also show model aliases

Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>

common/arg.cpp

@@ -718,9 +718,8 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
 
         // model is required (except for server)
         // TODO @ngxson : maybe show a list of available models in CLI in this case
-        if (params.model.path.empty()
-                && !params.usage
-                && !params.completion) {
+        bool can_skip_model = params.usage || params.completion || !params.server_base.empty();
+        if (!can_skip_model && params.model.path.empty()) {
             throw std::invalid_argument("error: --model is required\n");
         }
     }
@@ -1240,6 +1239,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.completion = true;
         }
     ));
+    add_opt(common_arg(
+        {"--server-base"}, "URL",
+        string_format("connect to this server instead of starting a new one, example: 'http://localhost:8080' (default: none)"),
+        [](common_params & params, const std::string & value) {
+            params.server_base = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"--verbose-prompt"},
         string_format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"),

common/common.h

@@ -644,6 +644,9 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
+    // CLI params
+    std::string server_base; // if set, connect to this server instead of starting a new one
+
     // UI configs
     bool ui = true;
     bool ui_mcp_proxy = false;

common/http.h

@@ -2,6 +2,16 @@
 
 #include <cpp-httplib/httplib.h>
 
+#ifdef _WIN32
+#include <winsock2.h>
+#include <windows.h>
+#else
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#endif
+
 struct common_http_url {
     std::string scheme;
     std::string user;
@@ -119,3 +129,63 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
 static std::string common_http_show_masked_url(const common_http_url & parts) {
     return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + common_http_format_host(parts.host) + parts.path;
 }
+
+static int common_http_get_free_port() {
+#ifdef _WIN32
+    WSADATA wsaData;
+    if (WSAStartup(MAKEWORD(2, 2), &wsaData) != 0) {
+        return -1;
+    }
+    typedef SOCKET native_socket_t;
+#define INVALID_SOCKET_VAL INVALID_SOCKET
+#define CLOSE_SOCKET(s) closesocket(s)
+#else
+    typedef int native_socket_t;
+#define INVALID_SOCKET_VAL -1
+#define CLOSE_SOCKET(s) close(s)
+#endif
+
+    native_socket_t sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock == INVALID_SOCKET_VAL) {
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+    struct sockaddr_in serv_addr;
+    std::memset(&serv_addr, 0, sizeof(serv_addr));
+    serv_addr.sin_family = AF_INET;
+    serv_addr.sin_addr.s_addr = htonl(INADDR_ANY);
+    serv_addr.sin_port = htons(0);
+
+    if (bind(sock, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) != 0) {
+        CLOSE_SOCKET(sock);
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+#ifdef _WIN32
+    int namelen = sizeof(serv_addr);
+#else
+    socklen_t namelen = sizeof(serv_addr);
+#endif
+    if (getsockname(sock, (struct sockaddr*)&serv_addr, &namelen) != 0) {
+        CLOSE_SOCKET(sock);
+#ifdef _WIN32
+        WSACleanup();
+#endif
+        return -1;
+    }
+
+    int port = ntohs(serv_addr.sin_port);
+
+    CLOSE_SOCKET(sock);
+#ifdef _WIN32
+    WSACleanup();
+#endif
+
+    return port;
+}

ggml/src/ggml-cpu/ops.cpp

@@ -5025,8 +5025,8 @@ void ggml_compute_forward_get_rows(
     //}
 }
 
-template<typename src_t, typename idx_t>
-static void ggml_compute_forward_set_rows_impl(
+template<typename idx_t>
+static void ggml_compute_forward_set_rows_f32(
         const ggml_compute_params * params,
               ggml_tensor * dst) {
 
@@ -5041,7 +5041,7 @@ static void ggml_compute_forward_set_rows_impl(
     assert(ne0  == nc);
     assert(ne2  == ne02);
     assert(ne3  == ne03);
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
+    assert(src0->type == GGML_TYPE_F32);
     assert(ne02 % ne11 == 0);
     assert(ne03 % ne12 == 0);
 
@@ -5055,8 +5055,6 @@ static void ggml_compute_forward_set_rows_impl(
     const int64_t ir0 = dr*ith;
     const int64_t ir1 = std::min(ir0 + dr, nr);
 
-    const size_t rs = ggml_row_size(src0->type, nc);
-
     ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
 
     for (int64_t i03 = 0; i03 < ne03; ++i03) {
@@ -5070,18 +5068,9 @@ static void ggml_compute_forward_set_rows_impl(
 
                 GGML_ASSERT(i1 >= 0 && i1 < ne1);
 
-                if constexpr (std::is_same_v<src_t, float>) {
-                    from_float(
-                            (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
-                                            ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
-                } else if constexpr (std::is_same_v<src_t, ggml_fp16_t>) {
-                    memcpy(
-                                            ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3),
-                                            ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
-                                            rs);
-                } else {
-                    GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
-                }
+                from_float(
+                        (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
+                                        ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
             }
         }
     }
@@ -5098,27 +5087,13 @@ void ggml_compute_forward_set_rows(
         case GGML_TYPE_F32:
             {
                 if (src1->type == GGML_TYPE_I64) {
-                    ggml_compute_forward_set_rows_impl<float, int64_t>(params, dst);
+                    ggml_compute_forward_set_rows_f32<int64_t>(params, dst);
                 } else if (src1->type == GGML_TYPE_I32) {
-                    ggml_compute_forward_set_rows_impl<float, int32_t>(params, dst);
+                    ggml_compute_forward_set_rows_f32<int32_t>(params, dst);
                 } else {
                     GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
                 }
             } break;
-        case GGML_TYPE_F16:
-            {
-                if (dst->type == GGML_TYPE_F16) {
-                    if (src1->type == GGML_TYPE_I64) {
-                        ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
-                    } else if (src1->type == GGML_TYPE_I32) {
-                        ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
-                    } else {
-                        GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
-                    }
-                } else {
-                    GGML_ABORT("dst->type = %d (%s) not supported with src0->type = %d (%s)", dst->type, ggml_type_name(dst->type), src0->type, ggml_type_name(src0->type));
-                }
-            } break;
         default:
             {
                 GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));

ggml/src/ggml-cpu/simd-gemm.h

@@ -78,7 +78,7 @@ static void simd_gemm(
             for (int64_t i = 0; i < GEMM_RM; i++) {
                 float a = C[i * N + jj];
                 for (int64_t kk = 0; kk < K; kk++) {
-                    a += A[i * K + kk] * B[kk * N + jj];
+                    a += A[i + kk] * B[kk * N + jj];
                 }
                 C[i * N + jj] = a;
             }

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -160,15 +160,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows(ggml_me
     return res;
 }
 
-ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_metal_library_t lib, const ggml_tensor * op) {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_metal_library_t lib, ggml_type tidx, ggml_type tdst) {
     char base[256];
     char name[256];
 
-    const auto tsrc = op->src[0]->type;
-    const auto tidx = op->src[1]->type;
-    const auto tdst = op->type;
-
-    snprintf(base, 256, "kernel_set_rows_%s_%s_%s", ggml_type_name(tsrc), ggml_type_name(tidx), ggml_type_name(tdst));
+    snprintf(base, 256, "kernel_set_rows_%s_%s", ggml_type_name(tdst), ggml_type_name(tidx));
     snprintf(name, 256, "%s", base);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);

ggml/src/ggml-metal/ggml-metal-device.h

@@ -112,7 +112,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cpy
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_1d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
-struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_concat            (ggml_metal_library_t lib, enum ggml_type tsrc);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1334,7 +1334,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return op->src[0]->type != GGML_TYPE_NVFP4;
         case GGML_OP_SET_ROWS:
             {
-                if (op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) {
+                if (op->src[0]->type != GGML_TYPE_F32) {
                     return false;
                 }
 

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -1202,7 +1202,7 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    auto pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op);
+    auto pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
 
     const int32_t nk0 = ne0/ggml_blck_size(op->type);
 

ggml/src/ggml-metal/ggml-metal.metal

@@ -42,8 +42,6 @@ typedef matrix<bfloat, 4, 4> bfloat4x4;
 typedef matrix<bfloat, 2, 4> bfloat2x4;
 #endif
 
-#define QK_NL 16
-
 constexpr constant static float kvalues_iq4nl_f[16] = {
     -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
 };
@@ -9388,40 +9386,7 @@ kernel void kernel_get_rows_f(
     }
 }
 
-typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
-
-template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
-template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
-template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
-#if defined(GGML_METAL_HAS_BF16)
-template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
-#endif
-
-typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
-
-template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
-template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
-template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
-template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
-template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
-template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
-template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
-template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
-template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-
-template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+template<typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
 kernel void kernel_set_rows_q32(
         constant ggml_metal_kargs_set_rows & args,
         device const  void * src0,
@@ -9445,14 +9410,14 @@ kernel void kernel_set_rows_q32(
     const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
 
           device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
-    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    const device float   * src_row = (const device float   *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
 
     for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
         quantize_func(src_row + 32*ind, dst_row[ind]);
     }
 }
 
-template<typename TS, typename TI, typename TD>
+template<typename T, typename TI>
 kernel void kernel_set_rows_f(
         constant ggml_metal_kargs_set_rows & args,
         device const  void * src0,
@@ -9475,47 +9440,14 @@ kernel void kernel_set_rows_f(
     const int32_t i10 = i01;
     const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
 
-          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
-    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+          device T     * dst_row = (      device T     *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float * src_row = (const device float *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
 
     for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
-        dst_row[ind] = (TD) src_row[ind];
+        dst_row[ind] = (T) src_row[ind];
     }
 }
 
-typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
-
-template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
-template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
-template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
-template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
-#if defined(GGML_METAL_HAS_BF16)
-template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
-template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
-#endif
-
-template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
-template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
-#if defined(GGML_METAL_HAS_BF16)
-template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
-template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
-#endif
-
-typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
-
-template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
-template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
-template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
-template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
-template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
-template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
-template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
-template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
-template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
-template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
-template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
-template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
-
 kernel void kernel_diag_f32(
         constant ggml_metal_kargs_diag & args,
         device   const char * src0,
@@ -10258,6 +10190,75 @@ kernel void kernel_mul_mm_id(
     }
 }
 
+#define QK_NL 16
+
+//
+// get rows
+//
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+//
+// set rows
+//
+
+typedef decltype(kernel_set_rows_f<float, int64_t>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t>;
+template [[host_name("kernel_set_rows_f32_i32")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t>;
+template [[host_name("kernel_set_rows_f16_i64")]]  kernel set_rows_f_t kernel_set_rows_f<half, int64_t>;
+template [[host_name("kernel_set_rows_f16_i32")]]  kernel set_rows_f_t kernel_set_rows_f<half, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t>;
+template [[host_name("kernel_set_rows_bf16_i32")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_q8_0_i64")]]   kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_q8_0_i32")]]   kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_q4_0_i64")]]   kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_q4_0_i32")]]   kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_q4_1_i64")]]   kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_q4_1_i32")]]   kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_q5_0_i64")]]   kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_q5_0_i32")]]   kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_q5_1_i64")]]   kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_q5_1_i32")]]   kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_iq4_nl_i64")]] kernel set_rows_q32_t kernel_set_rows_q32<int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_iq4_nl_i32")]] kernel set_rows_q32_t kernel_set_rows_q32<int32_t, block_iq4_nl, quantize_iq4_nl>;
+
 //
 // matrix-matrix multiplication
 //

