Comment by aetherspawn

Gemma 30B with 256K context runs at 20 tok/sec on my M3 Max with 128GB RAM so I think there’s something wrong with your setup. This should run at ~30-40 toks. Maybe your inference engine is not optimised for Mac.