5 ms·

Unrelated to this, but related to RDNA3. I have a 7900 XT, and all my attempts at getting ROCm to work on different Linuxes seem to have failed at different po

Unrelated to this, but related to RDNA3. I have a 7900 XT, and all my attempts at getting ROCm to work on different Linuxes seem to have failed at different points. Does anyone have a pointer to a clear explanation on how to get it to work?

oynqr 3y ago
Tried it last summer with an RX 7600 and it just worked by using the Arch packages in the repo.
edg5000 3y ago
I have read many reports of people giving up on it, claiming that it is terribly buggy and has been for a while. The question is, under what circumstances did AMD test it and what does AMD consider "working". Probably it works on some specific setups under certain use conditions.
- redeeman 3y ago
  and thats the problem. Seriously, its not that hard, just make some goddamn software that you can build and use more or less anywhere, why is it that I can relatively easily build mesa with vulkan/opengl driver for AMD cards, but their "rocm" is an atrocious mess? Honestly, it just stinks of incompetence, or if one were otherwise inclined in beliefs, disgust for their customers. If AMD had been hiring japanese to make this, they would have had to close up shop as they would have all had to perform seppuku.
  - weebull 3y ago
    Rocm hasn't been the the problem in my experience. It's things like pytorch on rocm. It uses hippify to convert cuda code to hip rather than having native support. It "works", but it's an obvious second class citizen, and the developers don't seem to take bug reports / PRs seriously.
- roenxi 3y ago
  More than likely they are testing it on specific server setups with a dedicated GPU handling compute workloads. Think supercomputers and maybe cloud provider setups. That'd make sense if someone in AMD ranked the opportunities by $ value and started working down the list fixing bugs. My anecdotal evidence with an AMD GPU is there are serious problems trying to use a graphics card to run graphics and throwing a few compute tasks at it - because X11 will reliably crash and the kernel graphics driver enters some sort of corrupt state. I think that is a use case that was beyond their design considerations to start with. There is a hint that servers are targeted because they supported linux first with ROCm, and had a very tight support matrix of what OS they expected people to use. Unfortunately, this strategy generates a lot of bad publicity and I'd guess has trouble gaining traction because people can't experiment cheaply with AMD hardware. They're progressing though and even with a dose of cynicism about their wherewithal there is a lot in the pipe that has the potential to take on CUDA. Even APUs are promising.
kkielhofner 3y ago
I've been working in this space for ~10 years at this point and have been attempting ROCm since the Vega days (pretty much since initial releases of ROCm ~six years ago). "Hello world" in torch? On a new install/card/config/even upgrade with a bit of fiddling I can get there (I have it more-or-less down at this point). Actually using something in a real world use case? Hours, days, or give up. Often the third scenario where I end up going back to CUDA because I just need to get something done. Messing around with different driver versions, docker containers, ROCm versions, obscure environment variables, weird random cherry picks/patches and hacks from spelunking a bunch of GH repos/discussions/PRs/blog posts/etc. Feels like it never ends. I talk about it quite a bit and I keep buying AMD hardware and putting time into this because I'm rooting for AMD. However, consider the CEO of Nvidia saying 30% of their cost/spend is on software. He gets it. I'm pretty convinced at this point that AMD just doesn't have software in their DNA, with even die-hard "Team Red" Windows gaming users frequently complaining about the quality of drivers. Drivers. For gaming. On Windows. Enter >80% Nvidia market share in that market and >90% market share in ML. They just don't get it and I'm pretty sure they never will. I expect Intel, Apple, and even new entrants in the field will be the ones to offer viable alternatives to the Nvidia/CUDA monopoly.
- ahartmetz 3y ago
  For OpenGL and Vulkan drivers on Linux, AMD sort of fell into a good software culture by accident by hiring a bunch of good people that brought the Linux (kernel) / Xorg / Mesa culture with them. I am pretty happy with the results there! Not contradicting you about the other areas, although I do have some hope because things (well... mostly hardware) have already improved in the GPU department under Lisa Su.
  - kkielhofner 3y ago
