move prompt templates out of python bindings

cmd/cmd.go

@@ -59,7 +59,7 @@ func NewCLI() *cobra.Command {
 
 	rootCmd := &cobra.Command{
 		Use:   "ollama",
-		Short: "Run any large language model on any machine.",
+		Short: "Large language model runner",
 		CompletionOptions: cobra.CompletionOptions{
 			DisableDefaultCmd: true,
 		},
@@ -88,14 +88,13 @@ func NewCLI() *cobra.Command {
 		},
 	}
 
-/*
 	runCmd := &cobra.Command{
 		Use: "run",
 		Short: "Run a model and submit prompts.",
-		RunE: func(cmd *cobra.Command. args []string) error {
+		RunE: func(cmd *cobra.Command,args []string) error {
+			return nil
 		},
 	}
-*/
 
 	serveCmd := &cobra.Command{
 		Use:     "serve",
@@ -132,6 +131,7 @@ func NewCLI() *cobra.Command {
 	rootCmd.AddCommand(
 		modelsCmd,
 		serveCmd,
+		runCmd,
 	)
 
 	return rootCmd

llama/CMakeLists.txt

@@ -23,5 +23,6 @@ target_include_directories(binding PRIVATE ${llama_cpp_SOURCE_DIR})
 target_include_directories(binding PRIVATE ${llama_cpp_SOURCE_DIR}/examples)
 target_link_libraries(binding llama ggml_static)
 
+configure_file(${llama_cpp_SOURCE_DIR}/ggml-metal.metal ${CMAKE_CURRENT_BINARY_DIR}/ggml-metal.metal COPYONLY)
 configure_file(${llama_cpp_BINARY_DIR}/libllama.a ${CMAKE_CURRENT_BINARY_DIR}/libllama.a COPYONLY)
 configure_file(${llama_cpp_BINARY_DIR}/libggml_static.a ${CMAKE_CURRENT_BINARY_DIR}/libggml_static.a COPYONLY)

server/routes.go

@@ -36,6 +36,8 @@ func Serve(ln net.Listener) error {
 	})
 
 	r.POST("/api/generate", func(c *gin.Context) {
+		// TODO: set prompt from template
+
 		var req api.GenerateRequest
 		if err := c.ShouldBindJSON(&req); err != nil {
 			c.JSON(http.StatusBadRequest, gin.H{"message": err.Error()})
@@ -64,13 +66,10 @@ func Serve(ln net.Listener) error {
 			return true
 		})
 
-/*
+		// embeds, err := l.Embeddings(text)
-                embeds, err := l.Embeddings(text)
+		// if err != nil {
-                if err != nil {
+		//         fmt.Printf("Embeddings: error %s \n", err.Error())
-                        fmt.Printf("Embeddings: error %s \n", err.Error())
+		// }
-                }
-*/
-		
 	})
 
 	log.Printf("Listening on %s", ln.Addr())