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -16653,7 +16653,6 @@ static cl_mem ggml_cl_mul_mat_dequant_quant_to_f16(
         ? ggml_cl_is_q4_0_soa(tensor)
         : ggml_cl_is_q8_0_soa(tensor);
 
-    cl_mem aos = nullptr;
     if (is_soa) {
         // Reconstruct full parent AoS; view's own nb[] then index it correctly.
         const ggml_tensor * parent = tensor->view_src ? tensor->view_src : tensor;
@@ -16665,7 +16664,7 @@ static cl_mem ggml_cl_mul_mat_dequant_quant_to_f16(
         const size_t parent_nbytes = (size_t) ggml_nelements(parent) / blck_size * block_bytes;
 
         cl_int err;
-        aos = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, parent_nbytes, NULL, &err);
+        cl_mem aos = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, parent_nbytes, NULL, &err);
         CL_CHECK(err);
 
         // large q4_0/q8_0 WEIGHTS are stored transposed and small weights
@@ -16752,6 +16751,9 @@ static cl_mem ggml_cl_mul_mat_dequant_quant_to_f16(
 
         if (extra_reconstruct) {
             *extra_reconstruct = aos;
+        } else {
+            // OpenCL retains the memobj while queued kernels reference it.
+            CL_CHECK(clReleaseMemObject(aos));
         }
     } else {
         auto * extra = (ggml_tensor_extra_cl *) tensor->extra;
@@ -16815,13 +16817,6 @@ static cl_mem ggml_cl_mul_mat_dequant_quant_to_f16(
     size_t lws[3] = { 1, 1, 1 };
     CL_CHECK(clEnqueueNDRangeKernel(backend_ctx->queue, dq_kernel, 3, NULL, gws, lws, 0, NULL, NULL));
 
-    // release the reconstructed aos if
-    //  1. it was actually reconstructed
-    //  2. the caller didn't request it to be returned
-    // src_buf may refer to aos, so we should release after this enqueue
-    if (aos && !extra_reconstruct) {
-        CL_CHECK(clReleaseMemObject(aos));
-    }
     return out;
 }
 

ggml/src/ggml.c

@@ -3926,7 +3926,7 @@ struct ggml_tensor * ggml_set_rows(
     GGML_ASSERT(b->ne[2] % c->ne[1] == 0);
     GGML_ASSERT(b->ne[3] % c->ne[2] == 0);
     GGML_ASSERT(c->ne[3] == 1);
-    GGML_ASSERT(b->type == GGML_TYPE_F32 || b->type == GGML_TYPE_F16);
+    GGML_ASSERT(b->type == GGML_TYPE_F32);
     GGML_ASSERT(c->type == GGML_TYPE_I64 || c->type == GGML_TYPE_I32);
 
     GGML_ASSERT(ggml_is_contiguous_rows(a));

src/llama-vocab.cpp

@@ -887,6 +887,9 @@ struct llm_tokenizer_ugm : llm_tokenizer {
             // blob containing XOR-compressed compact double array (XCDA) entries
             uint32_t xcda_blob_size = *(const uint32_t *) &precompiled_charsmap[0];
             charsmap_offset += sizeof(xcda_blob_size);
+            if (xcda_blob_size + charsmap_offset >= precompiled_charsmap.size()) {
+                throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
+            }
 
             // Next xcda_blob_size bytes contain entries of XOR-compressed compact
             // double array (XCDA). Each entry is bit-packed into a 32-bit integer.
@@ -1202,15 +1205,7 @@ private:
                 throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
             }
             const char * prefix_replacement = &(tokenizer.prefix_replacements)[longest_prefix_offset];
-            size_t max_len = tokenizer.prefix_replacements_size - longest_prefix_offset;
-            size_t repl_len = 0;
-            while (repl_len < max_len && prefix_replacement[repl_len] != '\0') {
-                repl_len++;
-            }
-            if (repl_len == max_len) {
-                throw std::runtime_error("Unterminated string in precompiled charsmap!");
-            }
-            return { prefix_replacement, repl_len, longest_prefix_length };
+            return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
         }
 
         // check if the input prefix contains a valid sequence of UTF-8 code units
@@ -2023,18 +2018,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 const size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
                 const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                 precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
-                if (precompiled_charsmap.size() < sizeof(uint32_t)) {
-                    throw std::runtime_error("precompiled_charsmap too small for xcda_blob_size header!");
-                }
-                uint32_t * xcda_blob_size = (uint32_t *) &precompiled_charsmap[0];
-#if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-                *xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
-#endif
-                if (*xcda_blob_size + sizeof(uint32_t) >= precompiled_charsmap.size()) {
-                    throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
-                }
 #if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
                 // correct endianness of data in precompiled_charsmap binary blob
+                uint32_t * xcda_blob_size = (uint32_t *) &precompiled_charsmap[0];
+                *xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
+                assert(*xcda_blob_size + sizeof(uint32_t) < n_precompiled_charsmap);
                 size_t xcda_array_size = *xcda_blob_size / sizeof(uint32_t);
                 uint32_t * xcda_array = (uint32_t *) &precompiled_charsmap[sizeof(uint32_t)];
                 for (size_t i = 0; i < xcda_array_size; ++i) {

tests/test-backend-ops.cpp

@@ -2393,8 +2393,7 @@ static void init_set_rows_row_ids(ggml_tensor * t, int num_rows) {
 
 // GGML_OP_SET_ROWS
 struct test_set_rows : public test_case {
-    const ggml_type type_src;
-    const ggml_type type_dst;
+    const ggml_type type;
     const ggml_type type_idx;
     const std::array<int64_t, 4> ne;
     const std::array<int, 2> nr23; // broadcast only dims 2 and 3
@@ -2402,22 +2401,21 @@ struct test_set_rows : public test_case {
     const bool v; // view (non-contiguous src1)
 
     std::string vars() override {
-        return VARS_TO_STR7(type_src, type_dst, type_idx, ne, nr23, r, v);
+        return VARS_TO_STR6(type, type_idx, ne, nr23, r, v);
     }
 
-    test_set_rows(ggml_type type_src,
-            ggml_type type_dst,
+    test_set_rows(ggml_type type,
             ggml_type type_idx,
             std::array<int64_t, 4> ne,
             std::array<int, 2> nr23,
             int r, bool v = false)
-        : type_src(type_src), type_dst(type_dst), type_idx(type_idx), ne(ne), nr23(nr23), r(r), v(v) {}
+        : type(type), type_idx(type_idx), ne(ne), nr23(nr23), r(r), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * dst = ggml_new_tensor_4d(ctx, type_dst, ne[0], ne[1], ne[2]*nr23[0], ne[3]*nr23[1]);
+        ggml_tensor * dst = ggml_new_tensor_4d(ctx, type,          ne[0], ne[1], ne[2]*nr23[0], ne[3]*nr23[1]);
         ggml_set_name(dst, "dst");
 
-        ggml_tensor * src = ggml_new_tensor_4d(ctx, type_src, ne[0], r,     ne[2]*nr23[0], ne[3]*nr23[1]);
+        ggml_tensor * src = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], r,     ne[2]*nr23[0], ne[3]*nr23[1]);
         ggml_set_name(src, "src");
 
         ggml_tensor * row_idxs = ggml_new_tensor_3d(ctx, type_idx, r, ne[2], ne[3]);
@@ -2450,17 +2448,17 @@ struct test_set_rows : public test_case {
     }
 
     double max_nmse_err() override {
-        if (type_dst == GGML_TYPE_Q4_0 || type_dst == GGML_TYPE_Q4_1 || type_dst == GGML_TYPE_IQ4_NL ||
-            type_dst == GGML_TYPE_Q5_0 || type_dst == GGML_TYPE_Q5_1 || type_dst == GGML_TYPE_Q8_0) {
+        if (type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_IQ4_NL ||
+            type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1 || type == GGML_TYPE_Q8_0) {
             // estimate what the max nmse error would be if one quantized value is
             // off by one. The test values are distributed in [-1,1], so it'll be
             // roughly (2.0 / 2^bits)^2, divided by the mean square value of the reference,
             // which is roughly 0.25 times the number of elements.
             double err_estimate = 1.0f/8.0f;
-            if (type_dst == GGML_TYPE_Q5_0 || type_dst == GGML_TYPE_Q5_1) {
+            if (type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
                 err_estimate /= 2.0f;
             }
-            if (type_dst == GGML_TYPE_Q8_0) {
+            if (type == GGML_TYPE_Q8_0) {
                 err_estimate /= 8.0f;
             }
             err_estimate *= err_estimate;
@@ -2473,7 +2471,7 @@ struct test_set_rows : public test_case {
     // See dicussion here: https://github.com/ggml-org/llama.cpp/pull/23760#issuecomment-4566312209
     double max_nmse_err(ggml_backend_t backend) override {
         ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
-        if (type_dst == GGML_TYPE_Q8_0 && strcmp(ggml_backend_reg_name(reg), "WebGPU") == 0) {
+        if (type == GGML_TYPE_Q8_0 && strcmp(ggml_backend_reg_name(reg), "WebGPU") == 0) {
             return std::max(test_case::max_nmse_err(backend), 2e-7);
         }
         return test_case::max_nmse_err(backend);
@@ -7875,28 +7873,24 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_get_rows_back(GGML_TYPE_I32, 256, 5, 4, 1, v));
     }
 
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_Q8_0, GGML_TYPE_I32, { 256, 5, 1, 3 }, { 1, 1, }, 1, false));
+    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
+    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
+    test_cases.emplace_back(new test_set_rows(GGML_TYPE_Q8_0, GGML_TYPE_I32, { 256, 5, 1, 3 }, { 1, 1, }, 1, false));
     for (ggml_type type : all_types) {
         for (int b : {1, 7}) {
             for (bool v : {false, true}) {
-                test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 5,  b, 3 }, { 1, 1, }, 1, v));
-                test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v));
+                test_cases.emplace_back(new test_set_rows(type, GGML_TYPE_I64, { 256, 5,  b, 3 }, { 1, 1, }, 1, v));
+                test_cases.emplace_back(new test_set_rows(type, GGML_TYPE_I64, { 256, 11, 1, b }, { 2, 3, }, 7, v));
 
-                test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v));
+                test_cases.emplace_back(new test_set_rows(type, GGML_TYPE_I64, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v));
 
                 if (ggml_blck_size(type) == 1) {
-                    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v));
-                    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v));
+                    test_cases.emplace_back(new test_set_rows(type, GGML_TYPE_I64, { 31, 3, b, 1 }, { 2, 3, }, 2, v));
+                    test_cases.emplace_back(new test_set_rows(type, GGML_TYPE_I64, { 33, 5, 1, b }, { 2, 3, }, 1, v));
                 }
             }
         }
     }
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F16, GGML_TYPE_F16, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F16, GGML_TYPE_F16, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F16, GGML_TYPE_F16, GGML_TYPE_I64, { 1, 8, 1, 3 }, { 1, 1 }, 2, true));
-    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F16, GGML_TYPE_F16, GGML_TYPE_I32, { 1, 8, 1, 3 }, { 1, 1 }, 2, true));
 
     for (int mode : { GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_VISION }) {
         for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {

tools/cli/CMakeLists.txt

@@ -2,11 +2,13 @@
 
 set(TARGET llama-cli-impl)
 
