Maybe not the best place for this so happy to move it, I wanted to jot down some thoughts on testing inference levers.
K/V quant size, context size and quant, heterogenious K/V quant, convRot optimizaitons ( google turboquant derivatives), model quant, stock jinja template vs auto template (llama.cpp feature), speculative decoding, multi token prediction, quant scheme ( unsloth vs bartowski vs stock), model temp, internal looping, top-k, top-p, min-p, sample ordering, repetition penalty... there are a lot. I wonder if they all resolve into reliability and how well this testing maps to the more traditional perplexity and KLD benchmarks.
I'm also interested in how different attention schemes like SWA play, as well as MOE vs dense model performance.
The meta thing here I often wonder about is: since the API token vendors like anthropic control all of these levers I wonder how often the tweak them and how detectable those tweaks are, particularly knowing how compute constrained they are....
Maybe not the best place for this so happy to move it, I wanted to jot down some thoughts on testing inference levers.
K/V quant size, context size and quant, heterogenious K/V quant, convRot optimizaitons ( google turboquant derivatives), model quant, stock jinja template vs auto template (llama.cpp feature), speculative decoding, multi token prediction, quant scheme ( unsloth vs bartowski vs stock), model temp, internal looping, top-k, top-p, min-p, sample ordering, repetition penalty... there are a lot. I wonder if they all resolve into reliability and how well this testing maps to the more traditional perplexity and KLD benchmarks.
I'm also interested in how different attention schemes like SWA play, as well as MOE vs dense model performance.
The meta thing here I often wonder about is: since the API token vendors like anthropic control all of these levers I wonder how often the tweak them and how detectable those tweaks are, particularly knowing how compute constrained they are....