server : add arg for disabling prompt caching (#18776 )

* server : add arg for disabling prompt caching Disabling prompt caching is useful for clients who are restricted to sending only OpenAI-compat requests and want deterministic responses. * address review comments * address review comments
ci : use openssl for openEuler-latest-cmake-cann (#18779 )
2026-07-01 01:57:43 +02:00 · 2026-01-12 19:21:34 +02:00 · 2026-01-12 17:29:00 +01:00 · 2026-01-12 15:58:52 +01:00 · 2026-01-12 13:47:58 +01:00 · 2026-01-12 13:32:55 +01:00
12 changed files with 95 additions and 72 deletions
@@ -1463,12 +1463,14 @@ jobs:
            "${{ steps.cann-image.outputs.image }}" \
            bash -lc '
              set -e
-              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake libcurl-devel
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
              yum clean all && rm -rf /var/cache/yum
              git config --global --add safe.directory "/workspace"
              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
              cmake -S . -B build \
                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DLLAMA_CURL=OFF \
+                  -DLLAMA_OPENSSL=ON \
                  -DGGML_CANN=on \
                  -DSOC_TYPE=${SOC_TYPE}
              cmake --build build -j $(nproc)
@@ -1295,7 +1295,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.kv_unified = true;
        }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED}));
    add_opt(common_arg(
        {"--context-shift"},
        {"--no-context-shift"},
@@ -2877,10 +2877,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.n_threads_http = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
+    add_opt(common_arg(
+        {"--cache-prompt"},
+        {"--no-cache-prompt"},
+        string_format("whether to enable prompt caching (default: %s)", params.cache_prompt ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.cache_prompt = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CACHE_PROMPT"));
    add_opt(common_arg(
        {"--cache-reuse"}, "N",
        string_format(
-            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: %d)\n"
            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
        ),
        [](common_params & params, int value) {
@@ -80,6 +80,7 @@ int32_t cpu_get_num_math();
 //

 enum llama_example {
+    LLAMA_EXAMPLE_BATCHED,
    LLAMA_EXAMPLE_DEBUG,
    LLAMA_EXAMPLE_COMMON,
    LLAMA_EXAMPLE_SPECULATIVE,
@@ -475,6 +476,7 @@ struct common_params {
    int32_t timeout_write     = timeout_read; // http write timeout in seconds
    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 = 8;            // max number of context checkpoints per slot
    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.

@@ -21,7 +21,7 @@ int main(int argc, char ** argv) {
    params.prompt = "Hello my name is";
    params.n_predict = 32;

-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BATCHED, print_usage)) {
        return 1;
    }

@@ -238,9 +238,7 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
    /* .is_host          = */ NULL,
 };

-#ifdef GGML_VULKAN_MEMORY_DEBUG
 class vk_memory_logger;
-#endif
 class vk_perf_logger;
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
 static void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
@@ -825,9 +823,7 @@ struct vk_device_struct {
    bool allow_sysmem_fallback;
    bool disable_graph_optimize;

-#ifdef GGML_VULKAN_MEMORY_DEBUG
    std::unique_ptr<vk_memory_logger> memory_logger;
-#endif

    ~vk_device_struct() {
        VK_LOG_DEBUG("destroy device " << name);
@@ -1563,8 +1559,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
 static void ggml_vk_load_shaders(vk_device& device);
 static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);

-#if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
-#define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
+static bool vk_memory_logger_enabled = false;
+
+#define VK_LOG_MEMORY(msg) if (vk_memory_logger_enabled) { std::cerr << "ggml_vulkan memory: " << msg << std::endl; }

 static std::string format_size(size_t size) {
    const size_t kib = 1024;
@@ -1597,10 +1594,10 @@ private:
    std::map<vk::Buffer, size_t> allocations; // Track allocations
    size_t total_device;
    size_t total_host;
+    static std::mutex log_mutex;
 };
-#else
-#define VK_LOG_MEMORY(msg) ((void) 0)
-#endif // GGML_VULKAN_MEMORY_DEBUG
+
+std::mutex vk_memory_logger::log_mutex;

 static bool vk_perf_logger_enabled = false;
 static bool vk_perf_logger_concurrent = false;
@@ -1907,10 +1904,10 @@ struct ggml_backend_vk_buffer_context {
    }
 };

