Yeah I hear ya. How do you happen to be running it? I use NixOS and its a challenge there for me but I found atleast some success using docker since the dependencies are so out of control for AI at the moment.
Also give Ollama-rocm + Open Webgui a shot as an alternative to koboldcpp if you cant get atleast some text generation to work because that is the only thing I have got working with rocm.
The trick to nixos, in this instance, is to use a python venv. Python dependencies are fickle and nasty in the first place, triply so when talking about fast-churning AI code, I tried specifying everything with nix, I succeeded, and then you have random comfyui plugins assuming they can get a writeable location by constructing a path from comfyui’s main.py. It’s not worth it: Let python be the only dependency you feed in, let pip and general python jank do the rest.
At the moment I just don’t. I got kobolcpp to run through distrobox / boxbuddy but I can’t get it to compile with rocm, so I can only use CPU generation, which is abysmally slow. Might go back to NovelAI when they release their new model if I can’t find a solution.
What card do you use? I have a 6700XT and getting anything with ROCM running for me requires that I pass the HSA_OVERRIDE_GFX_VERSION=10.3.0 environmental variable to the related process, otherwise it just refuses to run properly. I wonder if it might be something similar for you too?
5500 here. I can’t use any recent rocm version because the GFX override I use is for a card that apparently has a couple more instructions and the newer kernels instantly crash with an illegal operation exception.
I found a build someone made buried in a docker image and it indeed does work, without override, for the 5500 but it’s using all generic code for the kernels and is like 4x slower than the ancient version.
What’s ultimately the worst thing about this isn’t that AMD isn’t supporting all cards for rocm – it’s that the support is all or nothing. There’s no “we won’t be spending time on this but it passes automated tests so ship it” kind of thing. “oh the new kernels broke that old card tough luck you don’t get new kernels”.
So in the meantime I’m living with the occasional (every couple of days?) freeze when using rocm because I can’t reasonably upgrade. Not just the driver crashes, the kernel tries to restart it, the whole card needs a reset before doing anything but display a vga console.
Yeah, I definitely am not a fan of how AMD handles rocm - there’s so many weird cases of “Well this card should work with rocm, but… [insert some weird quirk that you have to do, like the one I mentioned, or what you’ve run into]”.
Userspace/consumer side I enjoy AMD, but I fully understand why a lot of devs don’t make use of rocm and why Nvidia has such a tight hold on things in the GPU compute world with CUDA.
Ah, strange. I don’t suppose you specifically need a Fedora container? If not, I’ve been using this Ubuntu based distrobox container recipe for anything that requires ROCM and it has worked flawless for me.
If that still doesn’t work (I haven’t actually tried out kobolcpp yet), and you’re willing to try something other than kobolcpp, then I’d recommend the text-generation-webui project which supports a wide array of model types, including the GGUF types that Kobolcpp utilizes. Then if you really want to get deep into it, you can even pair it with SillyTavern (it is purely a frontend for a bunch of different LLM backends, text-generation-webui is one of the supported ones)!
I don’t think so, it’s just what I’m more familiar with and I usually try to avoid apt’s PPA hell as much as I can. But maybe I should try some others, as I couldn’t get Mullvad to run either yet. :/
Text gen web ui I tried quite a while before I went to koboldcpp on my previous distro and I could not get that to run without crashing whenever I tried to generate anything. Sillytavern is my standard frontend that I use, so any text gen software should inherently compatible with that anyway.
Hmm, gotcha. I just tried out a fresh copy of text-gen-webui and it seems like the latest version is borked with ROCM (I get the CUDA error: invalid device function error).
My next recommendation then would be LM Studio which to my knowledge can still output an OpenAI compatible API endpoint to be used in SillyTavern - I’ve used it in the past before and I didn’t even need to run it within Distrobox (I have all of the ROCM stuff installed locally, but I generally run most of the AI stuff in distrobox since it tends to require an older version of Python than Arch is currently using) - it seems they’ve recently started supporting running GGUF models via Vulkan, which I assume probably doesn’t require the ROCM stuff to be installed perhaps?
The advanced configuration settings no longer seem to directly mention GPU acceleration like it used to, however I can see it utilizing GPU resources in nvtop currently, and the speed it was generating at (the one in my screenshot was 83 tokens a second) couldn’t have possibly been done on the CPU so it seems to be fine on my side.
Yeah I hear ya. How do you happen to be running it? I use NixOS and its a challenge there for me but I found atleast some success using docker since the dependencies are so out of control for AI at the moment.
Also give Ollama-rocm + Open Webgui a shot as an alternative to koboldcpp if you cant get atleast some text generation to work because that is the only thing I have got working with rocm.
