Config: Read and expose max_position_embeddings

Author: turboderp

exllamav3/architecture/gemma3.py

@@ -138,6 +138,11 @@ def __init__(
         self.vision_pp.size = read_dict(read_prep_config, dict, ["size"], no_default)
 
 
+    def default_max_position_embeddings(self):
+        # Fixed for Gemma3, usually not present in config.json
+        return 131072
+
+
 class Gemma3TextConfig(Config):
     arch_string = "Gemma3ForCausalLM"
 
@@ -218,6 +223,11 @@ def __init__(
         self.final_logit_softcapping = self.read_cfg(float, "final_logit_softcapping", 0.0)
 
 
+    def default_max_position_embeddings(self):
+        # Fixed for Gemma2, usually not present in config.json
+        return 8192
+
+
 class Gemma3Model(Model):
     config_class = Gemma3Config
 

exllamav3/model/config.py

@@ -69,11 +69,16 @@ def __init__(
         self.num_q_heads = -1
         self.num_kv_heads = -1
         self.pos_encoding_mode = "NONE"
+        self.max_position_embeddings = self.read_cfg(int, "max_position_embeddings", self.default_max_position_embeddings())
 
         # Load parameters
         self.load_isq = False
 
 
+    def default_max_position_embeddings(self):
+        return 8192
+
+
     def read_cfg(self, *args):
         """
         Read from config.json, see read()