kv-cache : simplify + fix warning for recurrent models (#12756)

ggml-ci

.github/workflows/build-linux-cross.yml

@@ -0,0 +1,121 @@
+name: Build on Linux using cross-compiler
+on:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  ubuntu-latest-riscv64-cpu-cross:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo dpkg --add-architecture riscv64
+          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
+                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
+          sudo apt-get clean
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  ubuntu-latest-riscv64-vulkan-cross:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Riscv
+        run: |
+          sudo dpkg --add-architecture riscv64
+          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
+                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
+          sudo apt-get clean
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu \
+                  libvulkan-dev:riscv64
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  ubuntu-latest-arm64-vulkan-cross:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Arm64
+        run: |
+          sudo dpkg --add-architecture arm64
+          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
+                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
+          sudo apt-get clean
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  crossbuild-essential-arm64 \
+                  libvulkan-dev:arm64
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+                         -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
+                         -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -10,7 +10,7 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
   pull_request:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
@@ -467,6 +467,7 @@ jobs:
         run: |
           cmake -B build -S . \
             -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
@@ -476,6 +477,7 @@ jobs:
           cmake -B build2 -S . \
             -DCMAKE_C_COMPILER=hipcc \
             -DCMAKE_CXX_COMPILER=hipcc \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
             -DGGML_HIP=ON
           cmake --build build2 --config Release -j $(nproc)
 
@@ -604,6 +606,9 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
+  build-linux-cross:
+    uses: ./.github/workflows/build-linux-cross.yml
+
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -674,6 +679,35 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
+  macOS-latest-cmake-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
   macOS-latest-swift:
     runs-on: macos-latest
 
@@ -710,12 +744,11 @@ jobs:
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-          sudo cmake --install build --config Release
 
       - name: xcodebuild for swift package
         id: xcodebuild
         run: |
-          xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}"
+          ./build-xcframework.sh
 
   windows-msys2:
     runs-on: windows-latest
@@ -773,7 +806,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.3.261.1
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
@@ -1203,6 +1236,11 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
+        run: |
+          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+
       - name: Install
         id: depends
         run: |
@@ -1232,8 +1270,10 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_HIP=ON `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
@@ -1252,6 +1292,11 @@ jobs:
         with:
             fetch-depth: 0
 
+      - name: Clone rocWMMA repository
+        id: clone_rocwmma
+        run: |
+          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
         with:
@@ -1281,8 +1326,10 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
             -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
@@ -1321,6 +1368,8 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: Build
         id: cmake_build
@@ -1336,15 +1385,40 @@ jobs:
             -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-          sudo cmake --install build --config Release
 
       - name: xcodebuild for swift package
         id: xcodebuild
         run: |
-          xcodebuild -scheme llama-Package -destination 'generic/platform=iOS'
+          ./build-xcframework.sh
 
       - name: Build Xcode project
-        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
+        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
+          name: llama-${{ steps.tag.outputs.name }}-xcframework
 
   android-build:
     runs-on: ubuntu-latest

.github/workflows/server.yml

@@ -161,6 +161,8 @@ jobs:
       - name: Tests
         id: server_integration_tests
         if: ${{ matrix.sanitizer == '' }}
+        env:
+          GITHUB_ACTIONS: "true"
         run: |
           cd examples/server/tests
           ./tests.sh

.gitignore

@@ -45,6 +45,8 @@ lcov-report/
 tags
 .build/
 build*
+release
+debug
 !build-info.cmake
 !build-info.cpp.in
 !build-info.sh

AUTHORS

@@ -1,4 +1,4 @@
-# date: Tue Feb  4 13:04:05 EET 2025
+# date: Sat Mar  8 18:23:52 EET 2025
 # this file is auto-generated by scripts/gen-authors.sh
 
