vendor : update cpp-httplib to 0.30.1 (#18771)

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

SECURITY.md

@@ -38,7 +38,11 @@ Only vulnerabilities that fall within these parts of the project are considered
 - `src/**/*`
 - `ggml/**/*`
 - `gguf-py/**/*`
-- `tools/server/*` (note: Web UI is not covered)
+- `tools/server/*`, **excluding** the following topics:
+    - Web UI
+    - Features marked as experimental
+    - Features not recommended for use in untrusted environments (e.g., router, MCP)
+    - Bugs that can lead to Denial-of-Service attack
 
 Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.
 

common/arg.cpp

@@ -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"},

common/common.h

@@ -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,

examples/batched/batched.cpp

@@ -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;
     }
 

examples/debug/debug.cpp

@@ -57,11 +57,21 @@ struct callback_data {
     }
 };
 
+static bool has_pooling(llama_context * ctx) {
+    switch (llama_pooling_type(ctx)) {
+        case LLAMA_POOLING_TYPE_NONE:
+        case LLAMA_POOLING_TYPE_UNSPECIFIED:
+            return false;
+        default:
+            return true;
+    }
+}
+
 struct output_data {
     float *                  data_ptr    = nullptr;
     int                      data_size   = 0;
     std::string              type_suffix;
-    std::vector<float>       storage;
+    std::vector<float>       embd_norm;
     std::string              prompt;
     std::vector<llama_token> tokens;
 
@@ -73,24 +83,32 @@ struct output_data {
         prompt = params.prompt;
 
         if (params.embedding) {
-            const int  n_embd          = llama_model_n_embd_out(model);
-            const bool pooling_enabled = llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_NONE;
-            const int  n_embd_count    = pooling_enabled ? 1 : tokens.size();
-            const int  n_embeddings    = n_embd * n_embd_count;
+            const int n_embd       = llama_model_n_embd_out(model);
+            const bool pooling     = has_pooling(ctx);
+            const int n_embd_count = pooling ? 1 : tokens.size();
+            const int n_floats     = n_embd * n_embd_count;
 
-            float * embeddings;
-            if (pooling_enabled) {
-                embeddings = llama_get_embeddings_seq(ctx, 0);
-                storage.resize(n_embeddings);
-                common_embd_normalize(embeddings, storage.data(), n_embeddings, params.embd_normalize);
-                embeddings = storage.data();
-            } else {
-                embeddings = llama_get_embeddings(ctx);
+            float * embd_raw = pooling ? llama_get_embeddings_seq(ctx, 0) : llama_get_embeddings(ctx);
+            if (embd_raw == nullptr) {
+                throw std::runtime_error("failed to get embeddings from the model");
             }
 
-            data_ptr = embeddings;
-            data_size = n_embeddings;
+            LOG_DBG("pooling_enabled: %s\n", pooling ? "true" : "false");
+            LOG_DBG("n_embd: %d\n", n_embd);
+            LOG_DBG("n_floats: %d\n", n_floats);
+            LOG_DBG("n_embd_count: %d\n", n_embd_count);
+
+            data_ptr    = embd_raw;
+            data_size   = n_floats;
             type_suffix = "-embeddings";
+
+            if (params.embd_normalize >= 0) {
+                embd_norm.resize(n_floats);
+                for (int i = 0; i < n_embd_count; i++) {
+                    common_embd_normalize(embd_raw+i*n_embd, embd_norm.data()+i*n_embd, n_embd, params.embd_normalize);
+                }
+                data_ptr = embd_norm.data();
+            }
         } else {
             const float * logits = llama_get_logits_ith(ctx, tokens.size() - 1);
             const int n_logits = llama_vocab_n_tokens(vocab);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -119,6 +119,8 @@ struct ggml_backend_vk_context;
 // Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
 #define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 3)
 
+typedef std::shared_ptr<struct vk_pipeline_struct> vk_pipeline;
+
 struct vk_pipeline_struct {
     std::string name;
     vk::ShaderModule shader_module;
@@ -136,9 +138,15 @@ struct vk_pipeline_struct {
     std::atomic<bool> compiled {};
     // number of registers used, extracted from pipeline executable properties
     uint32_t register_count {};
+
+#if defined(VK_EXT_shader_64bit_indexing)
+    bool is_64b_indexing {};
+#endif
+    // linked list of pipelines for multiple compilation variants.
+    // currently only used to compile a 64-bit indexing variant.
+    vk_pipeline next;
 };
 
-typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline;
 typedef std::weak_ptr<vk_pipeline_struct> vk_pipeline_ref;
 
 static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
@@ -230,9 +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);
@@ -584,6 +590,8 @@ struct vk_device_struct {
     bool add_rms_fusion;
     uint32_t partials_binding_alignment;
 
+    bool shader_64b_indexing;
+
     bool integer_dot_product;
     // 0: default, 1: force mmvq, -1: disable mmvq
     int32_t mmvq_mode;
@@ -815,9 +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);
@@ -1553,8 +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;
@@ -1587,10 +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;
@@ -1897,10 +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);
@@ -1912,7 +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;
     }
 
@@ -1930,7 +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;
@@ -2080,6 +2086,19 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         compute_pipeline_create_info.setPNext(&rci);
     }
 
+#if defined(VK_EXT_shader_64bit_indexing)
+    vk::PipelineCreateFlags2CreateInfo pipelineFlags2CreateInfo;
+    if (pipeline->is_64b_indexing)
+    {
+        pipelineFlags2CreateInfo.flags = vk::PipelineCreateFlagBits2::e64BitIndexingEXT;
+        if (device->pipeline_executable_properties_support) {
+            pipelineFlags2CreateInfo.flags |= vk::PipelineCreateFlagBits2::eCaptureStatisticsKHR;
+        }
+        pipelineFlags2CreateInfo.setPNext(compute_pipeline_create_info.pNext);
+        compute_pipeline_create_info.setPNext(&pipelineFlags2CreateInfo);
+    }
+#endif
+
     try {
         pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
     } catch (const vk::SystemError& e) {
@@ -2570,9 +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;
 }
@@ -2629,11 +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();
 }
@@ -3002,6 +3017,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
             m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
             m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
+        } else if (device->vendor_id == VK_VENDOR_ID_AMD && device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary) {
+            // This is intentionally using tx_m values, slight performance increase
+            l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
         } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support && device->architecture == INTEL_XE2) {
             // Xe2/Xe3 with coopmat enabled - warptile performance tuning
             l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
@@ -3061,7 +3081,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     }
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const char *name, size_t spv_size, const void* spv_data, const char *entrypoint,
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& base_pipeline, const char *name, size_t spv_size, const void* spv_data, const char *entrypoint,
                                               uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
                                               uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
 
@@ -3069,35 +3089,49 @@ static void ggml_vk_load_shaders(vk_device& device) {
             required_subgroup_size = get_subgroup_size(name, device->architecture);
         }
 
-        if (!pipeline) {
-            pipeline = std::make_shared<vk_pipeline_struct>();
-        }
-        if (!pipeline->initialized) {
-            pipeline->name = name;
-            pipeline->parameter_count = parameter_count;
-            pipeline->push_constant_size = push_constant_size;
-            pipeline->wg_denoms = wg_denoms;
-            pipeline->align = align;
-            pipeline->initialized = true;
-        }
+        vk_pipeline *ptr = &base_pipeline;
 
-        if (!pipeline->needed || pipeline->compiled) {
-            return;
+        int num_pipelines = 1;
+#if defined(VK_EXT_shader_64bit_indexing)
+        if (device->shader_64b_indexing) {
+            num_pipelines = 2;
         }
-        // TODO: We're no longer benefitting from the async compiles (shaders are
-        // compiled individually, as needed) and this complexity can be removed.
-        {
-            // wait until fewer than N compiles are in progress
-            uint32_t N = std::max(1u, std::thread::hardware_concurrency());
-            std::unique_lock<std::mutex> guard(compile_count_mutex);
-            while (compile_count >= N) {
-                compile_count_cond.wait(guard);
+#endif
+        for (int i = 0; i < num_pipelines; ++i, ptr = &(*ptr)->next) {
+            vk_pipeline &pipeline = *ptr;
+            if (!pipeline) {
+                pipeline = std::make_shared<vk_pipeline_struct>();
+            }
+            if (!pipeline->initialized) {
+                pipeline->name = name;
+                pipeline->parameter_count = parameter_count;
+                pipeline->push_constant_size = push_constant_size;
+                pipeline->wg_denoms = wg_denoms;
+                pipeline->align = align;
+                pipeline->initialized = true;
+#if defined(VK_EXT_shader_64bit_indexing)
+                pipeline->is_64b_indexing = (i == 1);
+#endif
             }
-            compile_count++;
-        }
 
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint,
-                                      parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
+            if (!pipeline->needed || pipeline->compiled) {
+                continue;
+            }
+            // TODO: We're no longer benefitting from the async compiles (shaders are
+            // compiled individually, as needed) and this complexity can be removed.
+            {
+                // wait until fewer than N compiles are in progress
+                uint32_t N = std::max(1u, std::thread::hardware_concurrency());
+                std::unique_lock<std::mutex> guard(compile_count_mutex);
+                while (compile_count >= N) {
+                    compile_count_cond.wait(guard);
+                }
+                compile_count++;
+            }
+
+            compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint,
+                                          parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
+        }
     };
 
     auto const &ggml_vk_create_pipeline2 = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const char *entrypoint,
@@ -4435,9 +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];
 
