I have notes:
-
You need to switch to a sparse system with obfuscations like matrix rotation (depends on the stage) and dummy batches to stop nodes from tracking personal information.
-
Someone committed enough could use a quantized model and get just as many credits as if they ran at full precision. There should be bad actor detection beyond what you are doing; regular injection of canary tasks could go far.
If you had one person run the router and distributed experts among the rest of your users, you would need at least 385 people with more than 48GB of memory each to run KIMI K2.5+. Without more protections, few will contribute towards putting a dent in that lower bound.
I have notes:
You need to switch to a sparse system with obfuscations like matrix rotation (depends on the stage) and dummy batches to stop nodes from tracking personal information.
Someone committed enough could use a quantized model and get just as many credits as if they ran at full precision. There should be bad actor detection beyond what you are doing; regular injection of canary tasks could go far.
If you had one person run the router and distributed experts among the rest of your users, you would need at least 385 people with more than 48GB of memory each to run
KIMI K2.5+. Without more protections, few will contribute towards putting a dent in that lower bound.