vulkan: remove shell call from vulkan-shaders-gen tool, revert file check (#17219 )

* vulkan: remove shell call from vulkan-shaders-gen tool * use string vector for command execution * Fix condition * use string, remove const_cast * Fix dependency file quotation on Windows --------- Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
sched : fix reserve ignoring user tensor assignments (#17232 )
2026-07-16 09:25:56 +02:00 · 2025-11-13 14:51:21 +01:00 · 2025-11-13 13:14:02 +01:00 · 2025-11-13 13:13:32 +01:00 · 2025-11-13 13:32:44 +02:00 · 2025-11-13 12:59:37 +02:00
101 changed files with 53696 additions and 13173 deletions
@@ -49,7 +49,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
 # -- Organize build artifacts for copying in later stages --
 # Create a lib directory to store all .so files
 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 # Create a full directory to store all executables and Python scripts
 RUN mkdir -p /app/full && \
@@ -20,7 +20,7 @@ RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
    cmake --build build -j $(nproc)

 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -25,7 +25,7 @@ RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -21,7 +21,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -32,7 +32,7 @@ RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -34,6 +34,7 @@
  rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets,
  enableCurl ? true,
  useVulkan ? false,
+  useRpc ? false,
  llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake

  # It's necessary to consistently use backendStdenv when building with CUDA support,
@@ -175,6 +176,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
      (cmakeBool "GGML_METAL" useMetalKit)
      (cmakeBool "GGML_VULKAN" useVulkan)
      (cmakeBool "GGML_STATIC" enableStatic)
+      (cmakeBool "GGML_RPC" useRpc)
    ]
    ++ optionals useCuda [
      (
@@ -45,7 +45,7 @@ RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    && cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib \
-    && find build -name "*.so" -exec cp {} /app/lib \;
+    && find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -1,4 +1,4 @@
-ARG UBUNTU_VERSION=24.04
+ARG UBUNTU_VERSION=25.10

 FROM ubuntu:$UBUNTU_VERSION AS build

@@ -7,36 +7,20 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget xz-utils

-# Install Vulkan SDK
-ARG VULKAN_VERSION=1.4.321.1
-RUN ARCH=$(uname -m) && \
-    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
-    mkdir -p /opt/vulkan && \
-    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
-    mv /tmp/${ARCH}/* /opt/vulkan/ && \
-    rm -rf /tmp/*
-
 # Install cURL and Vulkan SDK dependencies
 RUN apt install -y libcurl4-openssl-dev curl \
-    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
-
-# Set environment variables
-ENV VULKAN_SDK=/opt/vulkan
-ENV PATH=$VULKAN_SDK/bin:$PATH
-ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
-ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
-ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan-dev glslc

 # Build it
 WORKDIR /app

 COPY . .

-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
    cmake --build build --config Release -j$(nproc)

 RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+    find build -name "*.so*" -exec cp -P {} /app/lib \;

 RUN mkdir -p /app/full \
    && cp build/bin/* /app/full \
@@ -50,7 +34,7 @@ RUN mkdir -p /app/full \
 FROM ubuntu:$UBUNTU_VERSION AS base

 RUN apt-get update \
-    && apt-get install -y libgomp1 curl libvulkan-dev \
+    && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
    && apt autoremove -y \
    && apt clean -y \
    && rm -rf /tmp/* /var/tmp/* \
@@ -60,3 +60,11 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[benches/**]
+indent_style = unset
+indent_size = unset
+end_of_line = unset
+charset = unset
+trim_trailing_whitespace = unset
+insert_final_newline = unset
@@ -1651,3 +1651,50 @@ jobs:
         run: |
           GG_BUILD_KLEIDIAI=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

+  ggml-ci-arm64-graviton4-kleidiai:
+     runs-on: ah-ubuntu_22_04-c8g_8x
+
+     steps:
+       - name: Clone
+         id: checkout
+         uses: actions/checkout@v4
+
+       - name: Dependencies
+         id: depends
+         run: |
+           set -euxo pipefail
+           sudo apt-get update
+           sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+           apt-get install -y \
+            build-essential \
+            libcurl4-openssl-dev \
+            python3-venv \
+            gpg \
+            wget \
+            time \
+            git-lfs
+
+           git lfs install
+
+           # install the latest cmake
+           sudo install -d /usr/share/keyrings
+           wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+            | gpg --dearmor \
+            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+           echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+            | sudo tee /etc/apt/sources.list.d/kitware.list
+           sudo apt-get update
+           sudo apt-get install -y cmake
+
+       - name: ccache
+         uses: ggml-org/ccache-action@v1.2.16
+         with:
+           key: ggml-ci-arm64-graviton4-kleidiai
+           evict-old-files: 1d
+
+       - name: Test
+         id: ggml-ci
+         run: |
+           GG_BUILD_KLEIDIAI=1 \
+           GG_BUILD_EXTRA_TESTS_0=1 \
+           bash ./ci/run.sh ./tmp/results ./tmp/mnt
@@ -0,0 +1,52 @@
+name: Check vendor
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      'vendor/**',
+      'scripts/sync_vendor.py'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      'vendor/**',
+      'scripts/sync_vendor.py'
+    ]
+
+jobs:
+  check-vendor:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.x'
+
+      - name: Run vendor sync
+        run: |
+          set -euo pipefail
+          python3 scripts/sync_vendor.py
+
+      - name: Check for changes
+        run: |
+          set -euo pipefail
+          # detect modified or untracked files
+          changed=$(git status --porcelain --untracked-files=all || true)
+          if [ -n "$changed" ]; then
+            echo "Vendor sync modified files:"
+            echo "$changed" | awk '{ print $2 }' | sed '/^$/d'
+            echo "Failing because vendor files mismatch. Please update scripts/sync_vendor.py"
+            exit 1
+          else
+            echo "Vendor files are up-to-date."
+          fi
@@ -209,7 +209,7 @@ jobs:
        working-directory: tools/server/webui

      - name: Run UI tests
-        run: npm run test:ui
+        run: npm run test:ui -- --testTimeout=60000
        working-directory: tools/server/webui

      - name: Run E2E tests
@@ -92,6 +92,7 @@ option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_

 # 3rd party libs
 option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
+option(LLAMA_HTTPLIB    "llama: if libcurl is disabled, use httplib to download model from an URL" ON)
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" OFF)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)

@@ -200,6 +201,9 @@ endif()

 if (LLAMA_BUILD_COMMON)
    add_subdirectory(common)
+    if (LLAMA_HTTPLIB)
+        add_subdirectory(vendor/cpp-httplib)
+    endif()
 endif()

 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -0,0 +1,6 @@
+{
+  "chars": 2296.1916666666666,
+  "chars:std": 986.051306946325,
+  "score": 0.925,
+  "score:std": 0.26339134382131846
+}
@@ -0,0 +1,264 @@
+## System info
+
+```bash
+uname --all
+Linux spark-17ed 6.11.0-1016-nvidia #16-Ubuntu SMP PREEMPT_DYNAMIC Sun Sep 21 16:52:46 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux
+
+g++ --version
+g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
+
+nvidia-smi
+Sun Nov  2 10:43:25 2025
+-----------------------------------------------------------------------------------------+
+| NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
+| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
+|                                         |                        |               MIG M. |
+|=========================================+========================+======================|
+|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
+| N/A   35C    P8              4W /  N/A  | Not Supported          |      0%      Default |
+|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
+```
+
+## ggml-org/gpt-oss-20b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.374 |  1369.01 |    0.383 |    83.64 |    0.757 |   719.01 |
+|   512 |     32 |    2 |   1088 |    0.274 |  3741.35 |    0.659 |    97.14 |    0.933 |  1166.66 |
+|   512 |     32 |    4 |   2176 |    0.526 |  3896.47 |    0.817 |   156.73 |    1.342 |  1621.08 |
+|   512 |     32 |    8 |   4352 |    1.044 |  3925.10 |    0.987 |   259.44 |    2.030 |  2143.56 |
+|   512 |     32 |   16 |   8704 |    2.076 |  3945.84 |    1.248 |   410.32 |    3.324 |  2618.60 |
+|   512 |     32 |   32 |  17408 |    4.170 |  3929.28 |    1.630 |   628.40 |    5.799 |  3001.76 |
+|  4096 |     32 |    1 |   4128 |    1.083 |  3782.66 |    0.394 |    81.21 |    1.477 |  2795.13 |
+|  4096 |     32 |    2 |   8256 |    2.166 |  3782.72 |    0.725 |    88.28 |    2.891 |  2856.14 |
+|  4096 |     32 |    4 |  16512 |    4.333 |  3780.88 |    0.896 |   142.82 |    5.230 |  3157.38 |
+|  4096 |     32 |    8 |  33024 |    8.618 |  3802.14 |    1.155 |   221.69 |    9.773 |  3379.08 |
+|  4096 |     32 |   16 |  66048 |   17.330 |  3781.73 |    1.598 |   320.34 |   18.928 |  3489.45 |
+|  4096 |     32 |   32 | 132096 |   34.671 |  3780.48 |    2.336 |   438.35 |   37.007 |  3569.51 |
+|  8192 |     32 |    1 |   8224 |    2.233 |  3668.56 |    0.438 |    72.98 |    2.671 |  3078.44 |
+|  8192 |     32 |    2 |  16448 |    4.425 |  3702.95 |    0.756 |    84.66 |    5.181 |  3174.95 |
+|  8192 |     32 |    4 |  32896 |    8.859 |  3698.64 |    0.967 |   132.38 |    9.826 |  3347.72 |
+|  8192 |     32 |    8 |  65792 |   17.714 |  3699.57 |    1.277 |   200.52 |   18.991 |  3464.35 |
+|  8192 |     32 |   16 | 131584 |   35.494 |  3692.84 |    1.841 |   278.12 |   37.335 |  3524.46 |
+|  8192 |     32 |   32 | 263168 |   70.949 |  3694.82 |    2.798 |   365.99 |   73.747 |  3568.53 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      3714.25 ± 20.36 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         86.58 ± 0.43 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |      3445.17 ± 17.85 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         81.72 ± 0.53 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      3218.78 ± 11.34 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.86 ± 0.64 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       2732.83 ± 7.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.57 ± 0.51 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      2119.75 ± 12.81 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         62.33 ± 0.24 |
+
+build: eeee367de (6989)
+
+## ggml-org/gpt-oss-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.571 |   897.18 |    0.543 |    58.96 |    1.113 |   488.60 |
+|   512 |     32 |    2 |   1088 |    0.593 |  1725.37 |    1.041 |    61.45 |    1.635 |   665.48 |
+|   512 |     32 |    4 |   2176 |    1.043 |  1963.15 |    1.334 |    95.95 |    2.377 |   915.36 |
+|   512 |     32 |    8 |   4352 |    2.099 |  1951.63 |    1.717 |   149.07 |    3.816 |  1140.45 |
+|   512 |     32 |   16 |   8704 |    4.207 |  1947.12 |    2.311 |   221.56 |    6.518 |  1335.35 |
+|   512 |     32 |   32 |  17408 |    8.422 |  1945.36 |    3.298 |   310.46 |   11.720 |  1485.27 |
+|  4096 |     32 |    1 |   4128 |    2.138 |  1915.88 |    0.571 |    56.09 |    2.708 |  1524.12 |
+|  4096 |     32 |    2 |   8256 |    4.266 |  1920.25 |    1.137 |    56.27 |    5.404 |  1527.90 |
+|  4096 |     32 |    4 |  16512 |    8.564 |  1913.02 |    1.471 |    86.99 |   10.036 |  1645.29 |
+|  4096 |     32 |    8 |  33024 |   17.092 |  1917.19 |    1.979 |   129.33 |   19.071 |  1731.63 |
+|  4096 |     32 |   16 |  66048 |   34.211 |  1915.65 |    2.850 |   179.66 |   37.061 |  1782.15 |
+|  4096 |     32 |   32 | 132096 |   68.394 |  1916.44 |    4.381 |   233.72 |   72.775 |  1815.13 |
+|  8192 |     32 |    1 |   8224 |    4.349 |  1883.45 |    0.620 |    51.65 |    4.969 |  1655.04 |
+|  8192 |     32 |    2 |  16448 |    8.674 |  1888.83 |    1.178 |    54.33 |    9.852 |  1669.48 |
+|  8192 |     32 |    4 |  32896 |   17.351 |  1888.55 |    1.580 |    81.01 |   18.931 |  1737.68 |
+|  8192 |     32 |    8 |  65792 |   34.743 |  1886.31 |    2.173 |   117.80 |   36.916 |  1782.20 |
+|  8192 |     32 |   16 | 131584 |   69.413 |  1888.29 |    3.297 |   155.28 |   72.710 |  1809.70 |
+|  8192 |     32 |   32 | 263168 |  138.903 |  1887.24 |    5.004 |   204.63 |  143.907 |  1828.73 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       1919.36 ± 5.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         60.40 ± 0.30 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       1825.30 ± 6.37 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         56.94 ± 0.29 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1739.19 ± 6.00 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         52.51 ± 0.42 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1536.75 ± 4.27 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         49.33 ± 0.27 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1255.85 ± 3.26 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.99 ± 0.18 |
+
+build: eeee367de (6989)
+
+## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.398 |  1285.90 |    0.530 |    60.41 |    0.928 |   586.27 |
+|   512 |     32 |    2 |   1088 |    0.386 |  2651.65 |    0.948 |    67.50 |    1.334 |   815.38 |
+|   512 |     32 |    4 |   2176 |    0.666 |  3076.37 |    1.209 |   105.87 |    1.875 |  1160.71 |
+|   512 |     32 |    8 |   4352 |    1.325 |  3091.39 |    1.610 |   158.98 |    2.935 |  1482.65 |
+|   512 |     32 |   16 |   8704 |    2.664 |  3075.58 |    2.150 |   238.19 |    4.813 |  1808.39 |
+|   512 |     32 |   32 |  17408 |    5.336 |  3070.31 |    2.904 |   352.59 |    8.240 |  2112.50 |
+|  4096 |     32 |    1 |   4128 |    1.444 |  2836.81 |    0.581 |    55.09 |    2.025 |  2038.81 |
+|  4096 |     32 |    2 |   8256 |    2.872 |  2852.14 |    1.084 |    59.06 |    3.956 |  2086.99 |
+|  4096 |     32 |    4 |  16512 |    5.744 |  2852.32 |    1.440 |    88.90 |    7.184 |  2298.47 |
+|  4096 |     32 |    8 |  33024 |   11.463 |  2858.68 |    2.068 |   123.78 |   13.531 |  2440.65 |
+|  4096 |     32 |   16 |  66048 |   22.915 |  2859.95 |    3.018 |   169.67 |   25.933 |  2546.90 |
+|  4096 |     32 |   32 | 132096 |   45.956 |  2852.10 |    4.609 |   222.18 |   50.565 |  2612.39 |
+|  8192 |     32 |    1 |   8224 |    3.063 |  2674.72 |    0.693 |    46.20 |    3.755 |  2189.92 |
+|  8192 |     32 |    2 |  16448 |    6.109 |  2681.87 |    1.214 |    52.71 |    7.323 |  2245.98 |
+|  8192 |     32 |    4 |  32896 |   12.197 |  2686.63 |    1.682 |    76.11 |   13.878 |  2370.30 |
+|  8192 |     32 |    8 |  65792 |   24.409 |  2684.94 |    2.556 |   100.17 |   26.965 |  2439.95 |
+|  8192 |     32 |   16 | 131584 |   48.753 |  2688.50 |    3.994 |   128.20 |   52.747 |  2494.64 |
+|  8192 |     32 |   32 | 263168 |   97.508 |  2688.42 |    6.528 |   156.86 |  104.037 |  2529.57 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2925.55 ± 4.25 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         62.80 ± 0.27 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2531.01 ± 6.79 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         55.86 ± 0.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       2244.39 ± 5.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         45.95 ± 0.33 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1783.17 ± 3.68 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.07 ± 0.10 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1241.90 ± 3.13 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         29.92 ± 0.06 |
+
+build: eeee367de (6989)
+
+## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.211 |  2421.57 |    1.055 |    30.33 |    1.266 |   429.57 |
+|   512 |     32 |    2 |   1088 |    0.419 |  2441.34 |    1.130 |    56.65 |    1.549 |   702.32 |
+|   512 |     32 |    4 |   2176 |    0.873 |  2345.54 |    1.174 |   108.99 |    2.048 |  1062.74 |
+|   512 |     32 |    8 |   4352 |    1.727 |  2371.85 |    1.254 |   204.22 |    2.980 |  1460.19 |
+|   512 |     32 |   16 |   8704 |    3.452 |  2373.22 |    1.492 |   343.16 |    4.944 |  1760.56 |
+|   512 |     32 |   32 |  17408 |    6.916 |  2368.93 |    1.675 |   611.51 |    8.591 |  2026.36 |
+|  4096 |     32 |    1 |   4128 |    1.799 |  2277.26 |    1.084 |    29.51 |    2.883 |  1431.91 |
+|  4096 |     32 |    2 |   8256 |    3.577 |  2290.01 |    1.196 |    53.50 |    4.774 |  1729.51 |
+|  4096 |     32 |    4 |  16512 |    7.172 |  2284.36 |    1.313 |    97.50 |    8.485 |  1946.00 |
+|  4096 |     32 |    8 |  33024 |   14.341 |  2284.96 |    1.520 |   168.46 |   15.860 |  2082.18 |
+|  4096 |     32 |   16 |  66048 |   28.675 |  2285.44 |    1.983 |   258.21 |   30.658 |  2154.33 |
+|  4096 |     32 |   32 | 132096 |   57.354 |  2285.32 |    2.640 |   387.87 |   59.994 |  2201.82 |
+|  8192 |     32 |    1 |   8224 |    3.701 |  2213.75 |    1.119 |    28.59 |    4.820 |  1706.34 |
+|  8192 |     32 |    2 |  16448 |    7.410 |  2211.19 |    1.272 |    50.31 |    8.682 |  1894.56 |
+|  8192 |     32 |    4 |  32896 |   14.802 |  2213.83 |    1.460 |    87.68 |   16.261 |  2022.96 |
+|  8192 |     32 |    8 |  65792 |   29.609 |  2213.35 |    1.781 |   143.74 |   31.390 |  2095.93 |
+|  8192 |     32 |   16 | 131584 |   59.229 |  2212.96 |    2.495 |   205.17 |   61.725 |  2131.79 |
+|  8192 |     32 |   32 | 263168 |  118.449 |  2213.15 |    3.714 |   275.75 |  122.162 |  2154.25 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2272.74 ± 4.68 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         30.66 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2107.80 ± 9.55 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         29.71 ± 0.05 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1937.80 ± 6.75 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         28.86 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1641.12 ± 1.78 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.24 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1296.02 ± 2.67 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.78 ± 0.03 |
+
+build: eeee367de (6989)
+
+## ggml-org/gemma-3-4b-it-qat-GGUF
+
+Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.094 |  5434.73 |    0.394 |    81.21 |    0.488 |  1114.15 |
+|   512 |     32 |    2 |   1088 |    0.168 |  6091.68 |    0.498 |   128.52 |    0.666 |  1633.41 |
+|   512 |     32 |    4 |   2176 |    0.341 |  6010.68 |    0.542 |   236.37 |    0.882 |  2466.43 |
+|   512 |     32 |    8 |   4352 |    0.665 |  6161.46 |    0.678 |   377.74 |    1.342 |  3241.72 |
+|   512 |     32 |   16 |   8704 |    1.323 |  6193.19 |    0.902 |   567.41 |    2.225 |  3911.74 |
+|   512 |     32 |   32 |  17408 |    2.642 |  6202.03 |    1.231 |   832.03 |    3.872 |  4495.36 |
+|  4096 |     32 |    1 |   4128 |    0.701 |  5840.49 |    0.439 |    72.95 |    1.140 |  3621.23 |
+|  4096 |     32 |    2 |   8256 |    1.387 |  5906.82 |    0.574 |   111.48 |    1.961 |  4210.12 |
+|  4096 |     32 |    4 |  16512 |    2.758 |  5940.33 |    0.651 |   196.58 |    3.409 |  4843.33 |
+|  4096 |     32 |    8 |  33024 |    5.491 |  5967.56 |    0.876 |   292.40 |    6.367 |  5187.12 |
+|  4096 |     32 |   16 |  66048 |   10.978 |  5969.58 |    1.275 |   401.69 |   12.253 |  5390.38 |
+|  4096 |     32 |   32 | 132096 |   21.944 |  5972.93 |    1.992 |   514.16 |   23.936 |  5518.73 |
+|  8192 |     32 |    1 |   8224 |    1.402 |  5841.91 |    0.452 |    70.73 |    1.855 |  4434.12 |
+|  8192 |     32 |    2 |  16448 |    2.793 |  5865.34 |    0.637 |   100.55 |    3.430 |  4795.51 |
+|  8192 |     32 |    4 |  32896 |    5.564 |  5889.64 |    0.770 |   166.26 |    6.334 |  5193.95 |
+|  8192 |     32 |    8 |  65792 |   11.114 |  5896.44 |    1.122 |   228.07 |   12.237 |  5376.51 |
+|  8192 |     32 |   16 | 131584 |   22.210 |  5901.38 |    1.789 |   286.15 |   24.000 |  5482.74 |
+|  8192 |     32 |   32 | 263168 |   44.382 |  5906.56 |    3.044 |   336.38 |   47.426 |  5549.02 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      5810.04 ± 21.71 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         84.54 ± 0.18 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       5288.04 ± 3.54 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         78.82 ± 1.37 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      4960.43 ± 16.64 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.13 ± 0.30 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |      4495.92 ± 31.11 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         72.37 ± 0.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      3746.90 ± 40.01 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         63.02 ± 0.20 |
+
+build: eeee367de (6989)
+
@@ -454,6 +454,8 @@ cmake -B build-visionos -G Xcode \
    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
    -DLLAMA_CURL=OFF \
+    -DLLAMA_HTTPLIB=OFF \
+    -DLLAMA_BUILD_SERVER=OFF \
    -S .
 cmake --build build-visionos --config Release -- -quiet

@@ -468,6 +470,8 @@ cmake -B build-visionos-sim -G Xcode \
    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
    -DLLAMA_CURL=OFF \
+    -DLLAMA_HTTPLIB=OFF \
+    -DLLAMA_BUILD_SERVER=OFF \
    -S .
 cmake --build build-visionos-sim --config Release -- -quiet

@@ -121,7 +121,12 @@ fi
 if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
    echo ">>===== Enabling KleidiAI support"

-    CANDIDATES=("armv9-a+dotprod+i8mm" "armv8.6-a+dotprod+i8mm" "armv8.2-a+dotprod")
+    CANDIDATES=(
+        "armv9-a+dotprod+i8mm+sve2"
+        "armv9-a+dotprod+i8mm"
+        "armv8.6-a+dotprod+i8mm"
+        "armv8.2-a+dotprod"
+    )
    CPU=""

    for cpu in "${CANDIDATES[@]}"; do
@@ -79,10 +79,11 @@ if (BUILD_SHARED_LIBS)
    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

+# TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
 set(LLAMA_COMMON_EXTRA_LIBS build_info)

-# Use curl to download model url
 if (LLAMA_CURL)
+    # Use curl to download model url
    find_package(CURL)
    if (NOT CURL_FOUND)
        message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF")
@@ -90,42 +91,10 @@ if (LLAMA_CURL)
    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
    include_directories(${CURL_INCLUDE_DIRS})
    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
-endif()
-
-if (LLAMA_OPENSSL)
-    find_package(OpenSSL)
-    if (OpenSSL_FOUND)
-        include(CheckCSourceCompiles)
-        set(SAVED_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
-        set(CMAKE_REQUIRED_INCLUDES ${OPENSSL_INCLUDE_DIR})
-        check_c_source_compiles("
-        #include <openssl/opensslv.h>
-        #if defined(OPENSSL_IS_BORINGSSL) || defined(LIBRESSL_VERSION_NUMBER)
-        #    if OPENSSL_VERSION_NUMBER < 0x1010107f
-        #        error bad version
-        #    endif
-        #else
-        #    if OPENSSL_VERSION_NUMBER < 0x30000000L
-        #        error bad version
-        #    endif
-        #endif
-        int main() { return 0; }
-        " OPENSSL_VERSION_SUPPORTED)
-        set(CMAKE_REQUIRED_INCLUDES ${SAVED_CMAKE_REQUIRED_INCLUDES})
-        if (OPENSSL_VERSION_SUPPORTED)
-            message(STATUS "OpenSSL found: ${OPENSSL_VERSION}")
-            target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT)
-            target_link_libraries(${TARGET} PUBLIC OpenSSL::SSL OpenSSL::Crypto)
-            if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")
-                target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN)
-                find_library(CORE_FOUNDATION_FRAMEWORK CoreFoundation REQUIRED)
-                find_library(SECURITY_FRAMEWORK Security REQUIRED)
-                target_link_libraries(${TARGET} PUBLIC ${CORE_FOUNDATION_FRAMEWORK} ${SECURITY_FRAMEWORK})
-            endif()
-        endif()
-    else()
-        message(STATUS "OpenSSL not found, SSL support disabled")
-    endif()
+elseif (LLAMA_HTTPLIB)
+    # otherwise, use cpp-httplib
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
 endif()

 if (LLAMA_LLGUIDANCE)
@@ -2253,6 +2253,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.is_pp_shared = true;
        }
    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
+    add_opt(common_arg(
+        {"-tgs"},
+        string_format("is the text generation separated across the different sequences (default: %s)", params.is_tg_separate ? "true" : "false"),
+        [](common_params & params) {
+            params.is_tg_separate = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
    add_opt(common_arg(
        {"-npp"}, "n0,n1,...",
        "number of prompt tokens",
@@ -460,7 +460,8 @@ struct common_params {
    float slot_prompt_similarity = 0.1f;

    // batched-bench params
-    bool is_pp_shared = false;
+    bool is_pp_shared   = false;
+    bool is_tg_separate = false;

    std::vector<int32_t> n_pp;
    std::vector<int32_t> n_tg;
@@ -20,7 +20,7 @@
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
 #include <curl/easy.h>
-#else
+#elif defined(LLAMA_USE_HTTPLIB)
 #include "http.h"
 #endif

@@ -467,7 +467,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
    return { res_code, std::move(res_buffer) };
 }

-#else
+#elif defined(LLAMA_USE_HTTPLIB)

 static bool is_output_a_tty() {
 #if defined(_WIN32)
@@ -713,6 +713,8 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string

 #endif // LLAMA_USE_CURL

+#if defined(LLAMA_USE_CURL) || defined(LLAMA_USE_HTTPLIB)
+
 static bool common_download_file_single(const std::string & url,
                                        const std::string & path,
                                        const std::string & bearer_token,
@@ -907,33 +909,6 @@ common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, cons
    return { hf_repo, ggufFile, mmprojFile };
 }

-std::vector<common_cached_model_info> common_list_cached_models() {
-    std::vector<common_cached_model_info> models;
-    const std::string cache_dir = fs_get_cache_directory();
-    const std::vector<common_file_info> files = fs_list_files(cache_dir);
-    for (const auto & file : files) {
-        if (string_starts_with(file.name, "manifest=") && string_ends_with(file.name, ".json")) {
-            common_cached_model_info model_info;
-            model_info.manifest_path = file.path;
-            std::string fname = file.name;
-            string_replace_all(fname, ".json", ""); // remove extension
-            auto parts = string_split<std::string>(fname, '=');
-            if (parts.size() == 4) {
-                // expect format: manifest=<user>=<model>=<tag>=<other>
-                model_info.user  = parts[1];
-                model_info.model = parts[2];
-                model_info.tag   = parts[3];
-            } else {
-                // invalid format
-                continue;
-            }
-            model_info.size = 0; // TODO: get GGUF size, not manifest size
-            models.push_back(model_info);
-        }
-    }
-    return models;
-}
-
 //
 // Docker registry functions
 //
@@ -1052,3 +1027,46 @@ std::string common_docker_resolve_model(const std::string & docker) {
        throw;
    }
 }
+
+#else
+
+common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool) {
+    throw std::runtime_error("download functionality is not enabled in this build");
+}
+
+bool common_download_model(const common_params_model &, const std::string &, bool) {
+    throw std::runtime_error("download functionality is not enabled in this build");
+}
+
+std::string common_docker_resolve_model(const std::string &) {
+    throw std::runtime_error("download functionality is not enabled in this build");
+}
+
+#endif // LLAMA_USE_CURL || LLAMA_USE_HTTPLIB
+
+std::vector<common_cached_model_info> common_list_cached_models() {
+    std::vector<common_cached_model_info> models;
+    const std::string cache_dir = fs_get_cache_directory();
+    const std::vector<common_file_info> files = fs_list_files(cache_dir);
+    for (const auto & file : files) {
+        if (string_starts_with(file.name, "manifest=") && string_ends_with(file.name, ".json")) {
+            common_cached_model_info model_info;
+            model_info.manifest_path = file.path;
+            std::string fname = file.name;
+            string_replace_all(fname, ".json", ""); // remove extension
+            auto parts = string_split<std::string>(fname, '=');
+            if (parts.size() == 4) {
+                // expect format: manifest=<user>=<model>=<tag>=<other>
+                model_info.user  = parts[1];
+                model_info.model = parts[2];
+                model_info.tag   = parts[3];
+            } else {
+                // invalid format
+                continue;
+            }
+            model_info.size = 0; // TODO: get GGUF size, not manifest size
+            models.push_back(model_info);
+        }
+    }
+    return models;
+}
@@ -218,8 +218,7 @@ class ModelBase:
            logger.info(f"gguf: indexing model part '{part_name}'")
            ctx: ContextManager[Any]
            if is_safetensors:
-                from safetensors import safe_open
-                ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
+                ctx = cast(ContextManager[Any], gguf.utility.SafetensorsLocal(self.dir_model / part_name))
            else:
                ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", mmap=True, weights_only=True))

@@ -228,18 +227,18 @@ class ModelBase:

                for name in model_part.keys():
                    if is_safetensors:
+                        data: gguf.utility.LocalTensor = model_part[name]
                        if self.lazy:
-                            data = model_part.get_slice(name)
-                            data_gen = lambda data=data: LazyTorchTensor.from_safetensors_slice(data)  # noqa: E731
+                            data_gen = lambda data=data: LazyTorchTensor.from_local_tensor(data)  # noqa: E731
                        else:
-                            data = model_part.get_tensor(name)
-                            data_gen = lambda data=data: data  # noqa: E731
+                            dtype = LazyTorchTensor._dtype_str_map[data.dtype]
+                            data_gen = lambda data=data, dtype=dtype: torch.from_numpy(data.mmap_bytes()).view(dtype).reshape(data.shape)  # noqa: E731
                    else:
-                        data = model_part[name]
+                        data_torch: Tensor = model_part[name]
                        if self.lazy:
-                            data_gen = lambda data=data: LazyTorchTensor.from_eager(data)  # noqa: E731
+                            data_gen = lambda data=data_torch: LazyTorchTensor.from_eager(data)  # noqa: E731
                        else:
-                            data_gen = lambda data=data: data  # noqa: E731
+                            data_gen = lambda data=data_torch: data  # noqa: E731
                    tensors[name] = data_gen

        # verify tensor name presence and identify potentially missing files
@@ -278,15 +277,14 @@ class ModelBase:
                # The scale is inverted
                return data / scale.float()

-            def dequant_simple(weight: Tensor, scale: Tensor) -> Tensor:
+            def dequant_simple(weight: Tensor, scale: Tensor, block_size: Sequence[int] | None = None) -> Tensor:
                scale = scale.float()

-                if (weight_block_size := quant_config.get("weight_block_size")):
-                    # TODO: make sure it's a list of integers
-                    for i, size in enumerate(weight_block_size):
+                if block_size is not None:
+                    for i, size in enumerate(block_size):
                        scale = scale.repeat_interleave(size, i)
-                # unpad the scale (e.g. when the tensor size isn't a multiple of the block size)
-                scale = scale[tuple(slice(0, size) for size in weight.shape)]
+                    # unpad the scale (e.g. when the tensor size isn't a multiple of the block size)
+                    scale = scale[tuple(slice(0, size) for size in weight.shape)]

                return weight.float() * scale

@@ -333,6 +331,40 @@ class ModelBase:

                return (scales[g_idx].float() * (weight - zeros[g_idx]).float()).T

+            def dequant_packed(w: Tensor, scale: Tensor, shape_tensor: Tensor, zero_point: Tensor | None, num_bits: int, group_size: int):
+                assert w.dtype == torch.int32
+                shape = tuple(shape_tensor.tolist())
+                assert len(shape) == 2
+                mask = (1 << num_bits) - 1
+
+                shifts = torch.arange(0, 32 - (num_bits - 1), num_bits, dtype=torch.int32)
+                if self.lazy:
+                    shifts = LazyTorchTensor.from_eager(shifts)
+
+                if zero_point is None:
+                    offset = 1 << (num_bits - 1)
+                else:
+                    assert len(zero_point.shape) == 2
+                    offset = (zero_point.unsqueeze(1) >> shifts.reshape(1, -1, 1)) & mask
+                    offset = offset.reshape(-1, zero_point.shape[1])
+                    # trim padding, and prepare for broadcast
+                    # NOTE: the zero-point is packed along dim 0
+                    offset = offset[:shape[0], :].unsqueeze(-1)
+
+                # extract values
+                # NOTE: the weights are packed along dim 1
+                unpacked = (w.unsqueeze(-1) >> shifts.reshape(1, 1, -1)) & mask
+                unpacked = unpacked.reshape(shape[0], -1)
+
+                # trim padding
+                unpacked = unpacked[:, :shape[1]]
+
+                # prepare for broadcast of the scale
+                unpacked = unpacked.reshape(shape[0], (unpacked.shape[-1] + group_size - 1) // group_size, group_size)
+                unpacked = unpacked - offset
+
+                return (unpacked * scale.unsqueeze(-1).float()).reshape(shape)
+
            if quant_method == "bitnet":
                for name in self.model_tensors.keys():
                    if name.endswith(".weight_scale"):
@@ -342,12 +374,13 @@ class ModelBase:
                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_bitnet(w(), s())
                        tensors_to_remove.append(name)
            elif quant_method == "fp8":
+                block_size = quant_config.get("weight_block_size")
                for name in self.model_tensors.keys():
                    if name.endswith(".weight_scale_inv"):
                        weight_name = name.removesuffix("_scale_inv")
                        w = self.model_tensors[weight_name]
                        s = self.model_tensors[name]
-                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s())
+                        self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                        tensors_to_remove.append(name)
            elif quant_method == "gptq":
                for name in self.model_tensors.keys():
@@ -371,6 +404,49 @@ class ModelBase:
                                ".scales",
                            )
                        ]
+            elif quant_method == "compressed-tensors":
+                quant_format = quant_config["format"]
+                groups = quant_config["config_groups"]
+                if len(groups) > 1:
+                    raise NotImplementedError("Can't handle multiple config groups for compressed-tensors yet")
+                weight_config = tuple(groups.values())[0]["weights"]
+
+                if quant_format == "float-quantized" or quant_format == "int-quantized" or quant_format == "naive-quantized":
+                    block_size = weight_config.get("block_structure", None)
+                    strategy = weight_config.get("strategy")
+                    assert strategy == "channel" or strategy == "block"
+                    assert weight_config.get("group_size") is None  # didn't find a model using this yet
+                    for name in self.model_tensors.keys():
+                        if name.endswith(".weight_scale"):
+                            weight_name = name.removesuffix("_scale")
+                            w = self.model_tensors[weight_name]
+                            s = self.model_tensors[name]
+                            self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), block_size)
+                            tensors_to_remove.append(name)
+                elif quant_format == "pack-quantized":
+                    assert weight_config.get("strategy") == "group"
+                    assert weight_config.get("type", "int") == "int"
+                    num_bits = weight_config.get("num_bits")
+                    group_size = weight_config.get("group_size")
+                    assert isinstance(num_bits, int)
+                    assert isinstance(group_size, int)
+                    for name in self.model_tensors.keys():
+                        if name.endswith(".weight_packed"):
+                            base_name = name.removesuffix("_packed")
+                            w = self.model_tensors[name]
+                            scale = self.model_tensors[base_name + "_scale"]
+                            shape = self.model_tensors[base_name + "_shape"]
+                            zero_point = self.model_tensors.get(base_name + "_zero_point", lambda: None)
+                            new_tensors[base_name] = (
+                                lambda w=w, scale=scale, shape=shape, zero_point=zero_point: dequant_packed(
+                                    w(), scale(), shape(), zero_point(), num_bits, group_size,
+                                )
+                            )
+                            tensors_to_remove += [base_name + n for n in ("_packed", "_shape", "_scale")]
+                            if (base_name + "_zero_point") in self.model_tensors:
+                                tensors_to_remove.append(base_name + "_zero_point")
+                else:
+                    raise NotImplementedError(f"Quant format {quant_format!r} for method {quant_method!r} is not yet supported")
            else:
                raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")

