A tokenizer_config.json with extra_special_tokens as a JSON array instead of
an object makes transformers raise "'list' object has no attribute 'keys'" in
SpecialTokensMixin._set_model_specific_special_tokens, before the model can
load (seen on unsloth/Qwen3.6-35B-A3B-MLX-8bit on Apple Silicon).
Add install_extra_special_tokens_compat() to coerce a non-dict value to {}
during tokenizer init, and call it on the MLX and transformers load paths
before from_pretrained. Idempotent, version guarded, and a no-op on
well-formed configs.
|
||
|---|---|---|
| .. | ||
| data_recipe | ||
| export | ||
| inference | ||
| rag | ||
| training | ||
| __init__.py | ||
| _torchao_stub.py | ||
| import_guards.py | ||
| tool_healing.py | ||