Merge 24bad77ebf into 32d6ee6385

examples/server/CMakeLists.txt

@@ -34,7 +34,6 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
-target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
 
 if (LLAMA_SERVER_SSL)

examples/server/README.md

@@ -450,8 +450,6 @@ These words will not be included in the completion, so make sure to add them to
 
 `post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
 
-`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error.
-
 **Response format**
 
 - Note: In streaming mode (`stream`), only `content`, `tokens` and `stop` will be returned until end of completion. Responses are sent using the [Server-sent events](https://html.spec.whatwg.org/multipage/server-sent-events.html) standard. Note: the browser's `EventSource` interface cannot be used due to its lack of `POST` request support.

examples/server/server.cpp

@@ -92,7 +92,6 @@ struct slot_params {
     int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
 
     std::vector<std::string> antiprompt;
-    std::vector<std::string> response_fields;
     bool timings_per_token = false;
     bool post_sampling_probs = false;
     bool ignore_eos = false;
@@ -210,7 +209,6 @@ struct server_task {
         params.n_discard        = json_value(data, "n_discard",          defaults.n_discard);
       //params.t_max_prompt_ms  = json_value(data, "t_max_prompt_ms",    defaults.t_max_prompt_ms); // TODO: implement
         params.t_max_predict_ms = json_value(data, "t_max_predict_ms",   defaults.t_max_predict_ms);
-        params.response_fields  = json_value(data, "response_fields",   std::vector<std::string>());
 
         params.sampling.top_k              = json_value(data, "top_k",              defaults.sampling.top_k);
         params.sampling.top_p              = json_value(data, "top_p",              defaults.sampling.top_p);
@@ -524,7 +522,6 @@ struct server_task_result_cmpl_final : server_task_result {
 
     bool post_sampling_probs;
     std::vector<completion_token_output> probs_output;
-    std::vector<std::string>  response_fields;
 
     slot_params generation_params;
 
@@ -571,7 +568,7 @@ struct server_task_result_cmpl_final : server_task_result {
         if (!stream && !probs_output.empty()) {
             res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs);
         }
-        return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
+        return res;
     }
 
     json to_json_oaicompat_chat() {
@@ -2069,7 +2066,6 @@ struct server_context {
         res->tokens          = slot.generated_tokens;
         res->timings         = slot.get_timings();
         res->prompt          = common_detokenize(ctx, slot.prompt_tokens, true);
-        res->response_fields = slot.params.response_fields;
 
         res->truncated           = slot.truncated;
         res->n_decoded           = slot.n_decoded;
@@ -3790,17 +3786,6 @@ int main(int argc, char ** argv) {
             return;
         }
 
-        bool use_base64 = false;
-        if (body.count("encoding_format") != 0) {
-            const std::string& format = body.at("encoding_format");
-            if (format == "base64") {
-                use_base64 = true;
-            } else if (format != "float") {
-                res_error(res, format_error_response("The format to return the embeddings in. Can be either float or base64", ERROR_TYPE_INVALID_REQUEST));
-                return;
-            }
-        }
-
         std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
@@ -3852,7 +3837,7 @@ int main(int argc, char ** argv) {
         }
 
         // write JSON response
-        json root = oaicompat ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses);
+        json root = oaicompat ? format_embeddings_response_oaicompat(body, responses) : json(responses);
         res_ok(res, root);
     };
 

examples/server/tests/unit/test_completion.py

@@ -95,7 +95,7 @@ def test_consistent_result_same_seed(n_slots: int):
         res = server.make_request("POST", "/completion", data={
             "prompt": "I believe the meaning of life is",
             "seed": 42,
-            "temperature": 0.0,
+            "temperature": 1.0,
             "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
         })
         if last_res is not None:
@@ -120,10 +120,9 @@ def test_different_result_different_seed(n_slots: int):
             assert res.body["content"] != last_res.body["content"]
         last_res = res
 
-# TODO figure why it don't work with temperature = 1
+
-# @pytest.mark.parametrize("temperature", [0.0, 1.0])
 @pytest.mark.parametrize("n_batch", [16, 32])
-@pytest.mark.parametrize("temperature", [0.0])
+@pytest.mark.parametrize("temperature", [0.0, 1.0])
 def test_consistent_result_different_batch_size(n_batch: int, temperature: float):
     global server
     server.n_batch = n_batch
@@ -258,40 +257,6 @@ def test_completion_parallel_slots(n_slots: int, n_requests: int):
         # assert match_regex(re_content, res.body["content"])
 
 
-@pytest.mark.parametrize(
-    "prompt,n_predict,response_fields",
-    [
-        ("I believe the meaning of life is", 8, []),
-        ("I believe the meaning of life is", 32, ["content", "generation_settings/n_predict", "prompt"]),
-    ],
-)
-def test_completion_response_fields(
-    prompt: str, n_predict: int, response_fields: list[str]
-):
-    global server
-    server.start()
-    res = server.make_request(
-        "POST",
-        "/completion",
-        data={
-            "n_predict": n_predict,
-            "prompt": prompt,
-            "response_fields": response_fields,
-        },
-    )
-    assert res.status_code == 200
-    assert "content" in res.body
-    assert len(res.body["content"])
-    if len(response_fields):
-        assert res.body["generation_settings/n_predict"] == n_predict
-        assert res.body["prompt"] == "<s> " + prompt
-        assert isinstance(res.body["content"], str)
-        assert len(res.body) == len(response_fields)
-    else:
-        assert len(res.body)
-        assert "generation_settings" in res.body
-
-
 def test_n_probs():
     global server
     server.start()

examples/server/tests/unit/test_embedding.py

@@ -1,5 +1,3 @@
-import base64
-import struct
 import pytest
 from openai import OpenAI
 from utils import *
@@ -196,42 +194,3 @@ def test_embedding_usage_multiple():
     assert res.status_code == 200
     assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
     assert res.body['usage']['prompt_tokens'] == 2 * 9
-
-
-def test_embedding_openai_library_base64():
-    server.start()
-    test_input = "Test base64 embedding output"
-
-    # get embedding in default format
-    res = server.make_request("POST", "/v1/embeddings", data={
-        "input": test_input
-    })
-    assert res.status_code == 200
-    vec0 = res.body["data"][0]["embedding"]
-
-    # get embedding in base64 format
-    res = server.make_request("POST", "/v1/embeddings", data={
-        "input": test_input,
-        "encoding_format": "base64"
-    })
-
-    assert res.status_code == 200
-    assert "data" in res.body
-    assert len(res.body["data"]) == 1
-
-    embedding_data = res.body["data"][0]
-    assert "embedding" in embedding_data
-    assert isinstance(embedding_data["embedding"], str)
-
-    # Verify embedding is valid base64
-    decoded = base64.b64decode(embedding_data["embedding"])
-    # Verify decoded data can be converted back to float array
-    float_count = len(decoded) // 4  # 4 bytes per float
-    floats = struct.unpack(f'{float_count}f', decoded)
-    assert len(floats) > 0
-    assert all(isinstance(x, float) for x in floats)
-    assert len(floats) == len(vec0)
-
-    # make sure the decoded data is the same as the original
-    for x, y in zip(floats, vec0):
-        assert abs(x - y) < EPSILON

examples/server/utils.hpp

@@ -3,7 +3,6 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "common/base64.hpp"
 
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@@ -91,28 +90,6 @@ static bool json_is_array_of_mixed_numbers_strings(const json & data) {
     return false;
 }
 
-// get value by path(key1 / key2)
-static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
-    json result = json::object();
-
-    for (const std::string & path : paths) {
-        json current = js;
-        const auto keys = string_split<std::string>(path, /*separator*/ '/');
-        bool valid_path = true;
-        for (const std::string & k : keys) {
-            if (valid_path && current.is_object() && current.contains(k)) {
-                current = current[k];
-            } else {
-                valid_path = false;
-            }
-        }
-        if (valid_path) {
-            result[path] = current;
-        }
-    }
-    return result;
-}
-
 /**
  * this handles 2 cases:
  * - only string, example: "string"
@@ -614,31 +591,16 @@ static json oaicompat_completion_params_parse(
     return llama_params;
 }
 
-static json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false) {
+static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
     json data = json::array();
     int32_t n_tokens = 0;
     int i = 0;
     for (const auto & elem : embeddings) {
-        json embedding_obj;
+        data.push_back(json{
-
+            {"embedding", json_value(elem, "embedding", json::array())},
-        if (use_base64) {
+            {"index",     i++},
-            const auto& vec = json_value(elem, "embedding", json::array()).get<std::vector<float>>();
+            {"object",    "embedding"}
-            const char* data_ptr = reinterpret_cast<const char*>(vec.data());
+        });
-            size_t data_size = vec.size() * sizeof(float);
-            embedding_obj = {
-                {"embedding", base64::encode(data_ptr, data_size)},
-                {"index", i++},
-                {"object", "embedding"},
-                {"encoding_format", "base64"}
-            };
-        } else {
-            embedding_obj = {
-                {"embedding", json_value(elem, "embedding", json::array())},
-                {"index", i++},
-                {"object", "embedding"}
-            };
-        }
-        data.push_back(embedding_obj);
 
         n_tokens += json_value(elem, "tokens_evaluated", 0);
     }

ggml/src/CMakeLists.txt

@@ -234,7 +234,6 @@ function(ggml_add_backend_library backend)
         # write the shared library to the output directory
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
-        add_dependencies(ggml ${backend})
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})

