I wasn't able to have very satisfying success until I bit the bullet and threw a GPU at the problem. Found an actually reasonably priced A4000 Ada generation 20GB GPU on eBay and never looked back. I still can't run the insanely large models, but 20GB should hold me over for a while, and I didn't have to upgrade my 10 year old Ivy Bridge vintage homelab.
replyAre mac kernels optimized compared to CUDA kernels? I know that the unified GPU approach is inherently slower, but I thought a ton of optimizations were at the kernel level too (CUDA itself is a moat)
replyDepending on what you do. If you are doing token generations, compute-dense kernel optimization is less interesting (as, it is memory-bounded) than latency optimizations else where (data transfers, kernel invocations etc). And for these, Mac devices actually have a leg than CUDA kernels (as pretty much Metal shaders pipelines are optimized for latencies (a.k.a. games) while CUDA shaders are not (until cudagraph introduction, and of course there are other issues).
replyMac kernels are almost always compute shaders written in Metal. That's the bare-minimum of acceleration, being done in a non-portable proprietary graphics API. It's optimized in the loosest sense of the word, but extremely far from "optimal" relative to CUDA (or hell, even Vulkan Compute).
replyMost people will not choose Metal if they're picking between the two moats. CUDA is far-and-away the better hardware architecture, not to mention better-supported by the community.