update scaling_alpha

vllm/config/model.py

@@ -1871,8 +1871,8 @@ def _get_and_verify_max_len(
                     "original_max_position_embeddings"]
                 
             # see DynamicNTKAlphaRotaryEmbedding
-            if rope_scaling["type"] == "dynamic" and "alpha" in rope_scaling:
-                scaling_factor = 1
+            # if "alpha" in rope_scaling and rope_scaling["type"] == "dynamic":
+            #     scaling_factor = 1
                 
             derived_max_model_len *= scaling_factor
 

vllm/model_executor/layers/rotary_embedding/__init__.py

@@ -137,10 +137,11 @@ def get_rope(
                     scaling_alpha, dtype)
             elif "factor" in rope_scaling:
                 scaling_factor = rope_scaling["factor"]
-                if "alpha" in rope_scaling:
+                scaling_alpha = rope_scaling["alpha"]
+                if scaling_alpha:
                     rotary_emb = DynamicNTKAlphaRotaryEmbedding(
                         head_size, rotary_dim, max_position, base, is_neox_style,
-                        rope_scaling["alpha"], dtype)
+                        scaling_alpha, dtype)
                 else:
                     rotary_emb = DynamicNTKScalingRotaryEmbedding(
                         head_size, rotary_dim, max_position, base, is_neox_style,