server: save and clear idle slots on new task (--clear-idle) (#20993)

* server: clear idle slots KV from VRAM (LLAMA_KV_KEEP_ONLY_ACTIVE)

* server: move idle slot KV clearing to slot release

The save "cost" is now paid by the finishing request.

* server: add --kv-clear-idle flag, enable by default

* server: skip clearing last idle slot, clear on launch

* server: test --no-kv-clear-idle flag

* server: simplify on-release clearing loop

* server: remove on-release KV clearing, keep launch-only

* cont : clean-up

* tests: update log strings after --clear-idle rename

* tests: use debug tags instead of log message matching

* test: fix Windows CI by dropping temp log file unlink

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

Author: yychyo

common/arg.cpp

@@ -1311,6 +1311,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.kv_unified = value;
         }
     ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
+    add_opt(common_arg(
+        {"--clear-idle"},
+        {"--no-clear-idle"},
+        "save and clear idle slots on new task (default: enabled, requires unified KV and cache-ram)",
+        [](common_params & params, bool value) {
+            params.clear_idle = value;
+        }
+    ).set_env("LLAMA_ARG_CLEAR_IDLE").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},

common/common.h

@@ -579,8 +579,9 @@ struct common_params {
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
     bool    cache_prompt        = true;  // whether to enable prompt caching
-    int32_t n_ctx_checkpoints   = 32;     // max number of context checkpoints per slot
-    int32_t checkpoint_every_nt = 8192;   // make a checkpoint every n tokens during prefill
+    bool    clear_idle          = true;  // save and clear idle slots upon starting a new task
+    int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
+    int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";

tools/cli/README.md

@@ -176,8 +176,8 @@
 | `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |
 | `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
 | `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--skip-chat-parsing, --no-skip-chat-parsing` | force a pure content parser, even if a Jinja template is specified; model will output everything in the content section, including any reasoning and/or tool calls (default: disabled)<br/>(env: LLAMA_ARG_SKIP_CHAT_PARSING) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |

tools/completion/README.md

@@ -255,8 +255,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |
 | `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
 | `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--skip-chat-parsing, --no-skip-chat-parsing` | force a pure content parser, even if a Jinja template is specified; model will output everything in the content section, including any reasoning and/or tool calls (default: disabled)<br/>(env: LLAMA_ARG_SKIP_CHAT_PARSING) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 

tools/server/README.md

@@ -167,6 +167,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-cpent, --checkpoint-every-n-tokens N` | create a checkpoint every n tokens during prefill (processing), -1 to disable (default: 8192)<br/>(env: LLAMA_ARG_CHECKPOINT_EVERY_NT) |
 | `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
 | `-kvu, --kv-unified, -no-kvu, --no-kv-unified` | use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)<br/>(env: LLAMA_ARG_KV_UNIFIED) |
+| `--clear-idle, --no-clear-idle` | save and clear idle slots on new task (default: enabled, requires unified KV and cache-ram)<br/>(env: LLAMA_ARG_CLEAR_IDLE) |
 | `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
 | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode |
 | `-sp, --special` | special tokens output enabled (default: false) |
@@ -221,8 +222,8 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |
 | `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
 | `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--skip-chat-parsing, --no-skip-chat-parsing` | force a pure content parser, even if a Jinja template is specified; model will output everything in the content section, including any reasoning and/or tool calls (default: disabled)<br/>(env: LLAMA_ARG_SKIP_CHAT_PARSING) |
 | `--prefill-assistant, --no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_PREFILL_ASSISTANT) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |

tools/server/server-context.cpp

@@ -605,6 +605,17 @@ struct server_context_impl {
         llama_batch_free(batch);
     }
 
+    void slot_save_and_clear(server_slot & slot) {
+        if (slot.prompt.n_tokens() == 0) {
+            return;
+        }
+        SLT_INF(slot, "%s", "saving idle slot to prompt cache\n");
+        SLT_DBG(slot, "%s", "__TEST_TAG_CLEAR_IDLE_SLOT__\n");
+        slot.prompt_save(*prompt_cache);
+        slot.prompt_clear(false);
+        prompt_cache->update();
+    }
+
     void handle_sleeping_state(bool new_state) {
         GGML_ASSERT(sleeping != new_state);
         if (new_state) {
@@ -864,6 +875,19 @@ struct server_context_impl {
 
         metrics.init();
 
+        if (params_base.clear_idle) {
+            if (!params_base.kv_unified) {
+                SRV_WRN("%s: --clear-idle requires --kv-unified, disabling\n", __func__);
+                params_base.clear_idle = false;
+            } else if (params_base.cache_ram_mib == 0) {
+                SRV_WRN("%s: --clear-idle requires --cache-ram, disabling\n", __func__);
+                params_base.clear_idle = false;
+            } else {
+                SRV_INF("%s: idle slots will be saved to prompt cache and cleared upon starting a new task\n", __func__);
+                SRV_DBG("%s", "__TEST_TAG_CLEAR_IDLE_ENABLED__\n");
+            }
+        }
+
         // populate webui settings
         {
             if (!params_base.webui_config_json.empty()) {
@@ -1010,15 +1034,15 @@ struct server_context_impl {
             // cache prompts only for completion tasks
             update_cache = update_cache && task.type == SERVER_TASK_TYPE_COMPLETION;
 
-            // don't update the cache if the slot's context is empty
-            update_cache = update_cache && tokens.size() > 0;
-
             if (update_cache) {
                 SRV_WRN("%s", "updating prompt cache\n");
 
                 const int64_t t_start = ggml_time_us();
 
-                ret->prompt_save(*prompt_cache);
+                // don't save the slot's state if its context is empty
+                if (tokens.size() > 0) {
+                    ret->prompt_save(*prompt_cache);
+                }
 
                 if (!ret->prompt_load(*prompt_cache, task.tokens)) {
                     ret->prompt_clear(false);
@@ -1692,9 +1716,7 @@ struct server_context_impl {
                     const int id_slot = task.id_slot;
                     const int id_task = task.id;
 
-                    server_slot * slot = id_slot != -1
-                                            ? get_slot_by_id(id_slot)
-                                            : get_available_slot(task);
+                    server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task);
 
                     //
                     // slot scheduling logic
@@ -1731,6 +1753,14 @@ struct server_context_impl {
                         SRV_ERR("failed to launch slot with task, id_task = %d\n", id_task);
                         break; // drop the task
                     }
+
+                    if (params_base.clear_idle) {
+                        for (auto & s : slots) {
+                            if (!s.is_processing()) {
+                                slot_save_and_clear(s);
+                            }
+                        }
+                    }
                 } break;
             case SERVER_TASK_TYPE_CANCEL:
                 {