DarkThoughts

I’m not using Fedora, I’m using Bazzite, which is immutable based on SilverBlue. I use an Arch distrobox for this since I can’t really install anything directly into the system. The script is what I tried originally in a Fedora distrobox which did not work at all.

ggml_cuda_compute_forward: ADD failed
CUDA error: shared object initialization failed
  current device: 0, in function ggml_cuda_compute_forward at ggml/src/ggml-cuda.cu:2365
  err
ggml/src/ggml-cuda.cu:107: CUDA error

I didn’t do anything past using yay to install the AUR koboldcpp-hipblas package, and customtkinter, since the UI wouldn’t work otherwise. The koboldcpp-rocm page very specifically does not mention any other steps in the Arch section and the AUR page only mentions the UI issue.

Well, anything remotely raunchy gets a “I cannot participate in explicit content” default reply.

I am using the rocm install of koboldcpp but as said, the ROCm option errors out with a CUDA error for some reason.

I distrohopped so much after each previous distro eventually broke and me clearly not being smart enough to recover. I’m honestly kinda sick of it, even if the immutable nature also annoys the shit out of me.

My GPU is a 6650 XT, which should in principle work with ROCm.

Which model specifically are you recommending? Llama-3.1-8B-Lexi-Uncensored-V2-GGUF? Because the original meta-llama ones are censored to all hell and Huggingface is not particularly easy to navigate, on top of figuring out the right model size & quantization being extremely confusing.

I just can’t get ROCm / gpu generation to work on Bazzite, like at all. It seems completely cursed. I tried koboldcpp through a Fedora distrobox and it didn’t even show any hardware options. Tried through an Arch AUR package through distrobox and the ROCm option is there but ends with a CUDA error. lol The Vulkan option works but seems to still use the CPU more than the GPU and is consequently still kinda slow and I struggle to find a good model for my 8GB card. Fimbulvetr-10.7B-v1-Q5_K_M for example was still too slow to be practical.

Tried LM Studio directly in Bazzite and it also just uses the CPU. It also is very obtuse on how to connect to it with SillyTavern, as it asks for an API key? I managed it once in the past but I can’t remember how but it also ended up stopping generating anything after a few replies.

Krita’s diffusion also only runs on the CPU, which is abysmally slow, but I’m not sure if they expect Krita to be build directly on the system for ROCm support to work.

I’m not even trying to get SDXL or something to run at this point, since that seems to be still complicated enough even on a regular distro.

No, like a duck.

It is probably dead but Easy Diffusion is imo the easiest for image generation.

KoboldCPP can be a bit weird here and there but was the first thing that worked for me for local text gen + gpu support.

No, I’m actually with them on that one. The he / they issue in of itself is tiny, I agree, and if they’d just changed it from gendered to gender neutral language then nobody would’ve even cared. Most of us tend to write in a gendered way out of habit or because we think about our own gender, and in a casual conversation that isn’t that important. But this is about a piece of software that, surely, is not just meant for male audiences. It’s just unprofessional to address someone as male by default. Most importantly though, being this stubborn on having the user specifically male is just a weird hill to die on, but even weirder if that particular action is the one that is actually causing the drama - which they allegedly claim wanting to prevent by dismissing “politics”. And I’m sorry, but changing a “he” to “they” is not politics, it’s just including non male users. Nothing more, nothing less. So why is it such an issue to not just address specifically male users? It really only would be because those people hold some very questionable views, which, in my opinion, clash heavily with the whole concept of free and open source software, which is supposedly for everyone. So if your actions and views are this flawed, how can you be trusted on such an important project?

Also, in regards to this news… “no code from rivals” also is just a stupid thing to say and do. There’s plenty of good open source code that they could and probably even SHOULD use. But whatever. I’m not gonna support this project and predict it will fail anyway.

I’m guessing it runs on Python, given the name. Either way, I think it’s generally great that more and more mod managers finally are natively Linux compatible. Still hoping for a native one for Bethesda games though.

Crucial MX 500 & Samsung 870 Evo are reliable / good & “cheap” SATA SSDs. For NVMe there’s the WD Blue SN570 and the Kioxia Exceria G2 but keep in mind that they tend to have smaller storage sizes too and depending on your use case you might not really notice a performance difference between SATA and NVMe anyway. Personally, I stay away from all native Chinese products. They tend to have terrible quality and fall apart quickly. I’m sure there’s exceptions here and there but wading through all the garbage and having to buy twice does not seem worth it and I rather support that country as little as possible anyway.

Those are bots. They all use the same or very similar titles. This is the third account that I’ve seen popping up on All, all of which post similar posts with some mature women and one also spammed some comments, also linking to a Telegram channel.