Hugging Face Transformers 5.16

Apply the Python 3 prompt first. These rules target Transformers 5.16.1 within the 5.16 release line, using Python 3.10–3.14 and PyTorch. The package requires PyTorch 2.5 or newer; guarded torch.load paths require at least 2.6. Use maintained, patched dependencies compatible with the deployed hardware.

Model artifacts and configuration

• Allowlist repositories and pass a full commit ID as revision to each from_pretrained call. Pin separately loaded tokenizers, processors, configurations, adapters, and their base models; one model revision does not pin another repository.
• Construct pipeline with an explicit approved task and reviewed model and preprocessing objects. Keep model IDs, local paths, model_kwargs, devices, dtypes, cache locations, and loading flags in server configuration rather than forwarding request dictionaries.
• Inspect config.json and adapter configuration before model construction. Allowlist architectures, nested model references, and quantization backends; bound layer counts, vocabulary sizes, tensor dimensions, and artifact sizes before allocation.
• Set use_safetensors=True for model loading and require approved safetensors artifacts. The default preference can fall back to PyTorch weights; also check explicit filenames, config.transformers_weights, and every selected shard rather than treating the flag as a universal format check.
• Keep weights_only=True and preserve check_torch_load_is_safe() failures. Never set weights_only=False or patch out the guard to make a checkpoint load; PyTorch 2.5 compatibility does not make