FIX Call super().__post_init__() in CPTConfig and RandLoraConfig

[DaoyuanLi2816]

What does this PR do?

CPTConfig and RandLoraConfig override __post_init__ but do not call super().__post_init__(), so PeftConfig.__post_init__ never runs for them. That base method:

• auto-fills peft_version (if self.peft_version is None: self.peft_version = self._get_peft_version()), and
• validates task_type.

As a result, CPTConfig(...).peft_version and RandLoraConfig(...).peft_version are left as None, so the saved adapter_config.json for CPT / RandLoRA adapters does not record the PEFT version that every other adapter type records. The base task_type validation is also skipped.

Of the tuner configs that override __post_init__, all but these two call super().__post_init__(). This PR adds the call (first, matching the convention) so CPT and RandLoRA behave the same.

How it was verified

Before:

• CPTConfig(task_type="CAUSAL_LM").peft_version → None
• RandLoraConfig().peft_version → None

After:

• both → e.g. '0.19.2.dev0@UNKNOWN', and the value round-trips through save_pretrained / from_pretrained.
• The configs' existing __post_init__ logic is unaffected (CPT still sets is_prompt_learning, num_virtual_tokens, …; RandLoRA still sets peft_type and the target_modules set).
• tests/test_cpt.py and tests/test_randlora.py pass (15 passed).
• ruff check passes on both files.

Note: #3227 also edits randlora/config.py (the save_projection field default), on different lines — no overlap with this change.

AI assistance

AI assistance was used to prepare this change. I reviewed every changed line, verified the behavior as described above, and take responsibility for the patch.

[BenjaminBossan]

Thanks for catching this omission.

I checked why our tests didn't catch this and found that ALL_CONFIG_CLASSES in tests/tst_config.py doesn't include those two (there might be others that are missing). Could you please add them to the test matrix?

[DaoyuanLi2816]

Thanks! Added them. Diffing ALL_CONFIG_CLASSES against all public PeftConfig subclasses turned up six missing: CPTConfig and RandLoraConfig, plus DeloraConfig, OFTConfig, PveraConfig, WaveFTConfig. All six are in the matrix now.

Two needed a little handling so the generic tests stay green:

• CPTConfig only supports task_type=CAUSAL_LM (enforced in its __post_init__), so it's pinned in mandatory_kwargs and skipped in test_valid_task_type / test_invalid_task_type. It's still exercised by the round-trip tests (test_methods / test_save_pretrained / test_to_dict / pickle / copy / ...), which is what would have caught the original __post_init__ omission.
• OFTConfig's from_pretrained has a backwards-compatibility guard that raises before the generic load path, so it's skipped in the four from_pretrained-based tests.

Locally: tests/test_config.py → 1137 passed, 12 skipped; ruff clean.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks for taking care of syncing the config tests. The PR LGTM.