similar in consoles.
Also, this will be important future: Heterogeneous Memory Management (HMM), allowing seamless sharing of data between host memory and accelerator devices.

In terms of drivers in Linux, I hope that the idea of being able to OC and undervolt both cores and memory to get the most out of each chip is maintained.

[workspace-public]

This is more of a hardware/driver request. Is it possible for ROCm to create an FPGA integration similar to Optane? Modern consumer motherboards for custom GPGPU workloads can have a glut of M.2 slots. Is it possible to have an M.2 FPGA for Instinct which can convert Crossfire capable GPUs into ROCm compatible systems?

[workspace-public]

Wrong target. This is for AMD's hardware team, not ROCm. @AMDGithubSCIMAdmin maybe?

[108806]

Using 7800xt and 23.11.1 on windows 10, and the latest hip rocm from the amd, it is not working at all. Nothing is able to detect the lib, it drops to OpenCL or CPU computing.

[Mhowser]

@108806 you should open a new issue for that.

[xiaobo1025]

Dear friends, recently, according to the ROCm deployment guide on the windows 10 system provided by the AMD official website, I have not been successful in my own windows 10 system. I hope you can give me some help so that I can get started with ROCm. I really appreciate your help!

[sorasoras]

You should add rocm to your env.

[Sarracen1a]

Are there any plan for support in the driver for QoS or resource isolation between containers？

[sorasoras]

ggerganov/llama.cpp#4742

On NVIDIA cards that are Volta or newer FP16 arithmetic instead of FP32 arithmetic is used. This is both faster and reduces the amount of shared memory required.
. For AMD FP16 arithmetic is disabled because the HIP documentation does not seem to contain an equivalent for __hmax2.
Could AMD add support for __hmax2？
so that AMD fp16 is usable in this application.

[hiepxanh]

All 3 links are broken, please fix it @saadrahim

[hiepxanh]

the link in https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html is also just stay at 5.7 version, I dont see where to download 6.0 version

[hiepxanh]

go to this page, this link die too:
https://rocm.docs.amd.com/projects/install-on-windows/en/latest/

[xiaobo1025]

Dear friends, I'm going to use CK's API to write a gemm.cpp program for matrix multiplication. How should I implement this task?

[xiaobo1025]

Dear friends, I'm going to use CK's API to write a gemm.cpp program for matrix multiplication. How should I implement this task?

[ChuloIva]

Please rocm support for rx7700xt on Ubuntu 🙏🙏🙏