Use _cast_input_dtype in FourierFTLinear.forward

[Chessing234]

Bug

FourierFTLinear.forward in src/peft/tuners/fourierft/layer.py casts the input to the adapter's dtype with a raw x.to(delta_w.dtype):

for active_adapter in self.active_adapters:
    if active_adapter not in self.fourierft_spectrum.keys():
        continue

    delta_w = self.get_delta_weight(active_adapter)
    x = x.to(delta_w.dtype)                        # <- raw cast
    result = result + F.linear(x, delta_w)

This bypasses BaseTunerLayer._cast_input_dtype, which is the method that consults cast_input_dtype_enabled (toggled by the peft.helpers.disable_lora_input_dtype_casting context manager introduced in #2353 and generalized to non-LoRA tuners in #2433).

Root cause

#2433 consolidated dtype handling across additive tuners onto self._cast_input_dtype. Every comparable tuner's forward uses the helper:

• lora/layer.py:967 — x = self._cast_input_dtype(x, lora_A.weight.dtype)
• waveft/layer.py:283 — x = self._cast_input_dtype(x, delta_w.dtype)
• hra/layer.py:260, 445 — x = self._cast_input_dtype(x, new_weight.dtype)
• psoft/layer.py:447 — x_cast = self._cast_input_dtype(x, A.dtype)

FourierFT was missed in that sweep, so toggling disable_lora_input_dtype_casting silently has no effect on FourierFT layers — the cast still happens unconditionally.

Fix

Route the cast through self._cast_input_dtype to match the established pattern. Behavior is identical when the casting context manager is not used (both call x.to(dtype)); the difference is that users can now opt out as documented.

[BenjaminBossan]

@Chessing234 Thanks for the PR, the change is correct. Instead of opening a separate PR for each PEFT method, could you please group all changes into a single PR?