`get_peft_model` and `add_adapter` do not preserve eval mode for newly injected adapter modules

System Info

• PEFT: 0.21.0 (main, commit 514b9e6d)
• Transformers: 5.3.0
• Python: 3.12.3

Who can help?

@BenjaminBossan

Reproduction

I noticed that injecting a LoRA adapter into an already-evaluated model can silently enable adapter dropout.

import torch
from torch import nn
from peft import LoraConfig, get_peft_model


class Tiny(nn.Module):
    def __init__(self):
        super().__init__()
        self.lin = nn.Linear(4, 4, bias=False)

    def forward(self, x):
        return self.lin(x)


config = LoraConfig(
    target_modules=["lin"],
    r=2,
    lora_alpha=2,
    lora_dropout=0.5,
    init_lora_weights=False,
)

model = get_peft_model(Tiny().eval(), config)
layer = model.base_model.model.lin

with torch.no_grad():
    layer.lora_A["default"].weight.fill_(1.0)
    layer.lora_B["default"].weight.fill_(1.0)

x = torch.ones(4, 4)

with torch.no_grad():
    output_1 = model(x)
    output_2 = model(x)

print(model.training)  # True
print(layer.lora_dropout["default"].training)  # True
print(torch.equal(output_1, output_2))  # False

The base model was in evaluation mode before adapter injection, so I expected the resulting PEFT model and adapter modules to remain in evaluation mode. Instead, the newly created LoRA dropout module stays in training mode, making repeated inference nondeterministic.

The same happens when adding an adapter to an existing evaluated PEFT model:

model = get_peft_model(
    Tiny(),
    LoraConfig(target_modules=["lin"], lora_dropout=0.0, init_lora_weights=False),
).eval()

model.add_adapter("other", config)
model.set_adapter("other")

print(model.training)  # False
print(model.base_model.model.lin.training)  # False
print(model.base_model.model.lin.lora_dropout["other"].training)  # True

In this case, the parent layer is correctly in evaluation mode, but the newly added adapter dropout is not.

Expected behavior

Expected behavior is that newly injected adapter modules inherit the current training/evaluation state of the model or parent module. This seems different from the existing adapter state-restoration issue in #3507 because the problem occurs during adapter injection rather than disable_adapter() context handling.

I’d be happy to open a PR with a shared fix and regression tests for both initial injection and add_adapter. Before doing that, would you prefer preserving the full incoming PEFT model’s mode, or explicitly inheriting the training state from each parent module when new adapter modules are created?