menaerus

35 ms·

721.
by menaerus 2y ago
Not true. icecream handles that case very well. Any machine in the network can trigger the build. And all the machines in the network can register themselves as build nodes. It also integrates with the ccache.
722.
by menaerus 2y ago
distcc was really slow IME as well. It did not know how to optimally distribute the work across different build nodes. AFAIR distcc simply assumed that all build nodes are created equal and it also disregarded the load on the nodes resultin
723.
by menaerus 2y ago
> The tldr is that a lot of the heavy lifting was done by an algorithm used called Deductive Database + Algebraic Relations. Let me make sure I understood this correctly. Author from reddit formed his opinion based on "examining the
724.
by menaerus 2y ago
> What is happening is the C module is getting a Rust wrapper, Yes, I understand that. I read the argument on the mailing list. The issue is that changing the DMA code will in turn potentially break the drivers written in Rust and the ma
725.
by menaerus 2y ago
Sure, engineering is all about the tradeoffs, there's no black and white. So, what is an advantage of mixing two different languages over one kernel module, in this example DMA? What problem does introducing another language solve and
726.
by menaerus 2y ago
So do you disagree with the fact that merging and testing the code will be much more difficult or you simply choose to ignore it for the sakes of idealogical wars?
727.
by menaerus 2y ago
It has certainly not been made simply because of the fact that Linus is not a boss to the people who are contributing or maintaining the code. There was however a decision to have Rust as an experiment in the kernel but not such that it wil
728.
by menaerus 2y ago
The reason is entirely technical and I think it has been formulated clearly - maintaining cross-language codebase, or one of the most important kernel modules FWIW in this case, certainly does not make things simpler but only more complex -
729.
by menaerus 2y ago
You don't need RNG since the whole transformer is an extremely large floating-point arithmetic unit. A wild guess - how about the source of non-determinism is coming from the fact that, on the HW level, tensor execution order is not gu
730.
by menaerus 2y ago
> innovation of the architecture was making it computation efficient to train. and > researchers demonstrated that simply scaling up training with more data yielded better models and > The fact that hardware was then optimized for
731.
by menaerus 2y ago
Language is improving (?), although IME it went besides the point I'm finding new features to be less useful for every day code. I'm perfectly happy with C++17/20 for 99% of the code I write. And keeping the backwards compati
732.
by menaerus 2y ago
What line of thinking you're referring to? Transformers were aimed to solve the "context" problem and authors, being aware that RNNs don't scale at all neither do they solve that particular problem, had to come up with t
733.
by menaerus 2y ago
I stand corrected, not sure why I was under the impression that this paper came out of DeepMind lab.
734.
by menaerus 2y ago
https://arxiv.org/abs/1706.03762
735.
by menaerus 2y ago
You can choose to be somewhat ignorant of the current state in AI, about which I could also agree that at certain moments it appears totally overhyped, but the reality is that there hasn't been a bigger technology breakthrough probably
736.
by menaerus 2y ago
If I had a startup where would I apply to get into this program? Where did your company apply to?
737.
by menaerus 2y ago
The criticism here is the system, not the people. And you seem to defend the system which has been proved to be inefficient and shady. You also seem to take this personally which you really shouldn't. People here are also sharing their
738.
by menaerus 2y ago
I won't disagree but it's the most effective way of getting an edge over the competition. Without the "reward and punishment" system in place, how else do you get that extra mile from your engineers? How else do you find
739.
by menaerus 2y ago
I am familiar with Horizon and EU grant funds and how these projects are managed - basically a big waste of time and money without concrete results with plenty of bureaucracy crawling the project ideas and resources to the halt. I am not sa
740.
by menaerus 2y ago
I largely see this type of collaboration as a very inefficient form of a distributed company (team) where members of that team do not have other incentive but to (mostly) collect points on research papers. There is no incentive to actually
741.
by menaerus 2y ago
So, if one has a well thought-through idea, what is the process of getting the resources ($$$) from OpenEuroLLM and the compute from EuroHPC? How do I become a partner as a long-standing engineer with plenty of industry practice in research
742.
by menaerus 2y ago
Another idea: AFAICS there have been a few memory-bw zen-related bugs reported to likwid and, in particular, https://github.com/RRZE-HPC/likwid/issues/535 may suggest that you could be hitting a similar bug b
743.
by menaerus 2y ago
I see. I am out of other ideas besides trying to play with BIOS tweaks wrt memory and CPU. I can see that there are plenty of them, for worse or for the better. At a quick glance, some of them look interesting such as "Workload tuning&
744.
by menaerus 2y ago
Mixed results. I suspect you might have an ES (engineering sample) of your CPU.
745.
by menaerus 2y ago
Yes, you're correct that your CPU has 8 CCDs but the bw with 8 threads is already too low. Those 8 cores should be able to get you at roughly half of the theoretical bw. 8x zen5 cores for comparison can reach the ~230 GB/s mark. C
746.
by menaerus 2y ago
To saturate the bandwidth, you would need ~16 zen4 cores but you could first try running lkwid -t load -i 100 -w S0:5GB:8:1:2 and see what you get. I think you should be able to get somewhere around ~200 GB/s.
747.
by menaerus 2y ago
> but self-hosting these wont be a viable alternative to using providers like OpenAI for business applications for a while. Why not? While 3-4 tok/s is still on the lower end, it is still usable to the point that I can use it for an
748.
by menaerus 2y ago
Assuming that you populated the channels correctly, which I believe you did, I can only think that this issue could be related to the motherboard itself or RAM. I think you could start by measuring the single-core RAM bandwidth and latency.
749.
by menaerus 2y ago
Right, perhaps I was a little bit too vague but what I was trying to say is that by merely sampling the CPU cycles we cannot infer that the foo() was burning CPU because it was memory-bound and which in itself is not an artifact of foo() im
750.
by menaerus 2y ago
Blog is packed with information, thanks! Isn't it the case that from stack traces it is rather impossible to read that function foo() is burning CPU cycles because it is memory-bound? And the reason could be rather somewhere else and n

More ›