313·
8 days agoStandardised censorship of general communication is scary Orwellian nonsense
𞋴𝛂𝛋𝛆
- 4 Posts
- 27 Comments
Joined 2 years ago
Cake day: June 9th, 2023
You are not logged in. If you use a Fediverse account that is able to follow users, you can follow this user.
11·20 days agodeleted by creator
I haven’t looked into the issue of PCIe lanes and the GPU.
I don’t think it should matter with a smaller PCIe bus, in theory, if I understand correctly (unlikely). The only time a lot of data is transferred is when the model layers are initially loaded. Like with Oobabooga when I load a model, most of the time my desktop RAM monitor widget does not even have the time to refresh and tell me how much memory was used on the CPU side. What is loaded in the GPU is around 90% static. I have a script that monitors this so that I can tune the maximum number of layers. I leave overhead room for the context to build up over time but there are no major changes happening aside from initial loading. One just sets the number of layers to offload on the GPU and loads the model. However many seconds that takes is irrelevant startup delay that only happens once when initiating the server.
So assuming the kernel modules and hardware support the more narrow bandwidth, it should work… I think. There are laptops that have options for an external FireWire GPU too, so I don’t think the PCIe bus is too baked in.
Anything under 16 is a no go. Your number of CPU cores are important. Use Oobabooga Textgen for an advanced llama.cpp setup that splits between the CPU and GPU. You'll need at least 64 GB of RAM or be willing to offload layers using the NVME with deepspeed. I can run up to a 72b model with 4 bit quantization in GGUF with a 12700 laptop with a mobile 3080Ti which has 16GB of VRAM (mobile is like that).
I prefer to run a 8×7b mixture of experts model because only 2 of the 8 are ever running at the same time. I am running that in 4 bit quantized GGUF and it takes 56 GB total to load. Once loaded it is about like a 13b model for speed but is ~90% of the capabilities of a 70b. The streaming speed is faster than my fastest reading pace.
A 70b model streams at my slowest tenable reading pace.
Both of these options are exponentially more capable than any of the smaller model sizes even if you screw around with training. Unfortunately, this streaming speed is still pretty slow for most advanced agentic stuff. Maybe if I had 24 to 48gb it would be different, I cannot say. If I was building now, I would be looking at what hardware options have the largest L1 cache, the most cores that include the most advanced AVX instructions. Generally, anything with efficiency cores are removing AVX and because the CPU schedulers in kernels are usually unable to handle this asymmetry consumer junk has poor AVX support. It is quite likely that all the problems Intel has had in recent years has been due to how they tried to block consumer stuff from accessing the advanced P-core instructions that were only blocked in microcode. It requires disabling the e-cores or setting up a CPU set isolation in Linux or BSD distros.
You need good Linux support even if you run windows. Most good and advanced stuff with AI will be done with WSL if you haven’t ditched doz for whatever reason. Use https://linux-hardware.org/ to see support for devices.
The reason I mentioned avoid consumer e-cores is because there have been some articles popping up lately about all p-core hardware.
The main constraint for the CPU is the L2 to L1 cache bus width. Researching this deeply may be beneficial.
Splitting the load between multiple GPUs may be an option too. As of a year ago, the cheapest option for a 16 GB GPU in a machine was a second hand 12th gen Intel laptop with a 3080Ti by a considerable margin when all of it is added up. It is noisy, gets hot, and I hate it many times, wishing I had gotten a server like setup for AI, but I have something and that is what matters.
Abstract solutions for content recognition with a bot on a server is not a platform specific issue. The dev is skilled and likely on Matrix too.
It is a bot that identifies CSAM images. They are a very skilled dev. The problem is content recognition on a server. So in abstract, it is the same problem.
Search for posts or contact db0. IIRC they worked with LW admin and others to create a filter for this using a very small AI model. It should be on their Git.
I’d maybe pay to have all IR junk and screen side camera removed. I’d definitely feel it adds value if they would clamp the speaker and mic to the audible range with a hardware filter. Give me a hardware switch to remove modem power and I’d be happy.