 0cc4m <picard12@live.de>
@@ -8,10 +8,12 @@
 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
 44670 <44670@users.noreply.github.com>
 65a <10104049+65a@users.noreply.github.com>
+708-145 <40387547+708-145@users.noreply.github.com>
 AN Long <aisk@users.noreply.github.com>
 AT <manyoso@users.noreply.github.com>
 Aarni Koskela <akx@iki.fi>
 Aaron Miller <apage43@ninjawhale.com>
+Aaron Teo <57927438+taronaeo@users.noreply.github.com>
 Aaryaman Vasishta <aaryaman.vasishta@amd.com>
 Abheek Gulati <abheekg@hotmail.com>
 Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
@@ -20,6 +22,7 @@ Adithya Balaji <adithya.b94@gmail.com>
 AdithyanI <adithyan.i4internet@gmail.com>
 Adrian <smith.adriane@gmail.com>
 Adrian Hesketh <a-h@users.noreply.github.com>
+Adrian Kretz <me@akretz.com>
 Adrien Gallouët <adrien@gallouet.fr>
 Adrien Gallouët <angt@huggingface.co>
 Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com>
@@ -28,15 +31,18 @@ AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
 AidanBeltonS <aidan.belton@codeplay.com>
 Aisuko <urakiny@gmail.com>
 Akarshan Biswas <akarshan.biswas@gmail.com>
+Akarshan Biswas <akarshan@menlo.ai>
 Akarshan Biswas <akarshanbiswas@fedoraproject.org>
 Al Mochkin <14274697+amochkin@users.noreply.github.com>
 Albert Jin <albert.jin@gmail.com>
 Alberto <57916483+albbus-stack@users.noreply.github.com>
 Alberto Cabrera Pérez <alberto.cabrera@codeplay.com>
 Alberto Cabrera Pérez <alberto.cabrera@intel.com>
+Aleksei Nikiforov <103434461+AlekseiNikiforovIBM@users.noreply.github.com>
 Alex <awhill19@icloud.com>
 Alex Azarov <alex@azarov.by>
 Alex Azarov <alexander.azarov@mapbox.com>
+Alex Brooks <alex.brooks@ibm.com>
 Alex Klinkhamer <from.github.com.917@grencez.dev>
 Alex Klinkhamer <git@grencez.dev>
 Alex Nguyen <tiendung@users.noreply.github.com>
@@ -67,6 +73,7 @@ Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com>
 Andy Salerno <andysalerno@gmail.com>
 Andy Tai <andy-tai@users.noreply.github.com>
 Anthony Van de Gejuchte <anthonyvdgent@gmail.com>
+Antoine Viallon <antoine@lesviallon.fr>
 Antonis Makropoulos <benuix@gmail.com>
 Arik Poznanski <arikpoz@users.noreply.github.com>
 Armen Kaleshian <kriation@users.noreply.github.com>
@@ -83,6 +90,7 @@ Atsushi Tatsuma <yoshoku@outlook.com>
 Austin <77757836+teleprint-me@users.noreply.github.com>
 AustinMroz <austinmroz@utexas.edu>
 BADR <contact@pythops.com>
+BB-fat <45072480+BB-fat@users.noreply.github.com>
 Bach Le <bach@bullno1.com>
 Bailey Chittle <39804642+bachittle@users.noreply.github.com>
 BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
@@ -101,6 +109,7 @@ Bert Wagner <github@bertwagner.com>
 Billel Mokeddem <billel.mokeddem.ml@gmail.com>
 Bingan <70050083+binganao@users.noreply.github.com>
 Bjarke Viksøe <164612031+bviksoe@users.noreply.github.com>
+Bodhi <3882561+BodhiHu@users.noreply.github.com>
 Bodo Graumann <mail@bodograumann.de>
 Bono Lv <lvscar@users.noreply.github.com>
 Borislav Stanimirov <b.stanimirov@abv.bg>
@@ -128,6 +137,7 @@ CentricStorm <CentricStorm@users.noreply.github.com>
 Chad Brewbaker <crb002@gmail.com>
 Changyeon Kim <cyzero.kim@samsung.com>
 Chao Jiang <jc19chaoj@zoho.com>
+Charles Duffy <charles@dyfis.net>
 Charles Xu <63788048+chaxu01@users.noreply.github.com>
 Charles Xu <charles.xu@arm.com>
 Chen Xi <xi2.chen@intel.com>
@@ -139,12 +149,14 @@ Chris Kuehl <ckuehl@ckuehl.me>
 Christian Demsar <christian@github.email.demsar.us>
 Christian Demsar <crasm@git.vczf.us>
 Christian Falch <875252+chrfalch@users.noreply.github.com>
+Christian Fillion <cfillion@users.noreply.github.com>
 Christian Kastner <ckk@kvr.at>
 Christian Kögler <ck3d@gmx.de>
 Christian Köhnenkamp <cvk5@me.com>
 Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com>
 Christopher Nielsen <62156882+mascguy@users.noreply.github.com>
 Clark Saben <76020733+csaben@users.noreply.github.com>
+Clauszy <zhangyub@uniontech.com>
 Clint Herron <hanclinto@gmail.com>
 Conrad Kramer <conrad@conradkramer.com>
 Corentin REGAL <corentin.regal@gmail.com>
@@ -163,6 +175,7 @@ Daniel Hiltgen <dhiltgen@users.noreply.github.com>
 Daniel Illescas Romero <illescas.daniel@protonmail.com>
 Daniel Kleine <53251018+d-kleine@users.noreply.github.com>
 Daniele <57776841+daniandtheweb@users.noreply.github.com>
+Danny Milosavljevic <dannym@friendly-machines.com>
 DannyDaemonic <DannyDaemonic@gmail.com>
 Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
 Dave <dave-fl@users.noreply.github.com>
@@ -170,6 +183,7 @@ Dave Airlie <airlied@gmail.com>
 Dave Airlie <airlied@redhat.com>
 Dave Della Costa <ddellacosta+github@gmail.com>
 David Friehs <david@friehs.info>
+David Huang <1969802+hjc4869@users.noreply.github.com>
 David Kennedy <dakennedyd@gmail.com>
 David Pflug <david@pflug.email>
 David Renshaw <dwrenshaw@gmail.com>
@@ -236,6 +250,7 @@ Felix <stenbackfelix@gmail.com>
 Finn Voorhees <finnvoorhees@gmail.com>
 Firat <firatkiral@gmail.com>
 FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
+Florent BENOIT <fbenoit@redhat.com>
 Folko-Ven <71110216+Folko-Ven@users.noreply.github.com>
 Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com>
 Francisco Melo <43780565+francis2tm@users.noreply.github.com>
@@ -254,6 +269,7 @@ Gary Mulder <gjmulder@gmail.com>
 Gavin Zhao <gavinzhaojw@protonmail.com>
 Genkagaku.GPT <hlhr202@163.com>
 Georgi Gerganov <ggerganov@gmail.com>
+Gian-Carlo Pascutto <gcp@sjeng.org>
 Gilad S <giladgd@users.noreply.github.com>
 Gilad S. <7817232+giladgd@users.noreply.github.com>
 Giuseppe Scrivano <giuseppe@scrivano.org>
@@ -267,7 +283,9 @@ Guspan Tanadi <36249910+guspan-tanadi@users.noreply.github.com>
 Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
 Haggai Nuchi <h.nuchi@gmail.com>
 Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
+Hale Chan <halechan@qq.com>
 Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com>
+Han Yin <han.yin@arm.com>
 HanishKVC <hanishkvc@gmail.com>
 Haohui Mai <ricetons@gmail.com>
 Haoxiang Fei <tonyfettes@tonyfettes.com>
@@ -278,6 +296,7 @@ Haus1 <haus.xda@gmail.com>
 Henk Poley <HenkPoley@gmail.com>
 Henri Vasserman <henv@hot.ee>
 Henrik Forstén <henrik.forsten@gmail.com>
+Henry Linjamäki <henry.linjamaki@gmail.com>
 Herman Semenov <GermanAizek@yandex.ru>
 Hesen Peng <hesen.peng@gmail.com>
 HimariO <dsfhe49854@gmail.com>
@@ -307,6 +326,7 @@ Ivan <nekotekina@gmail.com>
 Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
 Ivan Komarov <Ivan.Komarov@dfyz.info>
 Ivan Stepanov <ivanstepanovftw@gmail.com>
+JC <43374599+MrSMlT@users.noreply.github.com>
 JFLFY2255 <JFLFY2255@163.com>
 JH23X <165871467+JH23X@users.noreply.github.com>
 Jack Mousseau <jack@software.inc>
@@ -325,6 +345,7 @@ Jan Ploski <jpl@plosquare.com>
 Jannis Schönleber <joennlae@gmail.com>
 Jared Van Bortel <cebtenzzre@gmail.com>
 Jared Van Bortel <jared@nomic.ai>
+Jason C.H <ctrysbita@outlook.com>
 Jason McCartney <jmac@theroot.org>
 Jason Stillerman <jason.t.stillerman@gmail.com>
 Jean-Christophe Hoelt <hoelt@fovea.cc>
@@ -342,6 +363,7 @@ Jiahao Li <liplus17@163.com>
 Jian Liao <jianliao@users.noreply.github.com>
 JidongZhang-THU <1119708529@qq.com>
 Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
+Jinyang He <hejinyang@loongson.cn>
 Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
 Jiří Sejkora <Sejseloid@gmail.com>
 Joan Fontanals <jfontanalsmartinez@gmail.com>
@@ -379,6 +401,7 @@ Justine Tunney <jtunney@mozilla.com>
 Juuso Alasuutari <juuso.alasuutari@gmail.com>
 KASR <karim.asrih@gmail.com>
 Kamil Tomšík <info@tomsik.cz>
+Kante Yin <kerthcet@gmail.com>
 Karol Kontny <82021046+kkontny@users.noreply.github.com>
 Karsten Weiss <knweiss@gmail.com>
 Karthick <j.karthic2004@gmail.com>
@@ -419,6 +442,7 @@ LoganDark <github@logandark.mozmail.com>
 Loïc Carrère <loic.carrere@gmail.com>
 LostRuins <39025047+LostRuins@users.noreply.github.com>
 LostRuins Concedo <39025047+LostRuins@users.noreply.github.com>
+Lucas Moura Belo <lucas.belo@live.com>
 Luciano <lucianostrika44@gmail.com>
 Luo Tian <lt@basecity.com>
 Lyle Dean <dean@lyle.dev>
@@ -463,6 +487,7 @@ Matthew Tejo <matthew.tejo@gmail.com>
 Matvey Soloviev <blackhole89@gmail.com>
 Max Krasnyansky <max.krasnyansky@gmail.com>
 Max Krasnyansky <quic_maxk@quicinc.com>
+Maxim Evtush <154841002+maximevtush@users.noreply.github.com>
 Maxime <672982+maximegmd@users.noreply.github.com>
 Maximilian Winter <maximilian.winter.91@gmail.com>
 Meng Zhang <meng@tabbyml.com>
@@ -494,6 +519,7 @@ Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
 Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
 Mohammadreza Hendiani <mohammad.r.hendiani@gmail.com>
 Molly Sophia <mollysophia379@gmail.com>
+MoonRide303 <130458190+MoonRide303@users.noreply.github.com>
 MorganRO8 <47795945+MorganRO8@users.noreply.github.com>
 Murilo Santana <mvrilo@gmail.com>
 Musab Gultekin <musabgultekin@users.noreply.github.com>
@@ -524,6 +550,7 @@ Nikolas <127742645+nneubacher@users.noreply.github.com>
 Nindaleth <Nindaleth@users.noreply.github.com>
 Nuno <rare-magma@posteo.eu>
 OSecret <135510162+OLSecret@users.noreply.github.com>
+Oleksandr Kuvshynov <661042+okuvshynov@users.noreply.github.com>
 Oleksandr Nikitin <oleksandr@tvori.info>
 Oleksii Maryshchenko <oleksii.maryshchenko@gmail.com>
 Olivier Chafik <ochafik@users.noreply.github.com>
