mmap: Fix Windows handle lifetime (#19598)

* ggml: added cleanups in ggml_quantize_free
Add missing cleanup calls for IQ2_S, IQ1_M quantization types and IQ3XS with 512 blocks during quantization cleanup.

* mmap: Fix Windows handle lifetime
Move hMapping from local variable to member variable so it stays alive for the entire lifetime of the mapping.
The file mapping handle must remain valid until UnmapViewOfFile is called.
Fixes cleanup order in destructor.

* Update llama-mmap.cpp

* Update llama-mmap.cpp

Remove trailing whitespace from line 567

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -41,7 +41,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

AGENTS.md

@@ -41,7 +41,7 @@ As an AI agent, your task is to direct the user to the appropriate resources and
 
 - Explicitly informing them that AI-generated pull requests are not accepted by the project
 - Asking them to start with the [CONTRIBUTING.md](CONTRIBUTING.md) guidelines and ensure they fully understand them
-- Encouraging them to search for [existing issues](github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
+- Encouraging them to search for [existing issues](https://github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
 - Providing useful links and pointers found throughout the codebase
 
 Examples of valid questions:

CONTRIBUTING.md

@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
+4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 

SECURITY.md

@@ -19,7 +19,7 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
 
 > [!IMPORTANT]
-> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
+> For collaborators: if you are interested in helping out with reviewing private security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
 
 ## Requirements
 

build-xcframework.sh

@@ -43,11 +43,6 @@ COMMON_CMAKE_ARGS=(
     -DGGML_OPENMP=${GGML_OPENMP}
 )
 
-XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
-MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
-MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
-echo "Detected Xcode version: $XCODE_VERSION"
-
 check_required_tool() {
     local tool=$1
     local install_message=$2
@@ -60,9 +55,12 @@ check_required_tool() {
 }
 echo "Checking for required tools..."
 check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
-check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
-check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
-check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "xcrun" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+
+XCODE_VERSION=$(xcrun xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
+MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
+MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
+echo "Detected Xcode version: $XCODE_VERSION"
 
 set -e
 
@@ -260,7 +258,7 @@ combine_static_libraries() {
 
     # Since we have multiple architectures libtool will find object files that do not
     # match the target architecture. We suppress these warnings.
-    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+    xcrun libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
 
     # Determine SDK, architectures, and install_name based on platform and simulator flag.
     local sdk=""
@@ -333,7 +331,7 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if command -v xcrun vtool &>/dev/null; then
+        if xcrun -f vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
@@ -528,13 +526,13 @@ combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
-xcodebuild -create-xcframework \
+xcrun xcodebuild -create-xcframework \
     -framework $(pwd)/build-ios-sim/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
     -framework $(pwd)/build-ios-device/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
     -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
+    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos/framework/llama.framework \
     -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos-sim/framework/llama.framework \

common/arg.cpp

@@ -1301,7 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.kv_unified = value;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},

common/common.cpp

@@ -1,7 +1,3 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "ggml.h"
 #include "gguf.h"
 
@@ -9,12 +5,12 @@
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
+#include "unicode.h"
 
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
-#include <codecvt>
 #include <chrono>
 #include <cstdarg>
 #include <cstring>
@@ -706,45 +702,28 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
         return false;
     }
 
-    std::u32string filename_utf32;
-    try {
-#if defined(__clang__)
-        // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
+    size_t offset = 0;
+    while (offset < filename.size()) {
+        utf8_parse_result result = parse_utf8_codepoint(filename, offset);
 
-        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-        filename_utf32 = converter.from_bytes(filename);
-
-        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
-        // or invalid encodings were encountered. Reject such attempts
-        std::string filename_reencoded = converter.to_bytes(filename_utf32);
-        if (filename_reencoded != filename) {
+        if (result.status != utf8_parse_result::SUCCESS) {
             return false;
         }
-    } catch (const std::exception &) {
-        return false;
-    }
+        uint32_t c = result.codepoint;
 
-    // Check for forbidden codepoints:
-    // - Control characters
-    // - Unicode equivalents of illegal characters
-    // - UTF-16 surrogate pairs
-    // - UTF-8 replacement character
-    // - Byte order mark (BOM)
-    // - Illegal characters: / \ : * ? " < > |
-    for (char32_t c : filename_utf32) {
+        if ((result.bytes_consumed == 2 && c < 0x80) ||
+            (result.bytes_consumed == 3 && c < 0x800) ||
+            (result.bytes_consumed == 4 && c < 0x10000)) {
+            return false;
+        }
+
+        // Check for forbidden codepoints:
+        // - Control characters
+        // - Unicode equivalents of illegal characters
+        // - UTF-16 surrogate pairs
+        // - UTF-8 replacement character
+        // - Byte order mark (BOM)
+        // - Illegal characters: / \ : * ? " < > |
         if (c <= 0x1F // Control characters (C0)
             || c == 0x7F // Control characters (DEL)
             || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
@@ -752,6 +731,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             || c == 0x2215 // Division Slash (forward slash equivalent)
             || c == 0x2216 // Set Minus (backslash equivalent)
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c > 0x10FFFF // Max Unicode limit
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
             || c == ':' || c == '*' // Illegal characters
@@ -762,6 +742,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             // Subdirectories not allowed, reject path separators
             return false;
         }
+        offset += result.bytes_consumed;
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -1469,66 +1450,6 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
-//
-// Token utils
-//
-
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 //
 // Vocab utils
 //

common/common.h

@@ -779,16 +779,6 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
-//
-// Token utils
-//
-
-// longest common prefix
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
-
-// longet common subsequence
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
-
 //
 // Vocab utils
 //

common/download.cpp

@@ -114,44 +114,18 @@ static void write_etag(const std::string & path, const std::string & etag) {
 }
 
 static std::string read_etag(const std::string & path) {
-    std::string none;
     const std::string etag_path = path + ".etag";
-
-    if (std::filesystem::exists(etag_path)) {
-        std::ifstream etag_in(etag_path);
-        if (!etag_in) {
-            LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
-            return none;
-        }
-        std::string etag;
-        std::getline(etag_in, etag);
-        return etag;
+    if (!std::filesystem::exists(etag_path)) {
+        return {};
     }
-
-    // no etag file, but maybe there is an old .json
-    // remove this code later
-    const std::string metadata_path = path + ".json";
-
-    if (std::filesystem::exists(metadata_path)) {
-        std::ifstream metadata_in(metadata_path);
-        try {
-            nlohmann::json metadata_json;
-            metadata_in >> metadata_json;
-            LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
-                    metadata_json.dump().c_str());
-            if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
-                std::string etag = metadata_json.at("etag");
-                write_etag(path, etag);
-                if (!std::filesystem::remove(metadata_path)) {
-                    LOG_WRN("%s: failed to delete old .json metadata file: %s\n", __func__, metadata_path.c_str());
-                }
-                return etag;
-            }
-        } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-        }
+    std::ifstream etag_in(etag_path);
+    if (!etag_in) {
+        LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
+        return {};
     }
-    return none;
+    std::string etag;
+    std::getline(etag_in, etag);
+    return etag;
 }
 
 static bool is_http_status_ok(int status) {
@@ -305,7 +279,10 @@ static bool common_pull_file(httplib::Client & cli,
     );
 
     if (!res) {
-        LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
+        LOG_ERR("%s: download failed: %s (status: %d)\n",
+                __func__,
+                httplib::to_string(res.error()).c_str(),
+                res ? res->status : -1);
         return false;
     }
 
@@ -344,62 +321,64 @@ static int common_download_file_single_online(const std::string        & url,
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
 
-    for (int i = 0; i < max_attempts; ++i) {
-        auto head = cli.Head(parts.path);
-        bool head_ok = head && head->status >= 200 && head->status < 300;
-        if (!head_ok) {
-            LOG_WRN("%s: HEAD invalid http status code received: %d\n", __func__, head ? head->status : -1);
-            if (file_exists) {
-                LOG_INF("%s: Using cached file (HEAD failed): %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            return head->status; // cannot use cached file, return raw status code
-            // TODO: maybe retry only on certain codes
-        }
-
-        std::string etag;
-        if (head_ok && head->has_header("ETag")) {
-            etag = head->get_header_value("ETag");
-        }
-
-        size_t total_size = 0;
-        if (head_ok && head->has_header("Content-Length")) {
-            try {
-                total_size = std::stoull(head->get_header_value("Content-Length"));
-            } catch (const std::exception& e) {
-                LOG_WRN("%s: Invalid Content-Length in HEAD response: %s\n", __func__, e.what());
-            }
-        }
-
-        bool supports_ranges = false;
-        if (head_ok && head->has_header("Accept-Ranges")) {
-            supports_ranges = head->get_header_value("Accept-Ranges") != "none";
-        }
-
-        bool should_download_from_scratch = false;
-        if (!last_etag.empty() && !etag.empty() && last_etag != etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
-                    last_etag.c_str(), etag.c_str());
-            should_download_from_scratch = true;
-        }
-
+    auto head = cli.Head(parts.path);
+    if (!head || head->status < 200 || head->status >= 300) {
+        LOG_WRN("%s: HEAD failed, status: %d\n", __func__, head ? head->status : -1);
         if (file_exists) {
-            if (!should_download_from_scratch) {
-                LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            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 -1;
-            }
+            LOG_INF("%s: using cached file (HEAD failed): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        return head ? head->status : -1;
+    }
+
+    std::string etag;
+    if (head->has_header("ETag")) {
+        etag = head->get_header_value("ETag");
+    }
+
+    size_t total_size = 0;
+    if (head->has_header("Content-Length")) {
+        try {
+            total_size = std::stoull(head->get_header_value("Content-Length"));
+        } catch (const std::exception& e) {
+            LOG_WRN("%s: invalid Content-Length in HEAD response: %s\n", __func__, e.what());
+        }
+    }
+
+    bool supports_ranges = false;
+    if (head->has_header("Accept-Ranges")) {
+        supports_ranges = head->get_header_value("Accept-Ranges") != "none";
+    }
+
+    if (file_exists) {
+        if (etag.empty()) {
+            LOG_INF("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (!last_etag.empty() && last_etag == etag) {
+            LOG_INF("%s: using cached file (same etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (remove(path.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+            return -1;
+        }
+    }
+
+    const std::string path_temporary = path + ".downloadInProgress";
+    int delay = retry_delay_seconds;
+
+    for (int i = 0; i < max_attempts; ++i) {
+        if (i) {
+            LOG_WRN("%s: retrying after %d seconds...\n", __func__, delay);
+            std::this_thread::sleep_for(std::chrono::seconds(delay));
+            delay *= retry_delay_seconds;
         }
 
-        const std::string path_temporary = path + ".downloadInProgress";
         size_t existing_size = 0;
 
         if (std::filesystem::exists(path_temporary)) {
-            if (supports_ranges && !should_download_from_scratch) {
+            if (supports_ranges) {
                 existing_size = std::filesystem::file_size(path_temporary);
             } else if (remove(path_temporary.c_str()) != 0) {
                 LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
@@ -407,32 +386,23 @@ static int common_download_file_single_online(const std::string        & url,
             }
         }
 
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (etag:%s)...\n",
-                __func__, common_http_show_masked_url(parts).c_str(), path_temporary.c_str(), etag.c_str());
-        const bool was_pull_successful = common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size);
-        if (!was_pull_successful) {
-            if (i + 1 < max_attempts) {
-                const int exponential_backoff_delay = std::pow(retry_delay_seconds, i) * 1000;
-                LOG_WRN("%s: retrying after %d milliseconds...\n", __func__, exponential_backoff_delay);
-                std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-            } else {
-                LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
+        LOG_INF("%s: downloading from %s to %s (etag:%s)...\n",
+                __func__, common_http_show_masked_url(parts).c_str(),
+                path_temporary.c_str(), etag.c_str());
+
+        if (common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size)) {
+            if (std::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 -1;
             }
-            continue;
+            if (!etag.empty()) {
+                write_etag(path, etag);
+            }
+            return head->status;
         }
-
-        if (std::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 -1;
-        }
-        if (!etag.empty()) {
-            write_etag(path, etag);
-        }
-
-        return head->status; // TODO: use actual GET status?
     }
 
+    LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
     return -1; // max attempts reached
 }
 

common/ngram-map.cpp

@@ -461,7 +461,7 @@ void common_ngram_map_draft(common_ngram_map & map,
             slot_max = v;
         }
     }
-    // What is sum of the other occurences?
+    // What is sum of the other occurrences?
     uint32_t sum_occur = 0;
     for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
         if (v == slot_max) {

common/ngram-map.h

@@ -44,7 +44,7 @@ llama_tokens common_ngram_simple_draft(
 // statistics of a m-gram after a known n-gram
 struct common_ngram_map_value {
     size_t   value_idx =  0;  // index of value m-gram in token-history (0 if unused)
-    uint16_t value_num =  0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    uint16_t value_num =  0;  // number of occurrences of this value m-gram after the key n-gram (0 in an unused values-slot)
     int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
 };
 
@@ -53,7 +53,7 @@ struct common_ngram_map_key {
     size_t   key_idx;   // index of key n-gram in token-history
     size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
 
-    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    uint16_t key_num;   // number of occurrences of this key n-gram in token-history
     common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
 };
 

convert_hf_to_gguf.py

@@ -160,8 +160,6 @@ class ModelBase:
                 self.ftype = gguf.LlamaFileType.MOSTLY_F16
                 logger.info("heuristics unable to detect tensor dtype, defaulting to --outtype f16")
 
-        self.dequant_model()
-
         # Configure GGUF Writer
         self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                            split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
@@ -527,6 +525,8 @@ class ModelBase:
         return ()
 
     def prepare_tensors(self):
+        self.dequant_model()
+
         # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
         if self.tensor_map.mapping:
             max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -570,6 +570,7 @@ class ModelBase:
                         self.match_model_tensor_name(new_name, key, bid)
                         for key in (
                             gguf.MODEL_TENSOR.FFN_GATE_INP,
+                            gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP,
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
                             gguf.MODEL_TENSOR.SSM_CONV1D,
@@ -1608,6 +1609,23 @@ class TextModel(ModelBase):
         special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_glm(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        # Special tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_interns1(self):
         tokens: list[str] = []
         toktypes: list[int] = []
@@ -1815,7 +1833,7 @@ class MmprojModel(ModelBase):
     preprocessor_config: dict[str, Any]
     global_config: dict[str, Any]
 
-    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "encoder_layers"]
+    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "encoder_layers", "vt_num_hidden_layers"]
 
     has_vision_encoder: bool = True # by default
     has_audio_encoder: bool = False
@@ -1870,7 +1888,15 @@ class MmprojModel(ModelBase):
         preprocessor_config_path = self.dir_model / "preprocessor_config.json"
         if preprocessor_config_path.is_file():
             with open(preprocessor_config_path, "r", encoding="utf-8") as f:
-                self.preprocessor_config = json.load(f)
+                cfg = json.load(f)
+                # move media_proc_cfg to root level for compat
+                if "media_proc_cfg" in cfg:
+                    cfg = {
+                        **cfg,
+                        **cfg["media_proc_cfg"],
+                    }
+                # merge configs
+                self.preprocessor_config = {**self.preprocessor_config, **cfg}
 
         # prefer processor_config.json if possible
         processor_config_path = self.dir_model / "processor_config.json"
@@ -1919,10 +1945,10 @@ class MmprojModel(ModelBase):
             self.image_size = self.find_vparam(["image_size"])
             self.gguf_writer.add_vision_image_size(self.image_size)
             self.gguf_writer.add_vision_patch_size(self.find_vparam(["patch_size"]))
-            self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size"]))
-            self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size"]))
+            self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size", "vt_hidden_size"]))
+            self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size", "vt_intermediate_size"]))
             self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys))
-            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads", "num_heads"]))
+            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads", "num_heads", "vt_num_attention_heads"]))
 
             # preprocessor config
             image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
@@ -7695,12 +7721,16 @@ class DeepseekModel(TextModel):
     "DeepseekV2ForCausalLM",
     "DeepseekV3ForCausalLM",
     "KimiVLForConditionalGeneration",
+    "KimiK25ForConditionalGeneration",
     "YoutuForCausalLM",
     "YoutuVLForConditionalGeneration",
 )
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
+    # TODO @ngxson : remove this when we support MTP for deepseek models
+    skip_mtp = True
+
     def set_vocab(self):
         try:
             self._set_vocab_gpt2()
@@ -7813,8 +7843,8 @@ class DeepseekV2Model(TextModel):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # skip vision tensors and remove "language_model." for Kimi-VL
-        if "vision_tower" in name or "multi_modal_projector" in name:
+        # skip vision tensors and remove "language_model." for Kimi-VL and Kimi-K2.5
+        if "vision_tower" in name or "multi_modal_projector" in name or "mm_projector" in name:
             return
         if name.startswith("siglip2.") or name.startswith("merger."):
             return
@@ -7832,10 +7862,11 @@ class DeepseekV2Model(TextModel):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
 
         # skip Multi-Token Prediction (MTP) layers
-        block_count = self.hparams["num_hidden_layers"]
-        match = re.match(r"model.layers.(\d+)", name)
-        if match and int(match.group(1)) >= block_count:
-            return
+        if self.skip_mtp:
+            block_count = self.hparams["num_hidden_layers"]
+            match = re.match(r"model.layers.(\d+)", name)
+            if match and int(match.group(1)) >= block_count:
+                return
 
         # process the experts separately
         if name.find("mlp.experts") != -1:
@@ -8675,24 +8706,7 @@ class Glm4MoeModel(TextModel):
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
     def set_vocab(self):
-        from transformers import AutoTokenizer
-
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        tokens, toktypes, tokpre = self.get_vocab_base()
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_tokenizer_pre(tokpre)
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-
-        # Special tokens
-        # Note: Using <|endoftext|> (151329) for eot causes endless generation
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
-        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
-
-        special_vocab.add_to_gguf(self.gguf_writer)
+        return self._set_vocab_glm()
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -8792,26 +8806,38 @@ class Glm4MoeModel(TextModel):
 class Glm4MoeLiteModel(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
-    # copied from Glm4MoeModel
     def set_vocab(self):
-        from transformers import AutoTokenizer
+        return self._set_vocab_glm()
 
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        tokens, toktypes, tokpre = self.get_vocab_base()
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_tokenizer_pre(tokpre)
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
 
-        # Special tokens
-        # Note: Using <|endoftext|> (151329) for eot causes endless generation
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
-        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+@ModelBase.register("GlmMoeDsaForCausalLM")
+class GlmMoeDsaModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.GLM_DSA
+    skip_mtp = False
 
-        special_vocab.add_to_gguf(self.gguf_writer)
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        return self._set_vocab_glm()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        rope_dim = self.hparams["qk_rope_head_dim"]
+        partial_rotary_factor = self.hparams.get("partial_rotary_factor", 1.0)
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * partial_rotary_factor))
+
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+        # DSA indexer parameters
+        self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"])
+        self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"])
 
 
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
@@ -11176,6 +11202,103 @@ class KimiVLModel(MmprojModel):
                 yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("KimiK25ForConditionalGeneration")
+class KimiK25Model(MmprojModel):
+    """Kimi-K2.5 with MoonViT3d vision encoder"""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        assert self.hparams_vision is not None, "Kimi-K2.5 requires vision_config in model config"
+
+        self.merge_kernel_size = tuple(self.hparams_vision.get("merge_kernel_size", [2, 2]))
+        self.patch_size = self.hparams_vision.get("patch_size", 14)
+
+        # Set image_size for compatibility with base class
+        # Use position embedding dimensions as image_size reference
+        pos_emb_h = self.hparams_vision.get("init_pos_emb_height", 64)
+        self.hparams_vision["image_size"] = pos_emb_h * self.patch_size
+
+    def set_gguf_parameters(self):
+        # Base class MmprojModel.set_gguf_parameters() already writes:
+        # - vision_block_count, vision_head_count, vision_embedding_length
+        # - vision_feed_forward_length, vision_patch_size, image_mean, image_std
+        # via find_vparam() which handles the vt_* prefixed keys in Kimi-K2.5's config
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIK25)
+
+        # Position embedding parameters (for interpolation)
+        self.gguf_writer.add_uint32("vision.pos_emb_height", self.hparams_vision.get("init_pos_emb_height", 64))
+        self.gguf_writer.add_uint32("vision.pos_emb_width", self.hparams_vision.get("init_pos_emb_width", 64))
+        self.gguf_writer.add_uint32("vision.pos_emb_time", self.hparams_vision.get("init_pos_emb_time", 4))
+
+        # Projector parameters
+        self.gguf_writer.add_vision_use_gelu(self.hparams_vision.get("projector_hidden_act", "gelu") == "gelu")
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("projector_ln_eps", 1e-5))
+        self.gguf_writer.add_vision_projector_scale_factor(self.merge_kernel_size[0])
+
+        # Image size limits
+        # Note: in_patch_limit is for images, in_patch_limit_each_frame is for video (not supported yet)
+        in_patch_limit = self.preprocessor_config.get("in_patch_limit", 16384)
+        min_patches = 8  # reasonable minimum
+        pixels_per_patch = self.patch_size ** 2
+        self.gguf_writer.add_vision_min_pixels(min_patches * pixels_per_patch)
+        self.gguf_writer.add_vision_max_pixels(in_patch_limit * pixels_per_patch)
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int) -> Tensor:
+        out_dim, in_dim = weights.shape
+        head_dim = out_dim // n_head
+        w = weights.reshape(n_head, head_dim // 4, 2, 2, in_dim)
+        w = w.permute(0, 2, 1, 3, 4)
+        return w.reshape(out_dim, in_dim)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Only process vision and projector tensors
+        is_vision = any(x in name for x in ["vision_tower", "mm_projector"])
+
+        if not is_vision:
+            return
+
+        assert self.hparams_vision is not None
+        n_head = self.hparams_vision.get("num_attention_heads", 16)
+
+        # Permute Q/K weights/biases from interleaved to split RoPE format
+        # This allows using build_rope_2d at runtime without post-permutation.
+        if "wqkv" in name:
+            out_dim = data_torch.shape[0]
+            qkv_dim = out_dim // 3
+            head_dim = qkv_dim // n_head
+
+            if "weight" in name:
+                wq, wk, wv = data_torch[:qkv_dim, :], data_torch[qkv_dim:2 * qkv_dim, :], data_torch[2 * qkv_dim:, :]
+                wq = self.permute(wq, n_head)
+                wk = self.permute(wk, n_head)
+                data_torch = torch.cat([wq, wk, wv], dim=0)
+            elif "bias" in name:
+                bq, bk, bv = data_torch[:qkv_dim], data_torch[qkv_dim:2 * qkv_dim], data_torch[2 * qkv_dim:]
+                bq = bq.reshape(n_head, head_dim // 4, 2, 2).permute(0, 2, 1, 3).reshape(-1)
+                bk = bk.reshape(n_head, head_dim // 4, 2, 2).permute(0, 2, 1, 3).reshape(-1)
+                data_torch = torch.cat([bq, bk, bv], dim=0)
+
+        # Temporal embeddings: (T, 1, C) → (T, C)
+        if "pos_emb.time_weight" in name:
+            T, _, C = data_torch.shape
+            data_torch = data_torch.reshape(T, C)
+
+        # PatchMergerMLP tensor name mapping
+        # proj.0.weight → proj.linear_1.weight
+        # proj.2.weight → proj.linear_2.weight
+        if "mm_projector.proj.0." in name:
+            name = name.replace(".proj.0.", ".proj.linear_1.")
+        elif "mm_projector.proj.2." in name:
+            name = name.replace(".proj.2.", ".proj.linear_2.")
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("CogVLMForCausalLM")
 class CogVLMVisionModel(MmprojModel):
 

