fix: handle FP8 model weights in LoRA adapters and merge#182

Open

shifusen329 wants to merge 7 commits intop-e-w:masterfrom

shifusen329:master

Commits on Feb 19, 2026

fix: handle FP8 model weights in LoRA adapters and merge

shifusen329
and
claude
committed
fix: resolve PEFT target modules by identity and dequantize FP8 weights

shifusen329
and
claude
committed
fix: handle zero base refusals in evaluator scoring

shifusen329
and
claude
committed
fix: treat NaN KL divergence as inf to guide Optuna away from bad trials

shifusen329
and
claude
committed
fix: lower max_weight search range to avoid model destabilization

shifusen329
and
claude
committed
fix: upcast logits to float32 before log_softmax for KL divergence

shifusen329
and
claude
committed
fix: clamp log_softmax output to prevent infinite KL divergence

shifusen329
and
claude
committed