@@ -4475,6 +4507,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool pipeline_executable_properties_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
+        device->shader_64b_indexing = false;
         bool bfloat16_support = false;
 
         for (const auto& properties : ext_props) {
@@ -4522,6 +4555,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 device->memory_priority = true;
             } else if (strcmp("VK_EXT_external_memory_host", properties.extensionName) == 0) {
                 device->external_memory_host = true;
+#if defined(VK_EXT_shader_64bit_indexing)
+            } else if (strcmp("VK_EXT_shader_64bit_indexing", properties.extensionName) == 0) {
+                device->shader_64b_indexing = true;
+#endif
             }
         }
 
@@ -4812,6 +4849,16 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_EXT_external_memory_host");
         }
 
+#if defined(VK_EXT_shader_64bit_indexing)
+        VkPhysicalDeviceShader64BitIndexingFeaturesEXT shader_64bit_indexing_features {};
+        shader_64bit_indexing_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_64_BIT_INDEXING_FEATURES_EXT;
+        if (device->shader_64b_indexing) {
+            last_struct->pNext = (VkBaseOutStructure *)&shader_64bit_indexing_features;
+            last_struct = (VkBaseOutStructure *)&shader_64bit_indexing_features;
+            device_extensions.push_back("VK_EXT_shader_64bit_indexing");
+        }
+#endif
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->pipeline_executable_properties_support = pipeline_executable_properties_support;
@@ -5078,7 +5125,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
             switch (device->vendor_id) {
 #ifndef GGML_VULKAN_RUN_TESTS
             case VK_VENDOR_ID_AMD:
-                device->mul_mat_l[i]    = false;
+                device->mul_mat_l[i]    = device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary;
                 device->mul_mat_m[i]    = true;
                 device->mul_mat_s[i]    = true;
                 device->mul_mat_id_l[i] = false;
@@ -5419,6 +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) {
@@ -6897,6 +6945,20 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
+static vk_pipeline ggml_vk_get_64b_indexing_pipeline(ggml_backend_vk_context * ctx, vk_pipeline &pipeline) {
+    GGML_UNUSED(ctx);
+#if defined(VK_EXT_shader_64bit_indexing)
+    vk_pipeline *ptr = &pipeline;
+    while (*ptr) {
+        if ((*ptr)->is_64b_indexing) {
+            return *ptr;
+        }
+        ptr = &(*ptr)->next;
+    }
+#endif
+    return pipeline;
+}
+
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool disable_split_k) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -6980,6 +7042,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
     const uint64_t x_ne = ggml_nelements(src0);
@@ -7289,6 +7355,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
     }
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        dmmv = ggml_vk_get_64b_indexing_pipeline(ctx, dmmv);
+    }
+
     const bool qx_needs_dequant = x_non_contig;
     const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
 
@@ -7484,9 +7554,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         gqa_ratio = 1;
     }
 
+    vk_pipeline pipeline = ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1];
+
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
@@ -7528,7 +7604,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         workgroups_z /= gqa_ratio;
     }
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1],
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
         {
             d_Qx,
             d_Qy,
@@ -7578,9 +7654,14 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_y = nb12 / sizeof(float);
 
+    vk_pipeline pipeline = ctx->device->pipeline_mul_mat_vec_nc_f16_f32;
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
@@ -7617,7 +7698,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     init_pushconst_tensor_offsets(ctx, pc, src0, src1, nullptr, nullptr, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
         {
             d_Qx,
             d_Qy,
@@ -7636,8 +7717,9 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     // Handle huge A matrix by splitting the M dimensions. This works well for convolution use cases
     // where the M dimension is very large.
     // Split_k doesn't work with M splitting.
+    // This only supports batchsize == 1.
     const size_t nbytes = ggml_nbytes(src0);
-    const bool needs_split = nbytes > ctx->device->properties.limits.maxStorageBufferRange;
+    const bool needs_split = dst->ne[2] == 1 && dst->ne[3] == 1 && nbytes > ctx->device->properties.limits.maxStorageBufferRange;
     if (needs_split) {
         // Choose the number of rows that can fit (and divide by two, to allow for any additional offsets)
         const uint32_t M_split = ctx->device->properties.limits.maxStorageBufferRange / (2 * src0->nb[1]);
@@ -7779,6 +7861,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
     const uint64_t x_ne = ggml_nelements(src0);
@@ -8040,6 +8125,10 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     const bool qx_needs_dequant = x_non_contig;
     const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        dmmv = ggml_vk_get_64b_indexing_pipeline(ctx, dmmv);
+    }
+
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -87,7 +87,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K;
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl

@@ -65,9 +65,9 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 
     a_offset =
 #ifdef MUL_MAT_ID
-            expert_id * p.batch_stride_a;
+            expert_id * (p.batch_stride_a / QUANT_K);
 #else
-            batch_idx_a * p.batch_stride_a;
+            batch_idx_a * (p.batch_stride_a / QUANT_K);
 #endif
     b_offset =
 #ifdef MUL_MAT_ID

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
                                const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
     // Compute starting index in matrix B for this superblock
     const uint y_idx = i * QUANT_K + 32 * ib32;
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
 
     // Precompute indices for quantization lookup tables
     const uint qh_base = 2 * ib32;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -17,7 +17,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
             const vec4 b_val_1 = vec4(data_b_v4[base_b_idx + 2 * l + 1]);
 
             // index for data_a
-            uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+            uint ibi = a_offset + first_row * num_blocks_per_row + i;
 
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
                 const float d = float(data_a[ibi].d);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -12,7 +12,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     // Precompute db multiplication factors
     float db_vals[NUM_ROWS];
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -22,7 +22,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         db_vals[n] = d * (0.125f + float(scale) * 0.25f);
         ibi += num_blocks_per_row;
     }
-    ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         // Preload grid and sign data for all l values
         vec4 grid0_vals[2], grid1_vals[2];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint signscale = pack32(u16vec2(

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

@@ -10,7 +10,7 @@ FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
     const uint y_idx = i * QUANT_K + 32 * ib32;
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint signscale = pack32(u16vec2(

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -14,7 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -13,7 +13,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -13,7 +13,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp

@@ -79,7 +79,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K_Q8_1;
+    a_offset *= QUANT_K / QUANT_K_Q8_1;
     b_offset /= QUANT_K_Q8_1;
 
     FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -234,13 +234,13 @@ void main() {
     const uint end_k = min(p.K, (ik + 1) * p.k_split);
 #endif
 
-    uint pos_a = (
+    uint pos_a =
 #ifdef MUL_MAT_ID
-        expert_idx * p.batch_stride_a +
+        expert_idx * (p.batch_stride_a / LOAD_VEC_A) +
 #else
-        batch_idx_a * p.batch_stride_a +
+        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A) +
 #endif
-        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
 #ifdef MUL_MAT_ID
     uint pos_b = 0;
 #else

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -250,10 +250,10 @@ void main() {
 #endif
 
 #ifdef MUL_MAT_ID
-    uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K;
+    uint pos_a = expert_idx * (p.batch_stride_a / QUANT_K);
     uint pos_b = 0;
 #else
-    uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K;
+    uint pos_a = batch_idx_a * (p.batch_stride_a / QUANT_K);
     uint pos_b = batch_idx * p.batch_stride_b;
     uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp

@@ -189,13 +189,13 @@ void main() {
     const uint end_k = min(p.K, (ik + 1) * p.k_split);
 #endif
 
-    uint pos_a_ib = (
+    uint pos_a_ib =
 #ifdef MUL_MAT_ID
-        expert_idx * p.batch_stride_a +
+        expert_idx * (p.batch_stride_a / BK) +
 #else
-        batch_idx_a * p.batch_stride_a +
+        batch_idx_a * (p.batch_stride_a / BK) +
 #endif
-        ir * BM * p.stride_a + start_k) / BK;
+        (ir * BM * p.stride_a + start_k) / BK;
 #ifdef MUL_MAT_ID
     uint pos_b_ib = 0;
 #else

scripts/sync_vendor.py

@@ -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",
 }

src/models/qwen3next.cpp

@@ -590,6 +590,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
         ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,
                                     mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
                                     (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * ggml_element_size(mixed_qkvz_reshaped));
+        z = ggml_cont(ctx0, z);
         cb(z, "z", il);
 
         // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions

tests/test-backend-ops.cpp

@@ -7560,6 +7560,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000, 96, 2592, {1, 1}, {1, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000,  3, 2592, {1, 1}, {1, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 1700000,  1, 2592, {1, 1}, {1, 1}));
+
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_Q8_0, GGML_TYPE_F32, 128, 128, false, 8192, 2, 5120)); // Llama-4-Maverick-17B-128E-PAB-Q8_0
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_Q8_0, GGML_TYPE_F32, 128, 128, false, 8192, 1, 5120)); // Llama-4-Maverick-17B-128E-PAB-Q8_0
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 8192, 1, 5120, {128, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 8192, 512, 5120, {128, 1}, {1, 1}));
 #endif
 
     for (ggml_type type_a : all_types) {

tools/cli/README.md

@@ -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) |

tools/completion/README.md

@@ -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 -->

tools/server/README.md

@@ -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
 

vendor/cpp-httplib/httplib.cpp

@@ -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);

vendor/cpp-httplib/httplib.h

@@ -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