rxnn 0.2.58__tar.gz → 0.2.60__tar.gz

{rxnn-0.2.58 → rxnn-0.2.60}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.2.58
+Version: 0.2.60
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.2.58 → rxnn-0.2.60}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.2.58"
+version = "0.2.60"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.2.58 → rxnn-0.2.60}/src/rxnn/memory/attention.py

@@ -48,7 +48,8 @@ class StmMemoryAttention(nn.Module):
             return layer_gate * new_layer_stm + (1 - layer_gate) * layer_stm
 
     def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None) -> torch.Tensor:
-        mem_mask = attention_mask.unsqueeze(1).unsqueeze(1).bool() if attention_mask else None
+        if attention_mask is not None:
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(1).bool()
         new_stm = torch.zeros_like(self.stm.memory)
         for i in range(self.num_layers):
             layer_stm = self.stm(i)
@@ -57,7 +58,7 @@ class StmMemoryAttention(nn.Module):
                 layer_stm = layer_stm.expand(x.size(0), -1, -1)
             encoded_layer_data = x[i]
             normalized_layer_stm = self.memory_norm_layers[i](layer_stm)
-            new_layer_stm = self.attention_layers[i](normalized_layer_stm, encoded_layer_data, encoded_layer_data, mask=mem_mask)
+            new_layer_stm = self.attention_layers[i](normalized_layer_stm, encoded_layer_data, encoded_layer_data, mask=attention_mask)
             if self.use_gated_residual:
                 new_stm[i] = self._residual_gate(self.gate[i], layer_stm, new_layer_stm) # gated residual
             else:

{rxnn-0.2.58 → rxnn-0.2.60}/src/rxnn/rxt/models.py

@@ -103,13 +103,13 @@ class RxTAlphaComponentBase(nn.Module, PyTorchModelHubMixin):
         if cross_att_type in ['mha', 'gqa', 'mqa']:
             cross_att_init = lambda: init_attention(embed_dim, att_heads, cross_att_type, att_groups, rope=rope,
                                                     use_flash_attention=use_flash_attention, dropout=att_dropout,
-                                                    max_seq_len=seq_len, is_causal=is_causal, rope_only_for_query=True)
+                                                    max_seq_len=seq_len, is_causal=False, rope_only_for_query=True)
         else:
             cross_att_init = lambda: init_experimental_attention(embed_dim, att_heads, cross_att_type,
                                                                  cross_att_groups or att_groups, rope=rope,
                                                                  use_flash_attention=use_flash_attention,
                                                                  dropout=att_dropout,
-                                                                 max_seq_len=seq_len, is_causal=is_causal,
+                                                                 max_seq_len=seq_len, is_causal=False,
                                                                  num_experts=att_experts,
                                                                  num_query_experts=att_query_experts,
                                                                  num_query_groups=cross_att_query_groups or att_query_groups,

{rxnn-0.2.58 → rxnn-0.2.60}/src/rxnn/transformers/models.py

@@ -213,6 +213,7 @@ class ClassicTransformerDecoder(ClassicTransformerBase):
             if attention_mask is not None:
                 mask &= attention_mask.unsqueeze(1).unsqueeze(1).bool()
         elif attention_mask is not None:
+            print(attention_mask.size())
             mask = attention_mask.unsqueeze(1).unsqueeze(1).bool()
         else:
             mask = None