docs/backend/snapdragon/README.md

@@ -35,7 +35,7 @@ Adapt below build commands accordingly.
 Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
 ```
-[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+[d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
 Preset CMake variables:

ggml/src/ggml-cpu/ops.cpp

@@ -2096,10 +2096,14 @@ static void ggml_compute_forward_gelu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2113,10 +2117,14 @@ static void ggml_compute_forward_gelu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2135,10 +2143,14 @@ static void ggml_compute_forward_gelu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2152,10 +2164,14 @@ static void ggml_compute_forward_gelu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2276,10 +2292,14 @@ static void ggml_compute_forward_gelu_erf_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2293,10 +2313,14 @@ static void ggml_compute_forward_gelu_erf_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2315,10 +2339,14 @@ static void ggml_compute_forward_gelu_erf_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2332,10 +2360,14 @@ static void ggml_compute_forward_gelu_erf_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2379,10 +2411,14 @@ static void ggml_compute_forward_gelu_quick_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2396,10 +2432,14 @@ static void ggml_compute_forward_gelu_quick_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2418,10 +2458,14 @@ static void ggml_compute_forward_gelu_quick_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2435,10 +2479,14 @@ static void ggml_compute_forward_gelu_quick_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2482,10 +2530,14 @@ static void ggml_compute_forward_silu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2499,10 +2551,14 @@ static void ggml_compute_forward_silu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2521,10 +2577,14 @@ static void ggml_compute_forward_silu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2538,10 +2598,14 @@ static void ggml_compute_forward_silu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {

ggml/src/ggml-cpu/repack.cpp

@@ -1916,9 +1916,10 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
         int src_offset = (i / 8) * blck_size_interleave;
         int dst_offset = i * blck_size_interleave;
 
+        // buffer large enough for the max interleave block size (8 bytes)
         uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&elems, &in[src_id].qs[src_offset], blck_size_interleave);
+        memcpy(&out.qs[dst_offset], &elems, blck_size_interleave);
     }
 
     // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K

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

@@ -111,7 +111,7 @@ 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_ASSERT(ggml_is_contiguous_rows(src0) && ggml_is_contiguous_rows(dst) && ggml_are_same_shape(src0, dst));
 
     GGML_TENSOR_UNARY_OP_LOCALS
 

ggml/src/ggml-cuda/convert.cu

@@ -7,7 +7,8 @@
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);
 
@@ -16,23 +17,27 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     }
 
     const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
 
-    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+        const int64_t i02 = dm.y;
+        const int64_t i03 = dm.x;
 
-    const int64_t ib = ibx0 + i00/qk; // block index
-    const int64_t iqs = (i00%qk)/qr; // quant index
-    const int64_t iybs = i00 - i00%qk; // y block start index
-    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+        const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
 
-    // dequantize
-    float2 v;
-    dequantize_kernel(vx, ib, iqs, v);
+        const int64_t ib = ibx0 + i00/qk; // block index
+        const int64_t iqs = (i00%qk)/qr; // quant index
+        const int64_t iybs = i00 - i00%qk; // y block start index
+        const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
-    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
-    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+        // dequantize
+        float2 v;
+        dequantize_kernel(vx, ib, iqs, v);
+
+        const int64_t iy0 = (i0203*ne01 + i01)*ne00 + iybs + iqs;
+        y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+        y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+    }
 }
 
 template <bool need_check>
@@ -485,9 +490,11 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, (int)std::min(ne0203, (int64_t)65535));
     dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -612,7 +619,8 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
 
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
-        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -621,23 +629,29 @@ static __global__ void convert_unary(
     }
 
     const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
 
     const src_t * x = (const src_t *) vx;
 
-    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
-    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+        const int64_t i02 = dm.y;
+        const int64_t i03 = dm.x;
+
+        const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+        const int64_t iy = (i0203*ne01 + i01)*ne00 + i00;
+        y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+    }
 }
 
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, (int)std::min(ne0203, (int64_t)65535));
     convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <typename src_t, typename dst_t>

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

@@ -63,11 +63,19 @@ static __global__ void flash_attn_ext_f16(
     constexpr int frag_m = ncols == 8 ? 32 : 16;
     constexpr int frag_n = ncols == 8 ?  8 : 16;
     static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+#if defined(GGML_USE_HIP)
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::row_major> frag_a_K;
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_a_V;
+    typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_b;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, _Float16>                          frag_c_VKQ;
+#else
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::row_major> frag_a_K;
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::col_major> frag_a_V;
     typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, half, wmma::col_major> frag_b;
     typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
     typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+#endif
 
     constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
     constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
@@ -126,6 +134,19 @@ static __global__ void flash_attn_ext_f16(
 
     __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
     half2 * VKQ2 = (half2 *) VKQ;
+
+#if defined(GGML_USE_HIP)
+    const _Float16 * K_h_f16  = reinterpret_cast<const _Float16 *>(K_h);
+    const _Float16 * V_h_f16  = reinterpret_cast<const _Float16 *>(V_h);
+    _Float16       * KQ_f16   = reinterpret_cast<_Float16 *>(KQ);
+    _Float16       * VKQ_f16  = reinterpret_cast<_Float16 *>(VKQ);
+#else
+    const half * K_h_f16  = K_h;
+    const half * V_h_f16  = V_h;
+    half       * KQ_f16   = KQ;
+    half       * VKQ_f16  = VKQ;
+#endif
+
 #pragma unroll
     for (int j0 = 0; j0 < ncols; j0 += nwarps) {
         const int j = j0 + threadIdx.y;
@@ -160,7 +181,7 @@ static __global__ void flash_attn_ext_f16(
     for (int i0 = 0; i0 < D; i0 += 16) {
 #pragma unroll
         for (int j0 = 0; j0 < ncols; j0 += frag_n) {
-            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ_f16 + j0*D_padded + i0, D_padded);
         }
     }
 
@@ -180,7 +201,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
             for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
                 frag_a_K K_a;
-                wmma::load_matrix_sync(K_a, K_h + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+                wmma::load_matrix_sync(K_a, K_h_f16 + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
@@ -310,7 +331,7 @@ static __global__ void flash_attn_ext_f16(
                 const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
                 wmma::load_matrix_sync(
                     KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
-                    KQ + j0*(kqar*kqs_padded) + k,
+                    KQ_f16 + j0*(kqar*kqs_padded) + k,
                     kqar*kqs_padded);
             }
         }
@@ -328,7 +349,7 @@ static __global__ void flash_attn_ext_f16(
                 const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
 
                 frag_a_V v_a;
-                wmma::load_matrix_sync(v_a, V_h + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+                wmma::load_matrix_sync(v_a, V_h_f16 + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
@@ -344,7 +365,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
             for (int j0 = 0; j0 < ncols; j0 += frag_n) {
                 wmma::store_matrix_sync(
-                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    KQ_f16 + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
                     VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
                     D_padded, wmma::mem_col_major);
             }

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

@@ -3640,11 +3640,13 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         n_fuse++;
 
                         if (n_fuse > 1) {
+                            ggml_tensor fused_add_node;
+                            memcpy(&fused_add_node, node, sizeof(ggml_tensor));
                             for (int j = 0; j < n_fuse - 1; ++j) {
-                                node->src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+                                fused_add_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
                             }
-                            cgraph->nodes[i + n_fuse - 1]->data = node->data;
-                            ggml_cuda_op_fused_add(*cuda_ctx, node, n_fuse);
+                            fused_add_node.data = cgraph->nodes[i + n_fuse - 1]->data;
+                            ggml_cuda_op_fused_add(*cuda_ctx, &fused_add_node, n_fuse);
                             i += n_fuse - 1;
 
                             continue;
@@ -4820,8 +4822,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
-        case GGML_OP_ACC:
             return true;
+        case GGML_OP_ACC:
+            // TODO: extend support like so:
+            //return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]);
+            return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
         case GGML_OP_SUM:
             return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_TOP_K:

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1935,11 +1935,6 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
         return false;
     }
 
-    // TODO: add support for non-contigiuos tensors
-    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
-        return false;
-    }
-
     return true;
 }
 
@@ -1991,6 +1986,25 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     return true;
 }
 
+static bool ggml_hexagon_supported_sum_rows(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (!hex_supported_src0_type(src0->type)) {
+        return false;
+    }
+    if (!hex_supported_dst_type(dst->type)) {
+        return false;
+    }
+
+    // TODO: add support for non-contigiuos tensors
+    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session * sess,
                                                const struct ggml_tensor *          op) {
     const struct ggml_tensor * src0 = op->src[0];
@@ -2111,6 +2125,26 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
     return true;
 }
 
+static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0]; // values
+    const struct ggml_tensor * dst  = op;         // indices
+
+    if (src0->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (dst->type != GGML_TYPE_I32) {
+        return false;
+    }
+
+    if (src0->ne[0] > (16*1024)) {
+        // reject tensors with huge rows for now
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const int32_t * op_params = &op->op_params[0];
 
@@ -2278,6 +2312,9 @@ static inline size_t init_binary_req(htp_general_req * req, dspqueue_buffer * bu
         case GGML_OP_SUB:
             req->op = HTP_OP_SUB;
             break;
+        case GGML_OP_DIV:
+            req->op = HTP_OP_DIV;
+            break;
         default:
             GGML_ABORT("ggml-hex: binary : unsupported op: %d\n", t->op);
             break;
@@ -2316,6 +2353,17 @@ static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer *
     return n_bufs;
 }
 
+static inline size_t init_argsort_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    req->op = HTP_OP_ARGSORT;
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 template <bool _is_src0_constant>
 static inline size_t init_binary_id_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
     switch (t->op) {
@@ -2370,6 +2418,16 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
             supported = true;
             break;
 
+        case GGML_OP_SQR:
+            req->op   = HTP_OP_SQR;
+            supported = true;
+            break;
+
+        case GGML_OP_SQRT:
+            req->op   = HTP_OP_SQRT;
+            supported = true;
+            break;
+
         case GGML_OP_UNARY:
             if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
                 req->op   = HTP_OP_UNARY_SILU;
@@ -2387,6 +2445,9 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
             } else if (ggml_get_glu_op(t) == GGML_GLU_OP_SWIGLU_OAI) {
                 req->op   = HTP_OP_GLU_SWIGLU_OAI;
                 supported = true;
+            } else if (ggml_get_glu_op(t) == GGML_GLU_OP_GEGLU) {
+                req->op   = HTP_OP_GLU_GEGLU;
+                supported = true;
             }
             break;
 
@@ -2411,6 +2472,17 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
     return n_bufs;
 }
 
+static inline size_t init_sum_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+    req->op = HTP_OP_SUM_ROWS;
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 static inline size_t init_rope_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
     memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
     req->op = HTP_OP_ROPE;
@@ -2519,6 +2591,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
             case GGML_OP_MUL:
             case GGML_OP_ADD:
             case GGML_OP_SUB:
+            case GGML_OP_DIV:
                 ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
                 break;
             case GGML_OP_ADD_ID:
@@ -2528,6 +2601,13 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
             case GGML_OP_SCALE:
                 ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                 break;
+            case GGML_OP_SQR:
+            case GGML_OP_SQRT:
+                ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
+                break;
+            case GGML_OP_SUM_ROWS:
+                ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
+                break;
             case GGML_OP_UNARY:
                 if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
                         (ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
@@ -2536,7 +2616,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 break;
             case GGML_OP_GLU:
                 if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
-                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI)) {
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
                     ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                 }
                 break;
@@ -2564,6 +2645,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
                 break;
 
+            case GGML_OP_ARGSORT:
+                ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
+                break;
+
             default:
                 GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
         }
@@ -2916,6 +3001,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_MUL:
         case GGML_OP_ADD:
         case GGML_OP_SUB:
+        case GGML_OP_DIV:
             supp = ggml_hexagon_supported_binary(sess, op);
             break;
 
@@ -2928,6 +3014,15 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_unary(sess, op);
             break;
 
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            supp = ggml_hexagon_supported_unary(sess, op);
+            break;
+
+        case GGML_OP_SUM_ROWS:
+            supp = ggml_hexagon_supported_sum_rows(sess, op);
+            break;
+
         case GGML_OP_SOFT_MAX:
             supp = ggml_hexagon_supported_softmax(sess, op);
             break;
@@ -2943,7 +3038,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_GLU:
             {
                 const auto glu_op = ggml_get_glu_op(op);
-                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI)) {
+                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
                     supp = ggml_hexagon_supported_activations(sess, op);
                 }
                 break;
@@ -2968,6 +3063,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_cpy(sess, op);
             break;
 
+        case GGML_OP_ARGSORT:
+            supp = ggml_hexagon_supported_argsort(sess, op);
+            break;
+
         default:
             break;
     }

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -6,6 +6,7 @@ include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
 include_directories(
     ${HEXAGON_SDK_ROOT}/incs
     ${HEXAGON_SDK_ROOT}/incs/stddef
+    ${CMAKE_CURRENT_SOURCE_DIR}/../../../include
     ${CMAKE_CURRENT_SOURCE_DIR}/../..
     ${CMAKE_CURRENT_SOURCE_DIR}/..
     ${CMAKE_CURRENT_SOURCE_DIR}
@@ -21,6 +22,7 @@ add_library(${HTP_LIB} SHARED
     matmul-ops.c
     binary-ops.c
     unary-ops.c
+    sum-rows-ops.c
     softmax-ops.c
     act-ops.c
     rope-ops.c
@@ -28,6 +30,7 @@ add_library(${HTP_LIB} SHARED
     set-rows-ops.c
     get-rows-ops.c
     cpy-ops.c
+    argsort-ops.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -410,7 +410,7 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
             // gelu = x * sigmoid(1.702 * x) // current implementation
             hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0);
             hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
         }
 
         dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -516,7 +516,7 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
 
             // silu = x * sigmoid(x)
             hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
         }
 
         dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -541,6 +541,143 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
          ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }
 
+static const float GELU_COEF_A     = 0.044715f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
+                                       const struct htp_tensor * src1,
+                                       struct htp_tensor *       dst,
+                                       const int32_t *           op_params,
+                                       struct htp_spad *         src0_spad,
+                                       struct htp_spad *         src1_spad,
+                                       struct htp_spad *         dst_spad,
+                                       uint32_t                  nth,
+                                       uint32_t                  ith,
+                                       uint32_t                  src0_nrows_per_thread,
+                                       dma_queue *               dma_queue) {
+    htp_act_preamble3;
+
+    size_t src0_row_size = nb01;
+    size_t src1_row_size = nb11;
+    size_t dst_row_size  = nb1;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    const bool src1_valid = src1->ne[0];
+    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
+    if (!src1_valid) {
+        const int32_t swapped = op_params[1];
+        data_src1             = data_src0;
+        src1_row_size         = src0_row_size;
+
+        const size_t nc_in_bytes = nc * SIZEOF_FP32;
+        data_src0 += swapped ? nc_in_bytes : 0;
+        data_src1 += swapped ? 0 : nc_in_bytes;
+    }
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
+    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
+    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+
+    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
+    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
+    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
+    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
+
+    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    if (BLOCK == 0) {
+        FARF(ERROR,
+             "geglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+             src0_spad->size_per_thread, src0_row_size_aligned);
+        return;
+    }
+
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);
+
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)),
+            src1_row_size_aligned, src1_row_size, block_size);
+    }
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
+        float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;
+
+        for (uint32_t ib = 0; ib < block_size; ib++) {
+            const uint8_t * src0_spad_ptr = (const uint8_t *)(src0_spad + ib * (src0_row_size_aligned / sizeof(float)));
+            const uint8_t * src1_spad_ptr = (const uint8_t *)(src1_spad + ib * (src1_row_size_aligned / sizeof(float)));
+            uint8_t *       dst_spad_ptr  = (uint8_t *)(dst_spad + ib * (dst_row_size_aligned / sizeof(float)));
+
+            // geglu tanh implementation
+            // geglu(x, g) = gelu(x) * g
+            // gelu(x) = 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)))
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, src0_spad_ptr, nc);                       // res = x*x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, GELU_COEF_A, nc);   // res = res * GELU_COEF_A
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 1.0f, nc);          // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, SQRT_2_OVER_PI, nc); // res = result * SQRT_2_OVER_PI
+            hvx_tanh_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, nc);         // res = tanh(res)
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, 1.0f, nc);           // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 0.5f, nc);          // res = res + 0.5f
+            hvx_mul_f32_aaa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, src1_spad_ptr, nc);       // res = res * g
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size,
+                                   dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)),
+                                       src0_row_size_aligned, src0_row_size, pref_block_size);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)),
+                                       src1_row_size_aligned, src1_row_size, pref_block_size);
+        }
+    }
+
+    dma_queue_flush(dma_queue);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "geglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
 static void unary_silu_f32(unsigned int n, unsigned int i, void * data) {
     struct htp_ops_context * octx = (struct htp_ops_context *) data;
     unary_silu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
@@ -559,6 +696,12 @@ static void glu_swiglu_oai_f32(unsigned int n, unsigned int i, void * data) {
                                    &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
 }
 
+static void glu_geglu_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    glu_geglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
+                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
+}
+
 static int execute_op_activations_f32(struct htp_ops_context * octx) {
     int err = HTP_STATUS_OK;
 
@@ -593,6 +736,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
             act_op_func = unary_gelu_f32;
             op_type     = "gelu-f32";
             break;
+
+        case HTP_OP_GLU_GEGLU:
+            act_op_func = glu_geglu_f32;
+            op_type     = "geglu-f32";
+            break;
         default:
             FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
             return HTP_STATUS_NO_SUPPORT;

ggml/src/ggml-hexagon/htp/argsort-ops.c

@@ -0,0 +1,281 @@
+#include <string.h>
+#include <stdlib.h>
+#include <math.h>
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "ggml.h"
+
+#include "hvx-utils.h"
+#include "hex-dma.h"
+
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+#ifndef MIN
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+struct htp_argsort_context {
+    struct htp_ops_context * octx;
+    uint32_t                 nrows_per_thread;
+};
+
+static inline bool all_greater_f32(HVX_Vector x, HVX_Vector y)
+{
+    const HVX_Vector one  = Q6_V_vsplat_R(1);
+    const HVX_Vector zero = Q6_V_vzero();
+
+    HVX_VectorPred pred = Q6_Q_vcmp_gt_VsfVsf(x, y);
+    HVX_Vector matches = Q6_V_vmux_QVV(pred, one, zero);
+    HVX_Vector sum = hvx_vec_reduce_sum_i32(matches);
+    return hvx_vec_get_i32(sum) == 32;
+}
+
+// Sorts values and mirrors swaps to indices.
+static void quicksort_values_indices_asc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+    while (i <= j) {
+        // Vectorized scan for i
+        while (i <= j) {
+            // Check if we have at least one full vector
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    // If all elements are < pivot, we can skip this whole block
+                    i += 32;
+                    continue;
+                }
+            }
+
+            // Scalar fallback / cleanup
+            if (values[i] < pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j
+        while (i <= j) {
+            if (j - 32 >= i) {
+                // Load 32 elements ending at j.
+                // Since we want `values[j] > pivot`, let's load from j-31 to j.
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] > pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_asc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_asc(values, indices, i, right);
+}
+
+static void quicksort_values_indices_desc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+
+    while (i <= j) {
+        // Vectorized scan for i (values[i] > pivot)
+        while (i <= j) {
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    i += 32;
+                    continue;
+                }
+            }
+
+            if (values[i] > pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j (values[j] < pivot)
+        while (i <= j) {
+            if (j - 32 >= i) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] < pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_desc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_desc(values, indices, i, right);
+}
+
+static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_argsort_context * actx = (struct htp_argsort_context *)data;
+    struct htp_ops_context * octx = actx->octx;
+
+    // Unpack context
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * dst = &octx->dst;
+
+    // Scratchpad memory
+    uint8_t * spad = octx->src0_spad.data + octx->src0_spad.size_per_thread * i;
+
+    // Dimensions
+    uint32_t ne00 = src0->ne[0];
+    uint32_t ne01 = src0->ne[1];
+    uint32_t ne02 = src0->ne[2];
+    uint32_t ne03 = src0->ne[3];
+
+    uint32_t nb01 = src0->nb[1];
+    //uint32_t nb02 = src0->nb[2];
+    //uint32_t nb03 = src0->nb[3];
+
+    uint32_t nb1 = dst->nb[1];
+    //uint32_t nb2 = dst->nb[2];
+    //uint32_t nb3 = dst->nb[3];
+
+    // Sort order
+    enum ggml_sort_order order = (enum ggml_sort_order) octx->op_params[0];
+
+    // Rows to process
+    uint32_t total_rows = ne01 * ne02 * ne03;
+    uint32_t rows_per_thread = actx->nrows_per_thread;
+    uint32_t start_row = rows_per_thread * i;
+    uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);
+
+    // Scratchpad layout:
+    // We need space for one row of float data (values) and one row of int32 indices.
+    // values: ne00 * sizeof(float)
+    // indices: ne00 * sizeof(int32_t)
+    // Padded to 128 bytes.
+
+    size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
+    float * values_buf = (float *) spad;
+    int32_t * indices_buf = (int32_t *) (spad + values_size);
+
+    for (uint32_t r = start_row; r < end_row; r++) {
+        uint32_t src_offset = r * nb01;
+        uint32_t dst_offset = r * nb1;
+
+        uint8_t * src_ptr = (uint8_t *) src0->data + src_offset;
+        uint8_t * dst_ptr = (uint8_t *) dst->data  + dst_offset;
+
+        hex_l2fetch(src_ptr, ne00 * sizeof(float), ne00 * sizeof(float), 1);
+        hvx_copy_f32_au((uint8_t*)values_buf, src_ptr, ne00);
+
+        // Initialize indices
+        for (uint32_t j = 0; j < ne00; j++) {
+            indices_buf[j] = j;
+        }
+
+        // Sort values and mirror swaps to indices
+        if (order == GGML_SORT_ORDER_ASC) {
+            quicksort_values_indices_asc(values_buf, indices_buf, 0, ne00 - 1);
+        } else {
+            quicksort_values_indices_desc(values_buf, indices_buf, 0, ne00 - 1);
+        }
+
+        // Copy indices back to DDR
+        hvx_copy_f32_ua(dst_ptr, (const uint8_t *) indices_buf, ne00);
+    }
+}
+
+int op_argsort(struct htp_ops_context * octx) {
+    // Check supported types
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    // Allocate scratchpad
+    // We need 1 row of float + 1 row of int32 per thread.
+    uint32_t ne00 = octx->src0.ne[0];
+    size_t values_size  = hex_round_up(ne00 * sizeof(float), 128);
+    size_t indices_size = hex_round_up(ne00 * sizeof(int32_t), 128);
+    size_t spad_per_thread = values_size + indices_size;
+
+    // Make sure we round up to 256 for alignment requirements
+    spad_per_thread = hex_round_up(spad_per_thread, 256);
+
+    size_t total_spad_size = spad_per_thread * octx->n_threads;
+
+    if (octx->ctx->vtcm_size < total_spad_size) {
+        FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src0_spad.size = total_spad_size;
+    octx->src0_spad.size_per_thread = spad_per_thread;
+
+    FARF(HIGH, "argsort: %ux%ux%ux%u -> %ux%ux%ux%u (0x%x, 0x%x)",
+         octx->src0.ne[0], octx->src0.ne[1], octx->src0.ne[2], octx->src0.ne[3],
+         octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
+         octx->src0.data, octx->dst.data);
+
+    uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    uint32_t n_jobs = MIN(total_rows, octx->n_threads);
+
+    struct htp_argsort_context actx;
+    actx.octx = octx;
+    actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
+
+    // Run jobs
+    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
+
+    return HTP_STATUS_OK;
+}

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -17,121 +17,6 @@
 #include "htp-msg.h"
 #include "htp-ops.h"
 
-static inline HVX_Vector hvx_load_f32_to_f16(const HVX_Vector * restrict src, const HVX_Vector zero) {
-    HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(src[0], zero);  // 32 elements
-    HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(src[1], zero);  // 32 elements
-    return Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
-}
-
-// Dot product of FP32 and FP16 vectors, accumulating to float
-static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict y, const void * restrict x, unsigned int n, float s) {
-    const HVX_Vector * restrict vy = (const HVX_Vector * restrict) y; // fp32
-    const HVX_Vector * restrict vx = (const HVX_Vector * restrict) x; // fp16
-
-    uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16; // leftover elements
-
-    const HVX_Vector zero = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum = Q6_V_vsplat_R(0);
-
-    uint32_t i = 0;
-
-    #pragma unroll(4)
-    for (i = 0; i < nvec; i++) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x_hf  = vx[i];
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
-    }
-
-    if (nloe) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x_hf  = vx[i];
-
-        // Zero-out unused elements
-        // Note that we need to clear both x and y because they may contain NANs
-        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        x_hf = Q6_V_vand_QV(bmask, x_hf);
-        y_hf = Q6_V_vand_QV(bmask, y_hf);
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
-    }
-
-    rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
-    hvx_vec_store_u(r, 4, Q6_Vsf_equals_Vqf32(rsum));
-}
-
-// Dot product of FP32 and FP16 vectors, accumulating to float
-static inline void hvx_dot_f32_f16_aa_rx2(float * restrict r,
-                                          const void * restrict y,
-                                          const void * restrict x0,
-                                          const void * restrict x1,
-                                          unsigned int n,
-                                          float        s) {
-    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp32
-    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x0;  // fp16
-    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
-
-    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
-
-    const HVX_Vector zero  = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
-
-    uint32_t i = 0;
-
-    #pragma unroll(2)
-    for (i = 0; i < nvec; i++) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-        // Load x (fp16)
-        HVX_Vector x0_hf = vx0[i];
-        HVX_Vector x1_hf = vx1[i];
-
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
-    }
-
-    if (nloe) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x0_hf = vx0[i];
-        HVX_Vector x1_hf = vx1[i];
-
-        // Zero-out unused elements
-        // Note that we need to clear both x and y because they may contain NANs
-        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        x0_hf                = Q6_V_vand_QV(bmask, x0_hf);
-        x1_hf                = Q6_V_vand_QV(bmask, x1_hf);
-        y_hf                 = Q6_V_vand_QV(bmask, y_hf);
-
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
-    }
-
-    HVX_Vector rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32x2(rsum0, rsum1));
-    hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
-}
-
 // Dot product of two F16 vectors, accumulating to float
 static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict x, const void * restrict y, unsigned int n, float s) {
     const HVX_Vector * restrict vx = (const HVX_Vector * restrict) x; // fp16
@@ -140,8 +25,7 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16; // leftover elements
 
-    const HVX_Vector zero = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum = Q6_V_vsplat_R(0);
+    HVX_Vector rsum = Q6_V_vsplat_R(0);
 
     uint32_t i = 0;
 
@@ -156,11 +40,10 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     }
 
     if (nloe) {
-        HVX_Vector y_hf = vy[i];
-
         // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        HVX_Vector      x_hf = Q6_V_vand_QV(bmask, vx[i]);
+        HVX_Vector y_hf = Q6_V_vand_QV(bmask, vy[i]);
+        HVX_Vector x_hf = Q6_V_vand_QV(bmask, vx[i]);
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
@@ -181,12 +64,11 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
     const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp16
 
-    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
+    uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;  // leftover elements
 
-    const HVX_Vector zero  = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
+    HVX_Vector rsum0 = Q6_V_vsplat_R(0);
+    HVX_Vector rsum1 = Q6_V_vsplat_R(0);
 
     uint32_t i = 0;
 
@@ -204,12 +86,11 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     }
 
     if (nloe) {
-        HVX_Vector y_hf = vy[i];
-
         // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        HVX_Vector     x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
-        HVX_Vector     x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+        HVX_Vector x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
+        HVX_Vector x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+        HVX_Vector y_hf  = Q6_V_vand_QV(bmask, vy[i]);
 
         HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
         HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
@@ -222,7 +103,7 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
 }
 
-// MAD: y (F32) += x (F16) * s (float)
+// MAD: y (F32) += x (F16) * s (F32)
 static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, int n, float s) {
     const HVX_Vector * restrict ptr_x = (const HVX_Vector *) x;
     HVX_Vector * restrict ptr_y = (HVX_Vector *) y;
@@ -259,15 +140,125 @@ static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict
     }
 }
 
+// MAD: y (F32) += x0 (F16) * s0 (F32) + x1 (F16) * s1 (F32)
+static inline void hvx_mad_f32_f16_aa_rx2(float * restrict y,
+                                          const void * restrict x0,
+                                          const void * restrict x1,
+                                          float s0,
+                                          float s1,
+                                          int   n) {
+    const HVX_Vector * restrict ptr_x0 = (const HVX_Vector *) x0;
+    const HVX_Vector * restrict ptr_x1 = (const HVX_Vector *) x1;
+    HVX_Vector * restrict ptr_y        = (HVX_Vector *) y;
+
+    uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;  // leftover elements
+
+    HVX_Vector S0 = hvx_vec_splat_f16(s0);
+    HVX_Vector S1 = hvx_vec_splat_f16(s1);
+
+    uint32_t i = 0;
+    #pragma unroll(2)
+    for (i = 0; i < nvec; ++i) {
+        // Multiply x * s -> pair of F32 vectors
+        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
+        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
+
+        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
+        HVX_Vector xs_p_hi = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
+
+        ptr_y[i * 2]     = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_lo, ptr_y[i * 2]));
+        ptr_y[i * 2 + 1] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_hi, ptr_y[i * 2 + 1]));
+    }
+
+    if (nloe) {
+        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
+        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
+
+        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
+        HVX_Vector xs      = xs_p_lo;
+        i = 2 * i;  // index for ptr_y
+
+        if (nloe >= 32) {
+            ptr_y[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
+            nloe -= 32; ++i;
+            xs = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
+        }
+
+        if (nloe) {
+            HVX_Vector xy = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
+            hvx_vec_store_a(&ptr_y[i], nloe * 4, xy);
+        }
+    }
+}
+
 #define FLASH_ATTN_BLOCK_SIZE 128
 
-static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, int nth) {
+struct htp_fa_context {
+    const struct htp_ops_context * octx;
+
+    struct fastdiv_values src0_div21;
+    struct fastdiv_values src0_div1;
+
+    struct fastdiv_values broadcast_rk2;
+    struct fastdiv_values broadcast_rk3;
+    struct fastdiv_values broadcast_rv2;
+    struct fastdiv_values broadcast_rv3;
+
+    struct fastdiv_values src3_div2;
+    struct fastdiv_values src3_div3;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+
+    uint32_t n_blocks;
+
+    size_t size_q_row_padded;
+    size_t size_k_row_padded;
+    size_t size_v_row_padded;
+
+    size_t size_k_block;
+    size_t size_v_block;
+    size_t size_m_block;
+
+    bool is_q_fp32;
+};
+
+static inline void hvx_scale_vec_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, HVX_Vector vs) {
+    assert((size_t) dst % 128 == 0);
+    assert((size_t) src % 128 == 0);
+
+    const HVX_Vector * restrict vsrc = (const HVX_Vector * restrict) src;
+    HVX_Vector * restrict vdst       = (HVX_Vector * restrict) dst;
+
+    const uint32_t nvec = n / VLEN_FP32;
+    const uint32_t nloe = n % VLEN_FP32;
+
+    uint32_t i = 0;
+    #pragma unroll(4)
+    for (; i < nvec; ++i) {
+        vdst[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs));
+    }
+    if (nloe) {
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs);
+        hvx_vec_store_a(&vdst[i], nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v));
+    }
+}
+
+static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_fa_context * factx = (struct htp_fa_context *) data;
+    const struct htp_ops_context * octx = factx->octx;
     const struct htp_tensor * q = &octx->src0;
     const struct htp_tensor * k = &octx->src1;
     const struct htp_tensor * v = &octx->src2;
     const struct htp_tensor * mask  = (octx->src3.data) ? &octx->src3 : NULL;
     const struct htp_tensor * sinks = (octx->src4.data) ? &octx->src4 : NULL;
-    struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * dst = &octx->dst;
 
     const uint32_t neq0 = q->ne[0];
     const uint32_t neq1 = q->ne[1];
@@ -304,18 +295,6 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     const uint32_t nb2 = dst->nb[2];
     const uint32_t nb3 = dst->nb[3];
 
-    float scale         = 1.0f;
-    float max_bias      = 0.0f;
-    float logit_softcap = 0.0f;
-
-    memcpy(&scale,         (float *) octx->op_params + 0, sizeof(float));
-    memcpy(&max_bias,      (float *) octx->op_params + 1, sizeof(float));
-    memcpy(&logit_softcap, (float *) octx->op_params + 2, sizeof(float));
-
-    if (logit_softcap != 0) {
-        scale /= logit_softcap;
-    }
-
     // total rows in q
     const uint32_t nr = neq1*neq2*neq3;
 
@@ -331,18 +310,8 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     const uint32_t DV = nev0;
 
     const size_t size_q_row = DK * ((q->type == HTP_TYPE_F32) ? 4 : 2);
-    const size_t size_q_row_padded = hex_round_up(size_q_row, 128);
-
     const size_t size_k_row = DK * sizeof(__fp16);
     const size_t size_v_row = DV * sizeof(__fp16);
-    const size_t size_m_row = FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16); // Treat block as one row for mask
-
-    const size_t size_k_row_padded = hex_round_up(size_k_row, 128);
-    const size_t size_v_row_padded = hex_round_up(size_v_row, 128);
-
-    const size_t size_k_block = size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    const size_t size_v_block = size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    const size_t size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
 
     // Scratchpad buffers for Q, K, V, Mask, and VKQ32 accumulator
     uint8_t * spad_q = octx->src0_spad.data + octx->src0_spad.size_per_thread * ith;
@@ -351,31 +320,28 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     uint8_t * spad_m = octx->src3_spad.data + octx->src3_spad.size_per_thread * ith;
     uint8_t * spad_a = octx->dst_spad.data  + octx->dst_spad.size_per_thread  * ith;
 
-    const uint32_t n_head = neq2;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+    const HVX_Vector logit_cap = hvx_vec_splat_f32(factx->logit_softcap);
 
     for (uint32_t ir = ir0; ir < ir1; ++ir) {
-        const uint32_t iq3 = fastdiv(ir, &octx->src0_div21);
-        const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &octx->src0_div1);
+        const uint32_t iq3 = fastdiv(ir, &factx->src0_div21);
+        const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &factx->src0_div1);
         const uint32_t iq1 = (ir - iq3*neq2*neq1 - iq2 * neq1);
 
-        const uint32_t ik3 = fastdiv(iq3, &octx->broadcast_rk3);
-        const uint32_t ik2 = fastdiv(iq2, &octx->broadcast_rk2);
+        const uint32_t ik3 = fastdiv(iq3, &factx->broadcast_rk3);
+        const uint32_t ik2 = fastdiv(iq2, &factx->broadcast_rk2);
 
-        const uint32_t iv3 = fastdiv(iq3, &octx->broadcast_rv3);
-        const uint32_t iv2 = fastdiv(iq2, &octx->broadcast_rv2);
+        const uint32_t iv3 = fastdiv(iq3, &factx->broadcast_rv3);
+        const uint32_t iv2 = fastdiv(iq2, &factx->broadcast_rv2);
 
         // Fetch Q row
         const uint8_t * q_row_ptr = (const uint8_t *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3);
-        dma_queue_push(dma, dma_make_ptr(spad_q, q_row_ptr), size_q_row_padded, nbq1, size_q_row, 1);
+        dma_queue_push(dma, dma_make_ptr(spad_q, q_row_ptr), factx->size_q_row_padded, nbq1, size_q_row, 1);
 
         const uint32_t h = iq2; // head index
-        const float slope = (max_bias > 0.0f) ? (h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1)) : 1.0f;
+        const float slope = (factx->max_bias > 0.0f) ? (h < factx->n_head_log2 ? powf(factx->m0, h + 1) : powf(factx->m1, 2*(h - factx->n_head_log2) + 1)) : 1.0f;
 
-        float S = 0.0f;      // sum
-        float M = -INFINITY; // maximum KQ value
+        HVX_Vector S_vec = hvx_vec_splat_f32(0.0f);
+        HVX_Vector M_vec = hvx_vec_splat_f32(-INFINITY);
 
         // Clear accumulator
         hvx_splat_f32_a(spad_a, 0, DV);
@@ -383,40 +349,42 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
 
         const __fp16 * mp_base = NULL;
         if (mask) {
-            const uint32_t im2 = fastmodulo(iq2, mask->ne[2], &octx->src3_div2);
-            const uint32_t im3 = fastmodulo(iq3, mask->ne[3], &octx->src3_div3);
+            const uint32_t im2 = fastmodulo(iq2, mask->ne[2], &factx->src3_div2);
+            const uint32_t im3 = fastmodulo(iq3, mask->ne[3], &factx->src3_div3);
             mp_base = (const __fp16 *) ((const uint8_t *) mask->data + iq1*mask->nb[1] + im2*mask->nb[2] + im3*mask->nb[3]);
         }
 
-        const uint32_t n_blocks = (nek1 + FLASH_ATTN_BLOCK_SIZE - 1) / FLASH_ATTN_BLOCK_SIZE;
-
         // Prefetch first two blocks
-        for (uint32_t ib = 0; ib < MIN(n_blocks, 2); ++ib) {
+        for (uint32_t ib = 0; ib < MIN(factx->n_blocks, 2); ++ib) {
             const uint32_t ic_start = ib * FLASH_ATTN_BLOCK_SIZE;
             const uint32_t current_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - ic_start);
 
             // K
             const uint8_t * k_src = (const uint8_t *) k->data + (ic_start*nbk1 + ik2*nbk2 + ik3*nbk3);
-            uint8_t * k_dst = spad_k + (ib % 2) * size_k_block;
-            dma_queue_push(dma, dma_make_ptr(k_dst, k_src), size_k_row_padded, nbk1, size_k_row, current_block_size);
+            uint8_t * k_dst = spad_k + (ib % 2) * factx->size_k_block;
+            dma_queue_push(dma, dma_make_ptr(k_dst, k_src), factx->size_k_row_padded, nbk1, size_k_row, current_block_size);
 
             // V
             const uint8_t * v_src = (const uint8_t *) v->data + (ic_start*nbv1 + iv2*nbv2 + iv3*nbv3);
-            uint8_t * v_dst = spad_v + (ib % 2) * size_v_block;
-            dma_queue_push(dma, dma_make_ptr(v_dst, v_src), size_v_row_padded, nbv1, size_v_row, current_block_size);
+            uint8_t * v_dst = spad_v + (ib % 2) * factx->size_v_block;
+            dma_queue_push(dma, dma_make_ptr(v_dst, v_src), factx->size_v_row_padded, nbv1, size_v_row, current_block_size);
 
             // Mask
             if (mask) {
                 const uint8_t * m_src = (const uint8_t *) (mp_base + ic_start);
-                uint8_t * m_dst = spad_m + (ib % 2) * size_m_block;
+                uint8_t * m_dst = spad_m + (ib % 2) * factx->size_m_block;
                 // Mask is 1D contiguous for this row
                 dma_queue_push(dma, dma_make_ptr(m_dst, m_src), current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
             }
         }
 
-        const uint8_t * q_ptr_vtcm = dma_queue_pop(dma).dst;
+        uint8_t * q_ptr_vtcm = dma_queue_pop(dma).dst;
+        if (factx->is_q_fp32) {
+            hvx_copy_f16_f32_aa(q_ptr_vtcm, q_ptr_vtcm, DK);  // inplace convert f32 to f16
+        }
 
-        for (uint32_t ib = 0; ib < n_blocks; ++ib) {
+        const HVX_Vector slope_vec = hvx_vec_splat_f16(slope);
+        for (uint32_t ib = 0; ib < factx->n_blocks; ++ib) {
             const uint32_t ic_start = ib * FLASH_ATTN_BLOCK_SIZE;
             const uint32_t current_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - ic_start);
 
@@ -428,8 +396,6 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             // Inner loop processing the block from VTCM
             uint32_t ic = 0;
 
-            const bool is_q_fp32 = (q->type == HTP_TYPE_F32);
-
             // Process in blocks of 32 (VLEN_FP32)
             static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 <= 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
             HVX_Vector_x4 scores_x4;
@@ -437,22 +403,18 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
                 float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
-                for (int j = 0; j < VLEN_FP32; j += 2) {
+                for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
                     const uint32_t cur_ic = ic + j;
-                    const uint8_t * k_ptr = k_base + cur_ic * size_k_row_padded;
-                    if (is_q_fp32) {
-                        hvx_dot_f32_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
-                    } else {
-                        hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
-                    }
+                    const uint8_t * k_ptr = k_base + cur_ic * factx->size_k_row_padded;
+                    hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + factx->size_k_row_padded, DK, factx->scale);
                 }
 
                 HVX_Vector scores = *(HVX_Vector *) scores_arr;
 
                 // 2. Softcap
-                if (logit_softcap != 0.0f) {
+                if (factx->logit_softcap != 0.0f) {
                     scores = hvx_vec_tanh_f32(scores);
-                    scores = Q6_Vqf32_vmpy_VsfVsf(scores, hvx_vec_splat_f32(logit_softcap));
+                    scores = Q6_Vqf32_vmpy_VsfVsf(scores, logit_cap);
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
@@ -460,70 +422,59 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 if (mask) {
                     const __fp16 * mp = m_base + ic;
                     HVX_Vector m_vals_f16 = *(const HVX_UVector *) mp;
-
-                    HVX_Vector one_f16 = Q6_Vh_vsplat_R(0x3c00);
-                    HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), one_f16);
-
-                    HVX_Vector m_vals_f32 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(m_vals_f32_pair));
-
-                    HVX_Vector slope_vec = hvx_vec_splat_f32(slope);
-                    HVX_Vector add_val = Q6_Vqf32_vmpy_VsfVsf(m_vals_f32, slope_vec);
-                    scores = Q6_Vqf32_vadd_VsfVsf(scores, Q6_Vsf_equals_Vqf32(add_val));
+                    HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), slope_vec);
+                    HVX_Vector add_val = Q6_V_lo_W(m_vals_f32_pair);
+                    scores = Q6_Vqf32_vadd_Vqf32Vsf(add_val, scores);
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
                 scores_x4.v[iv] = scores;
-                v_max = Q6_Vsf_vmax_VsfVsf(scores, v_max);
+                v_max = hvx_vec_reduce_max2_f32(scores, v_max); // All lanes have block max
             }
 
             {
                 // 4. Online Softmax Update
-                v_max = hvx_vec_reduce_max_f32(v_max);
-                float m_block = hvx_vec_get_f32(v_max);
-                float M_old = M;
-                float M_new = (m_block > M) ? m_block : M;
-                M = M_new;
+                HVX_Vector M_new_vec = Q6_Vsf_vmax_VsfVsf(v_max, M_vec);
+                HVX_Vector diff_vec  = Q6_Vqf32_vsub_VsfVsf(M_vec, M_new_vec);
+                HVX_Vector ms_vec    = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(diff_vec));
+                M_vec = M_new_vec;
 
-                const float ms = expf(M_old - M_new);
-                hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
+                hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
-                HVX_Vector M_new_vec = hvx_vec_splat_f32(M_new);
                 HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
                 for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
                     HVX_Vector scores = scores_x4.v[iv];
-                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
+                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_vec);
                     HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
 
                     p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
 
                     // 5. Accumulate V
                     float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
-                    *(HVX_Vector*)p_arr = P;
+                    *(HVX_Vector *) p_arr = P;
 
-                    for (int j = 0; j < VLEN_FP32; ++j) {
-                        const uint32_t cur_ic = ic2 + j;
-                        const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
-                        hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
+                        const uint32_t  cur_ic = ic2 + j;
+                        const uint8_t * v_ptr  = v_base + cur_ic * factx->size_v_row_padded;
+                        hvx_mad_f32_f16_aa_rx2(VKQ32, v_ptr, v_ptr + factx->size_v_row_padded, p_arr[j], p_arr[j + 1], DV);
                     }
                 }
 
                 p_sum_vec = hvx_vec_reduce_sum_f32(p_sum_vec);
-                S = S * ms + hvx_vec_get_f32(p_sum_vec);
+                S_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(S_vec, ms_vec)), p_sum_vec));
             }
 
+            // Sync scalars for leftover/next block if needed
+            float M = hvx_vec_get_f32(M_vec);
+            float S = hvx_vec_get_f32(S_vec);
+
             // Leftover
             for (; ic < current_block_size; ++ic) {
                 float s_val;
-                const uint8_t * k_ptr = k_base + ic * size_k_row_padded;
-
-                if (is_q_fp32) {
-                    hvx_dot_f32_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
-                } else {
-                    hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
-                }
-
-                if (logit_softcap != 0.0f) {
-                    s_val = logit_softcap * tanhf(s_val);
+                const uint8_t * k_ptr = k_base + ic * factx->size_k_row_padded;
+                hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, factx->scale);
+                if (factx->logit_softcap != 0.0f) {
+                    s_val = factx->logit_softcap * tanhf(s_val);
                 }
 
                 if (mask) {
@@ -532,37 +483,42 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 }
 
                 const float Mold = M;
-                float ms = 1.0f;
                 float vs = 1.0f;
 
                 if (s_val > M) {
                     M = s_val;
-                    ms = expf(Mold - M);
-                    hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
+                    HVX_Vector diff_vec = hvx_vec_splat_f32(Mold - M);
+                    HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
+                    hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
+
+                    float ms = hvx_vec_get_f32(ms_vec);
+                    S = S * ms + vs;
                 } else {
-                    vs = expf(s_val - M);
+                    HVX_Vector diff_vec = hvx_vec_splat_f32(s_val - M);
+                    vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
+                    S += vs;
                 }
 
-                const uint8_t * v_ptr = v_base + ic * size_v_row_padded;
+                const uint8_t * v_ptr = v_base + ic * factx->size_v_row_padded;
 
                 hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, vs);
-
-                S = S * ms + vs;
             }
+            M_vec = hvx_vec_splat_f32(M);
+            S_vec = hvx_vec_splat_f32(S);
 
             // Issue DMA for next+1 block (if exists)
-            if (ib + 2 < n_blocks) {
+            if (ib + 2 < factx->n_blocks) {
                 const uint32_t next_ib = ib + 2;
                 const uint32_t next_ic_start = next_ib * FLASH_ATTN_BLOCK_SIZE;
                 const uint32_t next_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - next_ic_start);
 
                 // K
                 const uint8_t * k_src = (const uint8_t *) k->data + (next_ic_start*nbk1 + ik2*nbk2 + ik3*nbk3);
-                dma_queue_push(dma, dma_make_ptr(k_base, k_src), size_k_row_padded, nbk1, size_k_row, next_block_size);
+                dma_queue_push(dma, dma_make_ptr(k_base, k_src), factx->size_k_row_padded, nbk1, size_k_row, next_block_size);
 
                 // V
                 const uint8_t * v_src = (const uint8_t *) v->data + (next_ic_start*nbv1 + iv2*nbv2 + iv3*nbv3);
-                dma_queue_push(dma, dma_make_ptr(v_base, v_src), size_v_row_padded, nbv1, size_v_row, next_block_size);
+                dma_queue_push(dma, dma_make_ptr(v_base, v_src), factx->size_v_row_padded, nbv1, size_v_row, next_block_size);
 
                 // Mask
                 if (mask) {
@@ -573,20 +529,26 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
         }
 
         // sinks
+        float M = hvx_vec_get_f32(M_vec);
+        float S = hvx_vec_get_f32(S_vec);
+
         if (sinks) {
             const float s = ((float *)((char *) sinks->data))[h];
 
-            float ms = 1.0f;
             float vs = 1.0f;
 
             if (s > M) {
-                ms = expf(M - s);
-                hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
-            } else {
-                vs = expf(s - M);
-            }
+                HVX_Vector diff_vec = hvx_vec_splat_f32(M - s);
+                HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
+                hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
-            S = S * ms + vs;
+                float ms = hvx_vec_get_f32(ms_vec);
+                S = S * ms + vs;
+            } else {
+                HVX_Vector diff_vec = hvx_vec_splat_f32(s - M);
+                vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
+                S += vs;
+            }
         }
 
         const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
@@ -609,53 +571,73 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     }
 }
 
-static void htp_flash_attn_ext_job(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = data;
-    flash_attn_ext_f16_thread(octx, i, n);
-}
-
 int op_flash_attn_ext(struct htp_ops_context * octx) {
     const struct htp_tensor * q = &octx->src0;
     const struct htp_tensor * k = &octx->src1;
     const struct htp_tensor * v = &octx->src2;
-    const struct htp_tensor * mask = (octx->src3.type != HTP_TYPE_COUNT) ? &octx->src3 : NULL;
-    struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * mask = (octx->src3.data) ? &octx->src3 : NULL;
+    const struct htp_tensor * dst = &octx->dst;
 
     // Check support
-    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) ||
-        k->type != HTP_TYPE_F16 ||
-        v->type != HTP_TYPE_F16) {
+    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) || k->type != HTP_TYPE_F16 || v->type != HTP_TYPE_F16) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    octx->src0_div21 = init_fastdiv_values(q->ne[2] * q->ne[1]);
-    octx->src0_div1  = init_fastdiv_values(q->ne[1]);
+    struct htp_fa_context factx;
+    factx.octx = octx;
 
-    octx->broadcast_rk2 = init_fastdiv_values(q->ne[2]/k->ne[2]);
-    octx->broadcast_rk3 = init_fastdiv_values(q->ne[3]/k->ne[3]);
-    octx->broadcast_rv2 = init_fastdiv_values(q->ne[2]/v->ne[2]);
-    octx->broadcast_rv3 = init_fastdiv_values(q->ne[3]/v->ne[3]);
+    factx.src0_div21 = init_fastdiv_values(q->ne[2] * q->ne[1]);
+    factx.src0_div1  = init_fastdiv_values(q->ne[1]);
+
+    factx.broadcast_rk2 = init_fastdiv_values(q->ne[2]/k->ne[2]);
+    factx.broadcast_rk3 = init_fastdiv_values(q->ne[3]/k->ne[3]);
+    factx.broadcast_rv2 = init_fastdiv_values(q->ne[2]/v->ne[2]);
+    factx.broadcast_rv3 = init_fastdiv_values(q->ne[3]/v->ne[3]);
 
     if (mask) {
-        octx->src3_div2 = init_fastdiv_values(mask->ne[2]);
-        octx->src3_div3 = init_fastdiv_values(mask->ne[3]);
+        factx.src3_div2 = init_fastdiv_values(mask->ne[2]);
+        factx.src3_div3 = init_fastdiv_values(mask->ne[3]);
     }
 
-    size_t size_q_row_padded = hex_round_up(q->ne[0] * (q->type == HTP_TYPE_F32 ? 4 : 2), 128);
-    size_t size_k_row_padded = hex_round_up(k->ne[0] * sizeof(__fp16), 128);
-    size_t size_v_row_padded = hex_round_up(v->ne[0] * sizeof(__fp16), 128);
+    factx.is_q_fp32 = (q->type == HTP_TYPE_F32);
+    factx.size_q_row_padded = hex_round_up(q->ne[0] * (factx.is_q_fp32 ? 4 : 2), 128);
+    factx.size_k_row_padded = hex_round_up(k->ne[0] * sizeof(__fp16), 128);
+    factx.size_v_row_padded = hex_round_up(v->ne[0] * sizeof(__fp16), 128);
 
-    size_t size_q_block = size_q_row_padded * 1; // single row for now
-    size_t size_k_block = size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    size_t size_v_block = size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    size_t size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
+    size_t size_q_block = factx.size_q_row_padded * 1; // single row for now
+    factx.size_k_block = factx.size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
+    factx.size_v_block = factx.size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
+    factx.size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
+
+    factx.n_blocks = (k->ne[1] + FLASH_ATTN_BLOCK_SIZE - 1) / FLASH_ATTN_BLOCK_SIZE;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) octx->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) octx->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) octx->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
+
+    factx.scale = scale;
+    factx.max_bias = max_bias;
+    factx.logit_softcap = logit_softcap;
+
+    uint32_t n_head = q->ne[2];
+    factx.n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+    factx.m0 = powf(2.0f, -(max_bias       ) / factx.n_head_log2);
+    factx.m1 = powf(2.0f, -(max_bias / 2.0f) / factx.n_head_log2);
 
     size_t size_vkq_acc = hex_round_up(v->ne[0] * sizeof(float), 128); // VKQ32
 
     octx->src0_spad.size_per_thread = size_q_block * 1;
-    octx->src1_spad.size_per_thread = size_k_block * 2;
-    octx->src2_spad.size_per_thread = size_v_block * 2;
-    octx->src3_spad.size_per_thread = mask ? size_m_block * 2 : 0;
+    octx->src1_spad.size_per_thread = factx.size_k_block * 2;
+    octx->src2_spad.size_per_thread = factx.size_v_block * 2;
+    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * 2 : 0;
     octx->dst_spad.size_per_thread  = size_vkq_acc;
 
     octx->src0_spad.size = octx->src0_spad.size_per_thread * octx->n_threads;
@@ -677,7 +659,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size;
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        worker_pool_run_func(octx->ctx->worker_pool, htp_flash_attn_ext_job, octx, octx->n_threads);
+        worker_pool_run_func(octx->ctx->worker_pool, flash_attn_ext_f16_thread, &factx, octx->n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/htp-msg.h

@@ -42,32 +42,36 @@ enum htp_data_type {
     HTP_TYPE_COUNT
 };
 
-// These values are manually translated over to HTP
-// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
+// Do not reorder first 4 (used as an index)
 enum htp_op {
-    HTP_OP_MUL            = 0,
-    HTP_OP_ADD            = 1,
-    HTP_OP_SUB            = 2,
-    HTP_OP_DIV            = 3,
-    HTP_OP_MUL_MAT        = 4,
-    HTP_OP_MUL_MAT_ID     = 5,
-    HTP_OP_RMS_NORM       = 6,
-    HTP_OP_UNARY_SILU     = 7,
-    HTP_OP_UNARY_GELU     = 8,
-    HTP_OP_GLU_SWIGLU     = 9,
-    HTP_OP_GLU_SWIGLU_OAI = 10,
-    HTP_OP_SOFTMAX        = 11,
-    HTP_OP_ADD_ID         = 12,
-    HTP_OP_ROPE           = 13,
-    HTP_OP_FLASH_ATTN_EXT = 14,
-    HTP_OP_SET_ROWS       = 15,
-    HTP_OP_SCALE          = 16,
-    HTP_OP_GET_ROWS       = 17,
-    HTP_OP_CPY            = 18,
+    HTP_OP_MUL = 0,
+    HTP_OP_ADD = 1,
+    HTP_OP_SUB = 2,
+    HTP_OP_DIV = 3,
+    HTP_OP_MUL_MAT,
+    HTP_OP_MUL_MAT_ID,
+    HTP_OP_RMS_NORM,
+    HTP_OP_UNARY_SILU,
+    HTP_OP_UNARY_GELU,
+    HTP_OP_GLU_SWIGLU,
+    HTP_OP_GLU_SWIGLU_OAI,
+    HTP_OP_GLU_GEGLU,
+    HTP_OP_SOFTMAX,
+    HTP_OP_ADD_ID,
+    HTP_OP_ROPE,
+    HTP_OP_FLASH_ATTN_EXT,
+    HTP_OP_SET_ROWS,
+    HTP_OP_GET_ROWS,
+    HTP_OP_SCALE,
+    HTP_OP_CPY,
+    HTP_OP_ARGSORT,
+    HTP_OP_SQR,
+    HTP_OP_SQRT,
+    HTP_OP_SUM_ROWS,
     INVALID
 };
 
-static inline size_t htp_type_block_size(uint32_t t) {
+static inline size_t htp_t_block_size(uint32_t t) {
     switch (t) {
         case HTP_TYPE_F32:
             return 1;
@@ -103,22 +107,6 @@ static inline size_t htp_type_nbytes(uint32_t t) {
     return 0;
 }
 
-static const char * htp_type_name(uint32_t t) {
-    switch (t) {
-        case HTP_TYPE_F32:
-            return "fp32";
-        case HTP_TYPE_F16:
-            return "fp16";
-        case HTP_TYPE_Q4_0:
-            return "q4_0";
-        case HTP_TYPE_Q8_0:
-            return "q8_0";
-        case HTP_TYPE_MXFP4:
-            return "mxfp4";
-    }
-    return 0;
-}
-
 // Internal types
 #define QK_Q4_0x4x2  256  // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
 #define QK_Q8_0x4x2  256  // 4x Q8_0 blocks concat with next 4x Q8_0 blocks

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -64,25 +64,12 @@ struct htp_ops_context {
     struct fastdiv_values broadcast_rv2;
     struct fastdiv_values broadcast_rv3;
 
-    struct fastdiv_values mm_div_ne12_ne1; // fastdiv values for ne12 * ne1
-    struct fastdiv_values mm_div_ne1;      // fastdiv values for ne1
-    struct fastdiv_values mm_div_r2;       // fastdiv values for ne12 / ne02
-    struct fastdiv_values mm_div_r3;       // fastdiv values for ne13 / ne03
-
     struct fastdiv_values set_rows_div_ne12; // fastdiv values for ne12
     struct fastdiv_values set_rows_div_ne11; // fastdiv values for ne11
 
     struct fastdiv_values get_rows_div_ne10;      // fastdiv values for ne10
     struct fastdiv_values get_rows_div_ne10_ne11; // fastdiv values for ne10 * ne11
 
-    struct fastdiv_values cpy_div_ne01; // fastdiv values for ne01
-    struct fastdiv_values cpy_div_ne02; // fastdiv values for ne02
-    struct fastdiv_values cpy_div_ne03; // fastdiv values for ne03
-
-    struct fastdiv_values cpy_rshp_div_n0;       // fastdiv values for ne00
-    struct fastdiv_values cpy_rshp_div_n1n0;     // fastdiv values for ne00*ne01
-    struct fastdiv_values cpy_rshp_div_n2n1n0;   // fastdiv values for ne00*ne01*ne02
-
     uint32_t flags;
 };
 
@@ -90,6 +77,7 @@ int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
+int op_sum_rows(struct htp_ops_context * octx);
 int op_activations(struct htp_ops_context * octx);
 int op_softmax(struct htp_ops_context * octx);
 int op_add_id(struct htp_ops_context * octx);
@@ -98,5 +86,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx);
 int op_set_rows(struct htp_ops_context * octx);
 int op_get_rows(struct htp_ops_context * octx);
 int op_cpy(struct htp_ops_context * octx);
+int op_argsort(struct htp_ops_context * octx);
 
 #endif /* HTP_OPS_H */

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -46,127 +46,76 @@
 #define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
 #endif
 
-// ADD variants
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \
 
-static inline void hvx_add_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD);
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
+
+// Dispatcher logic
+#define HVX_BINARY_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128)) { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
+        } \
+    } else { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
+        } \
+    } \
 }
 
-static inline void hvx_add_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-// SUB variants
-
-static inline void hvx_sub_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-// MUL variants
-
-static inline void hvx_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-// Dispatchers
-
-static inline void hvx_add_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_add_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_add_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_add_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_add_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_sub_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_sub_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_sub_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_sub_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_sub_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_mul_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_mul_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_mul_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_mul_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_mul_f32_uu(dst, src0, src1, num_elems);
-    }
-}
+HVX_BINARY_DISPATCHER(hvx_add_f32)
+HVX_BINARY_DISPATCHER(hvx_sub_f32)
+HVX_BINARY_DISPATCHER(hvx_mul_f32)
 
 // Mul-Mul Optimized
-
 static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src0 % 128 == 0);
@@ -443,6 +392,68 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
     }
 }
 
+//
+// Square
+//
+
+#define hvx_sqr_loop_body(dst_type, src_type, vec_store)           \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(float);                          \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]);                        \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]);                   \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #undef HVX_OP_ADD
 #undef HVX_OP_SUB
 #undef HVX_OP_MUL
@@ -453,5 +464,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
+#undef DEFINE_HVX_BINARY_OP_VARIANTS
+#undef HVX_BINARY_DISPATCHER
 
 #endif // HVX_ARITH_H

ggml/src/ggml-hexagon/htp/hvx-base.h

@@ -66,6 +66,12 @@ static inline float hvx_vec_get_f32(HVX_Vector v) {
     return x;
 }
 
+static inline int32_t hvx_vec_get_i32(HVX_Vector v) {
+    int32_t __attribute__((aligned(128))) x;
+    hvx_vec_store_a(&x, 4, v);
+    return x;
+}
+
 static inline HVX_Vector hvx_vec_abs_f16(HVX_Vector v) {
     // abs by clearing the fp16 sign bit
     HVX_Vector mask = Q6_Vh_vsplat_R(0x7fff);

ggml/src/ggml-hexagon/htp/hvx-copy.h

@@ -136,8 +136,6 @@ static inline void hvx_copy_f32_uu(uint8_t * restrict dst, const uint8_t * restr
         dst_type * restrict vdst = (dst_type *) dst;                                \
         src_type * restrict vsrc = (src_type *) src;                                \
                                                                                     \
-        const HVX_Vector zero = Q6_V_vsplat_R(0);                                   \
-                                                                                    \
         const uint32_t elem_size = sizeof(__fp16);                                  \
         const uint32_t epv  = 128 / elem_size;                                      \
         const uint32_t nvec = n / epv;                                              \

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -0,0 +1,116 @@
+#ifndef HVX_DIV_H
+#define HVX_DIV_H
+
+#include <HAP_farf.h>
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "hvx-base.h"
+#include "hex-utils.h"
+#include "hvx-inverse.h"
+#include "hvx-arith.h"
+
+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
+#define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store)             \
+    do {                                                                             \
+        dst_type * restrict vdst = (dst_type *) dst;                                 \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                             \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                             \
+                                                                                     \
+        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                   \
+                                                                                     \
+        const uint32_t nvec = n / VLEN_FP32;                                         \
+        const uint32_t nloe = n % VLEN_FP32;                                         \
+                                                                                     \
+        uint32_t i = 0;                                                              \
+                                                                                     \
+        _Pragma("unroll(4)")                                                         \
+        for (; i < nvec; i++) {                                                      \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vdst[i] = res;                                                           \
+        }                                                                            \
+        if (nloe) {                                                                  \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res);                   \
+        }                                                                            \
+    } while(0)
+
+// 3-letter suffix variants
+static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_aau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_auu(dst, src0, src1, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uuu(dst, src0, src1, num_elems);
+        }
+    }
+}
+
+#undef HVX_OP_MUL
+
+#endif // HVX_DIV_H

ggml/src/ggml-hexagon/htp/hvx-sigmoid.h

@@ -91,6 +91,27 @@ static inline HVX_Vector hvx_vec_tanh_f32(HVX_Vector x) {
         }                                                       \
     } while(0)
 
+#define hvx_tanh_loop_body(dst_type, src_type, vec_store)       \
+    do {                                                        \
+        dst_type * restrict vdst = (dst_type *) dst;            \
+        src_type * restrict vsrc = (src_type *) src;            \
+                                                                \
+        const uint32_t epv  = 128 / sizeof(float);              \
+        const uint32_t nvec = n / epv;                          \
+        const uint32_t nloe = n % epv;                          \
+                                                                \
+        uint32_t i = 0;                                         \
+                                                                \
+        _Pragma("unroll(4)")                                    \
+        for (; i < nvec; i++) {                                 \
+             vdst[i] = hvx_vec_tanh_f32(vsrc[i]);               \
+        }                                                       \
+        if (nloe) {                                             \
+             HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]);        \
+             vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
+        }                                                       \
+    } while(0)
+
 static inline void hvx_sigmoid_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
@@ -111,4 +132,10 @@ static inline void hvx_sigmoid_f32_uu(uint8_t * restrict dst, const uint8_t * re
     hvx_sigmoid_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }
 
+static inline void hvx_tanh_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_tanh_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
 #endif /* HVX_SIGMOID_H */

ggml/src/ggml-hexagon/htp/hvx-sqrt.h

@@ -12,11 +12,17 @@
 #define RSQRT_ONE_HALF     0x3f000000  // 0.5
 #define RSQRT_THREE_HALVES 0x3fc00000  // 1.5
 
+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
 static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
     //Algorithm :
     //  x2 = input*0.5
     //  y  = * (long *) &input
-    //  y  = 0x5f3759df - (y>>2)
+    //  y  = 0x5f3759df - (y>>1)
     //  y  = y*(threehalfs - x2*y*y)
 
     HVX_Vector rsqrtconst = Q6_V_vsplat_R(RSQRT_CONST);
@@ -57,4 +63,64 @@ static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
     return Q6_Vsf_equals_Vqf32(temp);
 }
 
+// Compute sqrt(x) as x*inv_sqrt(x)
+#define hvx_sqrt_f32_loop_body(dst_type, src_type, vec_store)                \
+    do {                                                                     \
+        dst_type * restrict vdst = (dst_type *) dst;                         \
+        src_type * restrict vsrc = (src_type *) src;                         \
+                                                                             \
+        const uint32_t nvec = n / VLEN_FP32;                                 \
+        const uint32_t nloe = n % VLEN_FP32;                                 \
+                                                                             \
+        uint32_t i = 0;                                                      \
+                                                                             \
+        _Pragma("unroll(4)")                                                 \
+        for (; i < nvec; i++) {                                              \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vdst[i] = sqrt_res;                                              \
+        }                                                                    \
+        if (nloe) {                                                          \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, sqrt_res);      \
+        }                                                                    \
+    } while(0)
+
+static inline void hvx_sqrt_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
+    if ((unsigned long) dst % 128 == 0) {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #endif /* HVX_SQRT_H */

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -12,6 +12,7 @@
 #include "hvx-sigmoid.h"
 #include "hvx-sqrt.h"
 #include "hvx-arith.h"
+#include "hvx-div.h"
 #include "hvx-base.h"
 
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -189,7 +189,7 @@ static int vtcm_release_callback(unsigned int rctx, void * state) {
     // otherwise we'll release it once we're done with the current Op.
 
     if (ctx->vtcm_inuse) {
-        ctx->vtcm_needs_release = false;
+        ctx->vtcm_needs_release = true;
         return 0;
     }
 
@@ -440,6 +440,45 @@ static void proc_matmul_req(struct htp_context *     ctx,
     send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
+static void proc_argsort_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[1];
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[0].fd     = bufs[1].fd;
+    rsp_bufs[0].ptr    = bufs[1].ptr;
+    rsp_bufs[0].offset = bufs[1].offset;
+    rsp_bufs[0].size   = bufs[1].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_argsort(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_cpy_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
     struct dspqueue_buffer rsp_bufs[1];
 
@@ -679,6 +718,45 @@ static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * re
     send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
+static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[0].fd     = bufs[1].fd;
+    rsp_bufs[0].ptr    = bufs[1].ptr;
+    rsp_bufs[0].offset = bufs[1].offset;
+    rsp_bufs[0].size   = bufs[1].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_sum_rows(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_activations_req(struct htp_context *     ctx,
                                  struct htp_general_req * req,
                                  struct dspqueue_buffer * bufs,
@@ -951,6 +1029,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             case HTP_OP_MUL:
             case HTP_OP_ADD:
             case HTP_OP_SUB:
+            case HTP_OP_DIV:
                 if (n_bufs != 3) {
                     FARF(ERROR, "Bad binary-req buffer list");
                     continue;
@@ -968,6 +1047,25 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 proc_unary_req(ctx, &req, bufs);
                 break;
 
+            case HTP_OP_SQR:
+            case HTP_OP_SQRT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_unary_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_SUM_ROWS:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_sum_rows_req(ctx, &req, bufs);
+                break;
+
             case HTP_OP_UNARY_SILU:
             case HTP_OP_UNARY_GELU:
                 if (n_bufs != 2) {
@@ -980,6 +1078,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             case HTP_OP_GLU_SWIGLU:
             case HTP_OP_GLU_SWIGLU_OAI:
             case HTP_OP_SOFTMAX:
+            case HTP_OP_GLU_GEGLU:
                 if ((n_bufs != 2) && (n_bufs != 3)) {
                     FARF(ERROR, "Bad act-req buffer list");
                     continue;
@@ -1035,6 +1134,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 proc_cpy_req(ctx, &req, bufs);
                 break;
 
+            case HTP_OP_ARGSORT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad argsort-req buffer list");
+                    continue;
+                }
+                proc_argsort_req(ctx, &req, bufs);
+                break;
+
             default:
                 FARF(ERROR, "Unknown Op %u", req.op);
                 break;

ggml/src/ggml-hexagon/htp/sum-rows-ops.c

@@ -0,0 +1,115 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <string.h>
+#include <math.h>
+
+#include "hex-dma.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+
+#define sum_rows_preamble                       \
+    struct htp_tensor *src0 =  &octx->src0;\
+    struct htp_tensor *dst  = &octx->dst;  \
+                                           \
+    const uint32_t ne00 = src0->ne[0];     \
+    const uint32_t ne01 = src0->ne[1];     \
+    const uint32_t ne02 = src0->ne[2];     \
+    const uint32_t ne03 = src0->ne[3];     \
+                                           \
+    const uint32_t nb00 = src0->nb[0];     \
+    const uint32_t nb01 = src0->nb[1];     \
+    const uint32_t nb02 = src0->nb[2];     \
+    const uint32_t nb03 = src0->nb[3];     \
+                                           \
+    const uint32_t  ne0 = dst->ne[0];      \
+    const uint32_t  ne1 = dst->ne[1];      \
+    const uint32_t  ne2 = dst->ne[2];      \
+    const uint32_t  ne3 = dst->ne[3];      \
+                                           \
+    const uint32_t  nb0 = dst->nb[0];      \
+    const uint32_t  nb1 = dst->nb[1];      \
+    const uint32_t  nb2 = dst->nb[2];      \
+    const uint32_t  nb3 = dst->nb[3];      \
+
+static int sum_rows_thread_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+    sum_rows_preamble;
+
+    const uint32_t src0_nrows_per_thread  = octx->src0_nrows_per_thread;
+    const size_t src0_row_size = nb01;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return HTP_STATUS_OK;
+    }
+
+    int opt_path   = 0;
+    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src = (const uint8_t *) src0->data;
+    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+
+    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
+    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
+
+    for (uint32_t ir = 0; ir < src0_nrows_per_thread; ir++) {
+        const float * restrict src_local = src_th + (ir * ne00);
+
+        if (ir + 1 < src0_nrows_per_thread) {
+            hex_l2fetch(src_local + ne00, src0_row_size, src0_row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            dst_th[ir] = hvx_reduce_sum_f32_a((const uint8_t *) src_local, ne00);
+        } else {
+            dst_th[ir] = hvx_reduce_sum_f32((const uint8_t *) src_local, ne00);
+        }
+    }
+
+    return HTP_STATUS_OK;
+}
+
+static void sum_rows_work_f32(unsigned int n, unsigned int i, void *data) {
+    sum_rows_thread_f32((struct htp_ops_context *) data, n, i);
+}
+
+int op_sum_rows(struct htp_ops_context * octx) {
+    sum_rows_preamble;
+
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+
+    uint32_t n_jobs = MIN(n_threads, src0_nrows);
+    octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_work_f32, octx, n_jobs);
+
+    return HTP_STATUS_OK;
+}
+

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -132,6 +132,56 @@ static void rms_norm_htp_f32(const float * restrict src,
     }
 }
 
+static void sqr_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqr_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
+static void sqrt_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqrt_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
 static void unary_job_f32_per_thread(const struct htp_tensor * src,
                                      struct htp_tensor *       dst,
                                      uint8_t *                 spad,
@@ -181,6 +231,12 @@ static void unary_job_f32_per_thread(const struct htp_tensor * src,
         case HTP_OP_SCALE:
             scale_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
             break;
+        case HTP_OP_SQR:
+            sqr_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
+        case HTP_OP_SQRT:
+            sqrt_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
 
         default:
             break;
@@ -218,6 +274,14 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             unary_op_func = unary_job_dispatcher_f32;
             op_type       = "scale-f32";
             break;
+        case HTP_OP_SQR:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqr-f32";
+            break;
+        case HTP_OP_SQRT:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqrt-f32";
+            break;
 
         default:
             FARF(ERROR, "Unsupported unary Op %u\n", octx->op);

ggml/src/ggml-metal/ggml-metal-common.cpp

@@ -264,15 +264,25 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             case GGML_OP_NORM:
             case GGML_OP_RMS_NORM:
             case GGML_OP_GROUP_NORM:
+            case GGML_OP_L2_NORM:
             case GGML_OP_SUM_ROWS:
+            case GGML_OP_SSM_CONV:
+            case GGML_OP_SSM_SCAN:
+            case GGML_OP_CLAMP:
+            case GGML_OP_TRI:
+            case GGML_OP_DIAG:
             case GGML_OP_MUL:
             case GGML_OP_ADD:
             case GGML_OP_DIV:
             case GGML_OP_GLU:
             case GGML_OP_SCALE:
+            case GGML_OP_UNARY:
             case GGML_OP_GET_ROWS:
-            case GGML_OP_CPY:
             case GGML_OP_SET_ROWS:
+            case GGML_OP_SET:
+            case GGML_OP_CPY:
+            case GGML_OP_CONT:
+            case GGML_OP_REPEAT:
                 return true;
             default:
                 return ggml_op_is_empty(op);
@@ -312,7 +322,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             h_add(mrs1, node0);
 
             // that many nodes forward to search for a concurrent node
-            constexpr int N_FORWARD = 8;
+            constexpr int N_FORWARD = 64;
 
             for (int i1 = i0 + 1; i1 < i0 + N_FORWARD && i1 < n; i1++) {
                 if (used[i1]) {

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

@@ -328,31 +328,46 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum(ggml_metal_l
 }
 
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows(ggml_metal_library_t lib, const ggml_tensor * op) {
-    GGML_ASSERT(op->src[0]->nb[0] == ggml_type_size(op->src[0]->type));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
 
     char base[256];
     char name[256];
 
-    const char * op_str = "undefined";
+    int op_num = -1;
+
     switch (op->op) {
-        case GGML_OP_SUM_ROWS:
-            op_str = "sum_rows"; break;
-        case GGML_OP_MEAN:
-            op_str = "mean"; break;
+        case GGML_OP_SUM_ROWS: op_num = OP_SUM_ROWS_NUM_SUM_ROWS; break;
+        case GGML_OP_MEAN:     op_num = OP_SUM_ROWS_NUM_MEAN;     break;
         default: GGML_ABORT("fatal error");
     };
 
-    snprintf(base, 256, "kernel_%s_%s", op_str, ggml_type_name(op->src[0]->type));
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t_str  = ggml_type_name(op->type);
 
-    snprintf(name, 256, "%s", base);
+    const bool is_c4 = op->src[0]->ne[0] % 4 == 0;
+
+    snprintf(base, 256, "kernel_sum_rows_%s_%s%s", t0_str, t_str, is_c4 ? "_4" : "");
+    snprintf(name, 256, "%s_op=%d", base, op_num);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_SUM_ROWS + 0);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
     }
 
     res.smem = 32*sizeof(float);
 
+    if (is_c4) {
+        res.smem *= 4;
+    }
+
+    res.c4  = is_c4;
+
     return res;
 }
 
@@ -1480,13 +1495,15 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one(ggml_met
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_L2_NORM);
 
-    GGML_ASSERT(op->src[0]->ne[0] % 4 == 0);
-    GGML_ASSERT(ggml_is_contiguous_1(op->src[0]));
-
     char base[256];
     char name[256];
 
-    snprintf(base, 256, "kernel_l2_norm_f32");
+    const bool is_c4 = op->src[0]->ne[0] % 4 == 0;
+
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t_str  = ggml_type_name(op->type);
+
+    snprintf(base, 256, "kernel_l2_norm_%s_%s%s", t0_str, t_str, is_c4 ? "_4" : "");
     snprintf(name, 256, "%s", base);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
@@ -1494,6 +1511,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_met
         res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
     }
 
+    res.c4   = is_c4;
     res.smem = 32*sizeof(float);
 
     return res;

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

@@ -1019,7 +1019,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_LOG:
-            return ggml_is_contiguous_rows(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+            return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_TANH:
@@ -1039,7 +1039,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_EXPM1:
-                    return ggml_is_contiguous_rows(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                    return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
                 default:
                     return false;
             }
@@ -1067,8 +1067,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_ADD_ID:
-            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ACC:
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_REPEAT:
         case GGML_OP_CONV_TRANSPOSE_1D:
             return true;
@@ -1086,9 +1086,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
-            return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_L2_NORM:
-            return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
+            return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_COUNT_EQUAL:
             return has_simdgroup_reduction &&
                 op->src[0]->type == GGML_TYPE_I32 &&
@@ -1160,6 +1159,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return has_simdgroup_reduction;
+        case GGML_OP_SET:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:

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

@@ -82,6 +82,7 @@
 #define FC_COUNT_EQUAL                 1100
 #define FC_UNARY                       1200
 #define FC_BIN                         1300
+#define FC_SUM_ROWS                    1400
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPSG 8
@@ -118,6 +119,8 @@
 #define OP_UNARY_NUM_SOFTPLUS    115
 #define OP_UNARY_NUM_EXPM1       116
 
+#define OP_SUM_ROWS_NUM_SUM_ROWS 10
+#define OP_SUM_ROWS_NUM_MEAN     11
 
 // kernel argument structs
 //
@@ -539,8 +542,21 @@ typedef struct {
 
 typedef struct {
     int32_t  ne00;
-    int32_t  ne00_4;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
     uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    eps;
 } ggml_metal_kargs_l2_norm;
 

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

@@ -426,6 +426,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_flash_attn_ext(ctx, idx);
             } break;
+        case GGML_OP_SET:
+            {
+                n_fuse = ggml_metal_op_set(ctx, idx);
+            } break;
         case GGML_OP_DUP:
         case GGML_OP_CPY:
         case GGML_OP_CONT:
@@ -616,8 +620,8 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->type         == GGML_TYPE_F32);
 
-    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
-    GGML_ASSERT(ggml_is_contiguous(op->src[1]));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
     const size_t pnb1 = ((const int32_t *) op->op_params)[0];
     const size_t pnb2 = ((const int32_t *) op->op_params)[1];
@@ -667,10 +671,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     }
 
     ggml_metal_kargs_bin args = {
-        /*.ne00 =*/ ne00,
-        /*.ne01 =*/ ne01,
-        /*.ne02 =*/ ne02,
-        /*.ne03 =*/ ne03,
+        /*.ne00 =*/ ne10,
+        /*.ne01 =*/ ne11,
+        /*.ne02 =*/ ne12,
+        /*.ne03 =*/ ne13,
         /*.nb00 =*/ nb00,
         /*.nb01 =*/ pnb1,
         /*.nb02 =*/ pnb2,
@@ -683,10 +687,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
         /*.nb11 =*/ nb11,
         /*.nb12 =*/ nb12,
         /*.nb13 =*/ nb13,
-        /*.ne0  =*/ ne0,
-        /*.ne1  =*/ ne1,
-        /*.ne2  =*/ ne2,
-        /*.ne3  =*/ ne3,
+        /*.ne0  =*/ ne10,
+        /*.ne1  =*/ ne11,
+        /*.ne2  =*/ ne12,
+        /*.ne3  =*/ ne13,
         /*.nb0  =*/ nb0,
         /*.nb1  =*/ pnb1,
         /*.nb2  =*/ pnb2,
@@ -703,7 +707,13 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+    const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    int nth = 1;
+
+    while (2*nth < args.ne0 && nth < nth_max) {
+        nth *= 2;
+    }
 
     ggml_metal_encoder_dispatch_threadgroups(enc, ne11, ne12, ne13, nth, 1, 1);
 
@@ -904,6 +914,11 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
     ggml_metal_kargs_sum_rows args = {
         /*.ne00 =*/ ne00,
         /*.ne01 =*/ ne01,
@@ -925,21 +940,26 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
 
     auto pipeline = ggml_metal_library_get_pipeline_sum_rows(lib, op);
 
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }
+
     int nth = 32; // SIMD width
 
-    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    while (nth < args.ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
         nth *= 2;
     }
 
     nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, ne00);
+    nth = std::min(nth, (int) args.ne00);
 
     const size_t smem = pipeline.smem;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
@@ -1599,6 +1619,134 @@ int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_set(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    const size_t pnb1 = ((const int32_t *) op->op_params)[0];
+    const size_t pnb2 = ((const int32_t *) op->op_params)[1];
+    const size_t pnb3 = ((const int32_t *) op->op_params)[2];
+    const size_t offs = ((const int32_t *) op->op_params)[3];
+
+    const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
+
+    if (!inplace) {
+        // run a separete kernel to cpy src->dst
+        // not sure how to avoid this
+        // TODO: make a simpler cpy_bytes kernel
+
+        //const id<MTLComputePipelineState> pipeline = ctx->pipelines[GGML_METAL_PIPELINE_TYPE_CPY_F32_F32].obj;
+        auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
+
+        ggml_metal_kargs_cpy args = {
+            /*.nk0  =*/ ne00,
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.ne0  =*/ ne0,
+            /*.ne1  =*/ ne1,
+            /*.ne2  =*/ ne2,
+            /*.ne3  =*/ ne3,
+            /*.nb0  =*/ nb0,
+            /*.nb1  =*/ nb1,
+            /*.nb2  =*/ nb2,
+            /*.nb3  =*/ nb3,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+        const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
+
+        ggml_metal_op_concurrency_reset(ctx);
+    }
+
+    auto pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[1]->type, op->type);
+
+    GGML_ASSERT(ne10 % ggml_blck_size(op->src[1]->type) == 0);
+
+    int64_t nk0 = ne10;
+    if (ggml_is_quantized(op->src[1]->type)) {
+        nk0 = ne10/16;
+    } else if (ggml_is_quantized(op->type)) {
+        nk0 = ne10/ggml_blck_size(op->type);
+    }
+
+    int nth = std::min<int>(nk0, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+    // when rows are small, we can batch them together in a single threadgroup
+    int nrptg = 1;
+
+    // TODO: relax this constraint in the future
+    if (ggml_blck_size(op->src[1]->type) == 1 && ggml_blck_size(op->type) == 1) {
+        if (nth > nk0) {
+            nrptg = (nth + nk0 - 1)/nk0;
+            nth   = nk0;
+
+            if (nrptg*nth > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+                nrptg--;
+            }
+        }
+    }
+
+    nth = std::min<int>(nth, nk0);
+
+    ggml_metal_kargs_cpy args = {
+        /*.nk0  =*/ nk0,
+        /*.ne00 =*/ ne10,
+        /*.ne01 =*/ ne11,
+        /*.ne02 =*/ ne12,
+        /*.ne03 =*/ ne13,
+        /*.nb00 =*/ nb10,
+        /*.nb01 =*/ nb11,
+        /*.nb02 =*/ nb12,
+        /*.nb03 =*/ nb13,
+        /*.ne0  =*/ ne10,
+        /*.ne1  =*/ ne11,
+        /*.ne2  =*/ ne12,
+        /*.ne3  =*/ ne13,
+        /*.nb0  =*/ ggml_element_size(op),
+        /*.nb1  =*/ pnb1,
+        /*.nb2  =*/ pnb2,
+        /*.nb3  =*/ pnb3,
+    };
+
+    const int nw0 = nrptg == 1 ? (nk0 + nth - 1)/nth : 1;
+
+    bid_dst.offs += offs;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, nw0*(ne11 + nrptg - 1)/nrptg, ne12, ne13, nth, nrptg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -2979,39 +3127,59 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
 
-    int nth = 32; // SIMD width
-
     ggml_metal_kargs_l2_norm args = {
-        /*.ne00   =*/ ne00,
-        /*.ne00_4 =*/ ne00/4,
-        /*.nb01   =*/ nb01,
-        /*.eps    =*/ eps,
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.eps   =*/ eps,
     };
 
     auto pipeline = ggml_metal_library_get_pipeline_l2_norm(lib, op);
 
-    while (nth < ne00/4 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }
+
+    int nth = 32; // SIMD width
+
+    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
         nth *= 2;
     }
 
     nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, ne00/4);
 
     const size_t smem = pipeline.smem;
 
-    const int64_t nrows = ggml_nrows(op->src[0]);
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
 
     return 1;
 }

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -59,6 +59,7 @@ int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rwkv              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_set               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);

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

@@ -77,6 +77,14 @@ static inline float dot(float x, float y) {
     return x*y;
 }
 
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
 // NOTE: this is not dequantizing - we are simply fitting the template
 template <typename type4x4>
 void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
@@ -910,7 +918,7 @@ constant float a4_erf = -1.453152027f;
 constant float a5_erf = 1.061405429f;
 
 template<typename T>
-T erf_approx(T x) {
+inline T erf_approx(T x) {
     T sign_x = sign(x);
     x = fabs(x);
     T t = 1.0f / (1.0f + p_erf * x);
@@ -918,10 +926,27 @@ T erf_approx(T x) {
     return sign_x * y;
 }
 
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}
+
 constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
 constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
 
-template <typename T0, typename T>
+template <typename T0, typename T, typename TC>
 kernel void kernel_unary_impl(
         constant ggml_metal_kargs_unary & args,
         device const char * src0,
@@ -963,111 +988,111 @@ kernel void kernel_unary_impl(
             }
         }
 
-        device const T0 & x = src0_ptr[i0];
+        const TC x = (TC) src0_ptr[i0];
 
         if (FC_OP == OP_UNARY_NUM_SCALE) {
-            dst_ptr[i0] = args.scale * x + args.bias;
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
         }
 
         if (FC_OP == OP_UNARY_NUM_FILL) {
-            dst_ptr[i0] = args.val;
+            dst_ptr[i0] = (T) args.val;
         }
 
         if (FC_OP == OP_UNARY_NUM_CLAMP) {
-            dst_ptr[i0] = clamp(x, args.min, args.max);
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
         }
 
         if (FC_OP == OP_UNARY_NUM_SQR) {
-            dst_ptr[i0] = x * x;
+            dst_ptr[i0] = (T) (x * x);
         }
 
         if (FC_OP == OP_UNARY_NUM_SQRT) {
-            dst_ptr[i0] = sqrt(x);
+            dst_ptr[i0] = (T) sqrt(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_SIN) {
-            dst_ptr[i0] = sin(x);
+            dst_ptr[i0] = (T) sin(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_COS) {
-            dst_ptr[i0] = cos(x);
+            dst_ptr[i0] = (T) cos(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_LOG) {
-            dst_ptr[i0] = log(x);
+            dst_ptr[i0] = (T) log(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
-            dst_ptr[i0] = T(x > 0.0f)*x + T(x <= 0.0f)*(x * args.slope);
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
         }
 
         if (FC_OP == OP_UNARY_NUM_TANH) {
-            dst_ptr[i0] = precise::tanh(x);
+            dst_ptr[i0] = (T) precise::tanh(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_RELU) {
-            dst_ptr[i0] = fmax(0.0f, x);
+            dst_ptr[i0] = (T) fmax(0, x);
         }
 
         if (FC_OP == OP_UNARY_NUM_SIGMOID) {
-            dst_ptr[i0] = 1.0f / (1.0f + exp(-x));
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
         }
 
         if (FC_OP == OP_UNARY_NUM_GELU) {
-            dst_ptr[i0] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
         }
 
         if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
-            dst_ptr[i0] = 0.5f*x*(1.0f + erf_approx(SQRT_2_INV*x));
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
         }
 
         if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
-            dst_ptr[i0] = x * (1.0f/(1.0f + exp(GELU_QUICK_COEF*x)));
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
         }
 
         if (FC_OP == OP_UNARY_NUM_SILU) {
-            dst_ptr[i0] = x / (1.0f + exp(-x));
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
         }
 
         if (FC_OP == OP_UNARY_NUM_ELU) {
-            dst_ptr[i0] = T(x > 0.0f)*x + T(x <= 0.0f)*(exp(x) - 1.0f);
+            dst_ptr[i0] = (T) elu_approx(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_NEG) {
-            dst_ptr[i0] = -x;
+            dst_ptr[i0] = (T) -x;
         }
 
         if (FC_OP == OP_UNARY_NUM_ABS) {
-            dst_ptr[i0] = fabs(x);
+            dst_ptr[i0] = (T) fabs(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_SGN) {
-            dst_ptr[i0] = T(x > 0.0f) - T(x < 0.0f);
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
         }
 
         if (FC_OP == OP_UNARY_NUM_STEP) {
-            dst_ptr[i0] = T(x > 0.0f);
+            dst_ptr[i0] = T(x > 0);
         }
 
         if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
-            dst_ptr[i0] = x * fmax(0.0f, fmin(1.0f, x/6.0f + 0.5f));
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
         }
 
         if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
-            dst_ptr[i0] = fmax(0.0f, fmin(1.0f, x/6.0f + 0.5f));
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
         }
 
         if (FC_OP == OP_UNARY_NUM_EXP) {
-            dst_ptr[i0] = exp(x);
+            dst_ptr[i0] = (T) exp(x);
         }
 
         if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
-            dst_ptr[i0] = select(log(1.0f + exp(x)), x, x > 20.0f);
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
         }
 
         if (FC_OP == OP_UNARY_NUM_EXPM1) {
             // TODO: precise implementation
-            dst_ptr[i0] = exp(x) - 1.0f;
+            dst_ptr[i0] = (T) (exp(x) - 1);
         }
     }
 
@@ -1075,11 +1100,12 @@ kernel void kernel_unary_impl(
 #undef FC_CNT
 }
 
-typedef decltype(kernel_unary_impl<float, float>) kernel_unary_t;
-
-template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float>;
-template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4>;
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
 
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
 
 // OP: 0 - add, 1 - sub, 2 - mul, 3 - div
 constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
@@ -1483,33 +1509,35 @@ kernel void kernel_op_sum_f32(
     }
 }
 
-template <bool norm>
-kernel void kernel_sum_rows(
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
         constant ggml_metal_kargs_sum_rows & args,
-        device const float * src0,
-        device       float * dst,
-        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
         ushort3 tpitg[[thread_position_in_threadgroup]],
         ushort  sgitg[[simdgroup_index_in_threadgroup]],
         ushort  tiisg[[thread_index_in_simdgroup]],
         ushort3   ntg[[threads_per_threadgroup]]) {
-    int64_t i3 = tgpig.z;
-    int64_t i2 = tgpig.y;
-    int64_t i1 = tgpig.x;
+#define FC_OP  FC_sum_rows_op
 
-    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
-        return;
-    }
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
 
     if (sgitg == 0) {
-        shmem_f32[tiisg] = 0.0f;
+        shmem_t[tiisg] = 0.0f;
     }
 
-    device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
-    device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float sumf = 0;
+    T0 sumf = T0(0.0f);
 
     for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
         sumf += src_row[i0];
@@ -1520,23 +1548,33 @@ kernel void kernel_sum_rows(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     if (tiisg == 0) {
-        shmem_f32[sgitg] = sumf;
+        shmem_t[sgitg] = sumf;
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    sumf = shmem_f32[tiisg];
+    sumf = shmem_t[tiisg];
     sumf = simd_sum(sumf);
 
     if (tpitg.x == 0) {
-        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
     }
+
+#undef FC_OP
 }
 
-typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
 
-template [[host_name("kernel_sum_rows_f32")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
-template [[host_name("kernel_mean_f32")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
 
 template<typename T>
 kernel void kernel_cumsum_blk(
@@ -2417,9 +2455,6 @@ kernel void kernel_solve_tri_f32(
     const short K   = FC_solve_tri_k;
     const short NP  = PAD2(N, NW);
 
-    const int32_t ne02 = args.ne02;
-    const int32_t ne03 = args.ne03;
-
     const int32_t i03 = tgpig.z;
     const int32_t i02 = tgpig.y;
     const int32_t i01 = tgpig.x*NSG + sgitg;
@@ -2706,26 +2741,32 @@ template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_f
 template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
 template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
 
-kernel void kernel_l2_norm_f32(
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
         constant ggml_metal_kargs_l2_norm & args,
         device const char * src0,
         device       char * dst,
         threadgroup float * shmem_f32 [[threadgroup(0)]],
-        uint   tgpig[[threadgroup_position_in_grid]],
-        ushort tpitg[[thread_position_in_threadgroup]],
-        ushort sgitg[[simdgroup_index_in_threadgroup]],
-        ushort tiisg[[thread_index_in_simdgroup]],
-        ushort   ntg[[threads_per_threadgroup]]) {
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
     if (sgitg == 0) {
         shmem_f32[tiisg] = 0.0f;
     }
 
-    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
 
     float sumf = 0.0f;
 
     // parallel sum
-    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
         sumf += dot(x[i00], x[i00]);
     }
     sumf = simd_sum(sumf);
@@ -2743,12 +2784,16 @@ kernel void kernel_l2_norm_f32(
 
     const float scale = 1.0f/sqrt(max(sumf, args.eps));
 
-    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
-    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
         y[i00] = x[i00] * scale;
     }
 }
 
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
 kernel void kernel_group_norm_f32(
         constant ggml_metal_kargs_group_norm & args,
         device const float * src0,
@@ -5921,7 +5966,7 @@ kernel void kernel_flash_attn_ext_vec(
     static_assert(DK4 % NL == 0, "DK4 must be divisible by NL");
     static_assert(DV4 % NL == 0, "DV4 must be divisible by NL");
 
-    const short T = PK + NSG*SH; // shared memory size per query in (half)
+  //const short T = PK + NSG*SH; // shared memory size per query in (half)
 
   //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                      0*PK); // holds the query data
     threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                      0*PK); // same as above but in q4_t
@@ -8509,7 +8554,9 @@ kernel void kernel_mul_mm(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
+#ifdef GGML_METAL_HAS_TENSOR
     threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
 
     constexpr int NR0 = 64;
     constexpr int NR1 = 32;
@@ -8632,8 +8679,8 @@ kernel void kernel_mul_mm(
             const short sx = (tiitg%NL1);
             const short sy = (tiitg/NL1)/8;
 
-            const short dx = sx;
-            const short dy = sy;
+          //const short dx = sx;
+          //const short dy = sy;
 
             const short ly = (tiitg/NL1)%8;
 
@@ -8882,7 +8929,9 @@ kernel void kernel_mul_mm_id(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
+#ifdef GGML_METAL_HAS_TENSOR
     threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
 
     constexpr int NR0 = 64;
     constexpr int NR1 = 32;
@@ -9017,8 +9066,8 @@ kernel void kernel_mul_mm_id(
             const short sx = (tiitg%NL1);
             const short sy = (tiitg/NL1)/8;
 
-            const short dx = sx;
-            const short dy = sy;
+          //const short dx = sx;
+          //const short dy = sy;
 
             const short ly = (tiitg/NL1)%8;
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -85,6 +85,9 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_8x_flat
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q4_1_f32
+    mul_mv_q4_1_f32_flat
+    mul_mv_q4_k_f32
     mul_mv_q6_k_f32
     mul_mv_q6_k_f32_flat
     mul_mv_q8_0_f32
@@ -100,7 +103,10 @@ set(GGML_OPENCL_KERNELS
     gemv_moe_mxfp4_f32
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
+    mul_mm_q4_0_f32_l4_lm
+    mul_mm_q4_1_f32_l4_lm
     mul_mm_q8_0_f32_l4_lm
+    mul_mm_q6_k_f32_l4_lm
     mul_mm_q8_0_f32_8x4
     gemv_noshuffle_general_q8_0_f32
     mul

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

@@ -525,6 +525,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mm_f16_f32_kq;
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
+    cl_kernel kernel_convert_block_q4_1, kernel_restore_block_q4_1;
     cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
     cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
@@ -532,6 +533,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_restore_block_q4_0_noshuffle;
     cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     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_q4_1_f32;
+    cl_kernel kernel_mul_mv_q4_1_f32_flat;
+    cl_kernel kernel_mul_mv_q4_K_f32;
     cl_kernel kernel_mul_mv_q6_K_f32;
     cl_kernel kernel_mul_mv_q6_K_f32_flat;
     cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
@@ -563,7 +567,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mv_id_mxfp4_f32_flat;
     cl_kernel kernel_mul_mm_f32_f32_l4_lm;
     cl_kernel kernel_mul_mm_f16_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q4_0_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q4_1_f32_l4_lm;
     cl_kernel kernel_mul_mm_q8_0_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
 
     std::vector<ProfilingInfo> profiling_info;
 
@@ -886,6 +893,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_1  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_1", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_1  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_1", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_mxfp4_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4_trans", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_mxfp4_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4_trans", &err), err));
@@ -1117,6 +1126,57 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_q4_1_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_1_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_1_f32.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q4_1_f32 = clCreateKernel(prog, "kernel_mul_mv_q4_1_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_1_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_1_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_1_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q4_1_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q4_1_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_k_f32.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q4_K_f32 = clCreateKernel(prog, "kernel_mul_mv_q4_K_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_q6_k_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1342,6 +1402,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_q4_0_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q4_0_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q4_0_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q4_0_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q4_0_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_q4_1_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q4_1_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q4_1_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q4_1_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q4_1_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mm_q8_0_f32_l4_lm
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1358,6 +1450,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_q6_k_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q6_k_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q6_k_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q6_k_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q6_k_f32_l4_lm", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mm_f16_f32_kq_kqv
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2887,6 +2996,59 @@ struct ggml_tensor_extra_cl_q4_0 {
     }
 };
 
+struct ggml_tensor_extra_cl_q4_1 {
+    // Quantized values.
+    cl_mem q = nullptr;
+    // Quantized values in image1d_buffer_t.
+    cl_mem q_img = nullptr;
+    // Scales.
+    cl_mem d = nullptr;
+    // Scales in image1d_buffer_t.
+    cl_mem d_img = nullptr;
+    // Min
+    cl_mem m = nullptr;
+    // Min in image1d_buffer_t.
+    cl_mem m_img = nullptr;
+    // Size of quantized values.
+    size_t size_q = 0;
+    // Size of scales.
+    size_t size_d = 0;
+    // Size of min values.
+    size_t size_m = 0;
+
+    ~ggml_tensor_extra_cl_q4_1() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        if (m != nullptr) {
+            CL_CHECK(clReleaseMemObject(m));
+            m = nullptr;
+        }
+        // Currently, q_img and d_img are only initialized when SMALL_ALLOC is
+        // enabled. They point to the images in ggml_backend_opencl_buffer_context.
+        // So, there is no need to release them here.
+        // TODO: initialize them for non SMALL_PATH path, or remove them.
+        q_img = nullptr;
+        d_img = nullptr;
+        m_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+        size_m = 0;
+    }
+};
+
 struct ggml_tensor_extra_cl_mxfp4 {
     // Quantized values.
     cl_mem q = nullptr;
@@ -3363,7 +3525,9 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
-            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0  || op->src[0]->type == GGML_TYPE_Q4_1 ||
+                       op->src[0]->type == GGML_TYPE_MXFP4 ||
+                       op->src[0]->type == GGML_TYPE_Q4_K  ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
@@ -3592,6 +3756,21 @@ struct ggml_backend_opencl_buffer_context {
         return extra;
     }
 
+    ggml_tensor_extra_cl_q4_1 * ggml_opencl_alloc_temp_tensor_extra_q4_1() {
+        ggml_tensor_extra_cl_q4_1 * extra;
+        if (temp_tensor_extras_q4_1.empty()) {
+            extra = new ggml_tensor_extra_cl_q4_1();
+        } else {
+            extra = temp_tensor_extras_q4_1.back();
+            temp_tensor_extras_q4_1.pop_back();
+        }
+
+        temp_tensor_extras_q4_1_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
     ggml_tensor_extra_cl_mxfp4 * ggml_opencl_alloc_temp_tensor_extra_mxfp4() {
         ggml_tensor_extra_cl_mxfp4 * extra;
         if (temp_tensor_extras_mxfp4.empty()) {
@@ -3648,6 +3827,11 @@ struct ggml_backend_opencl_buffer_context {
         }
         temp_tensor_extras_q4_0_in_use.clear();
 
+        for (ggml_tensor_extra_cl_q4_1 * e : temp_tensor_extras_q4_1_in_use) {
+            temp_tensor_extras_q4_1.push_back(e);
+        }
+        temp_tensor_extras_q4_1_in_use.clear();
+
         for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
             temp_tensor_extras_mxfp4.push_back(e);
         }
@@ -3673,6 +3857,8 @@ struct ggml_backend_opencl_buffer_context {
     std::vector<ggml_tensor_extra_cl *> temp_tensor_extras_in_use;
     std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0;
     std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0_in_use;
+    std::vector<ggml_tensor_extra_cl_q4_1 *> temp_tensor_extras_q4_1;
+    std::vector<ggml_tensor_extra_cl_q4_1 *> temp_tensor_extras_q4_1_in_use;
     std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4;
     std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
@@ -4042,6 +4228,75 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         return;
 
     }
+    if (tensor->type == GGML_TYPE_Q4_1) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q4_1 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q4_1();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_m = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_d + size_m + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, mins, then quants.
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for mins.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_m;
+        extra->m = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_m, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_1;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->m));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
+        return;
+    }
     if (tensor->type == GGML_TYPE_MXFP4) {
         ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
         GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
@@ -4544,7 +4799,35 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             size, data, 0, NULL, NULL));
         CL_CHECK(clReleaseMemObject(data_device));
         return;
-    } else if (tensor->type == GGML_TYPE_MXFP4) {
+    }
+    if (tensor->type == GGML_TYPE_Q4_1) {
+        ggml_tensor_extra_cl_q4_1 * extra = (ggml_tensor_extra_cl_q4_1 *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q4_1;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->m));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_MXFP4) {
         ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *)tensor->extra;
 
         cl_int err;
@@ -8372,6 +8655,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
 #ifdef GGML_OPENCL_SOA_Q
     ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_q4_1 * extra0_q4_1 = (ggml_tensor_extra_cl_q4_1 *)src0->extra;
     ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
     ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
     ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
@@ -8885,6 +9169,91 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
                 return;
             }
+            case GGML_TYPE_Q4_0: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q4_0_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_Q4_1: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q4_1_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_1->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_1->d));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q4_1->m));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
             case GGML_TYPE_Q8_0: {
                 if (ne11 < 32) {
                     break;
@@ -8927,6 +9296,50 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
                 return;
             }
+            case GGML_TYPE_Q6_K: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q6_k_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q6_K->qh));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q6_K->s));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q6_K->d));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
             default:
                 break;
         }
@@ -9181,7 +9594,71 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
 #endif // GGML_OPENCL_SOA_Q
             break;
-        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q4_1: {
+#ifdef GGML_OPENCL_SOA_Q
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            kernel = backend_ctx->kernel_mul_mv_q4_1_f32_flat;
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_1->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_1->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q4_1->m));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r3));
+#else
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            kernel = backend_ctx->kernel_mul_mv_q4_1_f32;
+
+            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),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
         case GGML_TYPE_Q8_0: {
 #ifdef GGML_OPENCL_SOA_Q
             kernel = backend_ctx->kernel_mul_mv_q8_0_f32_flat;
@@ -9262,7 +9739,42 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         }
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K: {
+            kernel = backend_ctx->kernel_mul_mv_q4_K_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),     &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(int),        &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),     &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(int),        &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),     &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),        &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),        &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),        &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),   &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),   &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),   &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),        &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),   &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),   &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),   &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),        &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),        &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),        &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),        &r3));
+            break;
+        }
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
 #ifdef GGML_OPENCL_SOA_Q
@@ -9424,7 +9936,10 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else if (src0t == GGML_TYPE_Q4_K) {
-        GGML_ASSERT(false && "not implemented");
+        size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else if (src0t == GGML_TYPE_Q3_K) {
         GGML_ASSERT(false && "not implemented");
     } else if (src0t == GGML_TYPE_Q5_K) {

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

@@ -46,6 +46,15 @@ struct block_q4_0
     uint8_t qs[QK4_0 / 2];
 };
 
+//------------------------------------------------------------------------------
+// block_q4_1
+//------------------------------------------------------------------------------
+struct block_q4_1 {
+    half d; // delta
+    half m; // min
+    uchar qs[QK4_1 / 2]; // nibbles / quants
+};
+
 //------------------------------------------------------------------------------
 // block_q6_K
 //------------------------------------------------------------------------------
@@ -148,6 +157,48 @@ kernel void kernel_restore_block_q4_0_noshuffle(
     }
 }
 
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_1
+// Convert the block_q4_1 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_1(
+    global struct block_q4_1 * src0,
+    global uchar * dst_q,
+    global half  * dst_d,
+    global half  * dst_m
+) {
+    global struct block_q4_1 * b = (global struct block_q4_1 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_1/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+    global half  * m = (global half *) dst_m + get_global_id(0);
+
+    *d = b->d;
+    *m = b->m;
+
+    for (int i = 0; i < QK4_1/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_1(
+    global uchar * src_q,
+    global half  * src_d,
+    global half  * src_m,
+    global struct block_q4_1 * dst
+) {
+    global struct block_q4_1 * b = (global struct block_q4_1 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_1/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+    global half  * m = (global half *) src_m + get_global_id(0);
+
+    b->d = *d;
+    b->m = *m;
+    for (int i = 0; i < QK4_1/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
 //------------------------------------------------------------------------------
 // block_mxfp4
 //------------------------------------------------------------------------------

ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl

@@ -0,0 +1,163 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 8
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q4_0_f32_l4_lm(
+    global uchar4 * src0_q,
+    global half   * src0_d,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                int ib  = idx / 4;
+                int iqs = idx % 4;
+
+                float d = (float)src0_d[ib];
+                global uchar4 * qs = src0_q + ib*4 + iqs;
+                uchar4 q = *qs;
+                float4 v1 = (convert_float4((uchar4)((q.s0   )&0x0F, (q.s1   )&0x0F, (q.s2   )&0x0F, (q.s3   )&0x0F)) - 8.0f)*d;
+                float4 v2 = (convert_float4((uchar4)((q.s0>>4)&0x0F, (q.s1>>4)&0x0F, (q.s2>>4)&0x0F, (q.s3>>4)&0x0F)) - 8.0f)*d;
+
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = v1.s3;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = v2.s0;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = v2.s1;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = v2.s2;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = v2.s3;
+            } else {
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl

@@ -0,0 +1,165 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 8
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q4_1_f32_l4_lm(
+    global uchar4 * src0_q,
+    global half   * src0_d,
+    global half   * src0_m,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                int ib  = idx / 4;
+                int iqs = idx % 4;
+
+                float d = (float)src0_d[ib];
+                float m = (float)src0_m[ib];
+                global uchar4 * qs = src0_q + ib*4 + iqs;
+                uchar4 q = *qs;
+                float4 v1 = (convert_float4((uchar4)((q.s0   )&0x0F, (q.s1   )&0x0F, (q.s2   )&0x0F, (q.s3   )&0x0F)))*d + m;
+                float4 v2 = (convert_float4((uchar4)((q.s0>>4)&0x0F, (q.s1>>4)&0x0F, (q.s2>>4)&0x0F, (q.s3>>4)&0x0F)))*d + m;
+
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = v1.s3;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = v2.s0;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = v2.s1;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = v2.s2;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = v2.s3;
+            } else {
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl

@@ -0,0 +1,158 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 2
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q6_k_f32_l4_lm(
+    global uchar * src0_ql,
+    global uchar * src0_qh,
+    global char  * src0_s,
+    global half  * src0_d,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+
+                int ib = idx / 128;                  // 2 values per idx
+                int iqs = idx % 128;                 // 0..127
+
+                int n = iqs / 64;                    // 0,1
+                int b = (iqs % 64) / 32;             // 0,1
+                int is_b = (iqs % 16) / 8;           // 0,1
+                int qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+                int is = 8 * n + qhshift + is_b;     // 0..15
+                int qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+                int qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+                float dscale = (float)src0_d[ib] * (float)src0_s[ib*16 + is];
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 0] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 0] >> qhshift) & 3) << 4)) - 32);
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 1] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 1] >> qhshift) & 3) << 4)) - 32);
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl

@@ -0,0 +1,219 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#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)
+#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")))
+#endif
+
+#define QK4_1                   32
+
+struct block_q4_1 {
+    half d; // delta
+    half m; // min
+    uchar qs[QK4_1 / 2]; // nibbles / quants
+};
+
+inline float block_q4_1_dot_y(
+    global const struct block_q4_1 * qb_curr,
+    float sumy,
+    float16 yl,
+    int il
+) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float4 acc = (float4)(0.0f, 0.0f, 0.0f, 0.0f);
+
+    global const ushort * qs = ((global const ushort *) qb_curr + 2 + il/2);
+
+    acc.s0 += yl.s0 * (qs[0] & 0x000F);
+    acc.s0 += yl.s1 * (qs[0] & 0x0F00);
+    acc.s0 += yl.s8 * (qs[0] & 0x00F0);
+    acc.s3 += yl.s9 * (qs[0] & 0xF000);
+
+    acc.s0 += yl.s2 * (qs[1] & 0x000F);
+    acc.s1 += yl.s3 * (qs[1] & 0x0F00);
+    acc.s2 += yl.sa * (qs[1] & 0x00F0);
+    acc.s3 += yl.sb * (qs[1] & 0xF000);
+
+    acc.s0 += yl.s4 * (qs[2] & 0x000F);
+    acc.s1 += yl.s5 * (qs[2] & 0x0F00);
+    acc.s2 += yl.sc * (qs[2] & 0x00F0);
+    acc.s3 += yl.sd * (qs[2] & 0xF000);
+
+    acc.s0 += yl.s6 * (qs[3] & 0x000F);
+    acc.s1 += yl.s7 * (qs[3] & 0x0F00);
+    acc.s2 += yl.se * (qs[3] & 0x00F0);
+    acc.s3 += yl.sf * (qs[3] & 0xF000);
+
+    return d * (acc.s0 + acc.s1 + acc.s2 + acc.s3) + sumy * m;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each subgroup works on 4 rows
+#define N_SIMDGROUP 1 // number of subgroups in a thread group
+#define N_SIMDWIDTH 16 // assuming subgroup size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_1;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_1 * x = (global struct block_q4_1 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_1 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q4_1_dot_y(x+ib+0*nb, sumy, yl, il);
+        sumf.s1 += block_q4_1_dot_y(x+ib+1*nb, sumy, yl, il);
+        sumf.s2 += block_q4_1_dot_y(x+ib+2*nb, sumy, yl, il);
+        sumf.s3 += block_q4_1_dot_y(x+ib+3*nb, sumy, yl, il);
+
+        yb += QK4_1 * (N_SIMDWIDTH/2);
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_1_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl

@@ -0,0 +1,229 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#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)
+#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")))
+#endif
+
+#define QK4_1                   32
+
+struct block_q4_1 {
+    half d; // delta
+    half m; // min
+    uchar qs[QK4_1 / 2]; // nibbles / quants
+};
+
+inline float block_q4_1_dot_y_flat(
+    global const uchar * x,
+    global const half  * dh,
+    global const half  * mh,
+    float sumy,
+    float16 yl,
+    int il
+) {
+    float                 d   = *dh;
+    float                 m   = *mh;
+    global const ushort * qs = ((global const ushort *) x + il/2);
+
+    float4 acc = (float4)(0.0f, 0.0f, 0.0f, 0.0f);
+
+    acc.s0 += yl.s0 * (qs[0] & 0x000F);
+    acc.s0 += yl.s1 * (qs[0] & 0x0F00);
+    acc.s0 += yl.s8 * (qs[0] & 0x00F0);
+    acc.s3 += yl.s9 * (qs[0] & 0xF000);
+
+    acc.s0 += yl.s2 * (qs[1] & 0x000F);
+    acc.s1 += yl.s3 * (qs[1] & 0x0F00);
+    acc.s2 += yl.sa * (qs[1] & 0x00F0);
+    acc.s3 += yl.sb * (qs[1] & 0xF000);
+
+    acc.s0 += yl.s4 * (qs[2] & 0x000F);
+    acc.s1 += yl.s5 * (qs[2] & 0x0F00);
+    acc.s2 += yl.sc * (qs[2] & 0x00F0);
+    acc.s3 += yl.sd * (qs[2] & 0xF000);
+
+    acc.s0 += yl.s6 * (qs[3] & 0x000F);
+    acc.s1 += yl.s7 * (qs[3] & 0x0F00);
+    acc.s2 += yl.se * (qs[3] & 0x00F0);
+    acc.s3 += yl.sf * (qs[3] & 0xF000);
+
+    return d * (acc.s0 + acc.s1 + acc.s2 + acc.s3) + sumy * m;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each subgroup works on 4 rows
+#define N_SIMDGROUP 1 // number of subgroups in a thread group
+#define N_SIMDWIDTH 16 // assuming subgroup size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_flat(
+        global void * src0_q,
+        global void * src0_d,
+        global void * src0_m,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_1;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    // The number of scales/mins is the same as the number of blocks.
+    ulong offset0_dm = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02));
+    // Each block contains QK4_1/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q  = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_1/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_dm;
+    global half  * m = (global half  *) src0_m + offset0_dm;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_1 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 0*nb*QK4_1/2, d + ib + 0*nb, m + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 1*nb*QK4_1/2, d + ib + 1*nb, m + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 2*nb*QK4_1/2, d + ib + 2*nb, m + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 3*nb*QK4_1/2, d + ib + 3*nb, m + ib + 3*nb, sumy, yl, il);
+
+        yb += QK4_1 * (N_SIMDWIDTH/2);
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_1_f32_flat(
+        global void * src0_q,
+        global void * src0_d,
+        global void * src0_m,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_flat(src0_q, src0_d, src0_m, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl

@@ -0,0 +1,180 @@
+#ifdef cl_intel_required_subgroup_size
+#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)
+#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")))
+#endif
+
+//------------------------------------------------------------------------------
+// block_q4_K
+//------------------------------------------------------------------------------
+#define QK_K            256
+#define K_SCALE_SIZE    12
+
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+typedef struct {
+    half d;    // super-block scale for quantized scales
+    half dmin; // super-block scale for quantized mins
+
+    uchar scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uchar qs[QK_K/2];           // 4-bit quants
+} block_q4_K;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// number of (super) blocks each subgroup processes
+// each thread in a subgroup processes a block (32 weights)
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_K_f32(
+        global char * src0,
+        int offset0,
+        global char * src1,
+        int offset1,
+        global char * dst,
+        int offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    ushort kmask1 = 0x3f3f;
+    ushort kmask2 = 0x0f0f;
+    ushort kmask3 = 0xc0c0;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // block index (inside super block)
+    int iq = it/4;     // 0 or 1 - first or second half of the super block
+    int ir = it%4;     // 0...3 - block index in the half super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    int offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_q4_K * x = (global block_q4_K *) (src0 + offset_src0);
+    global float      * y = (global float      *) (src1 + offset_src1);
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    ushort  sc16[4];
+    uchar * sc8 = (uchar *)sc16;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+0];
+            sumy.s0 += yl[i+0];
+
+            yl[i+8] = y4[i+32];
+            sumy.s1 += yl[i+8];
+
+            yh[i+0] = y4[i+128];
+            sumy.s2 += yh[i+0];
+
+            yh[i+8] = y4[i+160];
+            sumy.s3 += yh[i+8];
+        }
+
+        global ushort * sc = (global ushort *)x[ib].scales + iq;
+        global ushort * q1 = (global ushort *)x[ib].qs + 16 * iq + 4 * ir;
+        global half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            global ushort * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1.s0 += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1.s1 += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1.s2 += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1.s3 += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2.s0 += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2.s1 += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2.s2 += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2.s3 += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1.s0 + 1.f/256.f * acc1.s1) * sc8[0] +
+                                 (acc1.s2 + 1.f/256.f * acc1.s3) * sc8[1] * 1.f/16.f +
+                                 (acc2.s0 + 1.f/256.f * acc2.s1) * sc8[4] +
+                                 (acc2.s2 + 1.f/256.f * acc2.s3) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy.s0 * sc8[2] + sumy.s1 * sc8[3] + sumy.s2 * sc8[6] + sumy.s3 * sc8[7]);
+
+            q1 += nb01/2;
+            sc += nb01/2;
+            dh += nb01/2;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -92,6 +92,7 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 #define VK_VENDOR_ID_APPLE 0x106b
 #define VK_VENDOR_ID_INTEL 0x8086
 #define VK_VENDOR_ID_NVIDIA 0x10de
+#define VK_VENDOR_ID_QUALCOMM 0x5143
 
 #define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256
 
@@ -687,6 +688,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_acc_f32;
+    vk_pipeline pipeline_set_f32;
 
     // [src0 0=fp32,1=fp16][src1 0=fp32,1=fp16][dst 0=fp32,1=fp16]
     vk_pipeline pipeline_add[2][2][2];
@@ -4080,7 +4082,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -4181,7 +4183,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_set_f32, "set_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0, 0}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -5641,6 +5644,10 @@ static void ggml_vk_instance_init() {
                             driver_priorities[vk::DriverId::eMesaNvk] = 2;
 #endif
                             break;
+                        case VK_VENDOR_ID_QUALCOMM:
+                            driver_priorities[vk::DriverId::eQualcommProprietary] = 1;
+                            driver_priorities[vk::DriverId::eMesaTurnip] = 2;
+                            break;
                     }
                     driver_priorities[vk::DriverId::eMesaDozen] = 100;
 
@@ -8422,6 +8429,8 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
+    const uint32_t tmpsh = (Bc / MatBc) * sizeof(float);
+
     const uint32_t qstride = hsk_pad / 4 + 2;
     const uint32_t Qf = Br * qstride * f16vec4;
 
@@ -8438,7 +8447,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
 
     const uint32_t slope = Br * acctype;
 
-    const uint32_t total_size = Qf + Psh + sfsh + ksh + slope;
+    const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
     VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", kv_type=" << kv_type << ", total_size=" << total_size << ", supported=" << supported);
@@ -8815,6 +8824,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_acc_f32;
         }
         return nullptr;
+    case GGML_OP_SET:
+        if (src0->type == src1->type && src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32)) {
+            return ctx->device->pipeline_set_f32;
+        }
+        return nullptr;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_MUL:
@@ -9801,16 +9816,16 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
+    int nb1 = dst->op_params[0] / src0_type_size; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / src0_type_size; // 4 bytes of float32
+    int nb3 = dst->op_params[2] / src0_type_size; // 4 bytes of float32
+    int offset = dst->op_params[3] / src0_type_size; // offset in bytes
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ACC, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, dst->op, {
         (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)nb3,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)nb3,
         0,
         0.0f, 0.0f, offset,
     });
@@ -10624,8 +10639,10 @@ static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& sub
 }
 
 static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
-    float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
+    const float * op_params = (const float *)dst->op_params;
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, std::move(p));
 }
 
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -12500,6 +12517,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ACC:
+    case GGML_OP_SET:
         ggml_vk_acc(ctx, compute_ctx, src0, src1, node);
 
         break;
@@ -14896,8 +14914,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return true;
         case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_L2_NORM:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_L2_NORM:
+            return ggml_is_contiguous_rows(op->src[0]) &&
+                   op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_ADD:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
@@ -14960,7 +14980,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             }
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ACC:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_SET:
+            return op->src[0]->type == op->src[1]->type && op->src[0]->type == op->type &&
+                   (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_I32);
         case GGML_OP_CONCAT:
             return ggml_type_size(op->src[0]->type) == ggml_type_size(GGML_TYPE_F32);
         case GGML_OP_ADD1:
@@ -15611,6 +15634,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]);
         } else if (tensor->op == GGML_OP_ACC) {
             tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]);
+        } else if (tensor->op == GGML_OP_SET) {
+            tensor_clone = ggml_set(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]);
         } else if (tensor->op == GGML_OP_NORM) {
             tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
         } else if (tensor->op == GGML_OP_GROUP_NORM) {

ggml/src/ggml-vulkan/vulkan-shaders/acc.comp

@@ -3,6 +3,9 @@
 #include "types.glsl"
 #include "generic_binary_head.glsl"
 
+// false for SET, true for ACC
+layout(constant_id = 1) const bool ACC = true;
+
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 void main() {
@@ -13,17 +16,22 @@ void main() {
 
     const uint offset = p.param3;
     const uint src1_i = idx - offset;
-    const uint oz = src1_i / p.nb02;
-    const uint oy = (src1_i - (oz * p.nb02)) / p.nb01;
-    const uint ox = src1_i % p.nb01;
+    const uint i3 = src1_i / p.nb03;
+    const uint rem2 = src1_i - i3 * p.nb03;
+    const uint i2 = rem2 / p.nb02;
+    const uint rem1 = rem2 - i2 * p.nb02;
+    const uint i1 = rem1 / p.nb01;
+    const uint i0 = rem1 % p.nb01;
 
     uint i00, i01, i02, i03;
-    get_indices(idx, i00, i01, i02, i03);
 
-    if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+    if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
+        if (ACC) {
+            data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)]));
+        } else {
+            data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)]));
+        }
     } else {
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]));
     }
 }
-

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -130,6 +130,7 @@ void main() {
         if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
             bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
 
+            float max_mask = NEG_FLT_MAX_OVER_2;
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
@@ -137,12 +138,25 @@ void main() {
                     if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
                         float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
                         masksh[c][r] = m;
+                        max_mask = max(max_mask, m);
                     } else {
                         masksh[c][r] = float(0);
                     }
                 }
             }
+            // skip the block if the mask is entirely -inf
+            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
             barrier();
+            if (gl_SubgroupInvocationID == 0) {
+                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+            }
+            barrier();
+            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                max_mask = max(max_mask, tmpsh[s]);
+            }
+            if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                continue;
+            }
         }
 
         float Sf[Br][cols_per_thread];
@@ -260,6 +274,9 @@ void main() {
         barrier();
     }
 
+    // prevent race on tmpsh
+    barrier();
+
     // reduce across threads
 
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -42,6 +42,8 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
+shared float tmpsh[row_split];
+
 const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
@@ -213,6 +215,19 @@ void main() {
                         }
                     }
                 }
+                // skip the block if the mask is entirely -inf
+                bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+                barrier();
+                if (gl_SubgroupInvocationID == 0) {
+                    tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+                }
+                barrier();
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    max_mask = max(max_mask, tmpsh[s]);
+                }
+                if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
             }
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -176,7 +176,14 @@ void main() {
                     tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
                     tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
 
+                    coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
+
                     coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+                    // skip the block if the mask is entirely -inf
+                    coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                    if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                        continue;
+                    }
                 } else {
                     tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
                     // Don't clamp against nem1 when GQA is enabled
@@ -184,7 +191,14 @@ void main() {
                     tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
                     tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
 
+                    coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
+
                     coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+                    // skip the block if the mask is entirely -inf
+                    coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                    if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                        continue;
+                    }
                 }
             }
         }

ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "generic_head.glsl"
+#include "generic_unary_head.glsl"
 #include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
@@ -8,19 +8,22 @@
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
-
 shared FLOAT_TYPE sum[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
+    const uint i3 = row / (p.ne11 * p.ne12);
+    const uint i3_offset = i3 * p.ne12 * p.ne11;
+    const uint i2 = (row - i3_offset) / p.ne11;
+    const uint i2_offset = i2 * p.ne11;
+    const uint i1 = row - i3_offset - i2_offset;
+
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
         sum[tid] += xi * xi;
     }
 
@@ -35,7 +38,7 @@ void main() {
 
     const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1)));
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
     }
 }

ggml/src/ggml.c

@@ -5749,7 +5749,7 @@ static struct ggml_tensor * ggml_unary_impl(
         struct ggml_tensor  * a,
         enum ggml_unary_op    op,
         bool                  inplace) {
-    GGML_ASSERT(ggml_is_contiguous_1(a));
+    GGML_ASSERT(ggml_is_contiguous_rows(a));
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 

gguf-py/gguf/constants.py

@@ -181,6 +181,11 @@ class Keys:
         SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
         TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
 
+        class Indexer:
+            HEAD_COUNT = "{arch}.attention.indexer.head_count"
+            KEY_LENGTH = "{arch}.attention.indexer.key_length"
+            TOP_K      = "{arch}.attention.indexer.top_k"
+
     class Rope:
         DIMENSION_COUNT           = "{arch}.rope.dimension_count"
         DIMENSION_SECTIONS        = "{arch}.rope.dimension_sections"
@@ -425,6 +430,7 @@ class MODEL_ARCH(IntEnum):
     CHATGLM          = auto()
     GLM4             = auto()
     GLM4_MOE         = auto()
+    GLM_DSA          = auto()
     BITNET           = auto()
     T5               = auto()
     T5ENCODER        = auto()
@@ -670,6 +676,10 @@ class MODEL_TENSOR(IntEnum):
     VISEXP_GATE          = auto()
     VISEXP_DOWN          = auto()
     VISEXP_UP            = auto()
+    INDEXER_K_NORM       = auto()
+    INDEXER_PROJ         = auto()
+    INDEXER_ATTN_K       = auto()
+    INDEXER_ATTN_Q_B     = auto()
     # vision
     V_MMPROJ             = auto()
     V_MMPROJ_FC          = auto()
@@ -858,6 +868,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.CHATGLM:          "chatglm",
     MODEL_ARCH.GLM4:             "glm4",
     MODEL_ARCH.GLM4_MOE:         "glm4moe",
+    MODEL_ARCH.GLM_DSA:          "glm-dsa",
     MODEL_ARCH.BITNET:           "bitnet",
     MODEL_ARCH.T5:               "t5",
     MODEL_ARCH.T5ENCODER:        "t5encoder",
@@ -1101,6 +1112,10 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.VISEXP_GATE:               "blk.{bid}.vis_gate",
     MODEL_TENSOR.VISEXP_DOWN:               "blk.{bid}.vis_down",
     MODEL_TENSOR.VISEXP_UP:                 "blk.{bid}.vis_up",
+    MODEL_TENSOR.INDEXER_K_NORM:            "blk.{bid}.indexer.k_norm",
+    MODEL_TENSOR.INDEXER_PROJ:              "blk.{bid}.indexer.proj",
+    MODEL_TENSOR.INDEXER_ATTN_K:            "blk.{bid}.indexer.attn_k",
+    MODEL_TENSOR.INDEXER_ATTN_Q_B:          "blk.{bid}.indexer.attn_q_b",
     # vision
     MODEL_TENSOR.V_MMPROJ:                  "mm.{bid}",
     MODEL_TENSOR.V_MMPROJ_FC:               "mm.model.fc",
@@ -2677,6 +2692,47 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
         MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
+    MODEL_ARCH.GLM_DSA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.INDEXER_K_NORM,
+        MODEL_TENSOR.INDEXER_PROJ,
+        MODEL_TENSOR.INDEXER_ATTN_K,
+        MODEL_TENSOR.INDEXER_ATTN_Q_B,
+        # NextN/MTP tensors - preserved but unused
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+    ],
     MODEL_ARCH.BITNET: [
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,
@@ -3766,6 +3822,7 @@ class VisionProjectorType:
     VOXTRAL = "voxtral"
     LFM2 = "lfm2"
     KIMIVL = "kimivl"
+    KIMIK25 = "kimik25"
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"

gguf-py/gguf/gguf_writer.py

@@ -771,6 +771,15 @@ class GGUFWriter:
     def add_value_length_mla(self, length: int) -> None:
         self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
 
+    def add_indexer_head_count(self, count: int) -> None:
+        self.add_uint32(Keys.Attention.Indexer.HEAD_COUNT.format(arch=self.arch), count)
+
+    def add_indexer_key_length(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.Indexer.KEY_LENGTH.format(arch=self.arch), length)
+
+    def add_indexer_top_k(self, top_k: int) -> None:
+        self.add_uint32(Keys.Attention.Indexer.TOP_K.format(arch=self.arch), top_k)
+
     def add_max_alibi_bias(self, bias: float) -> None:
         self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
 

gguf-py/gguf/tensor_mapping.py

@@ -1206,6 +1206,22 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.vision_expert_query_key_value",  # cogvlm
         ),
 
+        MODEL_TENSOR.INDEXER_K_NORM: (
+            "model.layers.{bid}.self_attn.indexer.k_norm", # DSA
+        ),
+
+        MODEL_TENSOR.INDEXER_PROJ: (
+            "model.layers.{bid}.self_attn.indexer.weights_proj", # DSA
+        ),
+
+        MODEL_TENSOR.INDEXER_ATTN_K: (
+            "model.layers.{bid}.self_attn.indexer.wk", # DSA
+        ),
+
+        MODEL_TENSOR.INDEXER_ATTN_Q_B: (
+            "model.layers.{bid}.self_attn.indexer.wq_b", # DSA
+        ),
+
         ############################################################################
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
@@ -1303,6 +1319,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ: (
             "multi_modal_projector.linear_{bid}",
+            "mm_projector.proj.linear_{bid}", # Kimi-K2.5
             "visual.merger.mlp.{bid}", # qwen2vl
             "merger.mlp.{bid}",
         ),
@@ -1364,6 +1381,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
             "visual.blocks.{bid}.attn.qkv", # qwen3vl
             "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm
+            "vision_tower.encoder.blocks.{bid}.wqkv" # Kimi-K2.5
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q: (
@@ -1538,6 +1556,7 @@ class TensorNameMap:
             "multi_modal_projector.norm",
             "multi_modal_projector.layer_norm",
             "multi_modal_projector.pre_norm",
+            "mm_projector.pre_norm", # Kimi-K2.5
             "pre_mm_projector_norm",
             "model.vision.linear_proj.norm1", # cogvlm
             "merger.ln_q",

include/llama.h

@@ -482,7 +482,7 @@ extern "C" {
     enum llama_params_fit_status {
         LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
         LLAMA_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
-        LLAMA_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occured, e.g. because no model could be found at the specified path
+        LLAMA_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
     };
 
     // fits mparams and cparams to free device memory (assumes system memory is unlimited)
@@ -1150,9 +1150,9 @@ extern "C" {
     //
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
-    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
+    ///
     /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-    /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
+    /// @param tmpl A Jinja template to use for this chat.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat
     /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.

scripts/pr2wt.sh

@@ -30,12 +30,18 @@ fi
 PR=$1
 [[ "$PR" =~ ^[0-9]+$ ]] || { echo "error: PR number must be numeric"; exit 1; }
 
+url_origin=$(git config --get remote.upstream.url 2>/dev/null) || \
 url_origin=$(git config --get remote.origin.url) || {
-    echo "error: no remote named 'origin' in this repository"
+    echo "error: no remote named 'upstream' or 'origin' in this repository"
     exit 1
 }
 
-org_repo=$(echo $url_origin | cut -d/ -f4-)
+# Extract org/repo from either https or ssh format.
+if [[ $url_origin =~ ^git@ ]]; then
+    org_repo=$(echo $url_origin | cut -d: -f2)
+else
+    org_repo=$(echo $url_origin | cut -d/ -f4-)
+fi
 org_repo=${org_repo%.git}
 
 echo "org/repo: $org_repo"

scripts/sync_vendor.py

@@ -1,6 +1,11 @@
 #!/usr/bin/env python3
 
 import urllib.request
+import os
+import sys
+import subprocess
+
+HTTPLIB_VERSION = "f80864ca031932351abef49b74097c67f14719c6"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -12,8 +17,9 @@ vendor = {
     # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
     "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
 
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/httplib.h": "vendor/cpp-httplib/httplib.h",
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/LICENSE":   "vendor/cpp-httplib/LICENSE",
+    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/httplib.h": "httplib.h",
+    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/split.py":  "split.py",
+    f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/LICENSE":   "vendor/cpp-httplib/LICENSE",
 
     "https://raw.githubusercontent.com/sheredom/subprocess.h/b49c56e9fe214488493021017bf3954b91c7c1f5/subprocess.h": "vendor/sheredom/subprocess.h",
 }
@@ -22,19 +28,16 @@ for url, filename in vendor.items():
     print(f"downloading {url} to {filename}") # noqa: NP100
     urllib.request.urlretrieve(url, filename)
 
-    # split cpp/h files for httplib
-    # see: https://github.com/yhirose/cpp-httplib/blob/master/split.py
-    if 'httplib.h' in filename:
-        border = '// ----------------------------------------------------------------------------'
-        with open(filename, 'r') as f:
-            content = f.read()
-        header, implementation, footer = content.split(border, 2)
-        fname_cpp = filename.replace('.h', '.cpp')
-        with open(filename, 'w') as fh:
-            fh.write(header)
-            fh.write(footer)
-        with open(fname_cpp, 'w') as fc:
-            fc.write('#include "httplib.h"\n')
-            fc.write('namespace httplib {\n')
-            fc.write(implementation.replace('\ninline ', '\n'))
-            fc.write('} // namespace httplib\n')
+print("Splitting httplib.h...") # noqa: NP100
+try:
+    subprocess.check_call([
+        sys.executable, "split.py",
+        "--extension", "cpp",
+        "--out", "vendor/cpp-httplib"
+    ])
+except Exception as e:
+    print(f"Error: {e}") # noqa: NP100
+    sys.exit(1)
+finally:
+    os.remove("split.py")
+    os.remove("httplib.h")

src/llama-arch.cpp

@@ -74,6 +74,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
+    { LLM_ARCH_GLM_DSA,          "glm-dsa"          },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -225,6 +226,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
+    { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
+    { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
+    { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,           "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,        "%s.rope.dimension_sections"              },
@@ -516,6 +520,10 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_VISEXP_FFN_GATE,                        "blk.%d.vis_gate" },
     { LLM_TENSOR_VISEXP_FFN_DOWN,                        "blk.%d.vis_down" },
     { LLM_TENSOR_VISEXP_FFN_UP,                          "blk.%d.vis_up" },
+    { LLM_TENSOR_INDEXER_K_NORM,                         "blk.%d.indexer.k_norm" },
+    { LLM_TENSOR_INDEXER_PROJ,                           "blk.%d.indexer.proj" },
+    { LLM_TENSOR_INDEXER_ATTN_K,                         "blk.%d.indexer.attn_k" },
+    { LLM_TENSOR_INDEXER_ATTN_Q_B,                       "blk.%d.indexer.attn_q_b" },
 };
 
 static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
@@ -1657,6 +1665,46 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
                 LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
             };
+        case LLM_ARCH_GLM_DSA:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q_A_NORM,
+                LLM_TENSOR_ATTN_KV_A_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_Q_A,
+                LLM_TENSOR_ATTN_Q_B,
+                LLM_TENSOR_ATTN_KV_A_MQA,
+                LLM_TENSOR_ATTN_KV_B,
+                LLM_TENSOR_ATTN_K_B,
+                LLM_TENSOR_ATTN_V_B,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_GATE_INP,
+                LLM_TENSOR_FFN_GATE_EXPS,
+                LLM_TENSOR_FFN_DOWN_EXPS,
+                LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_INP_SHEXP,
+                LLM_TENSOR_FFN_GATE_SHEXP,
+                LLM_TENSOR_FFN_DOWN_SHEXP,
+                LLM_TENSOR_FFN_UP_SHEXP,
+                LLM_TENSOR_FFN_EXP_PROBS_B,
+                LLM_TENSOR_INDEXER_K_NORM,
+                LLM_TENSOR_INDEXER_PROJ,
+                LLM_TENSOR_INDEXER_ATTN_K,
+                LLM_TENSOR_INDEXER_ATTN_Q_B,
+                LLM_TENSOR_NEXTN_EH_PROJ,
+                LLM_TENSOR_NEXTN_EMBED_TOKENS,
+                LLM_TENSOR_NEXTN_ENORM,
+                LLM_TENSOR_NEXTN_HNORM,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
+            };
         case LLM_ARCH_BITNET:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -2643,6 +2691,10 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_VISEXP_FFN_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_VISEXP_FFN_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_K_NORM,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_INDEXER_PROJ,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_ATTN_K,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
     {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},

src/llama-arch.h

@@ -78,6 +78,7 @@ enum llm_arch {
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
     LLM_ARCH_GLM4_MOE,
+    LLM_ARCH_GLM_DSA,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -229,6 +230,9 @@ enum llm_kv {
     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
+    LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
+    LLM_KV_ATTENTION_INDEXER_TOP_K,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -517,6 +521,10 @@ enum llm_tensor {
     LLM_TENSOR_VISEXP_FFN_GATE,
     LLM_TENSOR_VISEXP_FFN_DOWN,
     LLM_TENSOR_VISEXP_FFN_UP,
+    LLM_TENSOR_INDEXER_K_NORM,
+    LLM_TENSOR_INDEXER_PROJ,
+    LLM_TENSOR_INDEXER_ATTN_K,
+    LLM_TENSOR_INDEXER_ATTN_Q_B,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,

src/llama-hparams.h

@@ -193,6 +193,11 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // DSA (deepseek sparse attention)
+    uint32_t indexer_n_head    = 0;
+    uint32_t indexer_head_size = 0;
+    uint32_t indexer_top_k     = 0;
+
     // qwen3vl deepstack
     uint32_t n_deepstack_layers = 0;
 

src/llama-mmap.cpp

@@ -504,6 +504,8 @@ struct llama_mmap::impl {
         }
     }
 #elif defined(_WIN32)
+    HANDLE hMapping = nullptr;
+
     impl(struct llama_file * file, size_t prefetch, bool numa) {
         GGML_UNUSED(numa);
 
@@ -511,7 +513,7 @@ struct llama_mmap::impl {
 
         HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
 
-        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
 
         if (hMapping == NULL) {
             DWORD error = GetLastError();
@@ -520,9 +522,9 @@ struct llama_mmap::impl {
 
         addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
         DWORD error = GetLastError();
-        CloseHandle(hMapping);
 
         if (addr == NULL) {
+            CloseHandle(hMapping);
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
@@ -554,9 +556,17 @@ struct llama_mmap::impl {
     }
 
     ~impl() {
-        if (!UnmapViewOfFile(addr)) {
-            LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
+        if (hMapping) {
+            if (addr) {
+                if (!UnmapViewOfFile(addr)) {
+                    LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
+                            llama_format_win_err(GetLastError()).c_str());
+                }
+            }
+            if (!CloseHandle(hMapping)) {
+                LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
         }
     }
 #else

src/llama-model.cpp

@@ -137,6 +137,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
         case LLM_TYPE_310B_A15B:     return "310B.A15B";
         case LLM_TYPE_355B_A32B:     return "355B.A32B";
+        case LLM_TYPE_744B_A40B:     return "744B.A40B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
@@ -1822,6 +1823,50 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM_DSA:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+                // MoE parameters
+                ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
+                ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+                // deepseek MLA parameters
+                ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,      hparams.n_lora_q);
+                ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+
+                // DSA parameters
+                ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
+                ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
+                ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k);
+
+                // Expert gating function (GLM-4.5 uses sigmoid)
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+                    hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                }
+
+                // NextN/MTP parameters
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+
+                // TODO: when MTP is implemented, this should probably be updated if needed
+                hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+                switch (hparams.n_layer) {
+                    case 79: type = LLM_TYPE_744B_A40B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5492,6 +5537,108 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 }
                 break;
+            case LLM_ARCH_GLM_DSA:
+                {
+                    const bool is_mla = hparams.is_mla();
+                    if (!is_mla) {
+                        throw std::runtime_error("GLM_DSA architecture requires MLA");
+                    }
+
+                    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+                    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+                    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+                    const int64_t q_lora_rank  = hparams.n_lora_q;
+                    const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    // try to load output.weight, if not found, use token_embd (tied embeddings)
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        int flags = 0;
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            // skip all tensors in the NextN layers
+                            // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later
+                            flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED;
+                        }
+
+                        auto & layer = layers[i];
+
+                        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+                        layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);
+                        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
+
+                        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+                        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);
+
+                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
+
+                        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+                        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);
+                        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+                        // DSA indexer
+                        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
+                        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
+                        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
+                        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
+                        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+                        } else {
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+                            if (n_expert == 0) {
+                                throw std::runtime_error("n_expert must be > 0");
+                            }
+                            if (n_expert_used == 0) {
+                                throw std::runtime_error("n_expert_used must be > 0");
+                            }
+
+                            // MoE branch
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+                        }
+
+                        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+                            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+                            // Optional tensors
+                            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+                            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+                        }
+                    }
+                } break;
             case LLM_ARCH_NEMOTRON:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -7765,7 +7912,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
     }
 
-    if (arch == LLM_ARCH_DEEPSEEK2) {
+    if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA) {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
@@ -7965,7 +8112,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             cparams.n_seq_max,
                             nullptr);
                 } else if (llm_arch_is_hybrid(arch)) {
-
                     // The main difference between hybrid architectures is the
                     // layer filters, so pick the right one here
                     llama_memory_hybrid::layer_filter_cb filter_attn = nullptr;
@@ -7990,7 +8136,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             /* attn_type_v       */ params.type_v,
                             /* attn_v_trans      */ !cparams.flash_attn,
                             /* attn_swa_full     */ params.swa_full,
-                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_kv_size      */ cparams.n_ctx_seq,
                             /* attn_n_ubatch     */ cparams.n_ubatch,
                             /* attn_n_pad        */ 1,
                             /* recurrent_type_r  */ GGML_TYPE_F32,
@@ -8007,7 +8153,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             /* attn_type_k       */ params.type_k,
                             /* attn_type_v       */ params.type_v,
                             /* attn_v_trans      */ !cparams.flash_attn,
-                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_kv_size      */ cparams.n_ctx_seq,
                             /* attn_n_pad        */ 1,
                             /* attn_n_swa        */ hparams.n_swa,
                             /* attn_swa_type     */ hparams.swa_type,
@@ -8338,6 +8484,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_deepseek>(*this, params);
             } break;
         case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_GLM_DSA:
             {
                 llm = std::make_unique<llm_build_deepseek2>(*this, params);
             } break;
@@ -8739,6 +8886,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MISTRAL3:
         case LLM_ARCH_LLAMA_EMBED:
         case LLM_ARCH_MAINCODER:
+        case LLM_ARCH_GLM_DSA:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

src/llama-model.h

@@ -130,6 +130,7 @@ enum llm_type {
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
     LLM_TYPE_355B_A32B, // GLM-4.5
+    LLM_TYPE_744B_A40B, // GLM-5
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };
@@ -429,6 +430,13 @@ struct llama_layer {
     struct ggml_tensor * ssm_g_b    = nullptr;
     struct ggml_tensor * ssm_o_norm = nullptr;
 
+    // DSA (deepseek sparse attention)
+    struct ggml_tensor * indexer_k_norm   = nullptr;
+    struct ggml_tensor * indexer_k_norm_b = nullptr;
+    struct ggml_tensor * indexer_proj     = nullptr;
+    struct ggml_tensor * indexer_attn_k   = nullptr;
+    struct ggml_tensor * indexer_attn_q_b = nullptr; // note: for lora a/b, not bias
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;

src/models/deepseek2.cpp

@@ -45,7 +45,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    for (int il = 0; il < n_layer; ++il) {
+    int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < effective_n_layers; ++il) {
         ggml_tensor * inpSA = inpL;
 
         // norm
@@ -188,7 +189,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                             Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == effective_n_layers - 1 && inp_out_ids) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }

src/models/kimi-linear.cpp

@@ -41,8 +41,11 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
         conv_x->nb[1], conv_x->nb[2], n_seq_tokens * conv_x->nb[0]);
     ggml_build_forward_expand(gf,
         ggml_cpy(ctx0, last_conv_x,
-            ggml_view_1d(ctx0, conv_states_all, conv_state_size * n_seqs,
-                (kv_head * n_embd_r_total + qkv * conv_state_size) * ggml_element_size(conv_states_all))));
+            ggml_view_3d(ctx0, conv_states_all,
+                d_conv - 1, d_inner, n_seqs,
+                (d_conv - 1) * ggml_element_size(conv_states_all),           // nb1: contiguous within one channel's conv taps
+                n_embd_r_total * ggml_element_size(conv_states_all),         // nb2: stride between sequences (skip over K,V states)
+                (kv_head * n_embd_r_total + qkv * conv_state_size) * ggml_element_size(conv_states_all))));  // offset to first seq's Q/K/V state
     // Reshape conv weight: GGUF [d_conv, 1, d_inner, 1] -> ggml_ssm_conv expects [d_conv, d_inner]
     // GGUF stores as [d_conv, 1, d_inner, 1] with memory layout w[conv_step + channel * d_conv]
     // vLLM stores as [d_inner, d_conv] with memory layout w[channel * d_conv + conv_step]

src/unicode.cpp

@@ -1,16 +1,10 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "unicode.h"
 #include "unicode-data.h"
 
 #include <algorithm>
 #include <cassert>
-#include <codecvt>
 #include <cstddef>
 #include <cstdint>
-#include <locale>
 #include <map>
 #include <regex>
 #include <stdexcept>
@@ -199,27 +193,6 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
     return map;
 }
 
-static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
-#if defined(__clang__)
-    // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-    return conv.from_bytes(s);
-}
-
 static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
     std::vector<std::string> bpe_encoded_words;
     for (const auto & word : bpe_words) {
@@ -1028,10 +1001,10 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                     break;
                 }
             }
+            const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
 
             if (use_collapsed) {
                 // sanity-check that the original regex does not contain any non-ASCII characters
-                const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
                 for (size_t i = 0; i < cpts_regex.size(); ++i) {
                     if (cpts_regex[i] >= 128) {
                         throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");
@@ -1087,7 +1060,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
             } else {
                 // no unicode category used, we can use std::wregex directly
-                const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
+                std::wstring wregex_expr(cpts_regex.begin(), cpts_regex.end());
 
                 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
                 std::wstring wtext(cpts.begin(), cpts.end());

tests/test-backend-ops.cpp

@@ -1943,7 +1943,11 @@ struct test_unary : public test_case {
 
         ggml_tensor * a;
         if (v & 1) {
-            auto ne = ne_a; ne[0] *= 3;
+            auto ne = ne_a;
+            ne[0] *= 3;
+            ne[1] *= 2;
+            ne[2] *= 5;
+            ne[3] *= 4;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (grad_supported) {
                 ggml_set_param(a);
@@ -2782,9 +2786,10 @@ struct test_set : public test_case {
     const ggml_type type_dst;
     const std::array<int64_t, 4> ne;
     const int dim;
+    const bool inplace;
 
     std::string vars() override {
-        return VARS_TO_STR4(type_src, type_dst, ne, dim);
+        return VARS_TO_STR5(type_src, type_dst, ne, dim, inplace);
     }
 
     size_t op_size(ggml_tensor * t) override {
@@ -2792,8 +2797,8 @@ struct test_set : public test_case {
     }
 
     test_set(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {6, 5, 4, 3}, int dim = 1)
-        : type_src(type_src), type_dst(type_dst), ne(ne), dim(dim) {}
+            std::array<int64_t, 4> ne = {6, 5, 4, 3}, int dim = 1, bool inplace = false)
+        : type_src(type_src), type_dst(type_dst), ne(ne), dim(dim), inplace(inplace) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -2804,7 +2809,7 @@ struct test_set : public test_case {
         for (int i = 0; i < dim; ++i) {
             ne_dst[i] *= 2;
         }
-        ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data());
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data());
         ggml_set_param(dst);
         ggml_set_name(dst, "dst");
 
@@ -2812,9 +2817,16 @@ struct test_set : public test_case {
         for (int i = 0; i < dim; ++i) {
             offset += ((ne_dst[i] - ne[i])/2)*dst->nb[i];
         }
-        ggml_tensor * out = ggml_set(ctx, dst, src,
-            // The backward pass requires setting a contiguous region:
-            src->nb[1], src->nb[2], src->nb[3], offset);
+        ggml_tensor * out;
+        if (inplace) {
+            out = ggml_set_inplace(ctx, dst, src,
+                    // The backward pass requires setting a contiguous region:
+                    src->nb[1], src->nb[2], src->nb[3], offset);
+        } else {
+            out = ggml_set(ctx, dst, src,
+                    // The backward pass requires setting a contiguous region:
+                    src->nb[1], src->nb[2], src->nb[3], offset);
+        }
         ggml_set_name(out, "out");
 
         return out;
@@ -5809,20 +5821,27 @@ struct test_l2_norm : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const float eps;
+    bool v;
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR4(type, ne, eps, v);
     }
 
     test_l2_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 64, 320, 1},
-            float eps = 1e-12f)
-        : type(type), ne(ne), eps(eps) {}
+            float eps = 1e-12f,
+            bool v = false)
+        : type(type), ne(ne), eps(eps), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_name(a, "a");
 
+        if (v) {
+            a = ggml_view_4d(ctx, a, a->ne[0]/2, a->ne[1]/2, a->ne[2]/2, a->ne[3]/2, a->nb[1], a->nb[2], a->nb[3], 0);
+            ggml_set_name(a, "view of a");
+        }
+
         ggml_tensor * out = ggml_l2_norm(ctx, a, eps);
         ggml_set_name(out, "out");
 
@@ -5835,26 +5854,46 @@ struct test_acc : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne_a;
     const std::array<int64_t, 4> ne_b;
+    const int64_t stride_dim;
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne_a, ne_b);
+        return VARS_TO_STR4(type, ne_a, ne_b, stride_dim);
     }
 
     test_acc(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne_a = {256, 17, 1, 1},
-            std::array<int64_t, 4> ne_b = {256, 16, 1, 1})
-        : type(type), ne_a(ne_a), ne_b(ne_b) {}
+            std::array<int64_t, 4> ne_a = {256, 17, 2, 3},
+            std::array<int64_t, 4> ne_b = {256, 16, 2, 3},
+            uint64_t stride_dim = -1)
+        : type(type), ne_a(ne_a), ne_b(ne_b), stride_dim(stride_dim) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
         ggml_set_param(a);
         ggml_set_name(a, "a");
 
-        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
-        ggml_set_param(b);
+        ggml_tensor * b;
+        if (stride_dim == 1 || stride_dim == 2 || stride_dim == 3) {
+            // Create a larger tensor and take a view at a non-zero offset.
+            // This tests that the backend correctly handles b's data offset
+            std::array<int64_t, 4> ne_b_pad = {ne_b[0], ne_b[1], ne_b[2], ne_b[3]};
+            ne_b_pad[stride_dim] += 1;
+            ggml_tensor * b_pad = ggml_new_tensor(ctx, type, 4, ne_b_pad.data());
+            ggml_set_param(b_pad);
+            ggml_set_name(b_pad, "b_pad");
+            // View that skips the first row, so b has a non-zero byte offset
+            b = ggml_view_4d(ctx, b_pad,
+                ne_b[0], ne_b[1], ne_b[2], ne_b[3],
+                b_pad->nb[1], b_pad->nb[2], b_pad->nb[3],
+                b_pad->nb[1]);
+        } else {
+            b = ggml_new_tensor(ctx, type, 4, ne_b.data());
+            ggml_set_param(b);
+        }
         ggml_set_name(b, "b");
 
-        ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]);
+        // When ne_b[0] < ne_a[0], a->nb[1] != b->nb[1], so the stride
+        // parameters to ggml_acc don't match b's natural stride.
+        ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], 0);
         ggml_set_name(out, "out");
 
         return out;
@@ -7424,11 +7463,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10,  8, 3, 1}, {1, 2, 0, 3}));
 
     for (int dim = 1; dim < GGML_MAX_DIMS; ++dim) {
-        test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim));
+        test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim, false));
+        test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim, true));
     }
 
     for (int dim = 1; dim < GGML_MAX_DIMS; ++dim) {
-        test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim));
+        test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim, false));
+        test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim, true));
     }
 
     // same-type copy
@@ -7562,7 +7603,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                 test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
             }
             test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, { n, 5, 4, 3 }, eps));
-            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps));
+            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false));
+            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
         }
     }
 
@@ -8128,29 +8170,40 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     test_cases.emplace_back(new test_sum());
-    test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_sum(GGML_TYPE_F32, {11, 5, 6, 3}, {0, 2, 1, 3}));  // row-contiguous but non-contiguous
     test_cases.emplace_back(new test_sum(GGML_TYPE_F32, {11, 5, 6, 3}, {0, 3, 2, 1}));
     test_cases.emplace_back(new test_sum(GGML_TYPE_F32, {11, 5, 6, 3}, {0, 1, 3, 2}));
+    test_cases.emplace_back(new test_mean());
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 256, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32769, 1, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32, 1, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32, 256, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32768, 1, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1024, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 256, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 256, 1, 1 }, { 1, 0, 2, 3 })); // sum dst not-contiguous
+    test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, false));
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, false, true));
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, true));
-    test_cases.emplace_back(new test_mean());
-    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 16, 5, 6, 3 }, true, false));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 16, 5, 6, 3 }, false, true));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 16, 5, 6, 3 }, true, true));
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 1, 1, 1 }));
-    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 1, 1, 1 }));
-    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1024, 1, 1 }));
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 1024, 1, 1 }));
-    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 256, 1, 1 }));
-    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 256, 1, 1 }, { 1, 0, 2, 3 })); // sum dst not-contiguous
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 256, 1, 1 }));
-    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 256, 1, 1 }));
-    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32769, 1, 1, 1 }));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_group_norm_mul_add(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm_mul_add(GGML_TYPE_F32, {9, 9, 1280, 1}));
-    test_cases.emplace_back(new test_acc());
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 1, 1}, {256, 16, 1, 1}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {256, 16, 2, 3}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {128, 16, 2, 3}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {256, 16, 2, 3}, 1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {128, 16, 2, 3}, 2));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {64, 16, 2, 3}, 3));
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_pad(GGML_TYPE_F32, {33, 17, 2, 1}, 4, 3, true)); // circular
     test_cases.emplace_back(new test_pad_ext());
@@ -8585,6 +8638,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 48, 1, 512, 1)); // prefill
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 48, 1, 1,   1)); // generate
 
+    // acc
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 1, 1}, {256, 16, 1, 1}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {256, 16, 2, 3}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {128, 16, 2, 3}, -1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {256, 16, 2, 3}, 1));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {128, 16, 2, 3}, 2));
+    test_cases.emplace_back(new test_acc(GGML_TYPE_F32, {256, 17, 2, 3}, {64, 16, 2, 3}, 3));
+
     return test_cases;
 }
 

tools/cli/cli.cpp

@@ -52,6 +52,7 @@ struct cli_context {
     json messages = json::array();
     std::vector<raw_buffer> input_files;
     task_params defaults;
+    bool verbose_prompt;
 
     // thread for showing "loading" animation
     std::atomic<bool> loading_show;
@@ -66,6 +67,8 @@ struct cli_context {
         defaults.stream = true; // make sure we always use streaming mode
         defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
         // defaults.return_progress = true; // TODO: show progress
+
+        verbose_prompt = params.verbose_prompt;
     }
 
     std::string generate_completion(result_timings & out_timings) {
@@ -91,6 +94,12 @@ struct cli_context {
             rd.post_task({std::move(task)});
         }
 
+        if (verbose_prompt) {
+            console::set_display(DISPLAY_TYPE_PROMPT);
+            console::log("%s\n\n", chat_params.prompt.c_str());
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+
         // wait for first result
         console::spinner::start();
         server_task_result_ptr result = rd.next(should_stop);

tools/mtmd/CMakeLists.txt

@@ -19,6 +19,7 @@ add_library(mtmd
             models/glm4v.cpp
             models/internvl.cpp
             models/kimivl.cpp
+            models/kimik25.cpp
             models/llama4.cpp
             models/llava.cpp
             models/minicpmv.cpp

tools/mtmd/clip-impl.h

@@ -235,6 +235,7 @@ enum projector_type {
     PROJECTOR_TYPE_LFM2A,
     PROJECTOR_TYPE_GLM4V,
     PROJECTOR_TYPE_YOUTUVL,
+    PROJECTOR_TYPE_KIMIK25,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -268,6 +269,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LFM2A,     "lfm2a"},
     { PROJECTOR_TYPE_GLM4V,     "glm4v"},
     { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
+    { PROJECTOR_TYPE_KIMIK25,   "kimik25"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip.cpp

@@ -673,8 +673,8 @@ ggml_tensor * clip_graph::build_rope_2d(
     {
         first = ggml_view_3d(ctx0, cur,
             n_dim/2, n_head, n_pos,
-            ggml_row_size(cur->type, n_dim),
-            ggml_row_size(cur->type, n_dim*n_head),
+            cur->nb[1],
+            cur->nb[2],
             0);
         first = ggml_rope_ext(
             ctx0,
@@ -692,8 +692,8 @@ ggml_tensor * clip_graph::build_rope_2d(
     {
         second = ggml_view_3d(ctx0, cur,
             n_dim/2, n_head, n_pos,
-            ggml_row_size(cur->type, n_dim),
-            ggml_row_size(cur->type, n_dim*n_head),
+            cur->nb[1],
+            cur->nb[2],
             n_dim/2 * ggml_element_size(cur));
         second = ggml_rope_ext(
             ctx0,
@@ -826,6 +826,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_kimivl>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_KIMIK25:
+            {
+                builder = std::make_unique<clip_graph_kimik25>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_COGVLM:
             {
                 builder = std::make_unique<clip_graph_cogvlm>(ctx, img);
@@ -1139,6 +1143,22 @@ struct clip_model_loader {
                         hparams.set_limit_image_tokens(8, 1024);
                         hparams.set_warmup_n_tokens(256); // avoid OOM on warmup
                     } break;
+                case PROJECTOR_TYPE_KIMIK25:
+                    {
+                        hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+
+                        int min_pixels = 0, max_pixels = 0;
+                        get_u32(KEY_IMAGE_MIN_PIXELS, min_pixels, false);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, max_pixels, false);
+                        if (min_pixels > 0 && max_pixels > 0) {
+                            hparams.image_min_pixels = min_pixels;
+                            hparams.image_max_pixels = max_pixels;
+                            hparams.warmup_image_size = static_cast<int>(std::sqrt(max_pixels));
+                        } else {
+                            hparams.set_limit_image_tokens(2, 4096);
+                        }
+                    } break;
                 case PROJECTOR_TYPE_GEMMA3:
                     {
                         // default value (used by all model sizes in gemma 3 family)
@@ -1668,6 +1688,7 @@ struct clip_model_loader {
                     model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                 } break;
             case PROJECTOR_TYPE_KIMIVL:
+            case PROJECTOR_TYPE_KIMIK25:
                 {
                     model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
                     model.mm_input_norm_b = get_tensor(TN_MM_INP_NORM_B);
@@ -3165,6 +3186,23 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 res_imgs->entries.push_back(std::move(res));
             } break;
 
+        case PROJECTOR_TYPE_KIMIK25:
+            {
+                GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
+                const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
+                    original_size,
+                    params.patch_size * params.n_merge,
+                    params.image_min_pixels,
+                    params.image_max_pixels);
+                const std::array<uint8_t, 3> pad_color = {0, 0, 0};
+
+                clip_image_u8 resized_img;
+                img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BICUBIC, true, pad_color);
+                clip_image_f32_ptr res(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(res));
+            } break;
+
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_MLP_NORM:
         case PROJECTOR_TYPE_LDP:
@@ -3373,6 +3411,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
             } break;
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
+        case PROJECTOR_TYPE_KIMIK25:
             {
                 // dynamic size
                 int out_patch_size = params.patch_size * ctx->model.hparams.n_merge;
@@ -3714,6 +3753,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_KIMIVL:
+        case PROJECTOR_TYPE_KIMIK25:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
                 // set the 2D positions
@@ -3850,6 +3890,47 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
     }
 
+    // Debug: dump final embeddings if MTMD_DEBUG_EMBEDDINGS is set
+    if (std::getenv("MTMD_DEBUG_EMBEDDINGS") != nullptr) {
+        const int64_t n_embd = embeddings->ne[0];
+        const int64_t n_tokens = embeddings->ne[1];
+        std::vector<float> emb_data(n_embd * n_tokens);
+        ggml_backend_tensor_get(embeddings, emb_data.data(), 0, ggml_nbytes(embeddings));
+
+        LOG_INF("\n=== MTMD_DEBUG_EMBEDDINGS ===\n");
+        LOG_INF("Shape: [%lld, %lld]\n", (long long)n_embd, (long long)n_tokens);
+
+        // Print first few values of first token
+        LOG_INF("Token 0 (first 16 values): ");
+        for (int i = 0; i < std::min((int64_t)16, n_embd); i++) {
+            LOG_INF("%.6f ", emb_data[i]);
+        }
+        LOG_INF("\n");
+
+        // Print last few values of first token
+        if (n_embd > 16) {
+            LOG_INF("Token 0 (last 16 values):  ");
+            for (int64_t i = n_embd - 16; i < n_embd; i++) {
+                LOG_INF("%.6f ", emb_data[i]);
+            }
+            LOG_INF("\n");
+        }
+
+        // Compute and print statistics
+        float sum = 0.0f, sum_sq = 0.0f, min_val = emb_data[0], max_val = emb_data[0];
+        for (size_t i = 0; i < emb_data.size(); i++) {
+            sum += emb_data[i];
+            sum_sq += emb_data[i] * emb_data[i];
+            min_val = std::min(min_val, emb_data[i]);
+            max_val = std::max(max_val, emb_data[i]);
+        }
+        float mean = sum / emb_data.size();
+        float variance = (sum_sq / emb_data.size()) - (mean * mean);
+        LOG_INF("Stats: mean=%.6f, std=%.6f, min=%.6f, max=%.6f, sum=%.6f\n",
+                mean, sqrtf(variance), min_val, max_val, sum);
+        LOG_INF("=== END MTMD_DEBUG_EMBEDDINGS ===\n\n");
+    }
+
     return true;
 }
 
@@ -3896,6 +3977,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
+        case PROJECTOR_TYPE_KIMIK25:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_COGVLM:
             return ctx->model.mm_4h_to_h_w->ne[1];

tools/mtmd/models/kimik25.cpp

@@ -0,0 +1,101 @@
+#include "models.h"
+#include <cstring>
+#include <cmath>
+
+// note: this is similar to clip_graph::resize_position_embeddings, major difference is having
+// the w/h in ne[1] and ne[2] instead of assuming with sqrt. Could try storing the tensor in 2D instead
+// with a w*h? Also the permute is a bit different at (2, 1, 0, 3) instead of (2, 0, 1, 3).
+ggml_tensor * clip_graph_kimik25::resize_position_embeddings_3d(uint32_t interpolation_mode) {
+    ggml_tensor * pos_embd = model.position_embeddings;
+    const int height       = img.ny / patch_size;
+    const int width        = img.nx / patch_size;
+    const uint32_t mode    = interpolation_mode;
+
+    GGML_ASSERT(pos_embd);
+
+    const int64_t stored_c = pos_embd->ne[0];  // C = 1152
+    const int64_t orig_w = pos_embd->ne[1];    // W = 64
+    const int64_t orig_h = pos_embd->ne[2];    // H = 64
+
+    GGML_ASSERT(stored_c == n_embd);
+
+    if (height == (int)orig_h && width == (int)orig_w) {
+        // No interpolation needed, just flatten to [C, H*W]
+        return ggml_cont_2d(ctx0, pos_embd, n_embd, width * height);
+    }
+
+    pos_embd = ggml_permute(ctx0, pos_embd, 2, 1, 0, 3);
+    pos_embd = ggml_interpolate(ctx0, pos_embd, height, width, n_embd, 1, mode);
+    pos_embd = ggml_permute(ctx0, pos_embd, 2, 1, 0, 3);
+    pos_embd = ggml_cont_2d(ctx0, pos_embd, n_embd, width * height);
+    return pos_embd;
+}
+
+ggml_cgraph * clip_graph_kimik25::build() {
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * learned_pos_embd = resize_position_embeddings_3d(GGML_SCALE_MODE_BICUBIC);
+
+    // Kimi-K2.5 uses interleaved 2D RoPE pattern natively, but
+    // Q / K are permuted during conversion to use split format.
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        cur = build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+        return cur;
+    };
+
+    ggml_tensor * inp = build_inp();
+
+    // I don't know why, but doing this in the build_vit lead to the ggml_add not occurring?
+    // Doing it manually here does work.
+    inp = ggml_add(ctx0, inp, learned_pos_embd);
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            nullptr,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+
+    {
+        // patch_merger
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+
+        // projection norm
+        int proj_inp_dim = cur->ne[0];
+        int n_merged_patches = cur->ne[1];
+        cur = ggml_view_2d(ctx0, cur,
+            n_embd, n_merged_patches * scale_factor * scale_factor,
+            ggml_row_size(cur->type, n_embd), 0);
+        cur = ggml_norm(ctx0, cur, hparams.eps);
+        cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+        cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        cur = ggml_view_2d(ctx0, cur,
+            proj_inp_dim, n_merged_patches,
+            ggml_row_size(cur->type, proj_inp_dim), 0);
+        cb(cur, "proj_inp_normed", -1);
+
+        // projection mlp
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+
+        cb(cur, "proj_out", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/models.h

@@ -109,3 +109,10 @@ struct clip_graph_mobilenetv5 : clip_graph {
         ggml_tensor * inp,
         const mobilenetv5_block & block);
 };
+
+struct clip_graph_kimik25 : clip_graph {
+    clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
+};

tools/server/README.md

@@ -19,7 +19,7 @@ Set of LLM REST APIs and a web UI to interact with llama.cpp.
  * Speculative decoding
  * Easy-to-use web UI
 
-For the ful list of features, please refer to [server's changelog](https://github.com/ggml-org/llama.cpp/issues/9291)
+For the full list of features, please refer to [server's changelog](https://github.com/ggml-org/llama.cpp/issues/9291)
 
 ## Usage
 

tools/server/webui/docs/flows/settings-flow.md

@@ -139,6 +139,6 @@ sequenceDiagram
 
     Note over settingsStore: UI-only (not synced):
     rect rgb(255, 240, 240)
-        Note over settingsStore: systemMessage, custom (JSON)<br/>showStatistics, enableContinueGeneration<br/>autoMicOnEmpty, disableAutoScroll<br/>apiKey, pdfAsImage, disableReasoningFormat
+        Note over settingsStore: systemMessage, custom (JSON)<br/>showStatistics, enableContinueGeneration<br/>autoMicOnEmpty, disableAutoScroll<br/>apiKey, pdfAsImage, disableReasoningParsing, showRawOutputSwitch
     end
 ```

tools/server/webui/src/app.css

@@ -14,11 +14,11 @@
 	--popover-foreground: oklch(0.145 0 0);
 	--primary: oklch(0.205 0 0);
 	--primary-foreground: oklch(0.985 0 0);
-	--secondary: oklch(0.97 0 0);
+	--secondary: oklch(0.95 0 0);
 	--secondary-foreground: oklch(0.205 0 0);
 	--muted: oklch(0.97 0 0);
 	--muted-foreground: oklch(0.556 0 0);
-	--accent: oklch(0.97 0 0);
+	--accent: oklch(0.95 0 0);
 	--accent-foreground: oklch(0.205 0 0);
 	--destructive: oklch(0.577 0.245 27.325);
 	--border: oklch(0.875 0 0);
@@ -37,7 +37,7 @@
 	--sidebar-accent-foreground: oklch(0.205 0 0);
 	--sidebar-border: oklch(0.922 0 0);
 	--sidebar-ring: oklch(0.708 0 0);
-	--code-background: oklch(0.975 0 0);
+	--code-background: oklch(0.985 0 0);
 	--code-foreground: oklch(0.145 0 0);
 	--layer-popover: 1000000;
 }
@@ -51,7 +51,7 @@
 	--popover-foreground: oklch(0.985 0 0);
 	--primary: oklch(0.922 0 0);
 	--primary-foreground: oklch(0.205 0 0);
-	--secondary: oklch(0.269 0 0);
+	--secondary: oklch(0.29 0 0);
 	--secondary-foreground: oklch(0.985 0 0);
 	--muted: oklch(0.269 0 0);
 	--muted-foreground: oklch(0.708 0 0);
@@ -116,12 +116,62 @@
 	--color-sidebar-ring: var(--sidebar-ring);
 }
 
+:root {
+	--chat-form-area-height: 8rem;
+	--chat-form-area-offset: 2rem;
+	--max-message-height: max(24rem, min(80dvh, calc(100dvh - var(--chat-form-area-height) - 12rem)));
+}
+
+@media (min-width: 640px) {
+	:root {
+		--chat-form-area-height: 24rem;
+		--chat-form-area-offset: 12rem;
+	}
+}
+
 @layer base {
 	* {
 		@apply border-border outline-ring/50;
 	}
+
 	body {
 		@apply bg-background text-foreground;
+		scrollbar-width: thin;
+		scrollbar-gutter: stable;
+	}
+
+	/* Global scrollbar styling - visible only on hover */
+	* {
+		scrollbar-width: thin;
+		scrollbar-color: transparent transparent;
+		transition: scrollbar-color 0.2s ease;
+	}
+
+	*:hover {
+		scrollbar-color: hsl(var(--muted-foreground) / 0.3) transparent;
+	}
+
+	*::-webkit-scrollbar {
+		width: 6px;
+		height: 6px;
+	}
+
+	*::-webkit-scrollbar-track {
+		background: transparent;
+	}
+
+	*::-webkit-scrollbar-thumb {
+		background: transparent;
+		border-radius: 3px;
+		transition: background 0.2s ease;
+	}
+
+	*:hover::-webkit-scrollbar-thumb {
+		background: hsl(var(--muted-foreground) / 0.3);
+	}
+
+	*::-webkit-scrollbar-thumb:hover {
+		background: hsl(var(--muted-foreground) / 0.5);
 	}
 }
 

tools/server/webui/src/lib/components/app/actions/ActionIcon.svelte

@@ -0,0 +1,48 @@
+<script lang="ts">
+	import { Button } from '$lib/components/ui/button';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import type { Component } from 'svelte';
+
+	interface Props {
+		icon: Component;
+		tooltip: string;
+		variant?: 'default' | 'destructive' | 'outline' | 'secondary' | 'ghost' | 'link';
+		size?: 'default' | 'sm' | 'lg' | 'icon';
+		class?: string;
+		disabled?: boolean;
+		onclick: () => void;
+		'aria-label'?: string;
+	}
+
+	let {
+		icon,
+		tooltip,
+		variant = 'ghost',
+		size = 'sm',
+		class: className = '',
+		disabled = false,
+		onclick,
+		'aria-label': ariaLabel
+	}: Props = $props();
+</script>
+
+<Tooltip.Root>
+	<Tooltip.Trigger>
+		<Button
+			{variant}
+			{size}
+			{disabled}
+			{onclick}
+			class="h-6 w-6 p-0 {className} flex"
+			aria-label={ariaLabel || tooltip}
+		>
+			{@const IconComponent = icon}
+
+			<IconComponent class="h-3 w-3" />
+		</Button>
+	</Tooltip.Trigger>
+
+	<Tooltip.Content>
+		<p>{tooltip}</p>
+	</Tooltip.Content>
+</Tooltip.Root>

tools/server/webui/src/lib/components/app/actions/ActionIconCopyToClipboard.svelte

@@ -0,0 +1,18 @@
+<script lang="ts">
+	import { Copy } from '@lucide/svelte';
+	import { copyToClipboard } from '$lib/utils';
+
+	interface Props {
+		ariaLabel?: string;
+		canCopy?: boolean;
+		text: string;
+	}
+
+	let { ariaLabel = 'Copy to clipboard', canCopy = true, text }: Props = $props();
+</script>
+
+<Copy
+	class="h-3 w-3 flex-shrink-0 cursor-{canCopy ? 'pointer' : 'not-allowed'}"
+	aria-label={ariaLabel}
+	onclick={() => canCopy && copyToClipboard(text)}
+/>

tools/server/webui/src/lib/components/app/actions/ActionIconRemove.svelte

@@ -0,0 +1,26 @@
+<script lang="ts">
+	import { X } from '@lucide/svelte';
+	import { Button } from '$lib/components/ui/button';
+
+	interface Props {
+		id: string;
+		onRemove?: (id: string) => void;
+		class?: string;
+	}
+
+	let { id, onRemove, class: className = '' }: Props = $props();
+</script>
+
+<Button
+	type="button"
+	variant="ghost"
+	size="sm"
+	class="h-6 w-6 bg-white/20 p-0 hover:bg-white/30 {className}"
+	onclick={(e: MouseEvent) => {
+		e.stopPropagation();
+		onRemove?.(id);
+	}}
+	aria-label="Remove file"
+>
+	<X class="h-3 w-3" />
+</Button>

tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte

@@ -0,0 +1,46 @@
+<script lang="ts">
+	import { Eye } from '@lucide/svelte';
+	import ActionIconCopyToClipboard from '$lib/components/app/actions/ActionIconCopyToClipboard.svelte';
+	import { FileTypeText } from '$lib/enums';
+
+	interface Props {
+		code: string;
+		language: string;
+		disabled?: boolean;
+		onPreview?: (code: string, language: string) => void;
+	}
+
+	let { code, language, disabled = false, onPreview }: Props = $props();
+
+	const showPreview = $derived(language?.toLowerCase() === FileTypeText.HTML);
+
+	function handlePreview() {
+		if (disabled) return;
+		onPreview?.(code, language);
+	}
+</script>
+
+<div class="code-block-actions">
+	<div class="copy-code-btn" class:opacity-50={disabled} class:!cursor-not-allowed={disabled}>
+		<ActionIconCopyToClipboard
+			text={code}
+			canCopy={!disabled}
+			ariaLabel={disabled ? 'Code incomplete' : 'Copy code'}
+		/>
+	</div>
+
+	{#if showPreview}
+		<button
+			class="preview-code-btn"
+			class:opacity-50={disabled}
+			class:!cursor-not-allowed={disabled}
+			title={disabled ? 'Code incomplete' : 'Preview code'}
+			aria-label="Preview code"
+			aria-disabled={disabled}
+			type="button"
+			onclick={handlePreview}
+		>
+			<Eye size={16} />
+		</button>
+	{/if}
+</div>

tools/server/webui/src/lib/components/app/actions/index.ts

@@ -0,0 +1,19 @@
+/**
+ *
+ * ACTIONS
+ *
+ * Small interactive components for user actions.
+ *
+ */
+
+/** Styled icon button for action triggers with tooltip. */
+export { default as ActionIcon } from './ActionIcon.svelte';
+
+/** Code block actions component (copy, preview). */
+export { default as ActionIconsCodeBlock } from './ActionIconsCodeBlock.svelte';
+
+/** Copy-to-clipboard icon button with click handler. */
+export { default as ActionIconCopyToClipboard } from './ActionIconCopyToClipboard.svelte';
+
+/** Remove/delete icon button with X icon. */
+export { default as ActionIconRemove } from './ActionIconRemove.svelte';

tools/server/webui/src/lib/components/app/badges/BadgeChatStatistic.svelte

@@ -0,0 +1,44 @@
+<script lang="ts">
+	import { BadgeInfo } from '$lib/components/app';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { copyToClipboard } from '$lib/utils';
+	import type { Component } from 'svelte';
+
+	interface Props {
+		class?: string;
+		icon: Component;
+		value: string | number;
+		tooltipLabel?: string;
+	}
+
+	let { class: className = '', icon: Icon, value, tooltipLabel }: Props = $props();
+
+	function handleClick() {
+		void copyToClipboard(String(value));
+	}
+</script>
+
+{#if tooltipLabel}
+	<Tooltip.Root>
+		<Tooltip.Trigger>
+			<BadgeInfo class={className} onclick={handleClick}>
+				{#snippet icon()}
+					<Icon class="h-3 w-3" />
+				{/snippet}
+
+				{value}
+			</BadgeInfo>
+		</Tooltip.Trigger>
+		<Tooltip.Content>
+			<p>{tooltipLabel}</p>
+		</Tooltip.Content>
+	</Tooltip.Root>
+{:else}
+	<BadgeInfo class={className} onclick={handleClick}>
+		{#snippet icon()}
+			<Icon class="h-3 w-3" />
+		{/snippet}
+
+		{value}
+	</BadgeInfo>
+{/if}

tools/server/webui/src/lib/components/app/badges/BadgeInfo.svelte

@@ -0,0 +1,27 @@
+<script lang="ts">
+	import { cn } from '$lib/components/ui/utils';
+	import type { Snippet } from 'svelte';
+
+	interface Props {
+		children: Snippet;
+		class?: string;
+		icon?: Snippet;
+		onclick?: () => void;
+	}
+
+	let { children, class: className = '', icon, onclick }: Props = $props();
+</script>
+
+<button
+	class={cn(
+		'inline-flex cursor-pointer items-center gap-1 rounded-sm bg-muted-foreground/15 px-1.5 py-0.75',
+		className
+	)}
+	{onclick}
+>
+	{#if icon}
+		{@render icon()}
+	{/if}
+
+	{@render children()}
+</button>

tools/server/webui/src/lib/components/app/badges/BadgeModality.svelte

@@ -0,0 +1,39 @@
+<script lang="ts">
+	import { ModelModality } from '$lib/enums';
+	import { MODALITY_ICONS, MODALITY_LABELS } from '$lib/constants/icons';
+	import { cn } from '$lib/components/ui/utils';
+
+	type DisplayableModality = ModelModality.VISION | ModelModality.AUDIO;
+
+	interface Props {
+		modalities: ModelModality[];
+		class?: string;
+	}
+
+	let { modalities, class: className = '' }: Props = $props();
+
+	// Filter to only modalities that have icons (VISION, AUDIO)
+	const displayableModalities = $derived(
+		modalities.filter(
+			(m): m is DisplayableModality => m === ModelModality.VISION || m === ModelModality.AUDIO
+		)
+	);
+</script>
+
+{#each displayableModalities as modality, index (index)}
+	{@const IconComponent = MODALITY_ICONS[modality]}
+	{@const label = MODALITY_LABELS[modality]}
+
+	<span
+		class={cn(
+			'inline-flex items-center gap-1 rounded-md bg-muted px-2 py-1 text-xs font-medium',
+			className
+		)}
+	>
+		{#if IconComponent}
+			<IconComponent class="h-3 w-3" />
+		{/if}
+
+		{label}
+	</span>
+{/each}

tools/server/webui/src/lib/components/app/badges/index.ts

@@ -0,0 +1,16 @@
+/**
+ *
+ * BADGES & INDICATORS
+ *
+ * Small visual indicators for status and metadata.
+ *
+ */
+
+/** Badge displaying chat statistics (tokens, timing). */
+export { default as BadgeChatStatistic } from './BadgeChatStatistic.svelte';
+
+/** Generic info badge with optional tooltip and click handler. */
+export { default as BadgeInfo } from './BadgeInfo.svelte';
+
+/** Badge indicating model modality (vision, audio, tools). */
+export { default as BadgeModality } from './BadgeModality.svelte';

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatForm.svelte

@@ -27,11 +27,13 @@
 	interface Props {
 		class?: string;
 		disabled?: boolean;
+		initialMessage?: string;
 		isLoading?: boolean;
 		onFileRemove?: (fileId: string) => void;
 		onFileUpload?: (files: File[]) => void;
 		onSend?: (message: string, files?: ChatUploadedFile[]) => Promise<boolean>;
 		onStop?: () => void;
+		onSystemPromptAdd?: (draft: { message: string; files: ChatUploadedFile[] }) => void;
 		showHelperText?: boolean;
 		uploadedFiles?: ChatUploadedFile[];
 	}
@@ -39,11 +41,13 @@
 	let {
 		class: className,
 		disabled = false,
+		initialMessage = '',
 		isLoading = false,
 		onFileRemove,
 		onFileUpload,
 		onSend,
 		onStop,
+		onSystemPromptAdd,
 		showHelperText = true,
 		uploadedFiles = $bindable([])
 	}: Props = $props();
@@ -53,15 +57,28 @@
 	let currentConfig = $derived(config());
 	let fileInputRef: ChatFormFileInputInvisible | undefined = $state(undefined);
 	let isRecording = $state(false);
-	let message = $state('');
+	let message = $state(initialMessage);
 	let pasteLongTextToFileLength = $derived.by(() => {
 		const n = Number(currentConfig.pasteLongTextToFileLen);
 		return Number.isNaN(n) ? Number(SETTING_CONFIG_DEFAULT.pasteLongTextToFileLen) : n;
 	});
 	let previousIsLoading = $state(isLoading);
+	let previousInitialMessage = $state(initialMessage);
 	let recordingSupported = $state(false);
 	let textareaRef: ChatFormTextarea | undefined = $state(undefined);
 
+	// Sync message when initialMessage prop changes (e.g., after draft restoration)
+	$effect(() => {
+		if (initialMessage !== previousInitialMessage) {
+			message = initialMessage;
+			previousInitialMessage = initialMessage;
+		}
+	});
+
+	function handleSystemPromptClick() {
+		onSystemPromptAdd?.({ message, files: uploadedFiles });
+	}
+
 	// Check if model is selected (in ROUTER mode)
 	let conversationModel = $derived(
 		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
@@ -308,6 +325,7 @@
 			onFileUpload={handleFileUpload}
 			onMicClick={handleMicClick}
 			onStop={handleStop}
+			onSystemPromptClick={handleSystemPromptClick}
 		/>
 	</div>
 </form>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionFileAttachments.svelte

@@ -1,5 +1,6 @@
 <script lang="ts">
 	import { Paperclip } from '@lucide/svelte';
+	import { MessageSquare } from '@lucide/svelte';
 	import { Button } from '$lib/components/ui/button';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
@@ -11,6 +12,7 @@
 		hasAudioModality?: boolean;
 		hasVisionModality?: boolean;
 		onFileUpload?: () => void;
+		onSystemPromptClick?: () => void;
 	}
 
 	let {
@@ -18,7 +20,8 @@
 		disabled = false,
 		hasAudioModality = false,
 		hasVisionModality = false,
-		onFileUpload
+		onFileUpload,
+		onSystemPromptClick
 	}: Props = $props();
 
 	const fileUploadTooltipText = $derived.by(() => {
@@ -118,6 +121,23 @@
 					</Tooltip.Content>
 				{/if}
 			</Tooltip.Root>
+			<DropdownMenu.Separator />
+			<Tooltip.Root>
+				<Tooltip.Trigger class="w-full">
+					<DropdownMenu.Item
+						class="flex cursor-pointer items-center gap-2"
+						onclick={() => onSystemPromptClick?.()}
+					>
+						<MessageSquare class="h-4 w-4" />
+
+						<span>System Prompt</span>
+					</DropdownMenu.Item>
+				</Tooltip.Trigger>
+
+				<Tooltip.Content>
+					<p>Add a custom system message for this conversation</p>
+				</Tooltip.Content>
+			</Tooltip.Root>
 		</DropdownMenu.Content>
 	</DropdownMenu.Root>
 </div>

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte

@@ -27,6 +27,7 @@
 		onFileUpload?: () => void;
 		onMicClick?: () => void;
 		onStop?: () => void;
+		onSystemPromptClick?: () => void;
 	}
 
 	let {
@@ -39,7 +40,8 @@
 		uploadedFiles = [],
 		onFileUpload,
 		onMicClick,
-		onStop
+		onStop,
+		onSystemPromptClick
 	}: Props = $props();
 
 	let currentConfig = $derived(config());
@@ -170,6 +172,7 @@
 		{hasAudioModality}
 		{hasVisionModality}
 		{onFileUpload}
+		{onSystemPromptClick}
 	/>
 
 	<ModelsSelector