The trick to nixos, in this instance, is to use a python venv. Python dependencies are fickle and nasty in the first place, triply so when talking about fast-churning AI code, I tried specifying everything with nix, I succeeded, and then you have random comfyui plugins assuming they can get a writeable location by constructing a path from comfyui’s main.py. It’s not worth it: Let python be the only dependency you feed in, let pip and general python jank do the rest.
At the moment I just don’t. I got kobolcpp to run through distrobox / boxbuddy but I can’t get it to compile with rocm, so I can only use CPU generation, which is abysmally slow. Might go back to NovelAI when they release their new model if I can’t find a solution.
What card do you use? I have a 6700XT and getting anything with ROCM running for me requires that I pass the
HSA_OVERRIDE_GFX_VERSION=10.3.0environmental variable to the related process, otherwise it just refuses to run properly. I wonder if it might be something similar for you too?5500 here. I can’t use any recent rocm version because the GFX override I use is for a card that apparently has a couple more instructions and the newer kernels instantly crash with an illegal operation exception.
I found a build someone made buried in a docker image and it indeed does work, without override, for the 5500 but it’s using all generic code for the kernels and is like 4x slower than the ancient version.
What’s ultimately the worst thing about this isn’t that AMD isn’t supporting all cards for rocm – it’s that the support is all or nothing. There’s no “we won’t be spending time on this but it passes automated tests so ship it” kind of thing. “oh the new kernels broke that old card tough luck you don’t get new kernels”.
So in the meantime I’m living with the occasional (every couple of days?) freeze when using rocm because I can’t reasonably upgrade. Not just the driver crashes, the kernel tries to restart it, the whole card needs a reset before doing anything but display a vga console.
Yeah, I definitely am not a fan of how AMD handles rocm - there’s so many weird cases of “Well this card should work with rocm, but… [insert some weird quirk that you have to do, like the one I mentioned, or what you’ve run into]”.
Userspace/consumer side I enjoy AMD, but I fully understand why a lot of devs don’t make use of rocm and why Nvidia has such a tight hold on things in the GPU compute world with CUDA.
6650 XT. Honestly no idea. When I run
make LLAMA_HIPBLAS=1 GPU_TARGETS=gfx1032 -j$(nproc)in the Fedora distrobox on kobolcpp it throws a bunch offatal error: 'hip/hip_fp16.h' file not found 36 | #include <hip/hip_fp16.h>errors and koboldcpp does not give an option to use Vulkan.
Ah, strange. I don’t suppose you specifically need a Fedora container? If not, I’ve been using this Ubuntu based distrobox container recipe for anything that requires ROCM and it has worked flawless for me.
If that still doesn’t work (I haven’t actually tried out kobolcpp yet), and you’re willing to try something other than kobolcpp, then I’d recommend the text-generation-webui project which supports a wide array of model types, including the GGUF types that Kobolcpp utilizes. Then if you really want to get deep into it, you can even pair it with SillyTavern (it is purely a frontend for a bunch of different LLM backends, text-generation-webui is one of the supported ones)!
I don’t think so, it’s just what I’m more familiar with and I usually try to avoid apt’s PPA hell as much as I can. But maybe I should try some others, as I couldn’t get Mullvad to run either yet. :/
Text gen web ui I tried quite a while before I went to koboldcpp on my previous distro and I could not get that to run without crashing whenever I tried to generate anything. Sillytavern is my standard frontend that I use, so any text gen software should inherently compatible with that anyway.
Hmm, gotcha. I just tried out a fresh copy of text-gen-webui and it seems like the latest version is borked with ROCM (I get the
CUDA error: invalid device functionerror).My next recommendation then would be LM Studio which to my knowledge can still output an OpenAI compatible API endpoint to be used in SillyTavern - I’ve used it in the past before and I didn’t even need to run it within Distrobox (I have all of the ROCM stuff installed locally, but I generally run most of the AI stuff in distrobox since it tends to require an older version of Python than Arch is currently using) - it seems they’ve recently started supporting running GGUF models via Vulkan, which I assume probably doesn’t require the ROCM stuff to be installed perhaps?
Might be worth a shot, I just downloaded the latest version (the UI has definitely changed a bit since I last used it) and just grabbed a copy of the Gemma model and ran it, and it seemed to work without an issue for me directly on the host.
The advanced configuration settings no longer seem to directly mention GPU acceleration like it used to, however I can see it utilizing GPU resources in
nvtopcurrently, and the speed it was generating at (the one in my screenshot was 83 tokens a second) couldn’t have possibly been done on the CPU so it seems to be fine on my side.