@@ -533,6 +560,7 @@ PAB <pierreantoine.bannier@gmail.com>
 Pablo Duboue <pablo.duboue@gmail.com>
 Pascal Patry <ppatry@mtacitlabs.com>
 Patrice Ferlet <metal3d@gmail.com>
+Patrick Peng <retr0@retr0.blog>
 Paul Tsochantaris <ptsochantaris@icloud.com>
 Pavel Zloi <github.com@drteam.rocks>
 Pavol Rusnak <pavol@rusnak.io>
@@ -549,6 +577,7 @@ Pieter Ouwerkerk <pieter.ouwerkerk@gmail.com>
 Plamen Minev <pacominev@gmail.com>
 Prashant Vithule <119530321+Vithulep@users.noreply.github.com>
 Przemysław Pawełczyk <przemoc@gmail.com>
+PureJourney <edward.pong@qq.com>
 Qin Yue Chen <71813199+chenqiny@users.noreply.github.com>
 Qingyou Meng <meng.qingyou@gmail.com>
 Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
@@ -564,14 +593,17 @@ Rand Xie <randxiexyy29@gmail.com>
 Randall Fitzgerald <randall@dasaku.net>
 Random Fly <renfei8@live.cn>
 Reinforce-II <fate@eastal.com>
+Rémy O <remyoudompheng@gmail.com>
 Rémy Oudompheng <oudomphe@phare.normalesup.org>
 Ren Xuancheng <jklj077@users.noreply.github.com>
 Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com>
 Reza Kakhki <rezakakhki.de@gmail.com>
+Reza Rahemtola <49811529+RezaRahemtola@users.noreply.github.com>
 RhinoDevel <RhinoDevel@users.noreply.github.com>
 Riccardo Orlando <Riccorl@users.noreply.github.com>
 Riceball LEE <snowyu.lee@gmail.com>
 Rich Dougherty <rich@rd.nz>
+Richard <r-burton@hotmail.co.uk>
 Richard Kiss <him@richardkiss.com>
 Richard Roberson <richardr1126@gmail.com>
 Rick G <26732651+TheFlipbook@users.noreply.github.com>
@@ -588,6 +620,7 @@ Robert Sung-wook Shin <edp1096@users.noreply.github.com>
 Robey Holderith <robey@flaminglunchbox.net>
 Robyn <robyngraf@users.noreply.github.com>
 Roger Meier <r.meier@siemens.com>
+Rohanjames1997 <rohan.james4@gmail.com>
 Roland <14355895+rbur0425@users.noreply.github.com>
 Romain Biessy <romain.biessy@codeplay.com>
 Romain D <90720+Artefact2@users.noreply.github.com>
@@ -610,6 +643,7 @@ Ryan Landay <rlanday@gmail.com>
 Ryder Wishart <ryderwishart@gmail.com>
 Ryuei <louixs@users.noreply.github.com>
 Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
+SAMI <samuel.koesnadi@stud.uni-due.de>
 SRHMorris <69468379+SRHMorris@users.noreply.github.com>
 SXX <sxx1136965276@gmail.com>
 SakuraUmi <yukinon244@gmail.com>
@@ -634,6 +668,8 @@ Shane A <shanea@allenai.org>
 Shangning Xu <32517059+xushangning@users.noreply.github.com>
 Shankar <gshankar.87@gmail.com>
 Shanshan Shen <467638484@qq.com>
+Shelby Jenkins <47464908+ShelbyJenkins@users.noreply.github.com>
+Sheldon Robinson <sheldon.robinson@live.com>
 Shijie <821898965@qq.com>
 Shintarou Okada <kokuzen@gmail.com>
 Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
@@ -713,18 +749,24 @@ Victor Nogueira <felladrin@gmail.com>
 Victor Z. Peng <ziliangdotme@gmail.com>
 Viet-Anh NGUYEN (Andrew) <vietanh.dev@gmail.com>
 Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com>
+Vitali Lovich <vlovich+github@gmail.com>
+Vivian <vynride@gmail.com>
 Vlad <spitfireage@gmail.com>
 Vladimir <bogdad@gmail.com>
 Vladimir Malyutin <first-leon@yandex.ru>
+Vladimir Vuksanovic <109677816+vvuksanovic@users.noreply.github.com>
 Vladimir Zorin <vladimir@deviant.guru>
 VoidIsVoid <343750470@qq.com>
 Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com>
+Wagner Bruna <wbruna@users.noreply.github.com>
 Wang Qin <37098874+wangqin0@users.noreply.github.com>
 Wang Ran (汪然) <wangr@smail.nju.edu.cn>
 WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com>
 Weird Constructor <weirdconstructor@gmail.com>
+Weizhao Ouyang <o451686892@gmail.com>
 Welby Seely <welbyseely@gmail.com>
 Wentai Zhang <rchardx@gmail.com>
+Wilken Gottwalt <12194808+wgottwalt@users.noreply.github.com>
 WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com>
 William Tambellini <william.tambellini@gmail.com>
 William Tambellini <wtambellini@sdl.com>
@@ -816,6 +858,8 @@ chaihahaha <chai836275709@gmail.com>
 chiranko <96988916+chiranko@users.noreply.github.com>
 clibdev <52199778+clibdev@users.noreply.github.com>
 clyang <clyang@clyang.net>
+cmdr2 <secondary.cmdr2@gmail.com>
+cmdr2 <shashank.shekhar.global@gmail.com>
 cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
 codezjx <code.zjx@gmail.com>
 coezbek <c.oezbek@gmail.com>
@@ -835,6 +879,7 @@ deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
 devojony <61173062+devojony@users.noreply.github.com>
 ditsuke <ditsuke@protonmail.com>
 divinity76 <divinity76@gmail.com>
+dm4 <dm4@secondstate.io>
 dm4 <sunrisedm4@gmail.com>
 dotpy314 <33351922+dotpy314@users.noreply.github.com>
 drbh <david.richard.holtz@gmail.com>
@@ -849,6 +894,7 @@ fairydreaming <166155368+fairydreaming@users.noreply.github.com>
 fengerhu1 <2748250768@qq.com>
 fj-y-saito <85871716+fj-y-saito@users.noreply.github.com>
 fraxy-v <65565042+fraxy-v@users.noreply.github.com>
+fxzjshm <11426482+fxzjshm@users.noreply.github.com>
 github-actions[bot] <github-actions[bot]@users.noreply.github.com>
 gliptic <gliptic@users.noreply.github.com>
 gn64 <yukikaze.jp@gmail.com>
@@ -873,6 +919,7 @@ hydai <z54981220@gmail.com>
 iSma <ismail.senhaji@gmail.com>
 iacore <74560659+iacore@users.noreply.github.com>
 icppWorld <124377669+icppWorld@users.noreply.github.com>
+igardev <49397134+igardev@users.noreply.github.com>
 igarnier <igarnier@protonmail.com>
 intelmatt <61025942+intelmatt@users.noreply.github.com>
 iohub <rickyang.pro@gmail.com>
@@ -880,6 +927,7 @@ issixx <46835150+issixx@users.noreply.github.com>
 jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
 jaime-m-p <167997752+jaime-m-p@users.noreply.github.com>
 jameswu2014 <545426914@qq.com>
+jason_w <jason.wang@126.com>
 jdomke <28772296+jdomke@users.noreply.github.com>
 jiahao su <damow890@gmail.com>
 jiez <373447296@qq.com>
@@ -891,6 +939,7 @@ jon-chuang <9093549+jon-chuang@users.noreply.github.com>
 jp-x-g <jpxg-dev@protonmail.com>
 jukofyork <69222624+jukofyork@users.noreply.github.com>
 junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
+junchao-zhao <68935141+junchao-loongson@users.noreply.github.com>
 jwj7140 <32943891+jwj7140@users.noreply.github.com>
 k.h.lai <adrian.k.h.lai@outlook.com>
 kaizau <kaizau@users.noreply.github.com>
@@ -925,6 +974,7 @@ ltoniazzi <61414566+ltoniazzi@users.noreply.github.com>
 luoyu-intel <yu.luo@intel.com>
 m3ndax <adrian.goessl@outlook.com>
 maddes8cht <55592906+maddes8cht@users.noreply.github.com>
+magicse <magicse@users.noreply.github.com>
 mahorozte <41834471+mahorozte@users.noreply.github.com>
 makomk <makosoft@googlemail.com>
 manikbhandari <mbbhandarimanik2@gmail.com>
@@ -935,6 +985,7 @@ matt23654 <matthew.webber@protonmail.com>
 matteo <matteogeniaccio@yahoo.it>
 mdrokz <mohammadmunshi@gmail.com>
 mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
+midnight <midnightmagic@users.noreply.github.com>
 minarchist <minarchist@users.noreply.github.com>
 mj-shifu <77107165+mj-shifu@users.noreply.github.com>
 mmyjona <jonathan.gonse@gmail.com>
@@ -958,10 +1009,12 @@ omahs <73983677+omahs@users.noreply.github.com>
 oobabooga <112222186+oobabooga@users.noreply.github.com>
 opparco <parco.opaai@gmail.com>
 ostix360 <55257054+ostix360@users.noreply.github.com>
+pascal-lc <49066376+pascal-lc@users.noreply.github.com>
 pculliton <phillipculliton@gmail.com>
 peidaqi <peidaqi@gmail.com>
 pengxin99 <pengxin.yuan@intel.com>
 perserk <perserk@gmail.com>
+petterreinholdtsen <pere-github@hungry.com>
 piDack <104877312+piDack@users.noreply.github.com>
 pmysl <piotr.myslinski@outlook.com>
 postmasters <namnguyen@google.com>
@@ -983,6 +1036,7 @@ semidark <me@semidark.net>
 serhii-nakon <57632032+serhii-nakon@users.noreply.github.com>
 sharpHL <132747147+sharpHL@users.noreply.github.com>
 shibe2 <shibe@tuta.io>
+simon886212 <37953122+simon886212@users.noreply.github.com>
 singularity <12184989+singularity-s0@users.noreply.github.com>
 sjinzh <sjinzh@gmail.com>
 sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com>
@@ -1000,10 +1054,12 @@ tarcey <cey.tarik@gmail.com>
 tc-mb <157115220+tc-mb@users.noreply.github.com>
 texmex76 <40733439+texmex76@users.noreply.github.com>
 thement <40525767+thement@users.noreply.github.com>
+theraininsky <76763719+theraininsky@users.noreply.github.com>
 thewh1teagle <61390950+thewh1teagle@users.noreply.github.com>
 tjohnman <tjohnman@users.noreply.github.com>
 toyer <2042519524@qq.com>
 tslmy <tslmy@users.noreply.github.com>
+tv1wnd <55383215+tv1wnd@users.noreply.github.com>
 ubik2 <ubik2@users.noreply.github.com>
 uint256_t <konndennsa@gmail.com>
 uint256_t <maekawatoshiki1017@gmail.com>
@@ -1014,6 +1070,7 @@ valiray <133289098+valiray@users.noreply.github.com>
 vb <vaibhavs10@gmail.com>
 vik <vikhyatk@gmail.com>
 viric <viric@viric.name>
+vmobilis <75476228+vmobilis@users.noreply.github.com>
 vodkaslime <646329483@qq.com>
 vvhg1 <94630311+vvhg1@users.noreply.github.com>
 vxiiduu <73044267+vxiiduu@users.noreply.github.com>
@@ -1028,6 +1085,8 @@ wzy <32936898+Freed-Wu@users.noreply.github.com>
 xaedes <xaedes@gmail.com>
 xaedes <xaedes@googlemail.com>
 xctan <axunlei@gmail.com>
+xiaobing318 <71554036+xiaobing318@users.noreply.github.com>
+xiaofei <hbuxiaofei@gmail.com>
 xloem <0xloem@gmail.com>
 yangli2 <yangli2@gmail.com>
 ymcki <84055651+ymcki@users.noreply.github.com>

CMakeLists.txt

@@ -29,6 +29,8 @@ else()
     set(LLAMA_STANDALONE OFF)
 endif()
 
+option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
+
 if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
@@ -145,7 +147,13 @@ endif()
 # 3rd-party
 #
 
-if (NOT TARGET ggml)
+if (LLAMA_USE_SYSTEM_GGML)
+    message(STATUS "Using system-provided libggml, skipping ggml build")
+    find_package(ggml REQUIRED)
+    add_library(ggml ALIAS ggml::ggml)
+endif()
+
+if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()

CONTRIBUTING.md

@@ -39,7 +39,7 @@
 
     _(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline.)_
 
-- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code
+- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` (from clang-tools v15+) to format the added code
 - For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines)
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
 - Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggml-org/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$

Makefile

@@ -836,7 +836,7 @@ ifdef GGML_MUSA
 	else
 		MUSA_PATH ?= /opt/musa
 	endif
-	MUSA_ARCHITECTURES ?= 21;22
+	MUSA_ARCHITECTURES ?= 21;22;31
 
 	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
 	MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib

Package.swift

@@ -1,19 +0,0 @@
-// swift-tools-version:5.5
-
-import PackageDescription
-
-let package = Package(
-    name: "llama",
-    platforms: [
-        .macOS(.v12),
-        .iOS(.v14),
-        .watchOS(.v4),
-        .tvOS(.v14)
-    ],
-    products: [
-        .library(name: "llama", targets: ["llama"]),
-    ],
-    targets: [
-        .systemLibrary(name: "llama", pkgConfig: "llama"),
-    ]
-)

README.md

@@ -5,7 +5,7 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggml-org/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
@@ -25,7 +25,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 - **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
 - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
-- Universal tool call support in `llama-server`: https://github.com/ggml-org/llama.cpp/pull/9639
+- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -112,6 +112,8 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
+- [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
+- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 
 #### Multimodal
 
@@ -157,6 +159,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
+- Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
 
 </details>
 
@@ -171,6 +174,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0)
 - [janhq/jan](https://github.com/janhq/jan) (AGPL)
+- [johnbean393/Sidekick](https://github.com/johnbean393/Sidekick) (MIT)
 - [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0)
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
@@ -526,6 +530,35 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
+## XCFramework
+The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
+and macOS. It can be used in Swift projects without the need to compile the
+library from source. For example:
+```swift
+// swift-tools-version: 5.10
+// The swift-tools-version declares the minimum version of Swift required to build this package.
+
+import PackageDescription
+
+let package = Package(
+    name: "MyLlamaPackage",
+    targets: [
+        .executableTarget(
+            name: "MyLlamaPackage",
+            dependencies: [
+                "LlamaFramework"
+            ]),
+        .binaryTarget(
+            name: "LlamaFramework",
+            url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
+            checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
+        )
+    ]
+)
+```
+The above example is using an intermediate build `b5046` of the library. This can be modified
+to use a different version by changing the URL and checksum.
+
 ## Completions
 Command-line completion is available for some environments.
 

Sources/llama/llama.h

@@ -1,4 +0,0 @@
-#pragma once
-
-#include <llama.h>
-

Sources/llama/module.modulemap

@@ -1,5 +0,0 @@
-module llama [system] {
-    header "llama.h"
-    link "llama"
-    export *
-}