    Su has been CEO of AMD for nearly 10 years. Does not contribute to my confidence. Of course with the AI goldrush they're paying a lot of lip-service to ROCm (pumping stock price?) but a couple of examples: 1) They just added official support to ROCm for the top-end consumer card they released over a year ago. Meanwhile in CUDA land Nvidia back-ported support to CUDA 11.8 well before Ada Lovelace/Hopper even launched. Worked with everything out there day of launch. Even newest CUDA 12 and their drivers support anything with Nvidia stamped on it from the past 6-7 years. Literally everything. 2) Just check their repos, Docker Hub, etc. It's embarrassing for them. They used Python 3.10 for ROCm 5.7 because its basically the bare-minimum/standard for any broader AI, ML, etc project. Oh neat, they released ROCm 6 to support their year old ~ $1000 card! Python 3.9? Ubuntu 20.04? What the hell? Pytorch "hello world" usable but almost nothing else even runs. The further you dig the more clownish it gets.
ckastner 3y ago
> I have a 7900 XT, and all my attempts at getting ROCm to work on different Linuxes seem to have failed at different points. Debian formed a ROCm Team a while ago and we are working hard on getting this to work out of the box, as in: apt-get install and you're done. For various reasons, our initial focus was on RDNA2 and CDNA2 and earlier, mainly because we are building a CI network in which we test all of our packages (and their reverse dependencies) on actual cards, and we had to bootstrap the network infrastructure first. CI is central to the Debian ecosystem but all our official infrastructure cannot deal with the requirements for specific hardware attached to a machine -- that was never a factor, until now. AMD has generously supported us with hardware donations, among which are RDNA3 cards. These are already in our possession, and it's just a question of person-time until these are integrated into our CI. We are already spec'ing the systems, though.
- hackerlight 3y ago
  For a fresh Linux install with a 7900 XT/XTX, would Debian or Ubuntu be better (for Pytorch)?
  - hnarayanan 3y ago
    I would love the answer to this question. I’ve tried with Debian and Fedora and gotten to different stages of the journey, but not successfully in either case.
    - hackerlight 3y ago
      How long ago did you try? Before or after official AMD support? https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch
      - hnarayanan 3y ago
        After. I am still currently trying. (Apologies I missed your response earlier as this thread suddenly had a lot going on.)
hackerlight 3y ago
AMD has had official ROCm support for 7900 XT on Linux for ~2 months. How long ago did you try? https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch
- jsheard 3y ago
  Worth noting that the 7900 XT was released over a year ago, though. Having to wait half a hardware generation to actually use your hardware isn't ideal.
  - oynqr 3y ago
    It worked before that. Official just means that you can pay AMD for support.
- hnarayanan 3y ago
  I have been trying different things for a few weeks now.
klowner 3y ago
I do know that I would love it if amdgpu would stop throwing null pointer dereference errors when I open Blender settings with my 7800XT.
FoodWThrow 3y ago
> pacman -S rocm-hip-sdk rocm-opencl-sdk I have the same card, the above works with everything I threw at it so far. Haven't even installed amdgpu-pro drivers btw, only have radv (that steam installed by default).
CapsAdmin 3y ago
You want to use rocm directly or with something like pytorch? If pytorch, keep in mind that most of the pytorch+rocm version is self contained, and you don't need to install every roc* package on your system for it to work. This confused me in the beginning at least.
- hnarayanan 3y ago
  I would like to use PyTorch. If you got it working, could you please share some steps? I feel like I am losing my mind and the overall quality of the software and docs seem really poor.
rwmj 3y ago
I have a similar problem where it worked through Stable Diffusion / PyTorch (worked as in, didn't crash), but the resulting images are just full of garbage. I wish there was some sort of test tool or hello world that could be use to diagnose this stuff, but I didn't find anything. Edit: Example garbage image: http://oirase.annexia.org/tmp/stablediffusion/sunflower-2.png http://oirase.annexia.org/tmp/stablediffusion/sunflower-2.pn...
JonChesterfield 3y ago
The pieces all have exciting versions-must-roughly-match properties. The official distribution idea is pick exactly the recommended OS and install the binary package from amd and it'll be fairly solid. I totally ignore that recommendation and sometimes have a happily working system and sometimes it's a wreck. In particular, the upstream Linux driver (what you have anyway) and the dkms built one are sometimes rather divergent. Llvm in rocm has different behaviour to the upstream llvm too. I'm cautiously optimistic that Debian is packaging will end up as turn key just-works. Part of that is persuading AND developers that breaking ABI has consequences, which it kind of doesn't on the install all of rocm together model. Right now, I use Debian with the same kernel version that the ROCm recommended Ubuntu ships with and build the dkms driver source intended for Ubuntu on it. That be the wrong thing to do but also totally solid in ymmv fashion.