-#ifdef GGML_VULKAN_MEMORY_DEBUG
-static std::mutex log_mutex;
-
 void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
+    if (!vk_memory_logger_enabled) {
+        return;
+    }
    std::lock_guard<std::mutex> guard(log_mutex);
    vk_buffer buf = buf_ref.lock();
    const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -1922,7 +1919,7 @@ void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
 }

 void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
-    if (buf_ref.expired() || buf_ref.lock()->size == 0) {
+    if (buf_ref.expired() || buf_ref.lock()->size == 0 || !vk_memory_logger_enabled) {
        return;
    }

@@ -1940,7 +1937,6 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
        VK_LOG_MEMORY("ERROR " << buf->device->name << ": Attempted to deallocate unknown " << type << " memory at " << buf->buffer);
    }
 }
-#endif // GGML_VULKAN_MEMORY_DEBUG

 struct vk_instance_t {
    vk::Instance instance;
@@ -2593,9 +2589,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
        buf->bda_addr = device->device.getBufferAddress(addressInfo);
    }

-#ifdef GGML_VULKAN_MEMORY_DEBUG
    device->memory_logger->log_allocation(buf, size);
-#endif

    return buf;
 }
@@ -2652,11 +2646,9 @@ static void ggml_vk_destroy_buffer(vk_buffer& buf) {
        return;
    }

-#ifdef GGML_VULKAN_MEMORY_DEBUG
    if (buf->device != nullptr) {
        buf->device->memory_logger->log_deallocation(buf);
    }
-#endif

    buf.reset();
 }
@@ -4477,9 +4469,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
        vk_device device = std::make_shared<vk_device_struct>();
        vk_instance.devices[idx] = device;

-#ifdef GGML_VULKAN_MEMORY_DEBUG
        device->memory_logger = std::unique_ptr<vk_memory_logger>(new vk_memory_logger());
-#endif

        size_t dev_num = vk_instance.device_indices[idx];

@@ -5476,6 +5466,7 @@ static void ggml_vk_instance_init() {
    vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
    vk_perf_logger_concurrent = getenv("GGML_VK_PERF_LOGGER_CONCURRENT") != nullptr;
    vk_enable_sync_logger = getenv("GGML_VK_SYNC_LOGGER") != nullptr;
+    vk_memory_logger_enabled = getenv("GGML_VK_MEMORY_LOGGER") != nullptr;
    const char* GGML_VK_PERF_LOGGER_FREQUENCY = getenv("GGML_VK_PERF_LOGGER_FREQUENCY");

    if (GGML_VK_PERF_LOGGER_FREQUENCY != nullptr) {
@@ -16,8 +16,8 @@ vendor = {
    # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
    "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",

-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.0/httplib.h": "vendor/cpp-httplib/httplib.h",
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.0/LICENSE":   "vendor/cpp-httplib/LICENSE",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/httplib.h": "vendor/cpp-httplib/httplib.h",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/LICENSE":   "vendor/cpp-httplib/LICENSE",

    "https://raw.githubusercontent.com/sheredom/subprocess.h/b49c56e9fe214488493021017bf3954b91c7c1f5/subprocess.h": "vendor/sheredom/subprocess.h",
 }
@@ -12,6 +12,7 @@
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `--license` | show source code license and dependencies |
 | `-cl, --cache-list` | show list of models in cache |
 | `--completion-bash` | print source-able bash completion script for llama.cpp |
 | `--verbose-prompt` | print a verbose prompt before generation (default: false) |
@@ -56,22 +57,23 @@
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | whether to memory-map model (if disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
-| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
+| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type<br/>(env: LLAMA_ARG_OVERRIDE_TENSOR) |
 | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
 | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
-| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM (default: -1)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
 | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
 | `-fit, --fit [on\|off]` | whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')<br/>(env: LLAMA_ARG_FIT) |
-| `-fitt, --fit-target MiB` | target margin per device for --fit option, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
+| `-fitt, --fit-target MiB0,MiB1,MiB2,...` | target margin per device for --fit, comma-separated list of values, single value is broadcast across all devices, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
 | `-fitc, --fit-ctx N` | minimum ctx size that can be set by --fit option, default: 4096<br/>(env: LLAMA_ARG_FIT_CTX) |
 | `--check-tensors` | check model tensor data for invalid values (default: false) |
-| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated or repeat this argument.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
+| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated values.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
 | `--op-offload, --no-op-offload` | whether to offload host tensor operations to device (default: true) |
 | `--lora FNAME` | path to LoRA adapter (use comma-separated values to load multiple adapters) |
 | `--lora-scaled FNAME:SCALE,...` | path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)<br/>note: use comma-separated values |
@@ -134,6 +136,7 @@
 | `--grammar-file FNAME` | file to read grammar from |
 | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
 | `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-bs, --backend-sampling` | enable backend sampling (experimental) (default: disabled)<br/>(env: LLAMA_ARG_BACKEND_SAMPLING) |