build-xcframework.sh

@@ -0,0 +1,519 @@
+#!/bin/bash
+#
+# Options
+IOS_MIN_OS_VERSION=16.4
+MACOS_MIN_OS_VERSION=13.3
+VISIONOS_MIN_OS_VERSION=1.0
+TVOS_MIN_OS_VERSION=16.4
+
+BUILD_SHARED_LIBS=OFF
+LLAMA_BUILD_EXAMPLES=OFF
+LLAMA_BUILD_TESTS=OFF
+LLAMA_BUILD_SERVER=OFF
+GGML_METAL=ON
+GGML_METAL_EMBED_LIBRARY=ON
+GGML_BLAS_DEFAULT=ON
+GGML_METAL_USE_BF16=ON
+GGML_OPENMP=OFF
+
+COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
+COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
+
+# Common options for all builds
+COMMON_CMAKE_ARGS=(
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_REQUIRED=NO
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGN_IDENTITY=""
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_ALLOWED=NO
+    -DCMAKE_XCODE_ATTRIBUTE_DEBUG_INFORMATION_FORMAT="dwarf-with-dsym"
+    -DCMAKE_XCODE_ATTRIBUTE_GCC_GENERATE_DEBUGGING_SYMBOLS=YES
+    -DCMAKE_XCODE_ATTRIBUTE_COPY_PHASE_STRIP=NO
+    -DCMAKE_XCODE_ATTRIBUTE_STRIP_INSTALLED_PRODUCT=NO
+    -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+    -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
+    -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
+    -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
+    -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
+    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
+    -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
+    -DGGML_METAL=${GGML_METAL}
+    -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16}
+    -DGGML_NATIVE=OFF
+    -DGGML_OPENMP=${GGML_OPENMP}
+)
+
+check_required_tool() {
+    local tool=$1
+    local install_message=$2
+
+    if ! command -v $tool &> /dev/null; then
+        echo "Error: $tool is required but not found."
+        echo "$install_message"
+        exit 1
+    fi
+}
+echo "Checking for required tools..."
+check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
+check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
+check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+
+set -e
+
+## Clean up previous builds
+rm -rf build-apple
+rm -rf build-ios-sim
+rm -rf build-ios-device
+rm -rf build-macos
+rm -rf build-visionos
+rm -rf build-visionos-sim
+rm -rf build-tvos-sim
+rm -rf build-tvos-device
+
+# Setup the xcframework build directory structure
+setup_framework_structure() {
+    local build_dir=$1
+    local min_os_version=$2
+    local platform=$3  # "ios", "macos", "visionos", or "tvos"
+    local framework_name="llama"
+
+    echo "Creating ${platform}-style framework structure for ${build_dir}"
+
+    if [[ "$platform" == "macos" ]]; then
+        # macOS versioned structure uses versioned directories
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Headers
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Modules
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources
+
+        # Create symbolic links
+        ln -sf A ${build_dir}/framework/${framework_name}.framework/Versions/Current
+        ln -sf Versions/Current/Headers ${build_dir}/framework/${framework_name}.framework/Headers
+        ln -sf Versions/Current/Modules ${build_dir}/framework/${framework_name}.framework/Modules
+        ln -sf Versions/Current/Resources ${build_dir}/framework/${framework_name}.framework/Resources
+        ln -sf Versions/Current/${framework_name} ${build_dir}/framework/${framework_name}.framework/${framework_name}
+
+        # Set header and module paths
+        local header_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Headers/
+        local module_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Modules/
+    else
+        # iOS/VisionOS/tvOS use a flat structure
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Headers
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Modules
+
+        # Remove any existing structure to ensure clean build
+        rm -rf ${build_dir}/framework/${framework_name}.framework/Versions
+
+        # Set header and module paths
+        local header_path=${build_dir}/framework/${framework_name}.framework/Headers/
+        local module_path=${build_dir}/framework/${framework_name}.framework/Modules/
+    fi
+
+    # Copy all required headers (common for all platforms)
+    cp include/llama.h             ${header_path}
+    cp ggml/include/ggml.h         ${header_path}
+    cp ggml/include/ggml-alloc.h   ${header_path}
+    cp ggml/include/ggml-backend.h ${header_path}
+    cp ggml/include/ggml-metal.h   ${header_path}
+    cp ggml/include/ggml-cpu.h     ${header_path}
+    cp ggml/include/ggml-blas.h    ${header_path}
+    cp ggml/include/gguf.h         ${header_path}
+
+    # Create module map (common for all platforms)
+    cat > ${module_path}module.modulemap << EOF
+framework module llama {
+    header "llama.h"
+    header "ggml.h"
+    header "ggml-alloc.h"
+    header "ggml-backend.h"
+    header "ggml-metal.h"
+    header "ggml-cpu.h"
+    header "ggml-blas.h"
+    header "gguf.h"
+
+    link "c++"
+    link framework "Accelerate"
+    link framework "Metal"
+    link framework "Foundation"
+
+    export *
+}
+EOF
+
+    # Platform-specific settings for Info.plist
+    local platform_name=""
+    local sdk_name=""
+    local supported_platform=""
+
+    case "$platform" in
+        "ios")
+            platform_name="iphoneos"
+            sdk_name="iphoneos${min_os_version}"
+            supported_platform="iPhoneOS"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family='    <key>UIDeviceFamily</key>
+    <array>
+        <integer>1</integer>
+        <integer>2</integer>
+    </array>'
+            ;;
+        "macos")
+            platform_name="macosx"
+            sdk_name="macosx${min_os_version}"
+            supported_platform="MacOSX"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/Info.plist"
+            local device_family=""
+            ;;
+        "visionos")
+            platform_name="xros"
+            sdk_name="xros${min_os_version}"
+            supported_platform="XRPlatform"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family=""
+            ;;
+        "tvos")
+            platform_name="appletvos"
+            sdk_name="appletvos${min_os_version}"
+            supported_platform="AppleTVOS"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family='    <key>UIDeviceFamily</key>
+    <array>
+        <integer>3</integer>
+    </array>'
+            ;;
+    esac
+
+    # Create Info.plist
+    cat > ${plist_path} << EOF
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>CFBundleDevelopmentRegion</key>
+    <string>en</string>
+    <key>CFBundleExecutable</key>
+    <string>llama</string>
+    <key>CFBundleIdentifier</key>
+    <string>org.ggml.llama</string>
+    <key>CFBundleInfoDictionaryVersion</key>
+    <string>6.0</string>
+    <key>CFBundleName</key>
+    <string>llama</string>
+    <key>CFBundlePackageType</key>
+    <string>FMWK</string>
+    <key>CFBundleShortVersionString</key>
+    <string>1.0</string>
+    <key>CFBundleVersion</key>
+    <string>1</string>
+    <key>MinimumOSVersion</key>
+    <string>${min_os_version}</string>
+    <key>CFBundleSupportedPlatforms</key>
+    <array>
+        <string>${supported_platform}</string>
+    </array>${device_family}
+    <key>DTPlatformName</key>
+    <string>${platform_name}</string>
+    <key>DTSDKName</key>
+    <string>${sdk_name}</string>
+</dict>
+</plist>
+EOF
+}
+
+# Create dynamic libraries from static libraries.
+combine_static_libraries() {
+    local build_dir="$1"
+    local release_dir="$2"
+    local platform="$3"  # "ios", "macos", "visionos", or "tvos"
+    local is_simulator="$4"
+    local base_dir="$(pwd)"
+    local framework_name="llama"
+
+    # Determine output path based on platform
+    local output_lib=""
+    if [[ "$platform" == "macos" ]]; then
+        # macOS uses versioned structure
+        output_lib="${build_dir}/framework/${framework_name}.framework/Versions/A/${framework_name}"
+    else
+        # iOS, visionOS, and tvOS use a directory flat structure
+        output_lib="${build_dir}/framework/${framework_name}.framework/${framework_name}"
+    fi
+
+    local libs=(
+        "${base_dir}/${build_dir}/src/${release_dir}/libllama.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-base.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a"
+        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
+        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
+    )
+
+    # Create temporary directory for processing
+    local temp_dir="${base_dir}/${build_dir}/temp"
+    mkdir -p "${temp_dir}"
+
+    # Since we have multiple architectures libtool will find object files that do not
+    # match the target architecture. We suppress these warnings.
+    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+
+    # Determine SDK, architectures, and install_name based on platform and simulator flag.
+    local sdk=""
+    local archs=""
+    local min_version_flag=""
+    local install_name=""
+
+    case "$platform" in
+        "ios")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="iphonesimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mios-simulator-version-min=${IOS_MIN_OS_VERSION}"
+            else
+                sdk="iphoneos"
+                archs="arm64"
+                min_version_flag="-mios-version-min=${IOS_MIN_OS_VERSION}"
+            fi
+            install_name="@rpath/llama.framework/llama"
+            ;;
+        "macos")
+            sdk="macosx"
+            archs="arm64 x86_64"
+            min_version_flag="-mmacosx-version-min=${MACOS_MIN_OS_VERSION}"
+            install_name="@rpath/llama.framework/Versions/Current/llama"
+            ;;
+        "visionos")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="xrsimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}-simulator"
+            else
+                sdk="xros"
+                archs="arm64"
+                min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}"
+            fi
+            # Use flat structure for visionOS, same as iOS
+            install_name="@rpath/llama.framework/llama"
+            ;;
+        "tvos")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="appletvsimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mtvos-simulator-version-min=${TVOS_MIN_OS_VERSION}"
+            else
+                sdk="appletvos"
+                archs="arm64"
+                min_version_flag="-mtvos-version-min=${TVOS_MIN_OS_VERSION}"
+            fi
+            install_name="@rpath/llama.framework/llama"
+            ;;
+    esac
+
+    # Build architecture flags
+    local arch_flags=""
+    for arch in $archs; do
+        arch_flags+=" -arch $arch"
+    done
+
+    # Create dynamic library
+    echo "Creating dynamic library for ${platform}."
+    xcrun -sdk $sdk clang++ -dynamiclib \
+        -isysroot $(xcrun --sdk $sdk --show-sdk-path) \
+        $arch_flags \
+        $min_version_flag \
+        -Wl,-force_load,"${temp_dir}/combined.a" \
+        -framework Foundation -framework Metal -framework Accelerate \
+        -install_name "$install_name" \
+        -o "${base_dir}/${output_lib}"
+
+    # Platform-specific post-processing for device builds
+    if [[ "$is_simulator" == "false" ]]; then
+        if command -v vtool &>/dev/null; then
+            case "$platform" in
+                "ios")
+                    echo "Marking binary as a framework binary for iOS..."
+                    vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+                "visionos")
+                    echo "Marking binary as a framework binary for visionOS..."
+                    vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+                "tvos")
+                    echo "Marking binary as a framework binary for tvOS..."
+                    vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+            esac
+        else
+            echo "Warning: vtool not found. Binary may not pass App Store validation."
+        fi
+    fi
+
+    echo "Creating properly formatted dSYM..."
+    # Create a separate directory for dSYMs for all platforms
+    mkdir -p "${base_dir}/${build_dir}/dSYMs"
+
+    # iOS and visionOS style dSYM (flat structure)
+    if [[ "$platform" == "ios" || "$platform" == "visionos" || "$platform" == "tvos" ]]; then
+        # Generate dSYM in the dSYMs directory
+        xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/llama.dSYM"
+
+        # Create a copy of the binary that will be stripped
+        cp "${base_dir}/${output_lib}" "${temp_dir}/binary_to_strip"
+
+        # Strip debug symbols from the copy
+        xcrun strip -S "${temp_dir}/binary_to_strip" -o "${temp_dir}/stripped_lib"
+
+        # Replace the original with the stripped version
+        mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}"
+    else
+        # macOS style dSYM
+        # First strip debug info to a separate file
+        xcrun strip -S "${base_dir}/${output_lib}" -o "${temp_dir}/stripped_lib"
+
+        # Generate dSYM in the dSYMs directory
+        xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/llama.dSYM"
+
+        # Replace original binary with stripped version
+        mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}"
+    fi
+
+    # Remove any automatically generated dSYM files in the framework structure as they will
+    # otherwise case Invalid Bundle Structure validation errors.
+    if [ -d "${base_dir}/${output_lib}.dSYM" ]; then
+        echo "Removing generated dSYM file in framework structure: ${base_dir}/${output_lib}.dSYM"
+        rm -rf "${base_dir}/${output_lib}.dSYM"
+    fi
+
+    # Clean up
+    rm -rf "${temp_dir}"
+}
+
+echo "Building for iOS simulator..."
+cmake -B build-ios-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DIOS=ON \
+    -DCMAKE_SYSTEM_NAME=iOS \
+    -DCMAKE_OSX_SYSROOT=iphonesimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-ios-sim --config Release -- -quiet
+
+echo "Building for iOS devices..."
+cmake -B build-ios-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_SYSROOT=iphoneos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-ios-device --config Release -- -quiet
+
+echo "Building for macOS..."
+cmake -B build-macos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-macos --config Release -- -quiet
+
+echo "Building for visionOS..."
+cmake -B build-visionos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xros \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-visionos --config Release -- -quiet
+
+echo "Building for visionOS simulator..."
+cmake -B build-visionos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xrsimulator \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-visionos-sim --config Release -- -quiet
+
+# Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
+echo "Building for tvOS simulator..."
+cmake -B build-tvos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvsimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-tvos-sim --config Release -- -quiet
+
+echo "Building for tvOS devices..."
+cmake -B build-tvos-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-tvos-device --config Release -- -quiet
+
+# Setup frameworks and copy binaries and headers
+echo "Setting up framework structures..."
+setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos"
+setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos"
+setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos"
+
+# Create dynamic libraries from static libraries
+echo "Creating dynamic libraries from static libraries..."
+combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true"
+combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false"
+combine_static_libraries "build-macos" "Release" "macos" "false"
+combine_static_libraries "build-visionos" "Release-xros" "visionos" "false"
+combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true"
+combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true"
+combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
+
+# Create XCFramework with correct debug symbols paths
+echo "Creating XCFramework..."
+xcodebuild -create-xcframework \
+    -framework $(pwd)/build-ios-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-ios-device/framework/llama.framework \
+    -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-macos/framework/llama.framework \
+    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
+    -framework $(pwd)/build-visionos/framework/llama.framework \
+    -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-visionos-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-visionos-sim/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-tvos-device/framework/llama.framework \
+    -debug-symbols $(pwd)/build-tvos-device/dSYMs/llama.dSYM \
+    -framework $(pwd)/build-tvos-sim/framework/llama.framework \
+    -debug-symbols $(pwd)/build-tvos-sim/dSYMs/llama.dSYM \
+    -output $(pwd)/build-apple/llama.xcframework

ci/README.md

@@ -26,4 +26,43 @@ GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 # with SYCL support
 source /opt/intel/oneapi/setvars.sh
 GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with MUSA support
+GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```
+
+## Running MUSA CI in a Docker Container
+
+Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
+
+### 1. Create a local directory to store cached models, configuration files and venv:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-cache
+```
+
+### 2. Create a local directory to store CI run results:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-results
+```
+
+### 3. Start a Docker container and run the CI:
+
+```bash
+docker run --privileged -it \
+    -v $HOME/llama.cpp/ci-cache:/ci-cache \
+    -v $HOME/llama.cpp/ci-results:/ci-results \
+    -v $PWD:/ws -w /ws \
+    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+```
+
+Inside the container, execute the following commands:
+
+```bash
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
+git config --global --add safe.directory /ws
+GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
+```
+
+This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

ci/run.sh

@@ -16,6 +16,9 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with MUSA support
+# GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -52,13 +55,24 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
-
+    # Use only main GPU
+    export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+    # Enable sysman for correct memory reporting
+    export ZES_ENABLE_SYSMAN=1
+    # to circumvent precision issues on CPY operations
+    export SYCL_PROGRAM_COMPILE_OPTIONS="-cl-fp32-correctly-rounded-divide-sqrt"
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
+
+if [ ! -z ${GG_BUILD_MUSA} ]; then
+    # Use qy1 by default (MTT S80)
+    MUSA_ARCH=${MUSA_ARCH:-21}
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
+fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -352,10 +366,10 @@ function gg_run_open_llama_7b_v2 {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -808,7 +822,7 @@ export LLAMA_LOG_PREFIX=1
 export LLAMA_LOG_TIMESTAMPS=1
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
     rm -rf ${SRC}/models-mnt
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
@@ -826,8 +840,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-
-test $ret -eq 0 && gg_run ctest_debug
+if [ -z ${GG_BUILD_SYCL} ]; then
+    # SYCL build breaks with debug build flags
+    test $ret -eq 0 && gg_run ctest_debug
+fi
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -835,7 +851,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        test $ret -eq 0 && gg_run test_scripts_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run test_scripts_debug
+        fi
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -846,7 +864,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
             test $ret -eq 0 && gg_run pythia_2_8b
             #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        test $ret -eq 0 && gg_run ctest_with_model_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run ctest_with_model_debug
+        fi
         test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi

cmake/common.cmake

@@ -1,3 +1,5 @@
+include("ggml/cmake/common.cmake")
+
 function(llama_add_compile_flags)
     if (LLAMA_FATAL_WARNINGS)
         if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")

common/CMakeLists.txt

@@ -114,8 +114,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.6.12:
-        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
+        # v0.7.10:
+        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/arg.cpp

