Simplify and improve CUDA graphs through use of indirect copy pointers (#9017)

* CUDA: Simplify and improve CUDA graphs through use of indirect copy pointers

Previously there was complexity in the CUDA graphs implementation due
frequently changing parameters to copy kernels associated with K and V
cache pointers. This patch simplifies by using indirection to avoid
such parameters frequently changing, avoiding the need for frequent
graph updates.

Fixes #12152

* Addressed comments

* fix HIP builds

* properly sync to stream

* removed ggml_cuda_cpy_fn_ptrs

* move stream sync before free

* guard to only use indirection with graphs

* style fixes

* check for errors

---------

Co-authored-by: slaren <slarengh@gmail.com>

.github/workflows/build.yml

@@ -803,7 +803,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.304.1
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:

README.md

@@ -112,6 +112,8 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
+- [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
+- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 
 #### Multimodal
 

ci/README.md

@@ -60,7 +60,7 @@ docker run --privileged -it \
 Inside the container, execute the following commands:
 
 ```bash
-apt update -y && apt install -y bc cmake git python3.10-venv time unzip wget
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
 git config --global --add safe.directory /ws
 GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
 ```

ci/run.sh

@@ -69,7 +69,7 @@ fi
 if [ ! -z ${GG_BUILD_MUSA} ]; then
     # Use qy1 by default (MTT S80)
     MUSA_ARCH=${MUSA_ARCH:-21}
-    CMAKE_EXTRA="-DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
 fi
 ## helpers
 

common/arg.cpp

@@ -1,9 +1,20 @@
+#include "gguf.h" // for reading GGUF splits
 #include "arg.h"
 
+#include "common.h"
 #include "log.h"
 #include "sampling.h"
 #include "chat.h"
 
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 #include <algorithm>
 #include <climits>
 #include <cstdarg>
@@ -14,6 +25,14 @@
 #include <thread>
 #include <vector>
 
+//#define LLAMA_USE_CURL
+
+#if defined(LLAMA_USE_CURL)
+#include <curl/curl.h>
+#include <curl/easy.h>
+#include <future>
+#endif
+
 #include "json-schema-to-grammar.h"
 
 using json = nlohmann::ordered_json;
@@ -125,47 +144,549 @@ std::string common_arg::to_string() {
     return ss.str();
 }
 
+//
+// downloader
+//
+
+struct common_hf_file_res {
+    std::string repo; // repo name with ":tag" removed
+    std::string ggufFile;
+    std::string mmprojFile;
+};
+
+#ifdef LLAMA_USE_CURL
+
+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
+#else
+#include <sys/syslimits.h>
+#endif
+#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
+
+//
+// CURL utils
+//
+
+using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
+
+// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
+struct curl_slist_ptr {
+    struct curl_slist * ptr = nullptr;
+    ~curl_slist_ptr() {
+        if (ptr) {
+            curl_slist_free_all(ptr);
+        }
+    }
+};
+
+#define CURL_MAX_RETRY 3
+#define CURL_RETRY_DELAY_SECONDS 2
+
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+    int remaining_attempts = max_attempts;
+
+    while (remaining_attempts > 0) {
+        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+
+        CURLcode res = curl_easy_perform(curl);
+        if (res == CURLE_OK) {
+            return true;
+        }
+
+        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
+        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
+
+        remaining_attempts--;
+        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+    }
+
+    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+
+    return false;
+}
+
+// download one single file from remote URL to local path
+static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) {
+    // Initialize libcurl
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    if (!curl) {
+        LOG_ERR("%s: error initializing libcurl\n", __func__);
+        return false;
+    }
+
+    bool force_download = false;
+
+    // Set the URL, allow to follow http redirection
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+
+    // Check if hf-token or bearer-token was specified
+    if (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    }
+
+#if defined(_WIN32)
+    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
+    //   operating system. Currently implemented under MS-Windows.
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+
+    // Check if the file already exists locally
+    auto file_exists = std::filesystem::exists(path);
+
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    nlohmann::json metadata;
+    std::string etag;
+    std::string last_modified;
+
+    if (file_exists) {
+        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+        std::ifstream metadata_in(metadata_path);
+        if (metadata_in.good()) {
+            try {
+                metadata_in >> metadata;
+                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
+                if (metadata.contains("url") && metadata.at("url").is_string()) {
+                    auto previous_url = metadata.at("url").get<std::string>();
+                    if (previous_url != url) {
+                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
+                        return false;
+                    }
+                }
+                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                    etag = metadata.at("etag");
+                }
+                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                    last_modified = metadata.at("lastModified");
+                }
+            } catch (const nlohmann::json::exception & e) {
+            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+                return false;
+            }
+        }
+    } else {
+        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
+    }
+
+    // Send a HEAD request to retrieve the etag and last-modified headers
+    struct common_load_model_from_url_headers {
+        std::string etag;
+        std::string last_modified;
+    };
+
+    common_load_model_from_url_headers headers;
+
+    {
+        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
+        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
+            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
+
+            static std::regex header_regex("([^:]+): (.*)\r\n");
+            static std::regex etag_regex("ETag", std::regex_constants::icase);
+            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
+
+            std::string header(buffer, n_items);
+            std::smatch match;
+            if (std::regex_match(header, match, header_regex)) {
+                const std::string & key = match[1];
+                const std::string & value = match[2];
+                if (std::regex_match(key, match, etag_regex)) {
+                    headers->etag = value;
+                } else if (std::regex_match(key, match, last_modified_regex)) {
+                    headers->last_modified = value;
+                }
+            }
+            return n_items;
+        };
+
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
+
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code != 200) {
+            // HEAD not supported, we don't know if the file has changed
+            // force trigger downloading
+            force_download = true;
+            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        }
+    }
+
+    bool should_download = !file_exists || force_download;
+    if (!should_download) {
+        if (!etag.empty() && etag != headers.etag) {
+            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
+            should_download = true;
+        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
+            should_download = true;
+        }
+    }
+    if (should_download) {
+        std::string path_temporary = path + ".downloadInProgress";
+        if (file_exists) {
+            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
+            if (remove(path.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+                return false;
+            }
+        }
+
+        // Set the output file
+
+        struct FILE_deleter {
+            void operator()(FILE * f) const {
+                fclose(f);
+            }
+        };
+
+        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
+        if (!outfile) {
+            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
+            return false;
+        }
+
+        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
+        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
+            return fwrite(data, size, nmemb, (FILE *)fd);
+        };
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
+
+        //  display download progress
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
+
+        // helper function to hide password in URL
+        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
+            std::size_t protocol_pos = url.find("://");
+            if (protocol_pos == std::string::npos) {
+                return url;  // Malformed URL
+            }
+
+            std::size_t at_pos = url.find('@', protocol_pos + 3);
+            if (at_pos == std::string::npos) {
+                return url;  // No password in URL
+            }
+
+            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
+        };
+
+        // start the download
+        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code < 200 || http_code >= 400) {
+            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
+            return false;
+        }
+
+        // Causes file to be closed explicitly here before we rename it.
+        outfile.reset();
+
+        // Write the updated JSON metadata file.
+        metadata.update({
+            {"url", url},
+            {"etag", headers.etag},
+            {"lastModified", headers.last_modified}
+        });
+        std::ofstream(metadata_path) << metadata.dump(4);
+        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+
+        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
+            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+            return false;
+        }
+    }
+
+    return true;
+}
+
+// download multiple files from remote URLs to local paths
+// the input is a vector of pairs <url, path>
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token) {
+    // Prepare download in parallel
+    std::vector<std::future<bool>> futures_download;
+    for (auto const & item : urls) {
+        futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair<std::string, std::string> & it) -> bool {
+            return common_download_file_single(it.first, it.second, bearer_token);
+        }, item));
+    }
+
+    // Wait for all downloads to complete
+    for (auto & f : futures_download) {
+        if (!f.get()) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool common_download_model(
+        const common_params_model & model,
+        const std::string & bearer_token) {
+    // Basic validation of the model.url
+    if (model.url.empty()) {
+        LOG_ERR("%s: invalid model url\n", __func__);
+        return false;
+    }
+
+    if (!common_download_file_single(model.url, model.path, bearer_token)) {
+        return false;
+    }
+
+    // check for additional GGUFs split to download
+    int n_split = 0;
+    {
+        struct gguf_init_params gguf_params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ NULL,
+        };
+        auto * ctx_gguf = gguf_init_from_file(model.path.c_str(), gguf_params);
+        if (!ctx_gguf) {
+            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, model.path.c_str());
+            return false;
+        }
+
+        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+        if (key_n_split >= 0) {
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+        }
+
+        gguf_free(ctx_gguf);
+    }
+
+    if (n_split > 1) {
+        char split_prefix[PATH_MAX] = {0};
+        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+
+        // Verify the first split file format
+        // and extract split URL and PATH prefixes
+        {
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), model.path.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split);
+                return false;
+            }
+
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model.url.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model.url.c_str(), n_split);
+                return false;
+            }
+        }
+
+        std::vector<std::pair<std::string, std::string>> urls;
+        for (int idx = 1; idx < n_split; idx++) {
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
+
+            char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+            llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);
+
+            if (std::string(split_path) == model.path) {
+                continue; // skip the already downloaded file
+            }
+
+            urls.push_back({split_url, split_path});
+        }
+
+        // Download in parallel
+        common_download_file_multiple(urls, bearer_token);
+    }
+
+    return true;
+}
+
+/**
+ * Allow getting the HF file from the HF repo with tag (like ollama), for example:
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
+ * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
+ *
+ * Return pair of <repo, file> (with "repo" already having tag removed)
+ *
+ * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
+ */
+static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token) {
+    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
+    std::string tag = parts.size() > 1 ? parts.back() : "latest";
+    std::string hf_repo = parts[0];
+    if (string_split<std::string>(hf_repo, '/').size() != 2) {
+        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
+    }
+
+    // fetch model info from Hugging Face Hub API
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::string res_str;
+    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+    }
+    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        throw std::runtime_error("error: cannot make GET request to HF API");
+    }
+
+    long res_code;
+    std::string ggufFile   = "";
+    std::string mmprojFile = "";
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+    if (res_code == 200) {
+        // extract ggufFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                ggufFile = match[1].str();
+            }
+        }
+        // extract mmprojFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                mmprojFile = match[1].str();
+            }
+        }
+    } else if (res_code == 401) {
+        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
+    } else {
+        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
+    }
+
+    // check response
+    if (ggufFile.empty()) {
+        throw std::runtime_error("error: model does not have ggufFile");
+    }
+
+    return { hf_repo, ggufFile, mmprojFile };
+}
+
+#else
+
+static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from internet\n");
+    return false;
+}
+
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static bool common_download_model(
+        const common_params_model &,
+        const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return {};
+}
+
+#endif // LLAMA_USE_CURL
+
 //
 // utils
 //
 
-static void common_params_handle_model_default(
-        std::string & model,
-        const std::string & model_url,
-        std::string & hf_repo,
-        std::string & hf_file,
-        const std::string & hf_token,
-        const std::string & model_default) {
-    if (!hf_repo.empty()) {
-        // short-hand to avoid specifying --hf-file -> default it to --model
-        if (hf_file.empty()) {
-            if (model.empty()) {
-                auto auto_detected = common_get_hf_file(hf_repo, hf_token);
-                if (auto_detected.first.empty() || auto_detected.second.empty()) {
-                    exit(1); // built without CURL, error message already printed
+static void common_params_handle_model(
+        struct common_params_model & model,
+        const std::string & bearer_token,
+        const std::string & model_path_default,
+        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+    // handle pre-fill default model path and url based on hf_repo and hf_file
+    {
+        if (!model.hf_repo.empty()) {
+            // short-hand to avoid specifying --hf-file -> default it to --model
+            if (model.hf_file.empty()) {
+                if (model.path.empty()) {
+                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token);
+                    if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
+                        exit(1); // built without CURL, error message already printed
+                    }
+                    model.hf_repo = auto_detected.repo;
+                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                } else {
+                    model.hf_file = model.path;
                 }
-                hf_repo = auto_detected.first;
-                hf_file = auto_detected.second;
-            } else {
-                hf_file = model;
             }
+
+            // TODO: allow custom host
+            model.url = "https://huggingface.co/" + model.hf_repo + "/resolve/main/" + model.hf_file;
+
+            // make sure model path is present (for caching purposes)
+            if (model.path.empty()) {
+                // this is to avoid different repo having same file name, or same file name in different subdirs
+                std::string filename = model.hf_repo + "_" + model.hf_file;
+                // to make sure we don't have any slashes in the filename
+                string_replace_all(filename, "/", "_");
+                model.path = fs_get_cache_file(filename);
+            }
+
+        } else if (!model.url.empty()) {
+            if (model.path.empty()) {
+                auto f = string_split<std::string>(model.url, '#').front();
+                f = string_split<std::string>(f, '?').front();
+                model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            }
+
+        } else if (model.path.empty()) {
+            model.path = model_path_default;
         }
-        // make sure model path is present (for caching purposes)
-        if (model.empty()) {
-            // this is to avoid different repo having same file name, or same file name in different subdirs
-            std::string filename = hf_repo + "_" + hf_file;
-            // to make sure we don't have any slashes in the filename
-            string_replace_all(filename, "/", "_");
-            model = fs_get_cache_file(filename);
+    }
+
+    // then, download it if needed
+    if (!model.url.empty()) {
+        bool ok = common_download_model(model, bearer_token);
+        if (!ok) {
+            LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
+            exit(1);
         }
-    } else if (!model_url.empty()) {
-        if (model.empty()) {
-            auto f = string_split<std::string>(model_url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            model = fs_get_cache_file(string_split<std::string>(f, '/').back());
-        }
-    } else if (model.empty()) {
-        model = model_default;
     }
 }
 
@@ -300,10 +821,16 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // TODO: refactor model params in a common struct
-    common_params_handle_model_default(params.model,             params.model_url,             params.hf_repo,             params.hf_file,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
-    common_params_handle_model_default(params.vocoder.model,     params.vocoder.model_url,     params.vocoder.hf_repo,     params.vocoder.hf_file,     params.hf_token, "");
+    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
+    common_params_handle_model(params.speculative.model, params.hf_token, "");
+    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
+
+    // allow --mmproj to be set from -hf
+    // assuming that mmproj is always in the same repo as text model
+    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
+        params.mmproj.hf_repo = params.model.hf_repo;
+    }
+    common_params_handle_model(params.mmproj,            params.hf_token, "", true);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -322,6 +849,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.kv_overrides.back().key[0] = 0;
     }
 
+    if (!params.tensor_buft_overrides.empty()) {
+        params.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
+
     if (params.reranking && params.embedding) {
         throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
     }
@@ -1561,7 +2092,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--mmproj"}, "FILE",
         "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
-            params.mmproj = value;
+            params.mmproj.path = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    add_opt(common_arg(
+        {"--mmproj-url"}, "URL",
+        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        [](common_params & params, const std::string & value) {
+            params.mmproj.url = value;
         }
     ).set_examples({LLAMA_EXAMPLE_LLAVA}));
     add_opt(common_arg(
@@ -1647,6 +2185,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
+    add_opt(common_arg(
+        {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
+        "override tensor buffer type", [](common_params & params, const std::string & value) {
+            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+            if (buft_list.empty()) {
+                // enumerate all the devices and add their buffer types to the list
+                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    auto * buft = ggml_backend_dev_buffer_type(dev);
+                    if (buft) {
+                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                    }
+                }
+            }
+
+            for (const auto & override : string_split<std::string>(value, ',')) {
+                std::string::size_type pos = override.find('=');
+                if (pos == std::string::npos) {
+                    throw std::invalid_argument("invalid value");
+                }
+                std::string tensor_name = override.substr(0, pos);
+                std::string buffer_type = override.substr(pos + 1);
+
+                if (buft_list.find(buffer_type) == buft_list.end()) {
+                    printf("Available buffer types:\n");
+                    for (const auto & it : buft_list) {
+                        printf("  %s\n", ggml_backend_buft_name(it.second));
+                    }
+                    throw std::invalid_argument("unknown buffer type");
+                }
+                // FIXME: this leaks memory
+                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+            }
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1790,14 +2363,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
             ),
         [](common_params & params, const std::string & value) {
-            params.model = value;
+            params.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.model_url = value;
+            params.model.url = value;
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
@@ -1806,35 +2379,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "example: unsloth/phi-4-GGUF:q4_k_m\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_repo = value;
+            params.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO"));
     add_opt(common_arg(
         {"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
         "Same as --hf-repo, but for the draft model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.hf_repo = value;
+            params.speculative.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HFD_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
         "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_file = value;
+            params.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE"));
     add_opt(common_arg(
         {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
         "Hugging Face model repository for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_repo = value;
+            params.vocoder.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO_V"));
     add_opt(common_arg(
         {"-hffv", "--hf-file-v"}, "FILE",
         "Hugging Face model file for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_file = value;
+            params.vocoder.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE_V"));
     add_opt(common_arg(
@@ -1979,7 +2552,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
-        string_format("ip address to listen (default: %s)", params.hostname.c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
         [](common_params & params, const std::string & value) {
             params.hostname = value;
         }
@@ -2454,7 +3027,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.model = value;
+            params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
 
@@ -2462,7 +3035,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-mv", "--model-vocoder"}, "FNAME",
         "vocoder model for audio generation (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.model = value;
+            params.vocoder.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
      add_opt(common_arg(
@@ -2485,10 +3058,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf";
+            params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
+            params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
+            params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
+            params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
@@ -2496,8 +3069,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-bge-small-en-default"},
         string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
-            params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
+            params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2510,8 +3083,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-e5-small-en-default"},
         string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
-            params.hf_file = "e5-small-v2-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
+            params.model.hf_file = "e5-small-v2-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2524,8 +3097,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-gte-small-default"},
         string_format("use default gte-small model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
-            params.hf_file = "gte-small-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
+            params.model.hf_file = "gte-small-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2538,8 +3111,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-1.5b-default"},
         string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2554,8 +3127,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-3b-default"},
         string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2570,8 +3143,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-7b-default"},
         string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2586,10 +3159,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-7b-spec"},
         string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
-            params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.speculative.n_gpu_layers = 99;
             params.port = 8012;
             params.n_gpu_layers = 99;
@@ -2605,10 +3178,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-14b-spec"},
         string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
-            params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.speculative.n_gpu_layers = 99;
             params.port = 8012;
             params.n_gpu_layers = 99;

common/common.cpp

@@ -7,9 +7,6 @@
 
 #include "common.h"
 #include "log.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
 #include "llama.h"
 
 #include <algorithm>
@@ -51,47 +48,11 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #endif
-#if defined(LLAMA_USE_CURL)
-#include <curl/curl.h>
-#include <curl/easy.h>
-#include <future>
-#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if defined(LLAMA_USE_CURL)
-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#else
-#include <sys/syslimits.h>
-#endif
-#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
-//
-// CURL utils
-//
-
-using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
-
-// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
-struct curl_slist_ptr {
-    struct curl_slist * ptr = nullptr;
-    ~curl_slist_ptr() {
-        if (ptr) {
-            curl_slist_free_all(ptr);
-        }
-    }
-};
-#endif // LLAMA_USE_CURL
-
-using json = nlohmann::ordered_json;
-
 //
 // CPU utils
 //
@@ -900,22 +861,14 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
 
-    llama_model * model = nullptr;
-
-    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
-    } else if (!params.model_url.empty()) {
-        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
-    } else {
-        model = llama_model_load_from_file(params.model.c_str(), mparams);
-    }
-
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -950,7 +903,7 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
@@ -1089,15 +1042,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (!params.devices.empty()) {
         mparams.devices = params.devices.data();
     }
+
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1105,6 +1061,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
+    if (params.tensor_buft_overrides.empty()) {
+        mparams.tensor_buft_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
+    }
+
     return mparams;
 }
 
@@ -1164,451 +1127,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     return tpp;
 }
 
-#ifdef LLAMA_USE_CURL
-
-#define CURL_MAX_RETRY 3
-#define CURL_RETRY_DELAY_SECONDS 2
-
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
-    int remaining_attempts = max_attempts;
-
-    while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
-
-        CURLcode res = curl_easy_perform(curl);
-        if (res == CURLE_OK) {
-            return true;
-        }
-
-        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
-        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
-
-        remaining_attempts--;
-        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-    }
-
-    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
-
-    return false;
-}
-
-static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
-    // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    if (!curl) {
-        LOG_ERR("%s: error initializing libcurl\n", __func__);
-        return false;
-    }
-
-    bool force_download = false;
-
-    // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-
-    // Check if hf-token or bearer-token was specified
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-    }
-
-#if defined(_WIN32)
-    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
-    //   operating system. Currently implemented under MS-Windows.
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
-
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
-    std::string etag;
-    std::string last_modified;
-
-    if (file_exists) {
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
-            }
-        }
-    } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
-    }
-
-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
-    };
-
-    common_load_model_from_url_headers headers;
-
-    {
-        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
-
-            static std::regex header_regex("([^:]+): (.*)\r\n");
-            static std::regex etag_regex("ETag", std::regex_constants::icase);
-            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-            std::string header(buffer, n_items);
-            std::smatch match;
-            if (std::regex_match(header, match, header_regex)) {
-                const std::string & key = match[1];
-                const std::string & value = match[2];
-                if (std::regex_match(key, match, etag_regex)) {
-                    headers->etag = value;
-                } else if (std::regex_match(key, match, last_modified_regex)) {
-                    headers->last_modified = value;
-                }
-            }
-            return n_items;
-        };
-
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-        }
-    }
-
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
-        if (!etag.empty() && etag != headers.etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
-            should_download = true;
-        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
-            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
-            should_download = true;
-        }
-    }
-    if (should_download) {
-        std::string path_temporary = path + ".downloadInProgress";
-        if (file_exists) {
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
-            }
-        }
-
-        // Set the output file
-
-        struct FILE_deleter {
-            void operator()(FILE * f) const {
-                fclose(f);
-            }
-        };
-
-        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
-        if (!outfile) {
-            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
-            return false;
-        }
-
-        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
-        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
-            return fwrite(data, size, nmemb, (FILE *)fd);
-        };
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
-
-        //  display download progress
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
-
-        // helper function to hide password in URL
-        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
-            std::size_t protocol_pos = url.find("://");
-            if (protocol_pos == std::string::npos) {
-                return url;  // Malformed URL
-            }
-
-            std::size_t at_pos = url.find('@', protocol_pos + 3);
-            if (at_pos == std::string::npos) {
-                return url;  // No password in URL
-            }
-
-            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
-        };
-
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code < 200 || http_code >= 400) {
-            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-            return false;
-        }
-
-        // Causes file to be closed explicitly here before we rename it.
-        outfile.reset();
-
-        // Write the updated JSON metadata file.
-        metadata.update({
-            {"url", url},
-            {"etag", headers.etag},
-            {"lastModified", headers.last_modified}
-        });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
-
-        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
-        }
-    }
-
-    return true;
-}
-
-struct llama_model * common_load_model_from_url(
-        const std::string & model_url,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // Basic validation of the model_url
-    if (model_url.empty()) {
-        LOG_ERR("%s: invalid model_url\n", __func__);
-        return NULL;
-    }
-
-    if (!common_download_file(model_url, local_path, hf_token)) {
-        return NULL;
-    }
-
-    // check for additional GGUFs split to download
-    int n_split = 0;
-    {
-        struct gguf_init_params gguf_params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ NULL,
-        };
-        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
-        if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
-            return NULL;
-        }
-
-        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-        if (key_n_split >= 0) {
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-        }
-
-        gguf_free(ctx_gguf);
-    }
-
-    if (n_split > 1) {
-        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-
-        // Verify the first split file format
-        // and extract split URL and PATH prefixes
-        {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
-                return NULL;
-            }
-
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
-                return NULL;
-            }
-        }
-
-        // Prepare download in parallel
-        std::vector<std::future<bool>> futures_download;
-        for (int idx = 1; idx < n_split; idx++) {
-            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
-                char split_path[PATH_MAX] = {0};
-                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
-
-                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
-
-                return common_download_file(split_url, split_path, hf_token);
-            }, idx));
-        }
-
-        // Wait for all downloads to complete
-        for (auto & f : futures_download) {
-            if (!f.get()) {
-                return NULL;
-            }
-        }
-    }
-
-    return llama_model_load_from_file(local_path.c_str(), params);
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & repo,
-        const std::string & remote_path,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // construct hugging face model url:
-    //
-    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
-    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
-    //
-    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //
-
-    std::string model_url = "https://huggingface.co/";
-    model_url += repo;
-    model_url += "/resolve/main/";
-    model_url += remote_path;
-
-    return common_load_model_from_url(model_url, local_path, hf_token, params);
-}
-
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
-
-    // fetch model info from Hugging Face Hub API
-    json model_info;
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
-        model_info = json::parse(res_str);
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
-    } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
-    }
-
-    // check response
-    if (!model_info.contains("ggufFile")) {
-        throw std::runtime_error("error: model does not have ggufFile");
-    }
-    json & gguf_file = model_info.at("ggufFile");
-    if (!gguf_file.contains("rfilename")) {
-        throw std::runtime_error("error: ggufFile does not have rfilename");
-    }
-
-    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
-}
-
-#else
-
-struct llama_model * common_load_model_from_url(
-        const std::string & /*model_url*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
-    return nullptr;
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & /*repo*/,
-        const std::string & /*remote_path*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return nullptr;
-}
-
-std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return std::make_pair("", "");
-}
-
-#endif // LLAMA_USE_CURL
-
 //
 // Batch utils
 //
@@ -2032,26 +1550,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
-
-template <>
-json common_grammar_trigger::to_json() const {
-    json out {
-        {"type", (int) type},
-        {"value", value},
-    };
-    if (type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
-        out["token"] = (int) token;
-    }
-    return out;
-}
-
-template <>
-common_grammar_trigger common_grammar_trigger::from_json(const json & in) {
-    common_grammar_trigger out;
-    out.type = (common_grammar_trigger_type) in.at("type").get<int>();
-    out.value = in.at("value").get<std::string>();
-    if (out.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
-        out.token = (llama_token) in.at("token").get<int>();
-    }
-    return out;
-}

common/common.h

@@ -121,10 +121,6 @@ struct common_grammar_trigger {
     common_grammar_trigger_type type;
     std::string value;
     llama_token token = LLAMA_TOKEN_NULL;
-
-    // T can only be nlohmann::ordered_json
-    template <class T> T to_json() const;
-    template <class T> static common_grammar_trigger from_json(const T & in);
 };
 
 // sampling parameters
@@ -184,6 +180,13 @@ struct common_params_sampling {
     std::string print() const;
 };
 
+struct common_params_model {
+    std::string path    = ""; // model local path                                           // NOLINT
+    std::string url     = ""; // model url to download                                      // NOLINT
+    std::string hf_repo = ""; // HF repo                                                    // NOLINT
+    std::string hf_file = ""; // HF file                                                    // NOLINT
+};
+
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
@@ -197,19 +200,11 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model = "";     // draft model for speculative decoding                      // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    struct common_params_model model;
 };
 
 struct common_params_vocoder {
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model     = ""; // model path                                                // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    struct common_params_model model;
 
     std::string speaker_file = ""; // speaker file path                                      // NOLINT
 
@@ -267,12 +262,10 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    std::string model                = ""; // model path                                                    // NOLINT
+    struct common_params_model model;
+
     std::string model_alias          = ""; // model alias                                                   // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
@@ -286,6 +279,7 @@ struct common_params {
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
     std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
@@ -347,7 +341,7 @@ struct common_params {
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    struct common_params_model mmproj;
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -546,23 +540,6 @@ struct llama_model_params     common_model_params_to_llama  (      common_params
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
-struct llama_model * common_load_model_from_url(
-    const std::string & model_url,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-struct llama_model * common_load_model_from_hf(
-    const std::string & repo,
-    const std::string & remote_path,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-std::pair<std::string, std::string> common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & hf_token);
-
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 

common/minja/minja.hpp

@@ -240,7 +240,7 @@ public:
       auto index = key.get<int>();
       return array_->at(index < 0 ? array_->size() + index : index);
     } else if (object_) {
-      if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+      if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
       auto it = object_->find(key.primitive_);
       if (it == object_->end()) return Value();
       return it->second;
@@ -249,7 +249,7 @@ public:
   }
   void set(const Value& key, const Value& value) {
     if (!object_) throw std::runtime_error("Value is not an object: " + dump());
-    if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     (*object_)[key.primitive_] = value;
   }
   Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {

common/sampling.cpp

@@ -208,6 +208,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                                                         trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                         trigger_tokens.data(), trigger_tokens.size())
              :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
     }
 
     auto * result = new common_sampler {

convert_hf_to_gguf.py

@@ -708,6 +708,12 @@ class Model:
         if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f":
             # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k
             res = "superbpe"
+        if chkhsh == "1994ffd01900cfb37395608534236ecd63f2bd5995d6cb1004dda1af50240f15":
+            # ref: https://huggingface.co/trillionlabs/Trillion-7B-preview
+            res = "trillion"
+        if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224":
+            # ref: https://huggingface.co/inclusionAI/Ling-lite
+            res = "bailingmoe"
 
         if res is None:
             logger.warning("\n")
@@ -2269,7 +2275,7 @@ class Qwen2Model(Model):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
-@Model.register("Qwen2VLForConditionalGeneration")
+@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2VL
 
@@ -3551,8 +3557,8 @@ class RWKV6Qwen2Model(Rwkv6Model):
         head_size = hidden_size // num_attention_heads
         rms_norm_eps = self.hparams["rms_norm_eps"]
         intermediate_size = self.hparams["intermediate_size"]
-        time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
-        time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
+        time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32)
+        time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64)
 
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
@@ -4419,6 +4425,29 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("PLMForCausalLM")
+class PLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLM
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+
 @Model.register("T5WithLMHeadModel")
 @Model.register("T5ForConditionalGeneration")
 @Model.register("MT5ForConditionalGeneration")
@@ -5107,6 +5136,105 @@ class GraniteMoeModel(GraniteModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@Model.register("BailingMoeForCausalLM")
+class BailingMoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
+
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_weights_scale(1.0)
+        self.gguf_writer.add_expert_count(hparams["num_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
+        if n_head_kv is not None and n_head != n_head_kv:
+            n_head = n_head_kv
+        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_embd = self.hparams["hidden_size"]
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
+
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+
+        if name.endswith("attention.dense.weight"):
+            return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)]
+        elif name.endswith("query_key_value.weight"):
+            q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2)
+
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+            ]
+        elif name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            tensors: list[tuple[str, Tensor]] = []
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+
+            return tensors
+
+        new_name = self.map_tensor_name(name)
+
+        if new_name == output_name and self.hparams.get("norm_head"):
+            data_torch = data_torch.float()
+            data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("ChameleonForConditionalGeneration")
 @Model.register("ChameleonForCausalLM")  # obsolete
 class ChameleonModel(Model):

convert_hf_to_gguf_update.py

@@ -111,6 +111,8 @@ models = [
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
     {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
+    {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
+    {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
 ]
 
 

docs/backend/SYCL.md

@@ -20,7 +20,7 @@
 **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
-- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
@@ -227,16 +227,6 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 **oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
 
-
-**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
-
-```sh
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
-cmake --build buildWithCublas --config Release
-```
-
 **oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
 
 ```sh
@@ -250,16 +240,6 @@ cmake --build build-nvidia --config Release
 
 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
 
-**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
-
-```sh
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-# Find your HIPTARGET with rocminfo, under the key 'Name:'
-cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
-cmake --build buildWithrocBLAS --config Release
-```
-
 3. **Verify installation and environment**
 
 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -324,13 +304,10 @@ cmake --build build --config Release -j -v
 
 #### Nvidia GPU
 
-```sh
-# Export relevant ENV variables
-export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
-export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
+By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
 
+```sh
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 # Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
@@ -347,12 +324,10 @@ cmake --build build --config Release -j -v
 
 #### AMD GPU
 
-```sh
-# Export relevant ENV variables
-export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
-export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
+The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
+By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
 
+```sh
 # Build LLAMA with rocBLAS acceleration through SYCL
 
 ## AMD

examples/batched-bench/batched-bench.cpp

@@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);

examples/batched/batched.cpp

@@ -41,7 +41,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n" , __func__);

examples/export-lora/export-lora.cpp

@@ -421,7 +421,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity > 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
+        lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/gritlm/gritlm.cpp

@@ -168,7 +168,7 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
 
     // create generation context
     llama_context * ctx = llama_init_from_model(model, cparams);

examples/llava/README-gemma3.md

@@ -4,6 +4,26 @@
 >
 > This is very experimental, only used for demo purpose.
 
+## Quick started
+
+You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)'s Hugging Face account
+
+```bash
+# build
+cmake -B build
+cmake --build build --target llama-gemma3-cli
+
+# alternatively, install from brew (MacOS)
+brew install llama.cpp
+
+# run it
+llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# note: 1B model does not support vision
+```
+
 ## How to get mmproj.gguf?
 
 ```bash

examples/llava/clip.cpp

@@ -1396,14 +1396,16 @@ struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_p
         const int n_kv = gguf_get_n_kv(ctx);
         const int ftype = get_u32(ctx, KEY_FTYPE);
         const std::string ftype_str = get_ftype(ftype);
-        const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION);
-        const std::string description = gguf_get_val_str(ctx, idx_desc);
         const int idx_name = gguf_find_key(ctx, KEY_NAME);
         if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
             const std::string name = gguf_get_val_str(ctx, idx_name);
             LOG_INF("%s: model name:   %s\n", __func__, name.c_str());
         }
-        LOG_INF("%s: description:  %s\n", __func__, description.c_str());
+        const int idx_desc = gguf_find_key(ctx, KEY_DESCRIPTION);
+        if (idx_desc != -1) { // ditto
+            const std::string description = gguf_get_val_str(ctx, idx_desc);
+            LOG_INF("%s: description:  %s\n", __func__, description.c_str());
+        }
         LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
         LOG_INF("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
         LOG_INF("%s: n_tensors:    %d\n", __func__, n_tensors);

examples/llava/gemma3-cli.cpp

@@ -78,7 +78,7 @@ struct gemma3_context {
     }
 
     void init_clip_model(common_params & params) {
-        const char * clip_path = params.mmproj.c_str();
+        const char * clip_path = params.mmproj.path.c_str();
         ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
     }
 
@@ -232,13 +232,13 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty()) {
+    if (params.mmproj.path.empty()) {
         show_additional_info(argc, argv);
         return 1;
     }
 
     gemma3_context ctx(params);
-    printf("%s: %s\n", __func__, params.model.c_str());
+    printf("%s: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
 

examples/llava/llava-cli.cpp

@@ -225,7 +225,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -234,7 +234,7 @@ static struct llama_model * llava_init(common_params * params) {
 }
 
 static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.c_str();
+    const char * clip_path = params->mmproj.path.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
@@ -283,7 +283,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);
         return 1;
     }

examples/llava/minicpmv-cli.cpp

@@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -80,7 +80,7 @@ static void llava_free(struct llava_context * ctx_llava) {
 }
 
 static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.c_str();
+    const char * clip_path = params->mmproj.path.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty())) {
+    if (params.mmproj.path.empty() || (params.image.empty())) {
         show_additional_info(argc, argv);
         return 1;
     }

examples/llava/qwen2vl-cli.cpp

@@ -314,7 +314,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -323,7 +323,7 @@ static struct llama_model * llava_init(common_params * params) {
 }
 
 static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
-    const char * clip_path = params->mmproj.c_str();
+    const char * clip_path = params->mmproj.path.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
@@ -524,7 +524,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+    if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);
         return 1;
     }

examples/parallel/parallel.cpp

@@ -106,6 +106,8 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    params.n_predict = 128;
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
@@ -405,7 +407,7 @@ int main(int argc, char ** argv) {
         params.prompt_file = "used built-in defaults";
     }
     LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
-    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
+    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.path.c_str());
 
     LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
     LOG_INF("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);

examples/passkey/passkey.cpp

@@ -64,7 +64,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);

examples/rpc/CMakeLists.txt

@@ -1,2 +1,4 @@
-add_executable(rpc-server rpc-server.cpp)
-target_link_libraries(rpc-server PRIVATE ggml llama)
+set(TARGET rpc-server)
+add_executable(${TARGET} rpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE ggml)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/rpc/README.md

@@ -72,3 +72,14 @@ $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name
 
 This way you can offload model layers to both local and remote devices.
 
+### Local cache
+
+The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
+This can speed up model loading significantly, especially when using large models.
+To enable the cache, use the `-c` option:
+
+```bash
+$ bin/rpc-server -c
+```
+
+By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.

examples/rpc/rpc-server.cpp

@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "ggml-cpu.h"
 
 #ifdef GGML_USE_CUDA
@@ -18,26 +22,142 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  define DIRECTORY_SEPARATOR '\\'
+#  include <locale>
 #  include <windows.h>
+#  include <fcntl.h>
+#  include <io.h>
 #else
+#  define DIRECTORY_SEPARATOR '/'
 #  include <unistd.h>
+#  include <sys/stat.h>
 #endif
+#include <codecvt>
 #include <string>
 #include <stdio.h>
+#include <vector>
+#include <filesystem>
+
+namespace fs = std::filesystem;
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+// returns true if successful, false otherwise
+static bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::string fs_get_cache_directory() {
+    std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
+#ifdef __linux__
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        }
+#elif defined(__APPLE__)
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+#elif defined(_WIN32)
+        cache_directory = std::getenv("LOCALAPPDATA");
+#endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
+}
 
 struct rpc_server_params {
     std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
+    bool        use_cache   = false;
 };
 
 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
     fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
-    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
-    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
+    fprintf(stderr, "  -h, --help                show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
+    fprintf(stderr, "  -c,      --cache          enable local file cache\n");
     fprintf(stderr, "\n");
 }
 
@@ -58,6 +178,8 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
             if (params.port <= 0 || params.port > 65535) {
                 return false;
             }
+        } else if (arg == "-c" || arg == "--cache") {
+            params.use_cache = true;
         } else if (arg == "-m" || arg == "--mem") {
             if (++i >= argc) {
                 return false;
@@ -164,8 +286,20 @@ int main(int argc, char * argv[]) {
     } else {
         get_backend_memory(&free_mem, &total_mem);
     }
-    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
-    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
+    const char * cache_dir = nullptr;
+    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
+    if (params.use_cache) {
+        if (!fs_create_directory_with_parents(cache_dir_str)) {
+            fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
+            return 1;
+        }
+        cache_dir = cache_dir_str.c_str();
+    }
+    printf("Starting RPC server\n");
+    printf("  endpoint       : %s\n", endpoint.c_str());
+    printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
+    printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
     ggml_backend_free(backend);
     return 0;
 }

examples/server/server.cpp

@@ -133,7 +133,8 @@ struct slot_params {
 
         auto grammar_triggers = json::array();
         for (const auto & trigger : sampling.grammar_triggers) {
-            grammar_triggers.push_back(trigger.to_json<json>());
+            server_grammar_trigger ct(std::move(trigger));
+            grammar_triggers.push_back(ct.to_json());
         }
 
         return json {
@@ -372,9 +373,9 @@ struct server_task {
             const auto grammar_triggers = data.find("grammar_triggers");
             if (grammar_triggers != data.end()) {
                 for (const auto & t : *grammar_triggers) {
-                    auto ct = common_grammar_trigger::from_json(t);
-                    if (ct.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
-                        const auto & word = ct.value;
+                    server_grammar_trigger ct(t);
+                    if (ct.value.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
+                        const auto & word = ct.value.value;
                         auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true);
                         if (ids.size() == 1) {
                             auto token = ids[0];
@@ -392,7 +393,7 @@ struct server_task {
                             params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word});
                         }
                     } else {
-                        params.sampling.grammar_triggers.push_back(ct);
+                        params.sampling.grammar_triggers.push_back(std::move(ct.value));
                     }
                 }
             }
@@ -489,8 +490,12 @@ struct result_timings {
     double predicted_per_token_ms;
     double predicted_per_second;
 
+    // Optional speculative metrics - only included when > 0
+    int32_t draft_n = 0;
+    int32_t draft_n_accepted = 0;
+
     json to_json() const {
-        return {
+        json base = {
             {"prompt_n",               prompt_n},
             {"prompt_ms",              prompt_ms},
             {"prompt_per_token_ms",    prompt_per_token_ms},
@@ -501,6 +506,13 @@ struct result_timings {
             {"predicted_per_token_ms", predicted_per_token_ms},
             {"predicted_per_second",   predicted_per_second},
         };
+
+        if (draft_n > 0) {
+            base["draft_n"] = draft_n;
+            base["draft_n_accepted"] = draft_n_accepted;
+        }
+
+        return base;
     }
 };
 
@@ -1299,6 +1311,10 @@ struct server_slot {
 
     std::function<void(int)> callback_on_release;
 
+    // Speculative decoding stats
+    int32_t n_draft_total = 0;      // Total draft tokens generated
+    int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+
     void reset() {
         SLT_DBG(*this, "%s", "\n");
 
@@ -1315,6 +1331,10 @@ struct server_slot {
 
         generated_tokens.clear();
         generated_token_probs.clear();
+
+        // clear speculative decoding stats
+        n_draft_total = 0;
+        n_draft_accepted = 0;
     }
 
     bool is_non_causal() const {
@@ -1381,6 +1401,12 @@ struct server_slot {
         timings.predicted_per_token_ms = t_token_generation / n_decoded;
         timings.predicted_per_second = 1e3 / t_token_generation * n_decoded;
 
+        // Add speculative metrics
+        if (n_draft_total > 0) {
+            timings.draft_n = n_draft_total;
+            timings.draft_n_accepted = n_draft_accepted;
+        }
+
         return timings;
     }
 
@@ -1428,6 +1454,15 @@ struct server_slot {
                 t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                 t_token_generation, n_decoded, t_gen, n_gen_second,
                 t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
+
+        if (n_draft_total > 0) {
+            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            SLT_INF(*this,
+                    "\n"
+                    "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total
+            );
+        }
     }
 
     json to_json() const {
@@ -1842,7 +1877,7 @@ struct server_context {
     }
 
     bool load_model(const common_params & params) {
-        SRV_INF("loading model '%s'\n", params.model.c_str());
+        SRV_INF("loading model '%s'\n", params.model.path.c_str());
 
         params_base = params;
 
@@ -1852,7 +1887,7 @@ struct server_context {
         ctx   = llama_init.context.get();
 
         if (model == nullptr) {
-            SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
+            SRV_ERR("failed to load model, '%s'\n", params_base.model.path.c_str());
             return false;
         }
 
@@ -1863,16 +1898,13 @@ struct server_context {
         add_bos_token = llama_vocab_get_add_bos(vocab);
         has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
 
-        if (!params_base.speculative.model.empty() || !params_base.speculative.hf_repo.empty()) {
-            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
+        if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
 
             auto params_dft = params_base;
 
             params_dft.devices      = params_base.speculative.devices;
-            params_dft.hf_file      = params_base.speculative.hf_file;
-            params_dft.hf_repo      = params_base.speculative.hf_repo;
             params_dft.model        = params_base.speculative.model;
-            params_dft.model_url    = params_base.speculative.model_url;
             params_dft.n_ctx        = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx;
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
@@ -1886,12 +1918,12 @@ struct server_context {
             model_dft = llama_init_dft.model.get();
 
             if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
+                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
                 return false;
             }
 
             if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
-                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
+                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
 
                 return false;
             }
@@ -3290,6 +3322,9 @@ struct server_context {
 
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
+                // keep track of total number of tokens generated in the draft
+                slot.n_draft_total += draft.size();
+
                 // ignore small drafts
                 if (slot.params.speculative.n_min > (int) draft.size()) {
                     SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
@@ -3315,6 +3350,9 @@ struct server_context {
                 slot.n_past    += ids.size();
                 slot.n_decoded += ids.size();
 
+                // update how many tokens out of draft was accepted
+                slot.n_draft_accepted += ids.size() - 1;
+
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
 
@@ -3825,7 +3863,7 @@ int main(int argc, char ** argv) {
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
-            { "model_path",                  ctx_server.params_base.model },
+            { "model_path",                  ctx_server.params_base.model.path },
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -4091,7 +4129,7 @@ int main(int argc, char ** argv) {
             {"object", "list"},
             {"data", {
                 {
-                    {"id",       params.model_alias.empty() ? params.model : params.model_alias},
+                    {"id",       params.model_alias.empty() ? params.model.path : params.model_alias},
                     {"object",   "model"},
                     {"created",  std::time(0)},
                     {"owned_by", "llamacpp"},
@@ -4459,15 +4497,24 @@ int main(int argc, char ** argv) {
         llama_backend_free();
     };
 
-    // bind HTTP listen port
     bool was_bound = false;
-    if (params.port == 0) {
-        int bound_port = svr->bind_to_any_port(params.hostname);
-        if ((was_bound = (bound_port >= 0))) {
-            params.port = bound_port;
-        }
+    if (string_ends_with(std::string(params.hostname), ".sock")) {
+        LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
+        svr->set_address_family(AF_UNIX);
+        // bind_to_port requires a second arg, any value other than 0 should
+        // simply get ignored
+        was_bound = svr->bind_to_port(params.hostname, 8080);
     } else {
-        was_bound = svr->bind_to_port(params.hostname, params.port);
+        LOG_INF("%s: binding port with default address family\n", __func__);
+        // bind HTTP listen port
+        if (params.port == 0) {
+            int bound_port = svr->bind_to_any_port(params.hostname);
+            if ((was_bound = (bound_port >= 0))) {
+                params.port = bound_port;
+            }
+        } else {
+            was_bound = svr->bind_to_port(params.hostname, params.port);
+        }
     }
 
     if (!was_bound) {

examples/server/utils.hpp

@@ -58,6 +58,32 @@ static T json_value(const json & body, const std::string & key, const T & defaul
 
 const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
 
+// thin wrapper around common_grammar_trigger with (de)serialization functions
+struct server_grammar_trigger {
+    common_grammar_trigger value;
+
+    server_grammar_trigger() = default;
+    server_grammar_trigger(const common_grammar_trigger & value) : value(value) {}
+    server_grammar_trigger(const json & in) {
+        value.type = (common_grammar_trigger_type) in.at("type").get<int>();
+        value.value = in.at("value").get<std::string>();
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            value.token = (llama_token) in.at("token").get<int>();
+        }
+    }
+
+    json to_json() const {
+        json out {
+            {"type", (int) value.type},
+            {"value", value.value},
+        };
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            out["token"] = (int) value.token;
+        }
+        return out;
+    }
+};
+
 //
 // tokenizer and input processing utils
 //
@@ -627,7 +653,8 @@ static json oaicompat_completion_params_parse(
     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(trigger.to_json<json>());
+        server_grammar_trigger ct(trigger);
+        grammar_triggers.push_back(ct.to_json());
     }
     llama_params["grammar_triggers"] = grammar_triggers;
     llama_params["preserved_tokens"] = chat_params.preserved_tokens;

examples/speculative-simple/speculative-simple.cpp

@@ -24,7 +24,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.speculative.model.empty()) {
+    if (params.speculative.model.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }

examples/speculative/speculative.cpp

@@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.speculative.model.empty()) {
+    if (params.speculative.model.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }

examples/tts/tts.cpp

@@ -577,12 +577,7 @@ int main(int argc, char ** argv) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model_ttc);
 
-    // TODO: refactor in a common struct
-    params.model     = params.vocoder.model;
-    params.model_url = params.vocoder.model_url;
-    params.hf_repo   = params.vocoder.hf_repo;
-    params.hf_file   = params.vocoder.hf_file;
-
+    params.model = params.vocoder.model;
     params.embedding = true;
 
     common_init_result llama_init_cts = common_init_from_params(params);
@@ -699,11 +694,13 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
             const std::string voice_data = audio_data;
 
             auto tmp = common_tokenize(vocab, voice_data, false, true);
-            printf("\n\n");
+
+            std::ostringstream tokens_oss;
             for (size_t i = 0; i < tmp.size(); ++i) {
-                printf("%d, ", tmp[i]);
+                tokens_oss << tmp[i] << ", ";
             }
-            printf("\n\n");
+            LOG_INF("\n\n%s: llama tokens: %s\n\n", __func__, tokens_oss.str().c_str());
+
             prompt_add(prompt_inp, tmp);
 #else
             prompt_add(prompt_inp, llama_tokens {

ggml/CMakeLists.txt

@@ -100,6 +100,10 @@ else()
     set(INS_ENB ON)
 endif()
 
+message(DEBUG "GGML_NATIVE         : ${GGML_NATIVE}")
+message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
+message(DEBUG "INS_ENB             : ${INS_ENB}")
+
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
@@ -127,7 +131,8 @@ option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
-set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
 if (WIN32)

ggml/cmake/GitVars.cmake

@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

ggml/cmake/ggml-config.cmake.in

@@ -5,7 +5,7 @@
 
 set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
 set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
-set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
+#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
 
 find_package(Threads REQUIRED)
 

ggml/include/ggml-rpc.h

@@ -17,7 +17,9 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c
 
 GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                                    const char * cache_dir,
+                                                    size_t free_mem, size_t total_mem);
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 

ggml/include/ggml.h

@@ -1791,11 +1791,11 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd, n_batch,     n_head,    1]
-    // k:    [n_embd, n_kv,        n_head_kv, 1]
-    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,

ggml/src/CMakeLists.txt

@@ -65,7 +65,7 @@ if (GGML_LTO)
     endif()
 endif()
 
