AMD Engineers directly seeking ROCm feedback
Posted by FORLLM@reddit | LocalLLaMA | View on Reddit | 39 comments
Hoping there are folks here with ROCm experience they can share with those guys to improve the ecosystem. I haven't used it but I'd like there to be as many first class citizens in local ai hardware as possible.
der_pelikan@reddit
The focus on Ubuntu is a problem and hurting your stacks quality. Support at least one other, non .deb platform like Fedora or Arch to clean up the mess.
Unify your python repos for hip/rocm/torch*/onnx and make sure tools like llama.cpp and comfyui can actually be build and run against them before releasing.
Support all recent and upcoming hardware and continue supporting them.
Preconfigure sane defaults for your hardware. Having to lookup environment variables to configure your stack should not be necessary at all.
Nexter92@reddit
They should focus only UBUNTU for now. Even in Ubuntu, Rocm is shit. Focus the product and after the other distro release. Make the one thing why we install rocm great, expand after.
Anduin1357@reddit
Ubuntu is not even remotely a good distro lol
Might as well target Fedora for corpos or Arch for users.
Chupa-Skrull@reddit
You think Arch is designed to be used?
Anduin1357@reddit
Counterpoint: Any distro that defaults to Gnome is kinda miserable.
RoomyRoots@reddit
Counter-counterpoint. Ubuntu has near-official images with all the major DEs and Mint is still one of the best recommendations for beginners.
Anduin1357@reddit
Counter-counterpoint: Mint does not offer mainline Gnome, only a continuation of Gnome 2.
RoomyRoots@reddit
Mint I used as an non-official remix that is an alternative to the GTK people that don't want Gnome. It is not an official remix.
I don't use Gnome but when I tested they had better defaults since they included the extensions needed to make it closer to Gnome2/Unity.
Chupa-Skrull@reddit
I'm mostly just talking shit, I think even people intimidated by Arch would find it ~trivially accessible now that an LLM can answer most questions and instantly poop out bespoke scripts for anything under the sun.
But I also find Gnome very easily customizable to be exactly what I require out of a DE 🤷‍♂️ the performative hate on all fronts is kinda boring these days
Dr4kin@reddit
I like Gnome
RoomyRoots@reddit
As much as I hate Canonical, Ubuntu is the most accessible LTS distro and all work done there works with Debian too. That alone covers Workstation, Desktop and Server sub-markets. Even Steam still uses an ubuntu base image even if their distro is based on Arch.
Wiith RHEL they have been working a lot with them for the Pro series although I never tried running it there.
ShadyShroomz@reddit
no server is using a rolling release distro. these days almost all servers use debian. that should 100% be the main priority...
Anduin1357@reddit
Yup, Debian or Fedora for servers. Arch only if you're living in the daily driver itself.
Nexter92@reddit
Average shitpost from Arch Femboy community
RoomyRoots@reddit
I replied to them months ago in that GitHub tickets asking about the GPU support. Mine was added but I still think they haven't covered everything and they are still not working enough to keep older GPUs compatible with the newer ROCm releases.
RoomyRoots@reddit
OH BOI, do I have my mind to speak to them. I haven't had any issues with 7.11. It runs very well, was trivial to install and it's outperforming Vulkan whenever I remember to test but we still haven't got a public roadmap to when TheRock will replace the current SDK.
Weird-Consequence366@reddit
Stop only supporting Ubuntu.
Mickenfox@reddit
That's nice but the $500 billion company could afford to do actual market research and usability testing instead of asking on X dot com.
FORLLM@reddit (OP)
I have doubts about the benefits to tool improvement of formal market research vs speaking to real world users anywhere.
GoldenX86@reddit
How about supporting all goddamn hardware. It's been 3 years and RDNA3 is still not fully supported.
TeakTop@reddit
The install process and the segmentation between versions is extremely annoying. I even tried the new Ubuntu apt installer, but not only is it a six month old version, I couldn't even get it to work for compiling llama.cpp.
One of the major reasons nvidia is so far ahead is because right from the beginning they made CUDA work on all of their graphics cards, and just won the mind share of the tinkerers and the people that realized the potential of what it could do. AMD seems to be happy playing follow the leader and never making any kind of long term investment in the people that will be writing the next generation of AI.
bnolsen@reddit
with the amd vulkan backend i don't have to download anything special or proprietary, just have the correct dev packages on my system and llama.cpp compiles, installs and runs with little headache. With not running a mainstream distro keeping cuda up to date is a pain. wtf does their installer use curses 5? nvidia is very quick to drop driver support for recent cards but still requires a package that many distros don't even provide anymore?
IBM296@reddit
You should write this reply under the X/Twitter post (though I understand some people here don't use that platform).
