Load persistent RoPE buffers in lm-eval

modeling_tiger_alpha.py

@@ -49,8 +49,13 @@ class TigerAlphaRotaryEmbedding(nn.Module):
         positions = torch.arange(config.max_position_embeddings).float()
         frequencies = torch.outer(positions, inverse)
         angles = torch.cat((frequencies, frequencies), dim=-1)
-        self.register_buffer("cosine", angles.cos(), persistent=False)
-        self.register_buffer("sine", angles.sin(), persistent=False)
+        # These deterministic caches must be checkpointed. Hugging Face can
+        # instantiate remote-code models under a low-memory/meta-device
+        # context; non-persistent RoPE buffers can then remain invalid after
+        # loading and make every continuation log-probability NaN. This is the
+        # same export/load issue fixed in the original partial-Edu model.
+        self.register_buffer("cosine", angles.cos(), persistent=True)
+        self.register_buffer("sine", angles.sin(), persistent=True)
 
     def forward(
         self,