-if (GGML_CCACHE)
+if (GGML_CCACHE AND NOT CMAKE_C_COMPILER_LAUNCHER AND NOT CMAKE_CXX_COMPILER_LAUNCHER)
     find_program(GGML_CCACHE_FOUND ccache)
     find_program(GGML_SCCACHE_FOUND sccache)
 

ggml/src/ggml-cann/.clang-format

@@ -1,168 +0,0 @@
----
-Language:        Cpp
-# BasedOnStyle:  Google
-AccessModifierOffset: -1
-AlignAfterOpenBracket: Align
-AlignConsecutiveMacros: false
-AlignConsecutiveAssignments: false
-AlignConsecutiveDeclarations: false
-AlignEscapedNewlines: Left
-AlignOperands:   true
-AlignTrailingComments: true
-AllowAllArgumentsOnNextLine: true
-AllowAllConstructorInitializersOnNextLine: true
-AllowAllParametersOfDeclarationOnNextLine: true
-AllowShortBlocksOnASingleLine: Never
-AllowShortCaseLabelsOnASingleLine: false
-AllowShortFunctionsOnASingleLine: All
-AllowShortLambdasOnASingleLine: All
-AllowShortIfStatementsOnASingleLine: WithoutElse
-AllowShortLoopsOnASingleLine: true
-AlwaysBreakAfterDefinitionReturnType: None
-AlwaysBreakAfterReturnType: None
-AlwaysBreakBeforeMultilineStrings: true
-AlwaysBreakTemplateDeclarations: Yes
-BinPackArguments: true
-BinPackParameters: true
-BraceWrapping:
-  AfterCaseLabel:  false
-  AfterClass:      false
-  AfterControlStatement: false
-  AfterEnum:       false
-  AfterFunction:   false
-  AfterNamespace:  false
-  AfterObjCDeclaration: false
-  AfterStruct:     false
-  AfterUnion:      false
-  AfterExternBlock: false
-  BeforeCatch:     false
-  BeforeElse:      false
-  IndentBraces:    false
-  SplitEmptyFunction: true
-  SplitEmptyRecord: true
-  SplitEmptyNamespace: true
-BreakBeforeBinaryOperators: None
-BreakBeforeBraces: Attach
-BreakBeforeInheritanceComma: false
-BreakInheritanceList: BeforeColon
-BreakBeforeTernaryOperators: true
-BreakConstructorInitializersBeforeComma: false
-BreakConstructorInitializers: BeforeColon
-BreakAfterJavaFieldAnnotations: false
-BreakStringLiterals: true
-ColumnLimit:     80
-CommentPragmas:  '^ IWYU pragma:'
-CompactNamespaces: false
-ConstructorInitializerAllOnOneLineOrOnePerLine: true
-ConstructorInitializerIndentWidth: 4
-ContinuationIndentWidth: 4
-Cpp11BracedListStyle: true
-DeriveLineEnding: true
-DerivePointerAlignment: true
-DisableFormat:   false
-ExperimentalAutoDetectBinPacking: false
-FixNamespaceComments: true
-ForEachMacros:
-  - foreach
-  - Q_FOREACH
-  - BOOST_FOREACH
-IncludeBlocks:   Regroup
-IncludeCategories:
-  - Regex:           '^<ext/.*\.h>'
-    Priority:        2
-    SortPriority:    0
-  - Regex:           '^<.*\.h>'
-    Priority:        1
-    SortPriority:    0
-  - Regex:           '^<.*'
-    Priority:        2
-    SortPriority:    0
-  - Regex:           '.*'
-    Priority:        3
-    SortPriority:    0
-IncludeIsMainRegex: '([-_](test|unittest))?$'
-IncludeIsMainSourceRegex: ''
-IndentCaseLabels: true
-IndentGotoLabels: true
-IndentPPDirectives: None
-IndentWidth:     4
-IndentWrappedFunctionNames: false
-JavaScriptQuotes: Leave
-JavaScriptWrapImports: true
-KeepEmptyLinesAtTheStartOfBlocks: false
-MacroBlockBegin: ''
-MacroBlockEnd:   ''
-MaxEmptyLinesToKeep: 1
-NamespaceIndentation: None
-ObjCBinPackProtocolList: Never
-ObjCBlockIndentWidth: 2
-ObjCSpaceAfterProperty: false
-ObjCSpaceBeforeProtocolList: true
-PenaltyBreakAssignment: 2
-PenaltyBreakBeforeFirstCallParameter: 1
-PenaltyBreakComment: 300
-PenaltyBreakFirstLessLess: 120
-PenaltyBreakString: 1000
-PenaltyBreakTemplateDeclaration: 10
-PenaltyExcessCharacter: 1000000
-PenaltyReturnTypeOnItsOwnLine: 200
-PointerAlignment: Left
-RawStringFormats:
-  - Language:        Cpp
-    Delimiters:
-      - cc
-      - CC
-      - cpp
-      - Cpp
-      - CPP
-      - 'c++'
-      - 'C++'
-    CanonicalDelimiter: ''
-    BasedOnStyle:    google
-  - Language:        TextProto
-    Delimiters:
-      - pb
-      - PB
-      - proto
-      - PROTO
-    EnclosingFunctions:
-      - EqualsProto
-      - EquivToProto
-      - PARSE_PARTIAL_TEXT_PROTO
-      - PARSE_TEST_PROTO
-      - PARSE_TEXT_PROTO
-      - ParseTextOrDie
-      - ParseTextProtoOrDie
-    CanonicalDelimiter: ''
-    BasedOnStyle:    google
-ReflowComments:  true
-SortIncludes:    true
-SortUsingDeclarations: true
-SpaceAfterCStyleCast: false
-SpaceAfterLogicalNot: false
-SpaceAfterTemplateKeyword: true
-SpaceBeforeAssignmentOperators: true
-SpaceBeforeCpp11BracedList: false
-SpaceBeforeCtorInitializerColon: true
-SpaceBeforeInheritanceColon: true
-SpaceBeforeParens: ControlStatements
-SpaceBeforeRangeBasedForLoopColon: true
-SpaceInEmptyBlock: false
-SpaceInEmptyParentheses: false
-SpacesBeforeTrailingComments: 2
-SpacesInAngles:  false
-SpacesInConditionalStatement: false
-SpacesInContainerLiterals: true
-SpacesInCStyleCastParentheses: false
-SpacesInParentheses: false
-SpacesInSquareBrackets: false
-SpaceBeforeSquareBrackets: false
-Standard:        Auto
-StatementMacros:
-  - Q_UNUSED
-  - QT_REQUIRE_VERSION
-TabWidth:        8
-UseCRLF:         false
-UseTab:          Never
-...
-

ggml/src/ggml-cann/CMakeLists.txt

@@ -51,13 +51,11 @@ if (CANN_INSTALL_DIR)
         ${CANN_INSTALL_DIR}/acllib/include
     )
 
-    add_subdirectory(kernels)
     list(APPEND CANN_LIBRARIES
         ascendcl
         nnopbase
         opapi
         acl_op_compiler
-        ascendc_kernels
     )
 
     file(GLOB GGML_SOURCES_CANN "*.cpp")

ggml/src/ggml-cann/acl_tensor.cpp

@@ -54,9 +54,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     // added.
     int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
 
-    int64_t acl_storage_len = 0;
     if (ne == nullptr) {
-        acl_storage_len = ggml_nbytes(tensor);
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
             acl_ne[i] = tensor->ne[i];
             // The step size of acl is in elements.
@@ -65,14 +63,18 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     } else {
         // With bcast
         for (int i = 0; i < dims; i++) {
-            acl_storage_len += (ne[i] - 1) * nb[i];
             acl_ne[i] = ne[i];
             acl_stride[i] = nb[i] / ggml_element_size(tensor);
         }
     }
 
-    // Reverse ne and stride.
     int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
+    int64_t acl_storage_len = 1;
+    for (int i = 0; i < final_dims; i++) {
+        acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
+    }
+
+    // Reverse ne and stride.
     std::reverse(acl_ne, acl_ne + final_dims);
     std::reverse(acl_stride, acl_stride + final_dims);
 

ggml/src/ggml-cann/acl_tensor.h

@@ -101,14 +101,14 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
         tmp_stride[i] = nb[i] / type_size;
     }
 
+    int64_t acl_storage_len = 1;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i];
+    }
+
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
     aclTensor* acl_tensor =
         aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
                         format, &acl_storage_len, 1, data_ptr);

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -30,6 +30,7 @@
 #include <aclnnop/aclnn_copy.h>
 #include <aclnnop/aclnn_cos.h>
 #include <aclnnop/aclnn_div.h>
+#include <aclnnop/aclnn_embedding.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_group_norm.h>
@@ -58,7 +59,6 @@
 #include <vector>
 
 #include "ggml-impl.h"
-#include "kernels/ascendc_kernels.h"
 
 #define GGML_COMMON_DECL_C
 
@@ -99,6 +99,35 @@ static void aclnn_repeat(ggml_backend_cann_context& ctx, aclTensor* acl_src,
     ACL_CHECK(aclDestroyIntArray(repeats));
 }
 
+/**
+ * @brief Casts the elements of a tensor to a specified data type using the CANN backend.
+ *
+ * @details This function performs a type conversion on the elements of the input tensor `acl_src`
+ *          and stores the results in the destination tensor `acl_dst`. The conversion type is
+ *          determined based on the `dst` tensor's data type.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose elements will be cast.
+ * @param acl_dst The destination tensor that will store the casted elements.
+ * @param dst The ggml tensor specifying the target data type.
+ */
+static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                    aclTensor* acl_dst, ggml_tensor* dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+    ACL_CHECK(aclnnCastGetWorkspaceSize(acl_src,
+                                        ggml_cann_type_mapping(dst->type),
+                                        acl_dst, &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
 void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
     GGML_ASSERT(ggml_can_repeat(src, dst));
@@ -329,8 +358,6 @@ void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
 void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     float min;
     float max;
@@ -889,173 +916,76 @@ static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src,
 }
 
 void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src = dst->src[0];
+    ggml_tensor* src0 = dst->src[0];
 
-    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_src = ggml_cann_create_tensor(src0);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
-
-    ggml_cann_pool_alloc src_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
-    ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
-    src->extra = src_extra_allocator.get();
-    dst->extra = dst_extra_allocator.get();
-    ACL_CHECK(aclrtMemcpyAsync(src->extra, sizeof(ggml_tensor), src,
-                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
-                               ctx.stream()));
-    ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
-                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
-                               ctx.stream()));
-
-    if ((dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32) &&
-        ggml_are_same_shape(src, dst)) {
-        cann_copy(ctx, acl_src, acl_dst);
-        ACL_CHECK(aclDestroyTensor(acl_src));
-        ACL_CHECK(aclDestroyTensor(acl_dst));
-        return;
-    }
-    // TODO: simplify
-    if (src->type == GGML_TYPE_F16) {
-        if (dst->type == GGML_TYPE_Q8_0) {
-            aclrtlaunch_ascendc_quantize_f16_q8_0(
-                24, ctx.stream(), src->data, dst->data,
-                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
-                ((ggml_tensor*)dst->extra)->ne);
-            return;
-        }
-        if (dst->type == GGML_TYPE_Q4_0) {
-            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
-                24, ctx.stream(), src->data, dst->data,
-                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
-                ((ggml_tensor*)dst->extra)->ne);
-            return;
-        }
-        if (dst->type == GGML_TYPE_F16) {
-            if (ggml_are_same_shape(src, dst)) {
-                cann_copy(ctx, acl_src, acl_dst);
-                ACL_CHECK(aclDestroyTensor(acl_src));
-                ACL_CHECK(aclDestroyTensor(acl_dst));
-                return;
-            }
-            if (ggml_is_contiguous(dst)) {
-                const size_t src_type_size = ggml_type_size(src->type);
-                if (src->nb[0] == src_type_size) {
-                    // src0 is contigous on first dimension, copy by rows
-                    int64_t rows_num = ggml_nrows(src);
-
-                    aclrtlaunch_ascendc_dup_by_rows_fp16(
-                        rows_num, ctx.stream(), src->data, dst->data,
-                        ((ggml_tensor*)src->extra)->ne,
-                        ((ggml_tensor*)src->extra)->nb,
-                        ((ggml_tensor*)dst->extra)->ne,
-                        ((ggml_tensor*)dst->extra)->nb);
-                    return;
-                }
-                GGML_ABORT("fatal error");
-            }
-            GGML_ABORT("fatal error");
-        }
-        if (dst->type == GGML_TYPE_F32) {
-            if (ggml_are_same_shape(src, dst)) {
-                cann_copy(ctx, acl_src, acl_dst);
-                ACL_CHECK(aclDestroyTensor(acl_src));
-                ACL_CHECK(aclDestroyTensor(acl_dst));
-                return;
-            }
-            if (ggml_is_contiguous(dst)) {
-                const size_t src_type_size = ggml_type_size(src->type);
-                if (src->nb[0] == src_type_size) {
-                    // src0 is contigous on first dimension, copy by rows
-                    int64_t rows_num = ggml_nrows(src);
-                    aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32(
-                        rows_num, ctx.stream(), src->data, dst->data,
-                        ((ggml_tensor*)src->extra)->ne,
-                        ((ggml_tensor*)src->extra)->nb,
-                        ((ggml_tensor*)dst->extra)->ne,
-                        ((ggml_tensor*)dst->extra)->nb);
-                    return;
-                }
-                GGML_ABORT("fatal error");
-            }
-            GGML_ABORT("fatal error");
-        }
-        // TODO
-        GGML_ABORT("fatal error");
-    } else if (src->type == GGML_TYPE_F32) {
-        // TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size
-        //          && nb0 == type_size)
-        if (dst->type == GGML_TYPE_Q8_0) {
-            aclrtlaunch_ascendc_quantize_f32_q8_0(
-                24, ctx.stream(), src->data, dst->data,
-                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
-                ((ggml_tensor*)dst->extra)->ne);
-            return;
-        }
-        if (dst->type == GGML_TYPE_Q4_0) {
-            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
-                24, ctx.stream(), src->data, dst->data,
-                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
-                ((ggml_tensor*)dst->extra)->ne);
-            return;
-        }
-        if (dst->type == GGML_TYPE_F32) {
-            if (ggml_are_same_shape(src, dst)) {
-                cann_copy(ctx, acl_src, acl_dst);
-                ACL_CHECK(aclDestroyTensor(acl_src));
-                ACL_CHECK(aclDestroyTensor(acl_dst));
-                return;
-            }
-            if (ggml_is_contiguous(dst)) {
-                const size_t src_type_size = ggml_type_size(src->type);
-                if (src->nb[0] == src_type_size) {
-                    // src0 is contigous on first dimension, copy by rows
-                    int64_t rows_num = ggml_nrows(src);
-                    aclrtlaunch_ascendc_dup_by_rows_fp32(
-                        rows_num, ctx.stream(), src->data, dst->data,
-                        ((ggml_tensor*)src->extra)->ne,
-                        ((ggml_tensor*)src->extra)->nb,
-                        ((ggml_tensor*)dst->extra)->ne,
-                        ((ggml_tensor*)dst->extra)->nb);
-                    return;
-                }
-                GGML_ABORT("fatal error");
-            } else {
-                // TODO: dst not contiguous
-                GGML_ABORT("fatal error");
-            }
-        }
-        if (dst->type == GGML_TYPE_F16) {
-            if (ggml_are_same_shape(src, dst)) {
-                cann_copy(ctx, acl_src, acl_dst);
-                ACL_CHECK(aclDestroyTensor(acl_src));
-                ACL_CHECK(aclDestroyTensor(acl_dst));
-                return;
-            }
-            if (ggml_is_contiguous(dst)) {
-                const size_t src_type_size = ggml_type_size(src->type);
-                if (src->nb[0] == src_type_size) {
-                    // src0 is contigous on first dimension, copy by rows
-                    int64_t rows_num = ggml_nrows(src);
-                    aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16(
-                        rows_num, ctx.stream(), src->data, dst->data,
-                        ((ggml_tensor*)src->extra)->ne,
-                        ((ggml_tensor*)src->extra)->nb,
-                        ((ggml_tensor*)dst->extra)->ne,
-                        ((ggml_tensor*)dst->extra)->nb);
-                    return;
-                }
-                GGML_ABORT("fatal error");
-            }
-        }
-        // TODO
-        GGML_ABORT("fatal error");
-    } else {
-        if (ggml_are_same_shape(src, dst)) {
+    if (ggml_are_same_shape(src0, dst)) {
+        if (dst->type == src0->type) {
             cann_copy(ctx, acl_src, acl_dst);
-            ACL_CHECK(aclDestroyTensor(acl_src));
-            ACL_CHECK(aclDestroyTensor(acl_dst));
-            return;
+        } else {
+            aclnn_cast(ctx, acl_src, acl_dst, dst);
+        }
+    } else {
+        if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
+            if (dst->type == src0->type) {
+                size_t cpy_size = ggml_nbytes(dst);
+                ACL_CHECK(aclrtMemcpyAsync(
+                    dst->data, cpy_size, src0->data, cpy_size,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                return;
+            } else {
+                ggml_cann_pool_alloc src_buffer_allocator(
+                    ctx.pool(),
+                    ggml_nelements(dst) * ggml_type_size(dst->type));
+                void* src_trans_buffer = src_buffer_allocator.get();
+                size_t src_trans_nb[GGML_MAX_DIMS];
+                src_trans_nb[0] = ggml_type_size(dst->type);
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                }
+                aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                    src_trans_buffer, ggml_cann_type_mapping(dst->type),
+                    ggml_type_size(dst->type), src0->ne, src_trans_nb,
+                    GGML_MAX_DIMS);
+
+                aclnn_cast(ctx, acl_src, src_trans_tensor, dst);
+                size_t cpy_size = ggml_nbytes(dst);
+                ACL_CHECK(aclrtMemcpyAsync(
+                    dst->data, cpy_size, src_trans_buffer, cpy_size,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                ACL_CHECK(aclDestroyTensor(src_trans_tensor));
+                return;
+            }
+        } else if (ggml_is_contiguous(dst)) {
+            ggml_cann_pool_alloc src_buffer_allocator(
+                ctx.pool(), ggml_nelements(dst) * ggml_type_size(dst->type));
+            void* src_trans_buffer = src_buffer_allocator.get();
+            size_t src_trans_nb[GGML_MAX_DIMS];
+            src_trans_nb[0] = ggml_type_size(dst->type);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+            }
+            aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                src_trans_buffer, ggml_cann_type_mapping(dst->type),
+                ggml_type_size(dst->type), src0->ne, src_trans_nb,
+                GGML_MAX_DIMS);
+
+            aclnn_cast(ctx, acl_src, src_trans_tensor, dst);
+
+            size_t cpy_size = ggml_nbytes(dst);
+            ACL_CHECK(aclrtMemcpyAsync(dst->data, cpy_size, src_trans_buffer,
+                                       cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                       ctx.stream()));
+            ACL_CHECK(aclDestroyTensor(src_trans_tensor));
+            return;
+        } else {
+            GGML_ABORT("Unsupport dst is not tontiguous.");
         }
-        GGML_ABORT("fatal error");
     }
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
 }
 
 #ifdef __cplusplus
@@ -1158,8 +1088,6 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    GGML_ASSERT(eps > 0.0f);
-
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -2378,85 +2306,168 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ACL_CHECK(aclDestroyTensor(tmp_mask_tensor));
 }
 
-void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src0 = dst->src[0];
-    ggml_tensor* src1 = dst->src[1];
+/**
+ * @brief Performs embedding operation on a 4D tensor using the CANN backend.
+ *
+ * This function extracts slices from the source tensor (`src_buffer`),
+ * index tensor (`index`), and destination tensor (`dst`), and performs an
+ * embedding operation on them. The embedding operation is applied by iterating
+ * over the last two dimensions of the source tensor, creating the necessary
+ * tensors for the source, index, and output, and executing the embedding operation.
+ *
+ * @param ctx The context for CANN backend operations.
+ * @param src_buffer The source buffer holding the data for the source tensor.
+ * @param src_ne The dimensions of the source tensor.
+ * @param src_nb The strides (byte offsets) of the source tensor.
+ * @param index The index tensor used in the embedding operation.
+ * @param dst The destination tensor where the result will be stored.
+ */
+static void aclnn_embedding_4d(ggml_backend_cann_context& ctx, void* src_buffer,
+                            int64_t* src_ne, size_t* src_nb, ggml_tensor* index,
+                            ggml_tensor* dst) {
+    for (int64_t i = 0; i < src_ne[3]; i++) {
+        for (int64_t j = 0; j < src_ne[2]; j++) {
+            // src
+            int64_t acl_src_ne[2] = {src_ne[0], src_ne[1]};
+            size_t acl_src_nb[2] = {src_nb[0], src_nb[1]};
+            aclTensor* acl_src_tensor = ggml_cann_create_tensor(
+                (char*)src_buffer + i * src_nb[3] + j * src_nb[2],
+                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
+                acl_src_ne, acl_src_nb, 2);
 
-    ggml_cann_pool_alloc src0_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
-    ggml_cann_pool_alloc src1_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
-    ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
-    src0->extra = src0_extra_allocator.get();
-    src1->extra = src1_extra_allocator.get();
-    dst->extra = dst_extra_allocator.get();
-    ACL_CHECK(aclrtMemcpyAsync(src0->extra, sizeof(ggml_tensor), src0,
-                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
-                               ctx.stream()));
-    ACL_CHECK(aclrtMemcpyAsync(src1->extra, sizeof(ggml_tensor), src1,
-                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
-                               ctx.stream()));
-    ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
-                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
-                               ctx.stream()));
+            // index
+            int64_t acl_index_ne[1] = {index->ne[0]};
+            size_t acl_index_nb[1] = {index->nb[0]};
+            aclTensor* acl_index = ggml_cann_create_tensor(
+                (char*)index->data + i * index->nb[2] + j * index->nb[1],
+                ggml_cann_type_mapping(index->type), ggml_element_size(index),
+                acl_index_ne, acl_index_nb, 1);
+
+            // out
+            int64_t acl_out_ne[2] = {dst->ne[0], dst->ne[1]};
+            size_t acl_out_nb[2] = {dst->nb[0], dst->nb[1]};
+            aclTensor* acl_out = ggml_cann_create_tensor(
+                (char*)dst->data + i * dst->nb[3] + j * dst->nb[2],
+                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
+                acl_out_ne, acl_out_nb, 2);
+
+            uint64_t workspaceSize = 0;
+            aclOpExecutor* executor;
+            void* workspaceAddr = nullptr;
+
+            ACL_CHECK(aclnnEmbeddingGetWorkspaceSize(
+                acl_src_tensor, acl_index, acl_out, &workspaceSize, &executor));
+
+            if (workspaceSize > 0) {
+                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
+                                                         workspaceSize);
+                workspaceAddr = workspace_allocator.get();
+            }
+
+            ACL_CHECK(aclnnEmbedding(workspaceAddr, workspaceSize, executor,
+                                     ctx.stream()));
+
+            ACL_CHECK(aclDestroyTensor(acl_src_tensor));
+            ACL_CHECK(aclDestroyTensor(acl_index));
+            ACL_CHECK(aclDestroyTensor(acl_out));
+        }
+    }
+}
+
+void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];  // src
+    ggml_tensor* src1 = dst->src[1];  // index
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
-#ifdef ASCEND_310P
-            // Special operation for get_row_f32 kernel of 310P: clear the
-            // content of dest data buffer when row is not aligned to 32 bytes
-            if ((src0->ne[0] % 8) != 0) {
-                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
-                                 src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
-                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
-            }
-#endif
-            aclrtlaunch_ascendc_get_row_f32(
-                24, ctx.stream(), src0->data, src1->data, dst->data,
-                ((ggml_tensor*)src0->extra)->ne,
-                ((ggml_tensor*)src0->extra)->nb,
-                ((ggml_tensor*)src1->extra)->ne,
-                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
-                ((ggml_tensor*)dst->extra)->nb);
+            aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1,
+                                   dst);
             break;
         }
         case GGML_TYPE_F16: {
-#ifdef ASCEND_310P
-            // Special operation for get_row_f16 kernel of 310P: clear the
-            // content of dest data buffer when row is not aligned to 32 bytes
-            if ((src0->ne[0] % 16) != 0) {
-                size_t dst_len =
-                    src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
-                    ggml_type_size(
-                        GGML_TYPE_F32);  // out is also f32, even input is f16
-                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+            ggml_cann_pool_alloc src_buffer_allocator(
+                ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
+            void* src_trans_buffer = src_buffer_allocator.get();
+            size_t src_trans_nb[GGML_MAX_DIMS];
+            src_trans_nb[0] = sizeof(float_t);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
             }
-#endif
-            aclrtlaunch_ascendc_get_row_f16(
-                24, ctx.stream(), src0->data, src1->data, dst->data,
-                ((ggml_tensor*)src0->extra)->ne,
-                ((ggml_tensor*)src0->extra)->nb,
-                ((ggml_tensor*)src1->extra)->ne,
-                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
-                ((ggml_tensor*)dst->extra)->nb);
+            aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type),
+                src0->ne, src_trans_nb, GGML_MAX_DIMS);
+            aclnn_cast(ctx, acl_src0, src_trans_tensor, dst);
+            aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne,
+                                   src_trans_nb, src1, dst);
+            ACL_CHECK(aclDestroyTensor(acl_src0));
+            ACL_CHECK(aclDestroyTensor(src_trans_tensor));
             break;
         }
-        case GGML_TYPE_Q4_0:
-            aclrtlaunch_ascendc_get_row_q4_0(
-                24, ctx.stream(), src0->data, src1->data, dst->data,
-                ((ggml_tensor*)src0->extra)->ne,
-                ((ggml_tensor*)src1->extra)->ne,
-                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
-                ((ggml_tensor*)dst->extra)->nb);
-            break;
-        case GGML_TYPE_Q8_0:
-            aclrtlaunch_ascendc_get_row_q8_0(
-                24, ctx.stream(), src0->data, src1->data, dst->data,
-                ((ggml_tensor*)src0->extra)->ne,
-                ((ggml_tensor*)src1->extra)->ne,
-                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
-                ((ggml_tensor*)dst->extra)->nb);
+        case GGML_TYPE_Q8_0: {
+            // add 1 dim for bcast mul.
+            size_t weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1],
+                dequant_nb[GGML_MAX_DIMS + 1];
+            int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1],
+                *dequant_ne;
+            int64_t scale_offset = 0;
+
+            // [3,4,5,64] -> [3,4,5,2,32]
+            weight_ne[0] = QK8_0;
+            weight_ne[1] = src0->ne[0] / QK8_0;
+            weight_nb[0] = sizeof(int8_t);
+            weight_nb[1] = weight_nb[0] * weight_ne[0];
+            for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
+                weight_ne[i] = src0->ne[i - 1];
+                weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1];
+            }
+
+            // [3,4,5,64] -> [3,4,5,2,1]
+            scale_ne[0] = 1;
+            scale_ne[1] = src0->ne[0] / QK8_0;
+            scale_nb[0] = sizeof(uint16_t);
+            scale_nb[1] = scale_nb[0] * scale_ne[0];
+            for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
+                scale_ne[i] = src0->ne[i - 1];
+                scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1];
+            }
+
+            // [3,4,5,64] -> [3,4,5,2,32]
+            dequant_ne = weight_ne;
+            dequant_nb[0] = sizeof(float_t);
+            for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
+                dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
+            }
+
+            scale_offset = ggml_nelements(src0) * sizeof(int8_t);
+            ggml_cann_pool_alloc dequant_buffer_allocator(
+                ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
+
+            aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
+                src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb,
+                GGML_MAX_DIMS + 1);
+            aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
+                src0->data, ACL_FLOAT16, sizeof(float16_t), scale_ne, scale_nb,
+                GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
+            aclTensor* dequant_tensor = ggml_cann_create_tensor(
+                dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float_t),
+                dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
+
+            aclnn_mul(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor);
+            dequant_nb[0] = sizeof(float_t);
+            dequant_ne = src0->ne;
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
+            }
+
+            aclnn_embedding_4d(ctx, dequant_buffer_allocator.get(),
+                                   dequant_ne, dequant_nb, src1, dst);
+
+            ACL_CHECK(aclDestroyTensor(dequant_tensor));
             break;
+        }
         default:
-            GGML_ABORT("fatal error");
+            GGML_ABORT("Unsupported tensor type for GGML_OP_GET_ROWS");
             break;
     }
 }
@@ -2797,8 +2808,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                 acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
-                nullptr, nullptr, nullptr, antiquantGroupSize, acl_output_tensor,
-                &workspaceSize, &executor));
+                nullptr, nullptr, nullptr, antiquantGroupSize,
+                acl_output_tensor, &workspaceSize, &executor));
             if (workspaceAddr == nullptr) {
                 workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
@@ -3137,7 +3148,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
-    ggml_tensor* src2 = dst->src[2];  // freq_factors
+    // ggml_tensor* src2 = dst->src[2];  // freq_factors, not used now.
 
