add mask_first_token

README.md

@@ -50,13 +50,17 @@ You can change various parameters like :
 Default parameters work well in practice. If you are short on memory, reduce block sizes, increase sparsity factor and remove dropout in the attention score matrix.
 
 ```python:
+from transformers import AutoModel
+
 model = AutoModel.from_pretrained("ccdv/lsg-distilbert-base-uncased-4096", 
     trust_remote_code=True, 
     num_global_tokens=16,
     block_size=64,
     sparse_block_size=64,
-    sparsity_factor=4,
     attention_probs_dropout_prob=0.0
+    sparsity_factor=4,
+    sparsity_type="none",
+    mask_first_token=True
 )
 ``` 
 

config.json

@@ -22,6 +22,7 @@
   "hidden_dim": 3072,
   "initializer_range": 0.02,
   "lsh_num_pre_rounds": 1,
+  "mask_first_token": false,
   "max_position_embeddings": 4096,
   "model_type": "distilbert",
   "n_heads": 12,

modeling_lsg_distilbert.py

@@ -27,6 +27,7 @@ class LSGDistilBertConfig(DistilBertConfig):
         base_model_prefix="lsg",
         block_size=128,
         lsh_num_pre_rounds=1,
+        mask_first_token=False,
         num_global_tokens=1,
         pool_with_global=True,
         sparse_block_size=128,
@@ -42,6 +43,7 @@ class LSGDistilBertConfig(DistilBertConfig):
         self.base_model_prefix = base_model_prefix
         self.block_size = block_size
         self.lsh_num_pre_rounds = lsh_num_pre_rounds
+        self.mask_first_token = mask_first_token
         self.num_global_tokens = num_global_tokens
         self.pool_with_global = pool_with_global
         self.sparse_block_size = sparse_block_size
@@ -925,6 +927,7 @@ class LSGDistilBertModel(LSGDistilBertPreTrainedModel, DistilBertModel):
         assert hasattr(config, "block_size") and hasattr(config, "adaptive")
         self.block_size = config.block_size
         self.adaptive = config.adaptive
+        self.mask_first_token = config.first_token
         self.pool_with_global = config.pool_with_global
 
         # Initialize weights and apply final processing
@@ -946,6 +949,8 @@ class LSGDistilBertModel(LSGDistilBertPreTrainedModel, DistilBertModel):
 
         if attention_mask is None:
             attention_mask = torch.ones(n, t, device=inputs_.device)
+        if self.mask_first_token:
+            attention_mask[:,0] = 0
             
         b = self.block_size * 2
         pad = t % self.block_size