Upload sentiment-transformer model

modeling_sentiment_transformer.py

@@ -352,11 +352,17 @@ class SentimentTransformerForSequenceClassification(PreTrainedModel):
                 freqs = torch.outer(t, inv_freq)
                 module.cos_cached = freqs.cos()
                 module.sin_cached = freqs.sin()
-
-    def post_init(self) -> None:
-        """Override HF's post_init to also recompute RoPE buffers."""
-        super().post_init()
-        self._recompute_rope_buffers()
+        self._rope_valid = True
+
+    def _ensure_rope_valid(self) -> None:
+        """Lazily recompute RoPE buffers if they were corrupted by HF loading."""
+        if not getattr(self, "_rope_valid", False):
+            # Check if the backbone's RoPE buffers contain valid data
+            rope = self.backbone.rope
+            if not rope.cos_cached.isfinite().all():
+                self._recompute_rope_buffers()
+            else:
+                self._rope_valid = True
 
     def forward(
         self,
@@ -367,6 +373,8 @@ class SentimentTransformerForSequenceClassification(PreTrainedModel):
         **_kwargs,
     ) -> SequenceClassifierOutput | tuple[torch.Tensor, ...]:
         """Run sequence classification and return HF-style outputs."""
+        self._ensure_rope_valid()
+
         if input_ids is None:
             raise ValueError("`input_ids` is required.")
         if attention_mask is None: