metal : add set_rows with src0 f16 (#25434)

* preliminary extract common header

* op source split

* split metallib into 8 libs && load in parallel

* derive kernel->library routing from functionNames

* x-macro lib list + underscore filenames, dedup QK_NL, MRC fixes

* op source split 8 to 20

* improve robustness of source fallback

* clean up

* change bool -> atomic_bool

* only prepend headers that source actually includes

* no semaphore, use GCD global queue

* dedup library compile path, fix NSError lifetime, rename gla

* relocate upstream concat/rope_back/repeat kernel changes into split files

* move ggml-common.h from common.h into dequantize.h to shrink binary size

---------

Co-authored-by: lvyichen <lvyichen@stepfun.com>

.github/workflows/hip-quality-check.yml

@@ -9,6 +9,8 @@ on:
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
       '**/*.cuh',
+      'ggml/src/ggml-hip/CMakeLists.txt',
+      'ggml/src/ggml-cuda/vendors/hip.h',
       'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 
@@ -18,6 +20,8 @@ on:
       '.github/workflows/hip-quality-check.yml',
       '**/*.cu',
       '**/*.cuh',
+      'ggml/src/ggml-hip/CMakeLists.txt',
+      'ggml/src/ggml-cuda/vendors/hip.h',
       'scripts/hip/gcn-cdna-vgpr-check.py'
     ]
 

common/arg.cpp

@@ -27,6 +27,7 @@
 #include <cinttypes>
 #include <climits>
 #include <cstdarg>
+#include <filesystem>
 #include <fstream>
 #include <list>
 #include <regex>
@@ -718,9 +719,8 @@ 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 +1240,13 @@ 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"),
@@ -3451,9 +3458,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_LOG_FILE"));
     add_opt(common_arg(
         {"--log-prompts-dir"}, "PATH",
-        "Log prompts to directory (only used for debugging, default: disabled)",
+        "Log prompts to directory (auto-created if not present; only used for debugging, default: disabled)",
         [](common_params & params, const std::string & value) {
             params.path_prompts_log_dir = value;
+            std::error_code ec;
+            std::filesystem::create_directories(value, ec);
+            if (ec) {
+                fprintf(stderr, "warning: failed to create prompts-log-dir '%s': %s\n", value.c_str(), ec.message().c_str());
+            }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(

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;
+}

examples/llama-eval/llama-eval.py

@@ -362,7 +362,7 @@ class EvalState:
             case = cases.get(task_id, {})
             status = case.get("status", "pending")
             expected = case.get("expected", "")
-            answer = case.get("answer", "") if status == "ok" else ""
+            answer = case.get("answer") or "" if status == "ok" else ""
             is_correct = case.get("correct", False) if status == "ok" else False
             response = case.get("response", "") or ""
             prompt = case.get("prompt", "") or ""
@@ -647,7 +647,7 @@ class EvalState:
             question, prompt, expected = self.get_case(i)
             case = cases.get(task_id, {})
             status = case.get("status", "pending")
-            answer = case.get("answer", "N/A") if status == "ok" else "N/A"
+            answer = case.get("answer") or "N/A" if status == "ok" else "N/A"
             tokens = case.get("tokens")
             tokens_str = str(tokens) if tokens is not None else "N/A"
             tps_gen = case.get("tps_gen")

ggml/CMakeLists.txt

@@ -340,9 +340,6 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -4709,10 +4709,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             } break;
         case GGML_OP_SET_ROWS:
             {
-                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
-                       op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
-                       op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
-                       op->src[0]->type == GGML_TYPE_F32 &&
+                return (
+                           (
+                               (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                               op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
+                               op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
+                               op->src[0]->type == GGML_TYPE_F32
+                           ) || (
+                               op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F16
+                           )
+                       ) &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
         case GGML_OP_SET:

ggml/src/ggml-cuda/set-rows.cu

@@ -322,17 +322,77 @@ static void set_rows_cuda(ggml_backend_cuda_context & ctx, const ggml_tensor * s
     }
 }
 
+template<>
+void set_rows_cuda<half, int32_t>(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const half    * src0_d = (const half *)src0->data;
+    const int32_t * src1_d = (const int32_t *)src1->data;
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    cudaStream_t stream = ctx.stream();
+
+
+    if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
+    }
+}
+
+template<>
+void set_rows_cuda<half, int64_t>(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const half    * src0_d = (const half *)src0->data;
+    const int64_t * src1_d = (const int64_t *)src1->data;
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    cudaStream_t stream = ctx.stream();
+
+
+    if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
+    }
+}
+
 
 void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
     GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);
 
-    if (src1->type == GGML_TYPE_I64) {
-        set_rows_cuda<float, int64_t>(ctx, src0, src1, dst);
+    if (src0->type == GGML_TYPE_F32) {
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_cuda<float, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_cuda<float, int32_t>(ctx, src0, src1, dst);
+        }
+    } else if (src0->type == GGML_TYPE_F16) {
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_cuda<half, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_cuda<half, int32_t>(ctx, src0, src1, dst);
+        }
     } else {
-        set_rows_cuda<float, int32_t>(ctx, src0, src1, dst);
+        GGML_ABORT("unsupported type %s", ggml_type_name(src0->type));
     }
 }

ggml/src/ggml-metal/CMakeLists.txt

@@ -24,62 +24,119 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})
 
-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
     if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
-    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +147,46 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )
 
-    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )
 
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()

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

@@ -160,11 +160,15 @@ 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, ggml_type tidx, ggml_type tdst) {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_metal_library_t lib, const ggml_tensor * op) {
     char base[256];
     char name[256];
 
-    snprintf(base, 256, "kernel_set_rows_%s_%s", ggml_type_name(tdst), ggml_type_name(tidx));
+    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(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, enum ggml_type tidx, enum ggml_type tdst);
+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_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

@@ -94,8 +94,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;
 
     ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -103,160 +158,376 @@ struct ggml_metal_library {
     NSLock * lock;
 };
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}
 
-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];
 
-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }
 
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }
 
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}
 
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];
 
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }
 
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }
 
-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                 }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
             }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
         }
 
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }
 
-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}
 
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];
 
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];
 
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
             }
 
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;
 
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
             @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];
 
-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                 [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
             }
         }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
     }
 
-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);
 
-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
     res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
     return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }
 
 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -318,10 +589,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
         return NULL;
     }
 
-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];
 
     return res;
 }
@@ -331,8 +603,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
         return;
     }
 
-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
     }
 
     ggml_metal_pipelines_free(lib->pipelines);
@@ -393,11 +671,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
 
         GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);
 
+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
         id<MTLFunction> mtl_function;
         if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
         } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
             [lib->lock unlock];
@@ -1334,7 +1629,7 @@ 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) {
+                if (op->src[0]->type != GGML_TYPE_F32 && op->src[0]->type != GGML_TYPE_F16) {
                     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->src[1]->type, op->type);
+    auto pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op);
 
     const int32_t nk0 = ne0/ggml_blck_size(op->type);
 

ggml/src/ggml-metal/kernels/argsort.metal

@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;

ggml/src/ggml-metal/kernels/binbcast.metal

@@ -0,0 +1,226 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;

ggml/src/ggml-metal/kernels/common.h

@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+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
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}

ggml/src/ggml-metal/kernels/conv.metal

@@ -0,0 +1,527 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);

ggml/src/ggml-metal/kernels/dequantize.h

@@ -0,0 +1,686 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 16; i++) {
+        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+

ggml/src/ggml-metal/kernels/gated_delta_net.metal

@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif

ggml/src/ggml-metal/kernels/misc.metal

@@ -0,0 +1,347 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;

ggml/src/ggml-metal/kernels/mul_mm.metal

@@ -0,0 +1,838 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;

ggml/src/ggml-metal/kernels/norm.metal

@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}

ggml/src/ggml-metal/kernels/pool.metal

@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}

ggml/src/ggml-metal/kernels/quantize.h

@@ -0,0 +1,213 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}

ggml/src/ggml-metal/kernels/quantize.metal

@@ -0,0 +1,389 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+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 &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    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 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);
+
+    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>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    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);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) 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>;
+

ggml/src/ggml-metal/kernels/reduce.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;

ggml/src/ggml-metal/kernels/rope.metal

@@ -0,0 +1,318 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;

ggml/src/ggml-metal/kernels/softmax.metal

@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;

ggml/src/ggml-metal/kernels/solve_tri.metal

@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}

ggml/src/ggml-metal/kernels/ssm.metal

@@ -0,0 +1,279 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}

ggml/src/ggml-metal/kernels/tri.metal

@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif

ggml/src/ggml-metal/kernels/unary.metal

@@ -0,0 +1,360 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;

ggml/src/ggml-metal/kernels/upscale.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device const char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}

ggml/src/ggml-metal/kernels/wkv.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}

src/llama-batch.cpp

@@ -505,7 +505,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
     return ubatch_add(idxs, idxs.size(), false);
 }
 
-llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail) {
     if (sequential && has_cpl) {
         LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
@@ -548,7 +548,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
-    const uint32_t n_seqs = cur_seq_set.size();
+    uint32_t n_seqs = cur_seq_set.size();
 
     // we are done
     if (n_seqs == 0) {
@@ -569,7 +569,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
     std::vector<idx_vec_t> idxs_per_seq(n_seqs);
 
     while (true) {
-        // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and
+        // we can only add new n_seq_tokens tokens if all the sequence sets have at least 1 more unused tokens and
         //   if we haven't reached n_ubatch
         bool can_expand = true;
 
@@ -600,6 +600,72 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
+    // if n_keep_tail > 0, keep only the seqs that either finish in this ubatch or have at least
+    //   n_keep_tail tokens remaining for a future ubatch, so that the trailing n_keep_tail tokens
+    //   of each seq are never split across ubatches
+    if (n_keep_tail > 0) {
+        GGML_ASSERT(n_ubatch > n_keep_tail);
+
+        auto n_remaining = [&](uint32_t s) {
+            return (uint32_t) (seq_set_map[cur_seq_set[s]].size() - cur_idx[s]);
+        };
+
+        // keep the longest prefix of seqs that satisfy the constraint, to preserve sequential seq ids
+        uint32_t n_keep = 0;
+        while (n_keep < n_seqs) {
+            const uint32_t remaining = n_remaining(n_keep);
+
+            if (remaining != 0 && remaining < n_keep_tail) {
+                break;
+            }
+
+            n_keep++;
+        }
+
+        // all seqs violate the constraint - resolve the first one directly and emit it alone
+        if (n_keep == 0) {
+            auto & idxs = idxs_per_seq[0];
+
+            const auto & seq_idxs = seq_set_map[cur_seq_set[0]];
+
+            if (idxs.size() + n_remaining(0) <= n_ubatch) {
+                // extend the seq to completion
+                while (n_remaining(0) > 0) {
+                    const int32_t idx = seq_idxs[cur_idx[0]];
+
+                    idxs.push_back(idx);
+
+                    used[idx] = true;
+                    ++n_used;
+
+                    ++cur_idx[0];
+                }
+            } else {
+                // truncate the seq so that at least n_keep_tail tokens remain
+                while (n_remaining(0) < n_keep_tail) {
+                    used[idxs.back()] = false;
+                    --n_used;
+
+                    idxs.pop_back();
+
+                    --cur_idx[0];
+                }
+            }
+
+            n_keep = 1;
+        }
+
+        // return the tokens of the deferred seqs back to the pool
+        for (uint32_t s = n_keep; s < n_seqs; ++s) {
+            for (const int32_t idx : idxs_per_seq[s]) {
+                used[idx] = false;
+                --n_used;
+            }
+        }
+
+        n_seqs = n_keep;
+    }
+
     // concat the per-sequence-set lists
     std::vector<int32_t> idxs;
 
@@ -814,7 +880,7 @@ void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
         LLAMA_LOG_DEBUG("%s:   output     = %p\n", __func__, (void *) ubatch.output);
         LLAMA_LOG_DEBUG("%s:   n_outputs  = %d\n", __func__, n_outputs);
 
-        if (debug > 1) {
+        if (debug > 0) {
             int seq_id_max = 0;
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {

src/llama-batch.h

@@ -104,7 +104,8 @@ public:
 
     // make ubatches of equal-length sequences sets
     // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
-    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
+    // n_keep_tail = minimum trailing tokens of a seq that must land in the same ubatch
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);

src/llama-context.cpp

@@ -17,6 +17,7 @@
 #include <cstring>
 #include <limits>
 #include <stdexcept>
+#include <string>
 
 //
 // llama_context
@@ -30,6 +31,30 @@ static llm_graph_type ctx_type_to_graph_type(llama_context_type ctx_type) {
     throw std::runtime_error("Unsupported ctx type");
 }
 
+struct llm_fused_op_probe {
+    llm_fused_op op;
+    const char * name;
+    uint32_t n_tokens_per_seq;
+};
+
+static const llm_fused_op_probe llm_fused_op_flash_attn_probe = {
+    /*.op               =*/ LLM_FUSED_OP_FLASH_ATTN,
+    /*.name             =*/ "Flash Attention",
+    /*.n_tokens_per_seq =*/ 1,
+};
+
+static const llm_fused_op_probe llm_fused_op_gdn_ar_probe = {
+    /*.op               =*/ LLM_FUSED_OP_GDN_AR,
+    /*.name             =*/ "fused Gated Delta Net (autoregressive)",
+    /*.n_tokens_per_seq =*/ 1,
+};
+
+static const llm_fused_op_probe llm_fused_op_gdn_ch_probe = {
+    /*.op               =*/ LLM_FUSED_OP_GDN_CH,
+    /*.name             =*/ "fused Gated Delta Net (chunked)",
+    /*.n_tokens_per_seq =*/ 16,
+};
+
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
@@ -436,6 +461,69 @@ llama_context::~llama_context() {
     ggml_opt_free(opt_ctx);
 }
 
+void llama_context::resolve_fused_ops(const llama_memory_context_i * mctx, uint32_t n_seqs) {
+    const char * func = __func__;
+    auto resolve = [&](const llm_fused_op_probe & probe, bool & enabled) {
+        if (!enabled) {
+            return;
+        }
+
+        const uint32_t n_tokens_probe = probe.n_tokens_per_seq*n_seqs;
+
+        auto * gf = graph_reserve(n_tokens_probe, n_seqs, n_tokens_probe, mctx, true);
+        if (!gf) {
+            throw std::runtime_error(std::string("failed to reserve graph for ") + probe.name + " check");
+        }
+
+        bool device_mismatch = false;
+        for (const auto & node : get_gf_res_reserve()->get_fused_nodes()) {
+            if (node.op != probe.op) {
+                continue;
+            }
+
+            GGML_ASSERT(node.il >= 0);
+
+            ggml_backend_t backend_fused = ggml_backend_sched_get_tensor_backend(sched.get(), node.tensor);
+            ggml_backend_dev_t device_fused = backend_fused ? ggml_backend_get_device(backend_fused) : nullptr;
+
+            // TODO: make this descriptor-specific; model.dev_layer() preserves the current behavior,
+            // but is still wrong for cases like --no-kv-offload.
+            ggml_backend_dev_t device_layer = model.dev_layer(node.il);
+
+            if (device_fused != device_layer) {
+                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but %s "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        func, node.il,
+                        device_layer ? ggml_backend_dev_name(device_layer) : "none",
+                        probe.name,
+                        device_fused ? ggml_backend_dev_name(device_fused) : "none");
+                device_mismatch = true;
+                break;
+            }
+        }
+
+        if (device_mismatch) {
+            enabled = false;
+            LLAMA_LOG_WARN("%s: %s not supported, set to disabled\n", func, probe.name);
+        } else {
+            enabled = true;
+            LLAMA_LOG_INFO("%s: %s enabled\n", func, probe.name);
+        }
+    };
+
+    if (cparams.auto_fa) {
+        resolve(llm_fused_op_flash_attn_probe, cparams.flash_attn);
+        cparams.auto_fa = false;
+    }
+
+    if (cparams.auto_fgdn) {
+        LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", func);
+        resolve(llm_fused_op_gdn_ar_probe, cparams.fused_gdn_ar);
+        resolve(llm_fused_op_gdn_ch_probe, cparams.fused_gdn_ch);
+        cparams.auto_fgdn = false;
+    }
+}
+
 void llama_context::sched_reserve() {
     if (!sched_need_reserve) {
         return;
@@ -475,128 +563,7 @@ void llama_context::sched_reserve() {
 
     LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
 
-    // resolve automatic Flash Attention use
-    if (cparams.auto_fa) {
-        auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
-        if (!gf) {
-            throw std::runtime_error("failed to reserve graph for Flash Attention check");
-        }
-
-        const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
-        bool fa_device_mismatch = false;
-        for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-            ggml_tensor * n = ggml_graph_node(gf, i);
-            if (n->op != GGML_OP_FLASH_ATTN_EXT) {
-                continue;
-            }
-            ggml_backend_dev_t device_fa = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-            // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
-            GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
-            const int il = std::stoi(n->name + prefix_len);
-            ggml_backend_dev_t device_kv = model.dev_layer(il);
-            if (device_fa != device_kv) {
-                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
-                        "is assigned to device %s (usually due to missing support)\n",
-                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
-                // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
-                fa_device_mismatch = true;
-                break;
-            }
-        }
-
-        if (fa_device_mismatch) {
-            cparams.flash_attn = false;
-            LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
-        } else {
-            cparams.flash_attn = true;
-            LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
-        }
-
-        cparams.auto_fa = false;
-    }
-
-    if (cparams.auto_fgdn) {
-        LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", __func__);
-
-        if (cparams.fused_gdn_ar) {
-            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
-            if (!gf) {
-                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (autoregressive)");
-            }
-
-            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_AR) + 1;
-            bool gdn_device_mismatch = false;
-            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-                ggml_tensor * n = ggml_graph_node(gf, i);
-                if (n->op != GGML_OP_GATED_DELTA_NET) {
-                    continue;
-                }
-                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_AR "-", prefix_len) == 0);
-                const int il = std::stoi(n->name + prefix_len);
-                ggml_backend_dev_t device_kv = model.dev_layer(il);
-                if (device_gdn != device_kv) {
-                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
-                            "is assigned to device %s (usually due to missing support)\n",
-                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
-                    gdn_device_mismatch = true;
-                    break;
-                }
-            }
-
-            if (gdn_device_mismatch) {
-                cparams.fused_gdn_ar = false;
-                LLAMA_LOG_WARN("%s: fused Gated Delta Net (autoregressive) not supported, set to disabled\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: fused Gated Delta Net (autoregressive) enabled\n", __func__);
-            }
-        }
-
-        if (cparams.fused_gdn_ch) {
-            // more than one token in the batch per sequence in order to take the chunked path
-            // note: n_outputs must match n_tokens for embedding models with mean/rank pooling,
-            // because build_pooling creates inp_mean with shape [n_tokens, n_seqs] and multiplies
-            // it with t_embd which is reduced to [n_outputs, ...] via out_ids. if n_outputs != n_tokens,
-            // the ggml_mul_mat assertion fails.
-            const uint32_t n_tokens_ch = 16*n_seqs;
-            auto * gf = graph_reserve(n_tokens_ch, n_seqs, n_tokens_ch, mctx.get(), true);
-            if (!gf) {
-                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (chunked)");
-            }
-
-            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_CH) + 1;
-            bool gdn_device_mismatch = false;
-            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-                ggml_tensor * n = ggml_graph_node(gf, i);
-                if (n->op != GGML_OP_GATED_DELTA_NET) {
-                    continue;
-                }
-                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_CH "-", prefix_len) == 0);
-                const int il = std::stoi(n->name + prefix_len);
-                ggml_backend_dev_t device_kv = model.dev_layer(il);
-                if (device_gdn != device_kv) {
-                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
-                            "is assigned to device %s (usually due to missing support)\n",
-                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
-                    gdn_device_mismatch = true;
-                    break;
-                }
-            }
-
-            if (gdn_device_mismatch) {
-                cparams.fused_gdn_ch = false;
-                LLAMA_LOG_WARN("%s: fused Gated Delta Net (chunked) not supported, set to disabled\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: fused Gated Delta Net (chunked) enabled\n", __func__);
-            }
-        }
-
-        cparams.auto_fgdn = false;
-    }
+    resolve_fused_ops(mctx.get(), n_seqs);
 
     // reserve worst-case graph
     int n_splits_pp = -1;

src/llama-context.h

@@ -262,6 +262,10 @@ private:
 
     llm_graph_cb graph_get_cb() const;
 
+    // disable auto fused ops (Flash Attention, Gated Delta Net) whose op lands on a device
+    // that differs from the layer it belongs to (usually due to missing backend support)
+    void resolve_fused_ops(const llama_memory_context_i * mctx, uint32_t n_seqs);
+
     // TODO: read/write lora adapters and cvec
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);

src/llama-graph.cpp

@@ -1192,6 +1192,7 @@ void llm_graph_result::reset() {
     params = {};
 
     inputs.clear();
+    fused_nodes.clear();
 
     buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
 
@@ -1293,6 +1294,10 @@ llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) {
     return inputs.back().get();
 }
 
+void llm_graph_result::add_fused_node(llm_graph_fused_node result) {
+    fused_nodes.push_back(result);
+}
+
 void llm_graph_result::set_params(const llm_graph_params & params) {
     this->params = params;
 }
@@ -1352,6 +1357,8 @@ void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
     }
 }
 
+
+
 ggml_tensor * llm_graph_context::build_cvec(
          ggml_tensor * cur,
                  int   il) const {
@@ -2402,7 +2409,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
-        cb(cur, LLAMA_TENSOR_NAME_FATTN, il);
+        res->add_fused_node({LLM_FUSED_OP_FLASH_ATTN, cur, il});
 
         ggml_flash_attn_ext_add_sinks(cur, sinks);
         ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);

src/llama-graph.h

@@ -38,6 +38,12 @@ enum llm_graph_type {
     LLM_GRAPH_TYPE_DECODER_MTP,
 };
 
+enum llm_fused_op {
+    LLM_FUSED_OP_FLASH_ATTN,
+    LLM_FUSED_OP_GDN_AR,
+    LLM_FUSED_OP_GDN_CH,
+};
+
 enum llm_ffn_op_type : int {
     LLM_FFN_NONE = 0,           // sentinel: unset; archs must assign before use
     LLM_FFN_SILU,
@@ -775,6 +781,12 @@ struct llm_graph_params {
     }
 };
 
+struct llm_graph_fused_node {
+    llm_fused_op op;
+    ggml_tensor * tensor;
+    int il;
+};
+
 class llm_graph_result {
 public:
     llm_graph_result(int64_t max_nodes);
@@ -808,6 +820,10 @@ public:
 
     llm_graph_input_i * add_input(llm_graph_input_ptr input);
 
+    void add_fused_node(llm_graph_fused_node result);
+
+    const std::vector<llm_graph_fused_node> & get_fused_nodes() const { return fused_nodes; }
+
     void set_params(const llm_graph_params & params);
 
     // important graph nodes
@@ -826,6 +842,7 @@ public:
     std::map<llama_seq_id, ggml_tensor *> t_sampled_probs;
 
     std::vector<llm_graph_input_ptr> inputs;
+    std::vector<llm_graph_fused_node> fused_nodes;
 
     ggml_context_ptr ctx_compute;
 

src/llama-impl.h

@@ -103,7 +103,3 @@ std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
 std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
-
-#define LLAMA_TENSOR_NAME_FATTN   "__fattn__"
-#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
-#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"

src/llama-kv-cache-dsa.cpp

@@ -113,7 +113,7 @@ llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
+            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

src/llama-kv-cache-dsv4.cpp

@@ -1110,7 +1110,7 @@ llama_memory_context_ptr llama_kv_cache_dsv4::init_batch(
             if (has_coupled) {
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq);
+                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, 0);
             }
 
             if (ubatch.n_tokens == 0) {

src/llama-kv-cache-iswa.cpp

@@ -206,7 +206,7 @@ llama_memory_context_ptr llama_kv_cache_iswa::init_batch(llama_batch_allocr & ba
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch, !unified);
+            auto ubatch = balloc.split_equal(n_ubatch, !unified, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

src/llama-kv-cache.cpp

@@ -706,7 +706,7 @@ llama_memory_context_ptr llama_kv_cache::init_batch(
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
+            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

src/llama-memory-hybrid-iswa.cpp

@@ -77,15 +77,15 @@ llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (mem_recr->n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
-                    const bool unified = (mem_attn->get_base()->get_n_stream() == 1);
-                    ubatch = balloc.split_equal(n_ubatch, !unified);
-                }
+                // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
+                const bool unified = (mem_attn->get_base()->get_n_stream() == 1);
+
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                const uint32_t n_rs_seq = mem_recr->n_rs_seq;
+
+                ubatch = balloc.split_equal(n_ubatch, !unified, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

src/llama-memory-hybrid.cpp

@@ -78,15 +78,15 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (mem_recr->n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
-                    const bool unified = (mem_attn->get_n_stream() == 1);
-                    ubatch = balloc.split_equal(n_ubatch, !unified);
-                }
+                // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
+                const bool unified = (mem_attn->get_n_stream() == 1);
+
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                const uint32_t n_rs_seq = mem_recr->n_rs_seq;
+
+                ubatch = balloc.split_equal(n_ubatch, !unified, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

src/llama-memory-recurrent.cpp

@@ -416,15 +416,12 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // TODO: non-sequential equal split can be done if using unified KV cache
-                    //       for simplicity, we always use sequential equal split for now
-                    ubatch = balloc.split_equal(n_ubatch, true);
-                }
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                ubatch = balloc.split_equal(n_ubatch, true, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

src/llama-vocab.cpp

@@ -887,9 +887,6 @@ 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.
@@ -1205,7 +1202,15 @@ private:
                 throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
             }
             const char * prefix_replacement = &(tokenizer.prefix_replacements)[longest_prefix_offset];
-            return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
+            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 };
         }
 
         // check if the input prefix contains a valid sequence of UTF-8 code units
@@ -2018,11 +2023,18 @@ 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) {

src/models/delta-net-base.cpp

@@ -401,9 +401,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     // K=1: output carries the final state only. state s is 4D [S_v, S_v, H_v, n_seqs].
     ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, /*K=*/1);
     if (n_tokens == 1) {
-        cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_AR, result, il});
     } else {
-        cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_CH, result, il});
     }
 
     ggml_tensor * output = ggml_view_4d(ctx0, result,
@@ -496,8 +496,8 @@ ggml_tensor * llm_build_delta_net_base::build_conv_state(
         ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_state_last, conv_state_update));
     } else {
         // [TAG_RECURRENT_ROLLBACK_SPLITS]
-        // TODO: this logic incorrectly assumes that the last (n_rs_seq + 1) tokens of a sequence in a batch are
-        //       inside the same ubatch. currently with `split_equal()` this is not correct
+        // this logic assumes that the last (n_rs_seq + 1) tokens of a sequence in a batch are inside
+        //   the same ubatch, which `split_equal()` guarantees via its n_keep_tail argument
 
         const int64_t K = (int64_t) cparams.n_rs_seq + 1;
 
@@ -566,9 +566,9 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
     // state s is 4D [S_v, S_v, H_v, n_seqs]; K snapshot slots are written into the output.
     ggml_tensor * gdn_out = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, K);
     if (n_seq_tokens > 1) {
-        cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_CH, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_CH, gdn_out, il});
     } else {
-        cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_AR, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_AR, gdn_out, il});
     }
 
     const int64_t attn_score_elems    = S_v * H_v * n_seq_tokens * n_seqs;

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/README-dev.md

@@ -57,7 +57,7 @@ The core architecture consists of the following components:
 - `server_tokens`: Unified representation of token sequences (supports both text and multimodal tokens); used by `server_task` and `server_slot`.
 - `server_prompt_checkpoint`: For recurrent (e.g., RWKV) and SWA models, stores snapshots of KV cache state. Enables reuse when subsequent requests share the same prompt prefix, saving redundant computation.
 - `server_models`: Standalone component for managing multiple backend instances (used in router mode). It is completely independent of `server_context`.
-- `stream_session_manager`: Process wide owner of resumable SSE stream sessions (`g_stream_sessions`), keyed by conversation id. Backs the replay buffer that lets a client reattach to a generation after an HTTP disconnect. See the "Resumable streaming" section below.
+- `stream_session_manager`: process wide owner of resumable SSE stream sessions, keyed by conversation id. A file-static singleton inside `server-stream.cpp`, driven through `server_stream_session_manager_start/stop`. Backs the replay buffer that lets a client reattach to a generation after an HTTP disconnect. See the "Resumable streaming" section below.
 
 ```mermaid
 graph TD
@@ -127,10 +127,12 @@ It is opt in via the `X-Conversation-Id` header on `POST /v1/chat/completions`.
 The feature lives entirely in `server-stream.{h,cpp}` and rests on three types:
 
 - `stream_session`: a bounded ring buffer (4 MiB cap, oldest bytes drop first) plus a condvar. `append` pushes raw SSE bytes, `read_from` drains from any offset and blocks for live bytes or finalize, `finalize` wakes readers, `cancel` stops the producer. One conv maps to at most one live session.
-- `stream_session_manager` (`g_stream_sessions`): owns all sessions keyed by conv id, enforces the one conv one session invariant via `create_or_replace`, and runs a GC thread that drops completed sessions past their TTL.
+- `stream_session_manager`: a file-static singleton (`g_stream_sessions`) inside `server-stream.cpp`, owns all sessions keyed by conv id, enforces the one conv one session invariant via `create_or_replace`, and runs a GC thread that drops completed sessions past their TTL. Exposed to main only through `server_stream_session_manager_start/stop`.
 - `stream_pipe_producer` / `stream_pipe_consumer`: the write and read ends. The producer owns the session lifetime and finalizes it on destruction; the consumer is read only and never finalizes, so a reader detaching cannot kill a running generation.
 
-Producer side: `server_res_generator` attaches a producer pipe when the header is present. The HTTP content provider mirrors every chunk into the ring before writing it to the socket. While a pipe is attached, `stream_aware_should_stop` ignores peer disconnect, so a dropped socket does not stop generation: only an explicit `DELETE` does. When the peer leaves early, `on_complete` calls `close()`, which drains the rest of the generation into the ring on the http worker.
+The implementation is hidden in `server-stream.cpp` (pimpl). The header exposes only the route handler factories, `server_stream_session_attach_pipe`, `server_stream_aware_should_stop`, `server_stream_conv_id_from_headers` and the GC lifecycle; the session, manager and consumer types stay in the `.cpp`.
+
+Producer side: `server_res_generator` attaches a producer pipe when the header is present. The HTTP content provider mirrors every chunk into the ring before writing it to the socket. While a pipe is attached, `server_stream_aware_should_stop` ignores peer disconnect, so a dropped socket does not stop generation: only an explicit `DELETE` does. When the peer leaves early, `on_complete` calls `close()`, which drains the rest of the generation into the ring on the http worker.
 
 Lifetime safety: the producer pipe holds a shared `alive` flag also captured by the session cancel hook. `~server_res_generator` calls `cleanup()` to clear that hook while the reader is still alive, so a `cancel` arriving during teardown can never call `stop()` on a freed response. This ordering is the most fragile part of the feature: finalizing or destroying the producer before `cleanup()` runs reintroduces a use after free.
 
@@ -144,7 +146,7 @@ Routes:
 
 Router mode binds the same paths to proxy handlers. A `conv_id -> child` map (`conv_models`), populated when a POST is routed, resolves the owning child in one lookup with no polling. The lookup groups ids per child; GET and DELETE proxy straight to the owner. This loopback REST hop is expected to move to a websocket IPC later, swapping only the transport.
 
-Lifecycle: `g_stream_sessions.start_gc()` runs in main after common init, `stop_gc()` runs first in `clean_up()` and finalizes every live session so no reader hangs. Reader blocking and the post drop drain both run on httplib worker threads, which block on a condvar rather than spin.
+Lifecycle: `server_stream_session_manager_start()` runs in main after common init, `server_stream_session_manager_stop()` runs first in `clean_up()` and finalizes every live session so no reader hangs. Reader blocking and the post drop drain both run on httplib worker threads, which block on a condvar rather than spin.
 
 | Constant | Value | Role |
 | --- | --- | --- |

tools/server/README.md

@@ -228,7 +228,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 | `--sleep-idle-seconds SECONDS` | number of seconds of idleness after which the server will sleep (default: -1; -1 = disabled) |
-| `--log-prompts-dir PATH` | Log prompts to directory (only used for debugging, default: disabled) |
+| `--log-prompts-dir PATH` | Log prompts to directory (auto-created if not present; only used for debugging, default: disabled) |
 | `--spec-draft-hf, -hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_HF_REPO) |
 | `--spec-draft-threads, -td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
 | `--spec-draft-threads-batch, -tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |

tools/server/server-context.cpp

@@ -4188,7 +4188,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                 }
             };
 
-            auto effective_should_stop = stream_aware_should_stop(res_this, req.should_stop);
+            auto effective_should_stop = server_stream_aware_should_stop(res_this, req.should_stop);
 
             try {
                 if (effective_should_stop()) {
@@ -4284,7 +4284,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
 
     // attach a producer pipe to the response when X-Conversation-Id is present.
     // the pipe mirrors SSE chunks into the ring buffer and wires up the cancel hook.
-    stream_session_attach_pipe(*res, req.headers);
+    server_stream_session_attach_pipe(*res, req.headers);
 
     return res;
 }
@@ -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();
@@ -1681,7 +1615,7 @@ void server_models_routes::init_routes() {
         }
         // remember which child serves this conversation so the stream routes can route straight
         // to it without polling, keyed on the exact conv id from the header
-        std::string conv_id = stream_conv_id_from_headers(req.headers);
+        std::string conv_id = server_stream_conv_id_from_headers(req.headers);
         if (!conv_id.empty()) {
             models.conv_models.remember(conv_id, name);
         }
@@ -1896,7 +1830,7 @@ void server_models_routes::init_routes() {
         if (!from.empty()) {
             child_path += "?from=" + from;
         }
-        SRV_INF("proxying stream resume to model %s on port %d, path=%s\n",
+        SRV_TRC("proxying stream resume to model %s on port %d, path=%s\n",
                 owner->name.c_str(), owner->port, child_path.c_str());
         auto proxy = std::make_unique<server_http_proxy>(
                 "GET",

tools/server/server-stream.cpp

@@ -6,6 +6,12 @@
 #include <chrono>
 #include <memory>
 #include <utility>
+#include <shared_mutex>
+
+enum class stream_read_status {
+    OK,
+    OFFSET_LOST,
+};
 
 namespace {
 constexpr int64_t STREAM_SESSION_TTL_SECONDS         = 300;
@@ -13,7 +19,6 @@ constexpr size_t  STREAM_SESSION_MAX_BYTES           = 4 * 1024 * 1024;
 constexpr int64_t STREAM_SESSION_GC_INTERVAL_SECONDS = 60;
 constexpr int64_t STREAM_READ_WAKE_INTERVAL_MS       = 200;
 
-// returns unix time in seconds
 int64_t now_seconds() {
     return std::chrono::duration_cast<std::chrono::seconds>(
         std::chrono::system_clock::now().time_since_epoch()
@@ -21,6 +26,91 @@ int64_t now_seconds() {
 }
 }
 
+// owns all live sessions keyed by conversation_id, one conv = at most one live session.
+// a periodic GC evicts expired ones
+class stream_session_manager {
+public:
+    stream_session_manager();
+    ~stream_session_manager();
+
+    stream_session_manager(const stream_session_manager &)             = delete;
+    stream_session_manager & operator=(const stream_session_manager &) = delete;
+
+    // install a new session, evicting and cancelling any previous one. conversation_id must be non empty
+    stream_session_ptr create_or_replace(const std::string & conversation_id);
+
+    stream_session_ptr get(const std::string & conversation_id);
+
+    std::vector<stream_session_ptr> list_all() const;
+
+    void evict(const std::string & conversation_id);
+
+    void evict_and_cancel(const std::string & conversation_id);
+
+    void start_gc();
+    void stop_gc();
+
+private:
+    void gc_loop();
+
+    mutable std::shared_mutex                           map_mu;
+    std::unordered_map<std::string, stream_session_ptr> sessions; // key: conversation_id
+    std::thread                                         gc_thread;
+    bool                                                running;
+    std::mutex                                          gc_wake_mu;
+    std::condition_variable                             gc_wake_cv;
+};
+
+// process wide manager, lifecycle controlled by llama-server main() via start_gc/stop_gc
+static stream_session_manager g_stream_sessions;
+
+void server_stream_session_manager_start() {
+    g_stream_sessions.start_gc();
+}
+
+void server_stream_session_manager_stop() {
+    g_stream_sessions.stop_gc();
+}
+
+struct stream_session {
+    std::string conversation_id;
+    int64_t     started_ts; // unix seconds at construction
+
+    stream_session(std::string conversation_id_, size_t max_bytes_);
+    stream_session(const stream_session &)             = delete;
+    stream_session & operator=(const stream_session &) = delete;
+
+    bool append(const char * data, size_t len);
+
+    void finalize();
+
+    // drain from offset into sink, blocking for more bytes or finalize. OFFSET_LOST if offset
+    // fell below the dropped prefix
+    stream_read_status read_from(size_t offset,
+        const std::function<bool(const char *, size_t)> & sink,
+        const std::function<bool()> & should_stop);
+
+    bool    is_done() const;
+    bool    is_cancelled() const;
+    size_t  total_size() const;     // bytes that ever entered the session
+    size_t  dropped_prefix() const; // bytes evicted from the front due to cap
+    int64_t completed_at() const;   // 0 while alive, unix seconds after finalize
+
+    void set_stop_producer(std::function<void()> fn);
+
+    void cancel();
+
+private:
+    mutable std::mutex      mu;
+    std::condition_variable cv;
+    std::vector<char>       buffer;
+    size_t                  prefix_dropped;
+    size_t                  cap_bytes;
+    bool                    done;
+    std::atomic<bool>       cancelled; // polled lock-free by the should_stop closure, no mu
+    int64_t                 completed_ts;
+    std::function<void()>   stop_producer;
+};
 stream_session::stream_session(std::string conversation_id_, size_t max_bytes_)
     : conversation_id(std::move(conversation_id_))
     , started_ts(now_seconds())
@@ -38,7 +128,7 @@ bool stream_session::append(const char * data, size_t len) {
     }
     {
         std::lock_guard<std::mutex> lock(mu);
-        if (done.load(std::memory_order_relaxed)) {
+        if (done) {
             return false;
         }
         if (len >= cap_bytes) {
@@ -62,11 +152,14 @@ bool stream_session::append(const char * data, size_t len) {
 }
 
 void stream_session::finalize() {
-    bool was_done = done.exchange(true, std::memory_order_acq_rel);
-    if (was_done) {
-        return;
+    {
+        std::lock_guard<std::mutex> lock(mu);
+        if (done) {
+            return;
+        }
+        done         = true;
+        completed_ts = now_seconds();
     }
-    completed_ts.store(now_seconds(), std::memory_order_release);
     cv.notify_all();
 }
 
@@ -96,7 +189,7 @@ stream_read_status stream_session::read_from(size_t offset,
             lock.lock();
             continue;
         }
-        if (done.load(std::memory_order_acquire)) {
+        if (done) {
             return stream_read_status::OK;
         }
         // wait for new bytes, finalize, or a periodic wake to re check should_stop
@@ -105,7 +198,8 @@ stream_read_status stream_session::read_from(size_t offset,
 }
 
 bool stream_session::is_done() const {
-    return done.load(std::memory_order_acquire);
+    std::lock_guard<std::mutex> lock(mu);
+    return done;
 }
 
 size_t stream_session::total_size() const {
@@ -119,7 +213,8 @@ size_t stream_session::dropped_prefix() const {
 }
 
 int64_t stream_session::completed_at() const {
-    return completed_ts.load(std::memory_order_acquire);
+    std::lock_guard<std::mutex> lock(mu);
+    return completed_ts;
 }
 
 void stream_session::set_stop_producer(std::function<void()> fn) {
@@ -128,7 +223,7 @@ void stream_session::set_stop_producer(std::function<void()> fn) {
 }
 
 void stream_session::cancel() {
-    // flip cancelled first so the producer-side stream_aware_should_stop can break out of the
+    // flip cancelled first so the producer-side server_stream_aware_should_stop can break out of the
     // recv() wait even if remove_waiting_task_ids does not notify the condvar (the cancel task
     // posted by rd.stop() will eventually notify, but we do not want to depend on that timing)
     cancelled.store(true, std::memory_order_release);
@@ -237,18 +332,24 @@ void stream_session_manager::evict_and_cancel(const std::string & conversation_i
 }
 
 void stream_session_manager::start_gc() {
-    if (running.exchange(true)) {
-        return;
+    {
+        std::lock_guard<std::mutex> lock(gc_wake_mu);
+        if (running) {
+            return;
+        }
+        running = true;
     }
     gc_thread = std::thread([this] { gc_loop(); });
 }
 
 void stream_session_manager::stop_gc() {
-    bool was_running = running.exchange(false);
+    bool was_running;
+    {
+        std::lock_guard<std::mutex> lock(gc_wake_mu);
+        was_running = running;
+        running = false;
+    }
     if (was_running) {
-        {
-            std::lock_guard<std::mutex> lock(gc_wake_mu);
-        }
         gc_wake_cv.notify_all();
         if (gc_thread.joinable()) {
             gc_thread.join();
@@ -270,15 +371,15 @@ void stream_session_manager::stop_gc() {
 }
 
 void stream_session_manager::gc_loop() {
-    while (running.load(std::memory_order_acquire)) {
+    while (true) {
         {
             std::unique_lock<std::mutex> lock(gc_wake_mu);
             gc_wake_cv.wait_for(lock,
                 std::chrono::seconds(STREAM_SESSION_GC_INTERVAL_SECONDS),
-                [this] { return !running.load(std::memory_order_acquire); });
-        }
-        if (!running.load(std::memory_order_acquire)) {
-            return;
+                [this] { return !running; });
+            if (!running) {
+                return;
+            }
         }
         int64_t cutoff = now_seconds() - STREAM_SESSION_TTL_SECONDS;
         std::vector<stream_session_ptr> to_drop;
@@ -301,10 +402,19 @@ void stream_session_manager::gc_loop() {
     }
 }
 
-// process wide manager, lifecycle controlled by llama-server main() via start_gc/stop_gc
-stream_session_manager g_stream_sessions;
+// stream_pipe
 
-// stream_pipe ---------------------------------------------------------------------------------
+// consumer end: read-only replay of the ring buffer, the destructor does not finalize the session
+struct stream_pipe_consumer : stream_pipe {
+    stream_read_status read(size_t & offset,
+        const std::function<bool(const char *, size_t)> & sink,
+        const std::function<bool()> & should_stop);
+
+    static std::shared_ptr<stream_pipe_consumer> create(stream_session_ptr session);
+
+private:
+    explicit stream_pipe_consumer(stream_session_ptr session);
+};
 
 stream_pipe::stream_pipe(stream_session_ptr session)
     : session_(std::move(session)) {
@@ -408,12 +518,10 @@ static server_http_res_ptr make_error_response(int status, const std::string & m
     return res;
 }
 
-server_http_context::handler_t make_stream_get_handler() {
+server_http_context::handler_t server_stream_make_get_handler() {
     return [](const server_http_req & req) -> server_http_res_ptr {
-        // GET /v1/stream/<conv_id>?from=N replays the SSE bytes already buffered for the
-        // session, blocks for more bytes when the session is still running, returns when
-        // the session is finalized. the body is streamed back as text/event-stream so the
-        // browser EventSource can attach to it like a fresh request
+        // GET /v1/stream/<conv_id>?from=N replays buffered SSE bytes then blocks for live
+        // bytes until the session finalizes, streamed as text/event-stream for EventSource
         std::string conv_id = req.get_param("conv_id");
         if (conv_id.empty()) {
             return make_error_response(400, "Missing conversation id in path", ERROR_TYPE_INVALID_REQUEST);
@@ -459,11 +567,10 @@ server_http_context::handler_t make_stream_get_handler() {
     };
 }
 
-server_http_context::handler_t make_streams_lookup_handler() {
+server_http_context::handler_t server_stream_make_lookup_handler() {
     return [](const server_http_req & req) -> server_http_res_ptr {
-        // POST /v1/streams/lookup with body {"conversation_ids": ["X", "Y", ...]} returns the
-        // matching sessions, only for ids the caller already knows. each id matches the exact key
-        // and any "<id>::<model>" variant, so one lookup covers every per model session for a conv
+        // POST /v1/streams/lookup returns the matching sessions, only for ids the caller already
+        // knows. each id matches the exact key and any "<id>::<model>" per model variant
         std::vector<std::string> requested;
         try {
             json body = json::parse(req.body);
@@ -518,11 +625,10 @@ server_http_context::handler_t make_streams_lookup_handler() {
     };
 }
 
-server_http_context::handler_t make_stream_delete_handler() {
+server_http_context::handler_t server_stream_make_delete_handler() {
     return [](const server_http_req & req) -> server_http_res_ptr {
-        // DELETE /v1/stream/<conv_id> is the explicit user Stop, cancels the producer hook
-        // wired by handle_completions_impl and evicts the buffer. idempotent, a session that
-        // already finalized or was never created returns 204 either way
+        // DELETE /v1/stream/<conv_id> is the explicit user Stop, cancels the producer and evicts
+        // the buffer. idempotent, returns 204 even if the session was already gone
         std::string conv_id = req.get_param("conv_id");
         if (conv_id.empty()) {
             return make_error_response(400, "Missing conversation id in path", ERROR_TYPE_INVALID_REQUEST);
@@ -536,7 +642,7 @@ server_http_context::handler_t make_stream_delete_handler() {
     };
 }
 
-std::string stream_conv_id_from_headers(const std::map<std::string, std::string> & headers) {
+std::string server_stream_conv_id_from_headers(const std::map<std::string, std::string> & headers) {
     // case-insensitive scan for x-conversation-id
     static constexpr char   target[]   = "x-conversation-id";
     static constexpr size_t target_len = sizeof(target) - 1;
@@ -555,8 +661,8 @@ std::string stream_conv_id_from_headers(const std::map<std::string, std::string>
     return std::string();
 }
 
-void stream_session_attach_pipe(server_http_res & res, const std::map<std::string, std::string> & headers) {
-    std::string conversation_id = stream_conv_id_from_headers(headers);
+void server_stream_session_attach_pipe(server_http_res & res, const std::map<std::string, std::string> & headers) {
+    std::string conversation_id = server_stream_conv_id_from_headers(headers);
     SRV_TRC("conv_id=%s (empty=%d)\n", conversation_id.c_str(), conversation_id.empty() ? 1 : 0);
     if (conversation_id.empty()) {
         return;
@@ -565,7 +671,7 @@ void stream_session_attach_pipe(server_http_res & res, const std::map<std::strin
     res.spipe = stream_pipe_producer::create(session, res);
 }
 
-std::function<bool()> stream_aware_should_stop(server_http_res * res, std::function<bool()> fallback) {
+std::function<bool()> server_stream_aware_should_stop(server_http_res * res, std::function<bool()> fallback) {
     return [res, fallback = std::move(fallback)]() -> bool {
         if (res->spipe) {
             return res->spipe->is_cancelled();

tools/server/server-stream.h

@@ -3,81 +3,23 @@
 #include "server-http.h"
 
 #include <atomic>
-#include <condition_variable>
 #include <cstddef>
-#include <cstdint>
 #include <functional>
 #include <memory>
-#include <mutex>
-#include <shared_mutex>
 #include <string>
-#include <thread>
-#include <unordered_map>
-#include <vector>
 
-enum class stream_read_status {
-    OK,
-    OFFSET_LOST,
-};
+// streaming buffer for one generation, survives HTTP disconnect. the producer appends SSE bytes,
+// readers drain from any offset via read_from. keyed by conversation_id, one conv = one live session
 
-// streaming buffer for one generation, survives HTTP disconnect. the producer appends raw SSE
-// bytes, readers drain from any offset via read_from and block until more bytes or finalize.
-// keyed by conversation_id: one conv = at most one live session
-struct stream_session {
-    std::string conversation_id;
-    int64_t     started_ts; // unix seconds at construction, used by /v1/streams listing
-
-    stream_session(std::string conversation_id_, size_t max_bytes_);
-    stream_session(const stream_session &)             = delete;
-    stream_session & operator=(const stream_session &) = delete;
-
-    // append raw bytes, drops from the front if the cap is reached.
-    // returns false if the session is already finalized
-    bool append(const char * data, size_t len);
-
-    // mark the session as complete, wakes all pending readers
-    void finalize();
-
-    // drain bytes from offset, calling sink for each chunk. blocks until more
-    // bytes arrive or finalize is called. returns OK on clean exit, OFFSET_LOST
-    // if offset falls below the dropped prefix
-    stream_read_status read_from(size_t offset,
-        const std::function<bool(const char *, size_t)> & sink,
-        const std::function<bool()> & should_stop);
-
-    bool    is_done() const;
-    bool    is_cancelled() const;
-    size_t  total_size() const;     // bytes that ever entered the session
-    size_t  dropped_prefix() const; // bytes evicted from the front due to cap
-    int64_t completed_at() const;   // 0 while alive, unix seconds after finalize
-
-    // attach the producer stop hook used to cancel its reader, pass an empty function to detach
-    void set_stop_producer(std::function<void()> fn);
-
-    // signal the producer to abort its inference asap via the stop hook, idempotent
-    void cancel();
-
-private:
-    mutable std::mutex      mu;
-    std::condition_variable cv;
-    std::vector<char>       buffer;
-    size_t                  prefix_dropped;
-    size_t                  cap_bytes;
-    std::atomic<bool>       done;
-    std::atomic<bool>       cancelled;
-    std::atomic<int64_t>    completed_ts;
-    std::function<void()>   stop_producer; // protected by mu
-};
+struct stream_session;
 
 using stream_session_ptr = std::shared_ptr<stream_session>;
 
-// one end of a stream_session pipe. the base holds the session and the shared query, the
-// producer and consumer ends derive from it. virtual dtor so each end runs its own teardown:
+// base of the producer/consumer pipe ends. virtual dtor so each runs its own teardown:
 // the producer finalizes the session, the consumer leaves it untouched
 struct stream_pipe {
     virtual ~stream_pipe() = default;
 
-    // true if the session was cancelled (e.g. via DELETE /v1/stream/<conv_id>)
     bool is_cancelled() const;
 
 protected:
@@ -95,7 +37,6 @@ protected:
 struct stream_pipe_producer : stream_pipe {
     ~stream_pipe_producer() override;
 
-    // append raw bytes to the session's ring buffer, returns false if already finalized
     bool write(const char * data, size_t len);
 
     // mark the natural end on the wire so a later close() is a no-op
@@ -121,83 +62,21 @@ private:
     server_http_res *                   res_ = nullptr;
 };
 
-// consumer end: read-only replay of the ring buffer, the destructor does not finalize the session
-struct stream_pipe_consumer : stream_pipe {
-    // drain bytes from offset, calling sink for each available chunk. blocks until more data
-    // arrives or the session finalizes. should_stop is polled, returns OFFSET_LOST if offset
-    // fell below the dropped prefix
-    stream_read_status read(size_t & offset,
-        const std::function<bool(const char *, size_t)> & sink,
-        const std::function<bool()> & should_stop);
+void server_stream_session_manager_start();
+void server_stream_session_manager_stop();
 
-    static std::shared_ptr<stream_pipe_consumer> create(stream_session_ptr session);
+// route handler factories wired under /v1/stream/* by server.cpp
+server_http_context::handler_t server_stream_make_get_handler();
+server_http_context::handler_t server_stream_make_lookup_handler();
+server_http_context::handler_t server_stream_make_delete_handler();
 
-private:
-    explicit stream_pipe_consumer(stream_session_ptr session);
-};
+// extract the X-Conversation-Id header value (case-insensitive), empty when absent
+std::string server_stream_conv_id_from_headers(const std::map<std::string, std::string> & headers);
 
-// owns all live sessions, runs a periodic GC to evict expired ones.
-// the map is keyed by conversation_id, so the invariant "one conv = at most one
-// live session" is enforced at the type level
-class stream_session_manager {
-public:
-    stream_session_manager();
-    ~stream_session_manager();
-
-    stream_session_manager(const stream_session_manager &)             = delete;
-    stream_session_manager & operator=(const stream_session_manager &) = delete;
-
-    // install a new session for this conversation, evicting and cancelling any previous one.
-    // the conversation_id must be non empty, the caller is responsible for that check.
-    // returns the new session
-    stream_session_ptr create_or_replace(const std::string & conversation_id);
-
-    // lookup, returns null if unknown or already evicted
-    stream_session_ptr get(const std::string & conversation_id);
-
-    // list every live or recently completed session, used by GET /v1/streams without filter
-    std::vector<stream_session_ptr> list_all() const;
-
-    // remove from the map and finalize, wakes any pending readers
-    void evict(const std::string & conversation_id);
-
-    // signal the producer to cancel asap then evict, used by the explicit user Stop path
-    void evict_and_cancel(const std::string & conversation_id);
-
-    void start_gc();
-    void stop_gc();
-
-private:
-    void gc_loop();
-
-    mutable std::shared_mutex                           map_mu;
-    std::unordered_map<std::string, stream_session_ptr> sessions; // key: conversation_id
-    std::thread                                         gc_thread;
-    std::atomic<bool>                                   running;
-    std::mutex                                          gc_wake_mu;
-    std::condition_variable                             gc_wake_cv;
-};
-
-// process wide manager, linked by both llama-server and llama-cli. llama-server main() drives
-// start_gc/stop_gc, llama-cli leaves it idle. the dtor calls stop_gc() unconditionally so exit
-// is safe whether or not the GC thread ran
-extern stream_session_manager g_stream_sessions;
-
-// route handler factories operating on g_stream_sessions, wired under /v1/stream/* by server.cpp.
-// keeps the resumable stream surface confined to server-stream
-server_http_context::handler_t make_stream_get_handler();
-server_http_context::handler_t make_streams_lookup_handler();
-server_http_context::handler_t make_stream_delete_handler();
-
-// extract the X-Conversation-Id header value (case-insensitive), empty when absent. exposed so
-// the router can track which child serves a forwarded POST
-std::string stream_conv_id_from_headers(const std::map<std::string, std::string> & headers);
-
-// on an X-Conversation-Id header, create or replace the session and attach a producer pipe to
-// res. no-op when absent, called from the server_res_generator constructor
-void stream_session_attach_pipe(server_http_res & res, const std::map<std::string, std::string> & headers);
+// on an X-Conversation-Id header, create or replace the session and attach a producer pipe to res
+void server_stream_session_attach_pipe(server_http_res & res, const std::map<std::string, std::string> & headers);
 
 // should_stop closure that ignores peer disconnect when a pipe is attached, so only an explicit
 // DELETE stops the producer and generation keeps flowing into the ring buffer. without a pipe it
 // delegates to fallback, the legacy non-resumable flow
-std::function<bool()> stream_aware_should_stop(server_http_res * res, std::function<bool()> fallback);
+std::function<bool()> server_stream_aware_should_stop(server_http_res * res, std::function<bool()> fallback);

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");
 
@@ -85,7 +95,7 @@ int llama_server(int argc, char ** argv) {
 
     // start the stream session manager GC right after common init, before any HTTP route can
     // touch it. lifecycle is symmetric, stop_gc() runs in clean_up() before backend free
-    g_stream_sessions.start_gc();
+    server_stream_session_manager_start();
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) {
         return 1;
@@ -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
@@ -245,8 +265,8 @@ int llama_server(int argc, char ** argv) {
     ctx_http.post("/slots/:id_slot",           ex_wrapper(routes.post_slots));
 
     // resumable streaming, the conversation_id is the session identity end to end. router and
-    // child wire different handlers under the same paths: a child binds the local g_stream_sessions
-    // backed factories, the router binds proxies that resolve the owning child through the
+    // child wire different handlers under the same paths: a child binds the local session
+    // factories, the router binds proxies that resolve the owning child through the
     // conv_id -> model map
     server_http_context::handler_t stream_get_h;
     server_http_context::handler_t streams_lookup_h;
@@ -256,9 +276,9 @@ int llama_server(int argc, char ** argv) {
         streams_lookup_h = models_routes->router_streams_lookup;
         stream_delete_h  = models_routes->router_stream_delete;
     } else {
-        stream_get_h     = make_stream_get_handler();
-        streams_lookup_h = make_streams_lookup_handler();
-        stream_delete_h  = make_stream_delete_handler();
+        stream_get_h     = server_stream_make_get_handler();
+        streams_lookup_h = server_stream_make_lookup_handler();
+        stream_delete_h  = server_stream_make_delete_handler();
     }
     ctx_http.get ("/v1/stream/:conv_id",       ex_wrapper(stream_get_h));
     // POST /v1/streams/lookup with body {"conversation_ids": [...]}. you can only ask for ids
@@ -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) {
@@ -343,7 +364,7 @@ int llama_server(int argc, char ** argv) {
         clean_up = [&models_routes]() {
             SRV_INF("%s: cleaning up before exit...\n", __func__);
             // stop the session GC first, it finalizes live sessions and wakes pending readers
-            g_stream_sessions.stop_gc();
+            server_stream_session_manager_stop();
             if (models_routes.has_value()) {
                 models_routes->stopping.store(true); // maybe redundant, but just to be safe
                 models_routes->models.unload_all();
@@ -371,7 +392,7 @@ int llama_server(int argc, char ** argv) {
         clean_up = [&ctx_http, &ctx_server]() {
             SRV_INF("%s: cleaning up before exit...\n", __func__);
             // stop the session GC first, it finalizes live sessions and wakes pending readers
-            g_stream_sessions.stop_gc();
+            server_stream_session_manager_stop();
             ctx_http.stop();
             ctx_server.terminate();
             llama_backend_free();
@@ -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());
 

tools/ui/package-lock.json

@@ -11,7 +11,7 @@
 				"@chromatic-com/storybook": "5.0.0",
 				"@eslint/compat": "1.4.1",
 				"@eslint/js": "9.39.2",
-				"@internationalized/date": "3.10.1",
+				"@internationalized/date": "3.12.2",
 				"@lucide/svelte": "0.515.0",
 				"@modelcontextprotocol/sdk": "1.26.0",
 				"@playwright/test": "1.56.1",
@@ -2981,9 +2981,9 @@
 			}
 		},
 		"node_modules/@internationalized/date": {
-			"version": "3.10.1",
-			"resolved": "https://registry.npmjs.org/@internationalized/date/-/date-3.10.1.tgz",
-			"integrity": "sha512-oJrXtQiAXLvT9clCf1K4kxp3eKsQhIaZqxEyowkBcsvZDdZkbWrVmnGknxs5flTD0VGsxrxKgBCZty1EzoiMzA==",
+			"version": "3.12.2",
+			"resolved": "https://registry.npmjs.org/@internationalized/date/-/date-3.12.2.tgz",
+			"integrity": "sha512-FY1Y+H64NDs+HAF6omlnWxm3mEpfgaCSWtL5l551ZZfImA+kGjPFgrnJrGjH6lfmLL0g8Z/mBu1R3kufeCp6Jw==",
 			"dev": true,
 			"license": "Apache-2.0",
 			"dependencies": {

tools/ui/package.json

@@ -30,7 +30,7 @@
 		"@chromatic-com/storybook": "5.0.0",
 		"@eslint/compat": "1.4.1",
 		"@eslint/js": "9.39.2",
-		"@internationalized/date": "3.10.1",
+		"@internationalized/date": "3.12.2",
 		"@lucide/svelte": "0.515.0",
 		"@modelcontextprotocol/sdk": "1.26.0",
 		"@playwright/test": "1.56.1",

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddDropdown.svelte

@@ -11,7 +11,8 @@
 	} from '$lib/constants';
 	import {
 		ChatFormActionAddToolsSubmenu,
-		ChatFormActionAddMcpServersSubmenu
+		ChatFormActionAddMcpServersSubmenu,
+		ChatFormActionAddReasoningSubmenu
 	} from '$lib/components/app';
 	import { useAttachmentMenu } from '$lib/hooks/use-attachment-menu.svelte';
 
@@ -92,7 +93,11 @@
 			</Tooltip.Content>
 		</Tooltip.Root>
 
-		<DropdownMenu.Content align="start" class="w-48">
+		<DropdownMenu.Content align="start" class="w-52">
+			<ChatFormActionAddReasoningSubmenu />
+
+			<DropdownMenu.Separator />
+
 			<DropdownMenu.Sub>
 				<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
 					<File class="h-4 w-4" />

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte

@@ -2,7 +2,7 @@
 	import { Lightbulb, LightbulbOff, Check, Info } from '@lucide/svelte';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { ReasoningEffort, MessageRole } from '$lib/enums';
+	import { ReasoningEffort } from '$lib/enums';
 	import { REASONING_EFFORT_TOKENS } from '$lib/constants/reasoning-effort-tokens';
 	import { REASONING_EFFORT_LEVELS } from '$lib/constants/reasoning-effort';
 	import type { ReasoningEffortLevel } from '$lib/types';
@@ -18,31 +18,23 @@
 	import { isRouterMode } from '$lib/stores/server.svelte';
 	import type { DatabaseMessage } from '$lib/types/database';
 
-	let thinkingEnabled = $derived(conversationsStore.getThinkingEnabled());
-	let currentEffort = $derived(conversationsStore.getReasoningEffort());
-	let isOff = $derived(!thinkingEnabled);
-	let tooltipText = $derived(thinkingEnabled ? `${currentEffort} Reasoning` : 'Disabled Reasoning');
 	let subOpen = $state(false);
 
-	// Get conversation model from message history
 	let conversationModel = $derived(
 		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
 	);
 
-	// Fallback: if model props aren't available, check if any assistant messages
-	// for this model in the active conversation have reasoning content.
 	let modelSupportsThinkingFromMessages = $derived.by(() => {
 		const modelId = isRouterMode() ? modelsStore.selectedModelName || conversationModel : null;
 		if (!modelId) return false;
+
 		const messages = conversationsStore.activeMessages;
+
 		return messages.some(
-			(m: DatabaseMessage) =>
-				m.role === MessageRole.ASSISTANT && m.model === modelId && !!m.reasoningContent
+			(m) => m.role === 'assistant' && m.model === modelId && !!m.reasoningContent
 		);
 	});
 
-	// Check if model supports thinking. Primary: chat template from /props.
-	// Fallback: message history (reasoning content in assistant messages).
 	let modelSupportsThinking = $derived.by(() => {
 		loadedModelIds();
 		propsCacheVersion();
@@ -52,15 +44,15 @@
 			return checkModelSupportsThinking(modelId ?? '') || modelSupportsThinkingFromMessages;
 		}
 
-		// In non-router mode, use the built-in supportsThinking
 		return supportsThinking() || modelSupportsThinkingFromMessages;
 	});
 
-	// Check if current item is selected
+	let thinkingEnabled = $derived(conversationsStore.getThinkingEnabled());
+	let currentEffort = $derived(conversationsStore.getReasoningEffort());
+	let isOff = $derived(!thinkingEnabled);
+
 	function isSelected(item: ReasoningEffortLevel): boolean {
-		if (item.isOff) {
-			return isOff;
-		}
+		if (item.isOff) return isOff;
 		return thinkingEnabled && currentEffort === item.value;
 	}
 
@@ -76,39 +68,30 @@
 </script>
 
 {#if modelSupportsThinking}
-	<DropdownMenu.Root bind:open={subOpen}>
-		<Tooltip.Root>
-			<Tooltip.Trigger>
-				<DropdownMenu.Trigger
-					class={[
-						'flex h-6 w-6 cursor-pointer items-center justify-center rounded-full p-0 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2',
-						thinkingEnabled ? 'bg-amber-400/10 hover:bg-amber-400/20' : 'bg-muted'
-					]}
-					aria-label={`${tooltipText}. Click to configure.`}
-				>
-					{#if thinkingEnabled}
-						<Lightbulb class="h-3 w-3 text-amber-400" />
-					{:else}
-						<LightbulbOff class="h-3 w-3 text-muted-foreground" />
-					{/if}
-				</DropdownMenu.Trigger>
-			</Tooltip.Trigger>
+	<DropdownMenu.Sub bind:open={subOpen}>
+		<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
+			{#if thinkingEnabled}
+				<Lightbulb class="h-4 w-4 shrink-0 text-amber-400" />
+			{:else}
+				<LightbulbOff class="h-4 w-4 shrink-0 text-muted-foreground" />
+			{/if}
 
-			<Tooltip.Content>
-				<p class="capitalize">{tooltipText}</p>
-			</Tooltip.Content>
-		</Tooltip.Root>
+			<span class="text-sm inline-flex gap-2 {!thinkingEnabled ? 'text-muted-foreground' : ''}">
+				Reasoning
 
-		<DropdownMenu.Content
-			align="start"
-			class="w-60 rounded-xl bg-popover p-3 text-popover-foreground shadow-md outline-none"
+				<span class="capitalize text-muted-foreground">
+					{thinkingEnabled ? currentEffort : 'off'}
+				</span>
+			</span>
+		</DropdownMenu.SubTrigger>
+
+		<DropdownMenu.SubContent
+			class="w-60 bg-popover p-1.5 text-popover-foreground shadow-md outline-none"
 		>
-			<div class="mb-2 px-2.5 text-sm font-medium">Reasoning effort</div>
-
 			{#each REASONING_EFFORT_LEVELS as level (level.value)}
 				<button
 					type="button"
-					class="flex w-full cursor-pointer items-center gap-2 rounded-lg px-2.5 py-2 text-left text-sm transition-colors hover:bg-accent"
+					class="flex w-full cursor-pointer items-center gap-3 rounded-md px-2 py-1.75 text-left text-sm transition-colors hover:bg-accent"
 					class:bg-accent={isSelected(level)}
 					onclick={() => handleSelection(level)}
 				>
@@ -140,6 +123,6 @@
 					{/if}
 				</button>
 			{/each}
-		</DropdownMenu.Content>
-	</DropdownMenu.Root>
+		</DropdownMenu.SubContent>
+	</DropdownMenu.Sub>
 {/if}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte

@@ -7,14 +7,20 @@
 		ChatFormActionModels,
 		ChatFormActionRecord,
 		ChatFormActionSubmit,
-		ChatFormReasoningToggle
+		ChatFormContextGauge
 	} from '$lib/components/app';
-	import { FileTypeCategory } from '$lib/enums';
+	import { FileTypeCategory, MessageRole } from '$lib/enums';
 	import { mcpStore } from '$lib/stores/mcp.svelte';
 	import { config } from '$lib/stores/settings.svelte';
-	import { conversationsStore } from '$lib/stores/conversations.svelte';
+	import { activeMessages, conversationsStore } from '$lib/stores/conversations.svelte';
+	import {
+		activeProcessingState,
+		isChatStreaming,
+		isLoading as chatIsLoading
+	} from '$lib/stores/chat.svelte';
 	import { getFileTypeCategory } from '$lib/utils';
 	import { goto } from '$app/navigation';
+	import { page } from '$app/state';
 	import { ROUTES } from '$lib/constants/routes';
 
 	interface Props {
@@ -93,6 +99,36 @@
 	let activeMessage = $derived(
 		conversationsStore.activeMessages[conversationsStore.activeMessages.length - 1]
 	);
+
+	let hasProcessedTokens = $derived.by(() => {
+		if (!page.params.id) return false;
+
+		const messages = activeMessages() as DatabaseMessage[];
+		let totalHistoricalTokens = 0;
+		for (const m of messages) {
+			if (m.role !== MessageRole.ASSISTANT) continue;
+			const timings = m.timings;
+			if (!timings) continue;
+			const agenticLlm = timings.agentic?.llm;
+			if (agenticLlm?.prompt_n != null || agenticLlm?.predicted_n != null) {
+				totalHistoricalTokens += (agenticLlm?.prompt_n ?? 0) + (agenticLlm?.predicted_n ?? 0);
+			} else {
+				totalHistoricalTokens += (timings.prompt_n ?? 0) + (timings.predicted_n ?? 0);
+			}
+		}
+		if (totalHistoricalTokens > 0) return true;
+
+		if (!chatIsLoading() && !isChatStreaming()) return false;
+
+		const processingState = activeProcessingState();
+		if (!processingState) return false;
+		const livePromptTokens = Math.max(
+			processingState.promptTokens ?? 0,
+			processingState.promptProgress?.processed ?? 0
+		);
+		const liveOutputTokens = processingState.outputTokensUsed ?? 0;
+		return livePromptTokens > 0 || liveOutputTokens > 0;
+	});
 </script>
 
 <div
@@ -100,7 +136,7 @@
 	style="container-type: inline-size"
 >
 	{#if showAddButton}
-		<div class="mr-auto flex items-center gap-3">
+		<div class="mr-auto flex items-center gap-2">
 			<ChatFormActionsAdd
 				{disabled}
 				{hasAudioModality}
@@ -117,8 +153,10 @@
 		</div>
 	{/if}
 
-	<div class="flex items-center gap-2">
-		<ChatFormReasoningToggle />
+	<div class="flex items-center gap-1.5">
+		{#if hasProcessedTokens}
+			<ChatFormContextGauge />
+		{/if}
 
 		{#if showModelSelector}
 			<ChatFormActionModels

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormReasoningEffortSubmenu.svelte

@@ -6,6 +6,7 @@
 	import { REASONING_EFFORT_TOKENS } from '$lib/constants/reasoning-effort-tokens';
 	import { REASONING_EFFORT_LEVELS } from '$lib/constants/reasoning-effort';
 	import type { ReasoningEffortLevel } from '$lib/types';
+	import { DIALOG_SUBMENU_CONTENT } from '$lib/constants/css-classes';
 	import {
 		modelsStore,
 		checkModelSupportsThinking,
@@ -71,9 +72,7 @@
 
 {#if modelSupportsThinking}
 	<DropdownMenu.Sub bind:open={subOpen}>
-		<DropdownMenu.SubTrigger
-			class="flex cursor-pointer items-center gap-2 rounded-md px-2.5 py-1.5 text-sm transition-colors outline-none hover:bg-accent focus:bg-accent"
-		>
+		<DropdownMenu.SubTrigger class="flex cursor-pointer items-center gap-2">
 			{#if thinkingEnabled}
 				<Lightbulb class="h-4 w-4 shrink-0 text-amber-400" />
 			{:else}
@@ -89,23 +88,15 @@
 			{/if}
 		</DropdownMenu.SubTrigger>
 
-		<DropdownMenu.SubContent
-			class="w-60 rounded-xl bg-popover p-3 text-popover-foreground shadow-md outline-none data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95"
-		>
+		<DropdownMenu.SubContent class={DIALOG_SUBMENU_CONTENT}>
 			{#each REASONING_EFFORT_LEVELS as level (level.value)}
 				<button
 					type="button"
-					class="flex w-full cursor-pointer items-center gap-2 rounded-lg px-2.5 py-2 text-left text-sm transition-colors hover:bg-accent"
+					class="flex w-full cursor-pointer items-center gap-2"
 					class:bg-accent={isSelected(level)}
 					onclick={() => handleSelection(level)}
 				>
-					{#if isSelected(level)}
-						<Check class="h-4 w-4 shrink-0 text-foreground" />
-					{:else}
-						<div class="h-4 w-4 shrink-0"></div>
-					{/if}
-
-					<span class="flex-1">{level.label}</span>
+					<span class="flex-1 text-left">{level.label}</span>
 
 					{#if !level.isOff}
 						<span class="text-[11px] text-muted-foreground opacity-60">
@@ -125,6 +116,10 @@
 							</Tooltip.Content>
 						</Tooltip.Root>
 					{/if}
+
+					{#if isSelected(level)}
+						<Check class="h-4 w-4 shrink-0 text-foreground" />
+					{/if}
 				</button>
 			{/each}
 		</DropdownMenu.SubContent>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ChatFormContextGauge.svelte

@@ -0,0 +1,108 @@
+<script lang="ts">
+	import { untrack } from 'svelte';
+	import * as HoverCard from '$lib/components/ui/hover-card';
+	import { activeConversation, activeMessages } from '$lib/stores/conversations.svelte';
+	import { chatStore, isChatStreaming, isLoading } from '$lib/stores/chat.svelte';
+	import { formatParameters } from '$lib/utils/formatters';
+	import { useContextGauge } from '$lib/hooks/use-context-gauge.svelte';
+	import ContextGaugeDial from './ContextGaugeDial.svelte';
+	import ContextGaugeDetails from './ContextGaugeDetails.svelte';
+	import ContextGaugeLoadModel from './ContextGaugeLoadModel.svelte';
+	import { colorLevelBgClass, colorLevelTextClass } from './context-gauge';
+
+	const gauge = useContextGauge();
+
+	$effect(() => {
+		const conv = activeConversation();
+		untrack(() => chatStore.setActiveProcessingConversation(conv?.id ?? null));
+	});
+
+	$effect(() => {
+		const conv = activeConversation();
+		const messages = activeMessages() as DatabaseMessage[];
+		if (!conv) return;
+		if (isLoading() || isChatStreaming()) return;
+
+		if (messages.length === 0) {
+			untrack(() => chatStore.clearProcessingState(conv.id));
+			return;
+		}
+
+		untrack(() => chatStore.restoreProcessingStateFromMessages(messages, conv.id));
+	});
+
+	$effect(() => {
+		gauge.startMonitoring();
+	});
+
+	const showProgressBar = $derived(
+		gauge.contextTotal !== null &&
+			gauge.contextTotal > 0 &&
+			(gauge.activeModelId !== null || gauge.isActiveModelLoaded)
+	);
+</script>
+
+<HoverCard.Root>
+	<HoverCard.Trigger class="flex h-5 w-5 cursor-default items-center justify-center">
+		<ContextGaugeDial percent={gauge.contextPercent} level={gauge.colorLevel} />
+	</HoverCard.Trigger>
+
+	<HoverCard.Content
+		side="bottom"
+		class="z-50 w-64 rounded-lg border border-border/50 bg-popover p-3 text-popover-foreground shadow-lg"
+	>
+		<div class="flex flex-col gap-2">
+			<div class="flex items-center gap-2">
+				<span class="font-medium">Context</span>
+				<span class="text-muted-foreground">·</span>
+				<span class="font-mono text-muted-foreground">
+					{formatParameters(gauge.contextUsed)}
+					/ {gauge.contextTotal !== null ? formatParameters(gauge.contextTotal) : '-'}
+				</span>
+			</div>
+
+			{#if gauge.activeModelId !== null && !gauge.isActiveModelLoaded}
+				<ContextGaugeLoadModel
+					modelId={gauge.activeModelId}
+					isLoading={gauge.isActiveModelLoading}
+					onLoad={gauge.loadModel}
+				/>
+			{:else if showProgressBar}
+				<div class="h-1.5 w-full overflow-hidden rounded-full bg-muted">
+					<div
+						class="h-full rounded-full transition-all duration-300 {colorLevelBgClass(
+							gauge.colorLevel
+						)}"
+						style="width: {gauge.contextPercent}%"
+					></div>
+				</div>
+
+				<div class="flex justify-between text-xs text-muted-foreground">
+					<span>
+						<span class={colorLevelTextClass(gauge.colorLevel)}>{gauge.contextPercent}%</span> used
+					</span>
+					<span>
+						{formatParameters((gauge.contextTotal ?? 0) - gauge.contextUsed)} remaining
+					</span>
+				</div>
+			{:else}
+				<div class="text-xs text-muted-foreground">No context info available</div>
+			{/if}
+
+			{#if gauge.hasAnyUsage}
+				<ContextGaugeDetails
+					currentRead={gauge.currentRead}
+					currentFresh={gauge.currentFresh}
+					currentCache={gauge.currentCache}
+					currentOutput={gauge.currentOutput}
+					kvTotal={gauge.kvTotal}
+					cumulativeRead={gauge.cumulativeRead}
+					cumulativeOutput={gauge.cumulativeOutput}
+					cumulativeCacheTotal={gauge.cumulativeCacheTotal}
+					averageTokensPerSecond={gauge.averageTokensPerSecond}
+					transientDetails={gauge.transientDetails}
+				/>
+			{/if}
+		</div>
+	</HoverCard.Content>
+</HoverCard.Root>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeDetailRow.svelte

@@ -0,0 +1,20 @@
+<script lang="ts">
+	interface Props {
+		label: string;
+		value: string;
+		subtitle?: string;
+	}
+
+	let { label, value, subtitle }: Props = $props();
+</script>
+
+<div class="grid gap-1.5">
+	<div class="flex items-baseline justify-between">
+		<span class="text-muted-foreground">{label}</span>
+		<span class="font-mono text-muted-foreground">{value}</span>
+	</div>
+
+	{#if subtitle}
+		<div class="text-[10px] leading-tight text-muted-foreground/70">{subtitle}</div>
+	{/if}
+</div>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeDetails.svelte

@@ -0,0 +1,122 @@
+<script lang="ts">
+	import { ChevronDown } from '@lucide/svelte';
+	import * as Collapsible from '$lib/components/ui/collapsible';
+	import { STATS_UNITS } from '$lib/constants';
+	import ContextGaugeDetailRow from './ContextGaugeDetailRow.svelte';
+
+	interface Props {
+		currentRead: number;
+		currentFresh: number;
+		currentCache: number;
+		currentOutput: number;
+		kvTotal: number;
+		cumulativeRead: number;
+		cumulativeOutput: number;
+		cumulativeCacheTotal: number;
+		averageTokensPerSecond: number | null;
+		transientDetails: string[];
+	}
+
+	let {
+		currentRead,
+		currentFresh,
+		currentCache,
+		currentOutput,
+		kvTotal,
+		cumulativeRead,
+		cumulativeOutput,
+		cumulativeCacheTotal,
+		averageTokensPerSecond,
+		transientDetails
+	}: Props = $props();
+
+	let open = $state(false);
+
+	const hasCumulative = $derived(cumulativeRead > 0 || cumulativeOutput > 0);
+	const hasCurrent = $derived(currentRead > 0 || currentOutput > 0);
+</script>
+
+<Collapsible.Root bind:open class="mt-3 border-t border-border/50 pt-4">
+	<Collapsible.Trigger
+		class="flex w-full cursor-pointer items-center gap-1 text-xs text-muted-foreground hover:text-foreground"
+	>
+		<span>Token usage details</span>
+
+		<ChevronDown class={'ml-auto h-3 w-3 transition-transform' + (open ? ' rotate-180' : '')} />
+	</Collapsible.Trigger>
+
+	<Collapsible.Content class="flex flex-col gap-4 text-xs pt-4">
+		{#if hasCumulative}
+			<div>
+				<h3 class="text-[11px] font-medium uppercase tracking-wide text-muted-foreground/70 mb-2">
+					Across all turns
+				</h3>
+
+				<div class="flex flex-col gap-2">
+					{#if cumulativeRead > 0}
+						<ContextGaugeDetailRow
+							label="Prompt tokens evaluated"
+							value={`${cumulativeRead.toLocaleString()} tok`}
+							subtitle={cumulativeCacheTotal > 0
+								? `${cumulativeCacheTotal.toLocaleString()} reused from KV cache`
+								: undefined}
+						/>
+					{/if}
+					{#if cumulativeOutput > 0}
+						<ContextGaugeDetailRow
+							label="Tokens generated"
+							value={`${cumulativeOutput.toLocaleString()} tok`}
+						/>
+					{/if}
+				</div>
+			</div>
+		{/if}
+
+		{#if hasCurrent}
+			<div>
+				<h3 class="text-[11px] font-medium uppercase tracking-wide text-muted-foreground/70 mb-2">
+					This turn · KV cache
+				</h3>
+
+				<div class="flex flex-col gap-2">
+					{#if currentRead > 0}
+						<ContextGaugeDetailRow
+							label="Prompt"
+							value={`${currentRead.toLocaleString()} tok`}
+							subtitle={currentCache > 0
+								? `${currentFresh.toLocaleString()} fresh + ${currentCache.toLocaleString()} cached`
+								: undefined}
+						/>
+					{/if}
+
+					{#if currentOutput > 0}
+						<ContextGaugeDetailRow
+							label="Generated"
+							value={`${currentOutput.toLocaleString()} tok`}
+						/>
+					{/if}
+
+					<div class="pt-1 mt-0.5 border-t border-border/30">
+						<div class="flex justify-between">
+							<span class="text-muted-foreground">KV cache total</span>
+							<span class="font-mono font-medium">{kvTotal.toLocaleString()} tok</span>
+						</div>
+					</div>
+				</div>
+			</div>
+		{/if}
+
+		{#if averageTokensPerSecond !== null}
+			<div class="pt-1.5 mt-1 border-t border-border/30">
+				<ContextGaugeDetailRow
+					label="Avg speed"
+					value={`${averageTokensPerSecond.toFixed(1)}${STATS_UNITS.TOKENS_PER_SECOND}`}
+				/>
+			</div>
+		{/if}
+
+		{#each transientDetails as detail (detail)}
+			<div class="font-mono text-muted-foreground">{detail}</div>
+		{/each}
+	</Collapsible.Content>
+</Collapsible.Root>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeDial.svelte

@@ -0,0 +1,43 @@
+<script lang="ts">
+	import type { ColorLevel } from './context-gauge';
+	import { colorLevelTextClass } from './context-gauge';
+
+	interface Props {
+		percent: number | null;
+		level: ColorLevel;
+		size?: 'sm' | 'md';
+	}
+
+	let { percent, level, size = 'sm' }: Props = $props();
+
+	const RADIUS = 11;
+	const CIRCUMFERENCE = 2 * Math.PI * RADIUS;
+
+	const strokeLevelClass = $derived(colorLevelTextClass(level));
+	const dimensions = $derived(size === 'md' ? 'h-6 w-6' : 'h-5 w-5');
+	const strokeWidth = $derived(size === 'md' ? 4 : 3);
+</script>
+
+<svg viewBox="0 0 32 32" fill="none" class={dimensions}>
+	<circle
+		cx="16"
+		cy="16"
+		r={RADIUS}
+		stroke="currentColor"
+		stroke-opacity="0.1"
+		stroke-width={strokeWidth}
+	/>
+
+	<circle
+		cx="16"
+		cy="16"
+		r={RADIUS}
+		class="transition-colors duration-300 {strokeLevelClass}"
+		stroke="currentColor"
+		stroke-width={strokeWidth}
+		stroke-linecap="round"
+		stroke-dasharray={CIRCUMFERENCE}
+		stroke-dashoffset={percent !== null ? CIRCUMFERENCE * (1 - percent / 100) : CIRCUMFERENCE}
+		transform="rotate(-90 16 16)"
+	/>
+</svg>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/ContextGaugeLoadModel.svelte

@@ -0,0 +1,24 @@
+<script lang="ts">
+	import { Loader2 } from '@lucide/svelte';
+	import { Button } from '$lib/components/ui/button';
+
+	interface Props {
+		modelId: string | null;
+		isLoading: boolean;
+		onLoad: () => void;
+	}
+
+	let { modelId, isLoading, onLoad }: Props = $props();
+</script>
+
+{#if modelId !== null && !isLoading}
+	<div class="flex flex-col gap-2 border-t border-border/50 pt-2 text-xs text-muted-foreground">
+		<span>Available context size is only visible once the model is loaded.</span>
+		<Button size="sm" variant="secondary" class="self-start" onclick={onLoad}>Load model</Button>
+	</div>
+{:else if isLoading}
+	<div class="flex items-center gap-2 border-t border-border/50 pt-2 text-xs text-muted-foreground">
+		<Loader2 class="h-3.5 w-3.5 animate-spin" />
+		<span>Loading model...</span>
+	</div>
+{/if}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormContextGauge/context-gauge.ts

@@ -0,0 +1,37 @@
+export type ColorLevel = 'ok' | 'warning' | 'critical' | 'neutral';
+
+const WARNING_THRESHOLD = 80;
+const CRITICAL_THRESHOLD = 95;
+
+export function colorLevelFromPercent(percent: number | null): ColorLevel {
+	if (percent === null) return 'neutral';
+	if (percent >= CRITICAL_THRESHOLD) return 'critical';
+	if (percent >= WARNING_THRESHOLD) return 'warning';
+	return 'ok';
+}
+
+export function colorLevelTextClass(level: ColorLevel): string {
+	switch (level) {
+		case 'critical':
+			return 'text-red-400';
+		case 'warning':
+			return 'text-amber-400';
+		case 'ok':
+			return 'text-muted-foreground';
+		default:
+			return 'text-muted-foreground';
+	}
+}
+
+export function colorLevelBgClass(level: ColorLevel): string {
+	switch (level) {
+		case 'critical':
+			return 'bg-red-500';
+		case 'warning':
+			return 'bg-amber-500';
+		case 'ok':
+			return 'bg-green-500';
+		default:
+			return 'bg-muted';
+	}
+}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessage.svelte

@@ -24,6 +24,8 @@
 		message: DatabaseMessage;
 		toolMessages?: DatabaseMessage[];
 		isLastAssistantMessage?: boolean;
+		isLastUserMessage?: boolean;
+		nextAssistantMessage?: DatabaseMessage | null;
 		siblingInfo?: ChatMessageSiblingInfo | null;
 	}
 
@@ -32,6 +34,8 @@
 		message,
 		toolMessages = [],
 		isLastAssistantMessage = false,
+		isLastUserMessage = false,
+		nextAssistantMessage = null,
 		siblingInfo = null
 	}: Props = $props();
 
@@ -359,7 +363,9 @@
 		<ChatMessageUser
 			class={className}
 			{deletionInfo}
+			{isLastUserMessage}
 			{message}
+			{nextAssistantMessage}
 			onConfirmDelete={handleConfirmDelete}
 			onCopy={handleCopy}
 			onDelete={handleDelete}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte

@@ -11,11 +11,10 @@
 	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
 	import { isLoading, isChatStreaming } from '$lib/stores/chat.svelte';
 	import { copyToClipboard, deriveAgenticSections, modelLoadProgressText } from '$lib/utils';
-	import { AgenticSectionType } from '$lib/enums';
+	import { AgenticSectionType, ChatMessageStatisticsMode } from '$lib/enums';
 	import { REASONING_TAGS } from '$lib/constants/agentic';
-	import { tick } from 'svelte';
 	import { fade } from 'svelte/transition';
-	import { MessageRole, ChatMessageStatsView } from '$lib/enums';
+	import { MessageRole } from '$lib/enums';
 	import { config } from '$lib/stores/settings.svelte';
 	import { isRouterMode } from '$lib/stores/server.svelte';
 	import { modelsStore } from '$lib/stores/models.svelte';
@@ -122,62 +121,6 @@
 		return parts.join('\n\n\n');
 	});
 
-	let activeStatsView = $state<ChatMessageStatsView>(ChatMessageStatsView.GENERATION);
-	let statsContainerEl: HTMLDivElement | undefined = $state();
-
-	function getScrollParent(el: HTMLElement): HTMLElement | null {
-		let parent = el.parentElement;
-		while (parent) {
-			const style = getComputedStyle(parent);
-			if (/(auto|scroll)/.test(style.overflowY)) {
-				return parent;
-			}
-			parent = parent.parentElement;
-		}
-		return null;
-	}
-
-	async function handleStatsViewChange(view: ChatMessageStatsView) {
-		const el = statsContainerEl;
-		if (!el) {
-			activeStatsView = view;
-
-			return;
-		}
-
-		const scrollParent = getScrollParent(el);
-		if (!scrollParent) {
-			activeStatsView = view;
-
-			return;
-		}
-
-		const yBefore = el.getBoundingClientRect().top;
-
-		activeStatsView = view;
-
-		await tick();
-
-		const delta = el.getBoundingClientRect().top - yBefore;
-		if (delta !== 0) {
-			scrollParent.scrollTop += delta;
-		}
-
-		// Correct any drift after browser paint
-		requestAnimationFrame(() => {
-			const drift = el.getBoundingClientRect().top - yBefore;
-
-			if (Math.abs(drift) > 1) {
-				scrollParent.scrollTop += drift;
-			}
-		});
-	}
-
-	let highlightAgenticTurns = $derived(
-		isAgentic &&
-			(currentConfig.alwaysShowAgenticTurns || activeStatsView === ChatMessageStatsView.SUMMARY)
-	);
-
 	let displayedModel = $derived(message.model ?? null);
 
 	// model being switched to while it loads, so the selector bar tracks it
@@ -291,7 +234,6 @@
 				{toolMessages}
 				isStreaming={isChatStreaming()}
 				{isLastAssistantMessage}
-				highlightTurns={highlightAgenticTurns}
 			/>
 		{/if}
 	{:else}
@@ -315,10 +257,7 @@
 
 	<div class="info my-6 grid gap-4 tabular-nums">
 		{#if displayedModel}
-			<div
-				bind:this={statsContainerEl}
-				class="inline-flex flex-wrap items-start gap-2 text-xs text-muted-foreground"
-			>
+			<div class="inline-flex flex-wrap items-start gap-2 text-xs text-muted-foreground">
 				{#if isRouter}
 					<ModelsSelectorDropdown
 						currentModel={pendingModel ?? displayedModel}
@@ -347,28 +286,25 @@
 				{#if currentConfig.showMessageStats && message.timings && message.timings.predicted_n && message.timings.predicted_ms}
 					{@const agentic = message.timings.agentic}
 					<ChatMessageStatistics
+						mode={ChatMessageStatisticsMode.GENERATION}
 						promptTokens={agentic ? agentic.llm.prompt_n : message.timings.prompt_n}
 						promptMs={agentic ? agentic.llm.prompt_ms : message.timings.prompt_ms}
 						predictedTokens={agentic ? agentic.llm.predicted_n : message.timings.predicted_n}
 						predictedMs={agentic ? agentic.llm.predicted_ms : message.timings.predicted_ms}
 						agenticTimings={agentic}
-						onActiveViewChange={handleStatsViewChange}
 					/>
 				{:else if isLoading() && currentConfig.showMessageStats}
 					{@const liveStats = processingState.getLiveProcessingStats()}
 					{@const genStats = processingState.getLiveGenerationStats()}
-					{@const promptProgress = processingState.processingState?.promptProgress}
-					{@const isStillProcessingPrompt =
-						promptProgress && promptProgress.processed < promptProgress.total}
 
-					{#if liveStats || genStats}
+					{#if genStats}
 						<ChatMessageStatistics
+							mode={ChatMessageStatisticsMode.GENERATION}
 							isLive
-							isProcessingPrompt={!!isStillProcessingPrompt}
 							promptTokens={liveStats?.tokensProcessed}
 							promptMs={liveStats?.timeMs}
-							predictedTokens={genStats?.tokensGenerated}
-							predictedMs={genStats?.timeMs}
+							predictedTokens={genStats.tokensGenerated}
+							predictedMs={genStats.timeMs}
 						/>
 					{/if}
 				{/if}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUser.svelte

@@ -2,10 +2,14 @@
 	import {
 		ChatMessageActionIcons,
 		ChatMessageEditForm,
+		ChatMessageStatistics,
 		ChatMessageUserBubble
 	} from '$lib/components/app/chat';
 	import { getMessageEditContext } from '$lib/contexts';
-	import { MessageRole } from '$lib/enums';
+	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
+	import { isLoading } from '$lib/stores/chat.svelte';
+	import { MessageRole, ChatMessageStatisticsMode } from '$lib/enums';
+	import { config } from '$lib/stores/settings.svelte';
 
 	interface Props {
 		class?: string;
@@ -17,6 +21,8 @@
 			assistantMessages: number;
 			messageTypes: string[];
 		} | null;
+		isLastUserMessage?: boolean;
+		nextAssistantMessage?: DatabaseMessage | null;
 		showDeleteDialog: boolean;
 		onEdit: () => void;
 		onDelete: () => void;
@@ -32,6 +38,8 @@
 		message,
 		siblingInfo = null,
 		deletionInfo,
+		isLastUserMessage = false,
+		nextAssistantMessage = null,
 		showDeleteDialog,
 		onEdit,
 		onDelete,
@@ -44,6 +52,37 @@
 
 	// Get contexts
 	const editCtx = getMessageEditContext();
+	const processingState = useProcessingState();
+
+	const currentConfig = $derived(config());
+	const isActivelyProcessing = $derived(isLastUserMessage && isLoading());
+
+	// For agentic turns, prefer the cumulative agentic.llm totals over per-call timings.
+	let storedReadingStats = $derived.by(() => {
+		const timings = nextAssistantMessage?.timings;
+		if (!timings?.prompt_n || !timings?.prompt_ms) return null;
+
+		const agentic = timings.agentic;
+
+		return {
+			promptTokens: agentic ? agentic.llm.prompt_n : timings.prompt_n,
+			promptMs: agentic ? agentic.llm.prompt_ms : timings.prompt_ms
+		};
+	});
+
+	let showStoredReadingStats = $derived(
+		Boolean(currentConfig.showMessageStats) && storedReadingStats !== null
+	);
+
+	let showLiveReadingStats = $derived(
+		Boolean(currentConfig.showMessageStats) && isActivelyProcessing && storedReadingStats === null
+	);
+
+	$effect(() => {
+		if (showLiveReadingStats) {
+			processingState.startMonitoring();
+		}
+	});
 </script>
 
 <div
@@ -60,6 +99,37 @@
 			renderMarkdown={true}
 		/>
 
+		{#if showStoredReadingStats}
+			<!-- Reading stats sourced from the assistant message that followed this turn -->
+			<div class="info my-2 grid w-full justify-items-end gap-4 tabular-nums">
+				<div
+					class="inline-flex flex-wrap items-start justify-end gap-2 text-xs text-muted-foreground"
+				>
+					<ChatMessageStatistics
+						mode={ChatMessageStatisticsMode.READING}
+						promptTokens={storedReadingStats!.promptTokens}
+						promptMs={storedReadingStats!.promptMs}
+					/>
+				</div>
+			</div>
+		{:else if showLiveReadingStats}
+			{@const liveStats = processingState.getLiveProcessingStats()}
+			{#if liveStats}
+				<div class="info my-2 grid w-full justify-items-end gap-4 tabular-nums">
+					<div
+						class="inline-flex flex-wrap items-start justify-end gap-2 text-xs text-muted-foreground"
+					>
+						<ChatMessageStatistics
+							mode={ChatMessageStatisticsMode.READING}
+							isLive
+							promptTokens={liveStats.tokensProcessed}
+							promptMs={liveStats.timeMs}
+						/>
+					</div>
+				</div>
+			{/if}
+		{/if}
+
 		{#if message.timestamp}
 			<div class="max-w-[80%]">
 				<ChatMessageActionIcons

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserPending.svelte

@@ -2,7 +2,6 @@
 	import { ActionIcon, ChatMessageEditForm, ChatMessageUserBubble } from '$lib/components/app';
 	import { fadeInView } from '$lib/actions/fade-in-view.svelte';
 	import { ArrowUp, Edit, Trash2 } from '@lucide/svelte';
-	import { getProcessingInfoContext } from '$lib/contexts';
 	import { useMessageEditContext } from '$lib/hooks/use-message-edit-context.svelte';
 
 	interface Props {
@@ -23,9 +22,6 @@
 		onDelete
 	}: Props = $props();
 
-	const processingInfoCtx = getProcessingInfoContext();
-	let showProcessingInfo = $derived(processingInfoCtx.showProcessingInfo);
-
 	const editCtx = useMessageEditContext({
 		getContent: () => content,
 		getExtras: () => extras,
@@ -36,9 +32,7 @@
 <div
 	use:fadeInView
 	aria-label="Pending user message"
-	class="group flex flex-col items-end gap-3 transition-opacity hover:opacity-80 md:gap-2 {className} sticky {showProcessingInfo
-		? 'bottom-44'
-		: 'bottom-32'}"
+	class="group flex flex-col items-end gap-3 transition-opacity hover:opacity-80 md:gap-2 {className} sticky bottom-32"
 	role="group"
 >
 	{#if editCtx.isEditing}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageAgenticContent.svelte

@@ -41,15 +41,13 @@
 		toolMessages?: DatabaseMessage[];
 		isStreaming?: boolean;
 		isLastAssistantMessage?: boolean;
-		highlightTurns?: boolean;
 	}
 
 	let {
 		message,
 		toolMessages = [],
 		isStreaming = false,
-		isLastAssistantMessage = false,
-		highlightTurns = false
+		isLastAssistantMessage = false
 	}: Props = $props();
 
 	let expandedStates: Record<number, boolean> = $state({});
@@ -57,6 +55,7 @@
 	const showToolCallInProgress = $derived(config().showToolCallInProgress as boolean);
 	const showThoughtInProgress = $derived(config().showThoughtInProgress as boolean);
 	const renderThinkingAsMarkdown = $derived(config().renderThinkingAsMarkdown as boolean);
+	const showMessageStats = $derived(config().showMessageStats as boolean);
 
 	const hasReasoningError = $derived(
 		isLastAssistantMessage ? !!agenticLastError(message.convId) : false
@@ -354,16 +353,17 @@
 {/snippet}
 
 <div class="agentic-content">
-	{#if highlightTurns && turnGroups.length > 1}
+	{#if turnGroups.length > 1}
 		{#each turnGroups as turn, turnIndex (turnIndex)}
 			{@const turnStats = message?.timings?.agentic?.perTurn?.[turnIndex]}
-			<div class="agentic-turn my-2 hover:bg-muted/80 dark:hover:bg-muted/30">
-				<span class="agentic-turn-label">Turn {turnIndex + 1}</span>
+
+			<div class="agentic-turn group/turn grid gap-3 mb-4">
 				{#each turn.sections as section, sIdx (turn.flatIndices[sIdx])}
 					{@render renderSection(section, turn.flatIndices[sIdx])}
 				{/each}
-				{#if turnStats}
-					<div class="turn-stats">
+
+				{#if turnStats && showMessageStats}
+					<div class="turn-stats transition-opacity duration-150">
 						<ChatMessageStatistics
 							promptTokens={turnStats.llm.prompt_n}
 							promptMs={turnStats.llm.prompt_ms}
@@ -402,39 +402,21 @@
 	.agentic-content {
 		display: flex;
 		flex-direction: column;
-		gap: 0.5rem;
 		width: 100%;
 		max-width: 48rem;
+		gap: 1rem;
+	}
+
+	.agentic-content > :global(*),
+	.agentic-turn > :global(*) {
+		min-width: 0;
 	}
 
 	.agentic-text {
 		width: 100%;
 	}
 
-	.agentic-turn {
-		position: relative;
-		border: 1.5px dashed var(--muted-foreground);
-		border-radius: 0.75rem;
-		padding: 1rem;
-		transition: background 0.1s;
-	}
-
-	.agentic-turn-label {
-		position: absolute;
-		top: -1rem;
-		left: 0.75rem;
-		padding: 0 0.375rem;
-		background: var(--background);
-		font-size: 0.7rem;
-		font-weight: 500;
-		color: var(--muted-foreground);
-		text-transform: uppercase;
-		letter-spacing: 0.05em;
-	}
-
 	.turn-stats {
-		margin-top: 0.75rem;
-		padding-top: 0.5rem;
 		border-top: 1px solid hsl(var(--muted) / 0.5);
 	}
 </style>

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics/ChatMessageStatistics.svelte

@@ -2,7 +2,7 @@
 	import { Clock, Gauge, WholeWord, BookOpenText, Sparkles, Wrench, Layers } from '@lucide/svelte';
 	import { ChatMessageStatisticsBadge } from '$lib/components/app';
 	import * as Tooltip from '$lib/components/ui/tooltip';
-	import { ChatMessageStatsView } from '$lib/enums';
+	import { ChatMessageStatsView, ChatMessageStatisticsMode } from '$lib/enums';
 	import type { ChatMessageAgenticTimings } from '$lib/types/chat';
 	import { formatPerformanceTime } from '$lib/utils';
 	import { MS_PER_SECOND, DEFAULT_PERFORMANCE_TIME } from '$lib/constants';
@@ -19,6 +19,7 @@
 		agenticTimings?: ChatMessageAgenticTimings;
 		onActiveViewChange?: (view: ChatMessageStatsView) => void;
 		hideSummary?: boolean;
+		mode?: ChatMessageStatisticsMode;
 	}
 
 	let {
@@ -31,19 +32,30 @@
 		initialView = ChatMessageStatsView.GENERATION,
 		agenticTimings,
 		onActiveViewChange,
-		hideSummary = false
+		hideSummary = false,
+		mode = ChatMessageStatisticsMode.SWITCHABLE
 	}: Props = $props();
 
-	let activeView: ChatMessageStatsView = $derived(initialView);
+	let isSwitchable = $derived(mode === ChatMessageStatisticsMode.SWITCHABLE);
+
+	let activeView: ChatMessageStatsView = $derived(
+		mode === ChatMessageStatisticsMode.READING
+			? ChatMessageStatsView.READING
+			: mode === ChatMessageStatisticsMode.GENERATION
+				? ChatMessageStatsView.GENERATION
+				: initialView
+	);
 	let hasAutoSwitchedToGeneration = $state(false);
 
 	$effect(() => {
-		onActiveViewChange?.(activeView);
+		if (isSwitchable) {
+			onActiveViewChange?.(activeView);
+		}
 	});
 
 	// In live mode: auto-switch to GENERATION tab when prompt processing completes
 	$effect(() => {
-		if (isLive) {
+		if (isLive && isSwitchable) {
 			// Auto-switch to generation tab only when prompt processing is done (once)
 			if (
 				!hasAutoSwitchedToGeneration &&
@@ -91,8 +103,7 @@
 			formattedPromptTime !== undefined
 	);
 
-	// In live mode, generation tab is disabled until we have generation stats
-	let isGenerationDisabled = $derived(isLive && !hasGenerationStats);
+	let isGenerationDisabled = $derived(isLive && isSwitchable && !hasGenerationStats);
 
 	let hasAgenticStats = $derived(agenticTimings !== undefined && agenticTimings.toolCallsCount > 0);
 
@@ -153,44 +164,44 @@
 {/snippet}
 
 <div class="inline-flex items-center text-xs text-muted-foreground">
-	<div class="inline-flex items-center rounded-sm bg-muted-foreground/15 p-0.5">
-		{#if hasPromptStats || isLive}
-			{@render viewButton({
-				view: ChatMessageStatsView.READING,
-				icon: BookOpenText,
-				label: 'Reading',
-				tooltipText: 'Reading (prompt processing)'
-			})}
-		{/if}
-
-		{@render viewButton({
-			view: ChatMessageStatsView.GENERATION,
-			icon: Sparkles,
-			label: 'Generation',
-			tooltipText: isGenerationDisabled
-				? 'Generation (waiting for tokens...)'
-				: 'Generation (token output)',
-			disabled: isGenerationDisabled
-		})}
-
-		{#if hasAgenticStats}
-			{@render viewButton({
-				view: ChatMessageStatsView.TOOLS,
-				icon: Wrench,
-				label: 'Tools',
-				tooltipText: 'Tool calls'
-			})}
-
-			{#if !hideSummary}
+	{#if isSwitchable}
+		<div class="inline-flex items-center rounded-sm bg-muted-foreground/15 p-0.5">
+			{#if hasPromptStats || isLive}
 				{@render viewButton({
-					view: ChatMessageStatsView.SUMMARY,
-					icon: Layers,
-					label: 'Summary',
-					tooltipText: 'Agentic summary'
+					view: ChatMessageStatsView.READING,
+					icon: BookOpenText,
+					label: 'Reading',
+					tooltipText: 'Processing'
 				})}
 			{/if}
-		{/if}
-	</div>
+
+			{@render viewButton({
+				view: ChatMessageStatsView.GENERATION,
+				icon: Sparkles,
+				label: 'Generation',
+				tooltipText: isGenerationDisabled ? 'Waiting for tokens...' : 'Generation',
+				disabled: isGenerationDisabled
+			})}
+
+			{#if hasAgenticStats}
+				{@render viewButton({
+					view: ChatMessageStatsView.TOOLS,
+					icon: Wrench,
+					label: 'Tools',
+					tooltipText: 'Tool calls'
+				})}
+
+				{#if !hideSummary}
+					{@render viewButton({
+						view: ChatMessageStatsView.SUMMARY,
+						icon: Layers,
+						label: 'Summary',
+						tooltipText: 'Agentic summary'
+					})}
+				{/if}
+			{/if}
+		</div>
+	{/if}
 
 	<div class="flex items-center gap-1 px-2">
 		{#if activeView === ChatMessageStatsView.GENERATION && hasGenerationStats}
@@ -256,7 +267,7 @@
 				value={formattedAgenticTotalTime}
 				tooltipLabel="Total time (LLM + tools)"
 			/>
-		{:else if hasPromptStats}
+		{:else if hasPromptStats && (mode === ChatMessageStatisticsMode.READING || isSwitchable)}
 			<ChatMessageStatisticsBadge
 				class="bg-transparent"
 				icon={WholeWord}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessages.svelte

@@ -186,6 +186,8 @@
 			message: DatabaseMessage;
 			toolMessages: DatabaseMessage[];
 			isLastAssistantMessage: boolean;
+			isLastUserMessage: boolean;
+			nextAssistantMessage: DatabaseMessage | null;
 			siblingInfo: ChatMessageSiblingInfo;
 		}> = [];
 
@@ -236,18 +238,36 @@
 				message: msg,
 				toolMessages,
 				isLastAssistantMessage: false,
+				isLastUserMessage: false,
+				nextAssistantMessage: null,
 				siblingInfo
 			});
 		}
 
-		// Mark the last assistant message
+		let lastAssistantIdx = -1;
 		for (let i = result.length - 1; i >= 0; i--) {
 			if (result[i].message.role === MessageRole.ASSISTANT) {
 				result[i].isLastAssistantMessage = true;
+				lastAssistantIdx = i;
 				break;
 			}
 		}
 
+		if (lastAssistantIdx > 0 && result[lastAssistantIdx - 1].message.role === MessageRole.USER) {
+			result[lastAssistantIdx - 1].isLastUserMessage = true;
+		}
+
+		for (let i = 0; i < result.length; i++) {
+			if (result[i].message.role !== MessageRole.USER) continue;
+
+			for (let j = i + 1; j < result.length; j++) {
+				if (result[j].message.role === MessageRole.ASSISTANT) {
+					result[i].nextAssistantMessage = result[j].message;
+					break;
+				}
+			}
+		}
+
 		return result;
 	});
 </script>
@@ -257,12 +277,14 @@
 		{isVisible ? 'opacity-100' : 'opacity-0'}
 		{previousRouteId === '/(chat)/chat/[id]' ? '' : 'delay-300'}"
 >
-	{#each displayMessages as { message, toolMessages, isLastAssistantMessage, siblingInfo } (message.id)}
+	{#each displayMessages as { message, toolMessages, isLastAssistantMessage, isLastUserMessage, nextAssistantMessage, siblingInfo } (message.id)}
 		<ChatMessage
 			class="mx-auto mt-12 w-full max-w-3xl"
 			{message}
 			{toolMessages}
 			{isLastAssistantMessage}
+			{isLastUserMessage}
+			{nextAssistantMessage}
 			{siblingInfo}
 		/>
 	{/each}

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

@@ -4,12 +4,10 @@
 		ChatScreenForm,
 		ChatMessages,
 		ChatScreenDragOverlay,
-		ChatScreenProcessingInfo,
 		ChatScreenStreamResumeStatus,
 		ServerLoadingSplash,
 		ChatScreenServerError
 	} from '$lib/components/app';
-	import { setProcessingInfoContext } from '$lib/contexts';
 	import { createAutoScrollController } from '$lib/hooks/use-auto-scroll.svelte';
 	import { useChatScreenActiveModel } from '$lib/hooks/use-chat-screen-active-model.svelte';
 	import { useChatScreenDragAndDrop } from '$lib/hooks/use-chat-screen-drag-and-drop.svelte';
@@ -23,8 +21,7 @@
 		errorDialog,
 		isLoading,
 		isChatStreaming,
-		isEditing,
-		activeProcessingState
+		isEditing
 	} from '$lib/stores/chat.svelte';
 	import {
 		conversationsStore,
@@ -42,12 +39,6 @@
 
 	let { showCenteredEmpty = false } = $props();
 
-	setProcessingInfoContext({
-		get showProcessingInfo() {
-			return showProcessingInfo;
-		}
-	});
-
 	let disableAutoScroll = $derived(Boolean(config().disableAutoScroll) || isMobile.current);
 	let isMobileUserScrolledUp = $state(false);
 	let mobileScrollDownHint = $state(false);
@@ -63,11 +54,6 @@
 	let isServerLoading = $derived(serverLoading());
 	let hasPropsError = $derived(!!serverError());
 	let isCurrentConversationLoading = $derived(isLoading() || isChatStreaming());
-	let showProcessingInfo = $derived(
-		isCurrentConversationLoading ||
-			(config().keepStatsVisible && !!page.params.id) ||
-			activeProcessingState() !== null
-	);
 	let chatFormBottomPosition = $derived.by(() => {
 		if (!isMobile.current) return '1rem';
 		if (device.isStandalone) return '1.5rem';
@@ -298,10 +284,6 @@
 						}}
 					/>
 				{/if}
-
-				{#if showProcessingInfo}
-					<ChatScreenProcessingInfo />
-				{/if}
 			</div>
 
 			<ChatScreenForm

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenProcessingInfo.svelte

@@ -1,127 +0,0 @@
-<script lang="ts">
-	import { untrack } from 'svelte';
-	import { PROCESSING_INFO_TIMEOUT } from '$lib/constants';
-	import { useProcessingState } from '$lib/hooks/use-processing-state.svelte';
-	import { chatStore, isLoading, isChatStreaming } from '$lib/stores/chat.svelte';
-	import { activeMessages, activeConversation } from '$lib/stores/conversations.svelte';
-	import { config } from '$lib/stores/settings.svelte';
-
-	const processingState = useProcessingState();
-
-	let isCurrentConversationLoading = $derived(isLoading());
-	let isStreaming = $derived(isChatStreaming());
-	let processingDetails = $derived(processingState.getTechnicalDetails());
-
-	let processingVisible = $derived(processingDetails.length > 0);
-
-	let { onVisibilityChange }: { onVisibilityChange?: (visible: boolean) => void } = $props();
-
-	$effect(() => {
-		onVisibilityChange?.(processingVisible);
-	});
-
-	$effect(() => {
-		const conversation = activeConversation();
-
-		untrack(() => chatStore.setActiveProcessingConversation(conversation?.id ?? null));
-	});
-
-	$effect(() => {
-		const keepStatsVisible = config().keepStatsVisible;
-		const shouldMonitor = keepStatsVisible || isCurrentConversationLoading || isStreaming;
-
-		if (shouldMonitor) {
-			processingState.startMonitoring();
-		}
-
-		if (!isCurrentConversationLoading && !isStreaming && !keepStatsVisible) {
-			const timeout = setTimeout(() => {
-				if (!config().keepStatsVisible && !isChatStreaming()) {
-					processingState.stopMonitoring();
-				}
-			}, PROCESSING_INFO_TIMEOUT);
-
-			return () => clearTimeout(timeout);
-		}
-	});
-
-	$effect(() => {
-		const conversation = activeConversation();
-		const messages = activeMessages() as DatabaseMessage[];
-		const keepStatsVisible = config().keepStatsVisible;
-
-		if (keepStatsVisible && conversation) {
-			if (messages.length === 0) {
-				untrack(() => chatStore.clearProcessingState(conversation.id));
-				return;
-			}
-
-			if (!isCurrentConversationLoading && !isStreaming) {
-				untrack(() => chatStore.restoreProcessingStateFromMessages(messages, conversation.id));
-			}
-		}
-	});
-</script>
-
-<div
-	class={[
-		'chat-processing-info-container pointer-events-none relative w-full hidden md:block',
-		processingVisible && 'visible'
-	]}
->
-	<div class="chat-processing-info-content absolute bottom-4 left-1/2 -translate-x-1/2">
-		{#each processingDetails as detail (detail)}
-			<span class="chat-processing-info-detail pointer-events-auto backdrop-blur-sm">{detail}</span>
-		{/each}
-	</div>
-</div>
-
-<style>
-	.chat-processing-info-container {
-		position: sticky;
-		top: 0;
-		z-index: 10;
-		padding: 0 1rem 0.75rem;
-		opacity: 0;
-		transform: translateY(50%);
-		transition:
-			opacity 300ms ease-out,
-			transform 300ms ease-out;
-	}
-
-	.chat-processing-info-container.visible {
-		opacity: 1;
-		transform: translateY(0);
-	}
-
-	.chat-processing-info-content {
-		display: flex;
-		flex-wrap: wrap;
-		align-items: center;
-		gap: 1rem;
-		justify-content: center;
-		max-width: 48rem;
-		margin: 0 auto;
-	}
-
-	.chat-processing-info-detail {
-		color: var(--muted-foreground);
-		font-size: 0.75rem;
-		padding: 0.25rem 0.75rem;
-		border-radius: 0.375rem;
-		font-family:
-			ui-monospace, SFMono-Regular, 'SF Mono', Consolas, 'Liberation Mono', Menlo, monospace;
-		white-space: nowrap;
-	}
-
-	@media (max-width: 768px) {
-		.chat-processing-info-content {
-			gap: 0.5rem;
-		}
-
-		.chat-processing-info-detail {
-			font-size: 0.7rem;
-			padding: 0.2rem 0.5rem;
-		}
-	}
-</style>

tools/ui/src/lib/components/app/chat/index.ts

@@ -241,13 +241,18 @@ export { default as ChatFormActionAddToolsSubmenu } from './ChatForm/ChatFormAct
 export { default as ChatFormActionAddMcpServersSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte';
 
 /**
- * **ChatFormReasoningToggle** - Thinking toggle button with effort dropdown
+ * Dropdown submenu for selecting reasoning effort level.
  *
- * A toggle button with lightbulb icon that indicates thinking status.
- * Shows the reasoning effort dropdown when clicked.
+ * Shows a "Reasoning" sub-menu item with a lightbulb icon indicating
+ * thinking status, and a nested list of effort levels.
  * Only visible when the current model supports thinking.
  */
-export { default as ChatFormReasoningToggle } from './ChatForm/ChatFormActions/ChatFormReasoningToggle.svelte';
+export { default as ChatFormActionAddReasoningSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte';
+
+/**
+ * Compact context-usage gauge with per-turn and cumulative breakdown in the tooltip.
+ */
+export { default as ChatFormContextGauge } from './ChatForm/ChatFormContextGauge/ChatFormContextGauge.svelte';
 
 /**
  * Hidden file input element for programmatic file selection.
@@ -669,14 +674,6 @@ export { default as ChatScreenDragOverlay } from './ChatScreen/ChatScreenDragOve
  */
 export { default as ChatScreenForm } from './ChatScreen/ChatScreenForm.svelte';
 
-/**
- * Processing info display during generation. Shows real-time statistics:
- * tokens per second, prompt/completion token counts, and elapsed time.
- * Data sourced from slotsService polling during active generation.
- * Only visible when `isCurrentConversationLoading` is true.
- */
-export { default as ChatScreenProcessingInfo } from './ChatScreen/ChatScreenProcessingInfo.svelte';
-
 /**
  * Server error alert displayed when the server is unreachable.
  * Shows the error message with a retry button.

tools/ui/src/lib/components/app/content/CollapsibleContentBlock.svelte

@@ -76,7 +76,7 @@
 		open = value;
 		onToggle?.();
 	}}
-	class={className}
+	class="{className} my-0!"
 >
 	<Card class="gap-0 border-muted bg-muted/30 py-0">
 		<Collapsible.Trigger class="flex w-full cursor-pointer items-start justify-between gap-2 p-3">

tools/ui/src/lib/components/app/content/SyntaxHighlightedCode.svelte

@@ -72,8 +72,8 @@
 </script>
 
 <div
-	class="code-preview-wrapper rounded-lg border border-border bg-muted {className}"
-	style="max-height: {maxHeight}; max-width: {maxWidth};"
+	class="code-preview-wrapper min-w-0 max-w-full overflow-x-auto rounded-lg border border-border bg-muted {className}"
+	style="max-height: {maxHeight}; {maxWidth ? `max-width: ${maxWidth};` : ''}"
 >
 	<!-- Needs to be formatted as single line for proper rendering -->
 	<pre class="m-0"><code class="hljs text-sm leading-relaxed">{@html highlightedHtml}</code></pre>

tools/ui/src/lib/components/app/dialogs/DialogMcpServerRecommendations.svelte

@@ -4,9 +4,10 @@
 	import * as Dialog from '$lib/components/ui/dialog';
 	import { fly } from 'svelte/transition';
 	import { McpServerCardCompact, McpServerForm } from '$lib/components/app/mcp';
-	import { RECOMMENDED_MCP_SERVERS } from '$lib/constants';
+	import { RECOMMENDED_MCP_SERVERS, SETTINGS_KEYS } from '$lib/constants';
 	import { conversationsStore } from '$lib/stores/conversations.svelte';
 	import { mcpStore } from '$lib/stores/mcp.svelte';
+	import { settingsStore } from '$lib/stores/settings.svelte';
 	import { uuid } from '$lib/utils';
 	import { MCP_SERVERS_ADDED_TO_CHAT_LOCALSTORAGE_KEY, MCP_SERVER_ID_PREFIX } from '$lib/constants';
 	import type { MCPServerSettingsEntry } from '$lib/types';
@@ -24,6 +25,22 @@
 	);
 
 	let addedServers = $state<MCPServerSettingsEntry[]>([]);
+	let didAddAny = $state(false);
+
+	let selectedRecommendedCount = $derived.by(
+		() => RECOMMENDED_MCP_SERVERS.filter((server) => selected[server.id]).length
+	);
+
+	let footerLabel = $derived.by(() => {
+		const recommended = selectedRecommendedCount;
+		const custom = addedServers.length;
+		const total = recommended + custom;
+
+		if (total === 0) return 'Continue';
+		if (recommended === 0) return custom === 1 ? 'Add server' : `Add ${custom} servers`;
+		if (custom === 0) return recommended === 1 ? 'Add server' : `Add ${recommended} servers`;
+		return `Add ${recommended} servers and ${custom} custom`;
+	});
 
 	let showAddForm = $state(false);
 	let newServerUrl = $state('');
@@ -44,9 +61,14 @@
 			showAddForm = false;
 			newServerUrl = '';
 			newServerHeaders = '';
-			addedServers = [];
+
+			if (!didAddAny) {
+				settingsStore.updateConfig(SETTINGS_KEYS.MCP_SERVERS, []);
+			}
 
 			localStorage.setItem(MCP_SERVERS_ADDED_TO_CHAT_LOCALSTORAGE_KEY, 'true');
+			addedServers = [];
+			didAddAny = false;
 		}
 		open = value;
 		onOpenChange?.(value);
@@ -59,6 +81,7 @@
 	}
 
 	function enableSelected() {
+		didAddAny = true;
 		localStorage.setItem(MCP_SERVERS_ADDED_TO_CHAT_LOCALSTORAGE_KEY, 'true');
 
 		for (const server of RECOMMENDED_MCP_SERVERS) {
@@ -83,6 +106,8 @@
 	function saveNewServer() {
 		if (newServerUrlError) return;
 
+		didAddAny = true;
+
 		const newServerId = uuid() ?? `${MCP_SERVER_ID_PREFIX}-${Date.now()}`;
 
 		localStorage.setItem(MCP_SERVERS_ADDED_TO_CHAT_LOCALSTORAGE_KEY, 'true');
@@ -174,7 +199,12 @@
 		<Dialog.Footer>
 			<Button variant="secondary" size="sm" onclick={() => handleOpenChange(false)}>Not now</Button>
 
-			<Button variant="default" size="sm" onclick={enableSelected}>Add selected</Button>
+			<Button
+				variant="default"
+				size="sm"
+				onclick={enableSelected}
+				disabled={footerLabel === 'Continue'}>{footerLabel}</Button
+			>
 		</Dialog.Footer>
 	</Dialog.Content>
 </Dialog.Root>

tools/ui/src/lib/components/app/settings/SettingsChat/SettingsChatToolsTab.svelte

@@ -39,13 +39,17 @@
 					{@const faviconUrl = group.serverId ? mcpStore.getServerFavicon(group.serverId) : null}
 
 					<span class="inline-flex min-w-0 items-center gap-1.5 font-medium">
-						<McpServerIdentity
-							iconClass="h-4 w-4"
-							iconRounded="rounded-sm"
-							showVersion={false}
-							displayName={group.label}
-							{faviconUrl}
-						/>
+						{#if group.source === 'mcp'}
+							<McpServerIdentity
+								iconClass="h-4 w-4"
+								iconRounded="rounded-sm"
+								showVersion={false}
+								displayName={group.label}
+								{faviconUrl}
+							/>
+						{:else}
+							<TruncatedText text={group.label} class="font-medium" />
+						{/if}
 					</span>
 
 					<span class="ml-auto shrink-0 text-xs text-muted-foreground">

tools/ui/src/lib/components/ui/hover-card/hover-card-content.svelte

@@ -0,0 +1,31 @@
+<script lang="ts">
+	import { LinkPreview as HoverCardPrimitive } from 'bits-ui';
+	import { cn, type WithoutChildrenOrChild } from '$lib/components/ui/utils.js';
+	import HoverCardPortal from './hover-card-portal.svelte';
+	import type { ComponentProps } from 'svelte';
+
+	let {
+		ref = $bindable(null),
+		class: className,
+		align = 'center',
+		sideOffset = 4,
+		portalProps,
+		...restProps
+	}: HoverCardPrimitive.ContentProps & {
+		portalProps?: WithoutChildrenOrChild<ComponentProps<typeof HoverCardPortal>>;
+	} = $props();
+</script>
+
+<HoverCardPortal {...portalProps}>
+	<HoverCardPrimitive.Content
+		bind:ref
+		data-slot="hover-card-content"
+		{align}
+		{sideOffset}
+		class={cn(
+			'data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 ring-foreground/10 bg-popover text-popover-foreground w-64 rounded-lg p-2.5 text-sm shadow-md ring-1 duration-100 z-50 origin-(--transform-origin) outline-hidden',
+			className
+		)}
+		{...restProps}
+	/>
+</HoverCardPortal>

tools/ui/src/lib/components/ui/hover-card/hover-card-portal.svelte

@@ -0,0 +1,7 @@
+<script lang="ts">
+	import { LinkPreview as HoverCardPrimitive } from 'bits-ui';
+
+	let { ...restProps }: HoverCardPrimitive.PortalProps = $props();
+</script>
+
+<HoverCardPrimitive.Portal {...restProps} />

tools/ui/src/lib/components/ui/hover-card/hover-card-trigger.svelte

@@ -0,0 +1,7 @@
+<script lang="ts">
+	import { LinkPreview as HoverCardPrimitive } from 'bits-ui';
+
+	let { ref = $bindable(null), ...restProps }: HoverCardPrimitive.TriggerProps = $props();
+</script>
+
+<HoverCardPrimitive.Trigger bind:ref data-slot="hover-card-trigger" {...restProps} />

tools/ui/src/lib/components/ui/hover-card/hover-card.svelte

@@ -0,0 +1,7 @@
+<script lang="ts">
+	import { LinkPreview as HoverCardPrimitive } from 'bits-ui';
+
+	let { open = $bindable(false), ...restProps }: HoverCardPrimitive.RootProps = $props();
+</script>
+
+<HoverCardPrimitive.Root bind:open {...restProps} />

tools/ui/src/lib/components/ui/hover-card/index.ts

@@ -0,0 +1,15 @@
+import Root from './hover-card.svelte';
+import Content from './hover-card-content.svelte';
+import Trigger from './hover-card-trigger.svelte';
+import Portal from './hover-card-portal.svelte';
+
+export {
+	Root,
+	Content,
+	Trigger,
+	Portal,
+	Root as HoverCard,
+	Content as HoverCardContent,
+	Trigger as HoverCardTrigger,
+	Portal as HoverCardPortal
+};

tools/ui/src/lib/constants/context-keys.ts

@@ -1,4 +1,3 @@
 export const CONTEXT_KEY_MESSAGE_EDIT = 'chat-message-edit';
 export const CONTEXT_KEY_CHAT_ACTIONS = 'chat-actions';
 export const CONTEXT_KEY_CHAT_SETTINGS_CONFIG = 'chat-settings-config';
-export const CONTEXT_KEY_PROCESSING_INFO = 'processing-info';