@@ -7278,6 +7354,7 @@ class PLMModel(TextModel):
@ModelBase.register("T5ForConditionalGeneration")
@ModelBase.register("MT5ForConditionalGeneration")
@ModelBase.register("UMT5ForConditionalGeneration")
+@ModelBase.register("UMT5Model")
 class T5Model(TextModel):
    model_arch = gguf.MODEL_ARCH.T5

@@ -10002,6 +10079,16 @@ class LazyTorchTensor(gguf.LazyBase):
        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[...] if len(s.get_shape()) == 0 else s[:])
        return cast(torch.Tensor, lazy)

+    @classmethod
+    def from_local_tensor(cls, t: gguf.utility.LocalTensor) -> Tensor:
+        def load_tensor(tensor: gguf.utility.LocalTensor) -> Tensor:
+            dtype = cls._dtype_str_map[tensor.dtype]
+            return torch.from_numpy(tensor.mmap_bytes()).view(dtype).reshape(tensor.shape)
+        dtype = cls._dtype_str_map[t.dtype]
+        shape = t.shape
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(t,), func=lambda r: load_tensor(r))
+        return cast(torch.Tensor, lazy)
+
    @classmethod
    def from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor):
        dtype = cls._dtype_str_map[remote_tensor.dtype]
@@ -313,7 +313,12 @@ Converting the matmul weight format from ND to NZ to improve performance. Enable

 ### GGML_CANN_ACL_GRAPH

-Operators are executed using ACL graph execution, rather than in op-by-op (eager) mode. Enabled by default.
+Operators are executed using ACL graph execution, rather than in op-by-op (eager) mode. Enabled by default. This option is only effective if `USE_ACL_GRAPH` was enabled at compilation time. To enable it, recompile using:
+
+```sh
+cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release -DUSE_ACL_GRAPH=ON
+cmake --build build --config release
+```

 ### GGML_CANN_GRAPH_CACHE_CAPACITY

@@ -19,10 +19,10 @@ Legend:
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
@@ -42,7 +42,7 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -61,7 +61,7 @@ Legend:
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
@@ -77,18 +77,18 @@ Legend:
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | ❌ | ❌ |
 |                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
@@ -100,17 +100,17 @@ Legend:
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
 |                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -211,6 +211,11 @@ add_library(ggml-base
            ggml-quants.h
            gguf.cpp)

+set_target_properties(ggml-base PROPERTIES
+    VERSION ${GGML_VERSION}
+    SOVERSION ${GGML_VERSION_MAJOR}
+)
+
 target_include_directories(ggml-base PRIVATE .)
 if (GGML_BACKEND_DL)
    target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
@@ -220,6 +225,11 @@ add_library(ggml
            ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)

+set_target_properties(ggml PROPERTIES
+    VERSION ${GGML_VERSION}
+    SOVERSION ${GGML_VERSION_MAJOR}
+)
+
 if (GGML_BACKEND_DIR)
    if (NOT GGML_BACKEND_DL)
        message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
@@ -259,6 +269,12 @@ function(ggml_add_backend_library backend)
        target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
    endif()

+    # Set versioning properties for all backend libraries
+    set_target_properties(${backend} PROPERTIES
+        VERSION ${GGML_VERSION}
+        SOVERSION ${GGML_VERSION_MAJOR}
+    )
+
    if(NOT GGML_AVAILABLE_BACKENDS)
        set(GGML_AVAILABLE_BACKENDS "${backend}"
            CACHE INTERNAL "List of backends for cmake package")
@@ -1698,8 +1698,6 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
    GGML_ASSERT(sched);
    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);

-    ggml_backend_sched_reset(sched);
-
    ggml_backend_sched_synchronize(sched);

    ggml_backend_sched_split_graph(sched, measure_graph);
@@ -448,6 +448,121 @@ void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
    ggml_cann_release_resources(ctx, norm, acl_src, acl_dst);
 }

+void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+
+    aclTensor * acl_src = ggml_cann_create_tensor(src);
+    aclTensor * acl_dst = ggml_cann_create_tensor(dst);
+
+    size_t  type_size = ggml_type_size(src->type);
+    int64_t n_bytes   = src->ne[3]* src->ne[2]* src->ne[1]* type_size;
+    ggml_cann_pool_alloc temp_buffer_allocator(ctx.pool(), n_bytes);
+    void *               buffer       = temp_buffer_allocator.get();
+
+    int64_t div_ne[] = {1, src->ne[1], src->ne[2], src->ne[3]};
+    size_t  div_nb[GGML_MAX_DIMS];
+    div_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+        div_nb[i] = div_nb[i - 1] * div_ne[i - 1];
+    }
+    aclTensor *          acl_div      = ggml_cann_create_tensor(buffer, ACL_FLOAT, type_size, div_ne, div_nb, GGML_MAX_DIMS);
+
+    std::vector<int64_t> norm_dims = { 3 };
+    aclIntArray * dims_array = aclCreateIntArray(norm_dims.data(), norm_dims.size());
+
+    float p_value = 2.0f;
+    aclScalar * p_scalar = aclCreateScalar(&p_value, aclDataType::ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, Norm, acl_src, p_scalar, dims_array, true, acl_div);
+    GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src, acl_div, acl_dst);
+    ggml_cann_release_resources(ctx, dims_array, p_scalar, acl_src, acl_dst, acl_div);
+}
+
+void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+
+    const int64_t nc = src0->ne[0];
+    const int64_t nr = ggml_nrows(src0);
+
+    int64_t logits_ne[] = {nc, nr};
+    size_t logits_nb[2];
+    logits_nb[0] = ggml_type_size(src0->type);
+    logits_nb[1] = logits_nb[0] * logits_ne[0];
+    aclTensor * acl_logits = ggml_cann_create_tensor(src0->data, ACL_FLOAT, sizeof(float), logits_ne, logits_nb, 2);
+
+    size_t log_softmax_type_size = sizeof(float);
+    int64_t log_softmax_n_bytes = nr * nc * log_softmax_type_size;
+    ggml_cann_pool_alloc log_softmax_allocator(ctx.pool(), log_softmax_n_bytes);
+    void * log_softmax_buffer = log_softmax_allocator.get();
+
+    int64_t log_softmax_ne[] = {nc, nr};
+    size_t log_softmax_nb[2];
+    log_softmax_nb[0] = log_softmax_type_size;
+    log_softmax_nb[1] = log_softmax_nb[0] * log_softmax_ne[0];
+    aclTensor * acl_log_softmax = ggml_cann_create_tensor(log_softmax_buffer, ACL_FLOAT, log_softmax_type_size, log_softmax_ne, log_softmax_nb, 2);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, LogSoftmax, acl_logits, 1, acl_log_softmax);
+
+    int64_t labels_ne[] = {nc, nr};
+    size_t labels_nb[2];
+    labels_nb[0] = ggml_type_size(src1->type);
+    labels_nb[1] = labels_nb[0] * labels_ne[0];
+    aclTensor * acl_labels = ggml_cann_create_tensor(src1->data, ACL_FLOAT, sizeof(float), labels_ne, labels_nb, 2);
+
+    size_t mul_type_size = sizeof(float);
+    int64_t mul_n_bytes = nr * nc * mul_type_size;
+    ggml_cann_pool_alloc mul_allocator(ctx.pool(), mul_n_bytes);
+    void * mul_buffer = mul_allocator.get();
+
+    int64_t mul_ne[] = {nc, nr};
+    size_t mul_nb[2];
+    mul_nb[0] = mul_type_size;
+    mul_nb[1] = mul_nb[0] * mul_ne[0];
+    aclTensor * acl_mul_result = ggml_cann_create_tensor(mul_buffer, ACL_FLOAT, mul_type_size, mul_ne, mul_nb, 2);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, Mul, acl_log_softmax, acl_labels, acl_mul_result);
+
+    size_t sum_per_sample_type_size = sizeof(float);
+    int64_t sum_per_sample_n_bytes = nr * sum_per_sample_type_size;
+    ggml_cann_pool_alloc sum_per_sample_allocator(ctx.pool(), sum_per_sample_n_bytes);
+    void * sum_per_sample_buffer = sum_per_sample_allocator.get();
+
+    int64_t sum_per_sample_ne[] = {nr};
+    size_t sum_per_sample_nb[1];
+    sum_per_sample_nb[0] = sum_per_sample_type_size;
+    aclTensor * acl_sum_per_sample = ggml_cann_create_tensor(sum_per_sample_buffer, ACL_FLOAT, sum_per_sample_type_size, sum_per_sample_ne, sum_per_sample_nb, 1);
+
+    std::vector<int64_t> sum_dims = {1};
+    aclIntArray * dims_array = aclCreateIntArray(sum_dims.data(), sum_dims.size());
+    bool keep_dims = false;
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_mul_result, dims_array, keep_dims, ACL_FLOAT, acl_sum_per_sample);
+
+    size_t total_sum_type_size = sizeof(float);
+    int64_t total_sum_n_bytes = 1 * total_sum_type_size;
+    ggml_cann_pool_alloc total_sum_allocator(ctx.pool(), total_sum_n_bytes);
+    void * total_sum_buffer = total_sum_allocator.get();
+
+    int64_t total_sum_ne[] = {1};
+    size_t total_sum_nb[1];
+    total_sum_nb[0] = total_sum_type_size;
+
+    aclTensor * acl_total_sum = ggml_cann_create_tensor(total_sum_buffer, ACL_FLOAT, total_sum_type_size, total_sum_ne, total_sum_nb, 1);
+
+    std::vector<int64_t> total_sum_dims = {0};
+    aclIntArray * total_sum_dims_array = aclCreateIntArray(total_sum_dims.data(), total_sum_dims.size());
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_sum_per_sample, total_sum_dims_array, keep_dims, ACL_FLOAT, acl_total_sum);
+
+    float value = -1.0f / static_cast<float>(nr);
+    aclScalar * scale_factor = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+    aclTensor * acl_dst = ggml_cann_create_tensor(dst->data, ACL_FLOAT, sizeof(float), total_sum_ne, total_sum_nb, 1);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, Muls, acl_total_sum, scale_factor, acl_dst);
+
+    ggml_cann_release_resources(ctx, acl_logits, acl_log_softmax, acl_labels, acl_mul_result, acl_sum_per_sample, acl_total_sum, acl_dst, scale_factor, dims_array, total_sum_dims_array);
+}
+
 void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
    ggml_tensor * src = dst->src[0];

@@ -46,6 +46,8 @@
 #include <aclnnop/aclnn_cos.h>
 #include <aclnnop/aclnn_log.h>
 #include <aclnnop/aclnn_sign.h>
+#include <aclnnop/aclnn_norm.h>
+#include <aclnnop/aclnn_logsoftmax.h>
 #include "acl_tensor.h"
 #include "common.h"

@@ -187,6 +189,66 @@ void ggml_cann_argsort(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 */
 void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);

+/**
+ * @brief   Computes the L2 Normalization for a ggml tensor using the CANN
+ *          backend.
+ *
+ * @details This function applies the L2 Normalization operation on the
+ *          input tensor `src` and stores the result in the destination tensor
+ *          `dst`. L2 Normalization scales the input tensor such that the
+ *          L2 norm along the specified dimension equals 1. This operation
+ *          is commonly used in neural networks for feature normalization
+ *          and vector scaling.
+ *          The operation is defined as:
+ *          \f[
+ *              \text{out} = \frac{x}{\sqrt{\sum{x^2}}}
+ *          \f]
+ *          The normalization is performed along the last dimension by default.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the normalized values will be stored.
+ * @attention The normalization is performed along the last dimension of the
+ *            input tensor by default.
+ */
+void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Computes the Cross Entropy Loss for a ggml tensor using the CANN
+ *          backend.
+ *
+ * @details This function computes the cross entropy loss between the predicted
+ *          logits and target probability distributions. The operation follows
+ *          the same computation pattern as the CPU implementation:
+ *          1. Applies log_softmax to the logits along the class dimension
+ *          2. Element-wise multiplication with target distributions
+ *          3. Summation along the class dimension to get per-sample losses
+ *          4. Global summation and scaling by -1/nr to get final loss
+ *
+ *          The computation can be expressed as:
+ *          \f[
+ *              \text{loss} = -\frac{1}{N} \sum_{i=1}^{N} \sum_{j=1}^{C} y_{ij} \cdot \log(\text{softmax}(x_{ij}))
+ *          \f]
+ *          where \f$N\f$ is the total number of samples, \f$C\f$ is the number
+ *          of classes, \f$x\f$ are the logits, and \f$y\f$ are the target
+ *          probability distributions.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the computed loss will be stored.
+ *            This should be a scalar tensor containing the final loss value.
+ *
+ * @note This implementation computes cross entropy between probability
+ *       distributions, not the typical classification cross entropy that
+ *       expects class indices as targets. Both input tensors (src0 and src1)
+ *       should have the same shape and represent probability distributions
+ *       over the class dimension.
+ * @note The function expects two source tensors:
+ *       - dst->src[0]: Logits tensor (before softmax)
+ *       - dst->src[1]: Target probability distributions tensor
+ * @note The computation is performed using CANN backend operators including
+ *       LogSoftmax, Mul, ReduceSum, and Muls for the final scaling.
+ */
+void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
 * @brief  Computes the Group Normalization for a ggml tensor using the CANN
 *         backend.
@@ -1777,6 +1777,12 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
        case GGML_OP_GROUP_NORM:
            ggml_cann_group_norm(ctx, dst);
            break;
+        case GGML_OP_L2_NORM:
+            ggml_cann_l2_norm(ctx, dst);
+            break;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            ggml_cann_cross_entropy_loss(ctx, dst);
+            break;
        case GGML_OP_CONCAT:
            ggml_cann_concat(ctx, dst);
            break;
@@ -2515,6 +2521,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                // value of paddingW should be at most half of kernelW
                return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
            }
+        case GGML_OP_L2_NORM:
+        case GGML_OP_CROSS_ENTROPY_LOSS:
        case GGML_OP_DUP:
        case GGML_OP_SUM:
        case GGML_OP_IM2COL:
@@ -126,25 +126,36 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                )
                if (NOT ARM_MCPU_RESULT)
                    string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
+                    string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
+
+                    # on some old GCC we need to read -march=
+                    if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
+                    elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
+                    endif()
                endif()
-                if ("${ARM_MCPU_FLAG}" STREQUAL "")
-                    set(ARM_MCPU_FLAG -mcpu=native)
-                    message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
+
+                if ("${ARM_NATIVE_FLAG}" STREQUAL "")
+                    set(ARM_NATIVE_FLAG -mcpu=native)
+                    message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
+                else()
+                    message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
                endif()

                include(CheckCXXSourceRuns)

                function(check_arm_feature tag code)
                    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                    set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
+                    set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
                    check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
                    if (GGML_MACHINE_SUPPORTS_${tag})
-                        set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
+                        set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}" PARENT_SCOPE)
                    else()
-                        set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
+                        set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
                        check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
                        if (GGML_MACHINE_SUPPORTS_no${tag})
-                            set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                            set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}" PARENT_SCOPE)
                        endif()
                    endif()
                    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
@@ -155,7 +166,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                check_arm_feature(sve     "#include <arm_sve.h>\nint main()  { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
                check_arm_feature(sme     "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")

-                list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
+                list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
            else()
                if (GGML_CPU_ARM_ARCH)
                    list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
@@ -579,6 +590,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            ${KLEIDIAI_SRC}/kai/ukernels/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)

@@ -597,23 +609,34 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
-            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)

        if (NOT DOTPROD_ENABLED MATCHES -1)
            list(APPEND GGML_KLEIDIAI_SOURCES
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
        endif()

        if (NOT I8MM_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
        endif()

        if (NOT SME_ENABLED MATCHES -1)
            list(APPEND GGML_KLEIDIAI_SOURCES
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
@@ -2044,6 +2044,26 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

 }

+#ifdef __ARM_FEATURE_SVE
+static inline svuint32_t ggml_decode_q4scales_and_mins_for_mmla(const uint32_t * vx_scales) {
+    const svbool_t pg_all   = svptrue_pat_b32(SV_VL4);
+    const svbool_t pg_false = svpfalse_b();            // 0x0000
+    const svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8); // 0x00ff
+    const svbool_t pg_odd   = svzip1_b32(pg_false, pg_lo_8);
+
+    svuint32_t vutmp_hi, vutmp_lo;
+    svuint32_t vx01 = svld1_u32(pg_lo_8, vx_scales);
+    vutmp_hi = svzip1_u32(vx01, vx01);
+    vutmp_hi = svlsr_n_u32_m(pg_odd, vutmp_hi, 2);
+    vutmp_hi = svreinterpret_u32_u64(svand_n_u64_x(pg_all, svreinterpret_u64_u32(vutmp_hi), UINT64_C(0x303030303f3f3f3f)));
+    const svuint32_t vx2 = svdup_u32(vx_scales[2]);
+    vutmp_lo = svlsr_u32_x(pg_all, vx2, svreinterpret_u32_s32(svindex_s32(-2, 2)));
+    vutmp_lo = svand_n_u32_z(pg_odd, vutmp_lo, UINT32_C(0x0f0f0f0f));
+    svuint32_t vutmp = svorr_u32_z(pg_all, vutmp_hi, vutmp_lo);
+    return vutmp;
+}
+#endif
+
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
 #ifdef __ARM_FEATURE_MATMUL_INT8
@@ -2066,8 +2086,220 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    static const uint32_t kmask3 = 0x03030303;

    uint32_t utmp[4];
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif

-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        const block_q4_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q4_K * GGML_RESTRICT vx1 = (const block_q4_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        union {
+            uint32_t u32[8];
+            uint64_t u64[4];
+        } new_utmp;
+
+        svfloat32_t sumf1 = svdup_n_f32(0);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    svbool_t pg_false = svpfalse_b();
+                    svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8);
+                    svbool_t vmins_mask1= svzip1_b32(pg_lo_8, pg_false);
+                    svbool_t vmins_mask2 = svzip1_b32(pg_false, pg_lo_8);
+                    svbool_t pg128_all  = svptrue_pat_b8(SV_VL16);
+                    for (int i = 0; i < nb; ++i) {
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat32_t vy_dmins = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg128_all, svmul_f32_x(pg128_all, vy_dmins, vx_dmins), -1);
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint16_t lo = svld1_s16(pg128_all, vy0[i].bsums + 0);
+                        svint16_t hi = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        svint16_t sum_tmp1 = svuzp1_s16(lo, hi);
+                        svint16_t sum_tmp2 = svuzp2_s16(lo, hi);
+                        svint16_t svq8sums_0 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        lo = svld1_s16(pg128_all, vy1[i].bsums + 0);
+                        hi = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        sum_tmp1 = svuzp1(lo, hi);
+                        sum_tmp2 = svuzp2(lo, hi);
+                        svint16_t svq8sums_1 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg128_all, new_utmp.u32, decoded_scales);
+                        svint16_t svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp1_u32(svld1_u32(vmins_mask1, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint16_t svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp2_u32(svld1_u32(vmins_mask2, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint32_t svsumfs_tmp1 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_0));
+                        svint32_t svsumfs_tmp2 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_1));
+                        svint32_t svsumfs_tmp3 = svtrn1_s32(svsumfs_tmp1, svsumfs_tmp2);
+                        svint32_t svsumfs_tmp4 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_0));
+                        svint32_t svsumfs_tmp5 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_1));
+                        svint32_t svsumfs_tmp6 = svtrn1_s32(svsumfs_tmp4, svsumfs_tmp5);
+                        svint32_t svsumfs_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp = svadd_s32_x(pg128_all, svsumfs_tmp7, svsumfs_tmp8);
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t q4bytes_0_l, q4bytes_0_h, q4bytes_1_l, q4bytes_1_h, l0, l1, l2, l3,
+                                 q8bytes_0_h, q8bytes_0_l, q8bytes_1_h, q8bytes_1_l, r0, r1, r2, r3;
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            q4bytes_0_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 0xf));
+                            q4bytes_1_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 0xf));
+                            q4bytes_0_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 0xf));
+                            q4bytes_1_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 0xf));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+16);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+16);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi1 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg128_all, acc_sumif1, svscales, sumi1);
+
+                            q4bytes_0_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 4));
+                            q4bytes_1_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 4));
+                            q4bytes_0_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 4));
+                            q4bytes_1_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 4));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0+32);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1+32);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+48);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+48);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi2 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg128_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        sumf1 = svmla_f32_x(pg128_all,
+                                svmla_f32_x(pg128_all,
+                                    sumf1,
+                                    svcvt_f32_x(pg128_all,
+                                        svadd_s32_x(pg128_all, acc_sumif1, acc_sumif2)),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svcvt_f32_s32_x(pg128_all, svsumfs_tmp));
+                    }  //end of for nb
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    const svbool_t pg8_16 = svptrue_pat_b8(SV_VL16);
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t l0, l1, l2, l3, r0, r1, r2, r3;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_z(pg32_4, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat64_t vy_dmins_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_dmins = svreinterpret_f32_f64(svuzp1_f64(vy_dmins_tmp, vy_dmins_tmp));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg32_4, svmul_f32_x(pg32_4, vx_dmins, vy_dmins), -1);
+                        svint16_t rc1 = svuzp1_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t rc2 = svuzp2_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t svq8sums = svadd_s16_x(pg256_all, rc1, rc2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg8_16, new_utmp.u32, decoded_scales);
+                        svint16_t new_svq8sums_0 = svreinterpret_s16_u64(svtrn1_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svint16_t new_svq8sums_1 = svreinterpret_s16_u64(svtrn2_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svuint64_t new_mins_0 = svdup_u64(new_utmp.u64[2]);
+                        svuint64_t new_mins_1 = svdup_u64(new_utmp.u64[3]);
+                        svint16_t new_svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_0)));
+                        svint16_t new_svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_1)));
+                        svint64_t dot_prod_0 = svdot_s64(svdup_s64(0), new_svmins8_0, new_svq8sums_0);
+                        svint64_t dot_prod_1 = svdot_s64(dot_prod_0, new_svmins8_1, new_svq8sums_1);
+                        svfloat32_t converted_dot_prod_1 = svcvt_f32_s64_x(pg256_all, dot_prod_1);
+                        svfloat32_t svsumfs_tmp = svuzp1_f32(converted_dot_prod_1, converted_dot_prod_1);
+
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            svuint8_t q4bytes_0 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 0xf);
+                            svuint8_t q4bytes_1 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 0xf);
+                            svuint8_t q4bytes_2 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 4);
+                            svuint8_t q4bytes_3 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 4);
+                            l0 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l1 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l2 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            l3 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            svint8_t q8bytes_0 = svld1_s8(pg256_all, q8_0);
+                            svint8_t q8bytes_1 = svld1_s8(pg256_all, q8_1);
+                            svint8_t q8bytes_2 = svld1_s8(pg256_all, q8_0+32);
+                            svint8_t q8bytes_3 = svld1_s8(pg256_all, q8_1+32);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            sumi1 = svmmla(svmmla(svdup_n_s32(0), r0, l0), r1, l1);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg256_all, acc_sumif1, svscales, sumi1);
+                            sumi2 = svmmla(svmmla(svdup_n_s32(0), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg256_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        svint32_t acc_sumif = svadd_s32_x(pg256_all, acc_sumif1, acc_sumif2);
+                        svint32_t swap_acc_sumif = svext_s32(acc_sumif, acc_sumif, 4);
+                        acc_sumif = svadd_s32_x(pg32_4, acc_sumif, swap_acc_sumif);
+                        sumf1 = svmla_f32_x(pg32_4,
+                                svmla_f32_x(pg32_4,
+                                    sumf1,
+                                    svcvt_f32_x(pg32_4, acc_sumif),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svsumfs_tmp);
+                    } // end of for nb
+                } // end of case 256-512
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+
+        svst1_f32(pg32_2, s, sumf1);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sumf1), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
    if (nrc == 2) {
        const block_q4_K * GGML_RESTRICT x0 = x;
        const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
@@ -2235,7 +2467,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;

-        const int vector_length = ggml_cpu_get_sve_cnt()*8;
        const svuint8_t m4b = svdup_n_u8(0xf);
        const svint32_t mzero = svdup_n_s32(0);
        svint32_t sumi1 = svdup_n_s32(0);
@@ -2480,7 +2711,201 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    const int nb = n / QK_K;

-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        svfloat32_t sum = svdup_n_f32(0);
+
+        const block_q6_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q6_K * GGML_RESTRICT vx1 = (const block_q6_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    const svbool_t pg128_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        // process q8sum summation 128 bit route
+                        const svint16_t q8sums_01 = svld1_s16(pg128_all, vy0[i].bsums);
+                        const svint16_t q8sums_02 = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        const svint16_t q8sums_11 = svld1_s16(pg128_all, vy1[i].bsums);
+                        const svint16_t q8sums_12 = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        const svint64x2_t q6scales_0_tmp = svld2_s64(pg128_all, (const int64_t *)scale0);
+                        const svint16_t q6scales_01 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 0)));
+                        const svint16_t q6scales_02 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 1)));
+                        const svint64x2_t q6scales_1_tmp = svld2_s64(pg128_all, (const int64_t *)scale1);
+                        const svint16_t q6scales_11 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 0)));
+                        const svint16_t q6scales_12 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 1)));
+                        const svint64_t prod = svdup_n_s64(0);
+
+                        svint32_t isum_tmp1 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_01), q8sums_02, q6scales_02));
+                        svint32_t isum_tmp2 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_11), q8sums_02, q6scales_12));
+                        svint32_t isum_tmp3 = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp4 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_01), q8sums_12, q6scales_02));
+                        svint32_t isum_tmp5 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_11), q8sums_12, q6scales_12));
+                        svint32_t isum_tmp6 = svtrn1_s32(isum_tmp4, isum_tmp5);
+                        svint32_t isum_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t svisum_mins = svadd_s32_x(pg128_all, isum_tmp7, isum_tmp8);
+
+                        // process mmla
+                        svint8_t  l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; ++k) {
+                                svuint8_t qhbits_0 = svld1_u8(pg128_all, qh0+16*(k%2));
+                                svuint8_t qhbits_1 = svld1_u8(pg128_all, qh1+16*(k%2));
+                                svuint8_t q6bits_0 = svld1_u8(pg128_all, ql0+16*(k%4));
+                                svuint8_t q6bits_1 = svld1_u8(pg128_all, ql1+16*(k%4));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg128_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg128_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_0_lo, svlsr_n_u8_x(pg128_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_1_lo, svlsr_n_u8_x(pg128_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg128_all, q80+16*(k%8));
+                                svint8_t  q8bytes_1 = svld1_s8(pg128_all, q81+16*(k%8));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                isum_tmp = svmla_s32_x(pg128_all, isum_tmp, svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), svscale);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        }
+                        sum = svmla_f32_x(pg128_all, sum,
+                                svcvt_f32_x(pg128_all, svmla_s32_x(pg128_all, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg32_4, vy_d, vx_d);
+                        // process q8sum summation 256 bit route
+                        const svint16_t q8sums_0 = svld1_s16(pg256_all, vy0[i].bsums);
+                        const svint16_t q8sums_1 = svld1_s16(pg256_all, vy1[i].bsums);
+                        const svint16_t q6scales_0 = svunpklo_s16(svld1_s8(pg256_all, scale0));
+                        const svint16_t q6scales_1 = svunpklo_s16(svld1_s8(pg256_all, scale1));
+                        const svint64_t prod = svdup_n_s64(0);
+                        svint32_t isum_tmp1  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_0));
+                        svint32_t isum_tmp2  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_1));
+                        svint32_t isum_tmp3  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_0));
+                        svint32_t isum_tmp4  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_1));
+                        svint32_t isum_tmp5  = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp6  = svtrn1_s32(isum_tmp3, isum_tmp4);
+                        svint32_t isum_tmp7  = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8  = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp9  = svadd_s32_x(pg256_all, isum_tmp7, isum_tmp8);
+                        svint32_t isum_tmp10 = svreinterpret_s32_u8(svext_u8(svreinterpret_u8_s32(isum_tmp9), svreinterpret_u8_s32(isum_tmp9), 16));
+                        svint32_t svisum_mins = svadd_s32_z(pg32_4, isum_tmp9, isum_tmp10);
+
+                        // process mmla
+                        svint8_t l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; k+=2) { // process 2 block
+                                svuint8_t qhbits_0  = svld1_u8(pg256_all, qh0);
+                                svuint8_t qhbits_1  = svld1_u8(pg256_all, qh1);
+                                svuint8_t q6bits_0  = svld1_u8(pg256_all, ql0+32*((k%4)/2));
+                                svuint8_t q6bits_1  = svld1_u8(pg256_all, ql1+32*((k%4)/2));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg256_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg256_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_0_lo, svlsr_n_u8_x(pg256_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_1_lo, svlsr_n_u8_x(pg256_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg256_all, q80+32*(k/2));
+                                svint8_t  q8bytes_1 = svld1_s8(pg256_all, q81+32*(k/2));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale0 = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                svint32_t svscale1 = svzip1_s32(svdup_n_s32(scale0[k+1]), svdup_n_s32(scale1[k+1]));
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r0, l0), svscale0);
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r1, l1), svscale1);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        } // end of for
+                        svint32_t swap_isum_tmp = svext_s32(isum_tmp, isum_tmp, 4);
+                        isum_tmp = svadd_s32_x(pg32_4, isum_tmp, swap_isum_tmp);
+                        sum = svmla_f32_x(pg32_4, sum,
+                                svcvt_f32_x(pg32_4, svmla_s32_x(pg32_4, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 256
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        } // end of switch
+
+        svst1_f32(pg32_2, s, sum);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sum), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
    if (nrc == 2) {
        const block_q6_K * GGML_RESTRICT x0 = x;
        const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
@@ -2594,27 +3019,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            // adjust bias, apply superblock scale
            {
                int32_t bias[4];
-#ifdef __ARM_FEATURE_SVE
-                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
-                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
-                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
-                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
-                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
-                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
-                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
-                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
-                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
-                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
-                const svint64_t zero = svdup_n_s64(0);
-                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
-                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
-                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
-                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
-#else
                // NEON doesn't support int16 dot product, fallback to separated mul and add
                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
@@ -2646,7 +3050,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
                bias[3] = vaddvq_s32(prod);

-#endif
                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);

                const float32x4_t superblock_scale = {
@@ -2672,7 +3075,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif

 #ifdef __ARM_FEATURE_SVE
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
    float sum = 0;
    svuint8_t m4b = svdup_n_u8(0xf);
    svint32_t vzero = svdup_n_s32(0);
@@ -1807,22 +1807,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_cont(params, tensor);
            } break;
-        case GGML_OP_RESHAPE:
-            {
-                ggml_compute_forward_reshape(params, tensor);
-            } break;
-        case GGML_OP_VIEW:
-            {
-                ggml_compute_forward_view(params, tensor);
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                ggml_compute_forward_permute(params, tensor);
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                ggml_compute_forward_transpose(params, tensor);
-            } break;
        case GGML_OP_GET_ROWS:
            {
                ggml_compute_forward_get_rows(params, tensor);
@@ -2042,6 +2026,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                // nop
            } break;
+        case GGML_OP_RESHAPE:
+            {
+                // nop
+            } break;
+        case GGML_OP_PERMUTE:
+            {
+                // nop
+            } break;
+        case GGML_OP_VIEW:
+            {
+                // nop
+            } break;
+        case GGML_OP_TRANSPOSE:
+            {
+                // nop
+            } break;
        case GGML_OP_COUNT:
            {
                GGML_ABORT("fatal error");
@@ -2884,6 +2884,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
    for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
        struct ggml_tensor * node = cgraph->nodes[node_n];

+        if (ggml_op_is_empty(node->op)) {
+            // skip NOPs
+            continue;
+        }
+
        ggml_compute_forward(&params, node);

        if (state->ith == 0 && cplan->abort_callback &&
@@ -3269,6 +3274,13 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
        __m128 y_vec = _mm_cvtph_ps(x_vec);
        _mm_storeu_ps(y + i, y_vec);
    }
+#elif defined(__riscv_zvfh)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m1(n - i);
+        vfloat16m1_t vx = __riscv_vle16_v_f16m1((_Float16 *)&x[i], vl);
+        vfloat32m2_t vy = __riscv_vfwcvt_f_f_v_f32m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif

    for (; i < n; ++i) {
@@ -4,6 +4,7 @@

 // KleidiAI micro-kernels
 #include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
+#include "kai_matmul_clamp_f32_qai8dxp_qsi8cxp_interface.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
@@ -11,20 +12,31 @@
 #include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
 #include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"

 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
 #include "kai_lhs_quant_pack_qsi8d32p_f32.h"
 #include "kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.h"
 #include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
+#include "kai_lhs_quant_pack_qai8dxp_f32.h"

 #include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
 #include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
 #include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
+#include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h"

 #include "kai_common.h"

 #include "simd-mappings.h"

+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
 #include "kernels.h"

 #define NELEMS(x) sizeof(x) / sizeof(*x)
@@ -55,6 +67,14 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }

+template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
+static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
+                                     const void* lhs, const void* rhs, void* dst,
+                                     size_t dst_stride_row, size_t dst_stride_col,
+                                     float clamp_min, float clamp_max) {
+    Fn(m, n, k, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
 template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
 static inline size_t lhs_ps_fn6(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
    return Fn(m, k, bl, mr, kr, sr);
@@ -93,6 +113,12 @@ static inline void lhs_pack_void_fn9(size_t m, size_t k, size_t /*bl*/, size_t m
    Fn(m, k, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
 }

+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const float*,size_t,void*)>
+static inline void lhs_pack_float_fn9_no_bl(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr,
+                                            size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed) {
+    Fn(m, k, mr, kr, sr, m_idx_start, static_cast<const float*>(lhs), lhs_stride, lhs_packed);
+}
+
 template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
 static inline size_t rhs_ps_fn5(size_t n, size_t k, size_t nr, size_t kr, size_t bl) {
    return Fn(n, k, nr, kr, bl);
@@ -124,6 +150,18 @@ static inline void rhs_pack_fn12(size_t num_groups, size_t n, size_t k, size_t n
       static_cast<const kai_rhs_pack_qs4cxs1s0_param*>(params));
 }

+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const int8_t*,const float*,const float*,void*,size_t,const struct kai_rhs_pack_qsi8cx_params*)>
+static inline void rhs_pack_scale_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
+                                               size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* scale,
+                                               void* rhs_packed, size_t extra_bytes, const void* params) {
+    Fn(num_groups, n, k, nr, kr, sr,
+       static_cast<const int8_t*>(rhs),
+       static_cast<const float*>(bias),
+       static_cast<const float*>(scale),
+       rhs_packed, extra_bytes,
+       static_cast<const kai_rhs_pack_qsi8cx_params*>(params));
+}
+
 template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,const void*,const void*,void*,size_t,const void*)>
 static inline void rhs_pack_fn13(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
                                               size_t rhs_stride, const void* rhs, const void* bias, const void* scale,
@@ -213,6 +251,57 @@ static void dequantize_row_qsi4c32ps1s0scalef16(
    GGML_UNUSED(kr);
 }