-add_library(${TARGET} cli.cpp)
+add_library(${TARGET} cli.cpp
+                      cli-client.cpp
+                      cli-context.cpp)
 set_target_properties(${TARGET} PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 
 target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR} ../server)
-target_link_libraries(${TARGET} PUBLIC server-context llama-common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PUBLIC llama-server-impl llama-common ${CMAKE_THREAD_LIBS_INIT})
 
 if(LLAMA_TOOLS_INSTALL)
     install(TARGETS ${TARGET} LIBRARY)

tools/cli/cli-client.cpp

@@ -0,0 +1,130 @@
+#include "cli-client.h"
+
+#include "http.h"
+
+#include <algorithm>
+#include <chrono>
+#include <thread>
+
+// generation can stall for a long time during prompt processing, so the
+// read timeout must be generous
+static constexpr time_t CLI_HTTP_READ_TIMEOUT_SEC = 3600;
+
+// upper bound for the accumulated response body kept for error reporting
+static constexpr size_t CLI_HTTP_MAX_ERROR_BODY = 1024 * 1024;
+
+// returns the path with the base url's path prefix prepended (if any)
+static std::string join_path(const common_http_url & parts, const std::string & path) {
+    if (parts.path.empty() || parts.path == "/") {
+        return path;
+    }
+    std::string prefix = parts.path;
+    if (prefix.back() == '/') {
+        prefix.pop_back();
+    }
+    return prefix + path;
+}
+
+std::string cli_client::get(const std::string & path) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+    auto path_with_model = path + (model.empty() ? "" : ("?model=" + model));
+    auto res = cli.Get(join_path(parts, path_with_model));
+    if (!res) {
+        throw std::runtime_error("failed to connect to " + server_base + ": " + httplib::to_string(res.error()));
+    }
+    if (res->status < 200 || res->status >= 300) {
+        throw std::runtime_error("GET " + path + " failed with status " + std::to_string(res->status) + ": " + res->body);
+    }
+    return res->body;
+}
+
+std::string cli_client::post(const std::string & path, const std::string & body) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+    auto res = cli.Post(join_path(parts, path), body, "application/json");
+    if (!res) {
+        throw std::runtime_error("failed to connect to " + server_base + ": " + httplib::to_string(res.error()));
+    }
+    if (res->status < 200 || res->status >= 300) {
+        throw std::runtime_error("POST " + path + " failed with status " + std::to_string(res->status) + ": " + res->body);
+    }
+    return res->body;
+}
+
+std::string cli_client::post_sse(const std::string & path,
+                                  const std::string & body,
+                                  const std::function<bool()> & should_stop,
+                                  const std::function<void(const std::string &)> & on_data) {
+    auto [cli, parts] = common_http_client(server_base);
+    cli.set_read_timeout(CLI_HTTP_READ_TIMEOUT_SEC, 0);
+
+    std::string pending;  // buffer for incomplete SSE lines
+    std::string raw_body; // accumulated body, used only for error reporting
+
+    auto receiver = [&](const char * data, size_t len) -> bool {
+        if (should_stop()) {
+            return false; // aborts the request
+        }
+        if (raw_body.size() < CLI_HTTP_MAX_ERROR_BODY) {
+            raw_body.append(data, std::min(len, CLI_HTTP_MAX_ERROR_BODY - raw_body.size()));
+        }
+        pending.append(data, len);
+        size_t pos;
+        while ((pos = pending.find('\n')) != std::string::npos) {
+            std::string line = pending.substr(0, pos);
+            pending.erase(0, pos + 1);
+            if (!line.empty() && line.back() == '\r') {
+                line.pop_back();
+            }
+            if (line.rfind("data: ", 0) != 0) {
+                continue;
+            }
+            std::string payload = line.substr(6);
+            if (payload == "[DONE]") {
+                continue;
+            }
+            on_data(payload);
+        }
+        return true;
+    };
+
+    httplib::Headers headers = {{"Accept", "text/event-stream"}};
+    auto res = cli.Post(join_path(parts, path), headers, body, "application/json", receiver);
+
+    if (!res) {
+        if (res.error() == httplib::Error::Canceled && should_stop()) {
+            return ""; // cancelled by the user
+        }
+        return "failed to connect to " + server_base + ": " + httplib::to_string(res.error());
+    }
+    if (res->status < 200 || res->status >= 300) {
+        if (!raw_body.empty()) {
+            return raw_body;
+        }
+        return "request failed with status " + std::to_string(res->status);
+    }
+    return "";
+}
+
+bool cli_client::wait_health(const std::function<bool()> & is_aborted) {
+    int connect_attempts = 0;
+    while (!is_aborted()) {
+        auto [cli, parts] = common_http_client(server_base);
+        cli.set_connection_timeout(1, 0);
+        auto res = cli.Get(join_path(parts, "/health"));
+        if (res) {
+            if (res->status == 200) {
+                return true;
+            }
+            // any other status means the server is up but not ready yet
+            // (e.g. 503 while the model is still loading)
+        } else if (++connect_attempts >= 10) {
+            last_error = "failed to connect to " + server_base + ": " + httplib::to_string(res.error());
+            return false;
+        }
+        std::this_thread::sleep_for(std::chrono::milliseconds(300));
+    }
+    last_error = "aborted while waiting for the server to become ready";
+    return false;
+}

tools/cli/cli-client.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include <functional>
+#include <string>
+
+// openai-like client for CLI
+struct cli_client {
+    std::string server_base; // base url, for example "http://127.0.0.1:8080"
+    std::string last_error;  // set when wait_health() fails
+
+    std::string model; // optional, set when the server has multiple models (router mode)
+
+    // simple GET request, returns the raw response body
+    // throws std::runtime_error on transport error or non-2xx status
+    std::string get(const std::string & path);
+
+    // simple POST request, returns the raw response body
+    // throws std::runtime_error on transport error or non-2xx status
+    std::string post(const std::string & path, const std::string & body);
+
+    // POST request with an SSE streaming response
+    // on_data is invoked per "data:" event with the raw event payload
+    // returns after the stream is finished (empty string on graceful exit)
+    // otherwise, the raw error response body
+    std::string post_sse(const std::string & path,
+                          const std::string & body,
+                          const std::function<bool()> & should_stop,
+                          const std::function<void(const std::string &)> & on_data);
+
+    // poll /health until the server is ready to accept requests
+    // returns false if is_aborted returned true or the server is unreachable
+    bool wait_health(const std::function<bool()> & is_aborted);
+};