FullstackSensei@reddit
I'd love to share some feedback, but I'm not going on Twitter just for that. They should consider other venues for collecting feedback.
FORLLM@reddit (OP)
I posted a link there back to here since so many people are posting here, also encouraged them to make their own posts on reddit since there are so many local ai subs here. I fear though from their replies in that twitter thread that they might be stickler's for their own workflow (asking for github issues).
That'd be fine if they had a mature platform and dominant marketshare, but if they want to catch up I hope they realize they need to do a little more legwork, even if that only means spinning up some agents to collect reddit feedback, test it and create the kind of issues that meet their expectations.
I'm not optimistic, but I'm rooting for them. Better competition in hardware would be very nice to have.
no_no_no_oh_yes@reddit
vLLM is broken for multi GPU for their R9700 AI PRO cards, since version 0.19. "We are looking into it" doesn't cut it.
Quiet-Owl9220@reddit
7900 xtx, I really tried to make it work, but installation is a nightmare, my distro is not supported, and Vulkan works better for LLMs anyway so I gave up on it entirely. Every time I think about trying an audio or image gen model again I remember that I'll have to deal with ROCm and I begin to question if I care enough to bother.
hackiv@reddit
Maybe if you'd support more than few select cards, I'd actually be able to use it and give feedback.
Combinatorilliance@reddit
It's odd but my experience with rocm is actually quite smooth since I started using NixOS.
It just kinda works there with llama.cpp and I don't run into issues much.
DiscipleofDeceit666@reddit
Who’s manager do I have to talk to to get flash memory supported on my 16gb RCNA2 card?
OnlineParacosm@reddit
After Vega 56 I’m just never trusting AMD again until I see them promise an ML flagship card and then deliver on the ROCm side.
They just can’t compete with Nvidia shipping drivers for 20-year-oldcards. There’s no corollary.
My advice for them would be a buyback program for their first generation of laughably ridiculous ML cards I can’t even give to a person in a third world country in good conscience because of the HBMM and driver instability on everyday work.
I don’t know how they think they’re going to generate the goodwill here, frankly.
blackhawk00001@reddit
I found rocm 7.2.0 and 7.2.1 way easier to install than older versions on Ubuntu Linux. Uninstall old, run the listed command on their installation page, and then works great.
I’m using dual R9700s and have found ROCm to scale better with multi gpu than Vulkan. In testing I found Vulkan to have better tg than ROCm with single gpu, but rocm was faster with dual gpu in both using Q8 gguf. Lower quants did not see enough speed boost to justify downsizing.
ROCm vllm with multi-R9700 and mtp enabled is a different machine running Qwen3.6 27B FP8. Getting it working was a pain but there’s a few good posts recently for help.
ismaelgokufox@reddit
For me it has worked flawlessly using the ROCm build of lemonade-sdk of llama.cpp after support was merged for gfx103X-dgpu on https://github.com/lemonade-sdk/llamacpp-rocm/pull/69, by yours truly. (Because the main build on ggml/llama.cpp stopped support for my RX 6800 and I needed the full PP this card provides which was much more under ROCm than Vulcan at that time.)
I use only the ROCm build for daily inference. I do keep the Vulcan one available for testing/comparison from time to time (running llama-bench side by side after some builds).
All this under Windows. I don’t use it under any Linux at the moment. Not even WSL. Don’t know even the slightest on how to make it run there. Will try at some point for sure.
Mountain_Patience231@reddit
how did i gave feedback while its not working all the time
mr_tolkien@reddit
I’d love to reply if I could get Rocm to work reliably
fatboy93@reddit
and if they didnt murk the previous generations as the next one came out
the user should not worry about shit not working on rdna1, rdna2 or 3 or x y z
stop segmenting the fucking cards with basic ops
LagOps91@reddit
Never worked reliably for me and performance is worse than vulkan. Seems to be a pain to support, so that likely is at least partly responsible.
Equivalent-Repair488@reddit
Not an ROCm user but I considered an MI50 before I bought my 3090 around october (ish) 2025. Main reason was ROCm concerns
Two main reasons 1) I was afraid of the install complexity, plus the fact that I was considering adding it to my already existing 3080ti as a secondary GPU. Version specificity + forced into Linux when I'm already using Windows + installation complexity. I'm technologically literate, but not an expert, my budget was very small as a student, the 3090 ate into other expenses, and I was budgeting my meals very heavily, I wanted something reliable that I could trust myself to get working. I already had CUDA working and it is a lot easier. 2) Slower than CUDA, lacking robust support or native development focus for the fancy fast attention like flash and sage.
Also the 3090 helped with gaming as well, but that was hardware reasons, irrelevant to ROCm. I was really tempted on the MI50s though.
1ncehost@reddit
Anush is currently an executive