+static void dequantize_row_qsi8cxp(
+    const void *packed_data,
+    int32_t row_idx,
+    int64_t k,
+    float *out,
+    size_t nr,
+    size_t packed_row_stride,
+    size_t kr,
+    size_t bl,
+    size_t num_bytes_multiplier
+) {
+    GGML_UNUSED(bl);
+    GGML_UNUSED(num_bytes_multiplier);
+
+    const size_t k_internal = ((size_t) k + QK8_0 - 1) / QK8_0 * QK8_0;
+    const size_t group_idx = row_idx / nr;
+    const size_t row_in_group = row_idx % nr;
+
+    const uint8_t * group_ptr = static_cast<const uint8_t *>(packed_data) + group_idx * packed_row_stride;
+    const int8_t  * data_base = reinterpret_cast<const int8_t *>(group_ptr);
+
+    const size_t num_blocks = k_internal / kr;
+
+    for (size_t block = 0; block < num_blocks; ++block) {
+        const int8_t * block_ptr = data_base + (block * nr + row_in_group) * kr;
+        for (size_t i = 0; i < kr; ++i) {
+            const size_t k_idx = block * kr + i;
+            if (k_idx < (size_t) k) {
+                out[k_idx] = static_cast<float>(block_ptr[i]);
+            }
+        }
+    }
+
+    const uint8_t * sums_ptr = group_ptr + nr * k_internal;
+    GGML_UNUSED(sums_ptr);
+
+    const float * scale_ptr = reinterpret_cast<const float *>(sums_ptr + nr * sizeof(int32_t));
+    const float scale = scale_ptr[row_in_group];
+
+    if (scale == 0.0f) {
+        for (size_t i = 0; i < (size_t) k; ++i) {
+            out[i] = 0.0f;
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < (size_t) k; ++i) {
+        out[i] *= scale;
+    }
+}
+
 static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
 #if defined(__ARM_FEATURE_SME)
    {
@@ -548,6 +637,174 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
 #endif
 };

+static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
+#if defined(__ARM_FEATURE_SME)
+    {
+        /* SME GEMM */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* SME GEMV */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
+            /* .to_float              = */ dequantize_row_qsi8cxp,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q8_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    {
+        /* I8MM GEMM */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* I8MM GEMV (dotprod fallback) */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
+            /* .to_float              = */ dequantize_row_qsi8cxp,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q8_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
+#if defined(__ARM_FEATURE_DOTPROD)
+    {
+        /* DOTPROD GEMM */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* DOTPROD GEMV */
+        {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
+            /* .to_float              = */ dequantize_row_qsi8cxp,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q8_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
+};
+
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
    ggml_kleidiai_kernels * kernel = nullptr;

@@ -562,6 +819,17 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
                break;
            }
        }
+        if (!kernel) {
+            for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8); ++i) {
+                if ((cpu_features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu &&
+                    gemm_gemv_kernels_q8[i].lhs_type == tensor->src[1]->type &&
+                    gemm_gemv_kernels_q8[i].rhs_type == tensor->src[0]->type &&
+                    gemm_gemv_kernels_q8[i].op_type  == tensor->type) {
+                    kernel = &gemm_gemv_kernels_q8[i];
+                    break;
+                }
+            }
+        }
 #endif
    }

@@ -582,3 +850,18 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features)

    return kernels;
 }
+
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features) {
+    ggml_kleidiai_kernels * kernels = nullptr;
+
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
+    for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8); ++i) {
+        if ((features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu) {
+            kernels = &gemm_gemv_kernels_q8[i];
+            break;
+        }
+    }
+#endif
+
+    return kernels;
+}
@@ -87,3 +87,4 @@ struct ggml_kleidiai_kernels {

 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features);
@@ -5,10 +5,13 @@
 #include <assert.h>
 #include <atomic>
 #include <cfloat>
+#include <cmath>
+#include <algorithm>
 #include <stdexcept>
 #include <stdint.h>
 #include <string.h>
 #include <string>
+#include <vector>
 #if defined(__linux__)
 #include <asm/hwcap.h>
 #include <sys/auxv.h>
@@ -38,8 +41,9 @@

 struct ggml_kleidiai_context {
    cpu_feature features;
-    ggml_kleidiai_kernels * kernels;
-} static ctx = { CPU_FEATURE_NONE, NULL };
+    ggml_kleidiai_kernels * kernels_q4;
+    ggml_kleidiai_kernels * kernels_q8;
+} static ctx = { CPU_FEATURE_NONE, NULL, NULL };

 static const char* cpu_feature_to_string(cpu_feature f) {
    switch (f) {
@@ -73,10 +77,14 @@ static void init_kleidiai_context(void) {
        if (sme_enabled != 0) {
            ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
        }
-        ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
+        ctx.kernels_q4 = ggml_kleidiai_select_kernels_q4_0(ctx.features);
+        ctx.kernels_q8 = ggml_kleidiai_select_kernels_q8_0(ctx.features);
 #ifndef NDEBUG
-        if (ctx.kernels) {
-            GGML_LOG_DEBUG("kleidiai: using kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels->required_cpu));
+        if (ctx.kernels_q4) {
+            GGML_LOG_DEBUG("kleidiai: using q4 kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels_q4->required_cpu));
+        }
+        if (ctx.kernels_q8) {
+            GGML_LOG_DEBUG("kleidiai: using q8 kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels_q8->required_cpu));
        }
 #endif
    }
@@ -130,6 +138,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        if (kernels->rhs_type == GGML_TYPE_Q4_0) {
            if (!lhs_info->packed_size_ex) return false;
            size = lhs_info->packed_size_ex(m, k, QK4_0, mr, kr, sr);
+        } else if (kernels->rhs_type == GGML_TYPE_Q8_0) {
+            if (!lhs_info->packed_size_ex) return false;
+            size = lhs_info->packed_size_ex(m, k, QK8_0, mr, kr, sr);
        } else if (kernels->rhs_type == GGML_TYPE_F16) {
            if (!lhs_info->packed_size_ex || !kernels->rhs_info.packed_size_ex) return false;
            const int64_t lhs_batch_size0 = op->src[1]->ne[2];
@@ -149,11 +160,13 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        if (dst->op == GGML_OP_MUL_MAT) {
            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
                return compute_forward_q4_0(params, dst);
+            } else if (dst->src[0]->type == GGML_TYPE_Q8_0) {
+                return compute_forward_q8_0(params, dst);
            } else if (dst->src[0]->type == GGML_TYPE_F16) {
                return compute_forward_fp16(params, dst);
            }
        } else if (dst->op == GGML_OP_GET_ROWS) {
-            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+            if (dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0) {
                return compute_forward_get_rows(params, dst);
            }
        }
@@ -400,19 +413,120 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        return true;
    }

-    bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
-        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
-        if (!ctx.kernels) {
-            return false;
-        }
+    bool compute_forward_q8_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q8_0);

        const ggml_tensor * src0 = dst->src[0];
        const ggml_tensor * src1 = dst->src[1];

        GGML_TENSOR_BINARY_OP_LOCALS

-        rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
-        kernel_info * kernel        = &ctx.kernels->gemm;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        if (!kernels) {
+            return false;
+        }
+
+        bool is_gemv = src1->ne[1] == 1;
+        kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
+        lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
+
+        if (!kernel || !lhs_info->get_packed_offset_ex || !lhs_info->pack_func_ex ||
+            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
+            return false;
+        }
+
+        const int ith = params->ith;
+        const int nth_raw = params->nth;
+        const int nth = nth_raw > 0 ? nth_raw : 1;
+
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+
+        size_t mr = kernel->get_mr();
+        size_t kr = kernel->get_kr();
+        size_t sr = kernel->get_sr();
+
+        const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
+        uint8_t * lhs_packed       = static_cast<uint8_t *>(params->wdata);
+        const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+
+        const size_t n_step = kernel->get_n_step();
+        const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
+        const size_t n_start = ith * num_n_per_thread;
+
+        size_t n_to_process = 0;
+        if (n_start < n) {
+            n_to_process = num_n_per_thread;
+            if ((n_start + n_to_process) > n) {
+                n_to_process = n - n_start;
+            }
+        }
+
+        const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
+        const size_t m_start = ith * num_m_per_thread;
+        size_t m_to_process = num_m_per_thread;
+        if ((m_start + m_to_process) > m) {
+            m_to_process = m - m_start;
+        }
+
+        if (m_start < m) {
+            const size_t src_stride        = src1->nb[1];
+            const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
+            const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr);
+            void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
+
+            lhs_info->pack_func_ex(m_to_process, k, 0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const size_t dst_stride        = dst->nb[1];
+        const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+        const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
+        const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
+        const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
+        const void * lhs_ptr           = static_cast<const void *>(lhs_packed + lhs_packed_offset);
+        float * dst_ptr                = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
+
+        if (n_to_process > 0) {
+            kernel->run_kernel_ex(m, n_to_process, k, 0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                                  sizeof(float), -FLT_MAX, FLT_MAX);
+        }
+
+        return true;
+    }
+
+    bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        ggml_kleidiai_kernels * kernels = nullptr;
+        size_t block_len = 0;
+        size_t num_bytes_multiplier = 0;
+
+        if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+            if (!ctx.kernels_q4) {
+                return false;
+            }
+            kernels = ctx.kernels_q4;
+            block_len = QK4_0;
+            num_bytes_multiplier = sizeof(uint16_t);
+        } else if (dst->src[0]->type == GGML_TYPE_Q8_0) {
+            if (!ctx.kernels_q8) {
+                return false;
+            }
+            kernels = ctx.kernels_q8;
+            block_len = QK8_0;
+            num_bytes_multiplier = sizeof(float);
+        } else {
+            return false;
+        }
+
+        rhs_packing_info * rhs_info = &kernels->rhs_info;
+        kernel_info * kernel        = &kernels->gemm;
        if (!rhs_info->to_float || !kernel->get_nr) {
            return false;
        }
@@ -423,8 +537,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        const size_t block_rows = kernel->get_nr();
        const size_t kr         = kernel->get_kr();

-        const size_t num_bytes_multiplier = sizeof(uint16_t);
-        const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, QK4_0);
+        const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, block_len);

        const int ith = params->ith;
        const int nth = params->nth;
@@ -439,7 +552,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);

            float *out = (float *)((char *)dst->data + i * nb1);
-            rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, QK4_0, num_bytes_multiplier);
+            rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, block_len, num_bytes_multiplier);
        }

        return true;
@@ -447,21 +560,91 @@ class tensor_traits : public ggml::cpu::tensor_traits {

 public:
    int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
-        GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
-        GGML_ASSERT(ctx.kernels);
        const size_t n = tensor->ne[1];
        const size_t k = tensor->ne[0];
-        size_t nr      = ctx.kernels->gemm.get_nr();
-        size_t kr      = ctx.kernels->gemm.get_kr();
-        size_t sr      = ctx.kernels->gemm.get_sr();

-        struct kai_rhs_pack_qs4cxs1s0_param params;
-        params.lhs_zero_point = 1;
-        params.rhs_zero_point = 8;
-        ctx.kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, QK4_0, 0, (const uint8_t*)data, nullptr, nullptr, tensor->data, 0, &params);
+        if (tensor->type == GGML_TYPE_Q4_0) {
+            if (!ctx.kernels_q4) {
+                return -1;
+            }
+            size_t nr = ctx.kernels_q4->gemm.get_nr();
+            size_t kr = ctx.kernels_q4->gemm.get_kr();
+            size_t sr = ctx.kernels_q4->gemm.get_sr();
+
+            struct kai_rhs_pack_qs4cxs1s0_param params;
+            params.lhs_zero_point = 1;
+            params.rhs_zero_point = 8;
+            ctx.kernels_q4->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, QK4_0, 0,
+                                                  static_cast<const uint8_t *>(data),
+                                                  nullptr, nullptr, tensor->data, 0, &params);
+            GGML_UNUSED(data_size);
+            return 0;
+        } else if (tensor->type == GGML_TYPE_Q8_0) {
+            if (!ctx.kernels_q8) {
+                return -1;
+            }
+
+            const size_t row_stride = tensor->nb[1];
+            const size_t k_blocks   = (k + QK8_0 - 1) / QK8_0;
+
+            std::vector<int8_t> qdata(n * k, 0);
+            std::vector<float> scales(n, 0.0f);
+
+            for (size_t row = 0; row < n; ++row) {
+                const auto * row_blocks = reinterpret_cast<const block_q8_0 *>(
+                    static_cast<const uint8_t *>(data) + row * row_stride);
+
+                float max_abs = 0.0f;
+                for (size_t block = 0; block < k_blocks; ++block) {
+                    const block_q8_0 & blk = row_blocks[block];
+                    const float d = GGML_FP16_TO_FP32(blk.d);
+                    for (size_t l = 0; l < QK8_0; ++l) {
+                        const size_t linear_idx = block * QK8_0 + l;
+                        if (linear_idx >= k) {
+                            break;
+                        }
+                        const float value = d * blk.qs[l];
+                        max_abs = std::max(max_abs, std::fabs(value));
+                    }
+                }
+
+                float scale = max_abs > 0.0f ? max_abs / 127.0f : 0.0f;
+                scales[row] = scale;
+                const float inv_scale = scale > 0.0f ? 1.0f / scale : 0.0f;
+
+                for (size_t block = 0; block < k_blocks; ++block) {
+                    const block_q8_0 & blk = row_blocks[block];
+                    const float d = GGML_FP16_TO_FP32(blk.d);
+                    for (size_t l = 0; l < QK8_0; ++l) {
+                        const size_t linear_idx = block * QK8_0 + l;
+                        if (linear_idx >= k) {
+                            break;
+                        }
+                        const float value = d * blk.qs[l];
+                        int32_t q = scale > 0.0f ? static_cast<int32_t>(std::lround(value * inv_scale)) : 0;
+                        q = std::clamp(q, -127, 127);
+                        qdata[row * k + linear_idx] = static_cast<int8_t>(q);
+                    }
+                }
+            }
+
+            size_t nr = ctx.kernels_q8->gemm.get_nr();
+            size_t kr = ctx.kernels_q8->gemm.get_kr();
+            size_t sr = ctx.kernels_q8->gemm.get_sr();
+
+            struct kai_rhs_pack_qsi8cx_params params;
+            params.lhs_zero_point = 1;
+            params.scale_multiplier = 1.0f;
+
+            ctx.kernels_q8->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, 0, 0,
+                                                  qdata.data(), nullptr, scales.data(),
+                                                  tensor->data, 0, &params);
+            GGML_UNUSED(data_size);
+            return 0;
+        }

-        return 0;
        GGML_UNUSED(data_size);
+        return -1;
    }
 };

@@ -518,27 +701,45 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
 }

 static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
-    GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
-    GGML_ASSERT(ctx.kernels);
-
-    const size_t n  = tensor->ne[1];
-    const size_t k  = tensor->ne[0];
-    const size_t nr = ctx.kernels->gemm.get_nr();
-    const size_t kr = ctx.kernels->gemm.get_kr();
-
-    return ctx.kernels->rhs_info.packed_size_ex(n, k, nr, kr, QK4_0);
-
    GGML_UNUSED(buft);
+
+    const size_t n = tensor->ne[1];
+    const size_t k = tensor->ne[0];
+
+    ggml_kleidiai_kernels * kernels = nullptr;
+    size_t block_len = 0;
+
+    if (tensor->type == GGML_TYPE_Q4_0) {
+        GGML_ASSERT(ctx.kernels_q4);
+        kernels = ctx.kernels_q4;
+        block_len = QK4_0;
+    } else if (tensor->type == GGML_TYPE_Q8_0) {
+        GGML_ASSERT(ctx.kernels_q8);
+        kernels = ctx.kernels_q8;
+        block_len = QK8_0;
+    } else {
+        return 0;
+    }
+
+    const size_t nr = kernels->gemm.get_nr();
+    const size_t kr = kernels->gemm.get_kr();
+    const size_t packed = kernels->rhs_info.packed_size_ex(n, k, nr, kr, block_len);
+    const size_t raw     = ggml_nbytes(tensor);
+
+    return packed > raw ? packed : raw;
 }

 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
    bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
        if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
-            op->src[0]->type == GGML_TYPE_Q4_0 &&
+            (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0) &&
            op->src[0]->buffer &&
            (ggml_n_dims(op->src[0]) == 2) &&
-            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
+            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
+            if (((op->src[0]->type == GGML_TYPE_Q4_0) ? ctx.kernels_q4 : ctx.kernels_q8) == nullptr) {
+                return false;
+            }
            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                return false;
            }
@@ -7,8 +7,9 @@
 #include "unary-ops.h"
 #include "vec.h"

-#include <float.h>
+#include <cfloat>
 #include <algorithm>
+#include <functional>

 // ggml_compute_forward_dup

@@ -4455,46 +4456,6 @@ void ggml_compute_forward_cont(
    ggml_compute_forward_dup(params, dst);
 }

-// ggml_compute_forward_reshape
-
-void ggml_compute_forward_reshape(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_view
-
-void ggml_compute_forward_view(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_permute
-
-void ggml_compute_forward_permute(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_transpose
-
-void ggml_compute_forward_transpose(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
 // ggml_compute_forward_get_rows

 static void ggml_compute_forward_get_rows_q(
@@ -5543,7 +5504,28 @@ static void ggml_mrope_cache_init(
    }
 }

-static void ggml_compute_forward_rope_f32(
+
+template<typename T>
+static void rotate_pairs(const int64_t n, const int64_t n_offset, const float * cache, const T * src_data, T * dst_data, const int scale = 2) {
+  for (int64_t i0 = 0; i0 < n; i0 += 2) {
+    const int64_t ic = i0/scale; // hack for GGML_ROPE_TYPE_NORMAL, where we need ic = i0; for all other cases, ic = i0/2
+
+    const float cos_theta = cache[i0 + 0];
+    const float sin_theta = cache[i0 + 1];
+
+    const T * const src = src_data + ic;
+    T * dst             = dst_data + ic;
+
+    const float x0 = type_conversion_table<T>::to_f32(src[0]);
+    const float x1 = type_conversion_table<T>::to_f32(src[n_offset]);
+
+    dst[0]        = type_conversion_table<T>::from_f32(x0*cos_theta - x1*sin_theta);
+    dst[n_offset] = type_conversion_table<T>::from_f32(x0*sin_theta + x1*cos_theta);
+  }
+}
+
+template<typename T> //float or ggml_fp16_t
+static void ggml_compute_forward_rope_flt(
        const ggml_compute_params * params,
        ggml_tensor * dst,
        const bool forward) {
@@ -5552,6 +5534,9 @@ static void ggml_compute_forward_rope_f32(
    const ggml_tensor * src1 = dst->src[1];
    const ggml_tensor * src2 = dst->src[2];

+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+
    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
    int sections[4];

@@ -5574,7 +5559,8 @@ static void ggml_compute_forward_rope_f32(
    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
    //printf("n_past = %d, ne2 = %d\n", n_past, ne2);

-    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb0 == nb00);
+    GGML_ASSERT(nb0 == sizeof(T));

    const int ith = params->ith;
    const int nth = params->nth;
@@ -5599,12 +5585,11 @@ static void ggml_compute_forward_rope_f32(
    float corr_dims[2];
    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);

-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;

-    if (is_mrope) {
+    if (mrope_used) {
        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
    }

@@ -5630,7 +5615,7 @@ static void ggml_compute_forward_rope_f32(
        for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len

            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!is_mrope) {
+            if (!mrope_used) {
                const int64_t p = pos[i2];
                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
            }
@@ -5648,269 +5633,36 @@ static void ggml_compute_forward_rope_f32(
                if (ir++ < ir0) continue;
                if (ir   > ir1) break;

-                if (is_neox || is_mrope) {
-                    if (is_vision){
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
+                T * src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+                T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1);

-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = src[0];
-                            const float x1 = src[n_dims];
-
-                            dst_data[0]      = x0*cos_theta - x1*sin_theta;
-                            dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
-                        }
-                    } else {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = src[0];
-                            const float x1 = src[n_dims/2];
-
-                            dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
-                        }
-                    }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = src[0];
-                        const float x1 = src[1];
-
-                        dst_data[0] = x0*cos_theta - x1*sin_theta;
-                        dst_data[1] = x0*sin_theta + x1*cos_theta;
-                    }
+                switch (mode) {
+                    case GGML_ROPE_TYPE_NORMAL:
+                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        break;
+                    case GGML_ROPE_TYPE_NEOX:
+                    case GGML_ROPE_TYPE_MROPE:
+                    case GGML_ROPE_TYPE_IMROPE:
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        break;
+                    case GGML_ROPE_TYPE_VISION:
+                        rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
+                        break;
+                    default:
+                        GGML_ABORT("rope type not supported");
                }

-                if (is_vision) {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const int64_t ic = i0/2;
-
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                        const float x0 = src[0];
-                        const float x1 = src[n_dims];
-
-                        dst_data[0]      = x0*cos_theta - x1*sin_theta;
-                        dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
-                    }
-                } else {
+                if (!is_vision) {
                    // fill the remain channels with data from src tensor
                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                        const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);

                        dst_data[0] = src[0];
                        dst_data[1] = src[1];
                    }
                }
-            }
-        }
-    }
-}
-
-// TODO: deduplicate f16/f32 code
-static void ggml_compute_forward_rope_f16(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const bool forward) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const ggml_tensor * src2 = dst->src[2];
-
-    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-    int sections[4];
-
-    //const int n_past     = ((int32_t *) dst->op_params)[0];
-    const int n_dims     = ((int32_t *) dst->op_params)[1];
-    const int mode       = ((int32_t *) dst->op_params)[2];
-    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
-    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
-    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
-
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
-    //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
-
-    GGML_ASSERT(nb0 == sizeof(ggml_fp16_t));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nr = ggml_nrows(dst);
-
-    GGML_ASSERT(n_dims <= ne0);
-    GGML_ASSERT(n_dims % 2 == 0);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    // row index used to determine which thread to use
-    int ir = 0;
-
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
-
-    float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
-
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
-    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
-    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
-
-    if (is_mrope) {
-        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
-    }
-
-    if (is_vision) {
-        GGML_ASSERT(n_dims == ne0/2);
-    }
-
-    const float * freq_factors = NULL;
-    if (src2 != NULL) {
-        GGML_ASSERT(src2->type == GGML_TYPE_F32);
-        GGML_ASSERT(src2->ne[0] >= n_dims / 2);
-        freq_factors = (const float *) src2->data;
-    }
-
-    // backward process uses inverse rotation by cos and sin.
-    // cos and sin build a rotation matrix, where the inverse is the transpose.
-    // this essentially just switches the sign of sin.
-    const float sin_sign = forward ? 1.0f : -1.0f;
-
-    const int32_t * pos = (const int32_t *) src1->data;
-
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-
-            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!is_mrope) {
-                const int64_t p = pos[i2];
-                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-            else {
-                const int64_t p_t = pos[i2];
-                const int64_t p_h = pos[i2 + ne2];
-                const int64_t p_w = pos[i2 + ne2 * 2];
-                const int64_t p_e = pos[i2 + ne2 * 3];
-                ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
-                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                if (ir++ < ir0) continue;
-                if (ir   > ir1) break;
-
-                if (is_neox || is_mrope) {
-                    if (is_vision) {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
-
-                            dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        }
-                    } else {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims/2]);
-
-                            dst_data[0]        = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        }
-                    }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_CPU_FP16_TO_FP32(src[1]);
-
-                        dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                    }
-                }
-
-                if (is_vision) {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const int64_t ic = i0/2;
-
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
-
-                        dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                    }
-                } else {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        dst_data[0] = src[0];
-                        dst_data[1] = src[1];
-                    }
-                }
-            }
+            } //attn-heads
        }
    }
 }
@@ -5924,11 +5676,11 @@ void ggml_compute_forward_rope(
    switch (src0->type) {
        case GGML_TYPE_F16:
            {
-                ggml_compute_forward_rope_f16(params, dst, true);
+                ggml_compute_forward_rope_flt<ggml_fp16_t>(params, dst, true);
            } break;
        case GGML_TYPE_F32:
            {
-                ggml_compute_forward_rope_f32(params, dst, true);
+                ggml_compute_forward_rope_flt<float>(params, dst, true);
            } break;
        default:
            {
@@ -5948,11 +5700,11 @@ void ggml_compute_forward_rope_back(
    switch (src0->type) {
        case GGML_TYPE_F16:
            {
-                ggml_compute_forward_rope_f16(params, dst, false);
+                ggml_compute_forward_rope_flt<ggml_fp16_t>(params, dst, false);
            } break;
        case GGML_TYPE_F32:
            {
-                ggml_compute_forward_rope_f32(params, dst, false);
+                ggml_compute_forward_rope_flt<float>(params, dst, false);
            } break;
        default:
            {
@@ -7913,6 +7665,18 @@ void ggml_compute_forward_timestep_embedding(

 // ggml_compute_forward_argsort

+template<enum ggml_sort_order order>
+struct argsort_cmp {
+    const float * data;
+    bool operator()(int32_t a, int32_t b) const {
+        if constexpr (order == GGML_SORT_ORDER_ASC) {
+            return data[a] < data[b];
+        } else {
+            return data[a] > data[b];
+        }
+    }
+};
+
 static void ggml_compute_forward_argsort_f32(
    const ggml_compute_params * params,
    ggml_tensor * dst) {
@@ -7931,23 +7695,25 @@ static void ggml_compute_forward_argsort_f32(
    ggml_sort_order order = (ggml_sort_order) ggml_get_op_params_i32(dst, 0);

    for (int64_t i = ith; i < nr; i += nth) {
-        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
        const float * src_data = (float *)((char *) src0->data + i*nb01);

+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+
        for (int64_t j = 0; j < ne0; j++) {
            dst_data[j] = j;
        }

-        // C doesn't have a functional sort, so we do a bubble sort instead
-        for (int64_t j = 0; j < ne0; j++) {
-            for (int64_t k = j + 1; k < ne0; k++) {
-                if ((order == GGML_SORT_ORDER_ASC  && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
-                    (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
-                    int32_t tmp = dst_data[j];
-                    dst_data[j] = dst_data[k];
-                    dst_data[k] = tmp;
-                }
-            }
+        switch (order) {
+            case GGML_SORT_ORDER_ASC:
+                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_ASC>{src_data});
+                break;
+
+            case GGML_SORT_ORDER_DESC:
+                std::sort(dst_data, dst_data + ne0, argsort_cmp<GGML_SORT_ORDER_DESC>{src_data});
+                break;
+
+            default:
+                GGML_ABORT("invalid sort order");
        }
    }
 }
@@ -51,10 +51,6 @@ void ggml_compute_forward_scale(const struct ggml_compute_params * params, struc
 void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -360,6 +360,13 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
    for (; i + 3 < n; i += 4) {
        vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
    }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif
    for (; i < n; ++i) {
        y[i] = ggml_silu_f32(x[i]);
@@ -460,6 +467,16 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
        val = vec_mul(val, val);
        sum += (ggml_float)vec_hsum_f32x4(val);
    }
+#elif defined(__riscv_v_intrinsic)
+    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
+        __riscv_vse32_v_f32m2(&y[i], val, vl);
+        val = __riscv_vfmul_vv_f32m2(val, val, vl);
+        vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
+    }
+    sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
 #endif
    for (; i < n; ++i) {
        float val = x[i] - mean;
@@ -586,6 +586,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
 //     If dst and src point at different address spaces then they are guaranteed to not be aliased.
 template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
+    static_assert(
+        nbytes <= ggml_cuda_get_max_cpy_bytes() || alignment == 0,
+        "You are misusing the alignment parameter for ggml_cuda_memcpy_1. "
+        "The intent is for the parameter is only as a workaround if either one of the pointers is not properly aligned. "
+        "If you use it to do more bytes per copy than ggml_cuda_max_cpy_bytes() the reads and writes may not be coalesced. "
+        "Call ggml_cuda_memcpy_1 in a loop instead.");
    if constexpr (alignment != 0) {
        static_assert(nbytes % alignment == 0, "bad alignment");
    }
@@ -2992,6 +2992,36 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif

+static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
+                                                const ggml_tensor * view,
+                                                const ggml_tensor * set_rows) {
+    // ne3 not tested
+    if (rope->src[0]->ne[3] != 1) {
+        return false;
+    }
+
+    if (set_rows->type != GGML_TYPE_F32 && set_rows->type != GGML_TYPE_F16) {
+        return false;
+    }
+
+    if (set_rows->src[1]->type != GGML_TYPE_I64) {
+        return false;
+    }
+
+    // The view should flatten two dims of rope into one dim
+    if (!ggml_is_contiguous(view) || view->ne[0] != rope->ne[0] * rope->ne[1]) {
+        return false;
+    }
+
+    // Only norm/neox shaders have the fusion code
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
 #ifndef NDEBUG
    const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
@@ -3067,6 +3097,16 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
        }
    }

+    if (ops.size() == 3 && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+        const ggml_tensor * rope     = cgraph->nodes[node_idx];
+        const ggml_tensor * view     = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
+            return true;
+        }
+    }
+
    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
        return false;
    }
@@ -3196,6 +3236,15 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                        continue;
                    }

+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+                        ggml_tensor * rope = cgraph->nodes[i];
+                        ggml_tensor * set_rows = cgraph->nodes[i + 2];
+
+                        ggml_cuda_op_rope_fused(*cuda_ctx, rope, set_rows);
+                        i += 2;
+                        continue;
+                    }
+
                    if (node->op == GGML_OP_ADD) {
                        int n_fuse = 0;
                        ggml_op ops[8];
@@ -1,3 +1,6 @@
+#include "convert.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
 #include "rope.cuh"

 struct rope_corr_dims {
@@ -37,11 +40,23 @@ static __device__ void rope_yarn(
    }
 }

-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_norm(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
+template <bool forward, bool has_ff, typename T, typename D>
+static __global__ void rope_norm(const T *            x,
+                                 D *                  dst,
+                                 const int            ne0,
+                                 const int            ne1,
+                                 const int            s1,
+                                 const int            s2,
+                                 const int            n_dims,
+                                 const int32_t *      pos,
+                                 const float          freq_scale,
+                                 const float          ext_factor,
+                                 const float          attn_factor,
+                                 const rope_corr_dims corr_dims,
+                                 const float          theta_scale,
+                                 const float *        freq_factors,
+                                 const int64_t *      row_indices,
+                                 const int            set_rows_stride) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

    if (i0 >= ne0) {
@@ -53,13 +68,27 @@ static __global__ void rope_norm(
    const int row_x     = row_dst % ne1;
    const int channel_x = row_dst / ne1;

-    const int idst = row_dst*ne0 + i0;
+    int       idst = row_dst * ne0 + i0;
    const int ix   = channel_x*s2 + row_x*s1 + i0;

-    if (i0 >= n_dims) {
-        dst[idst + 0] = x[ix + 0];
-        dst[idst + 1] = x[ix + 1];
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
+    if (set_rows_stride != 0) {
+        idst = row_x * ne0 + i0;
+        idst += row_indices[channel_x] * set_rows_stride;
+    }

+    const auto & store_coaelsced = [&](float x0, float x1) {
+        if constexpr (std::is_same_v<float, D>) {
+            float2 v = make_float2(x0, x1);
+            ggml_cuda_memcpy_1<8>(dst + idst, &v);
+        } else if constexpr (std::is_same_v<half, D>) {
+            half2 v = make_half2(x0, x1);
+            ggml_cuda_memcpy_1<4>(dst + idst, &v);
+        }
+    };
+    if (i0 >= n_dims) {
+        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

@@ -75,15 +104,26 @@ static __global__ void rope_norm(
    const float x0 = x[ix + 0];
    const float x1 = x[ix + 1];

-    dst[idst + 0] = x0*cos_theta - x1*sin_theta;
-    dst[idst + 1] = x0*sin_theta + x1*cos_theta;
+    store_coaelsced(x0 * cos_theta - x1 * sin_theta, x0 * sin_theta + x1 * cos_theta);
 }

-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_neox(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
+template <bool forward, bool has_ff, typename T, typename D>
+static __global__ void rope_neox(const T *            x,
+                                 D *                  dst,
+                                 const int            ne0,
+                                 const int            ne1,
+                                 const int            s1,
+                                 const int            s2,
+                                 const int            n_dims,
+                                 const int32_t *      pos,
+                                 const float          freq_scale,
+                                 const float          ext_factor,
+                                 const float          attn_factor,
+                                 const rope_corr_dims corr_dims,
+                                 const float          theta_scale,
+                                 const float *        freq_factors,
+                                 const int64_t *      row_indices,
+                                 const int            set_rows_stride) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

    if (i0 >= ne0) {
@@ -95,12 +135,19 @@ static __global__ void rope_neox(
    const int row_x     = row_dst % ne1;
    const int channel_x = row_dst / ne1;

-    const int idst = row_dst*ne0 + i0/2;
+    int       idst = row_dst * ne0 + i0 / 2;
    const int ix   = channel_x*s2 + row_x*s1 + i0/2;

+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
+    if (set_rows_stride != 0) {
+        idst = row_x * ne0 + i0 / 2;
+        idst += row_indices[channel_x] * set_rows_stride;
+    }
+
    if (i0 >= n_dims) {
-        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
-        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
+        dst[idst + i0 / 2 + 0] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 0]);
+        dst[idst + i0 / 2 + 1] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 1]);

        return;
    }
@@ -117,8 +164,8 @@ static __global__ void rope_neox(
    const float x0 = x[ix + 0];
    const float x1 = x[ix + n_dims/2];

-    dst[idst + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
+    dst[idst + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template<bool forward, bool has_ff, typename T>
@@ -238,11 +285,25 @@ static __global__ void rope_vision(
    dst[idst + n_dims] = x0*sin_theta + x1*cos_theta;
 }

-template<bool forward, typename T>
-static void rope_norm_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+template <bool forward, typename T, typename D>
+static void rope_norm_cuda(const T *            x,
+                           D *                  dst,
+                           const int            ne0,
+                           const int            ne1,
+                           const int            s1,
+                           const int            s2,
+                           const int            n_dims,
+                           const int            nr,
+                           const int32_t *      pos,
+                           const float          freq_scale,
+                           const float          freq_base,
+                           const float          ext_factor,
+                           const float          attn_factor,
+                           const rope_corr_dims corr_dims,
+                           const float *        freq_factors,
+                           const int64_t *      row_indices,
+                           const int            set_rows_stride,
+                           cudaStream_t         stream) {
    GGML_ASSERT(ne0 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -252,20 +313,34 @@ static void rope_norm_cuda(

    if (freq_factors == nullptr) {
        rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
    } else {
        rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
    }
 }

-template<bool forward, typename T>
-static void rope_neox_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+template <bool forward, typename T, typename D>
+static void rope_neox_cuda(const T *            x,
+                           D *                  dst,
+                           const int            ne0,
+                           const int            ne1,
+                           const int            s1,
+                           const int            s2,
+                           const int            n_dims,
+                           const int            nr,
+                           const int32_t *      pos,
+                           const float          freq_scale,
+                           const float          freq_base,
+                           const float          ext_factor,
+                           const float          attn_factor,
+                           const rope_corr_dims corr_dims,
+                           const float *        freq_factors,
+                           const int64_t *      row_indices,
+                           const int            set_rows_stride,
+                           cudaStream_t         stream) {
    GGML_ASSERT(ne0 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -274,13 +349,13 @@ static void rope_neox_cuda(
    const float theta_scale = powf(freq_base, -2.0f/n_dims);

    if (freq_factors == nullptr) {
-        rope_neox<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+        rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
    } else {
-        rope_neox<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+        rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
    }
 }

@@ -333,7 +408,9 @@ static void rope_vision_cuda(
 }

 template <bool forward>
-void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
+                            ggml_tensor *               dst,
+                            const ggml_tensor *         set_rows = nullptr) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
    const ggml_tensor * src2 = dst->src[2];
@@ -341,12 +418,25 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
    const float * src0_d = (const float *)src0->data;
    const float * src1_d = (const float *)src1->data;

-    float * dst_d = (float *)dst->data;
+    void *          dst_d           = dst->data;
+    const int64_t * row_indices     = nullptr;
+    ggml_type       dst_type        = dst->type;
+    int             set_rows_stride = 0;
+
+    if (set_rows != nullptr) {
+        GGML_ASSERT(forward);
+        dst_d           = set_rows->data;
+        row_indices     = (const int64_t *) set_rows->src[1]->data;
+        dst_type        = set_rows->type;
+        set_rows_stride = set_rows->nb[1] / ggml_type_size(set_rows->type);
+    }
    cudaStream_t stream = ctx.stream();

    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
-    GGML_ASSERT(src0->type == dst->type);
+    // When not fused, src0 and dst types must match
+    // When fused (ROPE+VIEW+SET_ROWS), src0 may be F32 and dst may be F16
+    GGML_ASSERT(src0->type == dst->type || (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16));

    const int64_t ne00 = src0->ne[0]; // head dims
    const int64_t ne01 = src0->ne[1]; // num heads
@@ -404,14 +494,18 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)