Plan 9
Need max AVX instructions. Anything with P/E cores is junk. Only enterprise P cores have the max AVX instructions. When P/E are mixed the advanced AVX is disabled in microcode because the CPU scheduler is unable to determine if a process thread contains an AVX instruction and there is no asymmetrical scheduler that handles this. Prior to early 12k series Intel, the microcode for P enterprise could allegedly run if swapped manually. This was “fused off” to prevent it, probably because Linux could easily be adapted to asymmetrical scheduling but Windows would probably not. The whole reason W11 had to be made was because of the E-cores and the way the scheduler and spin up of idol cores works, at least according to someone on Linux Plumbers for the CPU scheduler ~2020. There are already asymmetric schedulers in Android ARM.
Anyways I think it was on Gamer’s Nexus in the last week or two that Intel was doing some all P core consumer stuff. I’d look at that. According to chips and cheese, the primary CPU bottleneck for tensors is the bus width and clock management of the L2 to L1 cache.
I do alright with my laptop, but haven’t tried R1 stuff yet. The 70B llama2 stuff that I ran was untenable for CPU only with a 12700 with just CPU. It is a little slower than my reading pace when split with a 16 GB GPU, and that was running a 4 bit quantization version.
Not unless an http port is open too. If the only port is https, you have to have the certificate. Like with my AI stuff it acts like the host is down if I try to connect with http. You have to have the certificate to decrypt anything at all from the host.
Sorta, you have to install your certificate authority into the browser and it might complain about verifying that but it will still connect with the encryption.
deleted by creator
I mean more like a self signed TLS certificate with your own host manually set in the browser. Then only make the TLS port available, or something like that. If you have access to both(all) devices, you should be able to fully encrypt by bruit force and without registering the certificate with anyone. That is what I do with AI at home.
I’ve half ass thought about this but never have tried to actually self host. If you have access to all devices, why not just use your own self signed certificates to encrypt everything and require the certificate for all connections? Then there is never a way to log in or connect right? The only reason for any authentication is to make it possible to use any connection to dial into your server. So is that a bug or a feature. Maybe I’m missing something fundamental in this abstract concept that someone will tell me?
I’ve tried 3 times so far in Python/gradio/Oobabooga and never managed to get certs to work or found a complete visual reference guide that demonstrates a complete working example like what I am looking for in a home network. (Only really commenting to subscribe to watch this post develop, and solicit advice:)
I’m no expert. Have you looked at the processors that are used and the RAM listed in the OpenWRT table? That will tell you the real details if you look it up. Then you can git clone OpenWRT, and use the gource utility to see what kind of recent dev activity has been happening in the source code.
I know, it’s a bunch of footwork. But really, you’re not buying brands and models. You’re buying one of a couple dozen processors that have had various peripherals added. The radios are just integrated PCI bus cards. A lot of options sold still come with 15+ year old processors.
The last time I looked (a few months ago) the Asus stuff seemed interesting for a router. However, for the price, maybe go this route: https://piped.video/watch?v=uAxe2pAUY50
Open source project device support pages are always my first stop. If you have access to a git repo for the project, use gource to visualize who is doing what and where within the project recently. This will make it obvious what hardware the main devs are invested in the most, and therefore what will have the best support and user experience.
Thanks. I went down this rabbit hole already with large FreeCAD assemblies where single threaded operations had an enormous amount of non voluntary context switching. It helped some in that instance. I don’t have a benchmark, but I reduced the NVCS by an order of magnitude using Tuna to set affinity and pinning. Editing a large assembly in the middle of the tree went from 5++ minutes to around 3-4 minutes.
I’m not claiming I know better, or that this is a bright idea. As mentioned, it’s an abstract curiosity. Playing with FreeCAD got me interested in the OS on a deeper level, and got me reading as much as I can about the CFS and others. The LLM toolchain I’m working on is to help me follow the CS curriculum posted freely online from UC Berkeley. Ultimately I plan to use a langchain database to help me follow along with the books and transcribed lectures. It is just a hobby interest.
Thanks for the insights!
It is not about the initial application. There is enormous power in the ability to control such an app and roll changes in over time