chop down position embedding matrix by 2 (1+padding_idx)

Author: iamlemec

convert_hf_to_gguf.py

@@ -2429,6 +2429,17 @@ def set_vocab(self):
         self.gguf_writer.add_add_bos_token(True) # looks like we need this
         self.gguf_writer.add_add_eos_token(True)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at 2 (padding_idx = 1), so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            del bid # unused
+
+            data_torch[:-2,:] = data_torch[2:,:].clone()
+            data_torch[-2:,:] = 0.0
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return super().modify_tensors(data_torch, name, bid)
 
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):