I mean, there’s a real issue.

Say you were china, or the EU, or any other country/bloc and basically your entire youth was addicted to Twitter, Facebook or whatever, and officially manipulatable by the US government…. And you got into a real conflict. Maybe even a hot war.

Wouldn’t you be worried about the US propagandizing your population?

I would.

The US government’s solution is completely dysfunctional and not getting at the root of the issue because they are afraid of reducing the power projection of big tech, among other things. But the core issue doesn’t need to be trivialized.

Honestly, most LLMs suck at the full 128K. Look up benchmarks like RULER.

In my personal tests over API, LLama 70B is bad out there. Qwen (and any fine tune based on Qwen Instruct, with maybe an exception or two) not only sucks, but is impractical past 32K once its internal rope scaling kicks in. Even GPT-4 is bad out there, with Gemini and some other very large models being the only usable ones I found.

So, ask yourself… Do you really need 128K? Because 32K-64K is a boatload of code with modern tokenizers, and that is perfectly doable on a single 24G GPU like a 3090 or 7900 XTX, and that’s where models actually perform well.

Late to this post, but shoot for and AMD Strix Halo or Nvidia Digits mini PC.

Prompt processing is just too slow on Apple, and the Nvidia/AMD backends are so much faster with long context.

Otherwise, your only sane option for 128K context in a server with a bunch of big GPUs.

Also… what model are you trying to use? You can fit Qwen coder 32B with like 70K context on a single 3090, but honestly its not good above 32K tokens anyway.

Unfortunately Nvidia is, by fair, the best choice for local LLM coder hosting, and there are basically two tiers:

• Buy a used 3090, limit the clocks to like 1400 Mhz, and then host Qwen 2.5 coder 32B.

• Buy a used 3060, host Arcee Medius 14B.

Both these will expose an OpenAI endpoint.

Run tabbyAPI instead of ollama, as it’s far faster and more vram efficient.

You can use AMD, but the setup is more involved. The kernel has to be compatible with the rocm package, and you need a 7000 card and some extra hoops for TabbyAPI compatibility.

Aside from that, an Arc B570 is not a terrible option for 14B coder models.

It’s not functional yet.

The UN that the United States basically set up, and hosts. That literally the rest of the world agrees with on this issue…

Yep. I didn’t mean to process shame you or anything, just trying to point out obscure but potentially useful projects most don’t know about :P

https://en.m.wikipedia.org/wiki/External_memory_algorithm

Unfortunately that’s not really relevant to LLMs beyond inserting things into the text you feed them. For every single word they predict, they make a pass through the multi-gigabyte weights. Its largely memory bound, and not integrated with any kind of sane external memory algorithm.

There are some techniques that muddy this a bit, like MoE and dynamic lora loading, but the principle is the same.

A1111

Eh, this is a problem because the “engine” is messy and unoptimized. You could at least try to switch to the “reforged” version, which might preserve extension compatibility and let you run features like torch.compile.

Oh you should be able to batch the heck out of that on a 4080. Are you not using HF diffusers or something?

I’d check out stable-fast if you haven’t already:

https://github.com/chengzeyi/stable-fast

VoltaML is also old at this point, but it has really fast AITemplate implementation for SD 1.5: https://github.com/VoltaML/voltaML-fast-stable-diffusion

Oh, 16GB should be plenty for SDXL.

For flux, I actually use a script that quantizes it down to 8 bit (not FP8, but true quantization with huggingface quanto), but I would also highly recommend checking this project out. It should fit everything in vram and be dramatically faster: https://github.com/mit-han-lab/nunchaku

You don’t want it to anyway, as “automatic” spillover with an LLM painfully slow.

The RAM/VRAM split is manually configurable in llama.cpp, but if you have at least 10GB VRAM, generally you want to keep the whole model within that.

Try a new quantization as well! Like an IQ4-M depending on the size of your GPU, or even better, an 4.5bpw exl2 with Q6 cache if you can manage to set up TabbyAPI.

Depends which 14B. Arcee’s 14B SuperNova Medius model (which is a Qwen 2.5 with some training distilled from larger models) is really incrtedible, but old Llama 2-based 13B models are awful.

No, all the weights, all the “data” essentially has to be in RAM. If you “talk to” a LLM on your GPU, it is not making any calls to the internet, but making a pass through all the weights every time a word is generated.

There are system to augment the prompt with external data (RAG is one word for this), but fundamentally the system is closed.

Oh I didn’t mean “should cost $4000” just “would cost $4000”

Ah, yeah. Absolutely. The situation sucks though.

I wish that the vram on video cards was modular, there’s so much ewaste generated by these bottlenecks.

Not possible, the speeds are so high that GDDR physically has to be soldered. Future CPUs will be that way too, unfortunately. SO-DIMMs have already topped out at 5600, with tons of wasted power/voltage, and I believe desktop DIMMs are bumping against their limits too.

But look into CAMM modules and LPCAMMS. My hope is that we will get modular LPDDR5X-8533 on AMD Strix Halo boards.

GDDR is actually super cheap! I think it would only be like another $75 on paper to double the 4090’s VRAM to 48GB (like they do for pro cards already).

Nvidia just doesn’t do it for market segmentation. AMD doesn’t do it for… honestly I have no idea why? They basically have no pro market to lose, the only explanation I can come up with is that their CEOs are colluding because they are cousins. And Intel doesn’t do it because they didn’t make a (consumer) GPU that was eally worth it until the B580.

The issue with Macs is that Apple does price gouge for memory, your software stack is effectively limited to llama.cpp or MLX, and 70B class LLMs do start to chug, especially at high context.

Diffusion is kinda a different duck. It’s more compute heavy, yes, but the “generally accessible” software stack is also much less optimized for Macs than it is for transformers LLMs.

I view AMD Strix Halo as a solution to this, as its a big IGP with a wide memory bus like a Mac, but it can use the same CUDA software stacks that discrete GPUs use for that speed/feature advantage… albeit with some quirks. But I’m willing to put up with that if AMD doesn’t price gouge it.

second-hand TPU

From where? I keep a look out for used Gaudi/TPU setups, but they’re like impossible to find, and usually in huge full-server configs. I can’t find Xeon Max GPUs or CPUs either.

Also, Google’s software stack isn’t really accessible. TPUs are made for internal use at Google, not for resale.

You can find used AMD MI100s or MI210s, sometimes, but the go-to used server card is still the venerable Tesla P40.

You can’t let it overflow if you’re using LLMs on windows. There’s a toggle for it in the Nvidia settings, and get llama.cpp to offload though its settings (or better yet, use exllama instead).

But…. Yeah. Qwen 32B fits in 24GB perfectly, and it’s great, but 72B really feels like the intelligence tipping point where I can dump so many API models, and that won’t fit in 24GB.