karolherbst

karolherbst OP , vor 15 Tagen

that wasn't too bad.. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29896

karolherbst , vor 15 Tagen

@sima I'm just getting a rice cooker honestly

karolherbst OP , vor 15 Tagen

Yes, I have "ipv4.dns" and "ipv4.dns-search" set in NetworkManager, but somehow it still gets dropped after connecting to the VPN.

karolherbst OP , vor 15 Tagen

ohh, turns out a negative "ipv4.dns-priority" value on the VPN connections makes all other connections lose their DNS settings. So now I have it set to 0 for the VPN as I don't care about DNS leaks, or rather I prefer if all non redhat.com DNS requests go through my local DNS server 🙃

I might have to play around with the priorities a bit more, but finally that's figured out :)

karolherbst OP , vor 15 Tagen

oh, and there also seems to be a special "~." entry for ipv4.dns-search just redirecting everything to that connection and that was also set on my work VPN connection. So double the fun.

Anyway, I've fixed my DNS setup for realz now and I hope it stays that way as in the past I had to fight with my work DNS server breaking apart randomly and other fun things breaking my internet.

karolherbst OP , vor 15 Tagen

oh, one thing I should try out as well is, what if I'm on VPN with work and my home while on a different network and stuff. But I think I should be able to figure those things out now at least :)

karolherbst OP , vor 15 Tagen

@airlied could just add more custom domains to the vpn connection then, it's a list of values

karolherbst OP , vor 15 Tagen

@serebit it's not even bad stuff specifically, just uhm.. negative stuff and I kinda want to cut down on that.

But it often helps to just write things down without posting it.

karolherbst OP , vor 16 Tagen

@sonny yeah, that's I guess the least we could do with the current state.

But I do wonder what if Gnome, KDE, Cosmic, others... would all chime in a bit of money and maybe that's enough for a full/half time position.

But then there is sooooooo much to do and whatever we'd spend that money on, somebody will have some problems with the decision being made.

But then desktops participating is already a big help.

karolherbst OP , vor 16 Tagen

@sonny However... FDO being in a neutral position might allow to fund work specific desktops couldn't due, due to regulations on "non-profit" and them not wanting to risk losing their financial status by funding work not directly aligning with their goals.

karolherbst OP , vor 16 Tagen

@mupuf oh, I've never heard of that, but that also sounds like a terrible reason to not do it.

karolherbst OP , vor 16 Tagen

@mupuf I mean, I'm sure everybody had good people in mind we would be able to trust with that money. That's really the wrong question here.

The question is rather, do we really want to keep such a policy, because somebody abused it in the past?

Of course, us being a non-profit, we have to make sure the money is well spent and everything. And the person we'd pay also needs a mentor who would take care of that part.

The actual question is just if we want to do it or not.

karolherbst OP , vor 15 Tagen

@mupuf it doesn't have to be spec work. And you have to trust people at some point.

Sure the org has to decide on what to spend money on, but that doesn't have to be the board itself.

We could appoint a group of people we trust to decide what's worth spending money on which also addresses dire needs in the Linux desktop. Maybe even focus on underfunded areas.

Maybe the CoC team can do it to improve inclusion of people we left behind.

Again, we just have to want to do it.

karolherbst OP , vor 16 Tagen

@VegaHarmonia I knew I was doing something wrong

karolherbst OP , vor 16 Tagen

"tcache_thread_shutdown(): unaligned tcache chunk detected" 😢

and it happens on application exit and like.. ahhhhh....

karolherbst OP , vor 15 Tagen

anyway, I think this happens, because my threads continue to run even after everything else was released.

So yeah, that's not great, but also shouldn't cause any issues. What I think is causing issues is mostly just how things run at application exit and sometimes the order could mess things up?

Not quite sure :) Maybe it's just LLVM trashing the memory, but I'm also too lazy to dig why exactly this happens.

Anyway, I should terminate my threads early regardless.

karolherbst OP , vor 15 Tagen

tldr: JoinHandle doesn't stop your threads on drop.

karolherbst OP , vor 15 Tagen

that wasn't too bad: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29880

Though I still don't know what causing the memory issue, but now I'll just blame "undefined order" when exit is run and move on with life 🙃

karolherbst OP , vor 15 Tagen

@permik maybe? That certainly would make my use case easier to deal with.

karolherbst , vor 16 Tagen

@krh isn't that the auto-vectorization programming model?

karolherbst , vor 18 Tagen

@lethalbit kinda looks like a spacy navigation thing for space ships we all use in 10000 years

karolherbst OP , vor 18 Tagen

@bashbaug Do you know something or do I have to dig through Intel's compute runtime git to find the proper tests for it?

karolherbst OP , vor 18 Tagen

@bashbaug mostly just clDeviceMemAllocINTEL, clSetKernelArgMemPointerINTEL and potentially clEnqueue*Intel for now I think.

But my biggest concern is still, that even that much requires SVM semantics and that's really tough to implement in every driver.

And I think it would be helpful to have something like cl_ext_buffer_device_address taken care of first, as this could also be used by e.g. chipstar instead of SVM/USM.

And I think SyCL impls could also use it instead of SVM/USM.

karolherbst OP , vor 18 Tagen

@Man2Dev oh sure, but I was more interested in knowing what people are using and what acceleration APIs it's using.

Like.. what tools are people using to translate e.g. speech to text, or other related things.

karolherbst OP , vor 18 Tagen

@Man2Dev right. I think it should work with rusticl on all supported drivers atm, though it does need a very very very beefy GPU to get reliable results afaik

Or rather my Intel iGPU isn't beefy enough.

karolherbst OP , vor 18 Tagen

@Man2Dev right.. just doesn't really change the fact, that you still need a beefy GPU/system to make it somewhat working.

I've seen some of the llama.cpp and whisper.cpp OpenCL code and there is definitely a lot of room for improvement, like.. not busy waiting on the CPU side 🙃

but yeah... I kinda hope we'd get to the point where those things don't really rely on whatever hardware you have, but just works with the hardware you already got.

karolherbst OP , vor 19 Tagen

@dotstdy I do have a passive cooling solution which is able to deal with 100W TDP :)

And I was considering something like the i7-14700T with a turbo TDP of 106W so that would be fine.

I'm just mostly concerned about idle power draw as that setup would run 24/7.

karolherbst OP , vor 19 Tagen

@dotstdy I was considering the Streacom DB4 one, which seems to be good enough.

My system won't be a peak load all the time, so even if it drops after a while it's fine.

karolherbst OP , vor 19 Tagen

@dotstdy there are some people going crazy with that one e.g. https://www.reddit.com/r/silentpc/comments/13xdu6c/streacom_db4_ft_i913900f_rtx_4070/

karolherbst OP , vor 19 Tagen

@dotstdy because you get more perf by having more threads. I could get a "14600T" but then I have 20 instead of 28 threads and just save a little bit of money. Something like the 14900T would be overkill tho.

Storage and all the other stuff will be the more expensive part of all of this anyway.

karolherbst OP , vor 18 Tagen

@darkdragon @shironeko @serebit it's deterministic and you can do something about it though.

But it's also something you can plan ahead of time or verify what the legal requirements are, etc...

The issue with floats mostly is, that it's hard to know that any operation is causing any numeric errors.

karolherbst OP , vor 15 Tagen

update on my budget docker cluster: I've added the watchtower docker container, which updates other docker containers when the used image tag got updated on dockerhub 🙃

sadly the acme-companion tags don't have a "stable" thing, so I might have to update their tags manually from time to time, but at least all the nginx ones use "stable-alpine" and that's great :)