Comment by lambda
13 hours ago
I've got a 128 GiB unified memory Ryzen Ai Max+ 395 (aka Strix Halo) laptop.
Trying to run LLM models somehow makes 128 GiB of memory feel incredibly tight. I'm frequently getting OOMs when I'm running models that are pushing the limits of what this can fit, I need to leave more memory free for system memory than I was expecting. I was expecting to be able to run models of up to ~100 GiB quantized, leaving 28 GiB for system memory, but it turns out I need to leave more room for context and overhead. ~80 GiB quantized seems like a better max limit when trying not running on a headless system so I'm running a desktop environment, browser, IDE, compilers, etc in addition to the model.
And memory bandwidth limitations for running the models is real! 10B active parameters at 4-6 bit quants feels usable but slow, much more than that and it really starts to feel sluggish.
So this can fit models like Qwen3.5-122B-A10B but it's not the speediest and I had to use a smaller quant than expected. Qwen3-Coder-Next (80B/3B active) feels quite on speed, though not quite as smart. Still trying out models, Nemotron-3-Super-120B-A12B just came out, but looks like it'll be a bit slower than Qwen3.5 while not offering up any more performance, though I do really like that they have been transparent in releasing most of its training data.
There's been some very recent ongoing work in some local AI frameworks on enabling mmap by default, which can potentially obviate some RAM-driven limitations especially for sparse MoE models. Running with mmap and too little RAM will then still come with severe slowdowns since read-only model parameters will have to be shuttled in from storage as they're needed, but for hardware with fast enough storage and especially for models that "almost" fit in the RAM filesystem cache, this can be a huge unblock at negligible cost. Especially if it potentially enables further unblocks via adding extra swap for K-V cache and long context.