Temp PR that investigates whether scaling to 200 perturbations adds value - #56

Draft

gdadlaney wants to merge 3 commits into

dg-fork-mainfrom

cos_sim_analysis

Commits on Apr 9, 2026

Fix major bug in _distribute_weights() to send weights to trainer that has not contributed to the current model version when var = bad.
gdadlaney
committed
Fix the problem of GRAD_POOL going as None for clients that had stale updates & needed weights & the updated GRAD_POOL to train
gdadlaney
committed
Temp commit that investigates whether scaling to 200 perturbations adds value. Moved perturbation generation to Cuda from CPU to reduce worst waruntimes by 8x
gdadlaney
committed