tools/cli/cli-context.cpp

@@ -0,0 +1,622 @@
+#include "cli-context.h"
+#include "cli-ui.h"
+
+#include "arg.h"
+#include "base64.hpp"
+#include "log.h"
+#include "console.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <algorithm>
+#include <cctype>
+#include <filesystem>
+#include <fstream>
+#include <map>
+#include <set>
+
+using json = nlohmann::ordered_json;
+
+struct cli_context_impl {
+    json messages      = json::array();
+    json pending_media = json::array(); // staged multimodal content parts
+};
+
+cli_context::cli_context(const common_params & params) : params(params), impl(new cli_context_impl()) {}
+
+cli_context::~cli_context() {
+    shutdown();
+}
+
+std::atomic<bool> & cli_context::interrupted() {
+    static std::atomic<bool> flag = false;
+    return flag;
+}
+
+static bool should_stop() {
+    return cli_context::interrupted().load();
+}
+
+static constexpr size_t FILE_GLOB_MAX_RESULTS = 100;
+
+const char * LLAMA_ASCII_LOGO = R"(
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+)";
+
+// number of values an arg consumes on the command line
+static int arg_num_values(const common_arg & opt) {
+    if (opt.value_hint_2 != nullptr) {
+        return 2;
+    }
+    if (opt.value_hint != nullptr) {
+        return 1;
+    }
+    return 0;
+}
+
+static std::string format_error_message(const json & err) {
+    if (err.contains("error") && err.at("error").is_object()) {
+        const auto & e = err.at("error");
+        if (e.contains("message") && e.at("message").is_string()) {
+            return e.at("message").get<std::string>();
+        }
+    }
+    return err.dump();
+}
+
+// err is the raw response body of a failed request; it may or may not be JSON
+static std::string format_error_message(const std::string & err) {
+    json parsed = json::parse(err, nullptr, false);
+    if (!parsed.is_discarded()) {
+        return format_error_message(parsed);
+    }
+    return err;
+}
+
+static std::string media_type_from_ext(const std::string & fname) {
+    std::string ext = std::filesystem::path(fname).extension().string();
+    std::transform(ext.begin(), ext.end(), ext.begin(), [](unsigned char c) { return std::tolower(c); });
+    if (ext == ".wav" || ext == ".mp3") {
+        return "audio";
+    }
+    if (ext == ".mp4" || ext == ".avi" || ext == ".mkv" || ext == ".mov" || ext == ".webm") {
+        return "video";
+    }
+    return "image";
+}
+
+bool cli_context::init() {
+    ui::init(params);
+
+    std::optional<ui::spinner> spinner;
+
+    bool use_external_server = !params.server_base.empty();
+    if (use_external_server) {
+        std::string base = params.server_base;
+        while (!base.empty() && base.back() == '/') {
+            base.pop_back();
+        }
+        client.server_base = base;
+
+        spinner.emplace("Connecting to server at " + base);
+    } else {
+        if (params.model.path.empty() && params.model.url.empty() &&
+                params.model.hf_repo.empty() && params.model.docker_repo.empty()) {
+            ui::show_error(
+                "no model specified",
+                "use -m <file.gguf> or -hf <user/repo> to run a local model,\n"
+                "or --server-base <url> to connect to a running llama-server"
+            );
+            return false;
+        }
+
+        spinner.emplace("\n\nLoading model...");
+
+        server.emplace();
+        if (!server->start(params)) {
+            ui::show_error("server start failed");
+            return false;
+        }
+        if (!server->wait_ready(should_stop)) {
+            if (!should_stop()) {
+                ui::show_error("the server exited before becoming ready");
+            }
+            return false;
+        }
+        client.server_base = server->address();
+    }
+
+    // for --server-base this is the main availability check; for a spawned
+    // server it is a cheap sanity check on top of the ready signal
+    auto is_aborted = [this]() {
+        return should_stop() || (server && !server->alive());
+    };
+    bool healthy = false;
+    try {
+        healthy = client.wait_health(is_aborted);
+    } catch (const std::exception & e) {
+        client.last_error = e.what();
+    }
+    if (!healthy) {
+        if (!should_stop()) {
+            ui::show_error(client.last_error);
+        }
+        return false;
+    }
+
+    if (use_external_server) {
+        spinner.reset();
+        if (!list_and_ask_models()) {
+            return false;
+        }
+        // restore the spinner for the next step
+        spinner.emplace("Waiting for server...");
+    }
+
+    fetch_server_props();
+
+    return true;
+}
+
+void cli_context::fetch_server_props() {
+    try {
+        json props = json::parse(client.get("/props"));
+        model_name = props.value("model_alias", "");
+        if (model_name.empty()) {
+            const std::string path = props.value("model_path", "");
+            if (!path.empty()) {
+                model_name = std::filesystem::path(path).filename().string();
+            }
+        }
+        model_ftype = props.value("model_ftype", "");
+        build_info = props.value("build_info", "");
+        if (props.contains("modalities") && props.at("modalities").is_object()) {
+            const auto & modalities = props.at("modalities");
+            has_vision = modalities.value("vision", false);
+            has_audio  = modalities.value("audio", false);
+            has_video  = modalities.value("video", false);
+        }
+    } catch (const std::exception & e) {
+        // /props can be disabled on remote servers; not fatal
+        LOG_DBG("failed to fetch /props: %s\n", e.what());
+    }
+}
+
+bool cli_context::list_and_ask_models() {
+    json resp = json::parse(client.get("/v1/models"));
+    if (!resp.contains("data") || !resp.at("data").is_array()) {
+        throw std::runtime_error("invalid response from /v1/models");
+    }
+    std::vector<std::string> models;
+    std::vector<std::string> models_display;
+    for (const auto & m : resp.at("data")) {
+        if (!m.contains("id") || !m.at("id").is_string()) {
+            continue;
+        }
+        std::string name = m.at("id").get<std::string>();
+        std::string display = name;
+        if (m.contains("aliases") && m.at("aliases").is_array()) {
+            std::vector<std::string> aliases;
+            for (const auto & a : m.at("aliases")) {
+                if (a.is_string()) {
+                    aliases.push_back(a.get<std::string>());
+                }
+            }
+            if (!aliases.empty()) {
+                display += " (" + string_join(aliases, ", ") + ")";
+            }
+        }
+        models.push_back(name);
+        models_display.push_back(display);
+    }
+
+    // only one model: use it without asking
+    if (models.size() == 1) {
+        model_name = models[0];
+        client.model = model_name;
+        return true;
+    }
+
+    std::string message = "\nAvailable models:";
+    for (size_t i = 0; i < models_display.size(); ++i) {
+        message += "\n  " + std::to_string(i + 1) + ". " + models_display[i];
+    }
+    message += "\n";
+    ui::show_message(message);
+    std::string selection;
+    while (selection.empty()) {
+        if (should_stop()) {
+            return false;
+        }
+        ui::user_turn user_turn;
+        selection = user_turn.read_input(false, "Select model by number: ");
+        if (selection.empty()) {
+            continue;
+        }
+        try {
+            size_t idx = std::stoul(selection);
+            if (idx > 0 && idx <= models.size()) {
+                model_name = models[idx - 1];
+                client.model = model_name;
+                ui::show_message("Selected model: " + model_name);
+                break;
+            }
+        } catch (...) {
+            // ignore
+        }
+        ui::show_error("Invalid selection. Please enter a valid number.");
+        selection.clear();
+        continue;
+    }
+    return true;
+}
+
+void cli_context::add_system_prompt() {
+    if (!params.system_prompt.empty()) {
+        impl->messages.push_back({
+            {"role",    "system"},
+            {"content", params.system_prompt}
+        });
+    }
+}
+
+void cli_context::push_user_message(const std::string & text) {
+    json content;
+    if (impl->pending_media.empty()) {
+        content = text;
+    } else {
+        // multimodal message: media parts first, then the text
+        content = impl->pending_media;
+        content.push_back({
+            {"type", "text"},
+            {"text", text}
+        });
+        impl->pending_media = json::array();
+    }
+    impl->messages.push_back({
+        {"role",    "user"},
+        {"content", content}
+    });
+}
+
+bool cli_context::stage_media_file(const std::string & fname, const std::string & type) {
+    std::ifstream file(fname, std::ios::binary);
+    if (!file) {
+        return false;
+    }
+    std::string data((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    std::string encoded = base64::encode(data);
+
+    if (type == "audio") {
+        std::string ext = std::filesystem::path(fname).extension().string();
+        std::transform(ext.begin(), ext.end(), ext.begin(), [](unsigned char c) { return std::tolower(c); });
+        impl->pending_media.push_back({
+            {"type", "input_audio"},
+            {"input_audio", {
+                {"data",   encoded},
+                {"format", ext == ".mp3" ? "mp3" : "wav"}
+            }}
+        });
+    } else if (type == "video") {
+        impl->pending_media.push_back({
+            {"type", "input_video"},
+            {"input_video", {
+                {"data", encoded}
+            }}
+        });
+    } else {
+        // the server detects the actual image type from the data
+        impl->pending_media.push_back({
+            {"type", "image_url"},
+            {"image_url", {
+                {"url", "data:image/unknown;base64," + encoded}
+            }}
+        });
+    }
+    return true;
+}
+
+bool cli_context::generate_completion(std::string & assistant_content, cli_timings & timings) {
+    json body = {
+        {"messages",          impl->messages},
+        {"stream",            true},
+        // in order to get timings even when we cancel mid-way
+        {"timings_per_token", true},
+    };
+    if (!client.model.empty()) {
+        body["model"] = client.model;
+    }
+
+    bool stream_error = false;
+
+    ui::assistant_turn a;
+
+    std::string err = client.post_sse("/v1/chat/completions", body.dump(), should_stop, [&](const std::string & payload) {
+        json chunk = json::parse(payload, nullptr, false);
+        if (chunk.is_discarded()) {
+            return;
+        }
+        if (chunk.contains("error")) {
+            stream_error = true;
+            ui::show_error(format_error_message(chunk));
+            return;
+        }
+        if (chunk.contains("timings")) {
+            const auto & t = chunk.at("timings");
+            timings.prompt_per_second    = t.value("prompt_per_second",    0.0);
+            timings.predicted_per_second = t.value("predicted_per_second", 0.0);
+        }
+        if (!chunk.contains("choices") || !chunk.at("choices").is_array() || chunk.at("choices").empty()) {
+            return;
+        }
+        const auto & choice = chunk.at("choices").at(0);
+        if (!choice.contains("delta")) {
+            return;
+        }
+        const auto & delta = choice.at("delta");
+        if (delta.contains("reasoning_content") && delta.at("reasoning_content").is_string()) {
+            const std::string text = delta.at("reasoning_content").get<std::string>();
+            if (!text.empty()) {
+                a.push(ui::ASSISTANT_DISPLAY_MODE_REASONING, text);
+            }
+        }
+        if (delta.contains("content") && delta.at("content").is_string()) {
+            const std::string text = delta.at("content").get<std::string>();
+            if (!text.empty()) {
+                assistant_content += text;
+                a.push(ui::ASSISTANT_DISPLAY_MODE_CONTENT, text);
+            }
+        }
+    });
+
+    cli_context::interrupted().store(false);
+
+    if (!err.empty()) {
+        ui::show_error(format_error_message(err));
+        return false;
+    }
+    return !stream_error;
+}
+
+int cli_context::run() {
+    add_system_prompt();
+
+    std::string modalities = "text";
+    if (has_vision) {
+        modalities += ", vision";
+    }
+    if (has_audio) {
+        modalities += ", audio";
+    }
+    if (has_video) {
+        modalities += ", video";
+    }
+
+    std::string banner;
+    banner += "\n";
+    banner += LLAMA_ASCII_LOGO;
+    banner += "\n";
+    banner += "build      : " + build_info + "\n";
+    banner += "model      : " + model_name + "\n";
+    if (!model_ftype.empty()) {
+        banner += "ftype      : " + model_ftype + "\n";
+    }
+    banner += "modalities : " + modalities + "\n";
+    if (!params.system_prompt.empty()) {
+        banner += "using custom system prompt\n";
+    }
+    banner += "\n";
+    banner += "available commands:\n";
+    banner += "  /exit or Ctrl+C     stop or exit\n";
+    banner += "  /regen              regenerate the last response\n";
+    banner += "  /clear              clear the chat history\n";
+    banner += "  /read <file>        add a text file\n";
+    banner += "  /glob <pattern>     add text files using globbing pattern\n";
+    if (has_vision) {
+        banner += "  /image <file>       add an image file\n";
+    }
+    if (has_audio) {
+        banner += "  /audio <file>       add an audio file\n";
+    }
+    if (has_video) {
+        banner += "  /video <file>       add a video file\n";
+    }
+    banner += "\n";
+
+    ui::show_message(banner);
+
+    // interactive loop
+    std::string cur_msg;
+
+    auto add_text_file = [&](const std::string & fname) -> bool {
+        std::ifstream file(fname, std::ios::binary);
+        if (!file) {
+            ui::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+            return false;
+        }
+        std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+        cur_msg += "--- File: ";
+        cur_msg += fname;
+        cur_msg += " ---\n";
+        cur_msg += content;
+        ui::show_message(string_format("Loaded text from '%s'", fname.c_str()));
+        return true;
+    };
+
+    while (true) {
+        std::string buffer;
+        {
+            ui::user_turn user_turn;
+
+            if (params.prompt.empty()) {
+                buffer = user_turn.read_input(params.multiline_input);
+            } else {
+                // process input prompt from args
+                for (auto & fname : params.image) {
+                    if (!stage_media_file(fname, media_type_from_ext(fname))) {
+                        ui::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+                        break;
+                    }
+                    ui::show_message(string_format("Loaded media from '%s'", fname.c_str()));
+                }
+                buffer = params.prompt;
+                user_turn.echo(buffer);
+                params.prompt.clear(); // only use it once
+            }
+        }
+
+        if (should_stop()) {
+            cli_context::interrupted().store(false);
+            break;
+        }
+
+        // remove trailing newline
+        if (!buffer.empty() && buffer.back() == '\n') {
+            buffer.pop_back();
+        }
+
+        // skip empty messages
+        if (buffer.empty()) {
+            continue;
+        }
+
+        bool add_user_msg = true;
+
+        // process commands
+        if (string_starts_with(buffer, "/exit")) {
+            break;
+        } else if (string_starts_with(buffer, "/regen")) {
+            if (impl->messages.size() >= 2) {
+                size_t last_idx = impl->messages.size() - 1;
+                impl->messages.erase(last_idx);
+                add_user_msg = false;
+            } else {
+                ui::show_error("No message to regenerate.");
+                continue;
+            }
+        } else if (string_starts_with(buffer, "/clear")) {
+            impl->messages.clear();
+            add_system_prompt();
+
+            impl->pending_media = json::array();
+            ui::show_message("Chat history cleared.");
+            continue;
+        } else if (
+                (string_starts_with(buffer, "/image ") && has_vision) ||
+                (string_starts_with(buffer, "/audio ") && has_audio) ||
+                (string_starts_with(buffer, "/video ") && has_video)) {
+            std::string type = buffer.substr(1, 5);
+            // just in case (bad copy-paste for example), we strip all trailing/leading spaces
+            std::string fname = string_strip(buffer.substr(7));
+            if (!stage_media_file(fname, type)) {
+                ui::show_error(string_format("file does not exist or cannot be opened: '%s'", fname.c_str()));
+                continue;
+            }
+            ui::show_message(string_format("Loaded media from '%s'", fname.c_str()));
+            continue;
+        } else if (string_starts_with(buffer, "/read ")) {
+            std::string fname = string_strip(buffer.substr(6));
+            add_text_file(fname);
+            continue;
+        } else if (string_starts_with(buffer, "/glob ")) {
+            std::error_code ec;
+            size_t count = 0;
+            auto curdir = std::filesystem::current_path();
+            std::string pattern = string_strip(buffer.substr(6));
+            std::filesystem::path rel_path;
+
+            auto startglob = pattern.find_first_of("![*?");
+            if (startglob != std::string::npos && startglob != 0) {
+                auto endpath = pattern.substr(0, startglob).find_last_of('/');
+                if (endpath != std::string::npos) {
+                    std::string rel_pattern = pattern.substr(0, endpath);
+#if !defined(_WIN32)
+                    if (string_starts_with(rel_pattern, '~')) {
+                        const char * home = std::getenv("HOME");
+                        if (home && home[0]) {
+                            rel_pattern = home + rel_pattern.substr(1);
+                        }
+                    }
+#endif
+                    rel_path = rel_pattern;
+                    pattern.erase(0, endpath + 1);
+                    curdir /= rel_path;
+                }
+            }
+
+            for (const auto & entry : std::filesystem::recursive_directory_iterator(curdir,
+                    std::filesystem::directory_options::skip_permission_denied, ec)) {
+                if (!entry.is_regular_file()) {
+                    continue;
+                }
+
+                std::string rel = std::filesystem::relative(entry.path(), curdir, ec).string();
+                if (ec) {
+                    ec.clear();
+                    continue;
+                }
+                std::replace(rel.begin(), rel.end(), '\\', '/');
+
+                if (!glob_match(pattern, rel)) {
+                    continue;
+                }
+
+                if (!add_text_file((rel_path / rel).string())) {
+                    continue;
+                }
+
+                if (++count >= FILE_GLOB_MAX_RESULTS) {
+                    ui::show_error(string_format("Maximum number of globbed files allowed (%zu) reached.", FILE_GLOB_MAX_RESULTS));
+                    break;
+                }
+            }
+            continue;
+        } else {
+            // not a command
+            cur_msg += buffer;
+        }
+
+        // generate response
+        if (add_user_msg) {
+            push_user_message(cur_msg);
+            cur_msg.clear();
+        }
+        cli_timings timings;
+        std::string assistant_content;
+        generate_completion(assistant_content, timings);
+        impl->messages.push_back({
+            {"role",    "assistant"},
+            {"content", assistant_content}
+        });
+
+        if (params.show_timings) {
+            ui::show_info(string_format(
+                "\n[ Prompt: %.1f t/s | Generation: %.1f t/s ]",
+                timings.prompt_per_second,
+                timings.predicted_per_second
+            ));
+        }
+
+        if (params.single_turn) {
+            break;
+        }
+    }
+
+    ui::show_message("\n\nExiting...");
+
+    return 0;
+}
+
+void cli_context::shutdown() {
+    if (server) {
+        server->stop();
+        server.reset();
+    }
+}