 ### CLI-specific params
@@ -164,19 +167,19 @@
 | `-otd, --override-tensor-draft <tensor name pattern>=<buffer type>,...` | override tensor buffer type for draft model |
 | `-cmoed, --cpu-moe-draft` | keep all Mixture of Experts (MoE) weights in the CPU for the draft model<br/>(env: LLAMA_ARG_CPU_MOE_DRAFT) |
 | `-ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_N_CPU_MOE_DRAFT) |
-| `--chat-template-kwargs STRING` | sets additional params for the json template parser<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
+| `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--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, 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, 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, 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, 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, deepseek2, deepseek3, 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, 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) |
 | `--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) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
 | `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
-| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
 | `--spec-replace TARGET DRAFT` | translate the string in TARGET into DRAFT if the draft model and main model are not compatible |
 | `--gpt-oss-20b-default` | use gpt-oss-20b (note: can download weights from the internet) |
@@ -95,6 +95,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `--license` | show source code license and dependencies |
 | `-cl, --cache-list` | show list of models in cache |
 | `--completion-bash` | print source-able bash completion script for llama.cpp |
 | `--verbose-prompt` | print a verbose prompt before generation (default: false) |
@@ -139,22 +140,23 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | whether to memory-map model (if disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
-| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
+| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type<br/>(env: LLAMA_ARG_OVERRIDE_TENSOR) |
 | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
 | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
-| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM (default: -1)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
 | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
 | `-fit, --fit [on\|off]` | whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')<br/>(env: LLAMA_ARG_FIT) |
-| `-fitt, --fit-target MiB` | target margin per device for --fit option, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
+| `-fitt, --fit-target MiB0,MiB1,MiB2,...` | target margin per device for --fit, comma-separated list of values, single value is broadcast across all devices, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
 | `-fitc, --fit-ctx N` | minimum ctx size that can be set by --fit option, default: 4096<br/>(env: LLAMA_ARG_FIT_CTX) |
 | `--check-tensors` | check model tensor data for invalid values (default: false) |
-| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated or repeat this argument.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
+| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated values.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
 | `--op-offload, --no-op-offload` | whether to offload host tensor operations to device (default: true) |
 | `--lora FNAME` | path to LoRA adapter (use comma-separated values to load multiple adapters) |
 | `--lora-scaled FNAME:SCALE,...` | path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)<br/>note: use comma-separated values |
@@ -217,6 +219,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--grammar-file FNAME` | file to read grammar from |
 | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
 | `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-bs, --backend-sampling` | enable backend sampling (experimental) (default: disabled)<br/>(env: LLAMA_ARG_BACKEND_SAMPLING) |


 ### Completion-specific params
@@ -248,8 +251,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--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, 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, 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, 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, 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, deepseek2, deepseek3, 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, 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) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |

 <!-- HELP_END -->
@@ -33,6 +33,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `--license` | show source code license and dependencies |
 | `-cl, --cache-list` | show list of models in cache |
 | `--completion-bash` | print source-able bash completion script for llama.cpp |
 | `--verbose-prompt` | print a verbose prompt before generation (default: false) |
@@ -73,22 +74,23 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | whether to memory-map model (if disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
-| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
+| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type<br/>(env: LLAMA_ARG_OVERRIDE_TENSOR) |
 | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
 | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
-| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM (default: -1)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
 | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
 | `-fit, --fit [on\|off]` | whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')<br/>(env: LLAMA_ARG_FIT) |
-| `-fitt, --fit-target MiB` | target margin per device for --fit option, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
+| `-fitt, --fit-target MiB0,MiB1,MiB2,...` | target margin per device for --fit, comma-separated list of values, single value is broadcast across all devices, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
 | `-fitc, --fit-ctx N` | minimum ctx size that can be set by --fit option, default: 4096<br/>(env: LLAMA_ARG_FIT_CTX) |
 | `--check-tensors` | check model tensor data for invalid values (default: false) |
-| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated or repeat this argument.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
+| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated values.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
 | `--op-offload, --no-op-offload` | whether to offload host tensor operations to device (default: true) |
 | `--lora FNAME` | path to LoRA adapter (use comma-separated values to load multiple adapters) |
 | `--lora-scaled FNAME:SCALE,...` | path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)<br/>note: use comma-separated values |
@@ -151,6 +153,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--grammar-file FNAME` | file to read grammar from |
 | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
 | `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-bs, --backend-sampling` | enable backend sampling (experimental) (default: disabled)<br/>(env: LLAMA_ARG_BACKEND_SAMPLING) |