ggml/src/ggml-backend-reg.cpp

@@ -66,26 +66,6 @@
 #include "ggml-kompute.h"
 #endif
 
-// disable C++17 deprecation warning for std::codecvt_utf8
-#if defined(__clang__)
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
-static std::wstring utf8_to_utf16(const std::string & str) {
-    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-    return converter.from_bytes(str);
-}
-
-static std::string utf16_to_utf8(const std::wstring & str) {
-    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-    return converter.to_bytes(str);
-}
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#endif
-
 #ifdef _WIN32
 
 using dl_handle = std::remove_pointer_t<HMODULE>;
@@ -108,6 +88,11 @@ static dl_handle * dl_load_library(const std::wstring & path) {
     return handle;
 }
 
+static dl_handle * dl_load_library(const std::string & path) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return dl_load_library(converter.from_bytes(path));
+}
+
 static void * dl_get_sym(dl_handle * handle, const char * name) {
     DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
     SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
@@ -129,8 +114,8 @@ struct dl_handle_deleter {
     }
 };
 
-static void * dl_load_library(const std::wstring & path) {
+static void * dl_load_library(const std::string & path) {
-    dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);
+    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
 
     return handle;
 }
@@ -217,11 +202,11 @@ struct ggml_backend_registry {
         devices.push_back(device);
     }
 
-    ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
+    ggml_backend_reg_t load_backend(const char * path, bool silent) {
         dl_handle_ptr handle { dl_load_library(path) };
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
             }
             return nullptr;
         }
@@ -229,7 +214,7 @@ struct ggml_backend_registry {
         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
         if (score_fn && score_fn() == 0) {
             if (!silent) {
-                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
             }
             return nullptr;
         }
@@ -237,7 +222,7 @@ struct ggml_backend_registry {
         auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
         if (!backend_init_fn) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
             }
             return nullptr;
         }
@@ -246,16 +231,16 @@ struct ggml_backend_registry {
         if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
             if (!silent) {
                 if (!reg) {
-                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
                 } else {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                        __func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
                 }
             }
             return nullptr;
         }
 
-        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
 
         register_backend(reg, std::move(handle));
 
@@ -391,14 +376,14 @@ ggml_backend_t ggml_backend_init_best(void) {
 
 // Dynamic loading
 ggml_backend_reg_t ggml_backend_load(const char * path) {
-    return get_reg().load_backend(utf8_to_utf16(path), false);
+    return get_reg().load_backend(path, false);
 }
 
 void ggml_backend_unload(ggml_backend_reg_t reg) {
     get_reg().unload_backend(reg, true);
 }
 
-static std::wstring get_executable_path() {
+static std::string get_executable_path() {
 #if defined(__APPLE__)
     // get executable path
     std::vector<char> path;
@@ -416,7 +401,7 @@ static std::wstring get_executable_path() {
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    return utf8_to_utf16(base_path + "/");
+    return base_path + "/";
 #elif defined(__linux__) || defined(__FreeBSD__)
     std::string base_path = ".";
     std::vector<char> path(1024);
@@ -442,63 +427,57 @@ static std::wstring get_executable_path() {
         path.resize(path.size() * 2);
     }
 
-    return utf8_to_utf16(base_path + "/");
+    return base_path + "/";
 #elif defined(_WIN32)
-    std::vector<wchar_t> path(MAX_PATH);
+    std::vector<char> path(MAX_PATH);
-    DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
+    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
     if (len == 0) {
-        return {};
+        return "";
     }
-    std::wstring base_path(path.data(), len);
+    std::string base_path(path.data(), len);
     // remove executable name
     auto last_slash = base_path.find_last_of('\\');
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    return base_path + L"\\";
+    return base_path + "\\";
-#else
-    return {};
 #endif
 }
 
-static std::wstring backend_filename_prefix() {
+static std::string backend_filename_prefix() {
 #ifdef _WIN32
-    return L"ggml-";
+    return "ggml-";
 #else
-    return L"libggml-";
+    return "libggml-";
 #endif
 }
 
-static std::wstring backend_filename_suffix() {
+static std::string backend_filename_suffix() {
 #ifdef _WIN32
-    return L".dll";
+    return ".dll";
 #else
-    return L".so";
+    return ".so";
-#endif
-}
-
-static std::wstring path_separator() {
-#ifdef _WIN32
-    return L"\\";
-#else
-    return L"/";
 #endif
 }
 
 static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
     // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
      // TODO: search system paths
-    std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
+    std::string file_prefix = backend_filename_prefix() + name + "-";
-    std::vector<std::wstring> search_paths;
+    std::vector<std::string> search_paths;
     if (user_search_path == nullptr) {
-        search_paths.push_back(L"." + path_separator());
+        search_paths.push_back("./");
         search_paths.push_back(get_executable_path());
     } else {
-        search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
+#if defined(_WIN32)
+        search_paths.push_back(std::string(user_search_path) + "\\");
+#else
+        search_paths.push_back(std::string(user_search_path) + "/");
+#endif
     }
 
     int best_score = 0;
-    std::wstring best_path;
+    std::string best_path;
 
     namespace fs = std::filesystem;
     for (const auto & search_path : search_paths) {
@@ -508,27 +487,27 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
         for (const auto & entry : dir_it) {
             if (entry.is_regular_file()) {
-                std::wstring filename = entry.path().filename().wstring();
+                std::string filename = entry.path().filename().string();
-                std::wstring ext = entry.path().extension().wstring();
+                std::string ext = entry.path().extension().string();
                 if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
+                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
                     if (!handle && !silent) {
-                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
                     }
                     if (handle) {
                         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
                         if (score_fn) {
                             int s = score_fn();
 #ifndef NDEBUG
-                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
 #endif
                             if (s > best_score) {
                                 best_score = s;
-                                best_path = entry.path().wstring();
+                                best_path = entry.path().string();
                             }
                         } else {
                             if (!silent) {
-                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
                             }
                         }
                     }
@@ -540,15 +519,15 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
     if (best_score == 0) {
         // try to load the base backend
         for (const auto & search_path : search_paths) {
-            std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
+            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
             if (fs::exists(path)) {
-                return get_reg().load_backend(path, silent);
+                return get_reg().load_backend(path.c_str(), silent);
             }
         }
         return nullptr;
     }
 
-    return get_reg().load_backend(best_path, silent);
+    return get_reg().load_backend(best_path.c_str(), silent);
 }
 
 void ggml_backend_load_all() {

ggml/src/ggml-cpu/CMakeLists.txt

@@ -135,20 +135,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             endif()
 
             # show enabled features
-            if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
-                set(FEAT_INPUT_FILE "NUL")
-            else()
-                set(FEAT_INPUT_FILE "/dev/null")
-            endif()
-
             execute_process(
                 COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
-                INPUT_FILE ${FEAT_INPUT_FILE}
+                INPUT_FILE "/dev/null"
                 OUTPUT_VARIABLE ARM_FEATURE
                 RESULT_VARIABLE ARM_FEATURE_RESULT
             )
             if (ARM_FEATURE_RESULT)
-                message(WARNING "Failed to get ARM features")
+                message(FATAL_ERROR "Failed to get ARM features")
             else()
                 foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
                     string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
@@ -323,11 +317,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
 
     if (GGML_BACKEND_DL)
-        if (GGML_NATIVE)
-            # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
-            message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
-        endif()
-
         # The feature detection code is compiled as a separate target so that
         # it can be built without the architecture flags
         # Since multiple variants of the CPU backend may be included in the same

ggml/src/ggml-cpu/ggml-cpu.c

@@ -7419,14 +7419,14 @@ static void ggml_compute_forward_mul_mat(
     if (src1_cont) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(params,
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
-                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                      nb01/ggml_type_size(src0->type),
                                      (const char *)src1->data + i12*nb12 + i13*nb13,
                                      nb11/ggml_type_size(src1->type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
+                                     ith, nth,
                                      src0->type,
                                      src1->type,
                                      dst->type))
@@ -7471,14 +7471,14 @@ UseGgmlGemm1:;
 
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(params,
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
-                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                      nb01/ggml_type_size(src0->type),
                                      (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                      row_size/ggml_type_size(vec_dot_type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
+                                     ith, nth,
                                      src0->type,
                                      vec_dot_type,
                                      dst->type))

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -53,8 +53,6 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
-#include <atomic>
-
 #ifdef _MSC_VER
 #define NOINLINE __declspec(noinline)
 #else
@@ -136,16 +134,6 @@ inline __m512 madd(__m512 a, __m512 b, __m512 c) {
     return _mm512_fmadd_ps(a, b, c);
 }
 #endif
-#if defined(__AVX512BF16__)
-template <>
-inline __m512 madd(__m512bh a, __m512bh b, __m512 c) {
-    return _mm512_dpbf16_ps(c, a, b);
-}
-template <>
-inline __m256 madd(__m256bh a, __m256bh b, __m256 c) {
-    return _mm256_dpbf16_ps(c, a, b);
-}
-#endif
 #endif
 
 #if defined(__ARM_FEATURE_FMA)
@@ -238,13 +226,6 @@ template <> inline __m256 load(const float *p) {
 }
 #endif // __AVX__
 
