disable context shifting for gemma3

Author: LostRuins

gpttype_adapter.cpp

@@ -1792,6 +1792,7 @@ void PurgeMissingTokens(llama_context * ctx, llama_context * draft_ctx, std::vec
 
     auto shared = LongestCommonSubseq(curr_ctx_without_memory, new_ctx_without_memory);
 
+    printf("\nSharedSize: %d, LCSTokThreshold: %d, ArrPass: %d\n",shared.size(),LCSTokThreshold,ArrStartWith(new_ctx_without_memory, shared));
     if (shared.size() > LCSTokThreshold && ArrStartWith(new_ctx_without_memory, shared)) // enough tokens in common
     {
         int found = ArrFindIndexOf(current_context_tokens,shared);
@@ -2128,6 +2129,11 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
             printf("Qwen2VL detected! Mrope will be used, and context shift will be disabled!\n");
             kcpp_data->use_contextshift = false;
         }
+        if(file_format_meta.model_architecture == GGUFArch::ARCH_GEMMA3)
+        {
+            printf("Gemma3 detected! Context shift will be disabled!\n");
+            kcpp_data->use_contextshift = false;
+        }
         model_params.main_gpu = cu_parseinfo_maindevice;
 
         #if defined(GGML_USE_CUDA)

model_adapter.cpp

@@ -321,6 +321,10 @@ void print_tok_vec(std::vector<float> &embd)
             {
                 fileformatmeta->model_architecture = GGUFArch::ARCH_QWEN2VL;
             }
+            else if(modelarch=="gemma3")
+            {
+                fileformatmeta->model_architecture = GGUFArch::ARCH_GEMMA3;
+            }
             else if(modelarch=="rwkv6")
             {
                 fileformatmeta->model_architecture = GGUFArch::ARCH_RWKV;

model_adapter.h

@@ -60,6 +60,7 @@ enum GGUFArch
     ARCH_QWEN2 = 5,
     ARCH_RWKV = 6,
     ARCH_QWEN2VL = 7,
+    ARCH_GEMMA3 = 8,
 };
 
 struct FileFormatExtraMeta