    // compute
    if (is_neox) {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_neox_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
+            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                  freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
+            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                 freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
+            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
+                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                freq_factors, row_indices, set_rows_stride, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -440,14 +534,18 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
            GGML_ABORT("fatal error");
        }
    } else {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_norm_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
+            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                  freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
+            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                 freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
+            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
+                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                freq_factors, row_indices, set_rows_stride, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -461,3 +559,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    ggml_cuda_op_rope_impl<false>(ctx, dst);
 }
+
+void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rope, ggml_tensor * set_rows) {
+    ggml_cuda_op_rope_impl<true>(ctx, rope, set_rows);
+}
@@ -5,3 +5,5 @@
 void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
@@ -81,6 +81,70 @@ static __global__ void upscale_f32_bilinear(const float * x, float * dst,
    dst[index] = result;
 }

+namespace bicubic_interpolation {
+// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+__device__ const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+
+static __device__ float weight1(float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+static __device__ float weight2(float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+
+static __device__ float bicubic(float p0, float p1, float p2, float p3, float x) {
+    const float w0 = weight2(x + 1);
+    const float w1 = weight1(x + 0);
+    const float w2 = weight1(1 - x);
+    const float w3 = weight2(2 - x);
+    return p0 * w0 + p1 * w1 + p2 * w2 + p3 * w3;
+};
+} // namespace bicubic_interpolation
+
+static __global__ void upscale_f32_bicubic(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    using bicubic_interpolation::bicubic;
+
+    const int64_t index              = threadIdx.x + blockIdx.x * blockDim.x;
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    const int y0_src    = (int)floorf(y_src_f);
+    const float dy      = y_src_f - (float)y0_src;
+
+    const float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    const int x0_src    = (int)floorf(x_src_f);
+    const float dx      = x_src_f - (float)x0_src;
+
+    const char * x_base = (const char *)x + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03;
+
+    auto load = [=](int x_off, int y_off) -> float {
+        int i00_src = max(0, min(x0_src + x_off, ne00_src - 1));
+        int i01_src = max(0, min(y0_src + y_off, ne01_src - 1));
+        return *(const float *)(x_base + (int64_t)i00_src * nb00 + (int64_t)i01_src * nb01);
+    };
+
+    const float result = bicubic(
+        bicubic(load(-1,-1), load(0,-1), load(1,-1), load(2,-1), dx),
+        bicubic(load(-1, 0), load(0, 0), load(1, 0), load(2, 0), dx),
+        bicubic(load(-1, 1), load(0, 1), load(1, 1), load(2, 1), dx),
+        bicubic(load(-1, 2), load(0, 2), load(1, 2), load(2, 2), dx), dy);
+
+    dst[index] = result;
+}
+
 static void upscale_f32_cuda(const float * x, float * dst,
        const int nb00, const int nb01, const int nb02, const int nb03,
        const int ne10, const int ne11, const int ne12, const int ne13,
@@ -104,6 +168,18 @@ static void upscale_f32_bilinear_cuda(const float * x, float * dst,
    upscale_f32_bilinear<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
 }

+static void upscale_f32_bicubic_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset, cudaStream_t stream) {
+    const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    const int64_t num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32_bicubic<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+}
+
 void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const float * src0_d = (const float *)src0->data;
@@ -121,17 +197,22 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    float sf2 = (float)dst->ne[2]/src0->ne[2];
    const float sf3 = (float)dst->ne[3]/src0->ne[3];

+    float pixel_offset = 0.5f;
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        sf0          = dst->ne[0] > 1 && src0->ne[0] > 1 ? (float)(dst->ne[0] - 1) / (src0->ne[0] - 1) : sf0;
+        sf1          = dst->ne[1] > 1 && src0->ne[1] > 1 ? (float)(dst->ne[1] - 1) / (src0->ne[1] - 1) : sf1;
+        pixel_offset = 0.0f;
+    }
+
    if (mode == GGML_SCALE_MODE_NEAREST) {
        upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            sf0          = dst->ne[0] > 1 && src0->ne[0] > 1 ? (float)(dst->ne[0] - 1) / (src0->ne[0] - 1) : sf0;
-            sf1          = dst->ne[1] > 1 && src0->ne[1] > 1 ? (float)(dst->ne[1] - 1) / (src0->ne[1] - 1) : sf1;
-            pixel_offset = 0.0f;
-        }
        upscale_f32_bilinear_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                                 src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
                                 sf0, sf1, sf2, sf3, pixel_offset, stream);
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        upscale_f32_bicubic_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                                 src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+                                 sf0, sf1, sf2, sf3, pixel_offset, stream);
    }
 }
@@ -3156,26 +3156,17 @@ static inline bool op_reuse_src1(const ggml_tensor * op1, const ggml_tensor * op
    return (op0 && op0->src[1] == op1->src[1]);
 }

+static inline bool is_compute_op(ggml_tensor *node)
+{
+    return !(ggml_op_is_empty(node->op) || ggml_is_empty(node));
+}
+
 // scan the graph and figure out last compute op index
 static inline int last_compute_op(ggml_cgraph * graph) {
-    int last;
+    int last = 0;
    for (int i = 0; i < graph->n_nodes; ++i) {
-        ggml_tensor * node = graph->nodes[i];
-
-        switch (node->op) {
-            case GGML_OP_MUL_MAT:
-            case GGML_OP_MUL_MAT_ID:
-            case GGML_OP_MUL:
-            case GGML_OP_ADD:
-            case GGML_OP_SUB:
-            case GGML_OP_RMS_NORM:
-            case GGML_OP_GLU:
-            case GGML_OP_ADD_ID:
-                last = i;
-                break;
-
-            default:
-                break;
+        if (is_compute_op(graph->nodes[i])) {
+            last = i;
        }
    }

@@ -3194,6 +3185,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
    for (int i = 0; i < graph->n_nodes; ++i) {
        ggml_tensor * node = graph->nodes[i];

+        if (!is_compute_op(node)) {
+            continue;
+        }
+
        uint32_t flags = 0;

        // skip quantizer if src1 is reused
@@ -3245,14 +3240,6 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                ggml_hexagon_rope(node, flags);
                break;

-            // non-compute ops
-            case GGML_OP_NONE:
-            case GGML_OP_RESHAPE:
-            case GGML_OP_VIEW:
-            case GGML_OP_PERMUTE:
-            case GGML_OP_TRANSPOSE:
-                break;
-
            default:
                GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
        }
@@ -34,6 +34,11 @@ static hvx_elemwise_f32_func func_table_HVX[]     = { hvx_mul_f32, hvx_add_f32,
 static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_opt, hvx_add_f32_opt, hvx_sub_f32_opt };

 #define htp_binary_preamble            \
+    const struct htp_tensor * src0 = &octx->src0; \
+    const struct htp_tensor * src1 = &octx->src1; \
+    const struct htp_tensor * src2 = &octx->src2; \
+    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]; \
@@ -62,16 +67,15 @@ static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_opt, hvx_add_f
    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];
+    const uint32_t nb3 = dst->nb[3];   \
+                                       \
+    const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;

-static void binary_job_f32_per_thread(const struct htp_tensor * src0,
-                                      const struct htp_tensor * src1,
-                                      struct htp_tensor *       dst,
-                                      uint8_t *                 spad_data,
-                                      uint32_t                  nth,
-                                      uint32_t                  ith,
-                                      uint32_t                  src0_nrows_per_thread,
-                                      enum htp_op               op) {
+static void binary_job_f32_per_thread(struct htp_ops_context * octx,
+                                      uint8_t *                spad_data,
+                                      uint32_t                 nth,
+                                      uint32_t                 ith,
+                                      enum htp_op              op) {
    htp_binary_preamble;

    const size_t src0_row_size = nb01;
@@ -107,16 +111,23 @@ static void binary_job_f32_per_thread(const struct htp_tensor * src0,

    uint8_t * restrict spad_data_th = spad_data + (ith * src0_row_size);

-    const uint32_t nr0 = ne00 / ne10;
-
    const uint8_t * restrict src0_ptr = (const uint8_t *) src0->data + (src0_start_row * src0_row_size);
    uint8_t * restrict dst_ptr        = (uint8_t *) dst->data + (src0_start_row * dst_row_size);

    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
-    const uint8_t * restrict src1_ptr  = NULL;
+
+    const uint32_t ne02_ne01 = ne02 * ne01;

    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
-        src1_ptr = data_src1 + (ir % src1_nrows) * src1_row_size;
+        const uint32_t i03 = fastdiv(ir, &octx->src0_div21);
+        const uint32_t i02 = fastdiv(ir - i03 * ne02_ne01, &octx->src0_div1);
+        const uint32_t i01 = (ir - i03 * ne02_ne01 - i02 * ne01);
+
+        const uint32_t i13 = fastmodulo(i03, ne13, &octx->src1_div3);
+        const uint32_t i12 = fastmodulo(i02, ne12, &octx->src1_div2);
+        const uint32_t i11 = fastmodulo(i01, ne11, &octx->src1_div1);
+
+        const uint8_t * restrict src1_ptr = data_src1 + i13 * nb13 + i12 * nb12 + i11 * src1_row_size;

        if (ir + 1 < src0_end_row) {
            htp_l2fetch(src0_ptr + ne00, 1, src0_row_size, src0_row_size);
@@ -125,6 +136,7 @@ static void binary_job_f32_per_thread(const struct htp_tensor * src0,
            }
        }

+        const uint32_t nr0 = ne00 / ne10;
        if (nr0 > 1) {
            if ((1 == is_aligned) && (nr0 == ne00)) {
                hvx_bcast_fp32_a(spad_data_th, *(float *) src1_ptr, nr0);
@@ -149,22 +161,17 @@ static void binary_job_f32_per_thread(const struct htp_tensor * src0,
         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void binary_add_id_job_f32_per_thread(const struct htp_tensor * src0,
-                                             const struct htp_tensor * src1,
-                                             const struct htp_tensor * src2,
-                                             struct htp_tensor *       dst,
-                                             uint8_t *                 spad_data,
-                                             uint32_t                  nth,
-                                             uint32_t                  ith,
-                                             uint32_t                  src0_nrows_per_thread,
-                                             hvx_elemwise_f32_func     func_HVX) {
+static void binary_add_id_job_f32_per_thread(struct htp_ops_context * octx,
+                                             uint8_t *                spad_data,
+                                             uint32_t                 nth,
+                                             uint32_t                 ith,
+                                             hvx_elemwise_f32_func    func_HVX) {
    htp_binary_preamble;

    const size_t src0_row_size = nb01;
    const size_t src1_row_size = nb11;
    const size_t dst_row_size  = nb1;

-    const uint32_t ne02_ne01  = ne02 * ne01;
    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
@@ -187,10 +194,11 @@ static void binary_add_id_job_f32_per_thread(const struct htp_tensor * src0,
    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
    uint8_t * restrict data_dst        = (uint8_t *) dst->data;

+    const uint32_t ne02_ne01  = ne02 * ne01;
    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir++) {
        // src0 indices
-        const uint32_t i03 = ir / ne02_ne01;
-        const uint32_t i02 = (ir - i03 * ne02_ne01) / ne01;
+        const uint32_t i03 = fastdiv(ir, &octx->src0_div21);
+        const uint32_t i02 = fastdiv(ir - i03 * ne02_ne01, &octx->src0_div1);
        const uint32_t i01 = (ir - i03 * ne02_ne01 - i02 * ne01);

        // src1 indices
@@ -234,13 +242,11 @@ static void binary_job_dispatcher_f32(unsigned int n, unsigned int i, void * dat
        case HTP_OP_MUL:
        case HTP_OP_ADD:
        case HTP_OP_SUB:
-            binary_job_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->src1_spad.data, n, i,
-                                      octx->src0_nrows_per_thread, octx->op);
+            binary_job_f32_per_thread(octx, octx->src1_spad.data, n, i, octx->op);
            break;

        case HTP_OP_ADD_ID:
-            binary_add_id_job_f32_per_thread(&octx->src0, &octx->src1, &octx->src2, &octx->dst, octx->src0_spad.data, n,
-                                             i, octx->src0_nrows_per_thread, hvx_add_f32);
+            binary_add_id_job_f32_per_thread(octx, octx->src0_spad.data, n, i, hvx_add_f32);
            break;

        default:
@@ -321,6 +327,16 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {

        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;

+        octx->src0_div21 = init_fastdiv_values(src0->ne[2] * src0->ne[1]);
+        octx->src0_div3  = init_fastdiv_values(src0->ne[3]);
+        octx->src0_div2  = init_fastdiv_values(src0->ne[2]);
+        octx->src0_div1  = init_fastdiv_values(src0->ne[1]);
+
+        octx->src1_div21 = init_fastdiv_values(src1->ne[2] * src1->ne[1]);
+        octx->src1_div3  = init_fastdiv_values(src1->ne[3]);
+        octx->src1_div2  = init_fastdiv_values(src1->ne[2]);
+        octx->src1_div1  = init_fastdiv_values(src1->ne[1]);
+
        worker_pool_run_func(octx->ctx->worker_pool, binary_op_func, octx, n_jobs);
    }

@@ -119,10 +119,10 @@ static const char * htp_type_name(uint32_t t) {
 #define HTP_MAX_DIMS 4

 struct htp_tensor {
-    uint32_t data;              // Buffer offset in the messages, and data pointer on the NSP
-    uint32_t type;              // Data type
-    uint32_t ne[HTP_MAX_DIMS];  // Number of elements
-    uint32_t nb[HTP_MAX_DIMS];  // Stride in bytes (see ggml.h ggml_tensor)
+    uint32_t data;                // Buffer offset in the messages, and data pointer on the NSP
+    uint32_t type;                // Data type
+    uint32_t ne[HTP_MAX_DIMS];    // Number of elements
+    uint32_t nb[HTP_MAX_DIMS];    // Stride in bytes (see ggml.h ggml_tensor)
 };

 #define HTP_MAX_OP_PARAMS 64
@@ -4,6 +4,7 @@
 #include "htp-ctx.h"
 #include "htp-msg.h"
 #include "worker-pool.h"
+#include "ops-utils.h"

 #include <assert.h>
 #include <stdint.h>
@@ -38,6 +39,16 @@ struct htp_ops_context {
    uint32_t src0_nrows_per_thread;
    uint32_t src1_nrows_per_thread;

+    struct fastdiv_values src0_div1;  // fastdiv values for ne1
+    struct fastdiv_values src0_div2;  // fastdiv values for ne2
+    struct fastdiv_values src0_div3;  // fastdiv values for ne3
+    struct fastdiv_values src0_div21; // fastdiv values for ne2 * ne1
+
+    struct fastdiv_values src1_div1;  // fastdiv values for ne1
+    struct fastdiv_values src1_div2;  // fastdiv values for ne2
+    struct fastdiv_values src1_div3;  // fastdiv values for ne3
+    struct fastdiv_values src1_div21; // fastdiv values for ne2 * ne1
+
    uint32_t flags;
 };

@@ -31,6 +31,39 @@ static inline uint32_t htp_round_up(uint32_t n, uint32_t m) {
    return m * ((n + m - 1) / m);
 }

+// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
+// Precompute mp (m' in the paper) and L such that division
+// can be computed using a multiply (high 32b of 64b result)
+// and a shift:
+//
+// n/d = (mulhi(n, mp) + n) >> L;
+struct fastdiv_values {
+    uint32_t mp;
+    uint32_t l;
+};
+
+static inline struct fastdiv_values init_fastdiv_values(uint32_t d) {
+    struct fastdiv_values result = { 0, 0 };
+    // compute L = ceil(log2(d));
+    while (result.l < 32 && ((uint32_t) 1 << result.l) < d) {
+        ++(result.l);
+    }
+
+    result.mp = (uint32_t) (((uint64_t) 1 << 32) * (((uint64_t) 1 << result.l) - d) / d + 1);
+    return result;
+}
+
+static inline uint32_t fastdiv(uint32_t n, const struct fastdiv_values * vals) {
+    // Compute high 32 bits of n * mp
+    const uint32_t hi = (uint32_t) (((uint64_t) n * vals->mp) >> 32);  // mulhi(n, mp)
+    // add n, apply bit shift
+    return (hi + n) >> vals->l;
+}
+
+static inline uint32_t fastmodulo(uint32_t n, uint32_t d, const struct fastdiv_values * vals) {
+    return n - fastdiv(n, vals) * d;
+}
+
 static inline void htp_l2fetch(const void * p, uint32_t height, uint32_t width, uint32_t stride) {
    const uint64_t control = Q6_P_combine_RR(stride, Q6_R_combine_RlRl(width, height));
    asm volatile(" l2fetch(%0,%1) " : : "r"(p), "r"(control));
@@ -289,7 +289,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,

        // queue the copy operation into the queue of the Metal context
        // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
        id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];

        [encoder copyFromBuffer:buf_src
@@ -300,6 +300,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,

        [encoder endEncoding];
        [cmd_buf commit];
+        [buf_src release];

        // do not wait here for completion
        //[cmd_buf waitUntilCompleted];
@@ -330,7 +331,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te

        // queue the copy operation into the queue of the Metal context
        // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
        id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];

        [encoder copyFromBuffer:bid_src.metal
@@ -341,6 +342,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te

        [encoder endEncoding];
        [cmd_buf commit];
+        [buf_dst release];

        // do not wait here for completion
        //[cmd_buf waitUntilCompleted];
@@ -1438,6 +1438,30 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d(ggml_met
    return res;
 }

+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_CONV_2D);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type         == GGML_TYPE_F32);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_conv_2d_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale(ggml_metal_library_t lib, const ggml_tensor * op) {
    assert(op->op == GGML_OP_UPSCALE);

@@ -133,6 +133,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope              (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -564,8 +564,10 @@ ggml_metal_device_t ggml_metal_device_init(void) {
            // TODO: try to update the tensor API kernels to at least match the simdgroup performance
            if (getenv("GGML_METAL_TENSOR_ENABLE") == NULL &&
                ![[dev->mtl_device name] containsString:@"M5"] &&
-                ![[dev->mtl_device name] containsString:@"M6"]) {
-                GGML_LOG_WARN("%s: tensor API disabled for pre-M5 device\n", __func__);
+                ![[dev->mtl_device name] containsString:@"M6"] &&
+                ![[dev->mtl_device name] containsString:@"A19"] &&
+                ![[dev->mtl_device name] containsString:@"A20"]) {
+                GGML_LOG_WARN("%s: tensor API disabled for pre-M5 and pre-A19 devices\n", __func__);
                dev->props.has_tensor = false;
            }

@@ -883,6 +885,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
            return true;
        case GGML_OP_IM2COL:
            return ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
+        case GGML_OP_CONV_2D:
+            return ggml_is_contiguous(op->src[0]) &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   op->type == GGML_TYPE_F32 &&
+                   (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
        case GGML_OP_POOL_1D:
            return false;
        case GGML_OP_UPSCALE:
@@ -528,6 +528,36 @@ typedef struct {
    uint64_t nb2;
 } ggml_metal_kargs_conv_transpose_2d;

+typedef struct {
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  IC;
+    int32_t  OC;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  N;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  d0;
+    int32_t  d1;
+} ggml_metal_kargs_conv_2d;
+
 typedef struct {
    uint64_t  ofs0;
    uint64_t  ofs1;
@@ -10,6 +10,7 @@

 #include <cassert>
 #include <algorithm>
+#include <limits>

 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
    if (!t) {
@@ -364,6 +365,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_im2col(ctx, idx);
            } break;
+        case GGML_OP_CONV_2D:
+            {
+                n_fuse = ggml_metal_op_conv_2d(ctx, idx);
+            } break;
        case GGML_OP_CONV_TRANSPOSE_1D:
            {
                n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
@@ -3077,6 +3082,84 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

+int ggml_metal_op_conv_2d(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_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t *) op->op_params)[0];
+    const int32_t s1 = ((const int32_t *) op->op_params)[1];
+    const int32_t p0 = ((const int32_t *) op->op_params)[2];
+    const int32_t p1 = ((const int32_t *) op->op_params)[3];
+    const int32_t d0 = ((const int32_t *) op->op_params)[4];
+    const int32_t d1 = ((const int32_t *) op->op_params)[5];
+
+    ggml_metal_kargs_conv_2d args = {
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+        /*.IW   =*/ ne10,
+        /*.IH   =*/ ne11,
+        /*.KW   =*/ ne00,
+        /*.KH   =*/ ne01,
+        /*.IC   =*/ ne02,
+        /*.OC   =*/ ne03,
+        /*.OW   =*/ ne0,
+        /*.OH   =*/ ne1,
+        /*.N    =*/ ne3,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.p0   =*/ p0,
+        /*.p1   =*/ p1,
+        /*.d0   =*/ d0,
+        /*.d1   =*/ d1,
+    };
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_2d(lib, op);
+
+    int nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    nth = std::min(nth, 256);
+    nth = std::max(nth, 1);
+
+    const uint64_t n_out = ggml_nelements(op);
+
+    uint64_t tg = (n_out + nth - 1)/nth;
+    tg = std::max<uint64_t>(tg, 1);
+    tg = std::min<uint64_t>(tg, (uint64_t) std::numeric_limits<int>::max());
+
+    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->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, tg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -70,6 +70,7 @@ int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rope              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_im2col            (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_conv_2d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_1d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);
@@ -4146,6 +4146,120 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 //template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
 //template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;

+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
 typedef void (conv_transpose_1d_t)(
        constant ggml_metal_kargs_conv_transpose_1d & args,
        device const float * src0,
@@ -53,6 +53,37 @@

 bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor);

+// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
+// Precompute mp (m' in the paper) and L such that division
+// can be computed using a multiply (high 32b of 64b result)
+// and a shift:
+//
+// n/d = (mulhi(n, mp) + n) >> L;
+struct fastdiv_vals {
+    uint32_t mp;
+    uint32_t L;
+    uint32_t d;
+    uint32_t pad;
+};
+static_assert(sizeof(fastdiv_vals) == 16, "fastdiv_vals size incorrect");
+
+static fastdiv_vals init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
+
+    uint32_t d = (uint32_t)d_64;
+
+    // compute L = ceil(log2(d));
+    uint32_t L = 0;
+    while (L < 32 && (uint32_t{ 1 } << L) < d) {
+        L++;
+    }
+
+    uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
+    // pack divisor as well to reduce error surface
+    return { mp, L, d, 0 };
+}
+
 enum GPU_FAMILY {
    ADRENO,
    INTEL,
@@ -2944,8 +2975,11 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
        case GGML_OP_PAD:
            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
-        case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_UPSCALE: {
+            ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(op, 0) & 0xFF);
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   (mode == GGML_SCALE_MODE_NEAREST || mode == GGML_SCALE_MODE_BILINEAR);
+        }
        case GGML_OP_CONV_2D:
            return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
                   (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
@@ -4461,6 +4495,9 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c
            GGML_ABORT("not implemented");
    }

+    fastdiv_vals ne11_ = init_fastdiv_values(ne11);
+    fastdiv_vals ne12_ = init_fastdiv_values(ne12);
+
    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));
@@ -4471,8 +4508,8 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c
    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne11));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(fastdiv_vals), &ne11_));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(fastdiv_vals), &ne12_));
    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
@@ -1,5 +1,16 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable

+// v = { mp, L, d }
+inline uint fastdiv(uint n, uint4 v) {
+    uint msbs;
+    msbs = mul_hi(n, v.s0);
+    return (msbs + n) >> v.s1;
+}
+inline uint fastmod(uint n, uint4 v) {
+    uint q = fastdiv(n, v);
+    return n - q * v.s2;
+}
+
 kernel void kernel_set_rows_f32_i64(
        global char * src0,
        ulong         offset0,
@@ -11,8 +22,8 @@ kernel void kernel_set_rows_f32_i64(
        ulong         nb01,
        ulong         nb02,
        ulong         nb03,
-        int           ne11,
-        int           ne12,
+        uint4         ne11,
+        uint4         ne12,
        ulong         nb10,
        ulong         nb11,
        ulong         nb12,
@@ -33,8 +44,10 @@ kernel void kernel_set_rows_f32_i64(
        return;
    }

-    int i12 = i03%ne12;
-    int i11 = i02%ne11;
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);

    int i10 = i01;
    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
@@ -58,8 +71,8 @@ kernel void kernel_set_rows_f16_i64(
        ulong         nb01,
        ulong         nb02,
        ulong         nb03,
-        int           ne11,
-        int           ne12,
+        uint4         ne11,
+        uint4         ne12,
        ulong         nb10,
        ulong         nb11,
        ulong         nb12,
@@ -80,8 +93,10 @@ kernel void kernel_set_rows_f16_i64(
        return;
    }

-    int i12 = i03%ne12;
-    int i11 = i02%ne11;
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);

    int i10 = i01;
    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
@@ -105,8 +120,8 @@ kernel void kernel_set_rows_f32_i32(
        ulong         nb01,
        ulong         nb02,
        ulong         nb03,
-        int           ne11,
-        int           ne12,
+        uint4         ne11,
+        uint4         ne12,
        ulong         nb10,
        ulong         nb11,
        ulong         nb12,
@@ -127,8 +142,10 @@ kernel void kernel_set_rows_f32_i32(
        return;
    }

-    int i12 = i03%ne12;
-    int i11 = i02%ne11;
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);

    int i10 = i01;
    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
@@ -152,8 +169,8 @@ kernel void kernel_set_rows_f16_i32(
        ulong         nb01,
        ulong         nb02,
        ulong         nb03,
-        int           ne11,
-        int           ne12,
+        uint4         ne11,
+        uint4         ne12,
        ulong         nb10,
        ulong         nb11,
        ulong         nb12,
@@ -174,8 +191,10 @@ kernel void kernel_set_rows_f16_i32(
        return;
    }

-    int i12 = i03%ne12;
-    int i11 = i02%ne11;
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);

    int i10 = i01;
    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
@@ -3933,6 +3933,7 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
            break;
        case GGML_OP_SSM_CONV:
            ggml_sycl_ssm_conv(ctx, dst);
+            break;
        case GGML_OP_ROLL:
            ggml_sycl_roll(ctx, dst);
            break;
@@ -586,7 +586,6 @@ struct vk_device_struct {
    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_COUNT];

    vk_pipeline pipeline_matmul_split_k_reduce;
-    vk_pipeline pipeline_quantize_q8_1;
    vk_pipeline pipeline_quantize_q8_1_x4;

    vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
@@ -621,7 +620,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_add_id_f32;

    vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
-    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32;
+    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32;
    vk_pipeline pipeline_scale_f32;
    vk_pipeline pipeline_sqr_f32;
    vk_pipeline pipeline_sqrt_f32;
@@ -830,6 +829,7 @@ struct vk_mat_vec_push_constants {
    uint32_t batch_stride_b;
    uint32_t batch_stride_d;
    uint32_t enable_bias;
+    uint32_t enable_scale;
    uint32_t ne02;
    uint32_t ne12;
    uint32_t broadcast2;
@@ -852,6 +852,7 @@ struct vk_mat_vec_id_push_constants {
    uint32_t batch_stride_b;
    uint32_t batch_stride_d;
    uint32_t enable_bias;
+    uint32_t enable_scale;
    uint32_t nei0;
    uint32_t ne11;
 };
@@ -2158,17 +2159,18 @@ static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
    }
 }

+static std::vector<uint32_t> ggml_vk_find_memory_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    std::vector<uint32_t> indices;

-static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
    for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
        vk::MemoryType memory_type = mem_props->memoryTypes[i];
        if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
            (flags & memory_type.propertyFlags) == flags &&
            mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
-            return static_cast<int32_t>(i);
+            indices.push_back(i);
        }
    }
-    return UINT32_MAX;
+    return indices;
 }

 static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std::initializer_list<vk::MemoryPropertyFlags> & req_flags_list) {
@@ -2211,24 +2213,33 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
    for (auto it = req_flags_list.begin(); it != req_flags_list.end(); it++) {
        const auto & req_flags = *it;

-        uint32_t memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
+        const std::vector<uint32_t> memory_type_indices = ggml_vk_find_memory_properties(&mem_props, &mem_req, req_flags);

-        if (memory_type_index == UINT32_MAX) {
+        if (memory_type_indices.empty()) {
            continue;
        }
        buf->memory_property_flags = req_flags;

-        try {
-            buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index, &mem_flags_info });
-            break;
-        } catch (const vk::SystemError& e) {
-            // loop and retry
-            // during last attempt throw the exception
-            if (it + 1 == req_flags_list.end()) {
-                device->device.destroyBuffer(buf->buffer);
-                throw e;
+        bool done = false;
+
+        for (auto mtype_it = memory_type_indices.begin(); mtype_it != memory_type_indices.end(); mtype_it++) {
+            try {
+                buf->device_memory = device->device.allocateMemory({ mem_req.size, *mtype_it, &mem_flags_info });
+                done = true;
+                break;
+            } catch (const vk::SystemError& e) {
+                // loop and retry
+                // during last attempt throw the exception
+                if (it + 1 == req_flags_list.end() && mtype_it + 1 == memory_type_indices.end()) {
+                    device->device.destroyBuffer(buf->buffer);
+                    throw e;
+                }
            }
        }
+
+        if (done) {
+            break;
+        }
    }

    if (!buf->device_memory) {
@@ -3556,10 +3567,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);

    if (device->subgroup_clustered && device->subgroup_require_full_support) {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_subgroup_len, quantize_q8_1_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
    } else {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
    }

@@ -3693,6 +3702,7 @@ static void ggml_vk_load_shaders(vk_device& device) {

    ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);

@@ -6259,20 +6269,20 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
    ggml_vk_sync_buffers(ctx, subctx);
 }

-static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type, bool use_x4_blocks) {
+static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
    switch(type) {
        case GGML_TYPE_Q8_1:
-            return use_x4_blocks ? ctx->device->pipeline_quantize_q8_1_x4 : ctx->device->pipeline_quantize_q8_1;
+            return ctx->device->pipeline_quantize_q8_1_x4;
        default:
            std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl;
            GGML_ABORT("fatal error");
    }
 }

-static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne, bool use_x4_blocks = false) {
+static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne) {
    VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")");

-    vk_pipeline pipeline = use_x4_blocks ? ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true) : ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, false);
+    vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);

    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
    ggml_vk_sync_buffers(ctx, subctx);
@@ -6363,16 +6373,17 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub

    // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
    uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
-    const int x_ne = ne01 * ne00;
-    const int y_ne = padded_n * ne10;
-    const int d_ne = ne11 * ne01;
+    const uint64_t x_ne = ggml_nelements(src0);
+    // 128 elements per Q8_1 x4 block
+    const uint64_t y_ne = padded_n * ne10 * ne12 * ne13;
+    const uint64_t d_ne = ggml_nelements(dst);

    const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, disable_split_k, pipeline);

    const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
    const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
    const uint64_t d_sz = sizeof(float) * d_ne;

    vk_pipeline to_fp16_vk_0 = nullptr;
@@ -6393,28 +6404,23 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT

    if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
    }

    {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
-        const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0;
+        const uint64_t split_k_size = split_k > 1 ? d_sz * split_k : 0;
        if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
                (split_k > 1 && split_k_size > ctx->device->properties.limits.maxStorageBufferRange)) {
            GGML_ABORT("Requested preallocation size is too large");
        }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }
        if (split_k > 1 && ctx->prealloc_size_split_k < split_k_size) {
@@ -6441,7 +6447,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
    vk_buffer d_D = dst_buf_ctx->dev_buffer;
    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
    GGML_ASSERT(d_D != nullptr);
-    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
+    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz);
    vk_buffer d_X;
    uint64_t x_buf_offset = 0;
    vk_buffer d_Y;
@@ -6458,7 +6464,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
    }
    if (qx_needs_dequant) {
        d_X = ctx->prealloc_x;
-        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+        GGML_ASSERT(d_X->size >= x_sz);
    } else {
        d_X = d_Qx;
        x_buf_offset = qx_buf_offset;
@@ -6466,10 +6472,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
    }
    if (qy_needs_dequant) {
        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
+        GGML_ASSERT(d_Y->size >= y_sz);
    } else if (quantize_y) {
        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz, 144) * 144);
    } else {
        d_Y = d_Qy;
        y_buf_offset = qy_buf_offset;
@@ -6486,7 +6492,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0));
    } else if (qx_needs_dequant) {
        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)(x_ne), 1, 1});
        ggml_vk_sync_buffers(ctx, subctx);
    }
    if (y_non_contig) {
@@ -6506,7 +6512,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
            if (ctx->prealloc_y_need_sync) {
                ggml_vk_sync_buffers(ctx, subctx);
            }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
            ctx->prealloc_y_last_tensor_used = src1;
        }
@@ -6523,16 +6529,11 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
    }

-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-    }
-
    // compute
    ggml_vk_matmul(
        ctx, subctx, pipeline,
-        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
-        ggml_vk_subbuffer(ctx, d_D, d_buf_offset), { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
+        { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
+        ggml_vk_subbuffer(ctx, d_D, d_buf_offset), { ctx->prealloc_split_k, 0, d_sz * split_k },
        ne01, ne11, ne10,
        ne10, ne10, stride_d, stride_batch_x, stride_batch_y, stride_batch_d,
        split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n
@@ -6615,8 +6616,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&

    const uint64_t ne20 = dst->ne[0];
    const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];

    const uint64_t r2 = ne12 / ne02;
    const uint64_t r3 = ne13 / ne03;
@@ -6672,7 +6673,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
    }

    if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
    }

    const bool qx_needs_dequant = x_non_contig;
@@ -6685,33 +6686,29 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
    GGML_ASSERT(dmmv != nullptr);

-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne11 * ne01;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = ggml_nelements(src1);
+    const uint64_t d_ne = ggml_nelements(dst);

    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) :
+                         (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
    const uint64_t d_sz = sizeof(float) * d_ne;

    {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
        if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
            GGML_ABORT("Requested preallocation size is too large");
        }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }

@@ -6768,7 +6765,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
        d_Y = ctx->prealloc_y;
    } else if (quantize_y) {
        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz, 144) * 144);
    } else {
        d_Y = d_Qy;
        y_buf_offset = qy_buf_offset;
@@ -6801,7 +6798,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
            if (ctx->prealloc_y_need_sync) {
                ggml_vk_sync_buffers(ctx, subctx);
            }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
            ctx->prealloc_y_last_tensor_used = src1;
        }
@@ -6830,17 +6827,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
        groups_x = CEIL_DIV(groups_x, groups_z);
    }

-    // TODO: Clean up this whole sz * ne_2 * ne_3 thing, it hasn't been necessary for a long time
-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-    }
-
    uint32_t enable_bias = ctx->num_additional_fused_ops > 0;

    vk_buffer d_B = d_D;
    size_t b_buf_offset = 0;
-    uint64_t b_sz = 0;
+    uint64_t b_sz = 1;

    if (enable_bias) {
        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
@@ -6863,14 +6854,14 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
    // compute
    const vk_mat_vec_push_constants pc = {
        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        stride_batch_x, stride_batch_y, stride_batch_d, enable_bias,
+        stride_batch_x, stride_batch_y, stride_batch_d, enable_bias, 0,
        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
    };
    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                              {
-                                vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
-                                vk_subbuffer{ d_Y, y_buf_offset, y_sz_total },
-                                vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23},
+                                vk_subbuffer{ d_X, x_buf_offset, x_sz },
+                                vk_subbuffer{ d_Y, y_buf_offset, y_sz },
+                                vk_subbuffer{ d_D, d_buf_offset, d_sz },
                                vk_subbuffer{ d_B, b_buf_offset, b_sz },
                              },
                              pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
@@ -6974,7 +6965,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c

    vk_buffer d_B = d_D;
    size_t b_buf_offset = 0;
-    uint64_t b_sz = 0;
+    uint64_t b_sz = 1;

    if (enable_bias) {
        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
@@ -7110,7 +7101,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con

    vk_buffer d_B = d_D;
    size_t b_buf_offset = 0;
-    uint64_t b_sz = 0;
+    uint64_t b_sz = 1;

    if (enable_bias) {
        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
@@ -7208,7 +7199,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
    const uint64_t ne00 = src0->ne[0];
    const uint64_t ne01 = src0->ne[1];
    const uint64_t ne02 = src0->ne[2];
-    const uint64_t ne03 = src0->ne[3];
+    // const uint64_t ne03 = src0->ne[3];

    const uint64_t ne10 = src1->ne[0];
    const uint64_t ne11 = src1->ne[1];
@@ -7223,8 +7214,8 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&

    const uint64_t ne20 = dst->ne[0];
    const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];

    const uint64_t n_as = ne02;

