wassnameandClaudypoo 57c8d4b166 add clamp apply mode: pin v-component to C instead of accumulating ...

clamp: y += (C - <y,v_hat>)v_hat at all positions -- bounded perturbation
regardless of generation length, vs add's per-step accumulation via KV cache.
C=0 is directional ablation. Smoke (Qwen3-0.6B, happy/joy): clamp C=+20 stays
coherent and on-concept (drifts to 'happiness and joy of my childhood', in
Chinese) while add C=+8 already degenerates to 'joyjoyjoy...'.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

2026-07-10 14:42:22 +08:00

..

scratch

wip

2026-07-10 14:36:55 +08:00

fit.py

wip

2026-07-10 14:36:55 +08:00

smoke.py

add clamp apply mode: pin v-component to C instead of accumulating

2026-07-10 14:42:22 +08:00

u4_step3_fit4b.py

U4 step3: dim_batch 8->4 + expandable_segments after 2nd OOM

2026-07-10 14:10:39 +08:00

u4_step3_retry.sh

U4 step3: auto-resume wrapper to self-heal through user's GPU bursts

2026-07-10 14:31:20 +08:00