diff --git a/cmd/cmd.go b/cmd/cmd.go
index c8e5b4c45..c3c854fa2 100644
--- a/cmd/cmd.go
+++ b/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
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index b2ef7b398..11af62d7d 100644
--- a/llama/CMakeLists.txt
+++ b/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)
diff --git a/server/routes.go b/server/routes.go
index 6cea2eb81..4912d0c6c 100644
--- a/server/routes.go
+++ b/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)
-                if err != nil {
-                        fmt.Printf("Embeddings: error %s \n", err.Error())
-                }
-*/
-		
+		// embeds, err := l.Embeddings(text)
+		// if err != nil {
+		//         fmt.Printf("Embeddings: error %s \n", err.Error())
+		// }
 	})
 
 	log.Printf("Listening on %s", ln.Addr())
diff --git a/python/ollama/templates/alpaca.prompt b/templates/alpaca.prompt
similarity index 100%
rename from python/ollama/templates/alpaca.prompt
rename to templates/alpaca.prompt
diff --git a/python/ollama/templates/falcon.prompt b/templates/falcon.prompt
similarity index 100%
rename from python/ollama/templates/falcon.prompt
rename to templates/falcon.prompt
diff --git a/python/ollama/templates/gpt4.prompt b/templates/gpt4.prompt
similarity index 100%
rename from python/ollama/templates/gpt4.prompt
rename to templates/gpt4.prompt
diff --git a/python/ollama/templates/hermes.prompt b/templates/hermes.prompt
similarity index 100%
rename from python/ollama/templates/hermes.prompt
rename to templates/hermes.prompt
diff --git a/python/ollama/templates/mpt.prompt b/templates/mpt.prompt
similarity index 100%
rename from python/ollama/templates/mpt.prompt
rename to templates/mpt.prompt
diff --git a/python/ollama/templates/oasst.prompt b/templates/oasst.prompt
similarity index 100%
rename from python/ollama/templates/oasst.prompt
rename to templates/oasst.prompt
diff --git a/python/ollama/templates/orca.prompt b/templates/orca.prompt
similarity index 100%
rename from python/ollama/templates/orca.prompt
rename to templates/orca.prompt
diff --git a/python/ollama/templates/qlora.prompt b/templates/qlora.prompt
similarity index 100%
rename from python/ollama/templates/qlora.prompt
rename to templates/qlora.prompt
diff --git a/python/ollama/templates/tulu.prompt b/templates/tulu.prompt
similarity index 100%
rename from python/ollama/templates/tulu.prompt
rename to templates/tulu.prompt
diff --git a/python/ollama/templates/ultralm.prompt b/templates/ultralm.prompt
similarity index 100%
rename from python/ollama/templates/ultralm.prompt
rename to templates/ultralm.prompt
diff --git a/python/ollama/templates/vicuna.prompt b/templates/vicuna.prompt
similarity index 100%
rename from python/ollama/templates/vicuna.prompt
rename to templates/vicuna.prompt
diff --git a/python/ollama/templates/wizardcoder.prompt b/templates/wizardcoder.prompt
similarity index 100%
rename from python/ollama/templates/wizardcoder.prompt
rename to templates/wizardcoder.prompt
diff --git a/python/ollama/templates/wizardlm.prompt b/templates/wizardlm.prompt
similarity index 100%
rename from python/ollama/templates/wizardlm.prompt
rename to templates/wizardlm.prompt