So limiting max context length also reduces VRAM needs a bit? If cache is 20% of total, 1/10th of context as a limit would mean 18% total memory reduction.
replyYup exactly, in principle it helps with both inference speed by reducing memory bandwidth usage and also reduces the memory footprint of your kvcache.
reply