@@ -7294,14 +7285,14 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&

    // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
    uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = padded_n * ne10;
-    const uint64_t d_ne = ne21 * ne20;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = padded_n * ne10 * ne12 * ne13;
+    const uint64_t d_ne = ggml_nelements(dst);

    const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
    const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
    const uint64_t ids_sz = nbi2;
    const uint64_t d_sz = sizeof(float) * d_ne;

@@ -7323,26 +7314,21 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT

    if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
    }

    {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
        if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
            GGML_ABORT("Requested preallocation size is too large");
        }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }

@@ -7383,7 +7369,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
    }
    if (qx_needs_dequant) {
        d_X = ctx->prealloc_x;
-        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+        GGML_ASSERT(d_X->size >= x_sz);
    } else {
        d_X = d_Qx;
        x_buf_offset = qx_buf_offset;
@@ -7391,10 +7377,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
    }
    if (qy_needs_dequant) {
        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
+        GGML_ASSERT(d_Y->size >= y_sz);
    } else if (quantize_y) {
        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz, 144) * 144);
    } else {
        d_Y = d_Qy;
        y_buf_offset = qy_buf_offset;
@@ -7412,7 +7398,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
    } else if (qx_needs_dequant) {
        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
-            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)x_ne, 1, 1});
        ggml_vk_sync_buffers(ctx, subctx);
    }
    if (y_non_contig) {
@@ -7432,7 +7418,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
            if (ctx->prealloc_y_need_sync) {
                ggml_vk_sync_buffers(ctx, subctx);
            }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
            ctx->prealloc_y_last_tensor_used = src1;
        }
@@ -7449,16 +7435,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
    }

-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-    }
-
    // compute
    ggml_vk_matmul_id(
        ctx, subctx, pipeline,
-        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
-        { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
+        { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
+        { d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz },
        ne01, ne21, ne10, ne10, ne10, ne01,
        stride_batch_x, stride_batch_y, ne20*ne21,
        n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
@@ -7488,13 +7469,13 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte

    const uint64_t ne00 = src0->ne[0];
    const uint64_t ne01 = src0->ne[1];
-    const uint64_t ne02 = src0->ne[2];
-    const uint64_t ne03 = src0->ne[3];
+    // const uint64_t ne02 = src0->ne[2];
+    // const uint64_t ne03 = src0->ne[3];

    const uint64_t ne10 = src1->ne[0];
    const uint64_t ne11 = src1->ne[1];
-    const uint64_t ne12 = src1->ne[2];
-    const uint64_t ne13 = src1->ne[3];
+    // const uint64_t ne12 = src1->ne[2];
+    // const uint64_t ne13 = src1->ne[3];

    const uint64_t nei0 = ids->ne[0];
    const uint64_t nei1 = ids->ne[1];
@@ -7505,8 +7486,8 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte

    const uint64_t ne20 = dst->ne[0];
    const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];

    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
@@ -7543,9 +7524,9 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
    // Not implemented
    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT

-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne21 * ne20;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = ggml_nelements(src1);
+    const uint64_t d_ne = ggml_nelements(dst);

    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
@@ -7570,19 +7551,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
    GGML_ASSERT(dmmv != nullptr);

    {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
        if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
            GGML_ABORT("Requested preallocation size is too large");
        }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }
-        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
            ggml_vk_preallocate_buffers(ctx, subctx);
        }

@@ -7684,13 +7663,22 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
        groups_x = CEIL_DIV(groups_x, groups_z);
    }

-    uint32_t enable_bias = ctx->num_additional_fused_ops > 0;
+    uint32_t enable_bias = 0;
+    uint32_t enable_scale = 0;
+    if (ctx->num_additional_fused_ops > 0) {
+        if (cgraph->nodes[node_idx + 1]->op == GGML_OP_MUL) {
+            enable_scale = 1;
+        } else {
+            GGML_ASSERT(cgraph->nodes[node_idx + 1]->op == GGML_OP_ADD_ID);
+            enable_bias = 1;
+        }
+    }

    vk_buffer d_B = d_D;
    size_t b_buf_offset = 0;
-    uint64_t b_sz = 0;
+    uint64_t b_sz = 1;

-    if (enable_bias) {
+    if (enable_bias || enable_scale) {
        const ggml_tensor * bias = cgraph->nodes[node_idx + 1]->src[1];

        bool b_uma = false;
@@ -7710,17 +7698,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
    // compute
    const vk_mat_vec_id_push_constants pc = {
        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),

-        enable_bias,
+        enable_bias, enable_scale,

        (uint32_t)nei0, (uint32_t)ne11,
    };
    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
        {
-            vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
-            vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 },
-            vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23},
+            vk_subbuffer{ d_X, x_buf_offset, x_sz },
+            vk_subbuffer{ d_Y, y_buf_offset, y_sz },
+            vk_subbuffer{ d_D, d_buf_offset, d_sz },
            vk_subbuffer{ d_B, b_buf_offset, b_sz },
            vk_subbuffer{ d_ids, ids_buf_offset, ids_sz },
        },
@@ -8213,6 +8201,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                    return ctx->device->pipeline_upscale_nearest_f32;
                case GGML_SCALE_MODE_BILINEAR:
                    return ctx->device->pipeline_upscale_bilinear_f32;
+                case GGML_SCALE_MODE_BICUBIC:
+                    return ctx->device->pipeline_upscale_bicubic_f32;
                default:
                    return nullptr;
            }
@@ -12490,6 +12480,40 @@ static bool ggml_vk_can_fuse(const ggml_backend_vk_context * ctx, const struct g
        }
    }

+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_MUL_MAT_ID && ops.begin()[1] == GGML_OP_MUL) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *mmid = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+        const ggml_tensor *scale = mul->src[1];
+
+        if (mmid != mul->src[0]) {
+            return false;
+        }
+        // mat-vec only
+        if (!ggml_vk_use_mul_mat_vec_id(cgraph, node_idx)) {
+            return false;
+        }
+        // shaders assume the types match
+        if (mmid->type != scale->type) {
+            return false;
+        }
+        // shaders assume the bias is contiguous
+        if (!ggml_is_contiguous(scale)) {
+            return false;
+        }
+        // unaligned bias isn't handled
+        if (get_misalign_bytes(ctx, scale) != 0) {
+            return false;
+        }
+        // shader only indexes by expert index
+        if (scale->ne[0] != 1 ||
+            scale->ne[1] != mul->ne[1] ||
+            scale->ne[2] != 1 ||
+            scale->ne[3] != 1) {
+            return false;
+        }
+    }
+
    return true;
 }

@@ -12646,6 +12670,12 @@ static bool ggml_vk_can_fuse_rms_norm_mul_rope(ggml_backend_vk_context * ctx, co
        return false;
    }

+    // conditions for pipeline creation
+    if (!(ctx->device->float_controls_rte_fp16 &&
+        sizeof(vk_op_rms_norm_mul_rope_push_constants) <= ctx->device->properties.limits.maxPushConstantsSize)) {
+        return false;
+    }
+
    return true;
 }

@@ -12798,6 +12828,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
                ctx->num_additional_fused_ops = 1;
            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID })) {
                ctx->num_additional_fused_ops = 1;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT_ID, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 4 }) &&
                       ggml_check_edges(cgraph, i, rms_norm_mul_rope_view_set_rows_edges) &&
                       ggml_vk_can_fuse_rms_norm_mul_rope(ctx, cgraph, i) &&
@@ -13033,7 +13065,8 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
                    is_src_of(graph->nodes[j], graph->nodes[c]) &&
                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_RMS_NORM && graph->nodes[j]->op == GGML_OP_MUL) &&
                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT && graph->nodes[j]->op == GGML_OP_ADD) &&
-                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT_ID && graph->nodes[j]->op == GGML_OP_ADD_ID)) {
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT_ID && graph->nodes[j]->op == GGML_OP_ADD_ID) &&
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT_ID && graph->nodes[j]->op == GGML_OP_MUL)) {
                    ok = false;
                    break;
                }
@@ -13190,25 +13223,28 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total
    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
    vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
    vk::PhysicalDeviceMemoryProperties2 memprops = {};
-    bool membudget_supported = vk_instance.device_supports_membudget[device];
+    const bool membudget_supported = vk_instance.device_supports_membudget[device];
+    const bool is_integrated_gpu = vkdev.getProperties().deviceType == vk::PhysicalDeviceType::eIntegratedGpu;

    if (membudget_supported) {
        memprops.pNext = &budgetprops;
    }
    vkdev.getMemoryProperties2(&memprops);

+    *total = 0;
+    *free = 0;
+
    for (uint32_t i = 0; i < memprops.memoryProperties.memoryHeapCount; ++i) {
        const vk::MemoryHeap & heap = memprops.memoryProperties.memoryHeaps[i];

-        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-            *total = heap.size;
+        if (is_integrated_gpu || (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal)) {
+            *total += heap.size;

            if (membudget_supported && i < budgetprops.heapUsage.size()) {
-                *free = budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
+                *free += budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
            } else {
-                *free = heap.size;
+                *free += heap.size;
            }
-            break;
        }
    }
 }
@@ -49,6 +49,7 @@ layout (push_constant) uniform parameter
    uint batch_stride_d;

    uint enable_bias;
+    uint enable_scale;

 #ifdef MUL_MAT_ID
    uint nei0;
@@ -129,6 +130,12 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
 #endif
                }
+#ifdef MUL_MAT_ID
+                if (p.enable_scale != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_bias[expert_idx]);
+                }
+#endif
                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
            }
        }
@@ -171,6 +178,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
 #endif
                }
+#ifdef MUL_MAT_ID
+                if (p.enable_scale != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_bias[expert_idx]);
+                }
+#endif
                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
            }
        }
@@ -203,6 +216,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                    tmpsh[j][n][0] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
 #endif
                }
+#ifdef MUL_MAT_ID
+                if (p.enable_scale != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_bias[expert_idx]);
+                }
+#endif
                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
            }
        }
@@ -211,7 +211,9 @@ void main() {
            const uint iqs = loadr_a;

            [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-                block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
+                if (block + k_step * BK < end_k) {
+                    block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
+                }
            }
        }
        [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
@@ -226,7 +228,7 @@ void main() {
            const uint iqs = loadr_b;

            [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-                block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs);
+                block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs, block + k_step * BK < end_k);
            }
        }

@@ -469,19 +469,30 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
 #endif

 #ifdef MMQ_SHMEM
-void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-    const uint ib_outer = ib / 4;
-    const uint ib_inner = ib % 4;
+void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
+    if (is_in_bounds) {
+        const uint ib_outer = ib / 4;
+        const uint ib_inner = ib % 4;

-    if (iqs == 0) {
-        buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+        }
+
+        const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
+        buf_b[buf_ib].qs[iqs * 4    ] = values.x;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
+    } else {
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
+        }
+
+        buf_b[buf_ib].qs[iqs * 4    ] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
    }
-
-    const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
-    buf_b[buf_ib].qs[iqs * 4    ] = values.x;
-    buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
-    buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
-    buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
 }

 void block_b_to_registers(const uint ib) {
@@ -61,7 +61,7 @@ void quantize() {

    const uint a_idx = ib * 8 + iqs;

-    vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f);
+    vec4 vals = a_idx < p.ne / 4 ? data_a[a_idx] : vec4(0.0f);
    const vec4 abs_vals = abs(vals);

    // Find absolute max for each block
@@ -20,6 +20,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 // from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
 #define NEAREST  0
 #define BILINEAR 1
+#define BICUBIC  2

 layout (constant_id = 0) const uint scale_mode = 0;

@@ -61,6 +62,39 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
    return fetch_bilinear(c0, c1, d, i12, i13);
 }

+// Bicubic interpolation with alpha = -0.75
+// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+const vec4 bcoeffs1 = vec4( 1.25, -2.25,  0.0, 1.0);
+const vec4 bcoeffs2 = vec4(-0.75,  3.75, -6.0, 3.0);
+vec4 powers(float x) { return vec4(x*x*x, x*x, x, 1); }
+
+float bicubic(float p0, float p1, float p2, float p3, float x) {
+    return p0 * dot(bcoeffs2, powers(x + 1)) +
+           p1 * dot(bcoeffs1, powers(x    )) +
+           p2 * dot(bcoeffs1, powers(1 - x)) +
+           p3 * dot(bcoeffs2, powers(2 - x));
+}
+
+#define FETCH(a,b) data_a[base + clamp(i.x+(a), 0, res.x) * p.nb00 + clamp(i.y+(b), 0, res.y) * p.nb01]
+
+float interpolate_bicubic(uint i10, uint i11, uint i12, uint i13) {
+    const ivec2 res = ivec2(p.ne00 - 1, p.ne01 - 1);
+
+    const vec2 coord = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
+    const vec2 d = fract(coord);
+    const ivec2 i = ivec2(floor(coord));
+
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+    const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
+
+    return bicubic(
+        bicubic(FETCH(-1,-1), FETCH(0,-1), FETCH(1,-1), FETCH(2,-1), d.x),
+        bicubic(FETCH(-1, 0), FETCH(0, 0), FETCH(1, 0), FETCH(2, 0), d.x),
+        bicubic(FETCH(-1, 1), FETCH(0, 1), FETCH(1, 1), FETCH(2, 1), d.x),
+        bicubic(FETCH(-1, 2), FETCH(0, 2), FETCH(1, 2), FETCH(2, 2), d.x), d.y);
+}
+
 void main() {
    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;

@@ -81,6 +115,9 @@ void main() {
        case BILINEAR:
            result = interpolate_bilinear(i10, i11, i12, i13);
            break;
+        case BICUBIC:
+            result = interpolate_bicubic(i10, i11, i12, i13);
+            break;
    }

    data_d[p.d_offset + idx] = D_TYPE(result);
@@ -18,6 +18,7 @@
 #include <algorithm>
 #include <sys/stat.h>
 #include <sys/types.h>
+#include <filesystem>

 #ifdef _WIN32
    #define NOMINMAX
@@ -75,7 +76,7 @@ enum MatMulIdType {

 namespace {

-void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
+void execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
    HANDLE stdout_read, stdout_write;
    HANDLE stderr_read, stderr_write;
@@ -98,8 +99,10 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
    si.hStdOutput = stdout_write;
    si.hStdError = stderr_write;

-    std::vector<char> cmd(command.begin(), command.end());
-    cmd.push_back('\0');
+    std::string cmd;
+    for (const auto& part : command) {
+        cmd += part + " ";
+    }

    if (!CreateProcessA(NULL, cmd.data(), NULL, NULL, TRUE, 0, NULL, NULL, &si, &pi)) {
        throw std::runtime_error("Failed to create process");
@@ -137,6 +140,12 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
        throw std::runtime_error("Failed to fork process");
    }

+    std::vector<char*> argv;
+    for (std::string& part : command) {
+        argv.push_back(part.data());
+    }
+    argv.push_back(nullptr);
+
    if (pid == 0) {
        close(stdout_pipe[0]);
        close(stderr_pipe[0]);
@@ -144,7 +153,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
        dup2(stderr_pipe[1], STDERR_FILENO);
        close(stdout_pipe[1]);
        close(stderr_pipe[1]);
-        execl("/bin/sh", "sh", "-c", command.c_str(), (char*) nullptr);
+        execvp(argv[0], argv.data());
        _exit(EXIT_FAILURE);
    } else {
        close(stdout_pipe[1]);
@@ -315,21 +324,27 @@ compile_count_guard acquire_compile_slot() {
 void string_to_spv_func(std::string name, std::string in_path, std::string out_path, std::map<std::string, std::string> defines, bool coopmat, bool dep_file, compile_count_guard slot) {
    std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";

+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, in_path, "-o", out_path};
+    #endif
+
    // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
    // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
    // disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
-    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos || name.find("rope") != std::string::npos) ? "" : "-O";
-
-    #ifdef _WIN32
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
-    #else
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o",  out_path};
-    #endif
+    if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos) {
+        cmd.push_back("-O");
+    }

    if (dep_file) {
        cmd.push_back("-MD");
        cmd.push_back("-MF");
+#ifdef _WIN32
        cmd.push_back("\"" + target_cpp + ".d\"");
+#else
+        cmd.push_back(target_cpp + ".d");
+#endif
    }

    #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
@@ -353,9 +368,13 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
        // }
        // std::cout << std::endl;

-        execute_command(command, stdout_str, stderr_str);
+        execute_command(cmd, stdout_str, stderr_str);
        if (!stderr_str.empty()) {
-            std::cerr << "cannot compile " << name << "\n\n" << command << "\n\n" << stderr_str << std::endl;
+            std::cerr << "cannot compile " << name << "\n\n";
+            for (const auto& part : cmd) {
+                std::cerr << part << " ";
+            }
+            std::cerr << "\n\n" << stderr_str << std::endl;
            return;
        }

@@ -429,7 +448,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
    base_dict["ACC_TYPE"     ] = f16acc ? "float16_t" : "float";
    base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2"   : "vec2";
    if (f16acc) {
-        base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
+        base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
    }

    if (coopmat) {
@@ -609,7 +628,7 @@ void process_shaders() {
        fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
        fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
        if (f16acc) {
-            fa_base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
+            fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
        }

        for (const auto& tname : type_names) {
@@ -48,13 +48,18 @@ class LazyMeta(ABCMeta):
        # NOTE: doing this from a metaclass is very convenient
        # TODO: make this even more comprehensive
        for binary_op in (
-            "lt", "le", "eq", "ne", "ge", "gt", "not"
-            "abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul",
-            "neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor",
+            "lt", "le", "eq", "ne", "ge", "gt",
+            "add", "and", "floordiv", "lshift", "mod", "mul", "matmul",
+            "or", "pow", "rshift", "sub", "truediv", "xor",
            "iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor",
            "radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor",
        ):
            attr_name = f"__{binary_op}__"
+            # evaluation on the meta tensor is needed in case there's broadcasting
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=False)
+
+        for unary_op in ("not", "abs", "invert", "neg", "pos"):
+            attr_name = f"__{unary_op}__"
            # the result of these operators usually has the same shape and dtype as the input,
            # so evaluation on the meta tensor can be skipped.
            namespace[attr_name] = mk_wrap(attr_name, meta_noop=True)
@@ -1,10 +1,12 @@
 from __future__ import annotations

 from dataclasses import dataclass
+from pathlib import Path
 from typing import Literal

 import os
 import json
+import numpy as np


 def fill_templated_filename(filename: str, output_type: str | None) -> str:
@@ -177,6 +179,10 @@ class SafetensorRemote:
            except KeyError as e:
                raise ValueError(f"Missing key in metadata for tensor '{name}': {e}, meta = {meta}")

+        # order by name (same as default safetensors behavior)
+        # ref: https://github.com/huggingface/safetensors/blob/0816a1ae1d6b731cefd67f061d80d1cadd0dd7bb/bindings/python/src/lib.rs#L606
+        res = dict(sorted(res.items(), key=lambda t: t[0]))
+
        return res

    @classmethod
@@ -266,3 +272,77 @@ class SafetensorRemote:
        if os.environ.get("HF_TOKEN"):
            headers["Authorization"] = f"Bearer {os.environ['HF_TOKEN']}"
        return headers
+
+
+@dataclass
+class LocalTensorRange:
+    filename: Path
+    offset: int
+    size: int
+
+
+@dataclass
+class LocalTensor:
+    dtype: str
+    shape: tuple[int, ...]
+    data_range: LocalTensorRange
+
+    def mmap_bytes(self) -> np.ndarray:
+        return np.memmap(self.data_range.filename, offset=self.data_range.offset, shape=self.data_range.size)
+
+
+class SafetensorsLocal:
+    """
+        Read a safetensors file from the local filesystem.
+
+        Custom parsing gives a bit more control over the memory usage.
+        The official safetensors library doesn't expose file ranges.
+    """
+    ALIGNMENT = 8  # bytes
+
+    tensors: dict[str, LocalTensor]
+
+    def __init__(self, filename: Path):
+        with open(filename, "rb") as f:
+            metadata_length = int.from_bytes(f.read(8), byteorder='little')
+            file_size = os.stat(filename).st_size
+            if file_size < 8 + metadata_length:
+                raise ValueError(f"Could not read complete metadata. Need {8 + metadata_length} bytes, got {file_size}")
+
+            metadata_str = f.read(metadata_length).decode('utf-8')
+            try:
+                metadata = json.loads(metadata_str)
+            except json.JSONDecodeError as e:
+                raise ValueError(f"Failed to parse safetensors metadata as JSON: {e}")
+
+            data_start_offset = f.tell()
+            alignment = self.ALIGNMENT
+            if data_start_offset % alignment != 0:
+                data_start_offset += alignment - (data_start_offset % alignment)
+
+            tensors: dict[str, LocalTensor] = {}
+            for name, meta in metadata.items():
+                if name == "__metadata__":
+                    # ignore metadata, it's not a tensor
+                    continue
+
+                tensors[name] = LocalTensor(
+                    dtype=meta["dtype"],
+                    shape=tuple(meta["shape"]),
+                    data_range=LocalTensorRange(
+                        filename,
+                        data_start_offset + meta["data_offsets"][0],
+                        meta["data_offsets"][1] - meta["data_offsets"][0],
+                    ),
+                )
+
+            # order by name (same as default safetensors behavior)
+            # ref: https://github.com/huggingface/safetensors/blob/0816a1ae1d6b731cefd67f061d80d1cadd0dd7bb/bindings/python/src/lib.rs#L606
+            self.tensors = dict(sorted(tensors.items(), key=lambda t: t[0]))
+
+    def __enter__(self, *args, **kwargs):
+        del args, kwargs  # unused
+        return self.tensors
+
+    def __exit__(self, *args, **kwargs):
+        del args, kwargs  # unused
@@ -12,11 +12,30 @@ vendor = {

    "https://raw.githubusercontent.com/nothings/stb/refs/heads/master/stb_image.h": "vendor/stb/stb_image.h",

-    "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.22/miniaudio.h": "vendor/miniaudio/miniaudio.h",
+    # not using latest tag to avoid this issue: https://github.com/ggml-org/llama.cpp/pull/17179#discussion_r2515877926
+    # "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/v0.20.1/httplib.h": "vendor/cpp-httplib/httplib.h",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.27.0/httplib.h": "vendor/cpp-httplib/httplib.h",
 }

 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')
@@ -132,6 +132,11 @@ add_library(llama
            models/graph-context-mamba.cpp
            )

+set_target_properties(llama PROPERTIES
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
+)
+
 target_include_directories(llama PRIVATE .)
 target_include_directories(llama PUBLIC ../include)
 target_compile_features   (llama PRIVATE cxx_std_17) # don't bump
@@ -1592,9 +1592,10 @@ ggml_tensor * llm_graph_context::build_attn(
            int       il) const {
    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
    ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
    ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);

    const auto * mctx_cur = inp->mctx;

@@ -151,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
        p1 = std::numeric_limits<llama_pos>::max();
    }

-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
    if (seq_id >= (int64_t) size) {
        // could be fatal
        return false;
@@ -160,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
        int32_t & tail_id = cells[seq_id].tail;
        if (tail_id >= 0) {
            const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                return false;
            }
@@ -1013,7 +1013,7 @@ private:
        }
    private:
        uint32_t get_node(size_t index) {
-            if (index > xcda_array_size) {
+            if (index >= xcda_array_size) {
                throw std::runtime_error("Index out of array bounds in XCDA array!");
            }
            return xcda_array[index];
@@ -1,7 +1,5 @@
 #include "models.h"

-
-
 llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
    llm_graph_context(params) {
    const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -19,6 +17,8 @@ llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_grap

    auto * inp_attn = build_attn_inp_kv();

+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
    for (int il = 0; il < n_layer; ++il) {
        ggml_tensor * inpSA = inpL;

@@ -67,9 +67,8 @@ llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_grap
        }
        if (il == n_layer - 1) {
            // skip computing output for unused tokens
-            ggml_tensor * inp_out_ids = build_inp_out_ids();
-            cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
-            inpSA                     = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);
@@ -11,6 +11,8 @@ llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model,

    auto * inp_attn = build_attn_inp_kv_iswa();

+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
    for (int il = 0; il < n_layer; ++il) {
        ggml_tensor * inpSA = inpL;

@@ -69,7 +71,6 @@ llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model,
        }
        if (il == n_layer - 1) {
            // skip computing output for unused tokens
-            ggml_tensor * inp_out_ids = build_inp_out_ids();
            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }
@@ -272,6 +272,10 @@ static double mean_abs_asymm(const float * a, const float * b, const size_t n, c

 // utils for printing the variables of the test cases

+static std::string var_to_str(const std::string & x) {
+    return x;
+}
+
 template<typename T>
 static std::string var_to_str(const T & x) {
    return std::to_string(x);
@@ -323,7 +327,8 @@ static std::string var_to_str(ggml_scale_mode mode) {
    switch (mode) {
        case GGML_SCALE_MODE_NEAREST:  return "nearest";
        case GGML_SCALE_MODE_BILINEAR: return "bilinear";
-        default:                      return std::to_string(mode);
+        case GGML_SCALE_MODE_BICUBIC:  return "bicubic";
+        default:                       return std::to_string(mode);
    }
 }

@@ -3557,6 +3562,27 @@ struct test_mul_mat : public test_case {
    }
 };

+static void init_mul_mat_id_tensors(ggml_context * ctx, int n_mats) {
+    std::random_device rd;
+    std::default_random_engine rng(rd());
+    for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->type == GGML_TYPE_I32) {
+            if (ggml_is_view_op(t->op)) { continue; }
+            // ids
+            for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                std::vector<int32_t> data(t->ne[0]);
+                for (int i = 0; i < t->ne[0]; i++) {
+                    data[i] = i % n_mats;
+                }
+                std::shuffle(data.begin(), data.end(), rng);
+                ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
+            }
+        } else {
+            init_tensor_uniform(t);
+        }
+    }
+}
+
 // GGML_OP_MUL_MAT_ID
 struct test_mul_mat_id : public test_case {
    const ggml_type type_a;
@@ -3567,10 +3593,9 @@ struct test_mul_mat_id : public test_case {
    const int64_t m;
    const int64_t n;
    const int64_t k;
-    const uint32_t o; // number of outputs

    std::string vars() override {
-        return VARS_TO_STR9(type_a, type_b, n_mats, n_used, b, m, n, k, o);
+        return VARS_TO_STR8(type_a, type_b, n_mats, n_used, b, m, n, k);
    }

    double max_nmse_err() override {
@@ -3584,9 +3609,69 @@ struct test_mul_mat_id : public test_case {

    test_mul_mat_id(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
            int n_mats = 8, int n_used = 2, bool b = false,
-            int64_t m = 32, int64_t n = 32, int64_t k = 32, uint32_t o = 1)
+            int64_t m = 32, int64_t n = 32, int64_t k = 32)
        : type_a(type_a), type_b(type_b), n_mats(n_mats), n_used(n_used), b(b),
-            m(m), n(n), k(k), o(o) {
+            m(m), n(n), k(k) {
+            GGML_ASSERT(n_used <= n_mats);
+        }
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        // C^T = A * B^T: (k, m) * (k, n) => (m, n)
+        ggml_tensor * as = ggml_new_tensor_3d(ctx, type_a, k, m, n_mats);
+        ggml_set_name(as, "as");
+
+        ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_mats, n);
+        ggml_set_name(ids, "ids");
+        if (n_used != n_mats) {
+            ids = ggml_view_2d(ctx, ids, n_used, n, ids->nb[1], 0);
+            ggml_set_name(ids, "view_of_ids");
+        }
+
+        ggml_tensor * b = ggml_new_tensor_3d(ctx, type_b, k, this->b ? 1 : n_used, n);
+        ggml_set_name(b, "b");
+
+        ggml_tensor * out = ggml_mul_mat_id(ctx, as, b, ids);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        init_mul_mat_id_tensors(ctx, n_mats);
+    }
+};
+
+// GGML_OP_MUL_MAT_ID + GGML_OP_ADD or GGML_OP_MUL
+struct test_mul_mat_id_fusion : public test_case {
+    const ggml_type type_a;
+    const ggml_type type_b;
+    const int n_mats;
+    const int n_used;
+    const bool b; // broadcast b matrix
+    const int64_t m;
+    const int64_t n;
+    const int64_t k;
+    const uint32_t o; // number of outputs
+    const bool mul;
+
+    std::string vars() override {
+        return VARS_TO_STR10(type_a, type_b, n_mats, n_used, b, m, n, k, o, mul);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    uint64_t op_flops(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return 2 * m * k * n * n_used;
+    }
+
+    test_mul_mat_id_fusion(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
+            int n_mats = 8, int n_used = 2, bool b = false,
+            int64_t m = 32, int64_t n = 32, int64_t k = 32, uint32_t o = 1, bool mul = false)
+        : type_a(type_a), type_b(type_b), n_mats(n_mats), n_used(n_used), b(b),
+            m(m), n(n), k(k), o(o), mul(mul) {
            GGML_ASSERT(n_used <= n_mats);
        }

@@ -3615,35 +3700,25 @@ struct test_mul_mat_id : public test_case {
            out = ggml_add(ctx, out, out2);
        }

+        if (mul) {
+            std::array<int64_t, 4> ne { 1, out->ne[1], out->ne[2], out->ne[3] };
+            ne[0] = 1;
+            ggml_tensor * m = ggml_new_tensor(ctx, out->type, 4, ne.data());
+            out = ggml_mul(ctx, out, m);
+        }
+
        return out;
    }

    void initialize_tensors(ggml_context * ctx) override {
-        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-            if (t->type == GGML_TYPE_I32) {
-                if (ggml_is_view_op(t->op)) { continue; }
-                std::random_device rd;
-                std::default_random_engine rng(rd());
-                // ids
-                for (int64_t r = 0; r < ggml_nrows(t); r++) {
-                    std::vector<int32_t> data(t->ne[0]);
-                    for (int i = 0; i < t->ne[0]; i++) {
-                        data[i] = i % n_mats;
-                    }
-                    std::shuffle(data.begin(), data.end(), rng);
-                    ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
-                }
-            } else {
-                init_tensor_uniform(t);
-            }
-        }
+        init_mul_mat_id_tensors(ctx, n_mats);
    }

-    bool run_whole_graph() override { return o > 1; }
+    bool run_whole_graph() override { return true; }

    std::string op_desc(ggml_tensor * t) override {
        GGML_UNUSED(t);
-        return ggml_op_name(GGML_OP_MUL_MAT_ID);
+        return "MUL_MAT_ID_FUSION";
    }
 };

@@ -4992,24 +5067,7 @@ struct test_mul_mat_vec_fusion : public test_case {
                init_tensor_uniform(t);
            }
        } else {
-            std::random_device rd;
-            std::default_random_engine rng(rd());
-            for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-                if (t->type == GGML_TYPE_I32) {
-                    if (ggml_is_view_op(t->op)) { continue; }
-                    // ids
-                    for (int64_t r = 0; r < ggml_nrows(t); r++) {
-                        std::vector<int32_t> data(t->ne[0]);
-                        for (int i = 0; i < t->ne[0]; i++) {
-                            data[i] = i % n_mats;
-                        }
-                        std::shuffle(data.begin(), data.end(), rng);
-                        ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
-                    }
-                } else {
-                    init_tensor_uniform(t);
-                }
-            }
+            init_mul_mat_id_tensors(ctx, n_mats);
        }
    }

@@ -5165,7 +5223,9 @@ struct test_interpolate : public test_case {
    const uint32_t mode = GGML_SCALE_MODE_NEAREST;

    std::string vars() override {
-        return VARS_TO_STR4(type, ne, ne_tgt, mode);
+        ggml_scale_mode mode = (ggml_scale_mode)(this->mode & 0xFF);
+        std::string flags = (this->mode & GGML_SCALE_FLAG_ALIGN_CORNERS) ? "align_corners" : "none";
+        return VARS_TO_STR5(type, ne, ne_tgt, mode, flags);
    }

    test_interpolate(ggml_type type = GGML_TYPE_F32,
@@ -6934,6 +6994,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));

+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 576, 512, 576, {1,1}, {1,1}));
+
 #if 0
    // test the mat-mat path for Metal
    for (int k = 1; k < 512; ++k) {
@@ -6979,7 +7041,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
    }

    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));
-    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 32, 32, 32, 3));
+    test_cases.emplace_back(new test_mul_mat_id_fusion(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 32, 32, 32, 3));

    // gpt-oss issue with Vulkan mmq_id
    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_MXFP4, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
@@ -7016,6 +7078,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
        }
    }

+    for (int bs : {1, 4, 512}) {
+        for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q4_K}) {
+            for (ggml_type type_b : {GGML_TYPE_F32}) {
+                // test with mul after (ffn_moe_weighted)
+                test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 128, 8, false, 768, bs, 2048, 1, true));
+            }
+        }
+    }
+
    for (ggml_type type_a : base_types) {
        for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
            for (int n : {1, 16}) {
@@ -7224,15 +7295,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {2, 8, 8192, 1}, order)); // bailingmoe2 (group selection)
    }

-    for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
+    for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR, GGML_SCALE_MODE_BICUBIC}) {
        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, mode));
        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {5, 7, 11, 13}, {2, 5,  7, 11}, mode));
    }
-    test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS));
-    test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {1, 4, 3, 2}, {2, 8, 3, 2}, GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS));
-    test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {4, 1, 3, 2}, {1, 1, 3, 2}, GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS));
+    for (ggml_scale_mode mode : {GGML_SCALE_MODE_BILINEAR, GGML_SCALE_MODE_BICUBIC}) {
+        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS));
+        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {1, 4, 3, 2}, {2, 8, 3, 2}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS));
+        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {4, 1, 3, 2}, {1, 1, 3, 2}, mode | GGML_SCALE_FLAG_ALIGN_CORNERS));
+    }

    test_cases.emplace_back(new test_sum());
    test_cases.emplace_back(new test_sum_rows());
@@ -7472,7 +7545,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
    for (int bs : {1, 4, 8, 32, 64, 128, 256, 512}) {
        for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q4_K, GGML_TYPE_Q6_K, GGML_TYPE_IQ2_XS}) {
            for (ggml_type type_b : {GGML_TYPE_F32}) {
-                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 128, 8, false, 768, bs, 2048, 1));
+                test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 128, 8, false, 768, bs, 2048, 1));
            }
        }
    }
@@ -7480,7 +7553,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
    for (int bs : {1, 4, 8, 32, 64, 128, 256, 512}) {
        for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q4_K, GGML_TYPE_Q6_K, GGML_TYPE_IQ2_XS}) {
            for (ggml_type type_b : {GGML_TYPE_F32}) {
-                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 32, 4, false, 1792, bs, 2048, 1));
+                test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 32, 4, false, 1792, bs, 2048, 1));
            }
        }
    }
@@ -7490,7 +7563,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
    for (int bs : {1, 4, 8, 512}) {
        for (ggml_type type_a : {GGML_TYPE_MXFP4}) {
            for (ggml_type type_b : {GGML_TYPE_F32}) {
-                test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, 32, 4, false, 2880, bs, 2880, 1));
+                test_cases.emplace_back(new test_mul_mat_id_fusion(type_a, type_b, 32, 4, false, 2880, bs, 2880, 1));
            }
        }
    }
@@ -7530,6 +7603,22 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
        test_cases.emplace_back(new test_add_id(GGML_TYPE_F32, GGML_TYPE_F32, 2880, 32, 4, n_token));
    }

