finish bnbEmbedding4bit's __init__ method

src/peft/tuners/lora/bnb.py

@@ -19,7 +19,7 @@
 from bitsandbytes.nn import Params4bit
 import torch
 import torch.nn.functional as F
-from torch import Tensor
+from torch import Tensor, device
 
 from peft.import_utils import is_bnb_4bit_available, is_bnb_available
 from peft.tuners.tuners_utils import BaseTunerLayer, check_adapters_to_merge
@@ -368,8 +368,36 @@ def __repr__(self) -> str:
             return "lora." + rep
 
     class bnbEmbedding4bit(torch.nn.Embedding): # WORK IN PROGRESS
-        def __init__(self):
-            raise NotImplementedError
+        def __init__(
+                self,
+                num_embeddings: int,
+                embedding_dim: int,
+                padding_idx: Optional[int] = None,
+                max_norm: Optional[float] = None,
+                norm_type: float = 2.0,
+                scale_grad_by_freq: bool = False,
+                sparse: bool = False,
+                _weight: Optional[Tensor] = None,
+                device: Optional[device] = None,
+                compute_dtype=None,
+                compress_statistics=True,
+                quant_type='fp4' #
+            ) -> None:
+                super().__init__(
+                    num_embeddings,
+                    embedding_dim,
+                    padding_idx,
+                    max_norm,
+                    norm_type,
+                    scale_grad_by_freq,
+                    sparse,
+                    _weight,
+                    device=device
+                )
+                self.weight = Params4bit(self.weight.data, requires_grad=False, compress_statistics=compress_statistics, quant_type=quant_type) #
+                self.compute_dtype = compute_dtype #
+                self.compute_type_is_set = False #
+                self.quant_state = None #
 
         def reset_parameters(self) -> None:
             torch.nn.init.xavier_uniform_(self.weight)