feat(llama.cpp): bump and adapt to upstream changes

Makefile

@@ -8,7 +8,7 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=274ec65af6e54039eb95cb44904af5c945dca1fa
+CPPLLAMA_VERSION?=c27ac678dd393af0da9b8acf10266e760c8a0912
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp

backend/cpp/llama/grpc-server.cpp

@@ -2228,6 +2228,35 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     }
 // }
 
+const std::vector<ggml_type> kv_cache_types = {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_BF16,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_IQ4_NL,
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+};
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    for (const auto & type : kv_cache_types) {
+        if (ggml_type_name(type) == s) {
+            return type;
+        }
+    }
+    throw std::runtime_error("Unsupported cache type: " + s);
+}
+
+static std::string get_all_kv_cache_types() {
+    std::ostringstream msg;
+    for (const auto & type : kv_cache_types) {
+        msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 static void params_parse(const backend::ModelOptions* request,
                                 common_params & params) {
 
@@ -2242,10 +2271,10 @@ static void params_parse(const backend::ModelOptions* request,
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
     if (!request->cachetypekey().empty()) {
-        params.cache_type_k = request->cachetypekey();
+        params.cache_type_k = kv_cache_type_from_str(request->cachetypekey());
     }
     if (!request->cachetypevalue().empty()) {
-        params.cache_type_v = request->cachetypevalue();
+        params.cache_type_v = kv_cache_type_from_str(request->cachetypevalue());
     }
     params.n_ctx = request->contextsize();
     //params.memory_f16 = request->f16memory();