oynqr 3y ago

Tried it last summer with an RX 7600 and it just worked by using the Arch packages in the repo.

edg5000 3y ago

I have read many reports of people giving up on it, claiming that it is terribly buggy and has been for a while. The question is, under what circumstances did AMD test it and what does AMD consider "working". Probably it works on some specific setups under certain use conditions.

redeeman 3y ago
and thats the problem. Seriously, its not that hard, just make some goddamn software that you can build and use more or less anywhere, why is it that I can relatively easily build mesa with vulkan/opengl driver for AMD cards, but their "rocm" is an atrocious mess? Honestly, it just stinks of incompetence, or if one were otherwise inclined in beliefs, disgust for their customers. If AMD had been hiring japanese to make this, they would have had to close up shop as they would have all had to perform seppuku.
- weebull 3y ago
  Rocm hasn't been the the problem in my experience. It's things like pytorch on rocm. It uses hippify to convert cuda code to hip rather than having native support. It "works", but it's an obvious second class citizen, and the developers don't seem to take bug reports / PRs seriously.
roenxi 3y ago
More than likely they are testing it on specific server setups with a dedicated GPU handling compute workloads. Think supercomputers and maybe cloud provider setups. That'd make sense if someone in AMD ranked the opportunities by $ value and started working down the list fixing bugs. My anecdotal evidence with an AMD GPU is there are serious problems trying to use a graphics card to run graphics and throwing a few compute tasks at it - because X11 will reliably crash and the kernel graphics driver enters some sort of corrupt state. I think that is a use case that was beyond their design considerations to start with. There is a hint that servers are targeted because they supported linux first with ROCm, and had a very tight support matrix of what OS they expected people to use. Unfortunately, this strategy generates a lot of bad publicity and I'd guess has trouble gaining traction because people can't experiment cheaply with AMD hardware. They're progressing though and even with a dose of cynicism about their wherewithal there is a lot in the pipe that has the potential to take on CUDA. Even APUs are promising.

kkielhofner 3y ago

I've been working in this space for ~10 years at this point and have been attempting ROCm since the Vega days (pretty much since initial releases of ROCm ~six years ago). "Hello world" in torch? On a new install/card/config/even upgrade with a bit of fiddling I can get there (I have it more-or-less down at this point). Actually using something in a real world use case? Hours, days, or give up. Often the third scenario where I end up going back to CUDA because I just need to get something done. Messing around with different driver versions, docker containers, ROCm versions, obscure environment variables, weird random cherry picks/patches and hacks from spelunking a bunch of GH repos/discussions/PRs/blog posts/etc. Feels like it never ends. I talk about it quite a bit and I keep buying AMD hardware and putting time into this because I'm rooting for AMD. However, consider the CEO of Nvidia saying 30% of their cost/spend is on software. He gets it. I'm pretty convinced at this point that AMD just doesn't have software in their DNA, with even die-hard "Team Red" Windows gaming users frequently complaining about the quality of drivers. Drivers. For gaming. On Windows. Enter >80% Nvidia market share in that market and >90% market share in ML. They just don't get it and I'm pretty sure they never will. I expect Intel, Apple, and even new entrants in the field will be the ones to offer viable alternatives to the Nvidia/CUDA monopoly.

ahartmetz 3y ago
For OpenGL and Vulkan drivers on Linux, AMD sort of fell into a good software culture by accident by hiring a bunch of good people that brought the Linux (kernel) / Xorg / Mesa culture with them. I am pretty happy with the results there! Not contradicting you about the other areas, although I do have some hope because things (well... mostly hardware) have already improved in the GPU department under Lisa Su.
- kkielhofner 3y ago
  Su has been CEO of AMD for nearly 10 years. Does not contribute to my confidence. Of course with the AI goldrush they're paying a lot of lip-service to ROCm (pumping stock price?) but a couple of examples: 1) They just added official support to ROCm for the top-end consumer card they released over a year ago. Meanwhile in CUDA land Nvidia back-ported support to CUDA 11.8 well before Ada Lovelace/Hopper even launched. Worked with everything out there day of launch. Even newest CUDA 12 and their drivers support anything with Nvidia stamped on it from the past 6-7 years. Literally everything. 2) Just check their repos, Docker Hub, etc. It's embarrassing for them. They used Python 3.10 for ROCm 5.7 because its basically the bare-minimum/standard for any broader AI, ML, etc project. Oh neat, they released ROCm 6 to support their year old ~ $1000 card! Python 3.9? Ubuntu 20.04? What the hell? Pytorch "hello world" usable but almost nothing else even runs. The further you dig the more clownish it gets.

ckastner 3y ago

> I have a 7900 XT, and all my attempts at getting ROCm to work on different Linuxes seem to have failed at different points. Debian formed a ROCm Team a while ago and we are working hard on getting this to work out of the box, as in: apt-get install and you're done. For various reasons, our initial focus was on RDNA2 and CDNA2 and earlier, mainly because we are building a CI network in which we test all of our packages (and their reverse dependencies) on actual cards, and we had to bootstrap the network infrastructure first. CI is central to the Debian ecosystem but all our official infrastructure cannot deal with the requirements for specific hardware attached to a machine -- that was never a factor, until now. AMD has generously supported us with hardware donations, among which are RDNA3 cards. These are already in our possession, and it's just a question of person-time until these are integrated into our CI. We are already spec'ing the systems, though.

hackerlight 3y ago
For a fresh Linux install with a 7900 XT/XTX, would Debian or Ubuntu be better (for Pytorch)?
- hnarayanan 3y ago
  I would love the answer to this question. I’ve tried with Debian and Fedora and gotten to different stages of the journey, but not successfully in either case.
  - hackerlight 3y ago
    How long ago did you try? Before or after official AMD support? https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch
    - hnarayanan 3y ago
      After. I am still currently trying. (Apologies I missed your response earlier as this thread suddenly had a lot going on.)

hackerlight 3y ago

AMD has had official ROCm support for 7900 XT on Linux for ~2 months. How long ago did you try? https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch

jsheard 3y ago
Worth noting that the 7900 XT was released over a year ago, though. Having to wait half a hardware generation to actually use your hardware isn't ideal.
- oynqr 3y ago
  It worked before that. Official just means that you can pay AMD for support.
hnarayanan 3y ago
I have been trying different things for a few weeks now.

klowner 3y ago

I do know that I would love it if amdgpu would stop throwing null pointer dereference errors when I open Blender settings with my 7800XT.

FoodWThrow 3y ago

> pacman -S rocm-hip-sdk rocm-opencl-sdk I have the same card, the above works with everything I threw at it so far. Haven't even installed amdgpu-pro drivers btw, only have radv (that steam installed by default).

CapsAdmin 3y ago

You want to use rocm directly or with something like pytorch? If pytorch, keep in mind that most of the pytorch+rocm version is self contained, and you don't need to install every roc* package on your system for it to work. This confused me in the beginning at least.

hnarayanan 3y ago
I would like to use PyTorch. If you got it working, could you please share some steps? I feel like I am losing my mind and the overall quality of the software and docs seem really poor.

rwmj 3y ago

I have a similar problem where it worked through Stable Diffusion / PyTorch (worked as in, didn't crash), but the resulting images are just full of garbage. I wish there was some sort of test tool or hello world that could be use to diagnose this stuff, but I didn't find anything. Edit: Example garbage image: http://oirase.annexia.org/tmp/stablediffusion/sunflower-2.png http://oirase.annexia.org/tmp/stablediffusion/sunflower-2.pn...

JonChesterfield 3y ago

The pieces all have exciting versions-must-roughly-match properties. The official distribution idea is pick exactly the recommended OS and install the binary package from amd and it'll be fairly solid. I totally ignore that recommendation and sometimes have a happily working system and sometimes it's a wreck. In particular, the upstream Linux driver (what you have anyway) and the dkms built one are sometimes rather divergent. Llvm in rocm has different behaviour to the upstream llvm too. I'm cautiously optimistic that Debian is packaging will end up as turn key just-works. Part of that is persuading AND developers that breaking ABI has consequences, which it kind of doesn't on the install all of rocm together model. Right now, I use Debian with the same kernel version that the ROCm recommended Ubuntu ships with and build the dkms driver source intended for Ubuntu on it. That be the wrong thing to do but also totally solid in ymmv fashion.