@@ -1,9 +1,20 @@
+#include "gguf.h" // for reading GGUF splits
 #include "arg.h"
 
+#include "common.h"
 #include "log.h"
 #include "sampling.h"
 #include "chat.h"
 
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 #include <algorithm>
 #include <climits>
 #include <cstdarg>
@@ -14,6 +25,14 @@
 #include <thread>
 #include <vector>
 
+//#define LLAMA_USE_CURL
+
+#if defined(LLAMA_USE_CURL)
+#include <curl/curl.h>
+#include <curl/easy.h>
+#include <future>
+#endif
+
 #include "json-schema-to-grammar.h"
 
 using json = nlohmann::ordered_json;
@@ -125,47 +144,549 @@ std::string common_arg::to_string() {
     return ss.str();
 }
 
+//
+// downloader
+//
+
+struct common_hf_file_res {
+    std::string repo; // repo name with ":tag" removed
+    std::string ggufFile;
+    std::string mmprojFile;
+};
+
+#ifdef LLAMA_USE_CURL
+
+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
+#else
+#include <sys/syslimits.h>
+#endif
+#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
+
+//
+// CURL utils
+//
+
+using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
+
+// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
+struct curl_slist_ptr {
+    struct curl_slist * ptr = nullptr;
+    ~curl_slist_ptr() {
+        if (ptr) {
+            curl_slist_free_all(ptr);
+        }
+    }
+};
+
+#define CURL_MAX_RETRY 3
+#define CURL_RETRY_DELAY_SECONDS 2
+
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+    int remaining_attempts = max_attempts;
+
+    while (remaining_attempts > 0) {
+        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+
+        CURLcode res = curl_easy_perform(curl);
+        if (res == CURLE_OK) {
+            return true;
+        }
+
+        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
+        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
+
+        remaining_attempts--;
+        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+    }
+
+    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+
+    return false;
+}
+
+// download one single file from remote URL to local path
+static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) {
+    // Initialize libcurl
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    if (!curl) {
+        LOG_ERR("%s: error initializing libcurl\n", __func__);
+        return false;
+    }
+
+    bool force_download = false;
+
+    // Set the URL, allow to follow http redirection
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+
+    // Check if hf-token or bearer-token was specified
+    if (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    }
+
+#if defined(_WIN32)
+    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
+    //   operating system. Currently implemented under MS-Windows.
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+
+    // Check if the file already exists locally
+    auto file_exists = std::filesystem::exists(path);
+
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    nlohmann::json metadata;
+    std::string etag;
+    std::string last_modified;
+
+    if (file_exists) {
+        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+        std::ifstream metadata_in(metadata_path);
+        if (metadata_in.good()) {
+            try {
+                metadata_in >> metadata;
+                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
+                if (metadata.contains("url") && metadata.at("url").is_string()) {
+                    auto previous_url = metadata.at("url").get<std::string>();
+                    if (previous_url != url) {
+                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
+                        return false;
+                    }
+                }
+                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                    etag = metadata.at("etag");
+                }
+                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                    last_modified = metadata.at("lastModified");
+                }
+            } catch (const nlohmann::json::exception & e) {
+            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+                return false;
+            }
+        }
+    } else {
+        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
+    }
+
+    // Send a HEAD request to retrieve the etag and last-modified headers
+    struct common_load_model_from_url_headers {
+        std::string etag;
+        std::string last_modified;
+    };
+
+    common_load_model_from_url_headers headers;
+
+    {
+        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
+        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
+            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
+
+            static std::regex header_regex("([^:]+): (.*)\r\n");
+            static std::regex etag_regex("ETag", std::regex_constants::icase);
+            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
+
+            std::string header(buffer, n_items);
+            std::smatch match;
+            if (std::regex_match(header, match, header_regex)) {
+                const std::string & key = match[1];
+                const std::string & value = match[2];
+                if (std::regex_match(key, match, etag_regex)) {
+                    headers->etag = value;
+                } else if (std::regex_match(key, match, last_modified_regex)) {
+                    headers->last_modified = value;
+                }
+            }
+            return n_items;
+        };
+
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
+
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code != 200) {
+            // HEAD not supported, we don't know if the file has changed
+            // force trigger downloading
+            force_download = true;
+            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        }
+    }
+
+    bool should_download = !file_exists || force_download;
+    if (!should_download) {
+        if (!etag.empty() && etag != headers.etag) {
+            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
+            should_download = true;
+        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
+            should_download = true;
+        }
+    }
+    if (should_download) {
+        std::string path_temporary = path + ".downloadInProgress";
+        if (file_exists) {
+            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
+            if (remove(path.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+                return false;
+            }
+        }
+
+        // Set the output file
+
+        struct FILE_deleter {
+            void operator()(FILE * f) const {
+                fclose(f);
+            }
+        };
+
+        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
+        if (!outfile) {
+            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
+            return false;
+        }
+
+        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
+        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
+            return fwrite(data, size, nmemb, (FILE *)fd);
+        };
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
+
+        //  display download progress
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
+
+        // helper function to hide password in URL
+        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
+            std::size_t protocol_pos = url.find("://");
+            if (protocol_pos == std::string::npos) {
+                return url;  // Malformed URL
+            }
+
+            std::size_t at_pos = url.find('@', protocol_pos + 3);
+            if (at_pos == std::string::npos) {
+                return url;  // No password in URL
+            }
+
+            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
+        };
+
+        // start the download
+        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        if (!was_perform_successful) {
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code < 200 || http_code >= 400) {
+            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
+            return false;
+        }
+
+        // Causes file to be closed explicitly here before we rename it.
+        outfile.reset();
+
+        // Write the updated JSON metadata file.
+        metadata.update({
+            {"url", url},
+            {"etag", headers.etag},
+            {"lastModified", headers.last_modified}
+        });
+        std::ofstream(metadata_path) << metadata.dump(4);
+        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+
+        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
+            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+            return false;
+        }
+    }
+
+    return true;
+}
+
+// download multiple files from remote URLs to local paths
+// the input is a vector of pairs <url, path>
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token) {
+    // Prepare download in parallel
+    std::vector<std::future<bool>> futures_download;
+    for (auto const & item : urls) {
+        futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair<std::string, std::string> & it) -> bool {
+            return common_download_file_single(it.first, it.second, bearer_token);
+        }, item));
+    }
+
+    // Wait for all downloads to complete
+    for (auto & f : futures_download) {
+        if (!f.get()) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool common_download_model(
+        const common_params_model & model,
+        const std::string & bearer_token) {
+    // Basic validation of the model.url
+    if (model.url.empty()) {
+        LOG_ERR("%s: invalid model url\n", __func__);
+        return false;
+    }
+
+    if (!common_download_file_single(model.url, model.path, bearer_token)) {
+        return false;
+    }
+
+    // check for additional GGUFs split to download
+    int n_split = 0;
+    {
+        struct gguf_init_params gguf_params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ NULL,
+        };
+        auto * ctx_gguf = gguf_init_from_file(model.path.c_str(), gguf_params);
+        if (!ctx_gguf) {
+            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, model.path.c_str());
+            return false;
+        }
+
+        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+        if (key_n_split >= 0) {
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+        }
+
+        gguf_free(ctx_gguf);
+    }
+
+    if (n_split > 1) {
+        char split_prefix[PATH_MAX] = {0};
+        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+
+        // Verify the first split file format
+        // and extract split URL and PATH prefixes
+        {
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), model.path.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split);
+                return false;
+            }
+
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model.url.c_str(), 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model.url.c_str(), n_split);
+                return false;
+            }
+        }
+
+        std::vector<std::pair<std::string, std::string>> urls;
+        for (int idx = 1; idx < n_split; idx++) {
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
+
+            char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+            llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);
+
+            if (std::string(split_path) == model.path) {
+                continue; // skip the already downloaded file
+            }
+
+            urls.push_back({split_url, split_path});
+        }
+
+        // Download in parallel
+        common_download_file_multiple(urls, bearer_token);
+    }
+
+    return true;
+}
+
+/**
+ * Allow getting the HF file from the HF repo with tag (like ollama), for example:
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
+ * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
+ *
+ * Return pair of <repo, file> (with "repo" already having tag removed)
+ *
+ * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
+ */
+static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token) {
+    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
+    std::string tag = parts.size() > 1 ? parts.back() : "latest";
+    std::string hf_repo = parts[0];
+    if (string_split<std::string>(hf_repo, '/').size() != 2) {
+        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
+    }
+
+    // fetch model info from Hugging Face Hub API
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::string res_str;
+    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (!bearer_token.empty()) {
+        std::string auth_header = "Authorization: Bearer " + bearer_token;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+    }
+    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        throw std::runtime_error("error: cannot make GET request to HF API");
+    }
+
+    long res_code;
+    std::string ggufFile   = "";
+    std::string mmprojFile = "";
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+    if (res_code == 200) {
+        // extract ggufFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                ggufFile = match[1].str();
+            }
+        }
+        // extract mmprojFile.rfilename in json, using regex
+        {
+            std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
+            std::smatch match;
+            if (std::regex_search(res_str, match, pattern)) {
+                mmprojFile = match[1].str();
+            }
+        }
+    } else if (res_code == 401) {
+        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
+    } else {
+        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
+    }
+
+    // check response
+    if (ggufFile.empty()) {
+        throw std::runtime_error("error: model does not have ggufFile");
+    }
+
+    return { hf_repo, ggufFile, mmprojFile };
+}
+
+#else
+
+static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from internet\n");
+    return false;
+}
+
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static bool common_download_model(
+        const common_params_model &,
+        const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return false;
+}
+
+static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
+    return {};
+}
+
+#endif // LLAMA_USE_CURL
+
 //
 // utils
 //
 
