PiSSA / MiCA / CorDA init bypasses quantized-layer guard on 4-bit float-storage layers

Summary

pissa_init, mica_init, and corda_init (src/peft/tuners/lora/layer.py) gate on weight.dtype in [float32, float16, bfloat16] — but a bitsandbytes 4-bit layer with float quant storage (bnb_4bit_quant_storage float) reports a float dtype while its data is packed nibbles. The guard passes, SVD runs over packed-nibbles-as-floats, and the residual is written back into the Params4bit. The same file already does this correctly in reset_lora_parameters (:386-396: class detection via get_bnb_param_type + dequantize_module_weight). (LoftQ was initially suspected too, but loftq_utils.py:81 dequantizes properly — excluded.)

Minimal reproduction (CPU; proves the guard bypass — packed-data corruption itself needs a CUDA quant state)

import torch, bitsandbytes as bnb
from torch import nn
from peft import LoraConfig, get_peft_model
from peft.utils.integrations import get_bnb_param_type

class QNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.lin = nn.Linear(16, 16)
        self.lin.weight = bnb.nn.Params4bit(self.lin.weight.data, quant_type="nf4")

w = QNet().lin.weight
print(type(w).__name__, w.dtype)          # Params4bit torch.float32  <- the confusion
print(get_bnb_param_type(w))              # 4bit (class detection works)
get_peft_model(QNet(), LoraConfig(r=4, target_modules=["lin"], init_lora_weights="pissa"))
print("no TypeError — guard bypassed on a quantized-param class")

On a CUDA 4-bit layer the consequence is worse than a bypass: weight.to(torch.float32) materializes packed values as floats, the SVD factorizes quantization noise, and weight.data = residual stores garbage back into the Params4bit.

Root cause

Three dtype-only guards (pissa_init, mica_init ~:470, corda_init ~:499) where a class check is needed. get_bnb_param_type (class-name based, integrations.py:129) already exists and the correct dequant path (dequantize_module_weight) is used two functions away.

Expected behavior

SVD-based inits either dequantize-then-proceed (matching reset_lora_parameters) or raise the existing TypeError on quantized layers — never silently factorize packed storage.

Proposed fix

Replace the three dtype guards with the get_bnb_param_type + dequantize_module_weight pattern (or reject quantized layers explicitly), plus a CPU-runnable guard test using a float-storage Params4bit exactly as above. Happy to PR after a nod.

Environment: Python 3.12, torch 2.13.0+cpu, bitsandbytes 0.50.2, transformers 5.15.1, peft @ ab2db1e0.