 ### Server-specific params
@@ -187,11 +190,11 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--webui, --no-webui` | whether to enable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_WEBUI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--rerank, --reranking` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
-| `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
+| `--api-key KEY` | API key to use for authentication, multiple keys can be provided as a comma-separated list (default: none)<br/>(env: LLAMA_API_KEY) |
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
-| `--chat-template-kwargs STRING` | sets additional params for the json template parser<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
+| `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
 | `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
@@ -207,8 +210,8 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--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, 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, 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, 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, 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, deepseek2, deepseek3, 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, 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) |
 | `--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) |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
@@ -220,7 +223,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
-| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
 | `--spec-replace TARGET DRAFT` | translate the string in TARGET into DRAFT if the draft model and main model are not compatible |
 | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) |
@@ -779,7 +782,8 @@ By default, it is read-only. To make POST request to change global properties, y
  "modalities": {
    "vision": false
  },
-  "build_info": "b(build number)-(build commit hash)"
+  "build_info": "b(build number)-(build commit hash)",
+  "is_sleeping": false
 }
 ```

@@ -788,6 +792,7 @@ By default, it is read-only. To make POST request to change global properties, y
 - `model_path` - the path to model file (same with `-m` argument)
 - `chat_template` - the model's original Jinja2 prompt template
 - `modalities` - the list of supported modalities
+- `is_sleeping` - sleeping status, see [Sleeping on idle](#sleeping-on-idle)

 ### POST `/props`: Change server global properties.

@@ -1630,9 +1635,12 @@ The server supports an automatic sleep mode that activates after a specified per

 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.

+The sleeping status can be retrieved from the `GET /props` endpoint (or `/props?model=(model_name)` in router mode).
+
 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`
+- `GET /models`

 ## More examples

@@ -160,6 +160,7 @@ task_params server_task::params_from_json_cmpl(
    defaults.n_keep        = params_base.n_keep;
    defaults.n_predict     = params_base.n_predict;
    defaults.n_cache_reuse = params_base.n_cache_reuse;
+    defaults.cache_prompt  = params_base.cache_prompt;
    defaults.antiprompt    = params_base.antiprompt;

    // enabling this will output extra debug information in the HTTP responses from the server
@@ -169,7 +170,7 @@ task_params server_task::params_from_json_cmpl(
    params.stream           = json_value(data,       "stream",             false);
    auto stream_opt         = json_value(data,       "stream_options",     json::object());
    params.include_usage    = json_value(stream_opt, "include_usage",      false);
-    params.cache_prompt     = json_value(data,       "cache_prompt",       true);
+    params.cache_prompt     = json_value(data,       "cache_prompt",       defaults.cache_prompt);
    params.return_tokens    = json_value(data,       "return_tokens",      false);
    params.return_progress  = json_value(data,       "return_progress",    false);
    params.n_predict        = json_value(data,       "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
@@ -1138,6 +1138,7 @@ int getaddrinfo_with_timeout(const char *node, const char *service,

  return ret;
 #elif TARGET_OS_MAC
+  if (!node) { return EAI_NONAME; }
  // macOS implementation using CFHost API for asynchronous DNS resolution
  CFStringRef hostname_ref = CFStringCreateWithCString(
      kCFAllocatorDefault, node, kCFStringEncodingUTF8);
@@ -5569,14 +5570,11 @@ bool Server::read_content(Stream &strm, Request &req, Response &res) {
          strm, req, res,
          // Regular
          [&](const char *buf, size_t n) {
-            // Prevent arithmetic overflow when checking sizes.
-            // Avoid computing (req.body.size() + n) directly because
-            // adding two unsigned `size_t` values can wrap around and
-            // produce a small result instead of indicating overflow.
-            // Instead, check using subtraction: ensure `n` does not
-            // exceed the remaining capacity `max_size() - size()`.
-            if (req.body.size() >= req.body.max_size() ||
-                n > req.body.max_size() - req.body.size()) {
+            // Limit decompressed body size to payload_max_length_ to protect
+            // against "zip bomb" attacks where a small compressed payload
+            // decompresses to a massive size.
+            if (req.body.size() + n > payload_max_length_ ||
+                req.body.size() + n > req.body.max_size()) {
              return false;
            }
            req.body.append(buf, n);
@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H

-#define CPPHTTPLIB_VERSION "0.30.0"
-#define CPPHTTPLIB_VERSION_NUM "0x001E00"
+#define CPPHTTPLIB_VERSION "0.30.1"
+#define CPPHTTPLIB_VERSION_NUM "0x001E01"

 /*
 * Platform compatibility check
@@ -205,7 +205,10 @@

 #pragma comment(lib, "ws2_32.lib")

+#ifndef _SSIZE_T_DEFINED
 using ssize_t = __int64;
+#define _SSIZE_T_DEFINED
+#endif
 #endif // _MSC_VER

 #ifndef S_ISREG
@@ -2443,16 +2446,20 @@ namespace detail {

 #if defined(_WIN32)
 inline std::wstring u8string_to_wstring(const char *s) {
-  std::wstring ws;
+  if (!s) { return std::wstring(); }
+
  auto len = static_cast<int>(strlen(s));
+  if (!len) { return std::wstring(); }
+
  auto wlen = ::MultiByteToWideChar(CP_UTF8, 0, s, len, nullptr, 0);
-  if (wlen > 0) {
-    ws.resize(wlen);
-    wlen = ::MultiByteToWideChar(
-        CP_UTF8, 0, s, len,
-        const_cast<LPWSTR>(reinterpret_cast<LPCWSTR>(ws.data())), wlen);
-    if (wlen != static_cast<int>(ws.size())) { ws.clear(); }
-  }
+  if (!wlen) { return std::wstring(); }
+
+  std::wstring ws;
+  ws.resize(wlen);
+  wlen = ::MultiByteToWideChar(
+      CP_UTF8, 0, s, len,
+      const_cast<LPWSTR>(reinterpret_cast<LPCWSTR>(ws.data())), wlen);
+  if (wlen != static_cast<int>(ws.size())) { ws.clear(); }
  return ws;
 }
 #endif
Author	SHA1	Message	Date
Radoslav Gerganov	bcf7546160	server : add arg for disabling prompt caching (#18776 ) * server : add arg for disabling prompt caching Disabling prompt caching is useful for clients who are restricted to sending only OpenAI-compat requests and want deterministic responses. * address review comments * address review comments	2026-01-12 19:21:34 +02:00
Adrien Gallouët	36c5913c45	ci : use openssl for openEuler-latest-cmake-cann (#18779 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2026-01-12 17:29:00 +01:00
Adrien Gallouët	8e649571cd	vendor : update cpp-httplib to 0.30.1 (#18771 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2026-01-12 15:58:52 +01:00
Daniel Bevenius	4150da9a95	examples : add --kv-unified to batched example (#18774 ) This commit adds the --kv-unified flag to the batched example. This flag is currently specified in the README.md as required, but is currently not available as a command line option for the batched example. The motivation for this is that specifying this flag as the README instructs, will lead to an error about the flag not being recognized, and without this option the example fail with the following error: ```console split_equal: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag) decode: failed to find a memory slot for batch of size 4 main: llama_decode() failed ```	2026-01-12 13:47:58 +01:00
Jeff Bolz	8e2da778da	vulkan: change memory_logger to be controlled by an env var (#18769 )	2026-01-12 13:32:55 +01:00
Xuan-Son Nguyen	ce3bf9b1a4	server: update docs for sleeping [no ci] (#18777 )	2026-01-12 13:01:24 +01:00