-static void common_params_handle_model_default(
-        std::string & model,
-        const std::string & model_url,
-        std::string & hf_repo,
-        std::string & hf_file,
-        const std::string & hf_token,
-        const std::string & model_default) {
-    if (!hf_repo.empty()) {
-        // short-hand to avoid specifying --hf-file -> default it to --model
-        if (hf_file.empty()) {
-            if (model.empty()) {
-                auto auto_detected = common_get_hf_file(hf_repo, hf_token);
-                if (auto_detected.first.empty() || auto_detected.second.empty()) {
-                    exit(1); // built without CURL, error message already printed
+static void common_params_handle_model(
+        struct common_params_model & model,
+        const std::string & bearer_token,
+        const std::string & model_path_default,
+        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+    // handle pre-fill default model path and url based on hf_repo and hf_file
+    {
+        if (!model.hf_repo.empty()) {
+            // short-hand to avoid specifying --hf-file -> default it to --model
+            if (model.hf_file.empty()) {
+                if (model.path.empty()) {
+                    auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token);
+                    if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
+                        exit(1); // built without CURL, error message already printed
+                    }
+                    model.hf_repo = auto_detected.repo;
+                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                } else {
+                    model.hf_file = model.path;
                 }
-                hf_repo = auto_detected.first;
-                hf_file = auto_detected.second;
-            } else {
-                hf_file = model;
             }
+
+            // TODO: allow custom host
+            model.url = "https://huggingface.co/" + model.hf_repo + "/resolve/main/" + model.hf_file;
+
+            // make sure model path is present (for caching purposes)
+            if (model.path.empty()) {
+                // this is to avoid different repo having same file name, or same file name in different subdirs
+                std::string filename = model.hf_repo + "_" + model.hf_file;
+                // to make sure we don't have any slashes in the filename
+                string_replace_all(filename, "/", "_");
+                model.path = fs_get_cache_file(filename);
+            }
+
+        } else if (!model.url.empty()) {
+            if (model.path.empty()) {
+                auto f = string_split<std::string>(model.url, '#').front();
+                f = string_split<std::string>(f, '?').front();
+                model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            }
+
+        } else if (model.path.empty()) {
+            model.path = model_path_default;
         }
-        // make sure model path is present (for caching purposes)
-        if (model.empty()) {
-            // this is to avoid different repo having same file name, or same file name in different subdirs
-            std::string filename = hf_repo + "_" + hf_file;
-            // to make sure we don't have any slashes in the filename
-            string_replace_all(filename, "/", "_");
-            model = fs_get_cache_file(filename);
+    }
+
+    // then, download it if needed
+    if (!model.url.empty()) {
+        bool ok = common_download_model(model, bearer_token);
+        if (!ok) {
+            LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
+            exit(1);
         }
-    } else if (!model_url.empty()) {
-        if (model.empty()) {
-            auto f = string_split<std::string>(model_url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            model = fs_get_cache_file(string_split<std::string>(f, '/').back());
-        }
-    } else if (model.empty()) {
-        model = model_default;
     }
 }
 
@@ -300,10 +821,16 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // TODO: refactor model params in a common struct
-    common_params_handle_model_default(params.model,             params.model_url,             params.hf_repo,             params.hf_file,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
-    common_params_handle_model_default(params.vocoder.model,     params.vocoder.model_url,     params.vocoder.hf_repo,     params.vocoder.hf_file,     params.hf_token, "");
+    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
+    common_params_handle_model(params.speculative.model, params.hf_token, "");
+    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
+
+    // allow --mmproj to be set from -hf
+    // assuming that mmproj is always in the same repo as text model
+    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
+        params.mmproj.hf_repo = params.model.hf_repo;
+    }
+    common_params_handle_model(params.mmproj,            params.hf_token, "", true);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -322,6 +849,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.kv_overrides.back().key[0] = 0;
     }
 
+    if (!params.tensor_buft_overrides.empty()) {
+        params.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
+
     if (params.reranking && params.embedding) {
         throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
     }
@@ -764,7 +1295,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CTX_SIZE"));
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
-        string_format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict),
+        string_format(
+            ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL
+                ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)"
+                : "number of tokens to predict (default: %d, -1 = infinity)",
+            params.n_predict),
         [](common_params & params, int value) {
             params.n_predict = value;
         }
@@ -813,13 +1348,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_FLASH_ATTN"));
     add_opt(common_arg(
         {"-p", "--prompt"}, "PROMPT",
-        ex == LLAMA_EXAMPLE_MAIN
-            ? "prompt to start generation with\nif -cnv is set, this will be used as system prompt"
-            : "prompt to start generation with",
+        "prompt to start generation with; for system message, use -sys",
         [](common_params & params, const std::string & value) {
             params.prompt = value;
         }
     ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-sys", "--system-prompt"}, "PROMPT",
+        "system prompt to use with model (if applicable, depending on chat template)",
+        [](common_params & params, const std::string & value) {
+            params.system_prompt = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--no-perf"},
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@@ -844,6 +1384,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-sysf", "--system-prompt-file"}, "FNAME",
+        "a file containing the system prompt (default: none)",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
+                params.system_prompt.pop_back();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
         "an input file (repeat to specify multiple files)",
@@ -944,6 +1498,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    add_opt(common_arg(
+        {"-st", "--single-turn"},
+        "run conversation for a single turn only, then exit when done\n"
+        "will not be interactive if first turn is predefined with --prompt\n"
+        "(default: false)",
+        [](common_params & params) {
+            params.single_turn = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"-i", "--interactive"},
         string_format("run in interactive mode (default: %s)", params.interactive ? "true" : "false"),
@@ -1529,7 +2092,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--mmproj"}, "FILE",
         "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
-            params.mmproj = value;
+            params.mmproj.path = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    add_opt(common_arg(
+        {"--mmproj-url"}, "URL",
+        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        [](common_params & params, const std::string & value) {
+            params.mmproj.url = value;
         }
     ).set_examples({LLAMA_EXAMPLE_LLAVA}));
     add_opt(common_arg(
@@ -1615,6 +2185,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
+    add_opt(common_arg(
+        {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
+        "override tensor buffer type", [](common_params & params, const std::string & value) {
+            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+            if (buft_list.empty()) {
+                // enumerate all the devices and add their buffer types to the list
+                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    auto * buft = ggml_backend_dev_buffer_type(dev);
+                    if (buft) {
+                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                    }
+                }
+            }
+
+            for (const auto & override : string_split<std::string>(value, ',')) {
+                std::string::size_type pos = override.find('=');
+                if (pos == std::string::npos) {
+                    throw std::invalid_argument("invalid value");
+                }
+                std::string tensor_name = override.substr(0, pos);
+                std::string buffer_type = override.substr(pos + 1);
+
+                if (buft_list.find(buffer_type) == buft_list.end()) {
+                    printf("Available buffer types:\n");
+                    for (const auto & it : buft_list) {
+                        printf("  %s\n", ggml_backend_buft_name(it.second));
+                    }
+                    throw std::invalid_argument("unknown buffer type");
+                }
+                // FIXME: this leaks memory
+                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+            }
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1758,14 +2363,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
             ),
         [](common_params & params, const std::string & value) {
-            params.model = value;
+            params.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.model_url = value;
+            params.model.url = value;
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
     add_opt(common_arg(
@@ -1774,35 +2379,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "example: unsloth/phi-4-GGUF:q4_k_m\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_repo = value;
+            params.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO"));
     add_opt(common_arg(
         {"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
         "Same as --hf-repo, but for the draft model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.hf_repo = value;
+            params.speculative.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HFD_REPO"));
     add_opt(common_arg(
         {"-hff", "--hf-file"}, "FILE",
         "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.hf_file = value;
+            params.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE"));
     add_opt(common_arg(
         {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
         "Hugging Face model repository for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_repo = value;
+            params.vocoder.model.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HF_REPO_V"));
     add_opt(common_arg(
         {"-hffv", "--hf-file-v"}, "FILE",
         "Hugging Face model file for the vocoder model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.hf_file = value;
+            params.vocoder.model.hf_file = value;
         }
     ).set_env("LLAMA_ARG_HF_FILE_V"));
     add_opt(common_arg(
@@ -1853,18 +2458,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_PASSKEY}));
     add_opt(common_arg(
         {"-o", "--output", "--output-file"}, "FNAME",
-        string_format("output file (default: '%s')",
-            ex == LLAMA_EXAMPLE_EXPORT_LORA
-                ? params.lora_outfile.c_str()
-                : ex == LLAMA_EXAMPLE_CVECTOR_GENERATOR
-                    ? params.cvector_outfile.c_str()
-                    : params.out_file.c_str()),
+        string_format("output file (default: '%s')", params.out_file.c_str()),
         [](common_params & params, const std::string & value) {
             params.out_file = value;
-            params.cvector_outfile = value;
-            params.lora_outfile = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -1954,7 +2552,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
-        string_format("ip address to listen (default: %s)", params.hostname.c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
         [](common_params & params, const std::string & value) {
             params.hostname = value;
         }
@@ -2429,7 +3027,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.model = value;
+            params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
 
