Ram Disk to speed up inference
This revision is from 2024/10/29 05:42. You can Restore it.
Using a RAM disk to speed up inference.
Does not work.
LLM's require CPU/GPU operations.
1 token per second, and longer inferences are slower. 1 inferences comes back in about 1 or 2 minutes, for instance I say: Hello, LLM returns a minute later with the finished response. Too slow.
Tested with 14B quant 8
Aim for 15-30 tokens per second