+    for (bool fw : {true, false}) { // fw == forward
+        for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+            for (bool ff : {false, true}) { // freq_factors
+                for (float v : { 0, 1 }) {
+                    test_cases.emplace_back(new test_rope(type, {128,  32, 512, 1}, 128, GGML_ROPE_TYPE_NORMAL, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // llama 7B
+                    test_cases.emplace_back(new test_rope(type, {128,  64, 512, 1}, 128, GGML_ROPE_TYPE_NORMAL, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // llama 65B
+                    test_cases.emplace_back(new test_rope(type, { 80,  32, 512, 1},  20, GGML_ROPE_TYPE_NEOX, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // neox (stablelm)
+                    test_cases.emplace_back(new test_rope(type, { 64,   8, 512, 1},  64, GGML_ROPE_TYPE_NEOX, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // neox (falcon 40B)
+                    test_cases.emplace_back(new test_rope(type, {128,  12, 512, 1}, 128, GGML_ROPE_TYPE_MROPE,  512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,m-rope (qwen2vl 2B)
+                    test_cases.emplace_back(new test_rope(type, {128,  12, 2, 1}, 128, GGML_ROPE_TYPE_IMROPE,  512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,imrope (qwen3vl 2B)
+                    test_cases.emplace_back(new test_rope(type, { 80,  16, 2, 1},  80, GGML_ROPE_TYPE_VISION, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,m-rope (qwen2vl ViT)
+                }
+            }
+        }
+    }
+
    std::vector<std::array<int64_t, 4>> reduce_rows_cases = {
        { 8192, 1,    1, 1 },
        { 8192, 8192, 1, 1 },
@@ -7542,6 +7631,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
        test_cases.emplace_back(new test_sum(GGML_TYPE_F32, it));
    }

+    test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {65000, 16, 1, 1}));
+
    return test_cases;
 }

@@ -138,7 +138,7 @@ int main(int /*argc*/, const char ** /*argv*/) {
    struct ggml_tensor * x;

    // rope f32
-    for (int m = 0; m < 6; ++m) {
+    for (int m = 0; m < 5; ++m) {
        const int ndims = 4;

        const int64_t n_rot = 128;
@@ -153,7 +153,7 @@ int main(int /*argc*/, const char ** /*argv*/) {
        x = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
        int mode = -1;

-        if (m < 3) {
+        if (m < 2) {
            struct ggml_tensor * p0 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
            struct ggml_tensor * p1 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
            struct ggml_tensor * p2 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
@@ -163,8 +163,8 @@ int main(int /*argc*/, const char ** /*argv*/) {
                ((int32_t *) p1->data)[i] = n_past_2 - n_past_0;
                ((int32_t *) p2->data)[i] = n_past_2 + i;
            }
-            // test mode 0, 2, 4 (standard, GPT-NeoX, GLM)
-            mode = m == 0 ? 0 : m == 1 ? 2 : 4;
+            // test mode 0, 2  (standard, GPT-NeoX)
+            mode = m == 0 ? GGML_ROPE_TYPE_NORMAL : GGML_ROPE_TYPE_NEOX;

            // 100, 101, 102, ..., 172
            r0 = ggml_rope(ctx0, x,  p0, n_rot, mode);
@@ -180,7 +180,8 @@ int main(int /*argc*/, const char ** /*argv*/) {
            struct ggml_tensor * p2 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2] * 4);

            int sections[4] = {16, 24, 24, 0};
-            mode = (m == 3) ? GGML_ROPE_TYPE_MROPE : (m == 4) ? GGML_ROPE_TYPE_VISION : GGML_ROPE_TYPE_IMROPE;
+
+            mode = (m == 2) ? GGML_ROPE_TYPE_MROPE : (m == 3) ? GGML_ROPE_TYPE_VISION : GGML_ROPE_TYPE_IMROPE;

            for (int i = 0; i < ne[2]; ++i) {
                for (int j = 0; j < 4; ++j) {
@@ -23,7 +23,8 @@ int main(int argc, char ** argv) {

    common_init();

-    int is_pp_shared = params.is_pp_shared;
+    int is_pp_shared   = params.is_pp_shared;
+    int is_tg_separate = params.is_tg_separate;

    std::vector<int> n_pp = params.n_pp;
    std::vector<int> n_tg = params.n_tg;
@@ -72,8 +73,8 @@ int main(int argc, char ** argv) {

    // decode in batches of ctx_params.n_batch tokens
    auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch, bool synchronize) {
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+        for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
+            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);

            llama_batch batch_view = {
                n_tokens,
@@ -113,7 +114,7 @@ int main(int argc, char ** argv) {

    if (!params.batched_bench_output_jsonl) {
        LOG("\n");
-        LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, int(params.flash_attn_type), params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+        LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, is_tg_separate = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, int(params.flash_attn_type), is_pp_shared, is_tg_separate, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
        LOG("\n");
        LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
        LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
@@ -172,16 +173,35 @@ int main(int argc, char ** argv) {

                const auto t_tg_start = ggml_time_us();

-                for (int i = 0; i < tg; ++i) {
-                    common_batch_clear(batch);
-
+                if (is_tg_separate) {
+                    // decode pattern:
+                    // 0 0 0 ... 1 1 1 ... 2 2 2 ... 3 3 3 ...
                    for (int j = 0; j < pl; ++j) {
-                        common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
-                    }
+                        for (int i = 0; i < tg; ++i) {
+                            common_batch_clear(batch);

-                    if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
-                        LOG_ERR("%s: llama_decode() failed\n", __func__);
-                        return 1;
+                            common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
+
+                            if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+                                LOG_ERR("%s: llama_decode() failed\n", __func__);
+                                return 1;
+                            }
+                        }
+                    }
+                } else {
+                    // decode pattern:
+                    // 0123 0123 0123 ...
+                    for (int i = 0; i < tg; ++i) {
+                        common_batch_clear(batch);
+
+                        for (int j = 0; j < pl; ++j) {
+                            common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
+                        }
+
+                        if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+                            LOG_ERR("%s: llama_decode() failed\n", __func__);
+                            return 1;
+                        }
                    }
                }

@@ -354,7 +354,11 @@ int main(int argc, char ** argv) {
        }

        // remove any "future" tokens that we might have inherited from the previous session
-        llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1);
+        if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) {
+            LOG_INF("%s: unable to resuse common prefix\n", __func__);
+            n_matching_session_tokens = 0;
+            llama_memory_seq_rm(mem, -1, -1, -1);
+        }
    }

    LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
@@ -13,6 +13,11 @@ add_library(mtmd
            mtmd-helper.h
            )

+set_target_properties(mtmd PROPERTIES
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
+)
+
 target_link_libraries     (mtmd PUBLIC ggml llama)
 target_link_libraries     (mtmd PRIVATE Threads::Threads)
 target_include_directories(mtmd PUBLIC  .)
@@ -160,13 +160,13 @@ enum patch_merge_type {
 };

 struct clip_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t n_embd;
-    int32_t n_ff;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
+    int32_t image_size = 0;
+    int32_t patch_size = 0;
+    int32_t n_embd = 0;
+    int32_t n_ff = 0;
+    int32_t projection_dim = 0;
+    int32_t n_head = 0;
+    int32_t n_layer = 0;
    // idefics3
    int32_t image_longest_edge = 0;
    int32_t image_min_pixels = -1;
@@ -2683,6 +2683,9 @@ struct clip_model_loader {
                }
            } else if (is_audio) {
                get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
+                // some hparams are unused, but still need to set to avoid issues
+                hparams.image_size = 0;
+                hparams.patch_size = 1;

            } else {
                GGML_ASSERT(false && "unknown modality");
@@ -182,7 +182,7 @@ int32_t mtmd_helper_decode_image_chunk(
    }

    const llama_model * model = llama_get_model(lctx);
-    int n_mmproj_embd = llama_model_n_embd(model);
+    int n_mmproj_embd = llama_model_n_embd_inp(model);
    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;

    int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
@@ -2,3 +2,7 @@ set(TARGET rpc-server)
 add_executable(${TARGET} rpc-server.cpp)
 target_link_libraries(${TARGET} PRIVATE ggml)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()
@@ -7,6 +7,10 @@ if (MINGW)
    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()

+if (NOT LLAMA_HTTPLIB)
+    message(FATAL_ERROR "LLAMA_HTTPLIB is OFF, cannot build llama-server. Hint: to skip building server, set -DLLAMA_BUILD_SERVER=OFF")
+endif()
+
 set(TARGET_SRCS
    server.cpp
    utils.hpp
@@ -33,7 +37,7 @@ install(TARGETS ${TARGET} RUNTIME)

 target_include_directories(${TARGET} PRIVATE ../mtmd)
 target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
-target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common mtmd cpp-httplib ${CMAKE_THREAD_LIBS_INIT})

 if (WIN32)
    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
@@ -684,7 +684,7 @@ struct server_task_result {
    }
    virtual bool is_stop() {
        // only used by server_task_result_cmpl_*
-        return false;
+        return true;
    }
    virtual int get_index() {
        return -1;
@@ -1690,6 +1690,9 @@ struct server_slot {
        bool res = prompt_cache.load(prompt, tokens, ctx, id);
        if (!res) {
            SLT_WRN(*this, "%s", "failed to load prompt from cache\n");
+
+            llama_memory_seq_rm(llama_get_memory(ctx), id, -1, -1);
+            prompt.tokens.clear();
        }
    }

@@ -3235,105 +3238,6 @@ struct server_context {
        queue_results.send(std::move(res));
    }

-    //
-    // Functions to create new task(s) and receive result(s)
-    //
-
-    void cancel_tasks(const std::unordered_set<int> & id_tasks) {
-        std::vector<server_task> cancel_tasks;
-        cancel_tasks.reserve(id_tasks.size());
-        for (const auto & id_task : id_tasks) {
-            SRV_WRN("cancel task, id_task = %d\n", id_task);
-
-            server_task task(SERVER_TASK_TYPE_CANCEL);
-            task.id_target = id_task;
-            queue_results.remove_waiting_task_id(id_task);
-            cancel_tasks.push_back(std::move(task));
-        }
-        // push to beginning of the queue, so it has highest priority
-        queue_tasks.post(std::move(cancel_tasks), true);
-    }
-
-    // receive the results from task(s)
-    void receive_multi_results(
-            const std::unordered_set<int> & id_tasks,
-            const std::function<void(std::vector<server_task_result_ptr>&)> & result_handler,
-            const std::function<void(json)> & error_handler,
-            const std::function<bool()> & is_connection_closed) {
-        std::vector<server_task_result_ptr> results(id_tasks.size());
-        for (int i = 0; i < (int)id_tasks.size(); i++) {
-            server_task_result_ptr result = queue_results.recv_with_timeout(id_tasks, HTTP_POLLING_SECONDS);
-
-            if (is_connection_closed()) {
-                cancel_tasks(id_tasks);
-                return;
-            }
-
-            if (result == nullptr) {
-                i--; // retry
-                continue;
-            }
-
-            if (result->is_error()) {
-                error_handler(result->to_json());
-                cancel_tasks(id_tasks);
-                return;
-            }
-
-            GGML_ASSERT(
-                dynamic_cast<server_task_result_cmpl_final*>(result.get()) != nullptr
-                || dynamic_cast<server_task_result_embd*>(result.get()) != nullptr
-                || dynamic_cast<server_task_result_rerank*>(result.get()) != nullptr
-            );
-            const size_t idx = result->get_index();
-            GGML_ASSERT(idx < results.size() && "index out of range");
-            results[idx] = std::move(result);
-        }
-        result_handler(results);
-    }
-
-    // receive the results from task(s), in stream mode
-    void receive_cmpl_results_stream(
-            const std::unordered_set<int> & id_tasks,
-            const std::function<bool(server_task_result_ptr&)> & result_handler,
-            const std::function<void(json)> & error_handler,
-            const std::function<bool()> & is_connection_closed) {
-        size_t n_finished = 0;
-        while (true) {
-            server_task_result_ptr result = queue_results.recv_with_timeout(id_tasks, HTTP_POLLING_SECONDS);
-
-            if (is_connection_closed()) {
-                cancel_tasks(id_tasks);
-                return;
-            }
-
-            if (result == nullptr) {
-                continue; // retry
-            }
-
-            if (result->is_error()) {
-                error_handler(result->to_json());
-                cancel_tasks(id_tasks);
-                return;
-            }
-
-            GGML_ASSERT(
-                dynamic_cast<server_task_result_cmpl_partial*>(result.get()) != nullptr
-                || dynamic_cast<server_task_result_cmpl_final*>(result.get()) != nullptr
-            );
-            if (!result_handler(result)) {
-                cancel_tasks(id_tasks);
-                break;
-            }
-
-            if (result->is_stop()) {
-                if (++n_finished == id_tasks.size()) {
-                    break;
-                }
-            }
-        }
-    }
-
    //
    // Functions to process the task
    //
@@ -4415,6 +4319,104 @@ struct server_context {
    }
 };

+// generator-like API for server responses, support pooling connection state and aggregating results
+struct server_response_reader {
+    std::unordered_set<int> id_tasks;
+    server_context & ctx_server;
+    size_t received_count = 0;
+    bool cancelled = false;
+
+    server_response_reader(server_context & ctx_server) : ctx_server(ctx_server) {}
+    ~server_response_reader() {
+        stop();
+    }
+
+    void post_tasks(std::vector<server_task> && tasks) {
+        id_tasks = server_task::get_list_id(tasks);
+        ctx_server.queue_results.add_waiting_tasks(tasks);
+        ctx_server.queue_tasks.post(std::move(tasks));
+    }
+
+    bool has_next() {
+        return !cancelled && received_count < id_tasks.size();
+    }
+
+    // return nullptr if should_stop() is true before receiving a result
+    // note: if one error is received, it will stop further processing and return error result
+    server_task_result_ptr next(const std::function<bool()> & should_stop) {
+        while (true) {
+            server_task_result_ptr result = ctx_server.queue_results.recv_with_timeout(id_tasks, HTTP_POLLING_SECONDS);
+            if (result == nullptr) {
+                // timeout, check stop condition
+                if (should_stop()) {
+                    SRV_DBG("%s", "stopping wait for next result due to should_stop condition\n");
+                    return nullptr;
+                }
+            } else {
+                if (result->is_error()) {
+                    stop(); // cancel remaining tasks
+                    SRV_DBG("%s", "received error result, stopping further processing\n");
+                    return result;
+                }
+                if (result->is_stop()) {
+                    received_count++;
+                }
+                return result;
+            }
+        }
+
+        // should not reach here
+    }
+
+    struct batch_response {
+        bool is_terminated = false; // if true, indicates that processing was stopped before all results were received
+        std::vector<server_task_result_ptr> results;
+        server_task_result_ptr error; // nullptr if no error
+    };
+
+    batch_response wait_for_all(const std::function<bool()> & should_stop) {
+        batch_response batch_res;
+        batch_res.results.resize(id_tasks.size());
+        while (has_next()) {
+            auto res = next(should_stop);
+            if (res == nullptr) {
+                batch_res.is_terminated = true;
+                return batch_res;
+            }
+            if (res->is_error()) {
+                batch_res.error = std::move(res);
+                return batch_res;
+            }
+            const size_t idx = res->get_index();
+            GGML_ASSERT(idx < batch_res.results.size() && "index out of range");
+            GGML_ASSERT(batch_res.results[idx] == nullptr && "duplicate result received");
+            batch_res.results[idx] = std::move(res);
+        }
+        return batch_res;
+    }
+
+    void stop() {
+        ctx_server.queue_results.remove_waiting_task_ids(id_tasks);
+        if (has_next() && !cancelled) {
+            // if tasks is not finished yet, cancel them
+            cancelled = true;
+            std::vector<server_task> cancel_tasks;
+            cancel_tasks.reserve(id_tasks.size());
+            for (const auto & id_task : id_tasks) {
+                SRV_WRN("cancel task, id_task = %d\n", id_task);
+                server_task task(SERVER_TASK_TYPE_CANCEL);
+                task.id_target = id_task;
+                ctx_server.queue_results.remove_waiting_task_id(id_task);
+                cancel_tasks.push_back(std::move(task));
+            }
+            // push to beginning of the queue, so it has highest priority
+            ctx_server.queue_tasks.post(std::move(cancel_tasks), true);
+        } else {
+            SRV_DBG("%s", "all tasks already finished, no need to cancel\n");
+        }
+    }
+};
+
 static void log_server_request(const httplib::Request & req, const httplib::Response & res) {
    // skip GH copilot requests when using default port
    if (req.path == "/v1/health") {
@@ -4429,6 +4431,17 @@ static void log_server_request(const httplib::Request & req, const httplib::Resp
    SRV_DBG("response: %s\n", res.body.c_str());
 }

+static void res_error(httplib::Response & res, const json & error_data) {
+    json final_response {{"error", error_data}};
+    res.set_content(safe_json_to_str(final_response), MIMETYPE_JSON);
+    res.status = json_value(error_data, "code", 500);
+}
+
+static void res_ok(httplib::Response & res, const json & data) {
+    res.set_content(safe_json_to_str(data), MIMETYPE_JSON);
+    res.status = 200;
+}
+
 std::function<void(int)> shutdown_handler;
 std::atomic_flag is_terminating = ATOMIC_FLAG_INIT;

@@ -4498,19 +4511,7 @@ int main(int argc, char ** argv) {

    svr->set_default_headers({{"Server", "llama.cpp"}});
    svr->set_logger(log_server_request);
-
-    auto res_error = [](httplib::Response & res, const json & error_data) {
-        json final_response {{"error", error_data}};
-        res.set_content(safe_json_to_str(final_response), MIMETYPE_JSON);
-        res.status = json_value(error_data, "code", 500);
-    };
-
-    auto res_ok = [](httplib::Response & res, const json & data) {
-        res.set_content(safe_json_to_str(data), MIMETYPE_JSON);
-        res.status = 200;
-    };
-
-    svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) {
+    svr->set_exception_handler([](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) {
        std::string message;
        try {
            std::rethrow_exception(ep);
@@ -4529,7 +4530,7 @@ int main(int argc, char ** argv) {
        }
    });

-    svr->set_error_handler([&res_error](const httplib::Request &, httplib::Response & res) {
+    svr->set_error_handler([](const httplib::Request &, httplib::Response & res) {
        if (res.status == 404) {
            res_error(res, format_error_response("File Not Found", ERROR_TYPE_NOT_FOUND));
        }
@@ -4559,7 +4560,7 @@ int main(int argc, char ** argv) {
    // Middlewares
    //

-    auto middleware_validate_api_key = [&params, &res_error](const httplib::Request & req, httplib::Response & res) {
+    auto middleware_validate_api_key = [&params](const httplib::Request & req, httplib::Response & res) {
        static const std::unordered_set<std::string> public_endpoints = {
            "/health",
            "/v1/health",
@@ -4597,7 +4598,7 @@ int main(int argc, char ** argv) {
        return false;
    };

-    auto middleware_server_state = [&res_error, &state](const httplib::Request & req, httplib::Response & res) {
+    auto middleware_server_state = [&state](const httplib::Request & req, httplib::Response & res) {
        server_state current_state = state.load();
        if (current_state == SERVER_STATE_LOADING_MODEL) {
            auto tmp = string_split<std::string>(req.path, '.');
@@ -4785,7 +4786,7 @@ int main(int argc, char ** argv) {
        res.status = 200; // HTTP OK
    };

-    const auto handle_slots_save = [&ctx_server, &res_error, &res_ok, &params](const httplib::Request & req, httplib::Response & res, int id_slot) {
+    const auto handle_slots_save = [&ctx_server, &params](const httplib::Request & req, httplib::Response & res, int id_slot) {
        json request_data = json::parse(req.body);
        std::string filename = request_data.at("filename");
        if (!fs_validate_filename(filename)) {
@@ -4817,7 +4818,7 @@ int main(int argc, char ** argv) {
        res_ok(res, result->to_json());
    };

-    const auto handle_slots_restore = [&ctx_server, &res_error, &res_ok, &params](const httplib::Request & req, httplib::Response & res, int id_slot) {
+    const auto handle_slots_restore = [&ctx_server, &params](const httplib::Request & req, httplib::Response & res, int id_slot) {
        json request_data = json::parse(req.body);
        std::string filename = request_data.at("filename");
        if (!fs_validate_filename(filename)) {
@@ -4850,7 +4851,7 @@ int main(int argc, char ** argv) {
        res_ok(res, result->to_json());
    };

-    const auto handle_slots_erase = [&ctx_server, &res_error, &res_ok](const httplib::Request & /* req */, httplib::Response & res, int id_slot) {
+    const auto handle_slots_erase = [&ctx_server](const httplib::Request & /* req */, httplib::Response & res, int id_slot) {
        int task_id = ctx_server.queue_tasks.get_new_id();
        {
            server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
@@ -4873,7 +4874,7 @@ int main(int argc, char ** argv) {
        res_ok(res, result->to_json());
    };

-    const auto handle_slots_action = [&params, &res_error, &handle_slots_save, &handle_slots_restore, &handle_slots_erase](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_slots_action = [&params, &handle_slots_save, &handle_slots_restore, &handle_slots_erase](const httplib::Request & req, httplib::Response & res) {
        if (params.slot_save_path.empty()) {
            res_error(res, format_error_response("This server does not support slots action. Start it with `--slot-save-path`", ERROR_TYPE_NOT_SUPPORTED));
            return;
@@ -4902,7 +4903,7 @@ int main(int argc, char ** argv) {
        }
    };

-    const auto handle_props = [&params, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+    const auto handle_props = [&params, &ctx_server](const httplib::Request &, httplib::Response & res) {
        json default_generation_settings_for_props;

        {
@@ -4944,7 +4945,7 @@ int main(int argc, char ** argv) {
        res_ok(res, data);
    };

-    const auto handle_props_change = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_props_change = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        if (!ctx_server.params_base.endpoint_props) {
            res_error(res, format_error_response("This server does not support changing global properties. Start it with `--props`", ERROR_TYPE_NOT_SUPPORTED));
            return;
@@ -4957,7 +4958,7 @@ int main(int argc, char ** argv) {
        res_ok(res, {{ "success", true }});
    };

-    const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+    const auto handle_api_show = [&ctx_server](const httplib::Request &, httplib::Response & res) {
        bool has_mtmd = ctx_server.mctx != nullptr;
        json data = {
            {
@@ -4988,7 +4989,7 @@ int main(int argc, char ** argv) {

    // handle completion-like requests (completion, chat, infill)
    // we can optionally provide a custom format for partial results and final results
-    const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
+    const auto handle_completions_impl = [&ctx_server](
            server_task_type type,
            json & data,
            const std::vector<raw_buffer> & files,
@@ -4998,7 +4999,10 @@ int main(int argc, char ** argv) {
        GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);

        auto completion_id = gen_chatcmplid();
-        std::unordered_set<int> task_ids;
+        // need to store the reader as a pointer, so that it won't be destroyed when the handle returns
+        // use shared_ptr as it's shared between the chunked_content_provider() and on_complete()
+        const auto rd = std::make_shared<server_response_reader>(ctx_server);
+
        try {
            std::vector<server_task> tasks;

@@ -5016,17 +5020,8 @@ int main(int argc, char ** argv) {
                // Everything else, including multimodal completions.
                inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
            }
-            const size_t n_ctx_slot = ctx_server.slots.front().n_ctx;
            tasks.reserve(inputs.size());
            for (size_t i = 0; i < inputs.size(); i++) {
-                auto n_prompt_tokens = inputs[i].size();
-                if (n_prompt_tokens >= n_ctx_slot) {
-                    json error_data = format_error_response("the request exceeds the available context size, try increasing it", ERROR_TYPE_EXCEED_CONTEXT_SIZE);
-                    error_data["n_prompt_tokens"] = n_prompt_tokens;
-                    error_data["n_ctx"] = n_ctx_slot;
-                    res_error(res, error_data);
-                    return;
-                }
                server_task task = server_task(type);

                task.id    = ctx_server.queue_tasks.get_new_id();
@@ -5047,9 +5042,7 @@ int main(int argc, char ** argv) {
                tasks.push_back(std::move(task));
            }

-            task_ids = server_task::get_list_id(tasks);
-            ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(std::move(tasks));
+            rd->post_tasks(std::move(tasks));
        } catch (const std::exception & e) {
            res_error(res, format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
            return;
@@ -5058,54 +5051,95 @@ int main(int argc, char ** argv) {
        bool stream = json_value(data, "stream", false);

        if (!stream) {
-            ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-                if (results.size() == 1) {
-                    // single result
-                    res_ok(res, results[0]->to_json());
-                } else {
-                    // multiple results (multitask)
-                    json arr = json::array();
-                    for (auto & res : results) {
-                        arr.push_back(res->to_json());
-                    }
-                    res_ok(res, arr);
+            // non-stream, wait for the results
+            auto all_results = rd->wait_for_all(is_connection_closed);
+            if (all_results.is_terminated) {
+                return; // connection is closed
+            } else if (all_results.error) {
+                res_error(res, all_results.error->to_json());
+                return;
+            } else {
+                json arr = json::array();
+                for (auto & res : all_results.results) {
+                    GGML_ASSERT(dynamic_cast<server_task_result_cmpl_final*>(res.get()) != nullptr);
+                    arr.push_back(res->to_json());
                }
-            }, [&](const json & error_data) {
-                res_error(res, error_data);
-            }, is_connection_closed);
+                // if single request, return single object instead of array
+                res_ok(res, arr.size() == 1 ? arr[0] : arr);
+            }

-            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
        } else {
-            const auto chunked_content_provider = [task_ids, &ctx_server, oaicompat](size_t, httplib::DataSink & sink) {
-                ctx_server.receive_cmpl_results_stream(task_ids, [&](server_task_result_ptr & result) -> bool {
-                    json res_json = result->to_json();
-                    if (res_json.is_array()) {
-                        for (const auto & res : res_json) {
-                            if (!server_sent_event(sink, res)) {
-                                // sending failed (HTTP connection closed), cancel the generation
-                                return false;
-                            }
-                        }
-                        return true;
-                    } else {
-                        return server_sent_event(sink, res_json);
+            // in streaming mode, the first error must be treated as non-stream response
+            // this is to match the OAI API behavior
+            // ref: https://github.com/ggml-org/llama.cpp/pull/16486#discussion_r2419657309
+            server_task_result_ptr first_result = rd->next(is_connection_closed);
+            if (first_result == nullptr) {
+                return; // connection is closed
+            } else if (first_result->is_error()) {
+                res_error(res, first_result->to_json());
+                return;
+            } else {
+                GGML_ASSERT(
+                    dynamic_cast<server_task_result_cmpl_partial*>(first_result.get()) != nullptr
+                    || dynamic_cast<server_task_result_cmpl_final*>(first_result.get()) != nullptr
+                );
+            }
+
+            // next responses are streamed
+            json first_result_json = first_result->to_json();
+            const auto chunked_content_provider = [first_result_json, rd, oaicompat](size_t, httplib::DataSink & sink) mutable -> bool {
+                // flush the first result as it's not an error
+                if (!first_result_json.empty()) {
+                    if (!server_sent_event(sink, first_result_json)) {
+                        sink.done();
+                        return false; // sending failed, go to on_complete()
                    }
-                }, [&](const json & error_data) {
-                    server_sent_event(sink, json{{"error", error_data}});
-                }, [&sink]() {
-                    // note: do not use req.is_connection_closed here because req is already destroyed
-                    return !sink.is_writable();
-                });
-                if (oaicompat != OAICOMPAT_TYPE_NONE) {
-                    static const std::string ev_done = "data: [DONE]\n\n";
-                    sink.write(ev_done.data(), ev_done.size());
+                    first_result_json.clear(); // mark as sent
                }
-                sink.done();
-                return false;
+
+                // receive subsequent results
+                auto result = rd->next([&sink]{ return !sink.is_writable(); });
+                if (result == nullptr) {
+                    sink.done();
+                    return false; // connection is closed, go to on_complete()
+                }
+
+                // send the results
+                json res_json = result->to_json();
+                bool ok = false;
+                if (result->is_error()) {
+                    ok = server_sent_event(sink, json {{ "error", result->to_json() }});
+                    sink.done();
+                    return false; // go to on_complete()
+                } else {
+                    GGML_ASSERT(
+                        dynamic_cast<server_task_result_cmpl_partial*>(result.get()) != nullptr
+                        || dynamic_cast<server_task_result_cmpl_final*>(result.get()) != nullptr
+                    );
+                    ok = server_sent_event(sink, res_json);
+                }
+
+                if (!ok) {
+                    sink.done();
+                    return false; // sending failed, go to on_complete()
+                }
+
+                // check if there is more data
+                if (!rd->has_next()) {
+                    if (oaicompat != OAICOMPAT_TYPE_NONE) {
+                        static const std::string ev_done = "data: [DONE]\n\n";
+                        sink.write(ev_done.data(), ev_done.size());
+                    }
+                    sink.done();
+                    return false; // no more data, go to on_complete()
+                }
+
+                // has next data, continue
+                return true;
            };

-            auto on_complete = [task_ids, &ctx_server] (bool) {
-                ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+            auto on_complete = [rd](bool) {
+                rd->stop();
            };

            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
@@ -5136,7 +5170,7 @@ int main(int argc, char ** argv) {
            OAICOMPAT_TYPE_COMPLETION);
    };

-    const auto handle_infill = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_infill = [&ctx_server, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
        // check model compatibility
        std::string err;
        if (llama_vocab_fim_pre(ctx_server.vocab) == LLAMA_TOKEN_NULL) {
@@ -5235,7 +5269,7 @@ int main(int argc, char ** argv) {
    };

    // same with handle_chat_completions, but without inference part
-    const auto handle_apply_template = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_apply_template = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        auto body = json::parse(req.body);
        std::vector<raw_buffer> files; // dummy, unused
        json data = oaicompat_chat_params_parse(
@@ -5245,7 +5279,7 @@ int main(int argc, char ** argv) {
        res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
    };

-    const auto handle_models = [&params, &ctx_server, &state, &res_ok](const httplib::Request &, httplib::Response & res) {
+    const auto handle_models = [&params, &ctx_server, &state](const httplib::Request &, httplib::Response & res) {
        server_state current_state = state.load();
        json model_meta = nullptr;
        if (current_state == SERVER_STATE_READY) {
@@ -5290,7 +5324,7 @@ int main(int argc, char ** argv) {
        res_ok(res, models);
    };

-    const auto handle_tokenize = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_tokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        const json body = json::parse(req.body);

        json tokens_response = json::array();
@@ -5331,7 +5365,7 @@ int main(int argc, char ** argv) {
        res_ok(res, data);
    };

-    const auto handle_detokenize = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_detokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        const json body = json::parse(req.body);

        std::string content;
@@ -5344,7 +5378,7 @@ int main(int argc, char ** argv) {
        res_ok(res, data);
    };

-    const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
+    const auto handle_embeddings_impl = [&ctx_server](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
        if (!ctx_server.params_base.embedding) {
            res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
            return;
@@ -5399,8 +5433,7 @@ int main(int argc, char ** argv) {

        // create and queue the task
        json responses = json::array();
-        bool error = false;
-        std::unordered_set<int> task_ids;
+        server_response_reader rd(ctx_server);
        {
            std::vector<server_task> tasks;
            for (size_t i = 0; i < tokenized_prompts.size(); i++) {
@@ -5416,27 +5449,23 @@ int main(int argc, char ** argv) {

                tasks.push_back(std::move(task));
            }
-
-            task_ids = server_task::get_list_id(tasks);
-            ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(std::move(tasks));
+            rd.post_tasks(std::move(tasks));
        }

-        // get the result
-        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-            for (auto & res : results) {
+        // wait for the results
+        auto all_results = rd.wait_for_all(req.is_connection_closed);
+
+        // collect results
+        if (all_results.is_terminated) {
+            return; // connection is closed
+        } else if (all_results.error) {
+            res_error(res, all_results.error->to_json());
+            return;
+        } else {
+            for (auto & res : all_results.results) {
                GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
                responses.push_back(res->to_json());
            }
-        }, [&](const json & error_data) {
-            res_error(res, error_data);
-            error = true;
-        }, req.is_connection_closed);
-
-        ctx_server.queue_results.remove_waiting_task_ids(task_ids);
-
-        if (error) {
-            return;
        }

        // write JSON response
@@ -5454,7 +5483,7 @@ int main(int argc, char ** argv) {
        handle_embeddings_impl(req, res, OAICOMPAT_TYPE_EMBEDDING);
    };

-    const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_rerank = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        if (!ctx_server.params_base.embedding || ctx_server.params_base.pooling_type != LLAMA_POOLING_TYPE_RANK) {
            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
            return;
@@ -5490,8 +5519,7 @@ int main(int argc, char ** argv) {

        // create and queue the task
        json responses = json::array();
-        bool error = false;
-        std::unordered_set<int> task_ids;
+        server_response_reader rd(ctx_server);
        {
            std::vector<server_task> tasks;
            tasks.reserve(documents.size());
@@ -5503,24 +5531,23 @@ int main(int argc, char ** argv) {
                task.tokens = std::move(tmp);
                tasks.push_back(std::move(task));
            }
-
-            task_ids = server_task::get_list_id(tasks);
-            ctx_server.queue_results.add_waiting_tasks(tasks);
-            ctx_server.queue_tasks.post(std::move(tasks));
+            rd.post_tasks(std::move(tasks));
        }

-        ctx_server.receive_multi_results(task_ids, [&](std::vector<server_task_result_ptr> & results) {
-            for (auto & res : results) {
+        // wait for the results
+        auto all_results = rd.wait_for_all(req.is_connection_closed);
+
+        // collect results
+        if (all_results.is_terminated) {
+            return; // connection is closed
+        } else if (all_results.error) {
+            res_error(res, all_results.error->to_json());
+            return;
+        } else {
+            for (auto & res : all_results.results) {
                GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
                responses.push_back(res->to_json());
            }
-        }, [&](const json & error_data) {
-            res_error(res, error_data);
-            error = true;
-        }, req.is_connection_closed);
-
-        if (error) {
-            return;
        }

        // write JSON response
@@ -1,6 +1,8 @@
 import pytest
 import requests
 import time
+import random
+
 from openai import OpenAI
 from utils import *

@@ -564,3 +566,43 @@ def test_cancel_request():
    time.sleep(1) # wait for HTTP_POLLING_SECONDS
    res = server.make_request("GET", "/slots")
    assert res.body[0]["is_processing"] == False
+
+
+# this test exercises the host-memory prompt cache
+# ref: https://github.com/ggml-org/llama.cpp/pull/16391
+# ref: https://github.com/ggml-org/llama.cpp/pull/17078
+def test_completion_prompt_cache():
+    global server
+    server.n_slots = 2
+    server.kv_unified = True
+    server.start()
+
+    for _ in range(16):
+        # generate alternating random prompts with variable lengths in order to get them in and out of the cache
+        r = random.randint(0, 4)
+        prompt = (" Hello " +  str(r)) * (40 + r)
+        n_prompt = (40 + r)*5 + 2
+        n_predict = random.randint(1, 8)
+
+        res = server.make_request(
+            "POST",
+            "/completion",
+            data={
+                "prompt": prompt,
+                "n_predict": n_predict,
+            },
+        )
+
+        assert res.status_code == 200
+        assert "content" in res.body
+        content = res.body["content"]
+        assert isinstance(content, str)
+        assert len(content) > 0
+
+        assert type(res.body["has_new_line"]) == bool
+        assert "timings" in res.body
+        timings = res.body["timings"]
+
+        assert "prompt_n" in timings and timings["prompt_n"] + timings["cache_n"] == n_prompt
+        assert "predicted_n" in timings and timings["predicted_n"] == n_predict
+        assert "tokens" in res.body and isinstance(res.body["tokens"], list)
@@ -9,14 +9,6 @@
 #include "mtmd-helper.h"
 #include "chat.h"

-// increase max payload length to allow use of larger context size
-#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
-// increase backlog size to avoid connection resets for >> 1 slots
-#define CPPHTTPLIB_LISTEN_BACKLOG 512
-// increase max URI length to handle longer prompts in query string
-#define CPPHTTPLIB_REQUEST_URI_MAX_LENGTH 32768
-// disable Nagle's algorithm
-#define CPPHTTPLIB_TCP_NODELAY true
 #include <cpp-httplib/httplib.h>

 #define JSON_ASSERT GGML_ASSERT
@@ -461,15 +453,29 @@ static std::string tokens_to_output_formatted_string(const llama_context * ctx,
    return out;
 }

+// note: if data is a json array, it will be sent as multiple events, one per item
 static bool server_sent_event(httplib::DataSink & sink, const json & data) {
-    const std::string str =
-        "data: " +
-        data.dump(-1, ' ', false, json::error_handler_t::replace) +
-        "\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row).
+    static auto send_single = [](httplib::DataSink & sink, const json & data) -> bool {
+        const std::string str =
+            "data: " +
+            data.dump(-1, ' ', false, json::error_handler_t::replace) +
+            "\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row).

