I can run GLM 4.6 on a Ryzen/single RTX 3090 desktop at 7 tokens/s, and it blows lesser API models away. I can run 14-49Bs (or GLM Air) in more utilitarian cases that do just fine.
And I can reach for free/dirt cheap APIs called locally when needed.
But again, it’s all ‘special interest tinkerer’ tier. You can’t do that with ollama run, you have to mess with exotic libraries and tweaked setups and RAG chains to squeeze out that kind of performance. But all that getting simplified is inevitable.
I’ll look into it. OAI’s 30B model is the most I can run in my MacBook and it’s decent. I don’t think I can even run that on my desktop with a 3060 GPU. I have access to GLM 4.6 through a service but that’s the ~350B parameter model and I’m pretty sure that’s not what you’re running at home.
It’s pretty reasonable in capability. I want to play around with setting up RAG pipelines for specific domain knowledge, but I’m just getting started.
The $342 Crucial kit is kinda a no-brainer for 128. Its timings aren’t great when overclocked, but it’s 5600 MHz out of the box, low voltage, and significantly cheaper per gigabyte than many 64GB/96GB kits. See for yourself:
I appreciate it. I’m not going to overclock. I used to do that but these days I value stability over maximum performance. I’ll go with your suggestion, thank you.
You don’t have to overclock, but I’d at least look at your mobo’s settings. Many mobos set really, really bad, non stock settings by default, especially with XMP memory.
An example: they might default to 1.3V VSOC which is absolutely a “default overclock” and is going to make idle power skyrocket, and your CPU potentially unstable because infinity fabric doesn’t like that. For reference, I personally wouldn’t go over 1.2V VSOC myself and shoot for like 1.1V.
Also, if you don’t turn on XMP at least (aka the RAM’s rated speed), they will run at some slow default and hurt your inference speed rather significantly.
I’m have to check. It’s a pro, not air, but I think it’s only 40GB total. I’m really new to Macs so the memory situation is unclear. I requested it at work specifically for its capability to run local AI.
Not anymore.
I can run GLM 4.6 on a Ryzen/single RTX 3090 desktop at 7 tokens/s, and it blows lesser API models away. I can run 14-49Bs (or GLM Air) in more utilitarian cases that do just fine.
And I can reach for free/dirt cheap APIs called locally when needed.
But again, it’s all ‘special interest tinkerer’ tier. You can’t do that with
ollama run, you have to mess with exotic libraries and tweaked setups and RAG chains to squeeze out that kind of performance. But all that getting simplified is inevitable.I’ll look into it. OAI’s 30B model is the most I can run in my MacBook and it’s decent. I don’t think I can even run that on my desktop with a 3060 GPU. I have access to GLM 4.6 through a service but that’s the ~350B parameter model and I’m pretty sure that’s not what you’re running at home.
It’s pretty reasonable in capability. I want to play around with setting up RAG pipelines for specific domain knowledge, but I’m just getting started.
It is. I’m running this model, with hybrid CPU+GPU inference, specifically: https://huggingface.co/Downtown-Case/GLM-4.6-128GB-RAM-IK-GGUF
You can likely run GLM Air on your 3060 desktop if you have 48GB+ RAM, or a smaller MoE easily. Heck. I’ll make a quant just for you, if you want.
Depending on the use case, I’d recommend ERNIE 4.5 21B (or 28B for vision) on your Macbook, or a Qwen 30B variant. Look for DWQ MLX quants, specifically: https://huggingface.co/models?sort=modified&search=dwq
I’m going to upgrade my ram shortly because I found a bad stick and I’m down to 16GB currently. I’ll see if I can swing that order this weekend.
To what?
64G would be good, as that’s enough to fit GLM Air. There are some good 2x64GB 6000Mhz kits for 128GB as well.
I’ll see about 128, then, but I’ll probably do 64. Just depends on cost. Any recs?
For DDR5? Depends how much you care about latency:
https://pcpartpicker.com/products/memory/#ff=ddr5&Z=131072002&B=1000000000%2C1250000000&sort=price&page=1
The $342 Crucial kit is kinda a no-brainer for 128. Its timings aren’t great when overclocked, but it’s 5600 MHz out of the box, low voltage, and significantly cheaper per gigabyte than many 64GB/96GB kits. See for yourself:
https://www.igorslab.de/en/when-size-is-all-that-matters-crucial-2x-64-gb-ddr5-5600-kit-with-new-micron-32-gbit-ics-in-test-incl-gaming-and-overclocking/4/
The overclockability matters even less if you are on a 7000 series CPU.
I got the 1.25V Flare X5 kit because I wanted tighter timings for sim games, albeit at a MUCH lower price ($390) than it is now.
RAM prices seem to be rising (hence the price of my kit spiked by $200), so now is not a bad time to buy.
I appreciate it. I’m not going to overclock. I used to do that but these days I value stability over maximum performance. I’ll go with your suggestion, thank you.
You don’t have to overclock, but I’d at least look at your mobo’s settings. Many mobos set really, really bad, non stock settings by default, especially with XMP memory.
An example: they might default to 1.3V VSOC which is absolutely a “default overclock” and is going to make idle power skyrocket, and your CPU potentially unstable because infinity fabric doesn’t like that. For reference, I personally wouldn’t go over 1.2V VSOC myself and shoot for like 1.1V.
I’d recommend Buildzoid’s videos:
https://youtu.be/dlYxmRcdLVw
https://youtu.be/Xcn_nvWGj7U
And Igor’s Lab for general text info.
Also, if you don’t turn on XMP at least (aka the RAM’s rated speed), they will run at some slow default and hurt your inference speed rather significantly.
I’m have to check. It’s a pro, not air, but I think it’s only 40GB total. I’m really new to Macs so the memory situation is unclear. I requested it at work specifically for its capability to run local AI.