sinlib 0.3.1__tar.gz → 0.3.2__tar.gz

{sinlib-0.3.1 → sinlib-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: sinlib
-Version: 0.3.1
+Version: 0.3.2
 Summary: Sinhala NLP Toolkit
 Project-URL: Code, https://github.com/Ransaka/sinlib
 Project-URL: Docs, https://sinlib.readthedocs.io

{sinlib-0.3.1 → sinlib-0.3.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sinlib"
-version = "0.3.1"
+version = "0.3.2"
 description = "Sinhala NLP Toolkit"
 authors = [
     { name = "Ransaka", email = "ransaka.ravihara@gmail.com" }

{sinlib-0.3.1 → sinlib-0.3.2}/src/sinlib/__init__.py

@@ -62,4 +62,4 @@ __all__: List[str] = [
     "setup_matplotlib",
 ]
 
-__version__ = "0.3.1"
+__version__ = "0.3.2"

{sinlib-0.3.1 → sinlib-0.3.2}/src/sinlib/charbert/encoder.py

@@ -120,6 +120,25 @@ class SinhalaCharBERTEncoder(nn.Module):
         all_char_hidden_states: List[torch.Tensor] = []
         all_attentions: List[torch.Tensor] = []
 
+        # Normalize the attention mask to the 4-D extended additive format
+        # (batch, 1, 1, seq_len) expected by transformers' attention
+        # implementations. Newer versions of transformers (>= 4.56) index the
+        # mask as attention_mask[:, :, :, :key_len] and crash with
+        # "too many indices for tensor of dimension 2" if a 2-D padding mask
+        # is passed straight through.
+        if attention_mask is not None and attention_mask.dim() != 4:
+            if attention_mask.dim() == 2:
+                mask = attention_mask[:, None, None, :]
+            elif attention_mask.dim() == 3:
+                mask = attention_mask[:, None, :, :]
+            else:
+                raise ValueError(
+                    f"Unsupported attention_mask dims: {attention_mask.dim()}"
+                )
+            # Convert 1/0 padding mask to an additive mask (0 keep, -inf drop)
+            dtype = token_hidden_states.dtype
+            attention_mask = (1.0 - mask.to(dtype)) * torch.finfo(dtype).min
+
         curr_token = token_hidden_states
         curr_char = char_hidden_states