FIX Mixed: honor bias='<prefix>_only' for all compatible tuners (not just lora_only)

[DaoyuanLi2816]

What does this PR do?

PeftMixedModel's _mark_only_adapters_as_trainable hardcodes bias == "lora_only":

elif bias == "lora_only":
    ...
else:
    raise ValueError(f"Requested bias: {bias}, is not implemented.")

But MixedModel.COMPATIBLE_TUNER_TYPES includes OFT, and OFTConfig accepts bias="oft_only" (a valid Literal["none", "all", "oft_only"]). So a mixed model with an OFT adapter and bias="oft_only" raises ValueError: Requested bias: oft_only, is not implemented., even though the identical config trains fine in a standalone OFT model.

The body of that branch is already generic — it marks the bias of every compatible Layers instance trainable, not just LoRA — so only the guard was too narrow. This aligns it with BaseTuner._mark_only_adapters_as_trainable, which uses bias.endswith("_only") (and removes the now-resolved # TODO: check if this is needed for other supported types).

How it was verified

• Repro (CPU): a SimpleNet mixed model with OFTConfig(target_modules=["lin1"], oft_block_size=8, bias="oft_only") raises the ValueError on main and trains its bias after the fix; standalone OFT with the same config already works.
• Added test_bias_only_honored_for_non_lora_tuner — it fails on main with the ValueError above and passes with this change.
• tests/test_mixed.py passes (33 passed); ruff check / ruff format --check clean.

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.

[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]

Good catch. LGTM.