-    LOG_DBG("data stream, to_send: %s", str.c_str());
+        LOG_DBG("data stream, to_send: %s", str.c_str());
+        return sink.write(str.c_str(), str.size());
+    };

-    return sink.write(str.c_str(), str.size());
+    if (data.is_array()) {
+        for (const auto & item : data) {
+            if (!send_single(sink, item)) {
+                return false;
+            }
+        }
+    } else {
+        return send_single(sink, data);
+    }
+
+    return true;
 }

 //
@@ -11,8 +11,16 @@ const preview: Preview = {
 				date: /Date$/i
 			}
 		},
+
 		backgrounds: {
 			disable: true
+		},
+
+		a11y: {
+			// 'todo' - show a11y violations in the test UI only
+			// 'error' - fail CI on a11y violations
+			// 'off' - skip a11y checks entirely
+			test: 'todo'
 		}
 	},
 	decorators: [
@@ -1,8 +1,9 @@
+import * as a11yAddonAnnotations from '@storybook/addon-a11y/preview';
 import { setProjectAnnotations } from '@storybook/sveltekit';
 import * as previewAnnotations from './preview';
 import { beforeAll } from 'vitest';

-const project = setProjectAnnotations([previewAnnotations]);
+const project = setProjectAnnotations([a11yAddonAnnotations, previewAnnotations]);

 beforeAll(async () => {
 	if (project.beforeAll) {
@@ -22,20 +22,20 @@
 				"unist-util-visit": "^5.0.0"
 			},
 			"devDependencies": {
-				"@chromatic-com/storybook": "^4.0.1",
+				"@chromatic-com/storybook": "^4.1.2",
 				"@eslint/compat": "^1.2.5",
 				"@eslint/js": "^9.18.0",
 				"@internationalized/date": "^3.8.2",
 				"@lucide/svelte": "^0.515.0",
 				"@playwright/test": "^1.49.1",
-				"@storybook/addon-a11y": "^9.0.17",
-				"@storybook/addon-docs": "^9.0.17",
-				"@storybook/addon-svelte-csf": "^5.0.7",
-				"@storybook/addon-vitest": "^9.0.17",
-				"@storybook/sveltekit": "^9.0.17",
-				"@sveltejs/adapter-static": "^3.0.8",
-				"@sveltejs/kit": "^2.22.0",
-				"@sveltejs/vite-plugin-svelte": "^6.0.0",
+				"@storybook/addon-a11y": "^10.0.7",
+				"@storybook/addon-docs": "^10.0.7",
+				"@storybook/addon-svelte-csf": "^5.0.10",
+				"@storybook/addon-vitest": "^10.0.7",
+				"@storybook/sveltekit": "^10.0.7",
+				"@sveltejs/adapter-static": "^3.0.10",
+				"@sveltejs/kit": "^2.48.4",
+				"@sveltejs/vite-plugin-svelte": "^6.2.1",
 				"@tailwindcss/forms": "^0.5.9",
 				"@tailwindcss/typography": "^0.5.15",
 				"@tailwindcss/vite": "^4.0.0",
@@ -46,21 +46,21 @@
 				"dexie": "^4.0.11",
 				"eslint": "^9.18.0",
 				"eslint-config-prettier": "^10.0.1",
-				"eslint-plugin-storybook": "^9.0.17",
+				"eslint-plugin-storybook": "^10.0.7",
 				"eslint-plugin-svelte": "^3.0.0",
 				"fflate": "^0.8.2",
 				"globals": "^16.0.0",
 				"http-server": "^14.1.1",
 				"mdast": "^3.0.0",
 				"mdsvex": "^0.12.3",
-				"playwright": "^1.53.0",
+				"playwright": "^1.56.1",
 				"prettier": "^3.4.2",
 				"prettier-plugin-svelte": "^3.3.3",
 				"prettier-plugin-tailwindcss": "^0.6.11",
 				"rehype-katex": "^7.0.1",
 				"remark-math": "^6.0.0",
 				"sass": "^1.93.3",
-				"storybook": "^9.0.17",
+				"storybook": "^10.0.7",
 				"svelte": "^5.0.0",
 				"svelte-check": "^4.0.0",
 				"tailwind-merge": "^3.3.1",
@@ -71,7 +71,7 @@
 				"typescript-eslint": "^8.20.0",
 				"unified": "^11.0.5",
 				"uuid": "^13.0.0",
-				"vite": "^7.0.4",
+				"vite": "^7.2.2",
 				"vite-plugin-devtools-json": "^0.2.0",
 				"vitest": "^3.2.3",
 				"vitest-browser-svelte": "^0.1.0"
@@ -133,9 +133,9 @@
 			}
 		},
 		"node_modules/@chromatic-com/storybook": {
-			"version": "4.0.1",
-			"resolved": "https://registry.npmjs.org/@chromatic-com/storybook/-/storybook-4.0.1.tgz",
-			"integrity": "sha512-GQXe5lyZl3yLewLJQyFXEpOp2h+mfN2bPrzYaOFNCJjO4Js9deKbRHTOSaiP2FRwZqDLdQwy2+SEGeXPZ94yYw==",
+			"version": "4.1.2",
+			"resolved": "https://registry.npmjs.org/@chromatic-com/storybook/-/storybook-4.1.2.tgz",
+			"integrity": "sha512-QAWGtHwib0qsP5CcO64aJCF75zpFgpKK3jNpxILzQiPK3sVo4EmnVGJVdwcZWpWrGdH8E4YkncGoitw4EXzKMg==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
@@ -150,7 +150,7 @@
 				"yarn": ">=1.22.18"
 			},
 			"peerDependencies": {
-				"storybook": "^0.0.0-0 || ^9.0.0 || ^9.1.0-0"
+				"storybook": "^0.0.0-0 || ^9.0.0 || ^9.1.0-0 || ^9.2.0-0 || ^10.0.0-0 || ^10.1.0-0 || ^10.2.0-0 || ^10.3.0-0"
 			}
 		},
 		"node_modules/@esbuild/aix-ppc64": {
@@ -894,6 +894,17 @@
 				"@jridgewell/trace-mapping": "^0.3.24"
 			}
 		},
+		"node_modules/@jridgewell/remapping": {
+			"version": "2.3.5",
+			"resolved": "https://registry.npmjs.org/@jridgewell/remapping/-/remapping-2.3.5.tgz",
+			"integrity": "sha512-LI9u/+laYG4Ds1TDKSJW2YPrIlcVYOwi2fUC6xB43lueCjgxV4lffOCZCtYFiH6TNOX+tQKXx97T4IKHbhyHEQ==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"@jridgewell/gen-mapping": "^0.3.5",
+				"@jridgewell/trace-mapping": "^0.3.24"
+			}
+		},
 		"node_modules/@jridgewell/resolve-uri": {
 			"version": "3.1.2",
 			"resolved": "https://registry.npmjs.org/@jridgewell/resolve-uri/-/resolve-uri-3.1.2.tgz",
@@ -1502,13 +1513,13 @@
 			}
 		},
 		"node_modules/@playwright/test": {
-			"version": "1.54.1",
-			"resolved": "https://registry.npmjs.org/@playwright/test/-/test-1.54.1.tgz",
-			"integrity": "sha512-FS8hQ12acieG2dYSksmLOF7BNxnVf2afRJdCuM1eMSxj6QTSE6G4InGF7oApGgDb65MX7AwMVlIkpru0yZA4Xw==",
+			"version": "1.56.1",
+			"resolved": "https://registry.npmjs.org/@playwright/test/-/test-1.56.1.tgz",
+			"integrity": "sha512-vSMYtL/zOcFpvJCW71Q/OEGQb7KYBPAdKh35WNSkaZA75JlAO8ED8UN6GUNTm3drWomcbcqRPFqQbLae8yBTdg==",
 			"dev": true,
 			"license": "Apache-2.0",
 			"dependencies": {
-				"playwright": "1.54.1"
+				"playwright": "1.56.1"
 			},
 			"bin": {
 				"playwright": "cli.js"
@@ -1812,9 +1823,9 @@
 			"license": "MIT"
 		},
 		"node_modules/@storybook/addon-a11y": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/addon-a11y/-/addon-a11y-9.0.17.tgz",
-			"integrity": "sha512-9cXNK3q/atx3hwJAt9HkJbd9vUxCXfKKiNNuSACbf8h9/j6u3jktulKOf6Xjc3B8lwn6ZpdK/x1HHZN2kTqsvg==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/addon-a11y/-/addon-a11y-10.0.7.tgz",
+			"integrity": "sha512-JsYPpZ/n67/2bI1XJeyrAWHHQkHemPkPHjCA0tAUnMz1Shlo/LV2q1Ahgpxoihx4strbHwZz71bcS4MqkHBduA==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
@@ -1826,20 +1837,20 @@
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"storybook": "^9.0.17"
+				"storybook": "^10.0.7"
 			}
 		},
 		"node_modules/@storybook/addon-docs": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/addon-docs/-/addon-docs-9.0.17.tgz",
-			"integrity": "sha512-LOX/kKgQGnyulrqZHsvf77+ZoH/nSUaplGr5hvZglW/U6ak6fO9seJyXAzVKEnC6p+F8n02kFBZbi3s+znQhSg==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/addon-docs/-/addon-docs-10.0.7.tgz",
+			"integrity": "sha512-qQQMoeYZC4W+/8ubfOZiTrE8nYC/f4wWP1uq4peRyDy1N2nIN9SwhyxwMn0m3VpeGmRBga5dLvJY9ko6SnJekg==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"@mdx-js/react": "^3.0.0",
-				"@storybook/csf-plugin": "9.0.17",
-				"@storybook/icons": "^1.2.12",
-				"@storybook/react-dom-shim": "9.0.17",
+				"@storybook/csf-plugin": "10.0.7",
+				"@storybook/icons": "^1.6.0",
+				"@storybook/react-dom-shim": "10.0.7",
 				"react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
 				"react-dom": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
 				"ts-dedent": "^2.0.0"
@@ -1849,13 +1860,13 @@
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"storybook": "^9.0.17"
+				"storybook": "^10.0.7"
 			}
 		},
 		"node_modules/@storybook/addon-svelte-csf": {
-			"version": "5.0.7",
-			"resolved": "https://registry.npmjs.org/@storybook/addon-svelte-csf/-/addon-svelte-csf-5.0.7.tgz",
-			"integrity": "sha512-6Zmy5HjOlrrG6OoKRTGDr9LR6zRK4/Sa7raFzQRKHGASgMlfKsMdNTNO0sxnMUWCu2JMS6HsuoLtB3Ma8SlYtg==",
+			"version": "5.0.10",
+			"resolved": "https://registry.npmjs.org/@storybook/addon-svelte-csf/-/addon-svelte-csf-5.0.10.tgz",
+			"integrity": "sha512-poSvTS7VdaQ42ZoqW5e4+2Hv1iLO0mekH9fwn/QuBNse48R4WlTyR8XFbHRTfatl9gdc9ZYC4uWzazrmV6zGIA==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
@@ -1868,22 +1879,22 @@
 				"zimmerframe": "^1.1.2"
 			},
 			"peerDependencies": {
-				"@storybook/svelte": "^0.0.0-0 || ^8.2.0 || ^9.0.0 || ^9.1.0-0",
+				"@storybook/svelte": "^0.0.0-0 || ^8.2.0 || ^9.0.0 || ^9.1.0-0 || ^10.0.0-0",
 				"@sveltejs/vite-plugin-svelte": "^4.0.0 || ^5.0.0 || ^6.0.0",
-				"storybook": "^0.0.0-0 || ^8.2.0 || ^9.0.0 || ^9.1.0-0",
+				"storybook": "^0.0.0-0 || ^8.2.0 || ^9.0.0 || ^9.1.0-0 || ^10.0.0-0",
 				"svelte": "^5.0.0",
 				"vite": "^5.0.0 || ^6.0.0 || ^7.0.0"
 			}
 		},
 		"node_modules/@storybook/addon-vitest": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/addon-vitest/-/addon-vitest-9.0.17.tgz",
-			"integrity": "sha512-eogqcGbACR1sTedBSE2SP/4QV1ruicHYEhYjBtoPIjvYgymN1g5KSuQNysLx4f0SvAzczrcNjX2WVVLX2DVyzA==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/addon-vitest/-/addon-vitest-10.0.7.tgz",
+			"integrity": "sha512-i6v/mAl+elrUxb+1f4NdnM17t/fg+KGJWL1U9quflXTd3KiLY0xJB4LwNP6yYo7Imc5NIO2fRkJbGvNqLBRe2Q==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"@storybook/global": "^5.0.0",
-				"@storybook/icons": "^1.4.0",
+				"@storybook/icons": "^1.6.0",
 				"prompts": "^2.4.0",
 				"ts-dedent": "^2.2.0"
 			},
@@ -1892,15 +1903,19 @@
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"@vitest/browser": "^3.0.0",
-				"@vitest/runner": "^3.0.0",
-				"storybook": "^9.0.17",
-				"vitest": "^3.0.0"
+				"@vitest/browser": "^3.0.0 || ^4.0.0",
+				"@vitest/browser-playwright": "^4.0.0",
+				"@vitest/runner": "^3.0.0 || ^4.0.0",
+				"storybook": "^10.0.7",
+				"vitest": "^3.0.0 || ^4.0.0"
 			},
 			"peerDependenciesMeta": {
 				"@vitest/browser": {
 					"optional": true
 				},
+				"@vitest/browser-playwright": {
+					"optional": true
+				},
 				"@vitest/runner": {
 					"optional": true
 				},
@@ -1910,13 +1925,13 @@
 			}
 		},
 		"node_modules/@storybook/builder-vite": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/builder-vite/-/builder-vite-9.0.17.tgz",
-			"integrity": "sha512-lyuvgGhb0NaVk1tdB4xwzky6+YXQfxlxfNQqENYZ9uYQZdPfErMa4ZTXVQTV+CQHAa2NL+p/dG2JPAeu39e9UA==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/builder-vite/-/builder-vite-10.0.7.tgz",
+			"integrity": "sha512-wk2TAoUY5+9t78GWVBndu9rEo9lo6Ec3SRrLT4VpIlcS2GPK+5f26UC2uvIBwOF/N7JrUUKq/zWDZ3m+do9QDg==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"@storybook/csf-plugin": "9.0.17",
+				"@storybook/csf-plugin": "10.0.7",
 				"ts-dedent": "^2.0.0"
 			},
 			"funding": {
@@ -1924,7 +1939,7 @@
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"storybook": "^9.0.17",
+				"storybook": "^10.0.7",
 				"vite": "^5.0.0 || ^6.0.0 || ^7.0.0"
 			}
 		},
@@ -1939,20 +1954,38 @@
 			}
 		},
 		"node_modules/@storybook/csf-plugin": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/csf-plugin/-/csf-plugin-9.0.17.tgz",
-			"integrity": "sha512-6Q4eo1ObrLlsnB6bIt6T8+45XAb4to2pQGNrI7QPkLQRLrZinrJcNbLY7AGkyIoCOEsEbq08n09/nClQUbu8HA==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/csf-plugin/-/csf-plugin-10.0.7.tgz",
+			"integrity": "sha512-YaYYlCyJBwxaMk7yREOdz+9MDSgxIYGdeJ9EIq/bUndmkoj9SRo1P9/0lC5dseWQoiGy4T3PbZiWruD8uM5m3g==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"unplugin": "^1.3.1"
+				"unplugin": "^2.3.5"
 			},
 			"funding": {
 				"type": "opencollective",
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"storybook": "^9.0.17"
+				"esbuild": "*",
+				"rollup": "*",
+				"storybook": "^10.0.7",
+				"vite": "*",
+				"webpack": "*"
+			},
+			"peerDependenciesMeta": {
+				"esbuild": {
+					"optional": true
+				},
+				"rollup": {
+					"optional": true
+				},
+				"vite": {
+					"optional": true
+				},
+				"webpack": {
+					"optional": true
+				}
 			}
 		},
 		"node_modules/@storybook/global": {
@@ -1963,9 +1996,9 @@
 			"license": "MIT"
 		},
 		"node_modules/@storybook/icons": {
-			"version": "1.4.0",
-			"resolved": "https://registry.npmjs.org/@storybook/icons/-/icons-1.4.0.tgz",
-			"integrity": "sha512-Td73IeJxOyalzvjQL+JXx72jlIYHgs+REaHiREOqfpo3A2AYYG71AUbcv+lg7mEDIweKVCxsMQ0UKo634c8XeA==",
+			"version": "1.6.0",
+			"resolved": "https://registry.npmjs.org/@storybook/icons/-/icons-1.6.0.tgz",
+			"integrity": "sha512-hcFZIjW8yQz8O8//2WTIXylm5Xsgc+lW9ISLgUk1xGmptIJQRdlhVIXCpSyLrQaaRiyhQRaVg7l3BD9S216BHw==",
 			"dev": true,
 			"license": "MIT",
 			"engines": {
@@ -1977,9 +2010,9 @@
 			}
 		},
 		"node_modules/@storybook/react-dom-shim": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/react-dom-shim/-/react-dom-shim-9.0.17.tgz",
-			"integrity": "sha512-ak/x/m6MDDxdE6rCDymTltaiQF3oiKrPHSwfM+YPgQR6MVmzTTs4+qaPfeev7FZEHq23IkfDMTmSTTJtX7Vs9A==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/react-dom-shim/-/react-dom-shim-10.0.7.tgz",
+			"integrity": "sha512-bp4OnMtZGwPJQDqNRi4K5iibLbZ2TZZMkWW7oSw5jjPFpGSreSjCe8LH9yj/lDnK8Ox9bGMCBFE5RV5XuML29w==",
 			"dev": true,
 			"license": "MIT",
 			"funding": {
@@ -1987,126 +2020,75 @@
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0-beta",
-				"react-dom": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0-beta",
-				"storybook": "^9.0.17"
+				"react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
+				"react-dom": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
+				"storybook": "^10.0.7"
 			}
 		},
 		"node_modules/@storybook/svelte": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/svelte/-/svelte-9.0.17.tgz",
-			"integrity": "sha512-RwOswdq7S3+ZOuoM/oRrcmlsKdjcd/3wMHbuirzYoAhdwsjubSuRepMV64O9RnlXd3x7rZw4fXpq1M/SVo5XiQ==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/svelte/-/svelte-10.0.7.tgz",
+			"integrity": "sha512-rO+YQhHucy47Vh67z318pALmd6x+K1Kj30Fb4a6oOEw4xn4zCo9KTmkMWs24c4oduEXD/eJu3badlRmsVXzyfA==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"ts-dedent": "^2.0.0",
 				"type-fest": "~2.19"
 			},
-			"engines": {
-				"node": ">=20.0.0"
-			},
 			"funding": {
 				"type": "opencollective",
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"storybook": "^9.0.17",
+				"storybook": "^10.0.7",
 				"svelte": "^5.0.0"
 			}
 		},
-		"node_modules/@storybook/sveltekit": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/sveltekit/-/sveltekit-9.0.17.tgz",
-			"integrity": "sha512-CUOATuW5Qk3SjNvmjH+wyx2GCsMF1cvw3gwkujV9kehPebzV20NhgHpbzSoepvwF7+Bj6jl8V6UxiMWk0jJFmA==",
+		"node_modules/@storybook/svelte-vite": {
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/svelte-vite/-/svelte-vite-10.0.7.tgz",
+			"integrity": "sha512-q9/RtrhX1CnznO6AO9MDEy1bsccbGeRxW28FLpgUrztV4IGZ/dFUrFIFurKRyuA3/nFsbtzp1F5jFt3RExmmTw==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"@storybook/builder-vite": "9.0.17",
-				"@storybook/svelte": "9.0.17",
-				"@storybook/svelte-vite": "9.0.17"
-			},
-			"engines": {
-				"node": ">=20.0.0"
-			},
-			"funding": {
-				"type": "opencollective",
-				"url": "https://opencollective.com/storybook"
-			},
-			"peerDependencies": {
-				"storybook": "^9.0.17",
-				"svelte": "^5.0.0",
-				"vite": "^5.0.0 || ^6.0.0 || ^7.0.0"
-			}
-		},
-		"node_modules/@storybook/sveltekit/node_modules/@storybook/svelte-vite": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/@storybook/svelte-vite/-/svelte-vite-9.0.17.tgz",
-			"integrity": "sha512-fRIxOZy9IRI6BfL1LgFn+B+IckGOlT1SstD01y9ddO4pVKWih/l+vb44bnZs+Z0faJZbrG/LgfnXTOPj052Z8g==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"@storybook/builder-vite": "9.0.17",
-				"@storybook/svelte": "9.0.17",
+				"@storybook/builder-vite": "10.0.7",
+				"@storybook/svelte": "10.0.7",
 				"magic-string": "^0.30.0",
-				"svelte2tsx": "^0.7.35",
+				"svelte2tsx": "^0.7.44",
 				"typescript": "^4.9.4 || ^5.0.0"
 			},
-			"engines": {
-				"node": ">=20.0.0"
+			"funding": {
+				"type": "opencollective",
+				"url": "https://opencollective.com/storybook"
+			},
+			"peerDependencies": {
+				"@sveltejs/vite-plugin-svelte": "^2.0.0 || ^3.0.0 || ^4.0.0 || ^5.0.0 || ^6.0.0",
+				"storybook": "^10.0.7",
+				"svelte": "^5.0.0",
+				"vite": "^5.0.0 || ^6.0.0 || ^7.0.0"
+			}
+		},
+		"node_modules/@storybook/sveltekit": {
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/@storybook/sveltekit/-/sveltekit-10.0.7.tgz",
+			"integrity": "sha512-ujTW7PfWvgBrzd7jzaZe9JgjUeM5YvBKm+xru6t7Dr4bdfmkKqlZHPRdXn/sy+fQNyfg6JL2WKy2KIIeA+RvSg==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"@storybook/builder-vite": "10.0.7",
+				"@storybook/svelte": "10.0.7",
+				"@storybook/svelte-vite": "10.0.7"
 			},
 			"funding": {
 				"type": "opencollective",
 				"url": "https://opencollective.com/storybook"
 			},
 			"peerDependencies": {
-				"@sveltejs/vite-plugin-svelte": "^2.0.0 || ^3.0.0 || ^4.0.0 || ^5.0.0",
-				"storybook": "^9.0.17",
+				"storybook": "^10.0.7",
 				"svelte": "^5.0.0",
 				"vite": "^5.0.0 || ^6.0.0 || ^7.0.0"
 			}
 		},
-		"node_modules/@storybook/sveltekit/node_modules/@sveltejs/vite-plugin-svelte": {
-			"version": "5.1.1",
-			"resolved": "https://registry.npmjs.org/@sveltejs/vite-plugin-svelte/-/vite-plugin-svelte-5.1.1.tgz",
-			"integrity": "sha512-Y1Cs7hhTc+a5E9Va/xwKlAJoariQyHY+5zBgCZg4PFWNYQ1nMN9sjK1zhw1gK69DuqVP++sht/1GZg1aRwmAXQ==",
-			"dev": true,
-			"license": "MIT",
-			"peer": true,
-			"dependencies": {
-				"@sveltejs/vite-plugin-svelte-inspector": "^4.0.1",
-				"debug": "^4.4.1",
-				"deepmerge": "^4.3.1",
-				"kleur": "^4.1.5",
-				"magic-string": "^0.30.17",
-				"vitefu": "^1.0.6"
-			},
-			"engines": {
-				"node": "^18.0.0 || ^20.0.0 || >=22"
-			},
-			"peerDependencies": {
-				"svelte": "^5.0.0",
-				"vite": "^6.0.0"
-			}
-		},
-		"node_modules/@storybook/sveltekit/node_modules/@sveltejs/vite-plugin-svelte/node_modules/@sveltejs/vite-plugin-svelte-inspector": {
-			"version": "4.0.1",
-			"resolved": "https://registry.npmjs.org/@sveltejs/vite-plugin-svelte-inspector/-/vite-plugin-svelte-inspector-4.0.1.tgz",
-			"integrity": "sha512-J/Nmb2Q2y7mck2hyCX4ckVHcR5tu2J+MtBEQqpDrrgELZ2uvraQcK/ioCV61AqkdXFgriksOKIceDcQmqnGhVw==",
-			"dev": true,
-			"license": "MIT",
-			"peer": true,
-			"dependencies": {
-				"debug": "^4.3.7"
-			},
-			"engines": {
-				"node": "^18.0.0 || ^20.0.0 || >=22"
-			},
-			"peerDependencies": {
-				"@sveltejs/vite-plugin-svelte": "^5.0.0",
-				"svelte": "^5.0.0",
-				"vite": "^6.0.0"
-			}
-		},
 		"node_modules/@sveltejs/acorn-typescript": {
 			"version": "1.0.5",
 			"resolved": "https://registry.npmjs.org/@sveltejs/acorn-typescript/-/acorn-typescript-1.0.5.tgz",
@@ -2117,9 +2099,9 @@
 			}
 		},
 		"node_modules/@sveltejs/adapter-static": {
-			"version": "3.0.9",
-			"resolved": "https://registry.npmjs.org/@sveltejs/adapter-static/-/adapter-static-3.0.9.tgz",
-			"integrity": "sha512-aytHXcMi7lb9ljsWUzXYQ0p5X1z9oWud2olu/EpmH7aCu4m84h7QLvb5Wp+CFirKcwoNnYvYWhyP/L8Vh1ztdw==",
+			"version": "3.0.10",
+			"resolved": "https://registry.npmjs.org/@sveltejs/adapter-static/-/adapter-static-3.0.10.tgz",
+			"integrity": "sha512-7D9lYFWJmB7zxZyTE/qxjksvMqzMuYrrsyh1f4AlZqeZeACPRySjbC3aFiY55wb1tWUaKOQG9PVbm74JcN2Iew==",
 			"dev": true,
 			"license": "MIT",
 			"peerDependencies": {
@@ -2127,9 +2109,9 @@
 			}
 		},
 		"node_modules/@sveltejs/kit": {
-			"version": "2.37.0",
-			"resolved": "https://registry.npmjs.org/@sveltejs/kit/-/kit-2.37.0.tgz",
-			"integrity": "sha512-xgKtpjQ6Ry4mdShd01ht5AODUsW7+K1iValPDq7QX8zI1hWOKREH9GjG8SRCN5tC4K7UXmMhuQam7gbLByVcnw==",
+			"version": "2.48.4",
+			"resolved": "https://registry.npmjs.org/@sveltejs/kit/-/kit-2.48.4.tgz",
+			"integrity": "sha512-TGFX1pZUt9qqY20Cv5NyYvy0iLWHf2jXi8s+eCGsig7jQMdwZWKUFMR6TbvFNhfDSUpc1sH/Y5EHv20g3HHA3g==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
@@ -2166,16 +2148,15 @@
 			}
 		},
 		"node_modules/@sveltejs/vite-plugin-svelte": {
-			"version": "6.1.0",
-			"resolved": "https://registry.npmjs.org/@sveltejs/vite-plugin-svelte/-/vite-plugin-svelte-6.1.0.tgz",
-			"integrity": "sha512-+U6lz1wvGEG/BvQyL4z/flyNdQ9xDNv5vrh+vWBWTHaebqT0c9RNggpZTo/XSPoHsSCWBlYaTlRX8pZ9GATXCw==",
+			"version": "6.2.1",
+			"resolved": "https://registry.npmjs.org/@sveltejs/vite-plugin-svelte/-/vite-plugin-svelte-6.2.1.tgz",
+			"integrity": "sha512-YZs/OSKOQAQCnJvM/P+F1URotNnYNeU3P2s4oIpzm1uFaqUEqRxUB0g5ejMjEb5Gjb9/PiBI5Ktrq4rUUF8UVQ==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"@sveltejs/vite-plugin-svelte-inspector": "^5.0.0-next.1",
+				"@sveltejs/vite-plugin-svelte-inspector": "^5.0.0",
 				"debug": "^4.4.1",
 				"deepmerge": "^4.3.1",
-				"kleur": "^4.1.5",
 				"magic-string": "^0.30.17",
 				"vitefu": "^1.1.1"
 			},
@@ -3361,19 +3342,6 @@
 				"node": ">= 0.8"
 			}
 		},
-		"node_modules/better-opn": {
-			"version": "3.0.2",
-			"resolved": "https://registry.npmjs.org/better-opn/-/better-opn-3.0.2.tgz",
-			"integrity": "sha512-aVNobHnJqLiUelTaHat9DZ1qM2w0C0Eym4LPI/3JxOnSokGVdsl1T1kN7TFvsEAD8G47A6VKQ0TVHqbBnYMJlQ==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"open": "^8.0.4"
-			},
-			"engines": {
-				"node": ">=12.0.0"
-			}
-		},
 		"node_modules/bits-ui": {
 			"version": "2.8.11",
 			"resolved": "https://registry.npmjs.org/bits-ui/-/bits-ui-2.8.11.tgz",
@@ -3844,16 +3812,6 @@
 				"node": ">=0.10.0"
 			}
 		},
-		"node_modules/define-lazy-prop": {
-			"version": "2.0.0",
-			"resolved": "https://registry.npmjs.org/define-lazy-prop/-/define-lazy-prop-2.0.0.tgz",
-			"integrity": "sha512-Ds09qNh8yw3khSjiJjiUInaGX9xlqZDY7JVryGxdxV7NPeuqQfplOpQ66yJFZut3jLa5zOwkXw1g9EI2uKh4Og==",
-			"dev": true,
-			"license": "MIT",
-			"engines": {
-				"node": ">=8"
-			}
-		},
 		"node_modules/dequal": {
 			"version": "2.0.3",
 			"resolved": "https://registry.npmjs.org/dequal/-/dequal-2.0.3.tgz",
@@ -4042,19 +4000,6 @@
 				"@esbuild/win32-x64": "0.25.8"
 			}
 		},
-		"node_modules/esbuild-register": {
-			"version": "3.6.0",
-			"resolved": "https://registry.npmjs.org/esbuild-register/-/esbuild-register-3.6.0.tgz",
-			"integrity": "sha512-H2/S7Pm8a9CL1uhp9OvjwrBh5Pvx0H8qVOxNu8Wed9Y7qv56MPtq+GGM8RJpq6glYJn9Wspr8uw7l55uyinNeg==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"debug": "^4.3.4"
-			},
-			"peerDependencies": {
-				"esbuild": ">=0.12 <1"
-			}
-		},
 		"node_modules/escape-string-regexp": {
 			"version": "4.0.0",
 			"resolved": "https://registry.npmjs.org/escape-string-regexp/-/escape-string-regexp-4.0.0.tgz",
@@ -4146,20 +4091,17 @@
 			}
 		},
 		"node_modules/eslint-plugin-storybook": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/eslint-plugin-storybook/-/eslint-plugin-storybook-9.0.17.tgz",
-			"integrity": "sha512-IuTdlwCEwoDNobdygRCxNhlKXHmsDfPtPvHGcsY35x2Bx8KItrjfekO19gJrjc1VT2CMfcZMYF8OBKaxHELupw==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/eslint-plugin-storybook/-/eslint-plugin-storybook-10.0.7.tgz",
+			"integrity": "sha512-qOQq9KdT1jsBgT3qsxUH2n67aj1WR8D1XCoER8Q6yuVlS5TimNwk1mZeWkXVf/o4RQQT6flT2y5cG2gPLZPvJA==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"@typescript-eslint/utils": "^8.8.1"
 			},
-			"engines": {
-				"node": ">=20.0.0"
-			},
 			"peerDependencies": {
 				"eslint": ">=8",
-				"storybook": "^9.0.17"
+				"storybook": "^10.0.7"
 			}
 		},
 		"node_modules/eslint-plugin-svelte": {
@@ -4405,11 +4347,14 @@
 			}
 		},
 		"node_modules/fdir": {
-			"version": "6.4.6",
-			"resolved": "https://registry.npmjs.org/fdir/-/fdir-6.4.6.tgz",
-			"integrity": "sha512-hiFoqpyZcfNm1yc4u8oWCf9A2c4D3QjCrks3zmoVKVxpQRzmPNar1hUJcBG2RQHvEVGDN+Jm81ZheVLAQMK6+w==",
+			"version": "6.5.0",
+			"resolved": "https://registry.npmjs.org/fdir/-/fdir-6.5.0.tgz",
+			"integrity": "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==",
 			"dev": true,
 			"license": "MIT",
+			"engines": {
+				"node": ">=12.0.0"
+			},
 			"peerDependencies": {
 				"picomatch": "^3 || ^4"
 			},
@@ -5072,22 +5017,6 @@
 			"integrity": "sha512-0aO8FkhNZlj/ZIbNi7Lxxr12obT7cL1moPfE4tg1LkX7LlLfC6DeX4l2ZEud1ukP9jNQyNnfzQVqwbwmAATY4Q==",
 			"license": "MIT"
 		},
-		"node_modules/is-docker": {
-			"version": "2.2.1",
-			"resolved": "https://registry.npmjs.org/is-docker/-/is-docker-2.2.1.tgz",
-			"integrity": "sha512-F+i2BKsFrH66iaUFc0woD8sLy8getkwTwtOBjvs56Cx4CgJDeKQeqfz8wAYiSb8JOprWhHH5p77PbmYCvvUuXQ==",
-			"dev": true,
-			"license": "MIT",
-			"bin": {
-				"is-docker": "cli.js"
-			},
-			"engines": {
-				"node": ">=8"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/sindresorhus"
-			}
-		},
 		"node_modules/is-extglob": {
 			"version": "2.1.1",
 			"resolved": "https://registry.npmjs.org/is-extglob/-/is-extglob-2.1.1.tgz",
@@ -5133,19 +5062,6 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
-		"node_modules/is-wsl": {
-			"version": "2.2.0",
-			"resolved": "https://registry.npmjs.org/is-wsl/-/is-wsl-2.2.0.tgz",
-			"integrity": "sha512-fKzAra0rGJUUBwGBgNkHZuToZcn+TtXHpeCgmkMJMMYx1sQDYaCSyjJBSCa2nH1DGm7s3n1oBnohoVTBaN7Lww==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"is-docker": "^2.0.0"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
 		"node_modules/isexe": {
 			"version": "2.0.0",
 			"resolved": "https://registry.npmjs.org/isexe/-/isexe-2.0.0.tgz",
@@ -5591,16 +5507,6 @@
 			"dev": true,
 			"license": "MIT"
 		},
-		"node_modules/lower-case": {
-			"version": "2.0.2",
-			"resolved": "https://registry.npmjs.org/lower-case/-/lower-case-2.0.2.tgz",
-			"integrity": "sha512-7fm3l3NAF9WfN6W3JOmf5drwpVqX78JtoGJ3A6W0a6ZnldM41w2fV5D490psKFTpMds8TJse/eHLFFsNHHjHgg==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"tslib": "^2.0.3"
-			}
-		},
 		"node_modules/lowlight": {
 			"version": "3.3.0",
 			"resolved": "https://registry.npmjs.org/lowlight/-/lowlight-3.3.0.tgz",
@@ -6783,17 +6689,6 @@
 			"dev": true,
 			"license": "MIT"
 		},
-		"node_modules/no-case": {
-			"version": "3.0.4",
-			"resolved": "https://registry.npmjs.org/no-case/-/no-case-3.0.4.tgz",
-			"integrity": "sha512-fgAN3jGAh+RoxUGZHTSOLJIqUc2wmoBwGR4tbpNAKmmovFoWq0OdRkb0VkldReO2a2iBT/OEulG9XSUc10r3zg==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"lower-case": "^2.0.2",
-				"tslib": "^2.0.3"
-			}
-		},
 		"node_modules/node-addon-api": {
 			"version": "7.1.1",
 			"resolved": "https://registry.npmjs.org/node-addon-api/-/node-addon-api-7.1.1.tgz",
@@ -6815,24 +6710,6 @@
 				"url": "https://github.com/sponsors/ljharb"
 			}
 		},
-		"node_modules/open": {
-			"version": "8.4.2",
-			"resolved": "https://registry.npmjs.org/open/-/open-8.4.2.tgz",
-			"integrity": "sha512-7x81NCL719oNbsq/3mh+hVrAWmFuEYUqrq/Iw3kUzH8ReypT9QQ0BLoJS7/G9k6N81XjW4qHWtjWwe/9eLy1EQ==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"define-lazy-prop": "^2.0.0",
-				"is-docker": "^2.1.1",
-				"is-wsl": "^2.2.0"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/sindresorhus"
-			}
-		},
 		"node_modules/opener": {
 			"version": "1.5.2",
 			"resolved": "https://registry.npmjs.org/opener/-/opener-1.5.2.tgz",
@@ -6919,17 +6796,6 @@
 				"url": "https://github.com/inikulin/parse5?sponsor=1"
 			}
 		},
-		"node_modules/pascal-case": {
-			"version": "3.1.2",
-			"resolved": "https://registry.npmjs.org/pascal-case/-/pascal-case-3.1.2.tgz",
-			"integrity": "sha512-uWlGT3YSnK9x3BQJaOdcZwrnV6hPpd8jFH1/ucpiLRPh/2zCVJKS19E4GvYHvaCcACn3foXZ0cLB9Wrx1KGe5g==",
-			"dev": true,
-			"license": "MIT",
-			"dependencies": {
-				"no-case": "^3.0.4",
-				"tslib": "^2.0.3"
-			}
-		},
 		"node_modules/path-exists": {
 			"version": "4.0.0",
 			"resolved": "https://registry.npmjs.org/path-exists/-/path-exists-4.0.0.tgz",
@@ -7000,13 +6866,13 @@
 			}
 		},
 		"node_modules/playwright": {
-			"version": "1.54.1",
-			"resolved": "https://registry.npmjs.org/playwright/-/playwright-1.54.1.tgz",
-			"integrity": "sha512-peWpSwIBmSLi6aW2auvrUtf2DqY16YYcCMO8rTVx486jKmDTJg7UAhyrraP98GB8BoPURZP8+nxO7TSd4cPr5g==",
+			"version": "1.56.1",
+			"resolved": "https://registry.npmjs.org/playwright/-/playwright-1.56.1.tgz",
+			"integrity": "sha512-aFi5B0WovBHTEvpM3DzXTUaeN6eN0qWnTkKx4NQaH4Wvcmc153PdaY2UBdSYKaGYw+UyWXSVyxDUg5DoPEttjw==",
 			"dev": true,
 			"license": "Apache-2.0",
 			"dependencies": {
-				"playwright-core": "1.54.1"
+				"playwright-core": "1.56.1"
 			},
 			"bin": {
 				"playwright": "cli.js"
@@ -7019,9 +6885,9 @@
 			}
 		},
 		"node_modules/playwright-core": {
-			"version": "1.54.1",
-			"resolved": "https://registry.npmjs.org/playwright-core/-/playwright-core-1.54.1.tgz",
-			"integrity": "sha512-Nbjs2zjj0htNhzgiy5wu+3w09YetDx5pkrpI/kZotDlDUaYk0HVA5xrBVPdow4SAUIlhgKcJeJg4GRKW6xHusA==",
+			"version": "1.56.1",
+			"resolved": "https://registry.npmjs.org/playwright-core/-/playwright-core-1.56.1.tgz",
+			"integrity": "sha512-hutraynyn31F+Bifme+Ps9Vq59hKuUCz7H1kDOcBs+2oGguKkWTU50bBWrtz34OUWmIwpBTWDxaRPXrIXkgvmQ==",
 			"dev": true,
 			"license": "Apache-2.0",
 			"bin": {
@@ -7852,6 +7718,13 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/scule": {
+			"version": "1.3.0",
+			"resolved": "https://registry.npmjs.org/scule/-/scule-1.3.0.tgz",
+			"integrity": "sha512-6FtHJEvt+pVMIB9IBY+IcCJ6Z5f1iQnytgyfKMhDKgmzYG+TeH/wx1y3l27rshSbLiSanrR9ffZDrEsmjlQF2g==",
+			"dev": true,
+			"license": "MIT"
+		},
 		"node_modules/secure-compare": {
 			"version": "3.0.1",
 			"resolved": "https://registry.npmjs.org/secure-compare/-/secure-compare-3.0.1.tgz",
@@ -8052,26 +7925,26 @@
 			"license": "MIT"
 		},
 		"node_modules/storybook": {
-			"version": "9.0.17",
-			"resolved": "https://registry.npmjs.org/storybook/-/storybook-9.0.17.tgz",
-			"integrity": "sha512-O+9jgJ+Trlq9VGD1uY4OBLKQWHHDKM/A/pA8vMW6PVehhGHNvpzcIC1bngr6mL5gGHZP2nBv+9XG8pTMcggMmg==",
+			"version": "10.0.7",
+			"resolved": "https://registry.npmjs.org/storybook/-/storybook-10.0.7.tgz",
+			"integrity": "sha512-7smAu0o+kdm378Q2uIddk32pn0UdIbrtTVU+rXRVtTVTCrK/P2cCui2y4JH+Bl3NgEq1bbBQpCAF/HKrDjk2Qw==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"@storybook/global": "^5.0.0",
+				"@storybook/icons": "^1.6.0",
 				"@testing-library/jest-dom": "^6.6.3",
 				"@testing-library/user-event": "^14.6.1",
 				"@vitest/expect": "3.2.4",
+				"@vitest/mocker": "3.2.4",
 				"@vitest/spy": "3.2.4",
-				"better-opn": "^3.0.2",
 				"esbuild": "^0.18.0 || ^0.19.0 || ^0.20.0 || ^0.21.0 || ^0.22.0 || ^0.23.0 || ^0.24.0 || ^0.25.0",
-				"esbuild-register": "^3.5.0",
 				"recast": "^0.23.5",
 				"semver": "^7.6.2",
 				"ws": "^8.18.0"
 			},
 			"bin": {
-				"storybook": "bin/index.cjs"
+				"storybook": "dist/bin/dispatcher.js"
 			},
 			"funding": {
 				"type": "opencollective",
@@ -8418,14 +8291,14 @@
 			}
 		},
 		"node_modules/svelte2tsx": {
-			"version": "0.7.41",
-			"resolved": "https://registry.npmjs.org/svelte2tsx/-/svelte2tsx-0.7.41.tgz",
-			"integrity": "sha512-/TUwpyn/Qc1wcGuayf2GSwvZ7htdAOzpo0JFFm96srKnRXoTD0gy4n06g+XgH8w016S3lPtyFVtFAm+0yJ0BZw==",
+			"version": "0.7.45",
+			"resolved": "https://registry.npmjs.org/svelte2tsx/-/svelte2tsx-0.7.45.tgz",
+			"integrity": "sha512-cSci+mYGygYBHIZLHlm/jYlEc1acjAHqaQaDFHdEBpUueM9kSTnPpvPtSl5VkJOU1qSJ7h1K+6F/LIUYiqC8VA==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"dedent-js": "^1.0.1",
-				"pascal-case": "^3.1.1"
+				"scule": "^1.3.0"
 			},
 			"peerDependencies": {
 				"svelte": "^3.55 || ^4.0.0-next.0 || ^4.0 || ^5.0.0-next.0",
@@ -8535,14 +8408,14 @@
 			"license": "MIT"
 		},
 		"node_modules/tinyglobby": {
-			"version": "0.2.14",
-			"resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.14.tgz",
-			"integrity": "sha512-tX5e7OM1HnYr2+a2C/4V0htOcSQcoSTH9KgJnVvNm5zm/cyEWKJ7j7YutsH9CxMdtOkkLFy2AHrMci9IM8IPZQ==",
+			"version": "0.2.15",
+			"resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.15.tgz",
+			"integrity": "sha512-j2Zq4NyQYG5XMST4cbs02Ak8iJUdxRM0XI5QyxXuZOzKOINmWurp3smXu3y5wDcJrptwpSjgXHzIQxR0omXljQ==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"fdir": "^6.4.4",
-				"picomatch": "^4.0.2"
+				"fdir": "^6.5.0",
+				"picomatch": "^4.0.3"
 			},
 			"engines": {
 				"node": ">=12.0.0"
@@ -8918,17 +8791,19 @@
 			}
 		},
 		"node_modules/unplugin": {
-			"version": "1.16.1",
-			"resolved": "https://registry.npmjs.org/unplugin/-/unplugin-1.16.1.tgz",
-			"integrity": "sha512-4/u/j4FrCKdi17jaxuJA0jClGxB1AvU2hw/IuayPc4ay1XGaJs/rbb4v5WKwAjNifjmXK9PIFyuPiaK8azyR9w==",
+			"version": "2.3.10",
+			"resolved": "https://registry.npmjs.org/unplugin/-/unplugin-2.3.10.tgz",
+			"integrity": "sha512-6NCPkv1ClwH+/BGE9QeoTIl09nuiAt0gS28nn1PvYXsGKRwM2TCbFA2QiilmehPDTXIe684k4rZI1yl3A1PCUw==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"acorn": "^8.14.0",
+				"@jridgewell/remapping": "^2.3.5",
+				"acorn": "^8.15.0",
+				"picomatch": "^4.0.3",
 				"webpack-virtual-modules": "^0.6.2"
 			},
 			"engines": {
-				"node": ">=14.0.0"
+				"node": ">=18.12.0"
 			}
 		},
 		"node_modules/uri-js": {
@@ -9054,18 +8929,18 @@
 			}
 		},
 		"node_modules/vite": {
-			"version": "7.0.5",
-			"resolved": "https://registry.npmjs.org/vite/-/vite-7.0.5.tgz",
-			"integrity": "sha512-1mncVwJxy2C9ThLwz0+2GKZyEXuC3MyWtAAlNftlZZXZDP3AJt5FmwcMit/IGGaNZ8ZOB2BNO/HFUB+CpN0NQw==",
+			"version": "7.2.2",
+			"resolved": "https://registry.npmjs.org/vite/-/vite-7.2.2.tgz",
+			"integrity": "sha512-BxAKBWmIbrDgrokdGZH1IgkIk/5mMHDreLDmCJ0qpyJaAteP8NvMhkwr/ZCQNqNH97bw/dANTE9PDzqwJghfMQ==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"esbuild": "^0.25.0",
-				"fdir": "^6.4.6",
-				"picomatch": "^4.0.2",
+				"fdir": "^6.5.0",
+				"picomatch": "^4.0.3",
 				"postcss": "^8.5.6",
-				"rollup": "^4.40.0",
-				"tinyglobby": "^0.2.14"
+				"rollup": "^4.43.0",
+				"tinyglobby": "^0.2.15"
 			},
 			"bin": {
 				"vite": "bin/vite.js"
@@ -24,20 +24,20 @@
 		"cleanup": "rm -rf .svelte-kit build node_modules test-results"
 	},
 	"devDependencies": {
-		"@chromatic-com/storybook": "^4.0.1",
+		"@chromatic-com/storybook": "^4.1.2",
 		"@eslint/compat": "^1.2.5",
 		"@eslint/js": "^9.18.0",
 		"@internationalized/date": "^3.8.2",
 		"@lucide/svelte": "^0.515.0",
 		"@playwright/test": "^1.49.1",
-		"@storybook/addon-a11y": "^9.0.17",
-		"@storybook/addon-docs": "^9.0.17",
-		"@storybook/addon-svelte-csf": "^5.0.7",
-		"@storybook/addon-vitest": "^9.0.17",
-		"@storybook/sveltekit": "^9.0.17",
-		"@sveltejs/adapter-static": "^3.0.8",
-		"@sveltejs/kit": "^2.22.0",
-		"@sveltejs/vite-plugin-svelte": "^6.0.0",
+		"@storybook/addon-a11y": "^10.0.7",
+		"@storybook/addon-docs": "^10.0.7",
+		"@storybook/addon-svelte-csf": "^5.0.10",
+		"@storybook/addon-vitest": "^10.0.7",
+		"@storybook/sveltekit": "^10.0.7",
+		"@sveltejs/adapter-static": "^3.0.10",
+		"@sveltejs/kit": "^2.48.4",
+		"@sveltejs/vite-plugin-svelte": "^6.2.1",
 		"@tailwindcss/forms": "^0.5.9",
 		"@tailwindcss/typography": "^0.5.15",
 		"@tailwindcss/vite": "^4.0.0",
@@ -48,21 +48,21 @@
 		"dexie": "^4.0.11",
 		"eslint": "^9.18.0",
 		"eslint-config-prettier": "^10.0.1",
-		"eslint-plugin-storybook": "^9.0.17",
+		"eslint-plugin-storybook": "^10.0.7",
 		"eslint-plugin-svelte": "^3.0.0",
 		"fflate": "^0.8.2",
 		"globals": "^16.0.0",
 		"http-server": "^14.1.1",
 		"mdast": "^3.0.0",
 		"mdsvex": "^0.12.3",
-		"playwright": "^1.53.0",
+		"playwright": "^1.56.1",
 		"prettier": "^3.4.2",
 		"prettier-plugin-svelte": "^3.3.3",
 		"prettier-plugin-tailwindcss": "^0.6.11",
 		"rehype-katex": "^7.0.1",
 		"remark-math": "^6.0.0",
 		"sass": "^1.93.3",
-		"storybook": "^9.0.17",
+		"storybook": "^10.0.7",
 		"svelte": "^5.0.0",
 		"svelte-check": "^4.0.0",
 		"tailwind-merge": "^3.3.1",
@@ -73,7 +73,7 @@
 		"typescript-eslint": "^8.20.0",
 		"unified": "^11.0.5",
 		"uuid": "^13.0.0",
-		"vite": "^7.0.4",
+		"vite": "^7.2.2",
 		"vite-plugin-devtools-json": "^0.2.0",
 		"vitest": "^3.2.3",
 		"vitest-browser-svelte": "^0.1.0"
@@ -1,7 +1,7 @@
 <script module lang="ts">
 	import { defineMeta } from '@storybook/addon-svelte-csf';
 	import ChatForm from '$lib/components/app/chat/ChatForm/ChatForm.svelte';
-	import { expect } from 'storybook/internal/test';
+	import { expect } from 'storybook/test';
 	import { mockServerProps, mockConfigs } from './fixtures/storybook-mocks';
 	import jpgAsset from './fixtures/assets/1.jpg?url';
 	import svgAsset from './fixtures/assets/hf-logo.svg?url';
@@ -1,7 +1,7 @@
 <script module lang="ts">
 	import { defineMeta } from '@storybook/addon-svelte-csf';
 	import ChatSidebar from '$lib/components/app/chat/ChatSidebar/ChatSidebar.svelte';
-	import { waitFor } from 'storybook/internal/test';
+	import { waitFor } from 'storybook/test';
 	import { screen } from 'storybook/test';

