Upload modelforseminat_v5.py with huggingface_hub

modelforseminat_v5.py

@@ -211,7 +211,14 @@ class TwoLayerMLP(nn.Module):
 
 
 
-
+class Olmo2ConfigForSemiNAT(Olmo2Config):
+    def __init__(self, chunk_size_limit: int = 5, decoder_layers: int = 1, encoder_layer: int = 1, mlp: bool = False, position_embedding_type: str = "absolute", **kwargs):
+        super().__init__(**kwargs)
+        self.chunk_size_limit = chunk_size_limit
+        self.decoder_layers = decoder_layers
+        self.encoder_layer = encoder_layer
+        self.mlp = mlp
+        self.position_embedding_type = position_embedding_type
 
 
 
@@ -224,7 +231,7 @@ class TwoLayerMLP(nn.Module):
 class Olmo2AttentionForSemiNAT(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
-    def __init__(self, config: Olmo2Config, layer_idx: Optional[int] = None, is_causal: bool = True):
+    def __init__(self, config: Olmo2ConfigForSemiNAT, layer_idx: Optional[int] = None, is_causal: bool = True):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
@@ -330,7 +337,7 @@ class Olmo2DecoderLayerForSemiNAT(nn.Module):
 
     def __init__(
         self,
-        config: Olmo2Config,
+        config: Olmo2ConfigForSemiNAT,
         layer_idx: int,
         is_causal: bool = True,
     ):
@@ -395,7 +402,7 @@ class Olmo2DecoderLayerForSemiNAT(nn.Module):
 
 class NATEncoderForSemiNAT(nn.Module):
 
-    def __init__(self, config: Olmo2Config, num_layer: int = 1):
+    def __init__(self, config: Olmo2ConfigForSemiNAT, num_layer: int = 1):
         super().__init__()
         self.num_layer = num_layer
         self.encoder_layers = nn.ModuleList([
@@ -431,7 +438,7 @@ class NATEncoderForSemiNAT(nn.Module):
 
 class NATDecoderForSemiNAT(nn.Module):
 
-    def __init__(self, config: Olmo2Config, num_layer: int = 1):
+    def __init__(self, config: Olmo2ConfigForSemiNAT, num_layer: int = 1):
         super().__init__()
         self.num_layer = num_layer
         self.decoder_layers = nn.ModuleList([
@@ -471,8 +478,8 @@ class Olmo2ModelForSemiNAT(Olmo2Model):
             for layer_idx in range(config.num_hidden_layers)
         ])
 
-        self.decoder = NATDecoderForSemiNAT(config, 1)
-        self.encoder = NATEncoderForSemiNAT(config, 1)
+        self.decoder = NATDecoderForSemiNAT(config, config.decoder_layers)
+        self.encoder = NATEncoderForSemiNAT(config, config.encoder_layer)
 
 
         # pdb.set_trace()
@@ -487,8 +494,10 @@ class Olmo2ModelForSemiNAT(Olmo2Model):
 
         self.length_predictor = nn.Linear(config.hidden_size,
                                           self.chunk_size_limit)
-
-        # self.linear_projection = TwoLayerMLP(config.hidden_size)
+        self.mlp = config.mlp
+        if self.mlp:
+            self.linear_projection = TwoLayerMLP(config.hidden_size)
+        self.position_embedding_type = config.position_embedding_type
 
 
     def forward(
@@ -780,13 +789,10 @@ class Olmo2ModelForSemiNAT(Olmo2Model):
                 hidden_states, length_ground_truth, self.chunk_size_limit, skip_val=-100)
 
 
-
-
-
-
-
-
-
+            if self.mlp:
+                nat_input_embeddings = self.linear_projection(nat_input_embeddings)
+                
+            
             # pdb.set_trace()
             # for b in range(bs):
             #     for i in range(slice_num[b]):
@@ -812,22 +818,8 @@ class Olmo2ModelForSemiNAT(Olmo2Model):
 
 
             # pdb.set_trace()
-            # nar_chunk_position = torch.arange(
-            #     0, self.chunk_size_limit).unsqueeze(0).repeat(
-            #         accumu_num,
-            #         1).to(hidden_states.device)  # bs * max_chunk_num
-
-            # nar_position_embeddings = self.rotary_emb(nat_attention_mask,
-            #                                           nar_chunk_position)
-
-
-            # pdb.set_trace()
-
-            nat_input_embeddings = self.pos_encoder(nat_input_embeddings) # 加上绝对位置编码
 
 
-            self.decoder = self.decoder.to(dtype=nat_input_embeddings.dtype)
-
 
             # 处理attention
             mask_nat_attention_mask = self.nat_prepare_4d_full_attention_mask_without_causal(
@@ -835,19 +827,32 @@ class Olmo2ModelForSemiNAT(Olmo2Model):
                 dtype=nat_attention_mask.dtype,
                 device=nat_attention_mask.device)
 
+            # pdb.set_trace()
+
+            self.decoder = self.decoder.to(dtype=nat_input_embeddings.dtype)
+            if self.position_embedding_type == "relative":
+                nar_chunk_position = torch.arange(
+                0, self.chunk_size_limit).unsqueeze(0).repeat(
+                    accumu_num,
+                    1).to(hidden_states.device) 
+
+                pos = self.rotary_emb(nat_attention_mask, nar_chunk_position)
+
+            elif self.position_embedding_type == "absolute":
+                nat_input_embeddings = self.pos_encoder(nat_input_embeddings) # 加上绝对位置编码
+                pos = None  
 
             # pdb.set_trace()
             nar_hidden_states = self.decoder(
                 nat_input_embeddings,
                 attention_mask=mask_nat_attention_mask,
                 # attention_mask=None,
-                # position_embeddings=nar_position_embeddings,
-                position_embeddings=None, #使用绝对位置，不传相对位置
+                position_embeddings=pos,
+                # position_embeddings=None, #使用绝对位置，不传相对位置
                 output_attentions=output_attentions,
                 use_cache=use_cache,
                 cache_position=None,
             )
-
             nar_hidden_states = self.norm(
                 nar_hidden_states)  # bs * max_chunk_num * hidden_size
 
@@ -1895,7 +1900,7 @@ class Olmo2ForCausalLMForSemiNAT(Olmo2ForCausalLM):
         model_kwargs[
             'cache_position'] = new_cache_position  # 更新一下cache position
 
-        # pdb.set_trace()
+        pdb.set_trace()
         ############ prefilling ############
 
         is_prefill = True