tools/cli/cli-context.h

@@ -0,0 +1,66 @@
+#pragma once
+
+#include "common.h"
+
+#include "cli-client.h"
+#include "cli-server.h"
+
+#include <atomic>
+#include <memory>
+#include <optional>
+#include <string>
+
+struct cli_timings {
+    double prompt_per_second    = 0.0;
+    double predicted_per_second = 0.0;
+};
+
+struct cli_context_impl;
+
+struct cli_context {
+    common_params params;
+
+    cli_client client;                // always initialized
+    std::optional<cli_server> server; // only set when no --server-base is given
+
+    // properties of the connected server
+    // will be populated by fetch_server_props()
+    std::string model_name;
+    std::string model_ftype;
+    std::string build_info;
+    bool has_vision = false;
+    bool has_audio  = false;
+    bool has_video  = false;
+
+    cli_context(const common_params & params);
+    ~cli_context();
+
+    // connect to --server-base or spawn a local llama-server child;
+    // argc/argv are needed to forward the server-relevant args to the child
+    bool init();
+
+    // run the interactive chat loop, returns the process exit code
+    int run();
+
+    // stop the local server child (if any)
+    void shutdown();
+
+    // set by the SIGINT handler; cleared once the interrupt has been handled
+    static std::atomic<bool> & interrupted();
+
+private:
+    bool generate_completion(std::string & assistant_content, cli_timings & timings);
+    void fetch_server_props();
+    void add_system_prompt();
+    void push_user_message(const std::string & text);
+
+    // check if server have multiple models (router mode)
+    // if yes, list them then ask; do nothing otherwise
+    bool list_and_ask_models();
+
+    // read a file and stage it as a multimodal content part; type is one of
+    // "image", "audio", "video"; returns false if the file cannot be read
+    bool stage_media_file(const std::string & fname, const std::string & type);
+
+    std::unique_ptr<cli_context_impl> impl;
+};

tools/cli/cli-server.h

@@ -0,0 +1,89 @@
+#pragma once
+
+#include <thread>
+
+#include "http.h"
+
+// llama_server will be available as a dynamic library symbol
+int llama_server(common_params & params, int argc, char ** argv);
+void llama_server_terminate();
+
+struct cli_server {
+    std::thread th;
+    int port = -1;
+    std::atomic<bool> is_alive = false;
+    std::atomic<bool> is_stopping = false;
+
+    ~cli_server() {
+        stop();
+    }
+
+    void stop() {
+        if (is_stopping.exchange(true)) {
+            return;
+        }
+        if (alive()) {
+            llama_server_terminate();
+        }
+        if (th.joinable()) {
+            th.join();
+        }
+    }
+
+    // spawn llama-server in a thread and interact with it via a random port
+    bool start(common_params & params) {
+        port = common_http_get_free_port();
+        if (port <= 0) {
+            fprintf(stderr, "failed to get a free port\n");
+            exit(1);
+        }
+
+        is_alive.store(true, std::memory_order_release);
+
+        common_params server_params = params; // copy
+        server_params.port = port;
+
+        th = std::thread([this, server_params]() mutable {
+            // argc / argv are only used in router mode, we can skip them for now
+            int res = llama_server(server_params, 0, nullptr);
+            if (res != 0) {
+                fprintf(stderr, "llama_server exited with code %d\n", res);
+            }
+            is_alive.store(false, std::memory_order_release);
+        });
+
+        return true;
+    }
+
+    std::string address() const {
+        return "http://127.0.0.1:" + std::to_string(port);
+    }
+
+    bool wait_ready(std::function<bool()> should_stop) {
+        if (!alive()) {
+            return false;
+        }
+        while (!should_stop()) {
+            auto [cli, parts] = common_http_client(address());
+            cli.set_connection_timeout(1, 0);
+            auto res = cli.Get("/health");
+            if (res) {
+                if (res->status == 200) {
+                    return true;
+                }
+                // any other status means the server is up but not ready yet
+                // (e.g. 503 while the model is still loading)
+            }
+            if (!alive()) {
+                // in case server die permanently
+                return false;
+            }
+            std::this_thread::sleep_for(std::chrono::milliseconds(200));
+        }
+        return true;
+    }
+
+    bool alive() const {
+        return is_alive.load(std::memory_order_acquire);
+    }
+};

tools/cli/cli-ui.h

@@ -0,0 +1,251 @@
+#pragma once
+
+#include "common.h"
+#include "console.h"
+
+#include <array>
+#include <algorithm>
+#include <cctype>
+#include <filesystem>
+#include <string_view>
+
+// TODO?: Make this reusable, enums, docs
+static const std::array<std::string_view, 8> cmds = {
+    "/audio ",
+    "/clear",
+    "/exit",
+    "/glob ",
+    "/image ",
+    "/read ",
+    "/regen",
+    "/video ",
+};
+
+static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std::string_view line, size_t cursor_byte_pos) {
+    std::vector<std::pair<std::string, size_t>> matches;
+    std::string cmd;
+
+    if (line.length() > 1 && line.front() == '/' && !std::any_of(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
+        return string_starts_with(line, prefix);
+    })) {
+        auto it = cmds.begin();
+
+        while ((it = std::find_if(it, cmds.end(), [line](std::string_view cmd_line) {
+            return string_starts_with(cmd_line, line);
+        })) != cmds.end()) {
+            matches.emplace_back(*it, it->length());
+            ++it;
+        }
+    } else {
+        auto it = std::find_if(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
+            return prefix.back() == ' ' && string_starts_with(line, prefix);
+        });
+
+        if (it != cmds.end()) {
+            cmd = *it;
+        }
+    }
+
+    if (!cmd.empty() && cmd != "/glob " && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
+        const std::string path_prefix  = std::string(line.substr(cmd.length(), cursor_byte_pos - cmd.length()));
+        const std::string path_postfix = std::string(line.substr(cursor_byte_pos));
+        auto cur_dir = std::filesystem::current_path();
+        std::string cur_dir_str = cur_dir.string();
+        std::string expanded_prefix = path_prefix;
+
+#if !defined(_WIN32)
+        if (string_starts_with(path_prefix, '~')) {
+            const char * home = std::getenv("HOME");
+            if (home && home[0]) {
+                expanded_prefix = home + path_prefix.substr(1);
+            }
+        }
+        if (string_starts_with(expanded_prefix, '/')) {
+#else
+        if (std::isalpha(static_cast<unsigned char>(expanded_prefix[0])) && expanded_prefix.find(':') == 1) {
+#endif
+            cur_dir = std::filesystem::path(expanded_prefix).parent_path();
+            cur_dir_str.clear();
+        } else if (!path_prefix.empty()) {
+            cur_dir /= std::filesystem::path(path_prefix).parent_path();
+        }
+
+        std::error_code ec;
+        for (const auto & entry : std::filesystem::directory_iterator(cur_dir, ec)) {
+            if (ec) {
+                break;
+            }
+            if (!entry.exists(ec)) {
+                ec.clear();
+                continue;
+            }
+
+            const std::string path_full = entry.path().string();
+            std::string path_entry = !cur_dir_str.empty() && string_starts_with(path_full, cur_dir_str) ? path_full.substr(cur_dir_str.length() + 1) : path_full;
+
+            if (entry.is_directory(ec)) {
+                path_entry.push_back(std::filesystem::path::preferred_separator);
+            }
+
+            if (expanded_prefix.empty() || string_starts_with(path_entry, expanded_prefix)) {
+                const std::string updated_line = cmd + path_entry;
+                matches.emplace_back(updated_line + path_postfix, updated_line.length());
+            }
+
+            if (ec) {
+                ec.clear();
+            }
+        }
+
+        if (matches.empty()) {
+            const std::string updated_line = cmd + path_prefix;
+            matches.emplace_back(updated_line + path_postfix, updated_line.length());
+        }
+
+        // Add the longest common prefix
+        if (!expanded_prefix.empty() && matches.size() > 1) {
+            const std::string_view match0(matches[0].first);
+            const std::string_view match1(matches[1].first);
+            auto it = std::mismatch(match0.begin(), match0.end(), match1.begin(), match1.end());
+            size_t len = it.first - match0.begin();
+
+            for (size_t i = 2; i < matches.size(); ++i) {
+                const std::string_view matchi(matches[i].first);
+                auto cmp = std::mismatch(match0.begin(), match0.end(), matchi.begin(), matchi.end());
+                len = std::min(len, static_cast<size_t>(cmp.first - match0.begin()));
+            }
+
+            const std::string updated_line = std::string(match0.substr(0, len));
+            matches.emplace_back(updated_line + path_postfix, updated_line.length());
+        }
+
+        std::sort(matches.begin(), matches.end(), [](const auto & a, const auto & b) {
+            return a.first.compare(0, a.second, b.first, 0, b.second) < 0;
+        });
+    }
+
+    return matches;
+}
+
+// note: make this view implementation generic, so that we can move to TUI in the future if we want to
+namespace ui {
+    static void init(const common_params & params) {
+        // TODO: avoid using atexit() here by making `console` a singleton
+        console::init(params.simple_io, params.use_color);
+        atexit([]() { console::cleanup(); });
+
+        console::set_completion_callback(auto_completion_callback);
+    }
+
+    struct spinner {
+        spinner(const std::string & message) {
+            if (!message.empty()) {
+                console::log("%s ", message.c_str());
+            }
+            console::spinner::start();
+        }
+        ~spinner() {
+            console::spinner::stop();
+        }
+    };
+
+    struct user_turn {
+        user_turn() {
+            console::set_display(DISPLAY_TYPE_USER_INPUT);
+        }
+        ~user_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+        void echo(const std::string & buffer) {
+            if (buffer.size() > 500) {
+                console::log("\n> %s ... (truncated)\n", buffer.substr(0, 500).c_str());
+            } else {
+                console::log("\n> %s\n", buffer.c_str());
+            }
+        }
+        std::string read_input(bool multiline_input, const char * prompt = nullptr) {
+            if (prompt) {
+                console::log("%s", prompt);
+            } else {
+                console::log("\n> ");
+            }
+            std::string buffer;
+            std::string line;
+            bool another_line = true;
+            do {
+                another_line = console::readline(line, multiline_input);
+                buffer += line;
+            } while (another_line);
+            return buffer;
+        }
+    };
+
+    enum assistant_display_mode {
+        ASSISTANT_DISPLAY_MODE_REASONING,
+        ASSISTANT_DISPLAY_MODE_CONTENT,
+    };
+    struct assistant_turn {
+        assistant_display_mode mode = ASSISTANT_DISPLAY_MODE_CONTENT;
+        bool trailing_newline = true;
+        bool is_inside_reasoning = false;
+        assistant_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+        ~assistant_turn() {
+            console::set_display(DISPLAY_TYPE_RESET);
+            add_newline_if_needed();
+        }
+        void push(assistant_display_mode m, const std::string & buffer) {
+            if (m != mode) {
+                add_newline_if_needed();
+                switch (m) {
+                    case ASSISTANT_DISPLAY_MODE_CONTENT:
+                        {
+                            if (is_inside_reasoning) {
+                                console::log("[End thinking]\n\n");
+                                is_inside_reasoning = false;
+                            }
+                            console::set_display(DISPLAY_TYPE_RESET);
+                        } break;
+                    case ASSISTANT_DISPLAY_MODE_REASONING:
+                        {
+                            console::set_display(DISPLAY_TYPE_REASONING);
+                            is_inside_reasoning = true;
+                            console::log("\n[Start thinking]\n\n");
+                        } break;
+                }
+            }
+            mode = m;
+            if (buffer.empty()) {
+                return;
+            }
+            trailing_newline = buffer.back() == '\n';
+            console::log("%s", buffer.c_str());
+            console::flush();
+        }
+        void add_newline_if_needed() {
+            if (!trailing_newline) {
+                console::log("\n");
+                console::flush();
+            }
+        }
+    };
+
+    static void show_error(const std::string & title, const std::string & message = "") {
+        console::spinner::stop();
+        console::error("Error: %s\n", title.c_str());
+        if (!message.empty()) {
+            console::log("%s\n", message.c_str());
+        }
+    }
+
+    static void show_message(const std::string & message) {
+        console::log("%s\n", message.c_str());
+    }
+
+    static void show_info(const std::string & message) {
+        console::set_display(DISPLAY_TYPE_INFO);
+        console::log("%s\n", message.c_str());
+        console::set_display(DISPLAY_TYPE_RESET);
+    }
+}

