Legacy model compatibility¶

Many trust_remote_code models were written for transformers 4.x and break on 5.x. The built-in core/compat.py shims patch the gaps automatically so the model's own code runs unchanged - no version pinning.

Breakage	Shim
Moved tokenizer symbols (`PaddingStrategy`, …)	re-exposed on `transformers.tokenization_utils`
Removed `AutoModelForVision2Seq`	recreated as an alias of `AutoModelForImageTextToText`, registered for `auto_map` dispatch
`tie_weights()` signature drift	legacy overrides made kwarg-tolerant
Hard `timm` version pins	`compat.spoof_timm_version()` context manager
Broken `torchcodec` (audio decode)	disabled so stdlib/soundfile paths win

compat.apply() is invoked automatically by the tool and provider. Trigger it explicitly if you need to:

use_transformers(action="compat", parameters={"timm_version": "0.9.16"})

These shims are generic - they help any 4.x-era custom-code model, not just OpenVLA.