Allow component models to use learned pos embeddings without regarding LLM max_seq_len

exllamav2/pos_embedding.py

@@ -20,11 +20,13 @@ class ExLlamaV2PosEmbedding(ExLlamaV2Module):
     def __init__(
         self,
         model: ExLlamaV2,
-        key: str
+        key: str,
+        override_max_seq_len: bool = False
     ):
         super().__init__(model, key)
 
         self.native_ctx_size = model.config.max_seq_len
+        self.override_max_seq_len = override_max_seq_len
         self.embedding = None
 
 
@@ -34,8 +36,9 @@ class ExLlamaV2PosEmbedding(ExLlamaV2Module):
         w = self.load_weight()
         assert isinstance(w, nn.Parameter)
         self.native_ctx_size = w.shape[0]
-        assert self.model.config.max_seq_len <= self.native_ctx_size, \
-            f"Learned positional embeddings cannot be extended past native size of {self.native_ctx_size}."
+        if not self.override_max_seq_len:
+            assert self.model.config.max_seq_len <= self.native_ctx_size, \
+                f"Learned positional embeddings cannot be extended past native size of {self.native_ctx_size}."
 
         self.embedding = nn.Embedding(self.native_ctx_size, self.model.config.hidden_size, device = "meta")
         self.embedding.weight = w