Warm-start: overwrite a freshly-built LoRA's A/B leaves with the weights saved in
dir/adapters.safetensors (saveAdapter's ${modulePath}.lora_a/b names), so a run
continues from a checkpoint's WEIGHTS. The optimizer state and LR schedule restart —
Adam re-warms its moments in a few steps — so this is a weights warm-start, not a
bit-perfect resume. The checkpoint's rank/targets must match this LoRA (shapes are
checked). MUST be called BEFORE the optimizer is built so it tracks the loaded
handles. Returns the number of targets loaded.
Warm-start: overwrite a freshly-built LoRA's A/B leaves with the weights saved in
dir/adapters.safetensors(saveAdapter's${modulePath}.lora_a/bnames), so a run continues from a checkpoint's WEIGHTS. The optimizer state and LR schedule restart — Adam re-warms its moments in a few steps — so this is a weights warm-start, not a bit-perfect resume. The checkpoint's rank/targets must match this LoRA (shapes are checked). MUST be called BEFORE the optimizer is built so it tracks the loaded handles. Returns the number of targets loaded.