Use specified max context.py

Instead of original_max_position_embeddings.

This appears to be what transformers intended, and does not update dynamically with sequence leng there.

Author: Downtown-Case

exllamav2/device.py

@@ -187,7 +187,7 @@ def prepare_sincos(self):
 
         elif cfg.alt_rope_method == "yarn":
 
-            yarn_max_position_embeddings = cfg.yarn_rope_original_max_position_embeddings
+            yarn_max_position_embeddings = cfg.max_seq_len
 
             # Only activate if longer than original ctx
             if cfg.max_seq_len > cfg.yarn_rope_original_max_position_embeddings: