It seems it still has the old drivers on the flatpak sandboxes.

Maybe this could help, updating the runtime drivers to match your updated drivers and remove the not used ones:

flatpak update -y && flatpak remove --unused -y

So, I’m going to say what I saw specifically in these years. I’m arguing in a more nuanced approach to valve.

For windows compatibility, my opinion is that in the late 2010s (Wine is the overall GOAT, but I’m talking about recent history), the main driver of linux gaming was dxvk, there was gallium for directx 9 but… It was not that good. Dxvk changed the game, the moment I bought a GT 1030 and ran GTA IV on linux (like this guy did 6 years ago), I was impressed, holy shit. Vulkan is low level enough that it can be an “easy” translation layer for any graphics API.

Now we have d9vk, d7vk, vkd3d, etc, all based on that idea.

Mono did this for all .net framework programs (or CLR based programs, if I’m remembering correctly). Gecko did this to Internet Explorer, etc.

Then came Valve and integrated them all in a package called proton, then came the Steam Deck, a console on a muted Linux distro.

To me, Valve absorbed the glory that came from dxvk. They did a good thing by integrating it all and paying the devs, which is a good thing, but at the end, people forgot what doitsujin did. Now it seems that they did some shady stuff to waydroid too…

I’m not kidding, that guy (dxvk creator) deserves all the respect in the world.

And then there is the gambling, Valve took CS:GO, Dota 2 and TF2 and turned them into polymarket. That’s a big no no. They improved the early lootbox and item market mechanics (in their favour) to what it is today.

They helped enshittify gaming, they absorbed the attribution (even if helping, because let’s get real, they only built the bundle), but they sold a Linux console and helped pay some of the devs involved.

Edit: wording

It’s more like steering or hijacking the “privileged” part of the models execution, as it continues from these tokens in the reasoning stage.

It’s just an example, probably people can come with more interesting applications, steering the reasoning traces into different things.

I never used te 12B model, but some smaller models I tested seemed to think and loop a lot more than qwen 35B and gemma 26B. Ternary qwen 27B looped as well.

Maybe parameter amount or the quantization might be causing it (I’m using the QAT from unsloth in this test).

Yeah, the reads are A LOT faster than my ssd and the latency is nil.

A line of memory from Intel that’s like NAND, but it’s not NAND. It’s 3D Xpoint, very good at random reads or sequential reads.

sudo hdparm -t /dev/nvme0n1p1

/dev/nvme0n1p1:
 Timing buffered disk reads: 2682 MB in  3.00 seconds = 893.33 MB/sec

sudo hdparm -t /dev/sda

/dev/sda:
 Timing buffered disk reads: 670 MB in  3.01 seconds = 222.89 MB/sec

Edit: when the swap was at my main drive, the sequential reads were like 140 MB/s

I only tested with those applications, but I could benchmark using my ssd vs the drive tomorrow or later. Download a RAM heavy game or something.

Do you have a recommendation?