[zozbot234]

32B active is nothing special, there's local setups that will easily support that. 1T total parameters ultimately requires keeping the bulk of them on SSD. This need not be an issue if there's enough locality in expert choice for any given workload; the "hot" experts will simply be cached in available spare RAM.

[spmurrayzzz]

When I've measured this myself, I've never seen a medium-to-long task horizon that would have expert locality such that you wouldn't be hitting the SSD constantly to swap layers (not to say it doesn't exist, just that in the literature and in my own empirics, it doesn't seem to be observed in a way you could rely on it for cache performance).

Over any task that has enough prefill input diversity and a decode phase thats more than a few tokens, its at least intuitive that experts activate nearly uniformly in the aggregate, since they're activated per token. This is why when you do something more than bs=1, you see forward passes light up the whole network.

[zozbot234]

> hitting the SSD constantly to swap layers

Thing is, people in the local llm community are already doing that to run the largest MoE models, using mmap such that spare-RAM-as-cache is managed automatically by the OS. It's a drag on performance to be sure but still somewhat usable, if you're willing to wait for results. And it unlocks these larger models on what's effectively semi-pro if not true consumer hardware. On the enterprise side, high bandwidth NAND Flash is just around the corner and perfectly suited for storing these large read-only model parameters (no wear and tear issues with the NAND storage) while preserving RAM-like throughput.

[spmurrayzzz]

I've tested this myself often (as an aside: I'm in said community, I run 2x RTX Pro 6000 locally, 4x 3090 before that), and I think what you said re: "willing to wait" is probably the difference maker for me.

I can run Minimax 2.1 in 5bpw at 200k context fully offloaded to GPU. The 30-40 tk/s feels like a lifetime for long horizon tasks, especially with subagent delegation etc, but it's still fast enough to be a daily driver.

But that's more or less my cutoff. Whenever I've tested other setups that dip into the single and sub-single digit throughput rates, it becomes maddening and entirely unusable (for me).

[RandomTeaParty]

What is bpw?

[spmurrayzzz]

Bits per weight, its an average precision across all the weights. When you quantize these models, they don't just used a fixed precision size across all model layers/weights. There's a mix and it varies per quant method. This is why you can get bit precision that arent "real" in a strict computing sense.

e.g. A 4-bit quant can have half the attention and feed forward tensors in Q6, and the rest in Q4. Due to how block-scaling works, those k-quant dtypes (specifically for llama.cpp/gguf) have larger bpw than they suggest in their name. Q4 is around ~4.5 bpw, and Q6 is ~6.5.

[1dom]

I never said it was special.

I was trying to correct the record that a lot of people will be using models of this size locally because of the local LLM community.

The most commonly downloaded local LLMs are normally <30b (e.g. https://huggingface.co/unsloth/models?sort=downloads). The things you're saying, especially when combined together, make it not usable by a lot of people in the local LLM community at the moment.