From: Georgi Gerganov <redacted>
Date: Mon, 26 Feb 2024 06:30:17 +0000 (+0200)
Subject: llama : fix Gemma rope type (#5691)
X-Git-Tag: upstream/0.0.4488~2219
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=269de86ba073b5dc9ce687c11a3bc4d7d873b962;p=pkg%2Fggml%2Fsources%2Fllama.cpp

llama : fix Gemma rope type (#5691)
---

diff --git a/llama.cpp b/llama.cpp
index 3424b199..28430254 100644
--- 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