DOC Fix broken docstring examples in peft_model and tuner models

[blipbyte]

Fixes #3644. Approved by @BenjaminBossan in
#3644 (comment).

Docstrings only. No runtime code is touched.

peft_model.py — four examples that cannot run

Four of the six PeftModelFor* examples raise TypeError when run as
written, because they pass config fields that were removed in 2023:

• postprocess_past_key_value_function, removed in a7dd0347 (2023-02-08),
  appears in the PeftModelForSequenceClassification, PeftModelForCausalLM
  and PeftModelForTokenClassification examples.
• enable_lora, removed in c21afbe8 (2023-03-28), appears in the
  PeftModelForSeq2SeqLM example.

TypeError: PrefixTuningConfig.__init__() got an unexpected keyword argument 'postprocess_past_key_value_function'
TypeError: LoraConfig.__init__() got an unexpected keyword argument 'enable_lora'

Both fields were deleted rather than renamed, so there is no replacement key
and the lines are simply removed. The remaining keys still describe the
configuration each example demonstrates.

All six classes are pulled into the API reference via [[autodoc]], so these
are the examples a reader copies off the docs site.

PeftModelForTokenClassification had a second problem behind the first: it
imported AutoModelForSequenceClassification but called
AutoModelForTokenClassification, so it would still have failed with a
NameError once the config was valid. The import now matches the class.

lora, hira, beft — examples that mislead

These three construct the tuner class directly:

>>> lora_model = LoraModel(model, config, "default")

That silently writes a base-model checkpoint instead of an adapter: BaseTuner.__getattr__ forwards to the wrapped module, so save_pretrained resolves to PreTrainedModel.save_pretrained rather than failing. Switched to get_peft_model, matching the form you asked for in the #3254 review for ia3 and adalora; these three were never swept.

Also removed the unused PeftModel import from the lora and hira k-bit
examples.

beft additionally passed peft_type="Beft", which BeftConfig.__post_init__
overwrites with PeftType.BEFT. It was the only example where the value was
also wrong — the enum is "BEFT", not "Beft", and the mismatch is silent.
The same redundant peft_type= appears in prompt_tuning, p_tuning,
ia3, adalora and prefix_tuning, but there the string matches the enum,
so I left them out of scope. Happy to remove those too if you'd prefer.

Testing

Nothing in CI executes these examples — make test runs pytest tests/, and
no doctest collection touches src/. That is why they sat broken since 2023.
So I ran them directly.

Before: 4 of 6 fail with the TypeErrors above. After: all six build their
config, and five run end to end. The sixth, PeftModelForCausalLM, needs
gpt2-large, which is not in my local cache.

The three tuner examples were run as edited too. Each now returns a PeftModel, so save_pretrained writes an adapter. The direct form resolves save_pretrained to PreTrainedModel.save_pretrained and writes a base-model checkpoint instead — no error either way.

Also run: make quality (ruff, format, doc-builder, doc coverage 131/131) and
pytest tests/test_config.py tests/test_decoder_models.py — 7064 passed,
2718 skipped.

Per the issue discussion, no test that executes the examples was added —
@BenjaminBossan said it wasn't needed.

AI assistance: I used Claude Code to help audit and draft this. I reviewed
every changed line, ran every command above myself, and can defend the change
end-to-end.