tools/cli/cli.cpp

@@ -1,20 +1,9 @@
-#include "chat.h"
-#include "common.h"
 #include "arg.h"
-#include "console.h"
-#include "fit.h"
-// #include "log.h"
+#include "common.h"
+#include "log.h"
 
-#include "server-common.h"
-#include "server-context.h"
-#include "server-task.h"
+#include "cli-context.h"
 
-#include <array>
-#include <atomic>
-#include <algorithm>
-#include <filesystem>
-#include <fstream>
-#include <thread>
 #include <signal.h>
 
 #if defined(_WIN32)
@@ -25,342 +14,19 @@
 #include <windows.h>
 #endif
 
-const char * LLAMA_ASCII_LOGO = R"(
-▄▄ ▄▄
-██ ██
-██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
-██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
-██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
-                                    ██    ██
-                                    ▀▀    ▀▀
-)";
-
-static std::atomic<bool> g_is_interrupted = false;
-static bool should_stop() {
-    return g_is_interrupted.load();
-}
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void signal_handler(int) {
-    if (g_is_interrupted.load()) {
+    if (cli_context::interrupted().load()) {
         // second Ctrl+C - exit immediately
         // make sure to clear colors before exiting (not using LOG or console.cpp here to avoid deadlock)
         fprintf(stdout, "\033[0m\n");
         fflush(stdout);
         std::exit(130);
     }
-    g_is_interrupted.store(true);
+    cli_context::interrupted().store(true);
 }
 #endif
 
-struct cli_context {
-    server_context ctx_server;
-    json messages = json::array();
-    std::vector<raw_buffer> input_files;
-    task_params defaults;
-    bool verbose_prompt;
-
-    // thread for showing "loading" animation
-    std::atomic<bool> loading_show;
-
-    cli_context(const common_params & params) {
-        defaults.sampling    = params.sampling;
-        defaults.speculative = params.speculative;
-        defaults.n_keep      = params.n_keep;
-        defaults.n_predict   = params.n_predict;
-        defaults.antiprompt  = params.antiprompt;
-
-        defaults.stream = true; // make sure we always use streaming mode
-        defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
-        // defaults.return_progress = true; // TODO: show progress
-
-        verbose_prompt = params.verbose_prompt;
-    }
-
-    std::string generate_completion(result_timings & out_timings) {
-        server_response_reader rd = ctx_server.get_response_reader();
-        auto chat_params = format_chat();
-        {
-            // TODO: reduce some copies here in the future
-            server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
-            task.id         = rd.get_new_id();
-            task.index      = 0;
-            task.params     = defaults;           // copy
-            task.cli_prompt = chat_params.prompt; // copy
-            task.cli_files  = input_files;        // copy
-            task.cli        = true;
-
-            // chat template settings
-            task.params.chat_parser_params = common_chat_parser_params(chat_params);
-            task.params.chat_parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
-            if (!chat_params.parser.empty()) {
-                task.params.chat_parser_params.parser.load(chat_params.parser);
-            }
-
-            // Copy the preserved tokens into the sampling params
-            const llama_vocab * vocab = llama_model_get_vocab(
-                llama_get_model(ctx_server.get_llama_context()));
-            for (const auto & token : chat_params.preserved_tokens) {
-                auto ids = common_tokenize(vocab, token, false, true);
-                if (ids.size() == 1) {
-                    task.params.sampling.preserved_tokens.insert(ids[0]);
-                }
-            }
-
-            // reasoning budget sampler
-            if (!chat_params.thinking_end_tag.empty()) {
-                task.params.sampling.reasoning_budget_tokens = defaults.sampling.reasoning_budget_tokens;
-                task.params.sampling.generation_prompt = chat_params.generation_prompt;
-
-                if (!chat_params.thinking_start_tag.empty()) {
-                    task.params.sampling.reasoning_budget_start =
-                        common_tokenize(vocab, chat_params.thinking_start_tag, false, true);
-                }
-                task.params.sampling.reasoning_budget_end =
-                    common_tokenize(vocab, chat_params.thinking_end_tag, false, true);
-                task.params.sampling.reasoning_budget_forced =
-                    common_tokenize(vocab, defaults.sampling.reasoning_budget_message + chat_params.thinking_end_tag, false, true);
-            }
-
-            rd.post_task({std::move(task)});
-        }
-
-        if (verbose_prompt) {
-            console::set_display(DISPLAY_TYPE_PROMPT);
-            console::log("%s\n\n", chat_params.prompt.c_str());
-            console::set_display(DISPLAY_TYPE_RESET);
-        }
-
-        // wait for first result
-        console::spinner::start();
-        server_task_result_ptr result = rd.next(should_stop);
-
-        while (true) {
-            auto res_partial = dynamic_cast<server_task_result_cmpl_partial *>(result.get());
-            if (res_partial && res_partial->is_begin) {
-                // this is the "send 200 status to client" signal in streaming mode
-                // skip, do not stop the spinner
-                result = rd.next(should_stop);
-            } else {
-                console::spinner::stop();
-                break;
-            }
-        }
-
-        std::string curr_content;
-        bool is_thinking = false;
-
-        while (result) {
-            if (should_stop()) {
-                break;
-            }
-            if (result->is_error()) {
-                json err_data = result->to_json();
-                if (err_data.contains("message")) {
-                    console::error("Error: %s\n", err_data["message"].get<std::string>().c_str());
-                } else {
-                    console::error("Error: %s\n", err_data.dump().c_str());
-                }
-                return curr_content;
-            }
-            auto res_partial = dynamic_cast<server_task_result_cmpl_partial *>(result.get());
-            if (res_partial) {
-                out_timings = std::move(res_partial->timings);
-                for (const auto & diff : res_partial->oaicompat_msg_diffs) {
-                    if (!diff.content_delta.empty()) {
-                        if (is_thinking) {
-                            console::log("\n[End thinking]\n\n");
-                            console::set_display(DISPLAY_TYPE_RESET);
-                            is_thinking = false;
-                        }
-                        curr_content += diff.content_delta;
-                        console::log("%s", diff.content_delta.c_str());
-                        console::flush();
-                    }
-                    if (!diff.reasoning_content_delta.empty()) {
-                        console::set_display(DISPLAY_TYPE_REASONING);
-                        if (!is_thinking) {
-                            console::log("[Start thinking]\n");
-                        }
-                        is_thinking = true;
-                        console::log("%s", diff.reasoning_content_delta.c_str());
-                        console::flush();
-                    }
-                }
-            }
-            auto res_final = dynamic_cast<server_task_result_cmpl_final *>(result.get());
-            if (res_final) {
-                out_timings = std::move(res_final->timings);
-                break;
-            }
-            result = rd.next(should_stop);
-        }
-        g_is_interrupted.store(false);
-        // server_response_reader automatically cancels pending tasks upon destruction
-        return curr_content;
-    }
-
-    // TODO: support remote files in the future (http, https, etc)
-    std::string load_input_file(const std::string & fname, bool is_media) {
-        std::ifstream file = fs_open_ifstream(fname, std::ios::binary);
-        if (!file) {
-            return "";
-        }
-        if (is_media) {
-            raw_buffer buf;
-            buf.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
-            input_files.push_back(std::move(buf));
-            return get_media_marker();
-        } else {
-            std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
-            return content;
-        }
-    }
-
-    common_chat_params format_chat() {
-        auto meta = ctx_server.get_meta();
-        auto & chat_params = meta.chat_params;
-
-        auto caps = common_chat_templates_get_caps(chat_params.tmpls.get());
-
-        common_chat_templates_inputs inputs;
-        inputs.messages              = common_chat_msgs_parse_oaicompat(messages);
-        inputs.tools                 = {}; // TODO
-        inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
-        inputs.json_schema           = ""; // TODO
-        inputs.grammar               = ""; // TODO
-        inputs.use_jinja             = chat_params.use_jinja;
-        inputs.parallel_tool_calls   = caps["supports_parallel_tool_calls"];
-        inputs.add_generation_prompt = true;
-        inputs.reasoning_format      = COMMON_REASONING_FORMAT_DEEPSEEK;
-        inputs.force_pure_content    = chat_params.force_pure_content;
-        inputs.enable_thinking       = chat_params.enable_thinking ? common_chat_templates_support_enable_thinking(chat_params.tmpls.get()) : false;
-
-        // Apply chat template to the list of messages
-        return common_chat_templates_apply(chat_params.tmpls.get(), inputs);
-    }
-};
-
-// TODO?: Make this reusable, enums, docs
-static const std::array<std::string_view, 8> cmds = {
-    "/audio ",
-    "/clear",
-    "/exit",
-    "/glob ",
-    "/image ",
-    "/read ",
-    "/regen",
-    "/video ",
-};
-
-static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std::string_view line, size_t cursor_byte_pos) {
-    std::vector<std::pair<std::string, size_t>> matches;
-    std::string cmd;
-
-    if (line.length() > 1 && line.front() == '/' && !std::any_of(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
-        return string_starts_with(line, prefix);
-    })) {
-        auto it = cmds.begin();
-
-        while ((it = std::find_if(it, cmds.end(), [line](std::string_view cmd_line) {
-            return string_starts_with(cmd_line, line);
-        })) != cmds.end()) {
-            matches.emplace_back(*it, it->length());
-            ++it;
-        }
-    } else {
-        auto it = std::find_if(cmds.begin(), cmds.end(), [line](std::string_view prefix) {
-            return prefix.back() == ' ' && string_starts_with(line, prefix);
-        });
-
-        if (it != cmds.end()) {
-            cmd = *it;
-        }
-    }
-
-    if (!cmd.empty() && cmd != "/glob " && line.length() >= cmd.length() && cursor_byte_pos >= cmd.length()) {
-        const std::string path_prefix  = std::string(line.substr(cmd.length(), cursor_byte_pos - cmd.length()));
-        const std::string path_postfix = std::string(line.substr(cursor_byte_pos));
-        auto cur_dir = std::filesystem::current_path();
-        std::string cur_dir_str = cur_dir.string();
-        std::string expanded_prefix = path_prefix;
-
-#if !defined(_WIN32)
-        if (string_starts_with(path_prefix, '~')) {
-            const char * home = std::getenv("HOME");
-            if (home && home[0]) {
-                expanded_prefix = home + path_prefix.substr(1);
-            }
-        }
-        if (string_starts_with(expanded_prefix, '/')) {
-#else
-        if (std::isalpha(expanded_prefix[0]) && expanded_prefix.find(':') == 1) {
-#endif
-            cur_dir = std::filesystem::path(expanded_prefix).parent_path();
-            cur_dir_str.clear();
-        } else if (!path_prefix.empty()) {
-            cur_dir /= std::filesystem::path(path_prefix).parent_path();
-        }
-
-        std::error_code ec;
-        for (const auto & entry : std::filesystem::directory_iterator(cur_dir, ec)) {
-            if (ec) {
-                break;
-            }
-            if (!entry.exists(ec)) {
-                ec.clear();
-                continue;
-            }
-
-            const std::string path_full = entry.path().string();
-            std::string path_entry = !cur_dir_str.empty() && string_starts_with(path_full, cur_dir_str) ? path_full.substr(cur_dir_str.length() + 1) : path_full;
-
-            if (entry.is_directory(ec)) {
-                path_entry.push_back(std::filesystem::path::preferred_separator);
-            }
-
-            if (expanded_prefix.empty() || string_starts_with(path_entry, expanded_prefix)) {
-                const std::string updated_line = cmd + path_entry;
-                matches.emplace_back(updated_line + path_postfix, updated_line.length());
-            }
-
-            if (ec) {
-                ec.clear();
-            }
-        }
-
-        if (matches.empty()) {
-            const std::string updated_line = cmd + path_prefix;
-            matches.emplace_back(updated_line + path_postfix, updated_line.length());
-        }
-
-        // Add the longest common prefix
-        if (!expanded_prefix.empty() && matches.size() > 1) {
-            const std::string_view match0(matches[0].first);
-            const std::string_view match1(matches[1].first);
-            auto it = std::mismatch(match0.begin(), match0.end(), match1.begin(), match1.end());
-            size_t len = it.first - match0.begin();
-
-            for (size_t i = 2; i < matches.size(); ++i) {
-                const std::string_view matchi(matches[i].first);
-                auto cmp = std::mismatch(match0.begin(), match0.end(), matchi.begin(), matchi.end());
-                len = std::min(len, static_cast<size_t>(cmp.first - match0.begin()));
-            }
-
-            const std::string updated_line = std::string(match0.substr(0, len));
-            matches.emplace_back(updated_line + path_postfix, updated_line.length());
-        }
-
-        std::sort(matches.begin(), matches.end(), [](const auto & a, const auto & b) {
-            return a.first.compare(0, a.second, b.first, 0, b.second) < 0;
-        });
-    }
-
-    return matches;
-}
-
-static constexpr size_t FILE_GLOB_MAX_RESULTS = 100;
-
 // satisfies -Wmissing-declarations
 int llama_cli(int argc, char ** argv);
 
@@ -375,25 +41,6 @@ int llama_cli(int argc, char ** argv) {
         return 1;
     }
 
-    // TODO: maybe support it later?
-    if (params.conversation_mode == COMMON_CONVERSATION_MODE_DISABLED) {
-        console::error("--no-conversation is not supported by llama-cli\n");
-        console::error("please use llama-completion instead\n");
-    }
-
-    // struct that contains llama context and inference
-    cli_context ctx_cli(params);
-
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    // TODO: avoid using atexit() here by making `console` a singleton
-    console::init(params.simple_io, params.use_color);
-    atexit([]() { console::cleanup(); });
-
-    console::set_display(DISPLAY_TYPE_RESET);
-    console::set_completion_callback(auto_completion_callback);
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
     struct sigaction sigint_action;
     sigint_action.sa_handler = signal_handler;
@@ -408,276 +55,11 @@ int llama_cli(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    console::log("\nLoading model... "); // followed by loading animation
-    console::spinner::start();
-    if (!ctx_cli.ctx_server.load_model(params)) {
-        console::spinner::stop();
-        console::error("\nFailed to load the model\n");
+    cli_context ctx_cli(params);
+
+    if (!ctx_cli.init()) {
         return 1;
     }
 
-    ctx_cli.defaults.sampling = params.sampling;
-
-    console::spinner::stop();
-    console::log("\n");
-
-    std::thread inference_thread([&ctx_cli]() {
-        ctx_cli.ctx_server.start_loop();
-    });
-
-    auto inf = ctx_cli.ctx_server.get_meta();
-    std::string modalities = "text";
-    if (inf.has_inp_image) {
-        modalities += ", vision";
-    }
-    if (inf.has_inp_audio) {
-        modalities += ", audio";
-    }
-
-    auto add_system_prompt = [&]() {
-        if (!params.system_prompt.empty()) {
-            ctx_cli.messages.push_back({
-                {"role",    "system"},
-                {"content", params.system_prompt}
-            });
-        }
-    };
-    add_system_prompt();
-
-    console::log("\n");
-    console::log("%s\n", LLAMA_ASCII_LOGO);
-    console::log("build      : %s\n", inf.build_info.c_str());
-    console::log("model      : %s\n", inf.model_name.c_str());
-    if (!inf.model_ftype.empty()) {
-        console::log("ftype      : %s\n", inf.model_ftype.c_str());
-    }
-    console::log("modalities : %s\n", modalities.c_str());
-    if (!params.system_prompt.empty()) {
-        console::log("using custom system prompt\n");
-    }
-    console::log("\n");
-    console::log("available commands:\n");
-    console::log("  /exit or Ctrl+C     stop or exit\n");
-    console::log("  /regen              regenerate the last response\n");
-    console::log("  /clear              clear the chat history\n");
-    console::log("  /read <file>        add a text file\n");
-    console::log("  /glob <pattern>     add text files using globbing pattern\n");
-    if (inf.has_inp_image) {
-        console::log("  /image <file>       add an image file\n");
-    }
-    if (inf.has_inp_audio) {
-        console::log("  /audio <file>       add an audio file\n");
-    }
-    if (inf.has_inp_video) {
-        console::log("  /video <file>       add a video file\n");
-    }
-    console::log("\n");
-
-    // interactive loop
-    std::string cur_msg;
-
-    auto add_text_file = [&](const std::string & fname) -> bool {
-        std::string marker = ctx_cli.load_input_file(fname, false);
-        if (marker.empty()) {
-            console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-            return false;
-        }
-        if (inf.fim_sep_token != LLAMA_TOKEN_NULL) {
-            cur_msg += common_token_to_piece(ctx_cli.ctx_server.get_llama_context(), inf.fim_sep_token, true);
-            cur_msg += fname;
-            cur_msg.push_back('\n');
-        } else {
-            cur_msg += "--- File: ";
-            cur_msg += fname;
-            cur_msg += " ---\n";
-        }
-        cur_msg += marker;
-        console::log("Loaded text from '%s'\n", fname.c_str());
-        return true;
-    };
-
-    while (true) {
-        std::string buffer;
-        console::set_display(DISPLAY_TYPE_USER_INPUT);
-        if (params.prompt.empty()) {
-            console::log("\n> ");
-            std::string line;
-            bool another_line = true;
-            do {
-                another_line = console::readline(line, params.multiline_input);
-                buffer += line;
-            } while (another_line);
-        } else {
-            // process input prompt from args
-            for (auto & fname : params.image) {
-                std::string marker = ctx_cli.load_input_file(fname, true);
-                if (marker.empty()) {
-                    console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-                    break;
-                }
-                console::log("Loaded media from '%s'\n", fname.c_str());
-                cur_msg += marker;
-            }
-            buffer = params.prompt;
-            if (buffer.size() > 500) {
-                console::log("\n> %s ... (truncated)\n", buffer.substr(0, 500).c_str());
-            } else {
-                console::log("\n> %s\n", buffer.c_str());
-            }
-            params.prompt.clear(); // only use it once
-        }
-        console::set_display(DISPLAY_TYPE_RESET);
-        console::log("\n");
-
-        if (should_stop()) {
-            g_is_interrupted.store(false);
-            break;
-        }
-
-        // remove trailing newline
-        if (!buffer.empty() &&buffer.back() == '\n') {
-            buffer.pop_back();
-        }
-
-        // skip empty messages
-        if (buffer.empty()) {
-            continue;
-        }
-
-        bool add_user_msg = true;
-
-        // process commands
-        if (string_starts_with(buffer, "/exit")) {
-            break;
-        } else if (string_starts_with(buffer, "/regen")) {
-            if (ctx_cli.messages.size() >= 2) {
-                size_t last_idx = ctx_cli.messages.size() - 1;
-                ctx_cli.messages.erase(last_idx);
-                add_user_msg = false;
-            } else {
-                console::error("No message to regenerate.\n");
-                continue;
-            }
-        } else if (string_starts_with(buffer, "/clear")) {
-            ctx_cli.messages.clear();
-            add_system_prompt();
-
-            ctx_cli.input_files.clear();
-            console::log("Chat history cleared.\n");
-            continue;
-        } else if (
-                (string_starts_with(buffer, "/image ") && inf.has_inp_image) ||
-                (string_starts_with(buffer, "/audio ") && inf.has_inp_audio) ||
-                (string_starts_with(buffer, "/video ") && inf.has_inp_video)) {
-            // just in case (bad copy-paste for example), we strip all trailing/leading spaces
-            std::string fname = string_strip(buffer.substr(7));
-            std::string marker = ctx_cli.load_input_file(fname, true);
-            if (marker.empty()) {
-                console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
-                continue;
-            }
-            cur_msg += marker;
-            console::log("Loaded media from '%s'\n", fname.c_str());
-            continue;
-        } else if (string_starts_with(buffer, "/read ")) {
-            std::string fname = string_strip(buffer.substr(6));
-            add_text_file(fname);
-            continue;
-        } else if (string_starts_with(buffer, "/glob ")) {
-            std::error_code ec;
-            size_t count = 0;
-            auto curdir = std::filesystem::current_path();
-            std::string pattern = string_strip(buffer.substr(6));
-            std::filesystem::path rel_path;
-
-            auto startglob = pattern.find_first_of("![*?");
-            if (startglob != std::string::npos && startglob != 0) {
-                auto endpath = pattern.substr(0, startglob).find_last_of('/');
-                if (endpath != std::string::npos) {
-                    std::string rel_pattern = pattern.substr(0, endpath);
-#if !defined(_WIN32)
-                    if (string_starts_with(rel_pattern, '~')) {
-                        const char * home = std::getenv("HOME");
-                        if (home && home[0]) {
-                            rel_pattern = home + rel_pattern.substr(1);
-                        }
-                    }
-#endif
-                    rel_path = rel_pattern;
-                    pattern.erase(0, endpath + 1);
-                    curdir /= rel_path;
-                }
-            }
-
-            for (const auto & entry : std::filesystem::recursive_directory_iterator(curdir,
-                    std::filesystem::directory_options::skip_permission_denied, ec)) {
-                if (!entry.is_regular_file()) {
-                    continue;
-                }
-
-                std::string rel = std::filesystem::relative(entry.path(), curdir, ec).string();
-                if (ec) {
-                    ec.clear();
-                    continue;
-                }
-                std::replace(rel.begin(), rel.end(), '\\', '/');
-
-                if (!glob_match(pattern, rel)) {
-                    continue;
-                }
-
-                if (!add_text_file((rel_path / rel).string())) {
-                    continue;
-                }
-
-                if (++count >= FILE_GLOB_MAX_RESULTS) {
-                    console::error("Maximum number of globbed files allowed (%zu) reached.\n", FILE_GLOB_MAX_RESULTS);
-                    break;
-                }
-            }
-            continue;
-        } else {
-            // not a command
-            cur_msg += buffer;
-        }
-
-        // generate response
-        if (add_user_msg) {
-            ctx_cli.messages.push_back({
-                {"role",    "user"},
-                {"content", cur_msg}
-            });
-            cur_msg.clear();
-        }
-        result_timings timings;
-        std::string assistant_content = ctx_cli.generate_completion(timings);
-        ctx_cli.messages.push_back({
-            {"role",    "assistant"},
-            {"content", assistant_content}
-        });
-        console::log("\n");
-
-        if (params.show_timings) {
-            console::set_display(DISPLAY_TYPE_INFO);
-            console::log("\n");
-            console::log("[ Prompt: %.1f t/s | Generation: %.1f t/s ]\n", timings.prompt_per_second, timings.predicted_per_second);
-            console::set_display(DISPLAY_TYPE_RESET);
-        }
-
-        if (params.single_turn) {
-            break;
-        }
-    }
-
-    console::set_display(DISPLAY_TYPE_RESET);
-
-    console::log("\nExiting...\n");
-    ctx_cli.ctx_server.terminate();
-    inference_thread.join();
-
-    // bump the log level to display timings
-    common_log_set_verbosity_thold(LOG_LEVEL_INFO);
-    common_memory_breakdown_print(ctx_cli.ctx_server.get_llama_context());
-
-    return 0;
+    return ctx_cli.run();
 }

tools/server/server-context.cpp

@@ -4521,6 +4521,7 @@ void server_routes::init_routes() {
             { "default_generation_settings", default_generation_settings_for_props },
             { "total_slots",                 params.n_parallel },
             { "model_alias",                 meta->model_name },
+            { "model_ftype",                 meta->model_ftype },
             { "model_path",                  meta->model_path },
             { "modalities",                  json {
                 {"vision", meta->has_inp_image},

tools/server/server-models.cpp

@@ -7,6 +7,7 @@
 #include "build-info.h"
 #include "preset.h"
 #include "download.h"
+#include "http.h"
 
 #include <cpp-httplib/httplib.h> // TODO: remove this once we use HTTP client from download.h
 #include <optional>
@@ -28,14 +29,7 @@
 #include <sstream>
 #include <cstring>
 
-#ifdef _WIN32
-#include <winsock2.h>
-#include <windows.h>
-#else
-#include <sys/socket.h>
-#include <netinet/in.h>
-#include <arpa/inet.h>
-#include <unistd.h>
+#ifndef _WIN32
 extern char **environ;
 #endif
 
@@ -716,66 +710,6 @@ std::optional<server_model_meta> server_models::get_meta(const std::string & nam
     return std::nullopt;
 }
 
-static int get_free_port() {
-#ifdef _WIN32
-    WSADATA wsaData;
-    if (WSAStartup(MAKEWORD(2, 2), &wsaData) != 0) {
-        return -1;
-    }
-    typedef SOCKET native_socket_t;
-#define INVALID_SOCKET_VAL INVALID_SOCKET
-#define CLOSE_SOCKET(s) closesocket(s)
-#else
-    typedef int native_socket_t;
-#define INVALID_SOCKET_VAL -1
-#define CLOSE_SOCKET(s) close(s)
-#endif
-
-    native_socket_t sock = socket(AF_INET, SOCK_STREAM, 0);
-    if (sock == INVALID_SOCKET_VAL) {
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-    struct sockaddr_in serv_addr;
-    std::memset(&serv_addr, 0, sizeof(serv_addr));
-    serv_addr.sin_family = AF_INET;
-    serv_addr.sin_addr.s_addr = htonl(INADDR_ANY);
-    serv_addr.sin_port = htons(0);
-
-    if (bind(sock, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) != 0) {
-        CLOSE_SOCKET(sock);
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-#ifdef _WIN32
-    int namelen = sizeof(serv_addr);
-#else
-    socklen_t namelen = sizeof(serv_addr);
-#endif
-    if (getsockname(sock, (struct sockaddr*)&serv_addr, &namelen) != 0) {
-        CLOSE_SOCKET(sock);
-#ifdef _WIN32
-        WSACleanup();
-#endif
-        return -1;
-    }
-
-    int port = ntohs(serv_addr.sin_port);
-
-    CLOSE_SOCKET(sock);
-#ifdef _WIN32
-    WSACleanup();
-#endif
-
-    return port;
-}
-
 // helper to convert vector<string> to char **
 // pointers are only valid as long as the original vector is valid
 static std::vector<char *> to_char_ptr_array(const std::vector<std::string> & vec) {
@@ -879,7 +813,7 @@ void server_models::load(const std::string & name, const load_options & opts) {
     // prepare new instance info
     instance_t inst;
     inst.meta             = meta;
-    inst.meta.port        = get_free_port();
+    inst.meta.port        = common_http_get_free_port();
     inst.meta.status      = SERVER_MODEL_STATUS_LOADING;
     inst.meta.loaded_info = json{};
     inst.meta.last_used   = ggml_time_ms();

tools/server/server.cpp

@@ -36,6 +36,19 @@ static inline void signal_handler(int signal) {
     shutdown_handler(signal);
 }
 
+// satisfies -Wmissing-declarations (used by llama command)
+int llama_server(int argc, char ** argv);
+
+// to be used via CLI (argc / argv are used by router mode only)
+int llama_server(common_params & params, int argc, char ** argv);
+void llama_server_terminate();
+void llama_server_terminate() {
+    if (shutdown_handler) {
+        shutdown_handler(0);
+    }
+}
+
+
 // wrapper function that handles exceptions and logs errors
 // this is to make sure handler_t never throws exceptions; instead, it returns an error response
 static server_http_context::handler_t ex_wrapper(server_http_context::handler_t func) {
@@ -72,9 +85,6 @@ static server_http_context::handler_t ex_wrapper(server_http_context::handler_t
     };
 }
 
-// satisfies -Wmissing-declarations
-int llama_server(int argc, char ** argv);
-
 int llama_server(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
@@ -94,16 +104,26 @@ int llama_server(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    return llama_server(params, argc, argv);
+}
+
+int llama_server(common_params & params, int argc, char ** argv) {
+    bool is_run_by_cli = (argv == nullptr);
+
     common_models_handler models_handler;
-    try {
-        models_handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
-        if (common_models_handler_is_preset_repo(models_handler)) {
-            // apply the preset and start the server in router mode
-            common_models_handler_apply(models_handler, params);
+
+    // note: router mode also accepts -hf remote-preset, so we need to check that first
+    if (!is_run_by_cli && !params.model.hf_repo.empty()) {
+        try {
+            models_handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
+            if (common_models_handler_is_preset_repo(models_handler)) {
+                // apply the preset and start the server in router mode
+                common_models_handler_apply(models_handler, params);
+            }
+        } catch (const std::exception & e) {
+            SRV_ERR("failed to fetch model metadata: %s\n", e.what());
+            return 1;
         }
-    } catch (const std::exception & e) {
-        SRV_ERR("failed to fetch model metadata: %s\n", e.what());
-        return 1;
     }
 
     // router server never loads a model and must not touch the GPU
@@ -321,8 +341,9 @@ int llama_server(int argc, char ** argv) {
 
     if (child.is_child() && child.get_mode() == SERVER_CHILD_MODE_DOWNLOAD) {
         return child.run_download(params);
-    } else if (!is_router_server) {
+    } else if (!is_router_server && !is_run_by_cli) {
         // single-model mode (NOT spawned by router)
+        // if this is invoked by CLI, model downloading should be already handled
         try {
             common_models_handler_apply(models_handler, params);
         } catch (const std::exception & e) {
@@ -411,20 +432,22 @@ int llama_server(int argc, char ** argv) {
         };
     }
 
-    // TODO: refactor in common/console
+    // register signal handler if not running by CLI
+    if (!is_run_by_cli) {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-    struct sigaction sigint_action;
-    sigint_action.sa_handler = signal_handler;
-    sigemptyset (&sigint_action.sa_mask);
-    sigint_action.sa_flags = 0;
-    sigaction(SIGINT, &sigint_action, NULL);
-    sigaction(SIGTERM, &sigint_action, NULL);
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = signal_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+        sigaction(SIGTERM, &sigint_action, NULL);
 #elif defined (_WIN32)
-    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-    };
-    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
+    }
 
     SRV_INF("listening on %s\n", ctx_http.listening_address.c_str());