Xsn/server sleep

common/arg.cpp

@@ -2887,6 +2887,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.lora_init_without_apply = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--sleep-idle-seconds"}, "SECONDS",
+        string_format("number of seconds of idleness after which the server will sleep (default: %d; -1 = disabled)", params.sleep_idle_seconds),
+        [](common_params & params, int value) {
+            if (value == 0 || value < -1) {
+                throw std::invalid_argument("invalid value: cannot be 0 or less than -1");
+            }
+            params.sleep_idle_seconds = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--simple-io"},
         "use basic IO for better compatibility in subprocesses and limited consoles",

common/common.h

@@ -475,7 +475,8 @@ struct common_params {
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
-    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
+    bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
+    int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
     std::vector<std::string> api_keys;
 

tools/cli/cli.cpp

@@ -209,8 +209,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    ctx_cli.ctx_server.init();
-
     console::spinner::stop();
     console::log("\n");
 

tools/server/README-dev.md

@@ -107,6 +107,8 @@ For detailed instructions, see the [test documentation](./tests/README.md).
 - Large-scale code base split into smaller files: https://github.com/ggml-org/llama.cpp/pull/17362
 - Introduction of router mode: https://github.com/ggml-org/llama.cpp/pull/17470
 - Speculative decoding: https://github.com/ggml-org/llama.cpp/pull/17808 and rework in https://github.com/ggml-org/llama.cpp/pull/17808
+- INI presets: https://github.com/ggml-org/llama.cpp/pull/17859 (+ refactoring: https://github.com/ggml-org/llama.cpp/pull/18169)
+- Sleeping mode: https://github.com/ggml-org/llama.cpp/pull/18228
 
 
 

tools/server/README.md

@@ -1621,6 +1621,16 @@ Example of an error:
 }
 ```
 
+## Sleeping on Idle
+
+The server supports an automatic sleep mode that activates after a specified period of inactivity (no incoming tasks). This feature, introduced in [PR #18228](https://github.com/ggml-org/llama.cpp/pull/18228), can be enabled using the `--sleep-idle-seconds` command-line argument. It works seamlessly in both single-model and multi-model configurations.
+
+When the server enters sleep mode, the model and its associated memory (including the KV cache) are unloaded from RAM to conserve resources. Any new incoming task will automatically trigger the model to reload.
+
+Note that the following endpoints are exempt from being considered as incoming tasks. They do not trigger model reloading and do not reset the idle timer:
+- `GET /health`
+- `GET /props`
+
 ## More examples
 
 ### Interactive mode