A random number I checked feels kinda optimistic, or there's something wrong with my setup; Gemma 26B-A4B on a 16GB 5060Ti is reported to get 21-88tok/s with Q4_K_M, meanwhile I can just barely eke out 70tok/s with Q3 (which fully fits in vram), and 40tok/s with Q4 (with recently-merged MoE cache on llama-cpp). I guess the average fits, but the peak feels implausible.
No clue about details, but Sebastian Raschka's site reports its kv cache usage at 210kB per token, which is like 6x more than this site claims ("0.273 GB per 8k"). Also, the site claims qwen 27b to use 2x more memory per token than gemma, which is the opposite from what I know.
No clue about details, but Sebastian Raschka's site reports its kv cache usage at 210kB per token, which is like 6x more than this site claims ("0.273 GB per 8k"). Also, the site claims qwen 27b to use 2x more memory per token than gemma, which is the opposite from what I know.