1 files changed, 38 insertions, 39 deletions

diff --git a/models/lora.py b/models/lora.py
index 01a540b..e506cff 100644
--- a/models/lora.py
+++ b/models/lora.py
@@ -1,8 +1,8 @@
 from typing import Optional
+import math
 
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
 
 
 class LoraLayer():
@@ -42,14 +42,12 @@ class LoraEmbedding(nn.Embedding, LoraLayer):
             self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout, merge_weights=merge_weights
         )
 
-        self.register_buffer('trainable_ids', self.weight.new_zeros(num_embeddings, dtype=torch.long))
-        self.trainable_ids -= 1
+        self.register_buffer('trainable_ids', self.weight.new_zeros(num_embeddings, dtype=torch.long) - 1)
 
-        if r > 0:
-            self.lora_A = nn.ParameterList()
-            self.lora_B = nn.Linear(r, embedding_dim, bias=False)
-            self.scaling = self.lora_alpha / self.r
-            self.weight.requires_grad = False
+        self.lora_A = nn.ParameterList()
+        self.lora_B = nn.Linear(r, embedding_dim, bias=False)
+        self.scaling = self.lora_alpha / self.r
+        self.weight.requires_grad = False
 
         self.reset_parameters()
 
@@ -70,8 +68,9 @@ class LoraEmbedding(nn.Embedding, LoraLayer):
         else:
             nn.init.zeros_(new_emb.weight.data)
         new_emb.weight.data[:n, :] = self.weight.data[:n, :]
-        new_emb.lora_A = self.lora_A
-        new_emb.lora_B = self.lora_B
+        for param in self.lora_A:
+            new_emb.lora_A.append(param)
+        new_emb.lora_B.weight[:].data = self.lora_B.weight[:].data
         new_emb.trainable_ids[:n] = self.trainable_ids[:n]
 
         return new_emb
@@ -87,60 +86,60 @@ class LoraEmbedding(nn.Embedding, LoraLayer):
         n2 = n1 + new_ids.shape[0]
         self.trainable_ids[new_ids] = torch.arange(n1, n2)
         for _ in new_ids:
-            self.lora_A.append(nn.Parameter(self.weight.new_zeros(self.r)))
+            w = self.weight.new_zeros(self.r)
+            self.lora_A.append(w)
+
+        if len(self.lora_A) > 1:
+            elems = torch.stack([param for param in self.lora_A])
+            nn.init.kaiming_uniform_(elems, a=math.sqrt(5))
 
     def get_weights(self, input_ids: torch.Tensor):
         if len(input_ids.shape) != 1:
             return torch.stack([self.get_weights(batch) for batch in input_ids])
 
-        trainable_ids = self.trainable_ids[input_ids]
-        mask = ~(trainable_ids == -1)
-        trainable_ids = trainable_ids[mask]
-
         weights = self.weight.new_zeros((input_ids.shape[0], self.embedding_dim))
-        elems = [self.lora_A[id] for id in trainable_ids]
 
-        if len(elems) != 0:
-            w = self.lora_B(self.lora_dropout(torch.stack(elems))) * self.scaling
-            weights[mask] = w.to(dtype=weights.dtype)
+        if not self.merged:
+            trainable_ids = self.trainable_ids[input_ids]
+            mask = ~(trainable_ids == -1)
+            elems = [self.lora_A[id] for id in trainable_ids[mask]]
+
+            if len(elems) != 0:
+                w = self.lora_B(self.lora_dropout(torch.stack(elems))) * self.scaling
+                weights[mask] = w.to(dtype=weights.dtype)
 
         return weights
 
     def persist(self):
-        if self.r > 0:
-            weights = self.get_weights(torch.arange(self.trainable_ids.shape[0]))
-            self.weight.data += weights
-            self.trainable_ids[:] = -1
-            self.lora_A = nn.ParameterList()
+        self.weight.data += self.get_weights(torch.arange(self.trainable_ids.shape[0]))
+        self.trainable_ids[:] = -1
+        self.lora_A = nn.ParameterList()
+        nn.init.zeros_(self.lora_B.weight)
 
     def reset_parameters(self):
         nn.Embedding.reset_parameters(self)
-        if hasattr(self, 'lora_A'):
+        if hasattr(self, "lora_A"):
             self.trainable_ids[:] = -1
             self.lora_A = nn.ParameterList()
             nn.init.zeros_(self.lora_B.weight)
 
     def train(self, mode: bool = True):
         nn.Embedding.train(self, mode)
-        if self.merge_weights and self.merged:
-            if self.r > 0:
-                weights = self.get_weights(torch.arange(self.trainable_ids.shape[0]))
-                self.weight.data -= weights
+        self.lora_A.train(mode)
+        self.lora_B.train(mode)
+        if not mode and self.merge_weights and not self.merged:
+            self.weight.data += self.get_weights(torch.arange(self.trainable_ids.shape[0]))
+            self.merged = True
+        elif self.merge_weights and self.merged:
+            self.weight.data -= self.get_weights(torch.arange(self.trainable_ids.shape[0]))
             self.merged = False
 
     def eval(self):
         nn.Embedding.eval(self)
-        if self.merge_weights and not self.merged:
-            if self.r > 0:
-                weights = self.get_weights(torch.arange(self.trainable_ids.shape[0]))
-                self.weight.data += weights
-            self.merged = True
+        self.lora_A.eval()
+        self.lora_B.eval()
 
     def forward(self, input_ids: torch.LongTensor):
         result = nn.Embedding.forward(self, input_ids)
-
-        if self.r > 0 and not self.merged:
-            weights = self.get_weights(input_ids)
-            result += weights
-
+        result += self.get_weights(input_ids)
         return result