-#if defined(__AVX2__) || defined(__AVX512F__)
-template <> inline __m256 load(const ggml_bf16_t *p) {
-    return _mm256_castsi256_ps(
-        _mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)p)), 16));
-}
-#endif // __AVX2__
-
 #if defined(__F16C__)
 template <> inline __m256 load(const ggml_fp16_t *p) {
     return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
@@ -258,27 +239,8 @@ template <> inline __m512 load(const float *p) {
 template <> inline __m512 load(const ggml_fp16_t *p) {
     return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
 }
-template <> inline __m512 load(const ggml_bf16_t *p) {
-    return _mm512_castsi512_ps(
-        _mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)p)), 16));
-}
 #endif // __AVX512F__
 
-#if defined(__AVX512BF16__)
-template <> inline __m512bh load(const ggml_bf16_t *p) {
-    return (__m512bh)_mm512_loadu_ps((const float *)p);
-}
-template <> inline __m256bh load(const ggml_bf16_t *p) {
-    return (__m256bh)_mm256_loadu_ps((const float *)p);
-}
-template <> inline __m512bh load(const float *p) {
-    return _mm512_cvtne2ps_pbh(_mm512_loadu_ps(p + 16), _mm512_loadu_ps(p));
-}
-template <> inline __m256bh load(const float *p) {
-    return _mm512_cvtneps_pbh(_mm512_loadu_ps(p));
-}
-#endif
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // CONSTANTS
 
@@ -290,170 +252,199 @@ static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
-template <int M>
-static inline int64_t BLOCK_SIZE(size_t m) {
-    const int64_t NB_BLOC_M = (m + M - 1) / M;
-    return (m % NB_BLOC_M == 0) ? m / NB_BLOC_M : (m / NB_BLOC_M) + 1;
-}
-
-static constexpr inline int64_t BLOC_POS(int64_t ib, int64_t ibN, int64_t bloc_size) {
-    return ib < ibN ? ib * bloc_size : ibN * bloc_size + (ib - ibN) * (bloc_size - 1);
-}
-
 template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
 class tinyBLAS {
   public:
-    tinyBLAS(const ggml_compute_params * params, int64_t k,
+    tinyBLAS(int64_t k,
              const TA *A, int64_t lda,
              const TB *B, int64_t ldb,
-             TC *C, int64_t ldc)
+             TC *C, int64_t ldc,
-        : params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
+             int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
     }
 
-    bool matmul(int64_t m, int64_t n) {
+    void matmul(int64_t m, int64_t n) {
-        if (k % KN != 0)
+        mnpack(0, m, 0, n);
-            return false;
-        // compute RM for only need tile with size RM&RM-1
-#if VECTOR_REGISTERS == 32
-        if (m % 16 == 0 && (m/16 >= params->nth)) {
-            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
-            mnpack<4, 6, 4>(m, n, SIZE_N, 12);
-            return true;
-        }
-        if (m % 8 == 0 ) {
-            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
-            mnpack<4, 6, 2>(m, n, SIZE_N, 12);
-            return true;
-        }
-        if (m % 4 == 0) {
-            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
-            mnpack<4, 6, 1>(m, n, SIZE_N, 12);
-            return true;
-        }
-#else  // VECTOR_REGISTERS == 16
-        if (m % 16 == 0 && (m/16 >= params->nth)) {
-            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
-            mnpack<4, 3, 4>(m, n, SIZE_N, 24);
-            return true;
-        }
-        if (m % 8 == 0 ) {
-            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
-            mnpack<4, 3, 2>(m, n, SIZE_N, 24);
-            return true;
-        }
-        if (m % 4 == 0) {
-            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
-            mnpack<4, 3, 1>(m, n, SIZE_N, 24);
-            return true;
-        }
-#endif
-        return false;
     }
 
   private:
-    template <int RM, int RN, int BM>
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-    inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
+        int64_t mc, nc, mp, np;
-        if (SIZE_N == RN) {
+        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
-            return gemm<RM, RN, BM>(m, n, BN);
+#if VECTOR_REGISTERS == 32
-        }
+        case 0x55:
-        if constexpr (RN > 1) {
+            mc = 5;
-            return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
+            nc = 5;
-        } else {
+            gemm<5, 5>(m0, m, n0, n);
-            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            break;
-            GGML_ASSERT(false); // we have miss something.
+        case 0x45:
+            mc = 4;
+            nc = 5;
+            gemm<4, 5>(m0, m, n0, n);
+            break;
+        case 0x54:
+            mc = 5;
+            nc = 4;
+            gemm<5, 4>(m0, m, n0, n);
+            break;
+        case 0x44:
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
+            break;
+        case 0x53:
+            mc = 5;
+            nc = 3;
+            gemm<5, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+            mc = 3;
+            nc = 5;
+            gemm<3, 5>(m0, m, n0, n);
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+#else
+        case 0x55:
+        case 0x54:
+        case 0x53:
+        case 0x45:
+        case 0x44:
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+#endif
+        case 0x34:
+            mc = 3;
+            nc = 4;
+            gemm<3, 4>(m0, m, n0, n);
+            break;
+        case 0x52:
+            mc = 5;
+            nc = 2;
+            gemm<5, 2>(m0, m, n0, n);
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x25:
+            mc = 2;
+            nc = 5;
+            gemm<2, 5>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x51:
+            mc = 5;
+            nc = 1;
+            gemm<5, 1>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+            gemm<4, 1>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x15:
+            mc = 1;
+            nc = 5;
+            gemm<1, 5>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+            gemm<1, 4>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
         }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
     }
 
     template <int RM, int RN>
-    inline void gemm_bloc(int64_t ii, int64_t jj) {
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        D Cv[RN][RM] = {};
+        int64_t ytiles = (m - m0) / RM;
-        for (int64_t l = 0; l < k; l += KN) {
+        int64_t xtiles = (n - n0) / RN;
-            // help compiler for op order.
+        int64_t tiles = xtiles * ytiles;
-            if constexpr (RM <= RN) {
+        int64_t duty = (tiles + nth - 1) / nth;
-                V Av[RM];
+        int64_t start = duty * ith;
-                for (int64_t i = 0; i < RM; ++i) {
+        int64_t end = start + duty;
-                    Av[i] = load<V>(A + lda * (ii + i) + l);
+        if (end > tiles)
-                }
+            end = tiles;
-                for (int64_t j = 0; j < RN; ++j) {
+        for (int64_t job = start; job < end; ++job) {
-                    V Bv = load<V>(B + ldb * (jj + j) + l);
+            int64_t ii = m0 + job / xtiles * RM;
-                    for (int64_t i = 0; i < RM; ++i) {
+            int64_t jj = n0 + job % xtiles * RN;
-                        Cv[j][i] = madd(Av[i], Bv, Cv[j][i]);
+            D Cv[RN][RM] = {};
-                    }
+            for (int64_t l = 0; l < k; l += KN)
-                }
+                for (int64_t j = 0; j < RN; ++j)
-            } else {
+                    for (int64_t i = 0; i < RM; ++i)
-                V Bv[RN];
+                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
-                for (int64_t j = 0; j < RN; ++j) {
+                                        load<V>(B + ldb * (jj + j) + l),
-                    Bv[j] = load<V>(B + ldb * (jj + j) + l);
+                                        Cv[j][i]);
-                }
+            for (int64_t j = 0; j < RN; ++j)
-                for (int64_t i = 0; i < RM; ++i) {
+                for (int64_t i = 0; i < RM; ++i)
-                    V Av = load<V>(A + lda * (ii + i) + l);
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
-                    for (int64_t j = 0; j < RN; ++j) {
-                        Cv[j][i] = madd(Av, Bv[j], Cv[j][i]);
-                    }
-                }
-            }
         }
-        for (int64_t j = 0; j < RN; ++j)
-            for (int64_t i = 0; i < RM; ++i)
-                C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
     }
 
-    template <int RM, int RN, int BM>
-    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
-        static std::atomic<int64_t> current_chunk;
-
-        GGML_ASSERT(m % (RM * BM) == 0);
-        const int64_t ytiles = m / (RM * BM);
-        const int64_t xtiles = (n + RN -1) / RN;
-        const int64_t jj_RN = (xtiles - (xtiles * RN - n));
-
-        // "round" bloc_size to "nearest" BN
-        const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
-        const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
-        const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
-        const int64_t nb_job = ytiles * NB_BN;
-
-        if (params->ith == 0) {
-            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
-            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
-        }
-
-        ggml_barrier(params->threadpool);
-
-        int64_t job = params->ith;
-        while (job < nb_job) {
-            const int64_t ii = (job % ytiles) * RM * BM;
-            const int64_t jb =  job / ytiles;
-            const int64_t jr0 = BLOC_POS(jb  , jj_BN, SIZE_BN);
-            const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
-
-            const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
-            const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
-            const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
-
-            for (int64_t bi = 0; bi < BM * RM; bi += RM) {
-                int64_t jj = jj0;
-                for (; jj < jj1; jj += RN) {
-                    gemm_bloc<RM, RN>(ii + bi, jj);
-                }
-                if constexpr (RN > 1) {
-                    for (; jj < jj2; jj += RN - 1) {
-                        gemm_bloc<RM, RN-1>(ii + bi, jj);
-                    }
-                }
-                GGML_ASSERT(jj == jj2);
-            }
-
-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
-        }
-
-        ggml_barrier(params->threadpool);
-        return;
-    }
-
-    const ggml_compute_params * params;
     const TA *const A;
     const TB *const B;
     TC *const C;
@@ -461,6 +452,8 @@ class tinyBLAS {
     const int64_t lda;
     const int64_t ldb;
     const int64_t ldc;
+    const int ith;
+    const int nth;
 };
 
 //////////////////////////////////////////////////////////////////////////////////////////
@@ -1664,9 +1657,8 @@ class tinyBLAS_PPC {
  * @param Ctype is GGML data type of `C`
  * @return true if this function was able to service the matmul request
  */
-bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
+bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
-                     const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
-                     int64_t ldc, int Atype, int Btype, int Ctype) {
 
     assert(m >= 0);
     assert(n >= 0);
@@ -1674,8 +1666,8 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
     assert(lda >= k);
     assert(ldb >= k);
     assert(ldc >= m);
-    assert(params->nth > 0);
+    assert(nth > 0);
-    assert(params->ith < params->nth);
+    assert(ith < nth);
 
     // only enable sgemm for prompt processing
     if (n < 2)
@@ -1690,25 +1682,37 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
         if (Btype != GGML_TYPE_F32)
             return false;
 #if defined(__AVX512F__)
-        tinyBLAS<16, __m512, __m512, float, float, float> tb{ params,
+        if (k % 16)
+            return false;
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc};
+            (float *)C, ldc,
-        return tb.matmul(m, n);
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif defined(__AVX__) || defined(__AVX2__)
-        tinyBLAS<8, __m256, __m256, float, float, float> tb{ params,
+        if (k % 8)
+            return false;
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc};
+            (float *)C, ldc,
-        return tb.matmul(m, n);
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif defined(__ARM_NEON)
         if (n < 4)
             return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
+        if (k % 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc};
+            (float *)C, ldc,
-        return tb.matmul(m, n);
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif defined(__MMA__)
         if (k % 8)
             return false;
@@ -1716,7 +1720,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const float *)A, lda,
             (const float *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1724,71 +1728,60 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
 #endif
     }
 
-    case GGML_TYPE_BF16: {
-#if defined(__AVX512BF16__)
-        if (Btype == GGML_TYPE_BF16) {
-            tinyBLAS<32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
-                (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-            return tb.matmul(m, n);
-        }
-#elif defined(__AVX512F__)
-        if (Btype == GGML_TYPE_BF16) {
-            tinyBLAS<16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
-                (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-            return tb.matmul(m, n);
-        }
-#elif defined(__AVX2__)
-        if (Btype == GGML_TYPE_BF16) {
-            tinyBLAS<8, __m256, __m256, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
-                (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-            return tb.matmul(m, n);
-        }
-#endif
-        return false;
-    }
     case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
-        if (Btype == GGML_TYPE_F16) {
+        if (k % 16)
-            tinyBLAS<16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+            return false;
-                (const ggml_fp16_t *)A, lda,
+        if (Btype != GGML_TYPE_F32)
-                (const ggml_fp16_t *)B, ldb,
+            return false;
-                (float *)C, ldc};
+        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
-            return tb.matmul(m, n);
+            k, (const ggml_fp16_t *)A, lda,
-        }
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
-        if (Btype == GGML_TYPE_F16) {
+        if (k % 8)
-            tinyBLAS<8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+            return false;
-                (const ggml_fp16_t *)A, lda,
+        if (Btype != GGML_TYPE_F32)
-                (const ggml_fp16_t *)B, ldb,
+            return false;
-                (float *)C, ldc};
+        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
-            return tb.matmul(m, n);
+            k, (const ggml_fp16_t *)A, lda,
-        }
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
         if (n < 8)
             return false;
-        if (Btype == GGML_TYPE_F16) {
+        if (k % 8)
-            tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+            return false;
-                k, (const ggml_fp16_t *)A, lda,
+        if (Btype != GGML_TYPE_F16)
-                (const ggml_fp16_t *)B, ldb,
+            return false;
-                (float *)C, ldc};
+        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
-            return tb.matmul(m, n);
+            k, (const ggml_fp16_t *)A, lda,
-        }
+            (const ggml_fp16_t *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
 #elif defined(__ARM_NEON) && !defined(_MSC_VER)
-        if (Btype == GGML_TYPE_F32) {
+        if (k % 4)
-            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{ params,
+            return false;
-                k, (const ggml_fp16_t *)A, lda,
+        if (Btype != GGML_TYPE_F32)
-                (const float *)B, ldb,
+            return false;
-                (float *)C, ldc};
+        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
-            return tb.matmul(m, n);
+            k, (const ggml_fp16_t *)A, lda,
-        }
+            (const float *)B, ldb,
-#endif
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
         return false;
+#endif
     }
 
     case GGML_TYPE_Q8_0: {
@@ -1799,7 +1792,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_q8_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@@ -1807,7 +1800,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_q8_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1823,7 +1816,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_q4_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@@ -1831,7 +1824,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_q4_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1847,7 +1840,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_q5_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1863,7 +1856,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             k, (const block_iq4_nl *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            params->ith, params->nth};
+            ith, nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1875,7 +1868,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
         return false;
     }
 
-    (void)params;
     (void)m;
     (void)n;
     (void)k;
@@ -1885,6 +1877,8 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
     (void)ldb;
     (void)C;
     (void)ldc;
+    (void)ith;
+    (void)nth;
     (void)Atype;
     (void)Btype;
     (void)Ctype;

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -5,8 +5,8 @@
 extern "C" {
 #endif
 
-bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
+bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
-                     const void *, int64_t, const void *, int64_t, void *, int64_t,
+                     const void *, int64_t, void *, int64_t, int, int,
                      int, int, int);
 
 #ifdef __cplusplus

scripts/compare-llama-bench.py

@@ -126,8 +126,6 @@ connection = sqlite3.connect(input_file)
 cursor = connection.cursor()
 builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
 
-commit_short_len = len(builds[0][0])
-
 try:
     repo = git.Repo(".", search_parent_directories=True)
 except git.InvalidGitRepositoryError:
@@ -140,11 +138,11 @@ def find_parent_in_data(commit: git.Commit):
     seen_hexsha8 = set()
     while heap:
         depth, current_commit = heapq.heappop(heap)
-        current_hexsha8 = commit.hexsha[:commit_short_len]
+        current_hexsha8 = commit.hexsha[:8]
         if (current_hexsha8,) in builds:
             return current_hexsha8
         for parent in commit.parents:
-            parent_hexsha8 = parent.hexsha[:commit_short_len]
+            parent_hexsha8 = parent.hexsha[:8]
             if parent_hexsha8 not in seen_hexsha8:
                 seen_hexsha8.add(parent_hexsha8)
                 heapq.heappush(heap, (depth + 1, parent))
@@ -158,9 +156,9 @@ def get_all_parent_hexsha8s(commit: git.Commit):
 
     while unvisited:
         current_commit = unvisited.pop(0)
-        visited.append(current_commit.hexsha[:commit_short_len])
+        visited.append(current_commit.hexsha[:8])
         for parent in current_commit.parents:
-            if parent.hexsha[:commit_short_len] not in visited:
+            if parent.hexsha[:8] not in visited:
                 unvisited.append(parent)
 
     return visited
@@ -171,10 +169,10 @@ def get_commit_name(hexsha8):
     if repo is None:
         return hexsha8
     for h in repo.heads:
-        if h.commit.hexsha[:commit_short_len] == hexsha8:
+        if h.commit.hexsha[:8] == hexsha8:
             return h.name
     for t in repo.tags:
-        if t.commit.hexsha[:commit_short_len] == hexsha8:
+        if t.commit.hexsha[:8] == hexsha8:
             return t.name
     return hexsha8
 
@@ -185,13 +183,13 @@ def get_commit_hexsha8(name):
         return None
     for h in repo.heads:
         if h.name == name:
-            return h.commit.hexsha[:commit_short_len]
+            return h.commit.hexsha[:8]
     for t in repo.tags:
         if t.name == name:
-            return t.commit.hexsha[:commit_short_len]
+            return t.commit.hexsha[:8]
     for c in repo.iter_commits("--all"):
-        if c.hexsha[:commit_short_len] == name[:commit_short_len]:
+        if c.hexsha[:8] == name[:8]:
-            return c.hexsha[:commit_short_len]
+            return c.hexsha[:8]
     return None
 
 

scripts/hf.sh

@@ -26,7 +26,7 @@ function has_cmd {
 }
 
 if has_cmd wget; then
-    cmd="wget -q -c -O %s/%s %s"
+    cmd="wget -q --show-progress -c -O %s/%s %s"
 elif has_cmd curl; then
     cmd="curl -C - -f --output-dir %s -o %s -L %s"
 else

src/llama-vocab.cpp

@@ -1657,7 +1657,7 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
 }
 
 llama_token llama_token_bos_impl(const struct llama_vocab & vocab) {
-    return vocab.type != LLAMA_VOCAB_TYPE_WPM ? vocab.special_bos_id : vocab.special_cls_id;
+    return vocab.special_bos_id;
 }
 
 llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {

src/llama-vocab.h

@@ -45,7 +45,7 @@ struct llama_vocab {
     id special_unk_id  = 0;
     id special_sep_id  = LLAMA_TOKEN_NULL;
     id special_pad_id  = LLAMA_TOKEN_NULL;
-    id special_cls_id  = LLAMA_TOKEN_NULL; // TODO: revisit if this is really needed https://github.com/ggerganov/llama.cpp/pull/10930
+    id special_cls_id  = LLAMA_TOKEN_NULL;
     id special_mask_id = LLAMA_TOKEN_NULL;
 
     id linefeed_id = 13;