SSD offload is always a possibility with good software support. Of course you might easily object that the model would not be "running" then, more like crawling. Still you'd be able to execute it locally and get it to respond after some time.
replyMeanwhile we're even seeing emerging 'engram' and 'inner-layer embedding parameters' techniques where the possibility of SSD offload is planned for in advance when developing the architecture.
For conversational purposes that may be too slow, but as a coding assistant this should work, especially if many tasks are batched, so that they may progress simultaneously through a single pass over the SSD data.
replyThree hour coffee break while the LLM prepares scaffolding for the project.
replyLike computing used to be. When I first compiled a Linux kernel it ran overnight on a Pentium-S. I had little idea what I was doing, probably compiled all the modules by mistake.
replyI remember that time, where compiling Linux kernels was measured in hours. Then multi-core computing arrived, and after a few years it was down to 10 minutes.
replyWith LLMs it feels more like the old punchcards, though.
At least the compiler was free
replyThe point of doing local inference with huge models stored on an SSD is to do it free, even if slow.
replyYou are just trading opex for capex. Local GPUs aren't free.
replyTrue, but this is not only a trade-off between opex and capex.
replyLocal inference using open weight models provides guaranteed performance which will remain stable over time, and be available at any moment.
As many current HN threads show, depending on external AI inference providers is extremely risky, as their performance can be degraded unpredictably at any time or their prices can be raised at any time, equally unpredictably.
Rather, Imagine you have 2-3 of these working 24/7 on top of what you're doing today. What does your backlog look like a month from now?
replyBatching many disparate tasks together is good for compute efficiency, but makes it harder to keep the full KV-cache for each in RAM. You could handle this in an emergency by dumping some of that KV-cache to storage (this is how prompt caching works too, AIUI) and offloading loads for that too, but that adds a lot more overhead compared to just offloading sparsely-used experts, since KV-cache is far more heavily accessed.
reply