fix issues

llama.cpp

@@ -5164,12 +5164,11 @@ static int llama_decode_internal(
 
     // If all tensors can be run on the GPU then using more than 1 thread is detrimental.
     const bool full_offload_supported =
-        model.arch == LLM_ARCH_LLAMA      ||
-        model.arch == LLM_ARCH_BAICHUAN   ||
-        model.arch == LLM_ARCH_FALCON     ||
-        model.arch == LLM_ARCH_REFACT     ||
-        model.arch == LLM_ARCH_MPT        ||
-        model.arch == LLM_ARCH_STARCODER;
+        model.arch == LLM_ARCH_LLAMA    ||
+        model.arch == LLM_ARCH_BAICHUAN ||
+        model.arch == LLM_ARCH_FALCON   ||
+        model.arch == LLM_ARCH_REFACT   ||
+        model.arch == LLM_ARCH_MPT;
 
     const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
     if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {