@fugood/llama.node 0.3.9 → 0.3.11

package/src/llama.cpp/examples/server/utils.hpp

@@ -5,10 +5,6 @@
 #include "llama.h"
 #include "common/base64.hpp"
 
-#ifndef NDEBUG
-// crash the server in debug mode, otherwise send an http 500 error
-#define CPPHTTPLIB_NO_EXCEPTIONS 1
-#endif
 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
 #include "httplib.h"
@@ -16,6 +12,9 @@
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
+#include "minja.hpp"
+#include "chat.hpp"
+#include "chat-template.hpp"
 
 #include <random>
 #include <sstream>
@@ -349,7 +348,7 @@ static llama_tokens format_infill(
 }
 
 // Format given chat. If tmpl is empty, we take the template from model metadata
-inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
+inline std::string format_chat(const common_chat_template & tmpl, const std::vector<json> & messages) {
     std::vector<common_chat_msg> chat;
 
     for (size_t i = 0; i < messages.size(); ++i) {
@@ -374,10 +373,10 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
             throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
         }
 
-        chat.push_back({role, content});
+        chat.push_back({role, content, /* tool_calls= */ {}});
     }
 
-    const auto formatted_chat = common_chat_apply_template(model, tmpl, chat, true);
+    const auto formatted_chat = common_chat_apply_template(tmpl, chat, true, /* use_jinja= */ false);
     LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
 
     return formatted_chat;
@@ -576,14 +575,32 @@ static json oaicompat_completion_params_parse(const json & body) {
     return llama_params;
 }
 
-static json oaicompat_chat_completion_params_parse(
-        const struct llama_model * model,
-        const json & body, /* openai api json semantics */
-        const std::string & chat_template) {
+static json oaicompat_completion_params_parse(
+    const json & body, /* openai api json semantics */
+    bool use_jinja,
+    const common_chat_templates & chat_templates)
+{
     json llama_params;
+    const auto & tmpl = body.contains("tools") && chat_templates.template_tool_use
+        ? *chat_templates.template_tool_use
+        : *chat_templates.template_default;
 
-    // Apply chat template to the list of messages
-    llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
+    auto tools = json_value(body, "tools", json());
+    auto stream = json_value(body, "stream", false);
+
+    if (tools.is_array() && !tools.empty()) {
+        if (stream) {
+            throw std::runtime_error("Cannot use tools with stream");
+        }
+        if (!use_jinja) {
+            throw std::runtime_error("tools param requires --jinja flag");
+        }
+    }
+    if (!use_jinja) {
+        if (body.contains("tool_choice") && !body.at("tool_choice").is_null()) {
+            throw std::runtime_error("Unsupported param: tool_choice");
+        }
+    }
 
     // Handle "stop" field
     if (body.contains("stop") && body.at("stop").is_string()) {
@@ -606,6 +623,49 @@ static json oaicompat_chat_completion_params_parse(
         }
     }
 
+    // Apply chat template to the list of messages
+    if (use_jinja) {
+        auto tool_choice = json_value(body, "tool_choice", std::string("auto"));
+        if (tool_choice != "none" && tool_choice != "auto" && tool_choice != "required") {
+            throw std::runtime_error("Invalid tool_choice: " + tool_choice);
+        }
+        if (tool_choice != "none" && llama_params.contains("grammar")) {
+            throw std::runtime_error("Cannot use custom grammar constraints with tools.");
+        }
+        common_chat_inputs inputs;
+        inputs.messages = body.at("messages");
+        inputs.tools = tools;
+        inputs.tool_choice = tool_choice;
+        inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
+        if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) {
+            LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n");
+            inputs.parallel_tool_calls = false;
+        }
+        inputs.stream = stream;
+        // TODO: support mixing schema w/ tools beyond generic format.
+        inputs.json_schema = json_value(llama_params, "json_schema", json());
+        auto chat_params = common_chat_params_init(tmpl, inputs);
+
+        llama_params["chat_format"] = static_cast<int>(chat_params.format);
+        llama_params["prompt"] = chat_params.prompt;
+        llama_params["grammar"] = chat_params.grammar;
+        llama_params["grammar_lazy"] = chat_params.grammar_lazy;
+        auto grammar_triggers = json::array();
+        for (const auto & trigger : chat_params.grammar_triggers) {
+            grammar_triggers.push_back({
+                {"word", trigger.word},
+                {"at_start", trigger.at_start},
+            });
+        }
+        llama_params["grammar_triggers"] = grammar_triggers;
+        llama_params["preserved_tokens"] = chat_params.preserved_tokens;
+        for (const auto & stop : chat_params.additional_stops) {
+            llama_params["stop"].push_back(stop);
+        }
+    } else {
+        llama_params["prompt"] = format_chat(tmpl, body.at("messages"));
+    }
+
     // Handle "n" field
     int n_choices = json_value(body, "n", 1);
     if (n_choices != 1) {
@@ -620,14 +680,6 @@ static json oaicompat_chat_completion_params_parse(
         throw std::runtime_error("top_logprobs requires logprobs to be set to true");
     }
 
-    // Params supported by OAI but unsupported by llama.cpp
-    static const std::vector<std::string> unsupported_params { "tools", "tool_choice" };
-    for (const auto & param : unsupported_params) {
-        if (body.contains(param)) {
-            throw std::runtime_error("Unsupported param: " + param);
-        }
-    }
-
     // Copy remaining properties to llama_params
     // This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
     // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp

package/src/llama.cpp/examples/simple-chat/simple-chat.cpp

@@ -95,13 +95,15 @@ int main(int argc, char ** argv) {
     llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
 
     // helper function to evaluate a prompt and generate a response
-    auto generate = [&](const std::string & prompt, bool is_first) {
+    auto generate = [&](const std::string & prompt) {
         std::string response;
 
+        const bool is_first = llama_get_kv_cache_used_cells(ctx) == 0;
+
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
         std::vector<llama_token> prompt_tokens(n_prompt_tokens);
-        if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
+        if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), is_first, true) < 0) {
             GGML_ABORT("failed to tokenize the prompt\n");
         }
 
@@ -161,7 +163,7 @@ int main(int argc, char ** argv) {
             break;
         }
 
-        const char * tmpl = llama_model_chat_template(model);
+        const char * tmpl = llama_model_chat_template(model, /* name */ nullptr);
 
         // add the user input to the message list and format it
         messages.push_back({"user", strdup(user.c_str())});
@@ -180,7 +182,7 @@ int main(int argc, char ** argv) {
 
         // generate a response
         printf("\033[33m");
-        std::string response = generate(prompt, prev_len == 0);
+        std::string response = generate(prompt);
         printf("\n\033[0m");
 
         // add the response to the messages

package/src/llama.cpp/examples/simple-cmake-pkg/CMakeLists.txt

@@ -0,0 +1,11 @@
+cmake_minimum_required(VERSION 3.12)
+project(llama-simple-cmake-pkg)
+
+set(TARGET llama-simple-cmake-pkg)
+
+find_package(Llama REQUIRED)
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../simple/simple.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ggml::all ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

package/src/llama.cpp/ggml/CMakeLists.txt

@@ -58,7 +58,8 @@ else()
     set(GGML_BLAS_VENDOR_DEFAULT "Generic")
 endif()
 
-if (CMAKE_CROSSCOMPILING)
+if (CMAKE_CROSSCOMPILING OR DEFINED ENV{SOURCE_DATE_EPOCH})
+    message(STATUS "Setting GGML_NATIVE_DEFAULT to OFF")
     set(GGML_NATIVE_DEFAULT OFF)
 else()
     set(GGML_NATIVE_DEFAULT ON)
@@ -153,6 +154,8 @@ option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashA
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
 option(GGML_HIP                             "ggml: use HIP"                                   OFF)
+option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
+option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
@@ -264,3 +267,77 @@ if (GGML_STANDALONE)
     install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
         DESTINATION share/pkgconfig)
 endif()
+
+#
+# Create CMake package
+#
+
+# Generate version info based on git commit.
+
+if(NOT DEFINED GGML_BUILD_NUMBER)
+    find_program(GIT_EXE NAMES git git.exe REQUIRED NO_CMAKE_FIND_ROOT_PATH)
+    execute_process(COMMAND ${GIT_EXE} rev-list --count HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE GGML_BUILD_NUMBER
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+
+    if(GGML_BUILD_NUMBER EQUAL 1)
+        message(WARNING "GGML build version fixed at 1 likely due to a shallow clone.")
+    endif()
+
+    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE GGML_BUILD_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+endif()
+
+
+# Capture variables prefixed with GGML_.
+
+set(variable_set_statements
+"
+####### Expanded from @GGML_VARIABLES_EXPANED@ by configure_package_config_file() #######
+####### Any changes to this file will be overwritten by the next CMake run        #######
+
+")
+
+set(GGML_SHARED_LIB ${BUILD_SHARED_LIBS})
+
+get_cmake_property(all_variables VARIABLES)
+foreach(variable_name IN LISTS all_variables)
+    if(variable_name MATCHES "^GGML_")
+        string(REPLACE ";" "\\;"
+               variable_value "${${variable_name}}")
+
+        set(variable_set_statements
+            "${variable_set_statements}set(${variable_name} \"${variable_value}\")\n")
+    endif()
+endforeach()
+
+set(GGML_VARIABLES_EXPANDED ${variable_set_statements})
+
+# Create the CMake package and set install location.
+
+set(GGML_INSTALL_VERSION 0.0.${GGML_BUILD_NUMBER})
+set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(GGML_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(GGML_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/ggml-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml
+    PATH_VARS GGML_INCLUDE_INSTALL_DIR
+              GGML_LIB_INSTALL_DIR
+              GGML_BIN_INSTALL_DIR)
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
+    VERSION ${GGML_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)

package/src/llama.cpp/ggml/include/ggml.h

@@ -1775,7 +1775,7 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   k);
 
-#define GGML_KQ_MASK_PAD 32
+#define GGML_KQ_MASK_PAD 64
 
     // q:    [n_embd, n_batch,     n_head,    1]
     // k:    [n_embd, n_kv,        n_head_kv, 1]

package/src/llama.cpp/ggml/src/CMakeLists.txt

@@ -93,12 +93,18 @@ endif()
 
 if (GGML_CCACHE)
     find_program(GGML_CCACHE_FOUND ccache)
+    find_program(GGML_SCCACHE_FOUND sccache)
 
-    if (GGML_CCACHE_FOUND)
+    if (GGML_CCACHE_FOUND OR GGML_SCCACHE_FOUND)
+        if(GGML_CCACHE_FOUND)
+            set(GGML_CCACHE_VARIANT ccache)
+        else()
+            set(GGML_CCACHE_VARIANT sccache)
+        endif()
         # TODO: should not be set globally
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
+        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
         set(ENV{CCACHE_SLOPPINESS} time_macros)
-        message(STATUS "ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
+        message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
     else()
         message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF")
     endif ()
@@ -250,6 +256,17 @@ function(ggml_add_backend_library backend)
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_BUILD)
         target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
     endif()
+
+    if(NOT GGML_AVAILABLE_BACKENDS)
+        set(GGML_AVAILABLE_BACKENDS "${backend}"
+            CACHE INTERNAL "List of backends for cmake package")
+    else()
+        list(FIND GGML_AVAILABLE_BACKENDS "${backend}" has_backend)
+        if(has_backend EQUAL -1)
+            set(GGML_AVAILABLE_BACKENDS "${GGML_AVAILABLE_BACKENDS};${backend}"
+                CACHE INTERNAL "List of backends for cmake package")
+        endif()
+    endif()
 endfunction()
 
 function(ggml_add_backend backend)
@@ -297,7 +314,7 @@ if (GGML_CPU_ALL_VARIANTS)
         # MSVC doesn't support AMX
         ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
     endif()
-else ()
+elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")
 endif()
 

package/src/llama.cpp/ggml/src/ggml-alloc.c

@@ -989,19 +989,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
         }
 
-        if (this_size > max_size) {
-            GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
-                    __func__, t->name,
-                    ggml_backend_buft_name(buft),
-                    this_size, max_size);
-            for (size_t i = 0; i < n_buffers; i++) {
-                ggml_backend_buffer_free(buffers[i]);
-            }
-            free(buffers);
-            return NULL;
-        }
-
-        if ((cur_buf_size + this_size) > max_size) {
+        if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) {
             // allocate tensors in the current buffer
             if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
                 return NULL;

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -297,6 +297,90 @@ static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
 static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
 #endif
 
+#if defined(__loongarch_sx)
+
+static __m128i lsx_packs_w(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_w(a, 15);
+    tmp1 = __lsx_vsat_w(b, 15);
+    return __lsx_vpickev_h(tmp1, tmp);
+}
+
+static __m128i lsx_packs_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_h(a, 7);
+    tmp1 = __lsx_vsat_h(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packus_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_hu(a, 7);
+    tmp1 = __lsx_vsat_hu(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_h_b(a, b);
+    tmp2 = __lsx_vmulwod_h_b(a, b);
+    return __lsx_vsadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_madd_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_w_h(a, b);
+    tmp2 = __lsx_vmulwod_w_h(a, b);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
+    v4i32 __ret = {d, c, b, a};
+    return (__m128i)__ret;
+}
+
+static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
+    __m128i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lsx_vreplgr2vr_b(f);
+    zero = __lsx_vldi(0);
+    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
+    return __lsx_vshuf_b(a, zero, tmp2);
+}
+
+static __m128i lsx_hadd_h(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_h(b, a);
+    __m128i tmp2 = __lsx_vpickod_h(b, a);
+    return __lsx_vadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_hadd_w(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_w(b, a);
+    __m128i tmp2 = __lsx_vpickod_w(b, a);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128 lsx_hadd_s(__m128 a, __m128 b) {
+    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
+    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
+
+    return __lsx_vfadd_s(tmp1, tmp2);
+}
+
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =lsx_hadd_s(a, b);
+    __m128 res_1 =lsx_hadd_s(c, d);
+    __m128 res =lsx_hadd_s(res_0, res_1);
+    res =lsx_hadd_s(res, res);
+    res =lsx_hadd_s(res, res);
+
+    return ((v4f32)res)[0];
+}
+#endif
+
 #if defined(__loongarch_asx)
 
 #ifdef __clang__
@@ -395,11 +479,6 @@ static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1
     return (__m256i)__ret;
 }
 
-static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
-    v4i32 __ret = {d, c, b, a};
-    return (__m128i)__ret;
-}
-
 static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
     v4i64 __ret = {d, c, b, a};
     return (__m256i)__ret;
@@ -409,18 +488,6 @@ static __m256i lasx_insertf128( __m128i x, __m128i y) {
     return lasx_set_q(x, y);
 }
 
-static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
-    __m128i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lsx_vreplgr2vr_b(f);
-    zero = __lsx_vldi(0);
-    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
-    return __lsx_vshuf_b(a, zero, tmp2);
-}
-
 static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
     __m256i mask_f, zero, tmp0, tmp2, mask;
     int f = 0x8f;
@@ -482,25 +549,6 @@ static __m128 lasx_extractf128( __m256 a, int pos) {
     return ret;
 }
 
-static __m128i lsx_hadd_h(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_h(b, a);
-    __m128i tmp2 = __lsx_vpickod_h(b, a);
-    return __lsx_vadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_hadd_w(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_w(b, a);
-    __m128i tmp2 = __lsx_vpickod_w(b, a);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-static __m128 lsx_hadd_s(__m128 a, __m128 b) {
-    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
-    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
-
-    return __lsx_vfadd_s(tmp1, tmp2);
-}
-
 static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
     __m256i tmp1, tmp2;
     tmp1 = __lasx_xvmulwev_h_b(a, b);
@@ -529,42 +577,6 @@ static __m256i lasx_packs_h(__m256i a, __m256i b) {
     return __lasx_xvpickev_b(tmp1, tmp);
 }
 
-static __m128i lsx_packs_w(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_w(a, 15);
-    tmp1 = __lsx_vsat_w(b, 15);
-    return __lsx_vpickev_h(tmp1, tmp);
-}
-
-static __m128i lsx_packs_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_h(a, 7);
-    tmp1 = __lsx_vsat_h(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_packus_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_hu(a, 7);
-    tmp1 = __lsx_vsat_hu(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-
-static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_h_b(a, b);
-    tmp2 = __lsx_vmulwod_h_b(a, b);
-    return __lsx_vsadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_madd_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_w_h(a, b);
-    tmp2 = __lsx_vmulwod_w_h(a, b);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
 // multiply int8_t, add results pairwise twice
 static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
     // Get absolute values of x vectors
@@ -2232,21 +2244,22 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     }
 
     sumf = hsum_float_8(acc);
+
 #elif defined(__loongarch_sx)
     // set constants
     const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
     const __m128i off = __lsx_vreplgr2vr_b(8);
 
     // Initialize accumulator with zeros
-    __m128 acc_0 = __lsx_vldi(0);
-    __m128 acc_1 = __lsx_vldi(0);
-    __m128 acc_2 = __lsx_vldi(0);
-    __m128 acc_3 = __lsx_vldi(0);
+    __m128 acc_0 = (__m128)__lsx_vldi(0);
+    __m128 acc_1 = (__m128)__lsx_vldi(0);
+    __m128 acc_2 = (__m128)__lsx_vldi(0);
+    __m128 acc_3 = (__m128)__lsx_vldi(0);
 
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -2264,7 +2277,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c

@@ -1302,7 +1302,7 @@ struct ggml_threadpool {
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
     atomic_bool pause;        // Used for pausing the threadpool or individual threads
-    atomic_bool abort;        // Used for aborting processing of a graph
+    atomic_int abort;         // Used for aborting processing of a graph
 
     struct ggml_compute_state * workers;   // per thread state
     int          n_threads_max; // number of threads in the pool
@@ -7883,7 +7883,7 @@ static void ggml_compute_forward_out_prod_f32(
 
                     float * s0 = (float *) ((char *) src0->data + (          i01*nb01 + i02*nb02 + i03*nb03));
                     float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
-                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
+                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1   + i2*nb2   + i3*nb3));
 
                     ggml_vec_mad_f32_unroll(ne0, nb01, nb11, d, s0, s1);
                 }
@@ -7892,7 +7892,7 @@ static void ggml_compute_forward_out_prod_f32(
 
                     float * s0 = (float *) ((char *) src0->data + (          i01*nb01 + i02*nb02 + i03*nb03));
                     float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
-                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
+                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1   + i2*nb2   + i3*nb3));
 
                     ggml_vec_mad_f32(ne0, d, s0, *s1);
                 }
@@ -13851,14 +13851,14 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         /*.threadpool=*/ tp,
     };
 
-    for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
+    for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&
                 cplan->abort_callback(cplan->abort_callback_data)) {
-            tp->abort = true;
+            atomic_store_explicit(&tp->abort, node_n + 1, memory_order_relaxed);
             tp->ec    = GGML_STATUS_ABORTED;
         }
 
@@ -14031,7 +14031,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
         threadpool->current_chunk    = 0;
         threadpool->stop             = false;
         threadpool->pause            = tpp->paused;
-        threadpool->abort            = false;
+        threadpool->abort            = -1;
         threadpool->workers          = NULL;
         threadpool->n_threads_max    = tpp->n_threads;
         threadpool->n_threads_cur    = tpp->n_threads;
@@ -14110,7 +14110,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         threadpool->cgraph           = cgraph;
         threadpool->cplan            = cplan;
         threadpool->current_chunk    = 0;
-        threadpool->abort            = false;
+        threadpool->abort            = -1;
         threadpool->ec               = GGML_STATUS_SUCCESS;
     }
 

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -416,7 +416,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
         case GGML_OP_OUT_PROD:
-            return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
+            return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
+                src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         default:
             return true;
     }

package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt

@@ -28,7 +28,7 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 
     file(GLOB   GGML_SOURCES_CUDA "*.cu")
-    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    file(GLOB   SRCS "template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/mmq*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})