 	const { Story } = defineMeta({
@@ -1,5 +1,6 @@
 <script module lang="ts">
 	import { defineMeta } from '@storybook/addon-svelte-csf';
+	import { expect } from 'storybook/test';
 	import { MarkdownContent } from '$lib/components/app';
 	import { AI_TUTORIAL_MD } from './fixtures/ai-tutorial.js';
 	import { API_DOCS_MD } from './fixtures/api-docs.js';
@@ -68,64 +69,62 @@ All links should have \`target="_blank"\` and \`rel="noopener noreferrer"\` attr
 		class: 'max-w-[56rem] w-[calc(100vw-2rem)]'
 	}}
 	play={async ({ canvasElement }) => {
-		const { expect } = await import('storybook/internal/test');
-		
 		// Wait for component to render
-		await new Promise(resolve => setTimeout(resolve, 100));
-		
+		await new Promise((resolve) => setTimeout(resolve, 100));
+
 		// Find all links in the rendered content
 		const links = canvasElement.querySelectorAll('a[href]');
-		
+
 		// Test that we have the expected number of links
 		expect(links.length).toBeGreaterThan(0);
-		
+
 		// Test each link for proper attributes
 		links.forEach((link) => {
 			const href = link.getAttribute('href');
-			
+
 			// Test that external links have proper security attributes
 			if (href && (href.startsWith('http://') || href.startsWith('https://'))) {
 				expect(link.getAttribute('target')).toBe('_blank');
 				expect(link.getAttribute('rel')).toBe('noopener noreferrer');
 			}
 		});
-		
+
 		// Test specific links exist
-		const hugginFaceLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://huggingface.co'
+		const hugginFaceLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://huggingface.co'
 		);
 		expect(hugginFaceLink).toBeTruthy();
 		expect(hugginFaceLink?.textContent).toBe('Hugging Face Homepage');
-		
-		const githubLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://github.com/ggml-org/llama.cpp'
+
+		const githubLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://github.com/ggml-org/llama.cpp'
 		);
 		expect(githubLink).toBeTruthy();
 		expect(githubLink?.textContent).toBe('GitHub Repository');
-		
-		const openaiLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://openai.com'
+
+		const openaiLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://openai.com'
 		);
 		expect(openaiLink).toBeTruthy();
 		expect(openaiLink?.textContent).toBe('OpenAI Website');
-		
-		const googleLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://www.google.com'
+
+		const googleLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://www.google.com'
 		);
 		expect(googleLink).toBeTruthy();
 		expect(googleLink?.textContent).toBe('Google Search');
-		
+
 		// Test inline links (auto-linked URLs)
-		const exampleLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://example.com'
+		const exampleLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://example.com'
 		);
 		expect(exampleLink).toBeTruthy();
-		
-		const pythonDocsLink = Array.from(links).find(link => 
-			link.getAttribute('href') === 'https://docs.python.org'
+
+		const pythonDocsLink = Array.from(links).find(
+			(link) => link.getAttribute('href') === 'https://docs.python.org'
 		);
 		expect(pythonDocsLink).toBeTruthy();
-		
+
 		console.log(`✅ URL Links test passed - Found ${links.length} links with proper attributes`);
 	}}
 />
@@ -0,0 +1,60 @@
+set(TARGET cpp-httplib)
+
+find_package(Threads REQUIRED)
+
+add_library(${TARGET} STATIC httplib.cpp httplib.h)
+if (NOT MSVC)
+    # disable warnings in 3rd party code
+    target_compile_options(${TARGET} PRIVATE -w)
+endif()
+
+target_link_libraries  (${TARGET} PRIVATE Threads::Threads)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+target_compile_definitions(${TARGET} PRIVATE
+    # increase max payload length to allow use of larger context size
+    CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH=1048576
+    # increase backlog size to avoid connection resets for >> 1 slots
+    CPPHTTPLIB_LISTEN_BACKLOG=512
+    # increase max URI length to handle longer prompts in query string
+    CPPHTTPLIB_REQUEST_URI_MAX_LENGTH=32768
+    # disable Nagle's algorithm
+    CPPHTTPLIB_TCP_NODELAY=1
+)
+
+if (LLAMA_OPENSSL)
+    find_package(OpenSSL)
+    if (OpenSSL_FOUND)
+        include(CheckCSourceCompiles)
+        set(SAVED_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
+        set(CMAKE_REQUIRED_INCLUDES ${OPENSSL_INCLUDE_DIR})
+        check_c_source_compiles("
+        #include <openssl/opensslv.h>
+        #if defined(OPENSSL_IS_BORINGSSL) || defined(LIBRESSL_VERSION_NUMBER)
+        #    if OPENSSL_VERSION_NUMBER < 0x1010107f
+        #        error bad version
+        #    endif
+        #else
+        #    if OPENSSL_VERSION_NUMBER < 0x30000000L
+        #        error bad version
+        #    endif
+        #endif
+        int main() { return 0; }
+        " OPENSSL_VERSION_SUPPORTED)
+        set(CMAKE_REQUIRED_INCLUDES ${SAVED_CMAKE_REQUIRED_INCLUDES})
+        if (OPENSSL_VERSION_SUPPORTED)
+            message(STATUS "OpenSSL found: ${OPENSSL_VERSION}")
+            target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT)
+            target_link_libraries(${TARGET} PUBLIC OpenSSL::SSL OpenSSL::Crypto)
+            if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")
+                target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN)
+                find_library(CORE_FOUNDATION_FRAMEWORK CoreFoundation REQUIRED)
+                find_library(SECURITY_FRAMEWORK Security REQUIRED)
+                target_link_libraries(${TARGET} PUBLIC ${CORE_FOUNDATION_FRAMEWORK} ${SECURITY_FRAMEWORK})
+            endif()
+        endif()
+    else()
+        message(STATUS "OpenSSL not found, SSL support disabled")
+    endif()
+endif()
+
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Ruben Ortlam	a19bd6f7ce	vulkan: remove shell call from vulkan-shaders-gen tool, revert file check (#17219 ) * vulkan: remove shell call from vulkan-shaders-gen tool * use string vector for command execution * Fix condition * use string, remove const_cast * Fix dependency file quotation on Windows --------- Co-authored-by: Jeff Bolz <jbolz@nvidia.com>	2025-11-13 14:51:21 +01:00
Diego Devesa	dd091e52f8	sched : fix reserve ignoring user tensor assignments (#17232 )	2025-11-13 13:14:02 +01:00
ixgbe	1215dde7b0	ggml-cpu : add RISC-V vector intrinsic support for silu and cvar operations (#17227 ) Signed-off-by: Wang Yang <yangwang@iscas.ac.cn>	2025-11-13 13:13:32 +01:00
bagheera	0cfb19166b	metal: accelerated conv2d (#17175 ) * metal: accelerated conv2d * cont : cleanup --------- Co-authored-by: bghira <bghira@users.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2025-11-13 13:32:44 +02:00
Georgi Gerganov	2776db6c81	Revert "ggml-cpu: handle 3d tensors in repack mat_mul (#17030 )" (#17233 ) This reverts commit `1c398dc9ec`.	2025-11-13 12:59:37 +02:00
Diego Devesa	879dec341a	ggml-cpu : use template for argsort (#17222 )	2025-11-13 10:59:05 +02:00
TecJesh	97d5117217	CANN: Add cross_entropy_loss op support (#16886 ) * update L2_NORM op support * update L2_NORM op support * remove extra whitespace * cann: update cross_entropy_loss op support * remove trailing whitespaces * rebase the latest code in the main repository and remove the l2_norm operator that already exists in another pull request. * undo the l2_norm operator deletion	2025-11-13 09:39:51 +08:00
Aman Gupta	a90eb94ca9	CUDA: fuse rope + set_rows (#16884 ) * CUDA: add fused rope * move k forward_expand up * create helper function instead of re-using params * make assert statement more in line with comment * rope_norm: coalesced writes to global mem	2025-11-13 08:50:01 +08:00
Neo Zhang Jianyu	07751f8d44	update SYCL support OPs (#17208 ) Co-authored-by: Zhang Jianyu <zhang.jianyu@outlook.com>	2025-11-13 08:42:23 +08:00
o7si	ffb6f3d921	vocab : correct bounds check for UGM XCDA array access (#17215 )	2025-11-12 23:41:02 +01:00
Johannes Gäßler	5d6838b74f	CUDA: static assert to prevent misuse of memcpy_1 (#17198 )	2025-11-12 23:13:55 +01:00
Mike Abbott	92bb442ad9	docker : preserve .so symlinks for docker container builds (#17214 )	2025-11-12 20:33:55 +01:00
Georgi Gerganov	374fe09cdd	ggml : use std::sort in ggml_argsort CPU implementation (#17211 ) * ggml : use std::sort in ggml_argsort CPU implementation * cont : add missing header	2025-11-12 20:43:38 +02:00
Aleksander Grygier	8e878f0cb4	Update packages + upgrade Storybook to v10 (#17201 ) * chore: Update packages + upgrade Storybook to v10 * fix: Increase timeout for UI tests	2025-11-12 19:01:48 +01:00
Xuan-Son Nguyen	00c94083b3	server: (refactor) implement generator-based API for task results (#17174 ) * server: (refactor) implement generator-based API for task results * improve * moving some code * fix "Response ended prematurely" * add sink.done before return false * rm redundant check * rm unused var * rename generator --> reader	2025-11-12 18:50:52 +01:00
Xuan-Son Nguyen	017eceed61	ci: add check vendor job (#17179 ) * ci: add check vendor job * use dev version of miniaudio * move to dedicated workflow, only run on related files changed	2025-11-12 14:56:02 +01:00
Xuan-Son Nguyen	ee8dd5c658	server: move res_error/res_ok to static function (#17167 )	2025-11-12 14:17:24 +01:00
Alberto Cabrera Pérez	1c398dc9ec	ggml-cpu: handle 3d tensors in repack mat_mul (#17030 ) * ggml-cpu: handle 3d tensors in repack mul_mat * Removed unnecessary branch, removed need for <algorithm> * Fixed dst_ptr pointer in chunk + clang_format * GGML_ASSERT to check wdata within bounds * Accidental ggml.h inclusion * Improved GGML_ASSERT on wdata boundaries	2025-11-12 14:52:19 +02:00
Adrien Gallouët	52cf111b31	cmake : cleanup (#17199 )	2025-11-12 14:48:30 +02:00
Adrien Gallouët	78010a0d52	cmake : move OpenSSL linking to vendor/cpp-httplib (#17177 ) * cmake : move OpenSSL linking to vendor/cpp-httplib Signed-off-by: Adrien Gallouët <angt@huggingface.co> * bring back httplib 0.27.0 * add -DLLAMA_HTTPLIB * update cmake config for visionos --------- Signed-off-by: Adrien Gallouët <angt@huggingface.co> Co-authored-by: Xuan Son Nguyen <son@huggingface.co>	2025-11-12 12:32:50 +01:00
TecJesh	655cddd174	CANN: Add L2_NORM op support (#16856 ) * update L2_NORM op support * update L2_NORM op support * remove extra whitespace	2025-11-12 15:11:42 +08:00
Neo Zhang Jianyu	5da7664960	[SYCL]fix ci crash about SSM_CONV (#17169 ) * fix ci crash * Update ggml-sycl.cpp * Update ggml/src/ggml-sycl/ggml-sycl.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Zhang Jianyu <zhang.jianyu@outlook.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2025-11-12 14:44:29 +08:00
Raul Torres	23a46ce972	CANN: GGML_CANN_ACL_GRAPH works only USE_ACL_GRAPH enabled (#16861 ) The documentation should state that `GGML_CANN_ACL_GRAPH` is only effective if `USE_ACL_GRAPH` was enabled at compilation time.	2025-11-12 14:37:52 +08:00
Max Krasnyansky	c273d75375	hexagon: various Op fixes (#17135 ) * hexagon: explicitly check for ops with zero nrows llm_graph_context::build_inp_out_ids() can generate tensors with zero nrows. Somehow other backends seems to handle this without obvious explicit checks. In the hexagon case we need to check explicitly and skip them. * hexagon: introduce fastdiv, fix test-backend-ops for ADD/SUB/MUL Co-authored-by: chraac <chraac@gmail.com> * hexagon: use fastdiv in ADD_ID * hexagon: use ggml_op_is_empty and ggml_is_empty to check for NOPs --------- Co-authored-by: chraac <chraac@gmail.com>	2025-11-11 15:25:04 -08:00
Eve	7d019cff74	disable rms norm mul rope for chips with no fp16 rte (#17134 )	2025-11-11 12:53:30 -06:00
sudhiarm	3fe36c3238	ci: add Arm-hosted Graviton4 runner (#17021 ) * ci: add Arm-hosted Graviton4 runner * ci: add missing dependencies for graviton4 build * ci: enable LFS checkout on graviton4 * ci: move git-lfs install to dependencies in Graviton4 workflow	2025-11-11 17:58:05 +02:00
Xuan-Son Nguyen	1d45b4228f	vendor: split httplib to cpp/h files (#17150 ) * vendor: split httplib to cpp/h files * move defines * include httplib if curl is not used * add TODO * fix build ios * fix build visionos instead	2025-11-11 13:32:58 +01:00
ixgbe	ca4844062b	ggml-cpu : add RISC-V RVV (Zvfh) optimization for FP16 to FP32 conversion (#17161 ) Signed-off-by: Wang Yang <yangwang@iscas.ac.cn>	2025-11-11 13:41:51 +02:00
duduta	73460f6278	ggml-cpu: templateify ggml_compute_forward_rope_f32 and _f16 (#16805 ) * extract rotate_pairs logic from ggml_compute_forward_rope_f32 * templateify ggml_compute_forward_rope_f32 and _f16 * abort when rope type not supported, remove GLM from test-rope * add imrope branch to switch * add rope tests for perf * Update ggml/src/ggml-cpu/ops.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update ggml/src/ggml-cpu/ops.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2025-11-11 13:33:24 +02:00
Charles Xu	8c583242ad	kleidiai: add optimized per-channel kernels for Q8_0 (#16993 )	2025-11-11 13:20:31 +02:00
Mike Abbott	4a5b8aff40	cmake : add version to all shared object files (#17091 ) When compiling llama.cpp in Yocto, it fails QA checks because the generated so files aren't versioned. This applies a version to all generated so files, allowing the package to build without errors.	2025-11-11 13:19:50 +02:00
Nicolas B. Pierron	d2d626938a	Install rpc-server when GGML_RPC is ON. (#17149 )	2025-11-11 10:53:59 +00:00
levkropp	2fc392ce35	convert : register UMT5Model architecture for T5 conversion (#17160 ) Register UMT5Model as a supported architecture variant for T5 model conversion. This allows the conversion to work for models downloaded with AutoModel.	2025-11-11 09:38:30 +01:00
lhez	ece0f5c177	opencl: add fastdiv and use it in set_rows, ported from cuda (#17090 ) * opencl: add fastdiv for mm q8_0 * opencl: use uint4 for fastdiv vals * opencl: use fastdiv for set_rows * opencl: do not use fastdiv for q8_0 mm	2025-11-10 15:00:13 -08:00
Sigbjørn Skjæret	7bef684118	models : move build_inp_out_ids outside loop (#17151 ) * move build_inp_out_ids outside loop * realign	2025-11-10 22:55:30 +01:00
Max Krasnyansky	395e286bc9	cpu: skip NOPs to avoid barriers (#17133 ) * cpu: skip NOPs to avoid barriers * cpu: use ggml_op_is_empty	2025-11-10 12:44:49 -08:00
Georgi Gerganov	13730c183b	metal : cap threadgroups size of set_rows (#17146 )	2025-11-10 21:33:35 +02:00
Adrien Gallouët	967eb4b2bf	ggml-cpu : inspect -march and -mcpu to found the CPU (#16333 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2025-11-10 21:03:36 +02:00
Ruben Ortlam	f117be185e	vulkan: check glslc executable string (#17144 )	2025-11-10 16:59:26 +01:00
Ruben Ortlam	85234a4b3a	vulkan: fix validation issue introduced by #16868 (#17145 )	2025-11-10 16:59:10 +01:00
Gabe Goodhart	0c74f32632	memory: Hybrid context shift (#17009 ) * feat(memory): Only fail partial erasure of recurrent tail The recurrent state is always assumed to be the state as of the last update from the final token in the sequence. When doing a partial erasure, if the range does not include the final token, the erasure can be considered a success since any memory used for the sequence prior to the final token (which is no memory) has been successfully removed. There is one potential case that this doesn't address which is the pruning of cache to remove sensitive data from the context. This wouldn't work for attention cache partial removal (in the middle) either since the KV state is linearly-dependent and states in later sequence positions would still be based on the state from the sensitive data, even if that data is no longer cached, so I don't think this is relevant, but it is worth noting that the semantics of this change for a partial erasure in the middle of the cache are essentially "my context is already compressed" and not "all trace of the removed tokens has been removed." https://github.com/ggml-org/llama.cpp/issues/16768 Branch: HybridContextShift-16768 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(main): Check the output of seq_rm for prefix matching This prefix matching is explicitly attempting to remove the tokens at the end of the sequence that don't match. This is the operation that can't be performed on a recurrent cache due to the state being updated in place, so if this removal fails, we need to clear the whole cache. https://github.com/ggml-org/llama.cpp/issues/16768 Branch: HybridContextShift-16768 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(memory): Fix condition for partial erasure failure if p0 > pos Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> Co-authored-by: compilade <git@compilade.net> * style: Fix extra parens Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * fix(main.cpp): Set n_matching_session_tokens to 0 on cache clear https://github.com/ggml-org/llama.cpp/issues/16768 Branch: HybridContextShift-16768 Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> --------- Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> Co-authored-by: compilade <git@compilade.net> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2025-11-10 17:14:23 +02:00
Georgi Gerganov	c27efd2bd1	metal : enable tensor API for A19 (#17087 )	2025-11-10 15:38:42 +02:00
fj-y-saito	df70bedda7	arm64: add i8mm route with SVE ggml_vec_dot_q4_K_q8_K and ggml_vec_dot_q6_K_… (#15277 ) * add i8mm route with SVE ggml_vec_dot_q4_K_q8_K and ggml_vec_dot_q6_K_q8_K * Surround SVE function with compiler directive * fix compile switch * fix coding style * ggml : fix indent --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2025-11-10 15:12:59 +02:00
Georgi Gerganov	f914544b16	batched-bench : add "separate text gen" mode (#17103 )	2025-11-10 12:59:29 +02:00
Xuan-Son Nguyen	4b13a684c5	mtmd: fix patch_size initialized to random value in audio models (#17128 ) * mtmd: fix patch_size initialized to random value in audio models * add default hparams	2025-11-10 11:41:05 +01:00
Georgi Gerganov	9898b57cbe	editorconfig : ignore benches/ (#17140 ) [no ci]	2025-11-10 12:17:19 +02:00
Acly	1032256ec9	cuda/vulkan : bicubic interpolation (#17022 ) * vulkan : implement upscale with bicubic interpolation * cuda : implement upscale with bicubic interpolation * tests : add ggml_interpolate with GGML_SCALE_MODE_BICUBIC to backend tests * adapt OpenCL backend to not support the OP in that case so tests don't fail * print scale mode & flags in test-backend-ops	2025-11-10 10:19:39 +01:00
Georgi Gerganov	15274c0c50	benches : add eval results (#17139 ) [no ci]	2025-11-10 10:44:10 +02:00
Georgi Gerganov	b8595b16e6	mtmd : fix embedding size for image input (#17123 )	2025-11-09 18:31:02 +02:00
Ruben Ortlam	392e09a608	vulkan: fix memory allocations (#17122 )	2025-11-09 16:14:41 +01:00
compilade	802cef44bf	convert : parse safetensors directly (#15667 ) * convert : parse safetensors directly * gguf-py : order safetensors tensors by name Applies to both local and remote safetensors custom parsing. This matches the behavior of the official safetensors implementation. * convert : rename from_safetensors_meta to from_local_tensor For consistency with from_remote_tensor * convert : fix no-lazy dtypes from direct safetensors	2025-11-09 09:49:40 -05:00
compilade	1c07c0c68c	convert : handle compressed-tensors quant method (#17069 ) * convert : handle compressed-tensors quant method * convert : handle int-quantized models * convert : handle naive-quantized models * gguf-py : __pos__ is also unary * convert : fix flake8 lint * convert : use F32 for dequant of pack-quantized tensors	2025-11-09 09:45:50 -05:00
Georgi Gerganov	cb1adf8851	server : handle failures to restore host cache (#17078 ) * server : handle failures to restore host cache * server : add tests for the prompt cache	2025-11-09 14:27:05 +02:00
Georgi Gerganov	ef1d826997	benches : add folder with benchmarks (#16931 ) * benches : add folder with benchmarks * benches : update dgx-spark bench	2025-11-09 12:53:29 +02:00
Eric Curtin	86fde91e62	Switch to using Ubuntu 25.10 vulkan/mesa (#16497 ) Because "Ubuntu packages to be discontinued in Vulkan SDK" Signed-off-by: Eric Curtin <eric.curtin@docker.com>	2025-11-09 10:25:38 +01:00
Ruben Ortlam	7f3e9d339c	vulkan: iGPU memory reporting fix (#17110 ) * vulkan: use all device-local heaps for memory availability reporting Co-authored-by: Giuseppe Scrivano <gscrivan@redhat.com> * use all available heaps for iGPU memory reporting * Allow multiple memory types per buffer request for devices with split heaps --------- Co-authored-by: Giuseppe Scrivano <gscrivan@redhat.com>	2025-11-09 09:54:47 +01:00
Ruben Ortlam	8a3519b708	vulkan: fix mmq out of bounds reads (#17108 ) * vulkan: fix mmq out of bounds reads, streamline outdated matmul host code * fix mul_mat_id quantization call * Fix compiler warnings	2025-11-09 09:52:57 +01:00
Jeff Bolz	80a6cf6347	vulkan: fuse mul_mat_id + mul (#17095 ) * vulkan: fuse mul_mat_id + mul This comes up in qwen3 moe. * split mul_mat_id fusion tests into a separate class	2025-11-09 09:48:42 +01:00
Georgi Gerganov	0750a59903	metal : retain src and dst buffers during async ops (#17101 )	2025-11-09 08:28:51 +02:00