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;

ggml/src/ggml-cann/aclnn_ops.h

@@ -535,9 +535,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
 void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
@@ -566,9 +563,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
 void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1458,11 +1458,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     ACL_CHECK(aclrtSynchronizeDevice());
     ACL_CHECK(aclrtResetDevice(cann_ctx->device));
 
-    // finalize when last backend freed.
-    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
-        ACL_CHECK(aclFinalize());
-    }
-
     delete cann_ctx;
     delete backend;
 }
@@ -1688,11 +1683,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                case GGML_TYPE_Q8_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
-                case GGML_TYPE_Q4_0:
                     return true;
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                    // only support contiguous for quantized types.
+                    return ggml_is_contiguous(op->src[0]) &&
+                            ggml_is_contiguous(op->src[1]);
                 default:
                     return false;
             }
@@ -1704,7 +1702,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             switch (op->src[0]->type) {
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
-                case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q8_0:
                     return true;
                 default:
@@ -1712,16 +1709,21 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         } break;
         case GGML_OP_CPY: {
-            switch (op->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_0:
-                    return true;
-                default:
-                    return false;
+            ggml_tensor *src = op->src[0];
+            if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
+                  (src->type != GGML_TYPE_F32 &&
+                    src->type != GGML_TYPE_F16)) {
+                // only support F32 and F16.
+                return false;
             }
-        }
+
+            if (!ggml_are_same_shape(op, src) && !ggml_is_contiguous(op)) {
+                // unsupport dst is not contiguous.
+                return false;
+            }
+
+            return true;
+        } break;
         case GGML_OP_CONT: {
             // TODO: support GGML_TYPE_BF16
             switch (op->src[0]->type) {
@@ -1734,13 +1736,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
-            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            float ext_factor = 0.0f;
+            memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
             }
             // TODO: ext_factor != 0
-            if (*ext_factor != 0) {
+            if (ext_factor != 0) {
                 return false;
             }
 
@@ -1762,9 +1765,19 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
             return true;
         }
+        case GGML_OP_POOL_2D: {
+            const int32_t * opts = (const int32_t *) op->op_params;
+            const int       k0   = opts[1];
+            const int       k1   = opts[2];
+            const int       p0   = opts[5];
+            const int       p1   = opts[6];
+            // value of paddingH should be at most half of kernelH
+            // value of paddingW should be at most half of kernelW
+            return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
+        }
+        case GGML_OP_DUP:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:
-        case GGML_OP_DUP:
         case GGML_OP_REPEAT:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
@@ -1781,7 +1794,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_CLAMP:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -1,30 +0,0 @@
-file(GLOB SRC_FILES
-    get_row_f32.cpp
-    get_row_f16.cpp
-    get_row_q4_0.cpp
-    get_row_q8_0.cpp
-    quantize_f32_q8_0.cpp
-    quantize_f16_q8_0.cpp
-    quantize_float_to_q4_0.cpp
-    dup.cpp
-)
-
-set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
-set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
-
-if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
-    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
-elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
-    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
-else()
-    message(FATAL_ERROR "ascendc_kernel_cmake does not exist, please check whether the compiler package is installed.")
-endif()
-include(${ASCENDC_CMAKE_DIR}/ascendc.cmake)
-
-ascendc_library(ascendc_kernels STATIC
-    ${SRC_FILES}
-)
-
-message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
-ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
-# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/ascendc_kernels.h

@@ -1,19 +0,0 @@
-#ifndef ASCENDC_KERNELS_H
-#define ASCENDC_KERNELS_H
-
-#include "aclrtlaunch_ascendc_get_row_f32.h"
-#include "aclrtlaunch_ascendc_get_row_f16.h"
-#include "aclrtlaunch_ascendc_get_row_q8_0.h"
-#include "aclrtlaunch_ascendc_get_row_q4_0.h"
-
-#include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
-#include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
-#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
-#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
-
-#include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32.h"
-
-#endif  // ASCENDC_KERNELS_H

ggml/src/ggml-cann/kernels/dup.cpp

@@ -1,234 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
-
-template <typename SRC_T, typename DST_T>
-class DupByRows {
-   public:
-    __aicore__ inline DupByRows() {}
-    __aicore__ inline void init(GM_ADDR src, GM_ADDR dst, int64_t *input_ne_ub,
-                                size_t *input_nb_ub) {
-        /* Dup by rows when src is contigous on first dimension and dst is
-        contiguous, each kernel process one row.
-        */
-
-        // Input has four dims.
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        // param
-        num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
-        num_elem = input_ne_ub[0];
-
-        // index for (ne[1], ne[2], ne[3]): (idx_ne1, idx_ne2, idx_ne3)
-        idx_ne3 = op_block_idx / (input_ne_ub[1] * input_ne_ub[2]);
-        idx_ne2 = (op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2]))
-                  / (input_ne_ub[1]);
-        idx_ne1 = op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2])
-                - idx_ne2 * input_ne_ub[1];
-
-        // src may not contiguous in dim [1,2,3], so stride decited by ne&nb
-        src_stride = input_nb_ub[3] * idx_ne3 + input_nb_ub[2] * idx_ne2
-                     + input_nb_ub[1] * idx_ne1;
-
-        // dst is contiguous
-        dst_stride = op_block_idx * (input_ne_ub[0] * sizeof(DST_T));
-
-        src_gm.SetGlobalBuffer(reinterpret_cast<__gm__ SRC_T *>(src +
-                                                                src_stride));
-        dst_gm.SetGlobalBuffer(reinterpret_cast<__gm__ DST_T *>(dst +
-                                                                dst_stride));
-
-        pipe.InitBuffer(src_queue, BUFFER_NUM, (sizeof(SRC_T) * num_elem +
-                                                32 - 1) / 32 * 32);
-        pipe.InitBuffer(dst_queue, BUFFER_NUM, (sizeof(DST_T) * num_elem +
-                                                32 - 1) / 32 * 32);
-    }
-
-    __aicore__ inline void copy_in() {
-        LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-        const size_t elem_per_block = 32 / sizeof(SRC_T);
-        size_t tail = num_elem % elem_per_block;
-        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
-        DataCopy(src_local, src_gm, cpy_elements_len);
-        src_queue.EnQue(src_local);
-    }
-
-    __aicore__ inline void copy_out() {
-        LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-#ifdef ASCEND_310P
-        const size_t elem_per_block = 32 / sizeof(DST_T);
-        size_t tail = num_elem % elem_per_block;
-        size_t len = num_elem & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(dst_gm, dst_local, len);
-        }
-        if(tail != 0) {
-            for (size_t i = tail; i < elem_per_block; i++) {
-                dst_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
-            SetAtomicNone();
-        }
-#else
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(DST_T);
-        DataCopyPad(dst_gm, dst_local, dataCopyParams);
-#endif
-        dst_queue.FreeTensor(dst_local);
-    }
-
-    __aicore__ inline void dup() {
-        // main process, copy one row data from src to dst.
-        copy_in();
-
-        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
-        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
-
-        int32_t BLOCK_NUM = 32 / sizeof(DST_T);
-        DataCopy(dst_local, src_local, (num_elem + BLOCK_NUM - 1)
-                                        / BLOCK_NUM * BLOCK_NUM);
-        dst_queue.EnQue<DST_T>(dst_local);
-
-        src_queue.FreeTensor(src_local);
-        copy_out();
-    }
-
-    __aicore__ inline void dup_with_cast() {
-        // main process, copy one row data from src to dst.
-        // cast dtype from src to dst.
-        copy_in();
-
-        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
-        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
-
-        Cast(dst_local, src_local, RoundMode::CAST_NONE, num_elem);
-        dst_queue.EnQue<DST_T>(dst_local);
-
-        src_queue.FreeTensor(src_local);
-        copy_out();
-    }
-
-   private:
-
-    TPipe pipe;
-    GlobalTensor<SRC_T> src_gm;
-    GlobalTensor<DST_T> dst_gm;
-
-    int64_t num_rows;
-    int64_t num_elem;
-    int64_t idx_ne3;
-    int64_t idx_ne2;
-    int64_t idx_ne1;
-    int64_t src_stride;
-    int64_t dst_stride;
-
-    TQue<QuePosition::VECIN, BUFFER_NUM> src_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> dst_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<half, half> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<float, float> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<float, half> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup_with_cast();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    // copy params from gm to ub.
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<half, float> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup_with_cast();
-}

ggml/src/ggml-cann/kernels/get_row_f16.cpp

@@ -1,197 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-class GET_ROW_F16 {
-   public:
-    __aicore__ inline GET_ROW_F16() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
-                                int64_t *output_ne_ub, size_t *output_nb_ub) {
-        // TODO, use template for F16/f32
-        int64_t op_block_num = GetBlockNum();
-        op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ half *)input);
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        uint64_t input_local_buffer_size = ((input_ne[0] * sizeof(half) + 31)
-                                             & ~31);
-        uint64_t output_local_buffer_size = ((input_ne[0] * sizeof(float) + 31)
-                                              & ~31);
-
-        local_buffer_elems = input_local_buffer_size / sizeof(half);
-
-        // TODO, consider long row that can't put in UB.
-        // All data should asign to 32. It's ok because all data is align to 32.
-        pipe.InitBuffer(input_queue, BUFFER_NUM, input_local_buffer_size);
-        pipe.InitBuffer(output_queue, BUFFER_NUM, output_local_buffer_size);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
-        size_t origin_len = len;
-        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        const size_t elem_per_block = 32 / sizeof(half);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if(tail != 0) {
-            len += elem_per_block;
-        }
-        DataCopy(input_local, input_gm[offset], len);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(output_gm[offset], output_local, len);
-        }
-
-        if(tail != 0) {
-#ifdef ASCEND_310P
-            for (size_t i = tail; i < elem_per_block; i++) {
-                output_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
-            SetAtomicNone();
-#else
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPad(output_gm[offset + len], output_local[len],
-                        dataCopyParams);
-#endif
-        }
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_row(int64_t idx) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3];
-
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3];
-
-        copy_in(input_offset, input_ne[0]);
-        LocalTensor<half> input_local = input_queue.DeQue<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        Cast(output_local, input_local, RoundMode::CAST_NONE,
-             local_buffer_elems);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset, input_ne[0]);
-
-        input_queue.FreeTensor(input_local);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            calculate_row(i);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    size_t local_buffer_elems;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<half> input_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    int64_t op_block_idx;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_f16(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
-    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_F16 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
-            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/get_row_f32.cpp

@@ -1,190 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-class GET_ROW_F32 {
-   public:
-    __aicore__ inline GET_ROW_F32() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
-                                int64_t *output_ne_ub, size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ float *)input);
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        uint64_t local_buffer_size = ((input_ne[0] * sizeof(float) + 31) & ~31);
-        local_buffer_elems = local_buffer_size / sizeof(float);
-
-        // TODO, consider long row that can't put in UB.
-        // All data should asign to 32. It's ok because all data is align to 32.
-        pipe.InitBuffer(input_queue, BUFFER_NUM, local_buffer_size);
-        pipe.InitBuffer(output_queue, BUFFER_NUM, local_buffer_size);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
-        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if(tail != 0) {
-            len += elem_per_block;
-        }
-        DataCopy(input_local, input_gm[offset], len);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(output_gm[offset], output_local, len);
-        }
-
-        if(tail != 0) {
-#ifdef ASCEND_310P
-            for (size_t i = tail; i < elem_per_block; i++) {
-                output_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
-            SetAtomicNone();
-#else
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPad(output_gm[offset + len], output_local[len],
-                        dataCopyParams);
-#endif
-        }
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_row(int64_t idx) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3];
-
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3];
-
-        copy_in(input_offset, input_ne[0]);
-        LocalTensor<float> input_local = input_queue.DeQue<float>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        DataCopy(output_local, input_local, local_buffer_elems);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset, input_ne[0]);
-
-        input_queue.FreeTensor(input_local);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            calculate_row(i);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    size_t local_buffer_elems;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<float> input_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    int64_t op_block_idx;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_f32(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
-    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_F32 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
-            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/get_row_q4_0.cpp

@@ -1,204 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support 4bit get row
-    extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
-        GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-        GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-        GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support 4bit get row.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-
-#define QK4_0 32
-
-class GET_ROW_Q4_0 {
-   public:
-    __aicore__ inline GET_ROW_Q4_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
-                                size_t *indices_nb_ub, int64_t *output_ne_ub,
-                                size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-            scale_ne[i] = input_ne_ub[i];
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // one scale for a group.
-        scale_ne[0] /= QK4_0;
-
-        input_stride[0] = 1;
-        scale_stride[0] = 1;
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        group_size_in_row = input_ne[0] / QK4_0;
-        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
-                               input_ne[3] / 2;
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ int4b_t *)input);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK4_0 * sizeof(int4b_t));
-        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK4_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK4_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<int4b_t> input_local = input_queue.AllocTensor<int4b_t>();
-        // 32 * sizeof(int4b_t) = 16, which is not aligned to 32, why no error?
-        DataCopy(input_local, input_gm[offset], QK4_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        DataCopy(output_gm[offset], output_local, QK4_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3] +
-                                     group * QK4_0;
-        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
-                                     indices_ne1_idx * scale_stride[2] +
-                                     indices_ne2_idx * scale_stride[3] + group;
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3] +
-                                      group * QK4_0;
-
-        copy_in(input_offset);
-        LocalTensor<int4b_t> input_local = input_queue.DeQue<int4b_t>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        // TODO: cast more data to speed up.
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
-        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
-
-        // Only mul need compile by group.
-        half scale = scale_gm.GetValue(scale_offset);
-
-        Muls(output_local, output_local, (float)scale, QK4_0);
-
-        input_queue.FreeTensor(input_local);
-        cast_queue.FreeTensor(cast_local);
-        output_queue.EnQue(output_local);
-
-        copy_out(output_offset);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                calculate_group(i, j);
-            }
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t scale_ne[4];
-    size_t scale_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t ir;
-    int64_t dr;
-
-    int64_t group_size_in_row;
-
-    TPipe pipe;
-    GlobalTensor<int4b_t> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_Q4_0 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
-            indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/get_row_q8_0.cpp

@@ -1,191 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-#define QK8_0 32
-
-class GET_ROW_Q8_0 {
-   public:
-    __aicore__ inline GET_ROW_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
-                                size_t *indices_nb_ub, int64_t *output_ne_ub,
-                                size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-            scale_ne[i] = input_ne_ub[i];
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // one scale for a group.
-        scale_ne[0] /= QK8_0;
-
-        input_stride[0] = 1;
-        scale_stride[0] = 1;
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        group_size_in_row = input_ne[0] / QK8_0;
-        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
-                               input_ne[3] * sizeof(int8_t);
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ int8_t *)input);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK8_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<int8_t> input_local = input_queue.AllocTensor<int8_t>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3] +
-                                     group * QK8_0;
-        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
-                                     indices_ne1_idx * scale_stride[2] +
-                                     indices_ne2_idx * scale_stride[3] + group;
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3] +
-                                      group * QK8_0;
-
-        copy_in(input_offset);
-        LocalTensor<int8_t> input_local = input_queue.DeQue<int8_t>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        // TODO: cast more data to speed up.
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
-        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK8_0);
-
-        // Only mul need compile by group.
-        half scale = scale_gm.GetValue(scale_offset);
-        Muls(output_local, output_local, (float)scale, QK8_0);
-
-        input_queue.FreeTensor(input_local);
-        cast_queue.FreeTensor(cast_local);
-        output_queue.EnQue(output_local);
-
-        copy_out(output_offset);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                calculate_group(i, j);
-            }
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t scale_ne[4];
-    size_t scale_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t ir;
-    int64_t dr;
-
-    int64_t group_size_in_row;
-
-    TPipe pipe;
-    GlobalTensor<int8_t> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_q8_0(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_Q8_0 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
-            indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp

@@ -1,218 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P
-    extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f16->8bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define QK8_0 32
-
-class QUANTIZE_F16_Q8_0 {
-   public:
-    __aicore__ inline QUANTIZE_F16_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / QK8_0;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t);
-
-        input_gm.SetGlobalBuffer((__gm__ half *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + ir *
-                                                 group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(work_queue, 1, 32);
-        pipe.InitBuffer(max_queue, 1, 32);
-        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
-        pipe.InitBuffer(scale_queue, 1, 32);
-        pipe.InitBuffer(cast_queue ,1 ,QK8_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + QK8_0 * group;
-
-        const int64_t output_offset = i1 * output_stride[1] +
-                                      i2 * output_stride[2] +
-                                      i3 * output_stride[3] + QK8_0 * group;
-
-        copy_in(input_offset);
-        LocalTensor<half> input_local = input_queue.DeQue<half>();
-        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
-
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
-        Abs(abs_local, cast_local, QK8_0);
-        ReduceMax(max_local, abs_local, work_local, QK8_0);
-
-        pipe_barrier(PIPE_ALL);
-        float d = max_local.GetValue(0);
-        d = d / ((1 << 7) - 1);
-        if (d != 0) {
-            Muls(cast_local, cast_local, 1.0f / d, QK8_0);
-        }
-
-        Cast(cast_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(input_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(output_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        abs_queue.FreeTensor(abs_local);
-        max_queue.FreeTensor(max_local);
-        cast_queue.FreeTensor(cast_local);
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                if (scale_local_offset == 16) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += 16;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<half> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, 1> work_queue;
-    TQue<QuePosition::VECOUT, 1> max_queue;
-    TQue<QuePosition::VECIN, 1> abs_queue;
-    TQue<QuePosition::VECOUT, 1> scale_queue;
-    TQue<QuePosition::VECOUT, 1> cast_queue;
-
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_F16_Q8_0 op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp

@@ -1,216 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support f32->8bit quantization
-    extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f32->8bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define QK8_0 32
-
-class QUANTIZE_F32_Q8_0 {
-   public:
-    __aicore__ inline QUANTIZE_F32_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / QK8_0;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t);
-
-        input_gm.SetGlobalBuffer((__gm__ float *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size +
-                                                 ir * group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(float));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(work_queue, 1, 32);
-        pipe.InitBuffer(max_queue, 1, 32);
-        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
-        pipe.InitBuffer(cast_queue, 1, QK8_0 * sizeof(half));
-        pipe.InitBuffer(scale_queue, 1, 32);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + QK8_0 * group;
-
-        const int64_t output_offset = i1 * output_stride[1] +
-                                      i2 * output_stride[2] +
-                                      i3 * output_stride[3] + QK8_0 * group;
-
-        copy_in(input_offset);
-        LocalTensor<float> input_local = input_queue.DeQue<float>();
-        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-
-        Abs(abs_local, input_local, QK8_0);
-        ReduceMax(max_local, abs_local, work_local, QK8_0);
-        pipe_barrier(PIPE_ALL);
-        float d = max_local.GetValue(0);
-        d = d / ((1 << 7) - 1);
-        if (d != 0) {
-            Muls(input_local, input_local, 1.0f / d, QK8_0);
-        }
-
-        Cast(input_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(cast_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(output_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        abs_queue.FreeTensor(abs_local);
-        max_queue.FreeTensor(max_local);
-        cast_queue.FreeTensor(cast_local);
-
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                if (scale_local_offset == 16) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += 16;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<float> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, 1> work_queue;
-    TQue<QuePosition::VECOUT, 1> max_queue;
-    TQue<QuePosition::VECIN, 1> abs_queue;
-    TQue<QuePosition::VECIN, 1> cast_queue;
-    TQue<QuePosition::VECOUT, 1> scale_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_F32_Q8_0 op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -1,295 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support float->4bit quantization
-    extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f32->4bit quantization.\n");
-    }
-
-    extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f16->4bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define Group_Size 32
-
-template <typename SRC_T>
-class QUANTIZE_FLOAT_TO_Q4_0 {
-   public:
-    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
-        //                         permute=[0,0,0,0]):
-        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        // input stride of data elements
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        // output stride of data elements
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / Group_Size;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t) / 2;
-
-        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
-                                                 group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
-        pipe.InitBuffer(output_queue, BUFFER_NUM,
-                            Group_Size * sizeof(int8_t) / 2);
-        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
-        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
-        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
-        DataCopy(input_local, input_gm[offset], Group_Size);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
-        // and using DataCopyPad to avoid 32 bits align.
-        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
-        LocalTensor<int8_t> output_int8_local =
-                                    output_local.ReinterpretCast<int8_t>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
-        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
-
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<float> input_local) {
-        DataCopy(cast_local, input_local, Group_Size);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<half> input_local) {
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + Group_Size * group;
-
-        // output_offset is stride for output_gm which datatype is int8_t and
-        // divided by 2 is needed for int4b_t.
-        const int64_t output_offset = (i1 * output_stride[1] +
-                                       i2 * output_stride[2] +
-                                       i3 * output_stride[3] +
-                                       Group_Size * group) / 2;
-        copy_in(input_offset);
-
-        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
-        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
-        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
-        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
-        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
-
-        input_to_cast(cast_local, input_local);
-
-        ReduceMax(max_local, cast_local, work_local, Group_Size);
-        ReduceMin(min_local, cast_local, work_local, Group_Size);
-        const float max_value = max_local.GetValue(0);
-        const float min_value = min_local.GetValue(0);
-        float d = max_value;
-        if (min_value < 0 && (-1 * min_value) > max_value) {
-            d = min_value;
-        }
-
-        d = d / (-8);
-        if (d != 0) {
-            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
-        }
-
-        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
-        float scalar = 8.5f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
-        scalar = 15.0f;
-        Mins(cast_local, cast_local, scalar, Group_Size);
-        scalar = -8.0f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-
-        // float->half->int4b
-        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
-        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
-
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        max_queue.FreeTensor(max_local);
-        min_queue.FreeTensor(min_local);
-        int8_queue.FreeTensor(int8_local);
-        half_queue.FreeTensor(half_local);
-        cast_queue.FreeTensor(cast_local);
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                // Copy Group_Size/2 length data each time.
-                if (scale_local_offset == Group_Size / 2) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local,
-                                      Group_Size / 2);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += Group_Size / 2;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-        scale_queue.FreeTensor(scale_local);
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<SRC_T> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<half> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<float> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-common.h

@@ -158,6 +158,12 @@ typedef sycl::half2 ggml_half2;
 
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
+#ifdef _MSC_VER
+#define GGML_EXTENSION
+#else // _MSC_VER
+#define GGML_EXTENSION __extension__
+#endif // _MSC_VER
+
 #define QK4_0 32
 typedef struct {
     ggml_half d;           // delta
@@ -167,7 +173,7 @@ static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 b
 
 #define QK4_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
@@ -188,7 +194,7 @@ static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0
 
 #define QK5_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
@@ -209,7 +215,7 @@ static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block
 
 #define QK8_1 32
 typedef struct {
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d; // delta
             ggml_half s; // d * sum(qs[i])
@@ -250,7 +256,7 @@ static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0
 typedef struct {
     uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
     uint8_t qs[QK_K/4];      // quants
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
@@ -277,7 +283,7 @@ static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
 typedef struct {
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
@@ -294,7 +300,7 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2,
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
 typedef struct {
-    union {
+    GGML_EXTENSION union {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins

ggml/src/ggml-cpu/CMakeLists.txt

@@ -23,6 +23,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         ggml-cpu/amx/mmq.cpp
         ggml-cpu/amx/mmq.h
         ggml-cpu/ggml-cpu-impl.h
+        ggml-cpu/common.h
+        ggml-cpu/binary-ops.h
+        ggml-cpu/binary-ops.cpp
+        ggml-cpu/unary-ops.h
+        ggml-cpu/unary-ops.cpp
         )
 
     target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
@@ -289,23 +294,29 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
         message(STATUS "PowerPC detected")
-        if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
-            file(READ "/proc/cpuinfo" POWER10_M)
-        elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
-            execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
-        endif()
+        if (GGML_NATIVE)
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+                file(READ "/proc/cpuinfo" POWER10_M)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
+                execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
+            endif()
 
-        string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
-        string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
-        if (EXTRACTED_NUMBER GREATER_EQUAL 10)
-            list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
-        elseif (EXTRACTED_NUMBER EQUAL 9)
-            list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
-        elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-            list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+            if (EXTRACTED_NUMBER GREATER_EQUAL 10)
+                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+            elseif (EXTRACTED_NUMBER EQUAL 9)
+                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+                list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+            else()
+                list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
+            endif()
         else()
-            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
+            if (GGML_CPU_POWERPC_CPUTYPE)
+                list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
+            endif()
         endif()
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         message(STATUS "loongarch64 detected")

ggml/src/ggml-cpu/binary-ops.cpp

@@ -0,0 +1,158 @@
+#include "binary-ops.h"
+
+#if defined(GGML_USE_ACCELERATE)
+#include <Accelerate/Accelerate.h>
+
+using vDSP_fn_t = void (*)(const float *, vDSP_Stride, const float *, vDSP_Stride, float *, vDSP_Stride, vDSP_Length);
+#endif
+
+static inline float op_add(float a, float b) {
+    return a + b;
+}
+
+static inline float op_sub(float a, float b) {
+    return a - b;
+}
+
+static inline float op_mul(float a, float b) {
+    return a * b;
+}
+
+static inline float op_div(float a, float b) {
+    return a / b;
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static inline void vec_binary_op_contiguous(const int64_t n, dst_t * z, const src0_t * x, const src1_t * y) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(y[i])));
+    }
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static inline void vec_binary_op_non_contiguous(const int64_t n, const int64_t ne10, const int64_t nb10, dst_t * z, const src0_t * x, const src1_t * y) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        int i10 = i % ne10;
+        const src1_t * y_ptr = (const src1_t *)((const char *)y + i10*nb10);
+        z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(*y_ptr)));
+    }
+}
+
+template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
+static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT( nb0 == sizeof(dst_t));
+    GGML_ASSERT(nb00 == sizeof(src0_t));
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+    const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
+
+    if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
+        GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    }
+
+#ifdef GGML_USE_ACCELERATE
+    vDSP_fn_t vDSP_op = nullptr;
+    // TODO - avoid the f32-only check using type 'trait' lookup tables and row-based src-to-float conversion functions
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (op == op_add) {
+            vDSP_op = vDSP_vadd;
+        } else if (op == op_sub) {
+            vDSP_op = vDSP_vsub;
+        } else if (op == op_mul) {
+            vDSP_op = vDSP_vmul;
+        } else if (op == op_div) {
+            vDSP_op = vDSP_vdiv;
+        }
+    }
+#endif
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        const int64_t i13 = i03 % ne13;
+        const int64_t i12 = i02 % ne12;
+        const int64_t i11 = i01 % ne11;
+
+        dst_t        * dst_ptr  = (dst_t  *)       ((char *)       dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+        const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+        if (is_src1_contiguous) {
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t nr0 = ne00 / ne10;
+
+            for (int64_t r = 0; r < nr0; ++r) {
+#ifdef GGML_USE_ACCELERATE
+                if constexpr (std::is_same_v<src0_t, float> && std::is_same_v<src1_t, float> && std::is_same_v<dst_t, float>) {
+                    if (vDSP_op != nullptr) {
+                        vDSP_op(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
+                        continue;
+                    }
+                }
+#endif
+                vec_binary_op_contiguous<op>(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
+            }
+        } else {
+            vec_binary_op_non_contiguous<op>(ne0, ne10, nb10, dst_ptr, src0_ptr, src1_ptr);
+        }
+    }
+}
+
+// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
+template <float (*op)(float, float)>
+static void binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    /*  */ if (src0->type == GGML_TYPE_F32  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
+        apply_binary_op<op, float, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
+        apply_binary_op<op, ggml_fp16_t, ggml_fp16_t, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
+        apply_binary_op<op, ggml_bf16_t, ggml_bf16_t, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_BF16) {
+        apply_binary_op<op, ggml_bf16_t, float, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) {
+        apply_binary_op<op, ggml_bf16_t, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F16) {
+        apply_binary_op<op, ggml_fp16_t, float, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && src1->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) {
+        apply_binary_op<op, ggml_fp16_t, float, float>(params, dst);
+    } else {
+        GGML_ABORT("%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
+    }
+}
+
+void ggml_compute_forward_add_non_quantized(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_add>(params, dst);
+}
+
+void ggml_compute_forward_sub(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_sub>(params, dst);
+}
+
+void ggml_compute_forward_mul(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_mul>(params, dst);
+}
+
+void ggml_compute_forward_div(const ggml_compute_params * params, ggml_tensor * dst) {
+    binary_op<op_div>(params, dst);
+}

ggml/src/ggml-cpu/binary-ops.h

@@ -0,0 +1,16 @@
+#pragma once
+
+#include "common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ggml_compute_forward_add_non_quantized(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sub(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_mul(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_div(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/common.h

@@ -0,0 +1,72 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpu-traits.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-impl.h"
+
+#ifdef __cplusplus
+
+#include <utility>
+
+// convenience functions/macros for use in template calls
+// note: these won't be required after the 'traits' lookup table is used.
+static inline ggml_fp16_t f32_to_f16(float x) {
+    return GGML_FP32_TO_FP16(x);
+}
+
+static inline float f16_to_f32(ggml_fp16_t x) {
+    return GGML_FP16_TO_FP32(x);
+}
+
+static inline ggml_bf16_t f32_to_bf16(float x) {
+    return GGML_FP32_TO_BF16(x);
+}
+
+static inline float bf16_to_f32(ggml_bf16_t x) {
+    return GGML_BF16_TO_FP32(x);
+}
+
+static inline float f32_to_f32(float x) {
+    return x;
+}
+
+// TODO - merge this into the traits table, after using row-based conversions
+template <class T>
+struct type_conversion_table;
+
+template <>
+struct type_conversion_table<ggml_fp16_t> {
+    static constexpr float (*to_f32)(ggml_fp16_t) = f16_to_f32;
+    static constexpr ggml_fp16_t (*from_f32)(float) = f32_to_f16;
+};
+
+template <>
+struct type_conversion_table<float> {
+    static constexpr float (*to_f32)(float) = f32_to_f32;
+    static constexpr float (*from_f32)(float) = f32_to_f32;
+};
+
+template <>
+struct type_conversion_table<ggml_bf16_t> {
+    static constexpr float (*to_f32)(ggml_bf16_t) = bf16_to_f32;
+    static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
+};
+
+static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
+    const int64_t ith = params->ith;
+    const int64_t nth = params->nth;
+
+    const int64_t nr  = ggml_nrows(src0);
+
+    // rows per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    return {ir0, ir1};
+}
+
+#endif

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

@@ -2680,13 +2680,25 @@ class tinyBLAS_PPC {
             __builtin_mma_xxsetaccz(&acc_0);
             vec_t vec_A[4] {0}, vec_B[4] = {0};
             for (int l=0; l<k; l+=4) {
-                if (RN >= 4 && RM == 1) {
+                /* 'GEMV Forwarding' concept is used in first two conditional loops.
+                 * when one of the matrix has a single row/column, the elements are
+                 * broadcasted, instead of using packing routine to prepack the
+                 * matrix elements.
+                 */
+                if (RM == 1) {
                     TA* a = const_cast<TA*>(A+(ii)*lda+l);
-                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
+                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
                     vec_A[0] = (vec_t)vec_xl(0,a);
                     vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1));
                     vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2));
                     vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3));
+                } else if (RN == 1) {
+                    packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
+                    TB* b = const_cast<TB*>(B+(jj)*ldb+l);
+                    vec_B[0] = (vec_t)vec_xl(0,b);
+                    vec_B[1] = (vec_t)vec_splats(*((TB*)&vec_B+1));
+                    vec_B[2] = (vec_t)vec_splats(*((TB*)&vec_B+2));
+                    vec_B[3] = (vec_t)vec_splats(*((TB*)&vec_B+3));
                 } else {
                     packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
                     packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
@@ -2790,8 +2802,10 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
     assert(params->ith < params->nth);
 
     // only enable sgemm for prompt processing
+#if !defined(__MMA__)
     if (n < 2)
         return false;
+#endif
 
     if (Ctype != GGML_TYPE_F32)
         return false;

ggml/src/ggml-cpu/unary-ops.cpp

@@ -0,0 +1,186 @@
+#include "unary-ops.h"
+
+static inline float op_abs(float x) {
+    return fabsf(x);
+}
+
+static inline float op_sgn(float x) {
+    return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f);
+}
+
+static inline float op_neg(float x) {
+    return -x;
+}
+
+static inline float op_step(float x) {
+    return (x > 0.f) ? 1.f : 0.f;
+}
+
+static inline float op_tanh(float x) {
+    return tanhf(x);
+}
+
+static inline float op_elu(float x) {
+    return (x > 0.f) ? x : expm1f(x);
+}
+
+static inline float op_relu(float x) {
+    return (x > 0.f) ? x : 0.f;
+}
+
+static inline float op_sigmoid(float x) {
+    return 1.f / (1.f + expf(-x));
+}
+
+static inline float op_hardsigmoid(float x) {
+    return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static inline float op_exp(float x) {
+    return expf(x);
+}
+
+static inline float op_hardswish(float x) {
+    return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static inline float op_sqr(float x) {
+    return x * x;
+}
+
+static inline float op_sqrt(float x) {
+    return sqrtf(x);
+}
+
+static inline float op_sin(float x) {
+    return sinf(x);
+}
+
+static inline float op_cos(float x) {
+    return cosf(x);
+}
+
+static inline float op_log(float x) {
+    return logf(x);
+}
+
+template <float (*op)(float), typename src0_t, typename dst_t>
+static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
+    constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
+    constexpr auto f32_to_dst  = type_conversion_table<dst_t >::from_f32;
+
+    for (int i = 0; i < n; i++) {
+        y[i] = f32_to_dst(op(src0_to_f32(x[i])));
+    }
+}
+
+template <float (*op)(float), typename src0_t, typename dst_t>
+static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT( nb0 == sizeof(dst_t));
+    GGML_ASSERT(nb00 == sizeof(src0_t));
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        dst_t        * dst_ptr  = (dst_t  *)       ((char *)       dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+        vec_unary_op<op>(ne0, dst_ptr, src0_ptr);
+    }
+}
+
+// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
+template <float (*op)(float)>
+static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    /*  */ if (src0->type == GGML_TYPE_F32  && dst->type == GGML_TYPE_F32) { // all f32
+        apply_unary_op<op, float, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F16) { // all f16
+        apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
+        apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
+    } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
+        apply_unary_op<op, ggml_bf16_t, float>(params, dst);
+    } else if (src0->type == GGML_TYPE_F16  && dst->type == GGML_TYPE_F32) {
+        apply_unary_op<op, ggml_fp16_t, float>(params, dst);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type));
+        GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_abs>(params, dst);
+}
+
+void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sgn>(params, dst);
+}
+
+void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_neg>(params, dst);
+}
+
+void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_step>(params, dst);
+}
+
+void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_tanh>(params, dst);
+}
+
+void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_elu>(params, dst);
+}
+
+void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_relu>(params, dst);
+}
+
+void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sigmoid>(params, dst);
+}
+
+void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_hardsigmoid>(params, dst);
+}
+
+void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_exp>(params, dst);
+}
+
+void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_hardswish>(params, dst);
+}
+
+void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sqr>(params, dst);
+}
+
+void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sqrt>(params, dst);
+}
+
+void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_sin>(params, dst);
+}
+
+void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_cos>(params, dst);
+}
+
+void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_log>(params, dst);
+}

ggml/src/ggml-cpu/unary-ops.h

@@ -0,0 +1,28 @@
+#pragma once
+
+#include "common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ggml_compute_forward_abs(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sgn(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_neg(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_step(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_tanh(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_elu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_hardsigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_exp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_hardswish(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sqr(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cuda/common.cuh

@@ -288,6 +288,10 @@ static __device__ void no_device_code(
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
+
+#if defined(GGML_USE_MUSA)
+    __builtin_unreachable();
+#endif // defined(GGML_USE_MUSA)
 }
 
 #ifdef __CUDA_ARCH__
@@ -725,7 +729,13 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    std::vector<char **> updated_kernel_arg;
+    bool use_cpy_indirection = false;
+    std::vector<char *> cpy_dest_ptrs;
+    char ** dest_ptrs_d;
+    int dest_ptrs_size = 0;
+    // Index to allow each cpy kernel to be aware of it's position within the graph
+    // relative to other cpy nodes.
+    int graph_cpynode_index = -1;
 #endif
 };
 

ggml/src/ggml-cuda/concat.cu

@@ -38,7 +38,7 @@ static __global__ void concat_f32_dim1(const float * x, const float * y, float *
         blockIdx.y * ne0 +
         blockIdx.z * ne0 * gridDim.y;
 
-    if (blockIdx.y < ne01) { // src0
+    if (blockIdx.y < (unsigned)ne01) { // src0
         int offset_src =
             nidx +
             blockIdx.y * ne0 +
@@ -64,7 +64,7 @@ static __global__ void concat_f32_dim2(const float * x, const float * y, float *
         blockIdx.y * ne0 +
         blockIdx.z * ne0 * gridDim.y;
 
-    if (blockIdx.z < ne02) { // src0
+    if (blockIdx.z < (unsigned)ne02) { // src0
         int offset_src =
             nidx +
             blockIdx.y * ne0 +

ggml/src/ggml-cuda/conv-transpose-1d.cu

@@ -34,6 +34,10 @@ static  __global__ void conv_transpose_1d_kernel(
         }
     }
     dst[global_index] = accumulator;
+    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
+    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
+    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
+    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(
@@ -75,8 +79,6 @@ void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor
     const int p0 = 0;//opts[3];
     const int d0 = 1;//opts[4];
 
-    const int64_t kernel_size = ggml_nelements(src0);
-    const int64_t input_size = ggml_nelements(src1);
     const int64_t output_size = ggml_nelements(dst);
 
     conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size,

ggml/src/ggml-cuda/convert.cu

@@ -577,7 +577,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
         return;
     }
 
-    const src_t * x = (src_t *) vx;
+    const src_t * x = (const src_t *) vx;
 
     y[i] = x[i];
 }

ggml/src/ggml-cuda/cpy.cu

@@ -32,16 +32,18 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
                                    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                   const int nb12, const int nb13) {
+                                   const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
     const int64_t i03 = i/(ne00 * ne01 * ne02);
@@ -288,16 +290,18 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -314,16 +318,18 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -339,66 +345,84 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
+// Copy destination pointers to GPU to be available when pointer indirection is in use
+
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+        if (cuda_graph->dest_ptrs_d != nullptr) {
+            CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d));
+        }
+        CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *)));
+        cuda_graph->dest_ptrs_size = host_dest_ptrs_size;
+    }
+    // copy destination pointers to GPU
+    CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
+    cuda_graph->graph_cpynode_index = 0; // reset index
+#endif
+}
+
 static void ggml_cpy_f16_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q8_0_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -407,22 +431,22 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -431,22 +455,22 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -455,22 +479,22 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -479,32 +503,32 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_iq4_nl_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f16_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
@@ -541,46 +565,60 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
+    char ** dest_ptrs_d = nullptr;
+    int graph_cpynode_index = -1;
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection) {
+        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
+        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
+    }
+#endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
         CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection) {
+        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
+    }
+#endif
+
 }
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/cpy.cuh

@@ -7,3 +7,5 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
+
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream);

ggml/src/ggml-cuda/fattn-common.cuh

@@ -315,14 +315,14 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
 
     float vals[sizeof(int)] = {0.0f};
 #pragma unroll
-    for (int l = 0; l < sizeof(int); ++l) {
+    for (int l = 0; l < int(sizeof(int)); ++l) {
         vals[l] = scale * x[4*threadIdx.x + l];
     }
 
     float amax = fabsf(vals[0]);
     float sum  = vals[0];
 #pragma unroll
-    for (int l = 1; l < sizeof(int); ++l) {
+    for (int l = 1; l < int(sizeof(int)); ++l) {
         amax = fmaxf(amax, fabsf(vals[l]));
         sum += vals[l];
     }
@@ -338,7 +338,7 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
 
     if (d != 0.0f) {
 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
             q8[l] = roundf(vals[l] / d);
         }
     }
@@ -638,7 +638,7 @@ static __global__ void flash_attn_combine_results(
     float VKQ_denominator = 0.0f;
     for (int l = 0; l < parallel_blocks; ++l) {
         const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
+        float KQ_max_scale = expf(diff);
         const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
         *((uint32_t *) &KQ_max_scale) &= ftz_mask;
 
@@ -649,6 +649,7 @@ static __global__ void flash_attn_combine_results(
     dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
 }
 
+[[noreturn]]
 static void on_no_fattn_vec_case(const int D) {
     if (D == 64) {
         fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -406,6 +406,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #endif // CP_ASYNC_AVAILABLE
 
 #else
+    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
+    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
+    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_KV);
+    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
+    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
+    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
+    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
+    GGML_UNUSED(kb0);
     NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -797,6 +806,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         __syncthreads();
     }
 #else
+    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
+    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
+    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1);
+    GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_KV); GGML_UNUSED(stride_mask);
+    GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
     NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -931,6 +946,16 @@ static __global__ void flash_attn_ext_f16(
         (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
+    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
 }
@@ -985,38 +1010,38 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/4, 4); \
     extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/8, 8); \
 
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,   8);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,   8);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,   8)
 
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  16);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  16);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  16)
 
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  32);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  32);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  32)
 
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  64);
-DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  64);
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256,  64)
 
 // Kernels with ncols == 128 are only 4% faster due to register pressure.
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128);
-// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128); // Needs too much shared memory.
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128)
+// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128) // Needs too much shared memory.

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -282,7 +282,19 @@ static __global__ void flash_attn_tile_ext_f16(
         }
     }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
 

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -281,6 +281,18 @@ static __global__ void flash_attn_tile_ext_f32(
         }
     }
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -292,7 +292,19 @@ static __global__ void flash_attn_vec_ext_f16(
         dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
 

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -277,6 +277,16 @@ static __global__ void flash_attn_vec_ext_f32(
         dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
+    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
+    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -430,7 +430,17 @@ static __global__ void flash_attn_ext_f16(
         dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
     }
 #else
-   NO_DEVICE_CODE;
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
 }
 

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

@@ -31,6 +31,8 @@
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/scale.cuh"
 #include "ggml-cuda/softmax.cuh"
+#include "ggml-cuda/ssm-conv.cuh"
+#include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
 #include "ggml-cuda/tsembd.cuh"
@@ -2296,6 +2298,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
+        case GGML_OP_SSM_CONV:
+            ggml_cuda_op_ssm_conv(ctx, dst);
+            break;
+        case GGML_OP_SSM_SCAN:
+            ggml_cuda_op_ssm_scan(ctx, dst);
+            break;
         case GGML_OP_ARGSORT:
             ggml_cuda_op_argsort(ctx, dst);
             break;
@@ -2433,10 +2441,11 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
 
 #ifdef USE_CUDA_GRAPH
 static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
-    std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool use_cuda_graph) {
+    bool use_cuda_graph) {
 
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
-    cuda_ctx->cuda_graph->updated_kernel_arg.clear();
+    cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
+
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
@@ -2468,8 +2477,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
         }
 
         if (node->op == GGML_OP_CPY) {
-            // store the copy op parameter which changes with each token.
-            cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
+
+            // Store the pointers which are updated for each token, such that these can be sent
+            // to the device and accessed using indirection from CUDA graph
+            cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data);
+
             // store a pointer to each copy op CUDA kernel to identify it later
             void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
             if (!ptr) {
@@ -2477,10 +2489,6 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #ifndef NDEBUG
                 GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
 #endif
-            } else {
-                if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
-                    ggml_cuda_cpy_fn_ptrs.push_back(ptr);
-                }
             }
         }
 
@@ -2489,6 +2497,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
         }
     }
 
+    if (use_cuda_graph) {
+        cuda_ctx->cuda_graph->use_cpy_indirection = true;
+        // copy pointers to GPU so they can be accessed via indirection within CUDA graph
+        ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream());
+    }
+
     return use_cuda_graph;
 }
 
@@ -2543,51 +2557,6 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     return true;
 }
 
-static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool cuda_graph_update_required) {
-
-    if (cuda_graph_update_required) {
-        // Extract nodes from graph
-        // First call with null argument gets number of nodes in graph
-        CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
-        // Subsequent call with non-null argument gets nodes
-        cuda_ctx->cuda_graph->nodes.clear();
-        cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
-        cuda_ctx->cuda_graph->params.clear();
-        cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
-        if (cuda_ctx->cuda_graph->num_nodes > 0) {
-            CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
-
-            // Loop over nodes, and extract kernel parameters from each node
-            for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
-                cudaGraphNodeType node_type;
-                CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
-                if (node_type == cudaGraphNodeTypeKernel) {
-                    cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
-                    if (stat == cudaErrorInvalidDeviceFunction) {
-                        // Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
-                        // We don't need to update blas nodes, so clear error and move on.
-                        (void)cudaGetLastError();
-                    } else {
-                        GGML_ASSERT(stat == cudaSuccess);
-                    }
-                }
-            }
-        }
-    } else {
-        // One of the arguments to the copy kernel is updated for each token, hence we need to
-        // replace that argument with the updated value in the CUDA graph
-        // on update steps, the live parameters will already be captured
-        int k = 0;
-        for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
-            if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
-                char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
-                *(void**)cuda_ctx->cuda_graph->params[i].kernelParams[1] = *(void**)updated_kernel_arg_ptr;
-                CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
-            }
-        }
-    }
-}
-
 static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
 
     bool cuda_graph_update_required = false;
@@ -2647,8 +2616,7 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 #endif
 
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
-   [[maybe_unused]] std::vector<void *> & ggml_cuda_cpy_fn_ptrs,  bool & graph_evaluated_or_captured, bool & use_cuda_graph,
-    bool & cuda_graph_update_required) {
+    bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
 
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@@ -2698,13 +2666,9 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
             CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
         }
-
-        // Perform update to graph (if required for this token), and change copy parameter (required for every token)
-        maintain_cuda_graph(cuda_ctx, ggml_cuda_cpy_fn_ptrs, cuda_graph_update_required);
-
-        // Update graph executable
-        update_cuda_graph_executable(cuda_ctx);
-
+        if (cuda_graph_update_required) { // Update graph executable
+            update_cuda_graph_executable(cuda_ctx);
+        }
         // Launch graph
         CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
 #else
@@ -2718,10 +2682,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 
     ggml_cuda_set_device(cuda_ctx->device);
 
-    // vector of pointers to CUDA cpy kernels, which are required to identify
-    // kernel parameters which need updated in the graph for each token
-    std::vector<void *> ggml_cuda_cpy_fn_ptrs;
-
 #ifdef USE_CUDA_GRAPH
     static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
 
@@ -2755,8 +2715,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     if (use_cuda_graph) {
         cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
 
-        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph,
-                             ggml_cuda_cpy_fn_ptrs, use_cuda_graph);
+        use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph);
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
         if (use_cuda_graph && cuda_graph_update_required) {
@@ -2777,6 +2736,10 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
+    if (!use_cuda_graph) {
+        cuda_ctx->cuda_graph->use_cpy_indirection = false;
+    }
+
 #else
     bool use_cuda_graph = false;
     bool cuda_graph_update_required = false;
@@ -2784,7 +2747,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 
     bool graph_evaluated_or_captured = false;
 
-    evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, ggml_cuda_cpy_fn_ptrs, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
+    evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3193,6 +3156,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_LOG:
+        case GGML_OP_SSM_SCAN:
+        case GGML_OP_SSM_CONV:
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
@@ -3232,6 +3197,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #ifndef FLASH_ATTN_AVAILABLE
             return false;
 #endif // FLASH_ATTN_AVAILABLE
+            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
+                // different head sizes of K and V are not supported yet
+                return false;
+            }
+            if (op->src[0]->ne[0] == 192) {
+                return false;
+            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/mma.cuh

@@ -26,6 +26,7 @@ static __device__ __forceinline__ int ggml_cuda_movmatrix(const int x) {
     asm("movmatrix.sync.aligned.m8n8.trans.b16 %0, %1;"
         : "=r"(ret) : "r"(x));
 #else
+    GGML_UNUSED(x);
     NO_DEVICE_CODE;
 #endif // defined(NEW_MMA_AVAILABLE)
     return ret;
@@ -178,6 +179,7 @@ namespace ggml_cuda_mma {
             : "l"(xs));
 #else
         load_generic(xs0, stride);
+        GGML_UNUSED(t);
 #endif // NEW_MMA_AVAILABLE
     }
 

ggml/src/ggml-cuda/mmq.cuh

@@ -945,7 +945,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
     NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -1024,7 +1024,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
     }
 
 #pragma unroll
-    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+    for (int k01 = 0; k01 < WARP_SIZE/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
         const int k0 = k00 + k01;
 
 #pragma unroll
@@ -1035,19 +1035,34 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
             for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
                 const int i = i0 + threadIdx.x;
 
-                if (k01 < WARP_SIZE/2) {
-                    constexpr int ns = 2;
-                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
-                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
-                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
-                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
-                } else {
-                    constexpr int ns = 1;
-                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
-                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
-                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
-                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
-                }
+                constexpr int ns = 2;
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                    &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+            }
+        }
+    }
+
+    // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop.
+    // As a workaround 2 separate loops are used instead.
+#pragma unroll
+    for (int k01 = WARP_SIZE/2; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                constexpr int ns = 1;
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                    &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
             }
         }
     }
@@ -1176,7 +1191,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
     NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -1253,7 +1268,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
         const float d = bxi->d;
 
 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
             x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*(threadIdx.x % (WARP_SIZE/8)) + l] = d*sc8[l];
         }
 #else
@@ -1376,7 +1391,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
         const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
 
 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
             x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
         }
     }
@@ -1517,7 +1532,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
         const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
 
 #pragma unroll
-        for (int l = 0; l < sizeof(int); ++l) {
+        for (int l = 0; l < int(sizeof(int)); ++l) {
             x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
         }
     }
@@ -1810,7 +1825,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
     NO_DEVICE_CODE;
 #endif // NEW_MMA_AVAILABLE
 }
@@ -2570,6 +2585,8 @@ static __device__ void mul_mat_q_process_tile(
     } else {
         write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
     }
+
+    GGML_UNUSED(ne00); GGML_UNUSED(ne10);
 }
 
 
@@ -2695,7 +2712,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
         const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
 
         // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
-        if (it != blockIdx.x || jt != blockIdx.y) {
+        if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) {
             continue;
         }
 
@@ -2825,7 +2842,6 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 template <ggml_type type>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     const int id    = ggml_cuda_get_device();
-    const int nsm   = ggml_cuda_info().devices[id].nsm;
     const int cc    = ggml_cuda_info().devices[id].cc;
     const int smpbo = ggml_cuda_info().devices[id].smpbo;
 

ggml/src/ggml-cuda/mmv.cu

@@ -29,7 +29,7 @@ static __global__ void mul_mat_vec(
         __syncthreads();
     }
 
-    float sumf;
+    float sumf = 0.0f;
 
     if constexpr (std::is_same<T, half>::value) {
         const half2 * x2 = (const half2 *) x;

ggml/src/ggml-cuda/mmvq.cu

@@ -151,7 +151,7 @@ static __global__ void mul_mat_vec_q(
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
     // partial sum for each thread
-    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
+    float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = (const block_q8_1 *) vy;
 
@@ -197,10 +197,12 @@ static __global__ void mul_mat_vec_q(
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
         }
 
-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
             dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }
+
+    GGML_UNUSED(nrows_x);
 }
 
 static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {

ggml/src/ggml-cuda/pad.cu

@@ -14,7 +14,7 @@ static __global__ void pad_f32(const float * x, float * dst, const int ne0, cons
         nidx +
         blockIdx.y * ne0 +
         blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+    if (nidx < ne00 && blockIdx.y < (unsigned)ne01 && blockIdx.z < (unsigned)(ne02*ne03)) {
         int offset_src =
             nidx +
             blockIdx.y * ne00 +

ggml/src/ggml-cuda/ssm-conv.cu

@@ -0,0 +1,151 @@
+#include "ssm-conv.cuh"
+
+template <size_t split_d_inner, size_t d_conv>
+static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+                                    const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
+                                    float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
+                                    const int nc, const int ncs, const int nr, const int n_t, const int n_s) {
+    const int tid  = threadIdx.x;
+    const int bidx = blockIdx.x;
+    const int bidy = blockIdx.y;
+
+    const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1);
+    const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1);
+    float *       y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0);
+
+    const int stride_x = src0_nb1 / sizeof(float);
+    const int stride_w = src1_nb1 / sizeof(float);
+    const int stride_y = dst_nb1 / sizeof(float);
+
+    float x[d_conv] = { 0.0f };
+    float w[d_conv] = { 0.0f };
+
+#pragma unroll
+    for (int j = 0; j < d_conv; j++) {
+        w[j] = w_block[tid * stride_w + j];
+    }
+
+    for (int i = 0; i < n_t; i++) {
+        float sumf = 0.0f;
+
+        if (i == 0) {
+            for (int j = 0; j < d_conv; j++) {
+                x[j] = x_block[tid * stride_x + j];
+            }
+        } else {
+            x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
+        }
+
+#pragma unroll
+        for (int j = 0; j < d_conv; j++) {
+            sumf += x[(i + j) % d_conv] * w[j];
+        }
+        y_block[i * stride_y + tid] = sumf;
+    }
+}
+
+template <size_t split_d_inner, size_t d_conv, size_t split_n_t>
+static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+                                               const int src0_nb0, const int src0_nb1, const int src0_nb2,
+                                               const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
+                                               const int dst_nb1, const int dst_nb2, const int nc, const int ncs,
+                                               const int nr, const int n_t, const int n_s) {
+    const int tid  = threadIdx.x;
+    const int bidx = blockIdx.x;
+    const int bidy = blockIdx.y;
+    const int bidz = blockIdx.z;
+
+    const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 +
+                                             bidz * split_n_t * src0_nb0);
+    const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1);
+    float *       y_block =
+        (float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0);
+
+    const int stride_x = src0_nb1 / sizeof(float);
+    const int stride_w = src1_nb1 / sizeof(float);
+    const int stride_y = dst_nb1 / sizeof(float);
+
+    float x[d_conv] = { 0.0f };
+    float w[d_conv] = { 0.0f };
+
+#pragma unroll
+    for (int j = 0; j < d_conv; j++) {
+        w[j] = w_block[tid * stride_w + j];
+    }
+
+#pragma unroll
+    for (int i = 0; i < split_n_t; i++) {
+        if (bidz * split_n_t + i < n_t) {
+            float sumf = 0.0f;
+
+            if (i == 0) {
+                for (int j = 0; j < d_conv; j++) {
+                    x[j] = x_block[tid * stride_x + j];
+                }
+            } else {
+                x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
+            }
+
+#pragma unroll
+            for (int j = 0; j < d_conv; j++) {
+                sumf += x[(i + j) % d_conv] * w[j];
+            }
+            y_block[i * stride_y + tid] = sumf;
+        }
+    }
+}
+
+static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
+                              const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
+                              const int dst_nb2, const int nc, const int ncs, const int nr, const int n_t,
+                              const int n_s, cudaStream_t stream) {
+    const int threads = 128;
+    GGML_ASSERT(nr % threads == 0);
+
+    if (n_t <= 32) {
+        const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
+        if (nc == 4) {
+            ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+                                                                     dst, dst_nb0, dst_nb1, dst_nb2, nc, ncs, nr, n_t,
+                                                                     n_s);
+        } else {
+            GGML_ABORT("Only support kernel size = 4  now.");
+        }
+    } else {
+        if (nc == 4) {
+            const int split_n_t = 32;
+            dim3      blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
+            ssm_conv_long_token_f32<threads, 4, split_n_t>
+                <<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0,
+                                                 dst_nb1, dst_nb2, nc, ncs, nr, n_t, n_s);
+        } else {
+            GGML_ABORT("Only support kernel size = 4 right now.");
+        }
+    }
+}
+
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];  // conv_x
+    const struct ggml_tensor * src1 = dst->src[1];  // conv1d.weight
+
+    const int nc  = src1->ne[0];                    // d_conv
+    const int ncs = src0->ne[0];                    // d_conv - 1 + n_t
+    const int nr  = src0->ne[1];                    // d_inner
+    const int n_t = dst->ne[1];                     // tokens per sequence
+    const int n_s = dst->ne[2];                     // number of sequences in the batch
+
+    GGML_ASSERT(dst->ne[0] == nr);
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float *       dst_d  = (float *) dst->data;
+    cudaStream_t  stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1],
+                      dst->nb[2], nc, ncs, nr, n_t, n_s, stream);
+}

ggml/src/ggml-cuda/ssm-conv.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ssm-scan.cu

@@ -0,0 +1,155 @@
+#include "ssm-scan.cuh"
+
+// #include <cuda_runtime.h>
+// static __device__ void global_to_shared(const float *src, float *dst) {
+//   asm volatile("cp.async.");
+// }
+
+template <size_t splitD, size_t N>
+__global__ void __launch_bounds__(splitD, 2)
+    ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
+                 const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+                 const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
+                 const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                 const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
+                 float * __restrict__ dst, const int D, const int L, const int B) {
+    const int bidx = blockIdx.x;  // split along B
+    const int bidy = blockIdx.y;  // split along D
+    const int tid  = threadIdx.x;
+    const int wid  = tid / 32;
+    const int wtid = tid % 32;
+
+    extern __shared__ float smem[];
+    const int               stride_sA  = N + 1;
+    const int               stride_ss0 = N + 1;
+    float *                 smem_A     = smem;
+    float *                 smem_s0    = smem_A + splitD * stride_sA;
+
+    const float * s0_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+    const float * x_block  = (const float *) ((char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    const float * dt_block = (const float *) ((char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
+    const float * A_block  = (const float *) ((char *) src3 + bidy * splitD * src3_nb1);
+    const float * B_block  = (const float *) ((char *) src4 + (bidx * src4_nb2));
+    const float * C_block  = (const float *) ((char *) src5 + (bidx * src5_nb2));
+    float *       y_block  = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    float *       s_block  = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+
+    const int stride_s0 = src0_nb1 / sizeof(float);
+    const int stride_x  = src1_nb1 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_A  = src3_nb1 / sizeof(float);
+    const int stride_B  = src4_nb1 / sizeof(float);
+    const int stride_C  = src5_nb1 / sizeof(float);
+    const int stride_s  = stride_s0;
+    const int stride_y  = stride_x;
+
+    // can N not be 16? for example 32?
+    if (N == 16) {
+#pragma unroll
+        for (int i = 0; i < splitD / 4; i += 2) {
+            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
+            // todo: bank conflict
+            // I am always confused with how to use the swizzling method to solve
+            // bank conflit. Hoping somebody can tell me.
+            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+        }
+#pragma unroll
+        for (int i = 0; i < splitD / 4; i += 2) {
+            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
+            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+        }
+    }
+
+    __syncthreads();
+
+    for (int i = 0; i < L; i++) {
+        float dt_soft_plus = dt_block[i * stride_dt + tid];
+        if (dt_soft_plus <= 20.0f) {
+            dt_soft_plus = log1pf(exp(dt_soft_plus));
+        }
+        float x_dt = x_block[i * stride_x + tid] * dt_soft_plus;
+        float sumf = 0.0f;
+#pragma unroll
+        for (int j = 0; j < N; j++) {
+            float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) +
+                          (B_block[i * stride_B + j] * x_dt);
+            sumf += state * C_block[i * stride_C + j];
+            if (i == L - 1) {
+                s_block[tid * stride_s + j] = state;
+            } else {
+                smem_s0[tid * stride_ss0 + j] = state;
+            }
+        }
+        __syncthreads();
+        y_block[i * stride_y + tid] = sumf;
+    }
+}
+
+static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
+                              const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
+                              const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
+                              const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                              const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
+                              float * dst, const int N, const int D, const int L, const int B, cudaStream_t stream) {
+    const int threads = 128;
+    // todo: consider D cannot be divided,does this situation exist?
+    GGML_ASSERT(D % threads == 0);
+    const dim3 blocks(B, (D + threads - 1) / threads, 1);
+    const int  smem_size = (threads * (N + 1) * 2) * sizeof(float);
+    if (N == 16) {
+        ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+            src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
+            src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, D, L, B);
+    } else {
+        GGML_ABORT("doesn't support N!=16.");
+    }
+}
+
+void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];  // s
+    const struct ggml_tensor * src1 = dst->src[1];  // x
+    const struct ggml_tensor * src2 = dst->src[2];  // dt
+    const struct ggml_tensor * src3 = dst->src[3];  // A
+    const struct ggml_tensor * src4 = dst->src[4];  // B
+    const struct ggml_tensor * src5 = dst->src[5];  // C
+
+    //   const int64_t d_state = src0->ne[0];
+    //   const int64_t d_inner = src0->ne[1];
+    //   const int64_t l = src1->ne[1];
+    //   const int64_t b = src0->ne[2];
+
+    const int64_t nc  = src0->ne[0];  // d_state
+    const int64_t nr  = src0->ne[1];  // d_inner
+    const int64_t n_t = src1->ne[1];  // number of tokens per sequence
+    const int64_t n_s = src0->ne[2];  // number of sequences in the batch
+
+    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+    GGML_ASSERT(src2->nb[0] == sizeof(float));
+    GGML_ASSERT(src3->nb[0] == sizeof(float));
+    GGML_ASSERT(src4->nb[0] == sizeof(float));
+    GGML_ASSERT(src5->nb[0] == sizeof(float));
+    // required for the dot product between s and C
+    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
+    // required for per-sequence offsets for states
+    GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
+    // required to get correct offset for state destination (i.e. src1->nb[3])
+    GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    const float * src2_d = (const float *) src2->data;
+    const float * src3_d = (const float *) src3->data;
+    const float * src4_d = (const float *) src4->data;
+    const float * src5_d = (const float *) src5->data;
+    float *       dst_d  = (float *) dst->data;
+    cudaStream_t  stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
+                      src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
+                      src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
+}

ggml/src/ggml-cuda/ssm-scan.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/upscale.cu

@@ -19,7 +19,7 @@ static __global__ void upscale_f32(const float * x, float * dst,
     int i02 = i12 / sf2;
     int i03 = i13 / sf3;
 
-    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+    dst[index] = *( (const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00) );
 }
 
 static void upscale_f32_cuda(const float * x, float * dst,

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -219,9 +219,12 @@ typedef struct {
     int32_t  ne11;
     int32_t  ne_12_2; // assume K and V are same shape
     int32_t  ne_12_3;
-    uint64_t nb_12_1;
-    uint64_t nb_12_2;
-    uint64_t nb_12_3;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
     uint64_t nb31;
     int32_t  ne1;
     int32_t  ne2;

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

@@ -48,7 +48,7 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
 
 template <typename type4>
 void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
-    reg = (type4)(*(src + il));
+    reg = (type4)(*(src));
 }
 
 #if defined(GGML_METAL_USE_BF16)
@@ -56,6 +56,11 @@ 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>
@@ -3100,7 +3105,8 @@ template<
     typename vd4x4_t, // key type in device memory
     short nl_v,
     void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
-    short D,         // head size
+    short DK,        // K head size
+    short DV,        // V head size
     short Q  = 8,    // queries per threadgroup
     short KV = 8,    // key/value processed per each simdgroup
     short C  = 32>   // cache items per threadgroup
@@ -3122,20 +3128,24 @@ kernel void kernel_flash_attn_ext(
     const int iq2 = tgpig[1];
     const int iq1 = tgpig[0]*Q;
 
-    const short D4  = D/4;
-    const short D8  = D/8;
-    const short D16 = D/16;
-    const short NW  = N_SIMDWIDTH;
-    const short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
+    constexpr short DK4  = DK/4;
+    constexpr short DK8  = DK/8;
+    constexpr short DK16 = DK/16;
+    constexpr short DV4  = DV/4;
+    constexpr short DV8  = DV/8;
+    constexpr short DV16 = DV/16;
+
+    constexpr short NW  = N_SIMDWIDTH;
+    constexpr short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
 
     const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
-    const short T  = D + 2*TS; // shared memory size per query in (half)
+    const short T  = DK + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*D); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*D); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*D); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*D); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*DK); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*DK); // same as above but in q4_t
+    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*DK); // reuse query data for accumulation
+    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*DK); // same as above but in o4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
     threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
@@ -3144,23 +3154,23 @@ kernel void kernel_flash_attn_ext(
     threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t
 
     // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
-    o8x8_t lo[D8];
+    o8x8_t lo[DV8];
 
     // load heads from Q to shared memory
     for (short j = sgitg; j < Q; j += nsg) {
         device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
-        for (short i = tiisg; i < D4; i += NW) {
+        for (short i = tiisg; i < DK4; i += NW) {
             if (iq1 + j < args.ne01) {
-                sq4[j*D4 + i] = (q4_t) q4[i];
+                sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*D4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = (q4_t) 0.0f;
             }
         }
     }
 
     // zero out lo
-    for (short i = 0; i < D8; ++i) {
+    for (short i = 0; i < DV8; ++i) {
         lo[i] = make_filled_simdgroup_matrix<o_t, 8>((o_t) 0.0f);
     }
 
@@ -3174,8 +3184,8 @@ kernel void kernel_flash_attn_ext(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     {
-        half S[Q] = { [0 ... Q-1] = 0.0f };
-        half M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
+        float S[Q] = { [0 ... Q-1] = 0.0f };
+        float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
 
         // thread indices inside the simdgroup
         // TODO: see if we can utilize quad-group functions for better performance
@@ -3190,22 +3200,15 @@ kernel void kernel_flash_attn_ext(
         const short ikv2 = iq2/(args.ne02/args.ne_12_2);
         const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
-        // load the queries from shared memory into local memory
-        q8x8_t mq[D8];
-
-        for (short i = 0; i < D8; ++i) {
-            simdgroup_load(mq[i], sq + i*8, D);
-        }
-
         const bool has_mask = mask != q;
 
-        half slope = 1.0f;
+        float slope = 1.0f;
 
         // ALiBi
         if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const float base = h < args.n_head_log2 ? args.m0 : args.m1;
             const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
@@ -3221,14 +3224,14 @@ kernel void kernel_flash_attn_ext(
 
             if (has_mask) {
                 // used to detect blocks full of -INF
-                half smax = -INFINITY;
+                float smax = -INFINITY;
 
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
                     device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
 
-                    const half m = pm[ic + tiisg];
+                    const float m = pm[ic + tiisg];
 
                     ss[j*TS + C + tiisg] = m;
                     smax = max(smax, m);
@@ -3249,20 +3252,22 @@ kernel void kernel_flash_attn_ext(
                     // this is compile-time check, so it does not have runtime overhead
                     if (is_same<kd4x4_t, k4x4_t>::value) {
                         // we can read directly from global memory
-                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13));
 
-                        #pragma unroll(D8)
-                        for (short i = 0; i < D8; ++i) {
+                        #pragma unroll(DK8)
+                        for (short i = 0; i < DK8; ++i) {
                             k8x8_t mk;
-                            simdgroup_load(mk, pk + i*8, args.nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
+                            simdgroup_load(mk, pk + i*8, args.nb11/sizeof(k_t), 0, true); // transpose // TODO: use ne10
 
-                            simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
+                            q8x8_t mq;
+                            simdgroup_load(mq, sq + i*8, DK);
+                            simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
                         }
                     } else {
-                        for (short ii = 0; ii < D16; ii += 4) {
-                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                        for (short ii = 0; ii < DK16; ii += 4) {
+                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13));
 
-                            if (D16%4 == 0) {
+                            if (DK16%4 == 0) {
                                 // the head is evenly divisible by 4*16 = 64, so no need for bound checks
                                 {
                                     k4x4_t tmp;
@@ -3275,15 +3280,18 @@ kernel void kernel_flash_attn_ext(
                                 #pragma unroll(4)
                                 for (short k = 0; k < 4; ++k) {
                                     k8x8_t mk;
+                                    q8x8_t mq;
 
                                     simdgroup_load(mk, sk + 16*k + 0*8, 4*16, 0, true); // transpose
-                                    simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 0], mk, mqk);
+                                    simdgroup_load(mq, sq + (2*(ii + k) + 0)*8, DK);
+                                    simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
 
                                     simdgroup_load(mk, sk + 16*k + 1*8, 4*16, 0, true); // transpose
-                                    simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 1], mk, mqk);
+                                    simdgroup_load(mq, sq + (2*(ii + k) + 1)*8, DK);
+                                    simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
                                 }
                             } else {
-                                if (ii + tx < D16) {
+                                if (ii + tx < DK16) {
                                     k4x4_t tmp;
                                     deq_k(pk4x4 + (ii + tx)/nl_k, (ii + tx)%nl_k, tmp);
                                     sk4x4[4*ty + tx] = tmp;
@@ -3291,14 +3299,17 @@ kernel void kernel_flash_attn_ext(
 
                                 simdgroup_barrier(mem_flags::mem_threadgroup);
 
-                                for (short k = 0; k < 4 && ii + k < D16; ++k) {
+                                for (short k = 0; k < 4 && ii + k < DK16; ++k) {
                                     k8x8_t mk;
+                                    q8x8_t mq;
 
                                     simdgroup_load(mk, sk + 16*k + 0*8, 4*16, 0, true); // transpose
-                                    simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 0], mk, mqk);
+                                    simdgroup_load(mq, sq + (2*(ii + k) + 0)*8, DK);
+                                    simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
 
                                     simdgroup_load(mk, sk + 16*k + 1*8, 4*16, 0, true); // transpose
-                                    simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 1], mk, mqk);
+                                    simdgroup_load(mq, sq + (2*(ii + k) + 1)*8, DK);
+                                    simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
                                 }
                             }
                         }
@@ -3316,10 +3327,10 @@ kernel void kernel_flash_attn_ext(
             // online softmax
             {
                 for (ushort j = 0; j < Q; ++j) {
-                    const half m = M[j];
+                    const float m = M[j];
 
                     // scale and apply the logitcap / mask
-                    half s = ss[j*TS + tiisg]*args.scale;
+                    float s = ss[j*TS + tiisg]*args.scale;
 
                     if (args.logit_softcap != 0.0f) {
                         s = args.logit_softcap*precise::tanh(s);
@@ -3330,8 +3341,8 @@ kernel void kernel_flash_attn_ext(
 
                     M[j] = simd_max(max(M[j], s));
 
-                    const half ms = exp(m - M[j]);
-                    const half vs = exp(s - M[j]);
+                    const float ms = exp(m - M[j]);
+                    const float vs = exp(s - M[j]);
 
                     S[j] = S[j]*ms + simd_sum(vs);
 
@@ -3350,8 +3361,8 @@ kernel void kernel_flash_attn_ext(
                 s8x8_t mm;
                 simdgroup_load(mm, ss + 2*C, TS, 0, false);
 
-                #pragma unroll(D8)
-                for (short i = 0; i < D8; ++i) {
+                #pragma unroll(DV8)
+                for (short i = 0; i < DV8; ++i) {
                     simdgroup_multiply(lo[i], mm, lo[i]);
                 }
             }
@@ -3364,20 +3375,20 @@ kernel void kernel_flash_attn_ext(
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
-                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23));
 
-                        #pragma unroll(D8)
-                        for (short i = 0; i < D8; ++i) {
+                        #pragma unroll(DV8)
+                        for (short i = 0; i < DV8; ++i) {
                             v8x8_t mv;
-                            simdgroup_load(mv, pv + i*8, args.nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
+                            simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20
 
                             simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
                         }
                     } else {
-                        for (short ii = 0; ii < D16; ii += 4) {
-                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                        for (short ii = 0; ii < DV16; ii += 4) {
+                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23));
 
-                            if (D16%4 == 0) {
+                            if (DV16%4 == 0) {
                                 // no need for bound checks
                                 {
                                     v4x4_t tmp;
@@ -3398,7 +3409,7 @@ kernel void kernel_flash_attn_ext(
                                     simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
-                                if (ii + tx < D16) {
+                                if (ii + tx < DV16) {
                                     v4x4_t tmp;
                                     deq_v(pv4x4 + (ii + tx)/nl_v, (ii + tx)%nl_v, tmp);
                                     sv4x4[4*ty + tx] = tmp;
@@ -3406,7 +3417,7 @@ kernel void kernel_flash_attn_ext(
 
                                 simdgroup_barrier(mem_flags::mem_threadgroup);
 
-                                for (short k = 0; k < 4 && ii + k < D16; ++k) {
+                                for (short k = 0; k < 4 && ii + k < DV16; ++k) {
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
@@ -3433,15 +3444,15 @@ kernel void kernel_flash_attn_ext(
 
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        half S = { 0.0f };
-        half M = { -__FLT16_MAX__/2 };
+        float S = { 0.0f };
+        float M = { -__FLT16_MAX__/2 };
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // each simdgroup stores its output to shared memory, reusing sq
         if (sgitg == sg) {
-            for (short i = 0; i < D8; ++i) {
-                simdgroup_store(lo[i], so + i*8, D, 0, false);
+            for (short i = 0; i < DV8; ++i) {
+                simdgroup_store(lo[i], so + i*8, DV, 0, false);
             }
         }
 
@@ -3450,16 +3461,16 @@ kernel void kernel_flash_attn_ext(
         // the first simdgroup accumulates the results from the other simdgroups
         if (sgitg == 0) {
             for (short j = 0; j < Q; ++j) {
-                const half S0 = ss[j*TS +         0];
-                const half S1 = ss[j*TS + sg*SH + 0];
+                const float S0 = ss[j*TS +         0];
+                const float S1 = ss[j*TS + sg*SH + 0];
 
-                const half M0 = ss[j*TS +         1];
-                const half M1 = ss[j*TS + sg*SH + 1];
+                const float M0 = ss[j*TS +         1];
+                const float M1 = ss[j*TS + sg*SH + 1];
 
                 M = max(M0, M1);
 
-                const half ms0 = exp(M0 - M);
-                const half ms1 = exp(M1 - M);
+                const float ms0 = exp(M0 - M);
+                const float ms1 = exp(M1 - M);
 
                 S = S0*ms0 + S1*ms1;
 
@@ -3480,11 +3491,11 @@ kernel void kernel_flash_attn_ext(
                 simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
                 simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
 
-                #pragma unroll(D8)
-                for (short i = 0; i < D8; ++i) {
+                #pragma unroll(DV8)
+                for (short i = 0; i < DV8; ++i) {
                     o8x8_t t;
 
-                    simdgroup_load    (t, so + i*8, D, 0, false);
+                    simdgroup_load    (t, so + i*8, DV, 0, false);
                     simdgroup_multiply(t, ms1, t);
 
                     simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
@@ -3495,8 +3506,8 @@ kernel void kernel_flash_attn_ext(
 
     // store result to shared memory (reuse sq)
     if (sgitg == 0) {
-        for (short i = 0; i < D8; ++i) {
-            simdgroup_store(lo[i], so + i*8, D, 0, false);
+        for (short i = 0; i < DV8; ++i) {
+            simdgroup_store(lo[i], so + i*8, DV, 0, false);
         }
     }
 
@@ -3507,8 +3518,8 @@ kernel void kernel_flash_attn_ext(
         for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
             const float S = ss[j*TS + 0];
 
-            for (short i = tiisg; i < D4; i += NW) {
-                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*D4 + i] = (float4) so4[j*D4 + i]/S;
+            for (short i = tiisg; i < DV4; i += NW) {
+                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S;
             }
         }
     }
@@ -3525,80 +3536,94 @@ kernel void kernel_flash_attn_ext(
     float,          simdgroup_float8x8, \
     half,  half4,   simdgroup_half8x8
 
-typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64>) flash_attn_ext_t;
+typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
-template [[host_name("kernel_flash_attn_ext_f16_h64" )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64>;
-template [[host_name("kernel_flash_attn_ext_f16_h80" )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80>;
-template [[host_name("kernel_flash_attn_ext_f16_h96" )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96>;
-template [[host_name("kernel_flash_attn_ext_f16_h112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112>;
-template [[host_name("kernel_flash_attn_ext_f16_h128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  128>;
-template [[host_name("kernel_flash_attn_ext_f16_h256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256>;
+template [[host_name("kernel_flash_attn_ext_f16_h64" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f16_h80" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
+template [[host_name("kernel_flash_attn_ext_f16_h96" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
+template [[host_name("kernel_flash_attn_ext_f16_h112")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112, 112>;
+template [[host_name("kernel_flash_attn_ext_f16_h128")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  128, 128>;
+template [[host_name("kernel_flash_attn_ext_f16_h192")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 192>;
+template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
+template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
 
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
 #endif
 
-template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64>;
-template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80>;
-template [[host_name("kernel_flash_attn_ext_q4_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96>;
-template [[host_name("kernel_flash_attn_ext_q4_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112>;
-template [[host_name("kernel_flash_attn_ext_q4_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 128>;
-template [[host_name("kernel_flash_attn_ext_q4_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_q4_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
 
-template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64>;
-template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80>;
-template [[host_name("kernel_flash_attn_ext_q4_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96>;
-template [[host_name("kernel_flash_attn_ext_q4_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112>;
-template [[host_name("kernel_flash_attn_ext_q4_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 128>;
-template [[host_name("kernel_flash_attn_ext_q4_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_q4_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
 
-template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64>;
-template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80>;
-template [[host_name("kernel_flash_attn_ext_q5_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96>;
-template [[host_name("kernel_flash_attn_ext_q5_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112>;
-template [[host_name("kernel_flash_attn_ext_q5_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 128>;
-template [[host_name("kernel_flash_attn_ext_q5_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_q5_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
 
-template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64>;
-template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80>;
-template [[host_name("kernel_flash_attn_ext_q5_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96>;
-template [[host_name("kernel_flash_attn_ext_q5_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112>;
-template [[host_name("kernel_flash_attn_ext_q5_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 128>;
-template [[host_name("kernel_flash_attn_ext_q5_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_q5_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
 
-template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64>;
-template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80>;
-template [[host_name("kernel_flash_attn_ext_q8_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96>;
-template [[host_name("kernel_flash_attn_ext_q8_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112>;
-template [[host_name("kernel_flash_attn_ext_q8_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 128>;
-template [[host_name("kernel_flash_attn_ext_q8_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
 
 #undef FA_TYPES
 
 template<
-    typename q4_t,    // query types in shared memory
-    typename q4x4_t,
-    typename k4x4_t,  // key types in shared memory
-    typename v4x4_t,  // value types in shared memory
-    typename qk_t,    // Q*K types
-    typename s_t,     // soft-max types
+    typename q4_t,  // query types in shared memory
+    typename k4_t,  // key types in shared memory
+    typename v4_t,  // value types in shared memory
+    typename qk_t,  // Q*K types
+    typename s_t,   // soft-max types
     typename s4_t,
-    typename s4x4_t,
-    typename o4x4_t,  // attention accumulation types
-    typename kd4x4_t, // key type in device memory
+    typename o4_t,  // attention accumulation types
+    typename kd4_t, // key type in device memory
     short nl_k,
-    void (*deq_k)(device const kd4x4_t *, short, thread k4x4_t &),
-    typename vd4x4_t, // key type in device memory
+    void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &),
+    typename vd4_t, // key type in device memory
     short nl_v,
-    void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
-    short D,         // head size
-    short Q  = 1,    // queries per threadgroup
-    short C  = 32>   // cache items per threadgroup
+    void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
+    short DK,       // K head size
+    short DV,       // V head size
+    short NE = 4,   // head elements per thread
+    short Q  = 1,   // queries per threadgroup
+    short C  = 32>  // cache items per threadgroup
 kernel void kernel_flash_attn_ext_vec(
         constant ggml_metal_kargs_flash_attn_ext & args,
         device const char * q,
@@ -3617,29 +3642,28 @@ kernel void kernel_flash_attn_ext_vec(
     const int iq2 = tgpig[1];
     const int iq1 = tgpig[0];
 
-    const short D4  = D/4;
-    const short D16 = D/16;
-    const short NW  = N_SIMDWIDTH;
-    const short NL  = NW/4; // note: this can be adjusted to support D%64 == 0 and D%32 == 0
-    const short SH  = 2*C;  // shared memory per simdgroup
+    constexpr short DK4 = DK/4;
+    constexpr short DV4 = DV/4;
+    constexpr short NW  = N_SIMDWIDTH;
+    constexpr short NL  = NW/NE; // note: this can be adjusted to support different head sizes and simdgroup work loads
+    constexpr short SH  = 4*C;   // shared memory per simdgroup
 
-    const short T = D + nsg*SH; // shared memory size per query in (half)
+    const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t    * sq    = (threadgroup q_t    *) (shmem_f16 +                0*D); // holds the query data
-    threadgroup q4_t   * sq4   = (threadgroup q4_t   *) (shmem_f16 +                0*D); // same as above but in q4_t
-    threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shmem_f16 +                0*D); // same as above but in q4x4_t
-    threadgroup s_t    * ss    = (threadgroup s_t    *) (shmem_f16 + sgitg*SH     + Q*D); // scratch buffer for attention
-    threadgroup s4_t   * ss4   = (threadgroup s4_t   *) (shmem_f16 + sgitg*SH     + Q*D); // same as above but in s4_t
-    threadgroup half   * sm    = (threadgroup half   *) (shmem_f16 + sgitg*SH + C + Q*D); // scratch buffer for mask
-    threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shmem_f16 + sgitg*D      + Q*T); // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                  0*DK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                  0*DK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 + sgitg*SH       + Q*DK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 + sgitg*SH       + Q*DK); // same as above but in s4_t
+    threadgroup float * sm  = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
+    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + sgitg*DV       + Q*T);  // scratch buffer for the results
 
-    // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
-    o4x4_t lo[D16/NL];
+    // store the result for all queries in local memory (the O matrix from the paper)
+    o4_t lo[DV4/NL];
 
     // load heads from Q to shared memory
     device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
-    for (short i = tiisg; i < D4; i += NW) {
+    for (short i = tiisg; i < DK4; i += NW) {
         if (iq1 < args.ne01) {
             sq4[i] = (q4_t) q4[i];
         } else {
@@ -3648,8 +3672,8 @@ kernel void kernel_flash_attn_ext_vec(
     }
 
     // zero out lo
-    for (short i = 0; i < D16/NL; ++i) {
-        lo[i] = (o4x4_t) 0.0f;
+    for (short i = 0; i < DV4/NL; ++i) {
+        lo[i] = (o4_t) 0.0f;
     }
 
     // zero out shared memory SH
@@ -3660,8 +3684,8 @@ kernel void kernel_flash_attn_ext_vec(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     {
-        half S = 0.0f;
-        half M = -__FLT16_MAX__/2;
+        float S = 0.0f;
+        float M = -__FLT16_MAX__/2;
 
         // thread indices inside the simdgroup
         const short tx = tiisg%NL;
@@ -3674,26 +3698,18 @@ kernel void kernel_flash_attn_ext_vec(
         const short ikv2 = iq2/(args.ne02/args.ne_12_2);
         const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
-        // load the queries from shared memory into local memory
-        q4x4_t mq[D16/NL];
-
-        #pragma unroll(D16/NL)
-        for (short ii = 0; ii < D16; ii += NL) {
-            mq[ii/NL] = sq4x4[ii + tx];
-        }
-
         const bool has_mask = mask != q;
 
         // pointer to the mask
         device const half * pm = (device const half *) (mask + iq1*args.nb31);
 
-        half slope = 1.0f;
+        float slope = 1.0f;
 
         // ALiBi
         if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const float base = h < args.n_head_log2 ? args.m0 : args.m1;
             const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
@@ -3713,43 +3729,56 @@ kernel void kernel_flash_attn_ext_vec(
 
             // Q*K^T
             {
-                // each simdgroup processes 1 query and 4 (NW/NL) keys
-                for (short cc = 0; cc < C/4; ++cc) {
-                    qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 };
+                // each simdgroup processes 1 query and NE (NW/NL) head elements
+                for (short cc = 0; cc < C/NE; ++cc) {
+                    qk_t mqk = 0.0f;
 
-                    device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                    device const kd4_t * pk = (device const kd4_t *) ((device const char *) k + ((ic + NE*cc + ty)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13));
 
-                    #pragma unroll(D16/NL)
-                    for (short ii = 0; ii < D16; ii += NL) {
+                    #pragma unroll(DK4/NL)
+                    for (short ii = 0; ii < DK4; ii += NL) {
                         const short i = ii + tx;
 
-                        k4x4_t mk;
-                        deq_k(pk + i/nl_k, i%nl_k, mk);
+                        k4_t mk;
+                        deq_k_t4(pk + i/nl_k, i%nl_k, mk);
 
                         // note: this is less precise than the version below
-                        //mqka[0] += dot(mq[ii/NL][0], mk[0]);
-                        //mqka[1] += dot(mq[ii/NL][1], mk[1]);
-                        //mqka[2] += dot(mq[ii/NL][2], mk[2]);
-                        //mqka[3] += dot(mq[ii/NL][3], mk[3]);
+                        //mqka[0] += dot(mq[0], mk[0]);
+                        //mqka[1] += dot(mq[1], mk[1]);
+                        //mqka[2] += dot(mq[2], mk[2]);
+                        //mqka[3] += dot(mq[3], mk[3]);
 
-                        mqka[0] += dot((float4) mq[ii/NL][0], (float4) mk[0]);
-                        mqka[1] += dot((float4) mq[ii/NL][1], (float4) mk[1]);
-                        mqka[2] += dot((float4) mq[ii/NL][2], (float4) mk[2]);
-                        mqka[3] += dot((float4) mq[ii/NL][3], (float4) mk[3]);
+                        //q4x4_t mq = sq4x4[i];
+                        //mqka[0] += dot((float4) mq[0], (float4) mk[0]);
+                        //mqka[1] += dot((float4) mq[1], (float4) mk[1]);
+                        //mqka[2] += dot((float4) mq[2], (float4) mk[2]);
+                        //mqka[3] += dot((float4) mq[3], (float4) mk[3]);
+
+                        mqk += dot((float4) mk, (float4) sq4[i]);
                     }
 
-                    qk_t mqk = mqka[0] + mqka[1] + mqka[2] + mqka[3];
+                    static_assert(NE > 1, "NE must be > 1"); // note: not sure why NE == 1 fails
 
-                    // simdgroup reduce
+                    // simdgroup reduce (NE = 4)
                     // [ 0 ..  7] -> [ 0]
                     // [ 8 .. 15] -> [ 8]
                     // [16 .. 23] -> [16]
                     // [24 .. 31] -> [24]
-                  //mqk += simd_shuffle_down(mqk, 16);
-                  //mqk += simd_shuffle_down(mqk,  8);
-                    mqk += simd_shuffle_down(mqk,  4);
-                    mqk += simd_shuffle_down(mqk,  2);
-                    mqk += simd_shuffle_down(mqk,  1);
+                    if (NE <= 1) {
+                        mqk += simd_shuffle_down(mqk, 16);
+                    }
+                    if (NE <= 2) {
+                        mqk += simd_shuffle_down(mqk,  8);
+                    }
+                    if (NE <= 4) {
+                        mqk += simd_shuffle_down(mqk,  4);
+                    }
+                    if (NE <= 8) {
+                        mqk += simd_shuffle_down(mqk,  2);
+                    }
+                    if (NE <= 16) {
+                        mqk += simd_shuffle_down(mqk,  1);
+                    }
 
                     // mqk = mqk*scale + mask*slope
                     if (tx == 0) {
@@ -3759,9 +3788,9 @@ kernel void kernel_flash_attn_ext_vec(
                             mqk = args.logit_softcap*precise::tanh(mqk);
                         }
 
-                        mqk += sm[4*cc + ty]*slope;
+                        mqk += sm[NE*cc + ty]*slope;
 
-                        ss[4*cc + ty] = mqk;
+                        ss[NE*cc + ty] = mqk;
                     }
                 }
             }
@@ -3770,13 +3799,13 @@ kernel void kernel_flash_attn_ext_vec(
 
             // online softmax
             {
-                const half m = M;
-                const half s = ss[tiisg];
+                const float m = M;
+                const float s = ss[tiisg];
 
                 M = simd_max(max(M, s));
 
-                const half ms = exp(m - M);
-                const half vs = exp(s - M);
+                const float ms = exp(m - M);
+                const float vs = exp(s - M);
 
                 S = S*ms + simd_sum(vs);
 
@@ -3784,8 +3813,8 @@ kernel void kernel_flash_attn_ext_vec(
                 ss[tiisg] = vs;
 
                 // O = diag(ms)*O
-                #pragma unroll(D16/NL)
-                for (short ii = 0; ii < D16; ii += NL) {
+                #pragma unroll(DV4/NL)
+                for (short ii = 0; ii < DV4; ii += NL) {
                     lo[ii/NL] *= ms;
                 }
             }
@@ -3794,19 +3823,20 @@ kernel void kernel_flash_attn_ext_vec(
 
             // O = O + (Q*K^T)*V
             {
-                for (short cc = 0; cc < C/4; ++cc) {
-                    device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
+                //#pragma unroll(C/NE)
+                for (short cc = 0; cc < C/NE; ++cc) {
+                    device const vd4_t * pv4 = (device const vd4_t *) ((device const char *) v + ((ic + NE*cc + ty)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23));
 
-                    const s4x4_t ms(ss[4*cc + ty]);
+                    const s4_t ms(ss[NE*cc + ty]);
 
-                    #pragma unroll(D16/NL)
-                    for (short ii = 0; ii < D16; ii += NL) {
+                    #pragma unroll(DV4/NL)
+                    for (short ii = 0; ii < DV4; ii += NL) {
                         const short i = ii + tx;
 
-                        v4x4_t mv;
-                        deq_v(pv4 + i/nl_v, i%nl_v, mv);
+                        v4_t mv;
+                        deq_v_t4(pv4 + i/nl_v, i%nl_v, mv);
 
-                        lo[ii/NL] += mv*ms;
+                        lo[ii/NL] += o4_t(float4(mv)*float4(ms));
                     }
                 }
             }
@@ -3819,7 +3849,7 @@ kernel void kernel_flash_attn_ext_vec(
         }
     }
 
-    // simdgroup reduce
+    // simdgroup reduce (NE = 4)
     // [ 0,  8, 16, 24] -> [ 0]
     // [ 1,  9, 17, 25] -> [ 1]
     // [ 2, 10, 18, 26] -> [ 2]
@@ -3828,37 +3858,48 @@ kernel void kernel_flash_attn_ext_vec(
     // [ 5, 13, 21, 29] -> [ 5]
     // [ 6, 14, 22, 30] -> [ 6]
     // [ 7, 15, 23, 31] -> [ 7]
-    for (short ii = 0; ii < D16; ii += NL) {
-        lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 16);
-        lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  8);
-      //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  4);
-      //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  2);
-      //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  1);
+    for (short ii = 0; ii < DV4; ii += NL) {
+        if (NE > 1) {
+            lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 16);
+            lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 16);
+            lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 16);
+            lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 16);
+        }
 
-        lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 16);
-        lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  8);
-      //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  4);
-      //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  2);
-      //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  1);
+        if (NE > 2) {
+            lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  8);
+            lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  8);
+            lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  8);
+            lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  8);
+        }
 
-        lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 16);
-        lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  8);
-      //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  4);
-      //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  2);
-      //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  1);
+        if (NE > 4) {
+            lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  4);
+            lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  4);
+            lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  4);
+            lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  4);
+        }
 
-        lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 16);
-        lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  8);
-      //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  4);
-      //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  2);
-      //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  1);
+        if (NE > 8) {
+            lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  2);
+            lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  2);
+            lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  2);
+            lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  2);
+        }
+
+        if (NE > 16) {
+            lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0],  1);
+            lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1],  1);
+            lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2],  1);
+            lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3],  1);
+        }
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     // store results to shared memory
-    for (short i = tiisg; i < D16; i += NL) {
-        sr4x4[i] = lo[i/NL];
+    for (short i = tiisg; i < DV4; i += NL) {
+        sr4[i] = lo[i/NL];
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
@@ -3866,18 +3907,18 @@ kernel void kernel_flash_attn_ext_vec(
     // parallel reduce
     for (short r = nsg/2; r > 0; r >>= 1) {
         if (sgitg < r) {
-            const half S0 = ss[       0];
-            const half S1 = ss[r*SH + 0];
+            const float S0 = ss[           0];
+            const float S1 = ss[r*(SH/2) + 0];
 
-            const half M0 = ss[       1];
-            const half M1 = ss[r*SH + 1];
+            const float M0 = ss[           1];
+            const float M1 = ss[r*(SH/2) + 1];
 
-            const half M = max(M0, M1);
+            const float M = max(M0, M1);
 
-            const half ms0 = exp(M0 - M);
-            const half ms1 = exp(M1 - M);
+            const float ms0 = exp(M0 - M);
+            const float ms1 = exp(M1 - M);
 
-            const half S = S0*ms0 + S1*ms1;
+            const float S = S0*ms0 + S1*ms1;
 
             if (tiisg == 0) {
                 ss[0] = S;
@@ -3885,22 +3926,22 @@ kernel void kernel_flash_attn_ext_vec(
             }
 
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
-            for (short i = tiisg; i < D16; i += NW) {
-                sr4x4[i] = sr4x4[i]*ms0 + sr4x4[i + r*D16]*ms1;
+            for (short i = tiisg; i < DV4; i += NW) {
+                sr4[i] = sr4[i]*ms0 + sr4[i + r*DV4]*ms1;
             }
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    device float4x4 * dst44 = (device float4x4 *) dst;
+    device float4 * dst4 = (device float4 *) dst;
 
     // final rescale with 1/S and store to global memory
     if (sgitg == 0) {
         const float S = ss[0];
 
-        for (short i = tiisg; i < D16; i += NW) {
-            dst44[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*D16 + i] = (float4x4) sr4x4[i]/S;
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*DV4 + i] = (float4) sr4[i]/S;
         }
     }
 }
@@ -3909,34 +3950,54 @@ kernel void kernel_flash_attn_ext_vec(
 //       in the other (non-vec) kernel, we need s_t to also be float because we scale during the soft_max
 //
 #define FA_TYPES \
-           half4,  half4x4, \
-                   half4x4, \
-                   half4x4, \
-    float,                  \
-    half,  half4,  half4x4, \
-                   half4x4
+           half4,  \
+           half4,  \
+           half4,  \
+    float,         \
+    float, float4, \
+           half4
 
-typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 128>) flash_attn_ext_vec_t;
+typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
-template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,     1, dequantize_f16,  128>;
+template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  128, 128, 4>;
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,   1, dequantize_bf16, 128>;
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 128, 128, 4>;
 #endif
-template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0,  2, dequantize_q4_0, 128>;
-template [[host_name("kernel_flash_attn_ext_vec_q4_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1,  2, dequantize_q4_1, 128>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0,  2, dequantize_q5_0, 128>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1,  2, dequantize_q5_1, 128>;
-template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0,  2, dequantize_q8_0, 128>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 128, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 128, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 128, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 128, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 128, 128, 4>;
 
-template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,     1, dequantize_f16,  256>;
+template [[host_name("kernel_flash_attn_ext_vec_f16_h192")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  192, 192, 4>;
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,   1, dequantize_bf16, 256>;
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 192, 192, 4>;
 #endif
-template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0,  2, dequantize_q4_0, 256>;
-template [[host_name("kernel_flash_attn_ext_vec_q4_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1,  2, dequantize_q4_1, 256>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0,  2, dequantize_q5_0, 256>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1,  2, dequantize_q5_1, 256>;
-template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0,  2, dequantize_q8_0, 256>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 192, 192, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 192, 192, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 192, 192, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 192, 192, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 192, 192, 4>;
+
+template [[host_name("kernel_flash_attn_ext_vec_f16_hk192_hv128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  192, 128, 4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 192, 128, 4>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 192, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 192, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 192, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 192, 128, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 192, 128, 4>;
+
+template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  256, 256, 4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 256, 256, 4>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 256, 256, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 256, 256, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 256, 256, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 256, 256, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 256, 256, 4>;
 
 #undef FA_TYPES
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -63,6 +63,7 @@ set(GGML_OPENCL_KERNELS
     ggml-opencl_transpose_16
     ggml-opencl_transpose_32
     ggml-opencl_transpose_32_16
+    ggml-opencl_im2col
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -224,12 +224,14 @@ struct ggml_backend_opencl_context {
     cl_program program;
     cl_program program_1;
     cl_program program_2;
+    cl_program program_im2col;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
     cl_kernel kernel_scale;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_clamp;
     cl_kernel kernel_norm;
@@ -239,6 +241,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
+    cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
     cl_kernel kernel_mul_mat_f32_f32;
     cl_kernel kernel_mul_mat_f16_f16;
@@ -252,6 +255,7 @@ struct ggml_backend_opencl_context {
               kernel_mul_mat_q4_0_f32_flat_img_v0;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -708,6 +712,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->kernel_silu_4             = clCreateKernel(backend_ctx->program, "kernel_silu_4", &err), err));
     CL_CHECK((backend_ctx->kernel_gelu               = clCreateKernel(backend_ctx->program, "kernel_gelu", &err), err));
     CL_CHECK((backend_ctx->kernel_gelu_4             = clCreateKernel(backend_ctx->program, "kernel_gelu_4", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick         = clCreateKernel(backend_ctx->program, "kernel_gelu_quick", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick_4       = clCreateKernel(backend_ctx->program, "kernel_gelu_quick_4", &err), err));
     CL_CHECK((backend_ctx->kernel_relu               = clCreateKernel(backend_ctx->program, "kernel_relu", &err), err));
     CL_CHECK((backend_ctx->kernel_clamp              = clCreateKernel(backend_ctx->program, "kernel_clamp", &err), err));
     CL_CHECK((backend_ctx->kernel_norm               = clCreateKernel(backend_ctx->program, "kernel_norm", &err), err));
@@ -722,6 +728,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->kernel_rope_norm_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_neox_f32      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f32", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_neox_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f32     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f16     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f32    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f16    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f16_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f16_f32        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f32", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f32_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f32_f16", &err), err));
@@ -769,6 +779,19 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
 
     CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle     = clCreateKernel(backend_ctx->program_2, "kernel_convert_block_q4_0_noshuffle", &err), err));
 
+    // im2col kernels
+#ifdef GGML_OPENCL_EMBED_KERNELS
+    const std::string kernel_src_im2col {
+        #include "ggml-opencl_im2col.cl.h"
+    };
+#else
+    const std::string kernel_src_im2col = read_file("ggml-opencl_im2col.cl");
+#endif
+    backend_ctx->program_im2col = build_program_from_source(context, device, kernel_src_im2col.c_str(), compile_opts);
+
+    CL_CHECK((backend_ctx->kernel_im2col_f32 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_im2col_f16 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f16", &err), err));
+
     // Kernels for Adreno
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -898,10 +921,30 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     backend_ctx->program_CL_gemm = build_program_from_source(context, device, kernel_src_CL_gemm.c_str(), compile_opts);
     CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err));
 
+    // TODO: fixme: these sizes are hardcoded for now.
+    //  they should be allocated based on the model's size
+    //  and the device's max alloc size
     // Allocate intermediate buffers and images
-    size_t max_A_q_d_bytes = 311164928;
-    size_t max_A_s_d_bytes = 38895616;
-    size_t max_B_d_bytes = 45088768;
+    size_t required_A_q_d_bytes = 311164928;
+    size_t required_A_s_d_bytes = 38895616;
+    size_t required_B_d_bytes = 45088768;
+
+    // Ensure buffer sizes do not exceed the maximum allocation size
+    size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, backend_ctx->max_alloc_size);
+    size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, backend_ctx->max_alloc_size);
+    size_t max_B_d_bytes   = MIN(required_B_d_bytes, backend_ctx->max_alloc_size);
+    if (required_A_q_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: A_q_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_A_q_d_bytes, max_A_q_d_bytes);
+    }
+    if (required_A_s_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: A_s_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_A_s_d_bytes, max_A_s_d_bytes);
+    }
+    if (required_B_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: B_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_B_d_bytes, max_B_d_bytes);
+    }
 
     CL_CHECK((backend_ctx->A_q_d_max = clCreateBuffer(context, 0, max_A_q_d_bytes, NULL, &err), err));
     CL_CHECK((backend_ctx->A_s_d_max = clCreateBuffer(context, 0, max_A_s_d_bytes, NULL, &err), err));
@@ -1187,6 +1230,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_QUICK:
                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -1216,14 +1260,26 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->ne[3] == 1;
         case GGML_OP_ROPE: {
             const int mode = ((const int32_t *) op->op_params)[2];
-            if (mode & GGML_ROPE_TYPE_MROPE) {
+            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+            if (is_mrope && !is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                 return false;
             }
-            if (mode & GGML_ROPE_TYPE_VISION) {
+            if (is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                 return false;
             }
             return true;
         }
+        case GGML_OP_IM2COL:
+            return true;
         default:
             return false;
     }
@@ -2582,6 +2638,53 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }
 
+static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_quick_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_quick;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt);
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
+#endif
+}
+
 static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -3980,6 +4083,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     float attn_factor;
     float beta_fast;
     float beta_slow;
+    int32_t sections[4];
 
     memcpy(&freq_base,   (int32_t *) dst->op_params + 5, sizeof(float));
     memcpy(&freq_scale,  (int32_t *) dst->op_params + 6, sizeof(float));
@@ -3987,23 +4091,23 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int32_t)*4);
 
     const bool is_neox = mode & 2;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
 
     cl_kernel kernel;
 
-    if (!is_neox) {
-        switch (src0->type) {
-            case GGML_TYPE_F32:
-                kernel = backend_ctx->kernel_rope_norm_f32;
-                break;
-            case GGML_TYPE_F16:
-                kernel = backend_ctx->kernel_rope_norm_f16;
-                break;
-            default:
-                GGML_ASSERT(false);
-        };
-    } else {
+    if (is_neox) {
         switch (src0->type) {
             case GGML_TYPE_F32:
                 kernel = backend_ctx->kernel_rope_neox_f32;
@@ -4014,6 +4118,39 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
             default:
                 GGML_ASSERT(false);
         };
+    } else if (is_mrope && !is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_multi_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_multi_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    } else if (is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_vision_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_vision_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    } else {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_norm_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_norm_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
     }
 
     CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
@@ -4049,6 +4186,9 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
     CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
     CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    if (is_mrope || is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
+    }
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -4064,6 +4204,98 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }
 
+static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // src0 - filter, src1 - input
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const cl_long IC = src1->ne[is_2D ? 2 : 1];
+    const cl_long IH = is_2D ? src1->ne[1] : 1;
+    const cl_long IW =         src1->ne[0];
+
+    const cl_long KH = is_2D ? src0->ne[1] : 1;
+    const cl_long KW =         src0->ne[0];
+
+    const cl_long OH = is_2D ? dst->ne[2] : 1;
+    const cl_long OW =         dst->ne[1];
+
+    // nb is byte offset, src is type float32
+    const cl_ulong delta_offset = src1->nb[is_2D ? 2 : 1]/4;
+    const cl_long  batch        = src1->ne[is_2D ? 3 : 2];
+    const cl_ulong batch_offset = src1->nb[is_2D ? 3 : 2]/4;
+
+    const cl_long pelements = OW*KW*KH;
+    const cl_long CHW       = IC*KH*KW;
+
+    cl_kernel kernel;
+
+    if(dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_im2col_f16;
+    } else {
+        kernel = backend_ctx->kernel_im2col_f32;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(cl_ulong), &batch_offset));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(cl_ulong), &delta_offset));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(cl_long),  &IW));
+    CL_CHECK(clSetKernelArg(kernel,   7, sizeof(cl_long),  &IH));
+    CL_CHECK(clSetKernelArg(kernel,   8, sizeof(cl_long),  &IC));
+    CL_CHECK(clSetKernelArg(kernel,   9, sizeof(cl_long),  &OW));
+    CL_CHECK(clSetKernelArg(kernel,  10, sizeof(cl_long),  &OH));
+    CL_CHECK(clSetKernelArg(kernel,  11, sizeof(cl_long),  &KW));
+    CL_CHECK(clSetKernelArg(kernel,  12, sizeof(cl_long),  &KH));
+    CL_CHECK(clSetKernelArg(kernel,  13, sizeof(cl_long),  &pelements));
+    CL_CHECK(clSetKernelArg(kernel,  14, sizeof(cl_long),  &CHW));
+    CL_CHECK(clSetKernelArg(kernel,  15, sizeof(int),      &s0));
+    CL_CHECK(clSetKernelArg(kernel,  16, sizeof(int),      &s1));
+    CL_CHECK(clSetKernelArg(kernel,  17, sizeof(int),      &p0));
+    CL_CHECK(clSetKernelArg(kernel,  18, sizeof(int),      &p1));
+    CL_CHECK(clSetKernelArg(kernel,  19, sizeof(int),      &d0));
+    CL_CHECK(clSetKernelArg(kernel,  20, sizeof(int),      &d1));
+
+    const int num_blocks = (pelements + 256 - 1) / 256;
+    size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC};
+    size_t local_work_size[] = {256, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 //------------------------------------------------------------------------------
 // Op offloading
 //------------------------------------------------------------------------------
@@ -4122,6 +4354,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_gelu;
                     break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_quick;
+                    break;
                 case GGML_UNARY_OP_SILU:
                     if (!any_on_device) {
                         return false;
@@ -4194,6 +4432,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_rope;
             break;
+        case GGML_OP_IM2COL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_im2col;
+            break;
         default:
             return false;
     }

ggml/src/ggml-opencl/kernels/ggml-opencl.cl

@@ -404,6 +404,7 @@ kernel void kernel_scale(
 // gelu
 //------------------------------------------------------------------------------
 #define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
 #define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
 
 kernel void kernel_gelu(
@@ -434,6 +435,32 @@ kernel void kernel_gelu_4(
     dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
+kernel void kernel_gelu_quick(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
 //------------------------------------------------------------------------------
 // silu
 //------------------------------------------------------------------------------
@@ -1325,6 +1352,368 @@ kernel void kernel_rope_neox_f16(
     }
 }
 
+kernel void kernel_rope_multi_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_vision_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
+
+kernel void kernel_rope_vision_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
+
 //------------------------------------------------------------------------------
 // cpy
 //------------------------------------------------------------------------------

ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl

@@ -0,0 +1,146 @@
+#ifdef cl_khr_fp16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#elif defined(cl_amd_fp16)
+#pragma OPENCL EXTENSION cl_amd_fp16 : enable
+#else
+#error "Half precision floating point not supportedby OpenCL implementation on your device."
+#endif
+
+#ifdef cl_khr_subgroups
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#elif defined(cl_intel_subgroups)
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#error "Subgroup not supported on your device."
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+// Always use subgroup size of 32 on Intel.
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+// Always use subgroups size of 64 on Adreno.
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+// TODO: do not know how to choose subgroup size on other GPUs.
+#error "Selecting subgroup size is not supported on your device."
+#endif
+
+kernel void kernel_im2col_f32(
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    // threadIdx.x + blockIdx.x * blockDim.x
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}
+
+kernel void kernel_im2col_f16(
+        global float * src1,
+        ulong offset1,
+        global half  * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -26,6 +26,10 @@
 #  include <unistd.h>
 #endif
 #include <cstring>
+#include <fstream>
+#include <filesystem>
+
+namespace fs = std::filesystem;
 
 #ifdef _WIN32
 typedef SOCKET sockfd_t;
@@ -80,6 +84,7 @@ enum rpc_cmd {
     RPC_CMD_FREE_BUFFER,
     RPC_CMD_BUFFER_CLEAR,
     RPC_CMD_SET_TENSOR,
+    RPC_CMD_SET_TENSOR_HASH,
     RPC_CMD_GET_TENSOR,
     RPC_CMD_COPY_TENSOR,
     RPC_CMD_GRAPH_COMPUTE,
@@ -89,6 +94,9 @@ enum rpc_cmd {
     RPC_CMD_COUNT,
 };
 
+// Try RPC_CMD_SET_TENSOR_HASH first when data size is larger than this threshold
+const size_t HASH_THRESHOLD = 10 * 1024 * 1024;
+
 struct rpc_msg_get_alloc_size_req {
     rpc_tensor tensor;
 };
@@ -135,6 +143,10 @@ struct rpc_msg_buffer_clear_req {
     uint8_t value;
 };
 
+struct rpc_msg_set_tensor_hash_rsp {
+    uint8_t result;
+};
+
 struct rpc_msg_get_tensor_req {
     rpc_tensor tensor;
     uint64_t offset;
@@ -187,6 +199,18 @@ struct ggml_backend_rpc_buffer_context {
 
 // RPC helper functions
 
+// Computes FNV-1a hash of the data
+static uint64_t fnv_hash(const uint8_t * data, size_t len) {
+    const uint64_t fnv_prime = 0x100000001b3ULL;
+    uint64_t hash = 0xcbf29ce484222325ULL;
+
+    for (size_t i = 0; i < len; ++i) {
+        hash ^= data[i];
+        hash *= fnv_prime;
+    }
+    return hash;
+}
+
 static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
 #ifdef _WIN32
     if (fd == INVALID_SOCKET) {
@@ -483,10 +507,26 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
 
 static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
+    rpc_tensor rpc_tensor = serialize_tensor(tensor);
+    if (size > HASH_THRESHOLD) {
+        // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes)
+        size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t);
+        std::vector<uint8_t> input(input_size, 0);
+        uint64_t hash = fnv_hash((const uint8_t*)data, size);
+        memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
+        memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
+        memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash));
+        rpc_msg_set_tensor_hash_rsp response;
+        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response));
+        GGML_ASSERT(status);
+        if (response.result) {
+            // the server has the same data, no need to send it
+            return;
+        }
+    }
+    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes)
     size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
     std::vector<uint8_t> input(input_size, 0);
-    rpc_tensor rpc_tensor = serialize_tensor(tensor);
     memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
@@ -772,7 +812,9 @@ void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, si
 
 class rpc_server {
 public:
-    rpc_server(ggml_backend_t backend) : backend(backend) {}
+    rpc_server(ggml_backend_t backend, const char * cache_dir)
+        : backend(backend), cache_dir(cache_dir) {
+    }
     ~rpc_server();
 
     void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response);
@@ -782,6 +824,7 @@ public:
     bool free_buffer(const rpc_msg_free_buffer_req & request);
     bool buffer_clear(const rpc_msg_buffer_clear_req & request);
     bool set_tensor(const std::vector<uint8_t> & input);
+    bool set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response);
     bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
     bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
     bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
@@ -789,6 +832,7 @@ public:
     bool get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_msg_get_alloc_size_rsp & response);
 
 private:
+    bool get_cached_file(uint64_t hash, std::vector<uint8_t> & data);
     ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
     ggml_tensor * create_node(uint64_t id,
                               struct ggml_context * ctx,
@@ -797,6 +841,7 @@ private:
 
 
     ggml_backend_t backend;
+    const char * cache_dir;
     std::unordered_set<ggml_backend_buffer_t> buffers;
 };
 
@@ -960,11 +1005,85 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
     }
 
     const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
+    if (cache_dir && size > HASH_THRESHOLD) {
+        uint64_t hash = fnv_hash((const uint8_t*)data, size);
+        char hash_str[17];
+        snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash);
+        // save to cache_dir/hash_str
+        fs::path cache_file = fs::path(cache_dir) / hash_str;
+        std::ofstream ofs(cache_file, std::ios::binary);
+        ofs.write((const char *)data, size);
+        printf("[%s] saved to '%s'\n", __func__, cache_file.c_str());
+    }
     ggml_backend_tensor_set(tensor, data, offset, size);
     ggml_free(ctx);
     return true;
 }
 
+bool rpc_server::get_cached_file(uint64_t hash, std::vector<uint8_t> & data) {
+    if (!cache_dir) {
+        return false;
+    }
+    char hash_str[17];
+    snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash);
+    fs::path cache_file = fs::path(cache_dir) / hash_str;
+    if (!fs::exists(cache_file)) {
+        return false;
+    }
+    std::ifstream ifs(cache_file, std::ios::binary);
+    ifs.seekg(0, std::ios::end);
+    size_t size = ifs.tellg();
+    ifs.seekg(0, std::ios::beg);
+    data.resize(size);
+    ifs.read((char *)data.data(), size);
+    return true;
+}
+
+bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response)
+{
+    // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) |
+    if (input.size() != sizeof(rpc_tensor) + 16) {
+        return false;
+    }
+    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
+    uint64_t offset;
+    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
+    const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset));
+    std::vector<uint8_t> cached_file;
+    if (!get_cached_file(*hash, cached_file)) {
+        response.result = 0;
+        return true;
+    }
+    size_t size = cached_file.size();
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
+    if (tensor == nullptr) {
+        GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
+        ggml_free(ctx);
+        return false;
+    }
+    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+    ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
+    response.result = 1;
+    ggml_free(ctx);
+    return true;
+}
+
 bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
     struct ggml_init_params params {
         /*.mem_size   =*/ ggml_tensor_overhead(),
@@ -1148,8 +1267,9 @@ rpc_server::~rpc_server() {
     }
 }
 
-static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
-    rpc_server server(backend);
+static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
+                             sockfd_t sockfd, size_t free_mem, size_t total_mem) {
+    rpc_server server(backend, cache_dir);
     while (true) {
         uint8_t cmd;
         if (!recv_data(sockfd, &cmd, 1)) {
@@ -1260,6 +1380,20 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
                 }
                 break;
             }
+            case RPC_CMD_SET_TENSOR_HASH: {
+                std::vector<uint8_t> input;
+                if (!recv_msg(sockfd, input)) {
+                    return;
+                }
+                rpc_msg_set_tensor_hash_rsp response;
+                if (!server.set_tensor_hash(input, response)) {
+                    return;
+                }
+                if (!send_msg(sockfd, &response, sizeof(response))) {
+                    return;
+                }
+                break;
+            }
             case RPC_CMD_INIT_TENSOR: {
                 rpc_msg_init_tensor_req request;
                 if (!recv_msg(sockfd, &request,sizeof(request))) {
@@ -1335,7 +1469,9 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
     }
 }
 
-void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
+void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                   const char * cache_dir,
+                                   size_t free_mem, size_t total_mem) {
     std::string host;
     int port;
     if (!parse_endpoint(endpoint, host, port)) {
@@ -1364,7 +1500,7 @@ void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint
         }
         printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
         fflush(stdout);
-        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
+        rpc_serve_client(backend, cache_dir, client_socket->fd, free_mem, total_mem);
         printf("Client connection closed\n");
         fflush(stdout);
     }

ggml/src/ggml-sycl/CMakeLists.txt

@@ -23,6 +23,23 @@ ggml_add_backend_library(ggml-sycl
                          ../../include/ggml-sycl.h
                         )
 
+file(GLOB   GGML_HEADERS_SYCL "*.hpp")
+file(GLOB   GGML_SOURCES_SYCL "*.cpp")
+target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
+
+find_package(IntelSYCL)
+if (IntelSYCL_FOUND)
+    # Use oneAPI CMake when possible
+    target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX)
+else()
+    # Fallback to the simplest way of enabling SYCL when using intel/llvm nightly for instance
+    target_compile_options(ggml-sycl PRIVATE "-fsycl")
+    target_link_options(ggml-sycl PRIVATE "-fsycl")
+endif()
+
+target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
+
+# Link against oneDNN
 find_package(DNNL)
 set(GGML_SYCL_DNNL 0)
 if(DNNL_FOUND)
@@ -62,8 +79,6 @@ if (GGML_SYCL_F16)
     add_compile_definitions(GGML_SYCL_F16)
 endif()
 
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-
 if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
     add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
 elseif (GGML_SYCL_TARGET STREQUAL "AMD")
@@ -76,34 +91,84 @@ else()
     add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
 endif()
 
-file(GLOB   GGML_HEADERS_SYCL "*.hpp")
-file(GLOB   GGML_SOURCES_SYCL "*.cpp")
-target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
+if (GGML_SYCL_GRAPH)
+    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
+endif()
 
-
-if (WIN32)
-    find_package(IntelSYCL REQUIRED)
+# Link against Intel oneMKL or oneMath
+if (GGML_SYCL_TARGET STREQUAL "INTEL")
+    # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically
+    # See https://github.com/uxlfoundation/oneMath/issues/654
     find_package(MKL REQUIRED)
-    target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+    target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
+    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL)
 else()
-    if (GGML_SYCL_GRAPH)
-        add_compile_definitions(GGML_SYCL_GRAPH)
+    find_package(oneMath QUIET)
+    if (NOT oneMath_FOUND)
+        message(STATUS "oneMath not found: oneMath will be automatically downloaded")
+        # Use FetchContent to automatically pull and build oneMath
+        include(FetchContent)
+        set(BUILD_FUNCTIONAL_TESTS False)
+        set(BUILD_EXAMPLES False)
+        set(TARGET_DOMAINS blas)
+        if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+            set(ENABLE_MKLCPU_BACKEND False)
+            set(ENABLE_MKLGPU_BACKEND False)
+            set(ENABLE_CUBLAS_BACKEND True)
+        elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+            set(ENABLE_MKLCPU_BACKEND False)
+            set(ENABLE_MKLGPU_BACKEND False)
+            set(ENABLE_ROCBLAS_BACKEND True)
+            # Ensure setting a string variable here is not overriden by oneMath CACHE variables
+            cmake_policy(SET CMP0126 NEW)
+            # Setting the device architecture is only needed and useful for AMD devices in oneMath
+            set(HIP_TARGETS ${GGML_SYCL_DEVICE_ARCH} CACHE STRING "oneMath HIP target" FORCE)
+        endif()
+        FetchContent_Declare(
+            ONEMATH
+            GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
+            GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
+        )
+        FetchContent_MakeAvailable(ONEMATH)
+        # Create alias to match with find_package targets name
+        function(onemath_alias target)
+            if (TARGET ${target}_obj)
+                # Silence verbose warnings from external libraries
+                target_compile_options(${target}_obj PRIVATE -w)
+            endif()
+            if (TARGET ${target})
+                add_library(ONEMATH::${target} ALIAS ${target})
+            endif()
+        endfunction()
+        onemath_alias(onemath)
+        onemath_alias(onemath_blas_mklcpu)
+        onemath_alias(onemath_blas_mklgpu)
+        onemath_alias(onemath_blas_cublas)
+        onemath_alias(onemath_blas_rocblas)
     endif()
-    if (GGML_SYCL_TARGET STREQUAL "INTEL")
-        target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-    elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-        add_compile_definitions(GGML_SYCL_NVIDIA)
-        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas)
+
+    # Below oneMath compile-time dispatching is used for better performance
+    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_cublas)
+        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
+        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
+        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
     elseif (GGML_SYCL_TARGET STREQUAL "AMD")
         if (NOT GGML_SYCL_DEVICE_ARCH)
             message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
         endif()
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
-        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
+        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
+        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
+        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
+        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_AMD)
+    else()
+        # Fallback to oneMath runtime dispatcher
+        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath)
+        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GENERIC)
     endif()
-
-    if (GGML_SYCL_DEVICE_ARCH)
-      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
-  endif()
+endif()
+
+if (GGML_SYCL_DEVICE_ARCH)
+    target_compile_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
+    target_link_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
 endif()

ggml/src/ggml-sycl/common.cpp

@@ -66,41 +66,6 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
   return sycl_down_blk_size;
 }
 
-void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                 const ggml_tensor *src1, ggml_tensor *dst,
-                                 const ggml_sycl_op_flatten_t op) try {
-
-    const bool use_src1 = src1 != nullptr;
-    if(use_src1)
-      GGML_ASSERT(strcmp(src1->buffer->buft->iface.get_name(src1->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
-    GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
-
-    // dd = data device
-    float * src0_ddf = (float *) src0->data;
-    float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
-    float *  dst_ddf = (float *) dst->data;
-
-    ggml_sycl_pool_alloc<float> src0_f(ctx.pool());
-    ggml_sycl_pool_alloc<float> src1_f(ctx.pool());
-    ggml_sycl_pool_alloc<float>  dst_f(ctx.pool());
-
-    ggml_sycl_set_device(ctx.device);
-    queue_ptr main_stream = ctx.stream();
-    // GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n",
-        // ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device);
-
-    // do the computation
-    op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
-    // print_ggml_tensor("tensor", dst);
-}
-catch (sycl::exception const &exc) {
-
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-            << ", line:" << __LINE__ << std::endl;
-  std::exit(1);
-}
-
-
 void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams) {
     for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
         for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {

ggml/src/ggml-sycl/common.hpp

@@ -494,12 +494,6 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
 
 int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
 
-typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                       const ggml_tensor *src1,
-                                       ggml_tensor *dst, const float *src0_dd,
-                                       const float *src1_dd, float *dst_dd,
-                                       const queue_ptr &main_stream);
-
 template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
 static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
         int ne0, int ne1, int ne2, int ne3,
@@ -757,24 +751,22 @@ struct bin_bcast_sycl {
 
 template <class op>
 inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                   const ggml_tensor *src1, ggml_tensor *dst,
-                                   const float *src0_dd, const float *src1_dd,
-                                   float *dst_dd,
-                                   const queue_ptr &main_stream) {
+                                   const ggml_tensor *src1, ggml_tensor *dst) {
+    dpct::queue_ptr main_stream = ctx.stream();
 
     if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+        op()(ctx, src0, src1, dst, (const float *)src0->data, (const float *)src1->data, (float *)dst->data, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd,
-             (sycl::half *)dst_dd, main_stream);
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data,
+             (sycl::half *)dst->data, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd,
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data, (float *)dst->data,
              main_stream);
     } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
-        op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd,
+        op()(ctx, src0, src1, dst, (const int32_t *)src0->data, (const int32_t *)src1->data, (int32_t *)dst->data,
              main_stream);
     } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
-        op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd,
+        op()(ctx, src0, src1, dst, (const int16_t *)src0->data, (const int16_t *)src1->data, (int16_t *)dst->data,
              main_stream);
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
@@ -784,8 +776,4 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
 }
 
 bool gpu_has_xmx(sycl::device &dev);
-
-void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                 const ggml_tensor *src1, ggml_tensor *dst,
-                                 const ggml_sycl_op_flatten_t op);
 #endif // GGML_SYCL_COMMON_HPP