llama : fix Gemma rope type (#5691)

diff --git a/llama.cpp b/llama.cpp
index 3424b1999ebdda7fd2a4d4a4e14c99c59f7a1bb7..28430254f698f18c14e069c13cadd048653ef897 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -12089,7 +12089,6 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_ORION:
         case LLM_ARCH_INTERNLM2:
         case LLM_ARCH_MINICPM:
-        case LLM_ARCH_GEMMA:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -12101,6 +12100,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_PHI2:
+        case LLM_ARCH_GEMMA:
             return LLAMA_ROPE_TYPE_NEOX;
 
         // all model arches should be listed explicitly here