@@ -2437,7 +3035,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-mv", "--model-vocoder"}, "FNAME",
         "vocoder model for audio generation (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.vocoder.model = value;
+            params.vocoder.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
      add_opt(common_arg(
@@ -2447,16 +3045,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.vocoder.use_guide_tokens = true;
         }
     ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--tts-speaker-file"}, "FNAME",
+        "speaker file path for audio generation",
+        [](common_params & params, const std::string & value) {
+            params.vocoder.speaker_file = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_TTS}));
 
     // model-specific
     add_opt(common_arg(
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf";
+            params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
+            params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
+            params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
+            params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
@@ -2464,8 +3069,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-bge-small-en-default"},
         string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
-            params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
+            params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2478,8 +3083,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-e5-small-en-default"},
         string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
-            params.hf_file = "e5-small-v2-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
+            params.model.hf_file = "e5-small-v2-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2492,8 +3097,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--embd-gte-small-default"},
         string_format("use default gte-small model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
-            params.hf_file = "gte-small-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
+            params.model.hf_file = "gte-small-q8_0.gguf";
             params.pooling_type = LLAMA_POOLING_TYPE_NONE;
             params.embd_normalize = 2;
             params.n_ctx = 512;
@@ -2506,8 +3111,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-1.5b-default"},
         string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2522,8 +3127,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-3b-default"},
         string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;
@@ -2538,8 +3143,46 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--fim-qwen-7b-default"},
         string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
-            params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.port = 8012;
+            params.n_gpu_layers = 99;
+            params.flash_attn = true;
+            params.n_ubatch = 1024;
+            params.n_batch = 1024;
+            params.n_ctx = 0;
+            params.n_cache_reuse = 256;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--fim-qwen-7b-spec"},
+        string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.speculative.n_gpu_layers = 99;
+            params.port = 8012;
+            params.n_gpu_layers = 99;
+            params.flash_attn = true;
+            params.n_ubatch = 1024;
+            params.n_batch = 1024;
+            params.n_ctx = 0;
+            params.n_cache_reuse = 256;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--fim-qwen-14b-spec"},
+        string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
+            params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
+            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.speculative.n_gpu_layers = 99;
             params.port = 8012;
             params.n_gpu_layers = 99;
             params.flash_attn = true;

common/chat.cpp

@@ -60,7 +60,9 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
             }
             msg.role = message.at("role");
 
-            if (message.contains("content")) {
+            auto has_content = message.contains("content");
+            auto has_tool_calls = message.contains("tool_calls");
+            if (has_content) {
                 const auto & content = message.at("content");
                 if (content.is_string()) {
                     msg.content = content;
@@ -81,19 +83,8 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                 } else if (!content.is_null()) {
                     throw std::runtime_error("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
                 }
-            } else {
-                throw std::runtime_error("Expected 'content' (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
             }
-            if (message.contains("reasoning_content")) {
-                msg.reasoning_content = message.at("reasoning_content");
-            }
-            if (message.contains("name")) {
-                msg.tool_name = message.at("name");
-            }
-            if (message.contains("tool_call_id")) {
-                msg.tool_call_id = message.at("tool_call_id");
-            }
-            if (message.contains("tool_calls")) {
+            if (has_tool_calls) {
                 for (const auto & tool_call : message.at("tool_calls")) {
                     common_chat_tool_call tc;
                     if (!tool_call.contains("type")) {
@@ -118,6 +109,18 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                     msg.tool_calls.push_back(tc);
                 }
             }
+            if (!has_content && !has_tool_calls) {
+                throw std::runtime_error("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)");
+            }
+            if (message.contains("reasoning_content")) {
+                msg.reasoning_content = message.at("reasoning_content");
+            }
+            if (message.contains("name")) {
+                msg.tool_name = message.at("name");
+            }
+            if (message.contains("tool_call_id")) {
+                msg.tool_call_id = message.at("tool_call_id");
+            }
 
             msgs.push_back(msg);
         }
@@ -442,6 +445,7 @@ std::string common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
         case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
+        case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING: return "Hermes 2 Pro (extract reasoning)";
         case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
         case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: return "Command R7B (extract reasoning)";
         default:
@@ -449,12 +453,6 @@ std::string common_chat_format_name(common_chat_format format) {
     }
 }
 
-const common_grammar_options grammar_options {
-    /* .dotall = */ false,
-    /* .compact_spaces = */ false,
-    // /* .compact_spaces = */ true,
-};
-
 static bool parse_json(std::string::const_iterator & it, const std::string::const_iterator & end, json & out) {
     // // https://json.nlohmann.me/features/parsing/sax_interface/
     struct json_error_locator : public nlohmann::json_sax<json> {
@@ -500,6 +498,34 @@ static bool parse_json(std::string::const_iterator & it, const std::string::cons
     }
 }
 
+static bool parse_literal(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) {
+    auto expected_it = expected.begin();
+    auto tmp_it = it;
+    while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) {
+        ++tmp_it;
+        ++expected_it;
+    }
+    if (expected_it == expected.end()) {
+        it = tmp_it;
+        return true;
+    }
+    return false;
+}
+
+static std::optional<std::smatch> parse_pattern(std::string::const_iterator & it, const std::string::const_iterator & end, const std::regex & expected) {
+    std::smatch match;
+    if (std::regex_match(it, end, match, expected)) {
+        it = match.suffix().first;
+        return match;
+    }
+    return std::nullopt;
+}
+
+static void consume_spaces(std::string::const_iterator & it, const std::string::const_iterator & end) {
+    while (it != end && std::isspace(*it)) {
+        ++it;
+    }
+}
 
 /**
  * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between.
@@ -509,7 +535,8 @@ static common_chat_msg parse_json_tool_calls(
     const std::string& input,
     const std::optional<std::regex> & trigger_opt,
     const std::regex & function_regex,
-    const std::regex & close_regex) {
+    const std::regex & close_regex,
+    bool allow_raw_python = false) {
     std::smatch match;
 
     common_chat_msg result;
@@ -540,14 +567,19 @@ static common_chat_msg parse_json_tool_calls(
         it = rit->suffix().first;
 
         json arguments;
-        if (!parse_json(it, end, arguments)) {
+        if (parse_json(it, end, arguments)) {
+            if (!std::regex_search(it, end, match, close_regex)) {
+                throw std::runtime_error("Malformed input, missing closing pattern: " + input);
+            }
+            it = match.suffix().first;
+            result.tool_calls.push_back({name, arguments.is_string() ? arguments.get<std::string>() : arguments.dump(), /* id= */ ""});
+        } else {
+            if (allow_raw_python && name == "python") {
+                result.tool_calls.push_back({name, json({{"code", std::string(it, end)}}).dump(), /* id= */ ""});
+                break;
+            }
             throw std::runtime_error("Failed to parse json tool call arguments: " + input);
         }
-        if (!std::regex_search(it, end, match, close_regex)) {
-            throw std::runtime_error("Malformed input, missing closing pattern: " + input);
-        }
-        it = match.suffix().first;
-        result.tool_calls.push_back({name, arguments.is_string() ? arguments.get<std::string>() : arguments.dump(), /* id= */ ""});
     }
 
     if (!result.tool_calls.empty()) {
@@ -559,29 +591,29 @@ static common_chat_msg parse_json_tool_calls(
     return result;
 }
 
+static common_chat_tool_call process_tool_call(const json & tool_call) {
+    const auto & arguments = tool_call.at("arguments");
+    return {
+        /* .name = */ tool_call.at("name"),
+        /* .arguments = */ arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
+        /* .id = */ tool_call.contains("id") ? tool_call.at("id") : "",
+    };
+}
 static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& input, const std::string & prefix, size_t rstrip_prefix = 0) {
     auto content_end = input.find(prefix);
     size_t tc_start = std::string::npos;
 
     common_chat_msg result;
     result.role = "assistant";
-    const auto process_tool_calls = [&](const json & tool_calls) {
-        for (const auto & tool_call : tool_calls) {
-            const auto & arguments = tool_call.at("arguments");
-            result.tool_calls.push_back({
-                tool_call.at("name"),
-                arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
-                tool_call.contains("id") ? tool_call.at("id") : "",
-            });
-        }
-    };
     if (content_end == std::string::npos) {
         result.content = input;
     } else {
         tc_start = content_end + prefix.size() - rstrip_prefix;
         result.content = input.substr(0, content_end);
         auto tool_calls = json::parse(input.substr(tc_start));
-        process_tool_calls(tool_calls);
+        for (const auto & tool_call : tool_calls) {
+            result.tool_calls.emplace_back(process_tool_call(tool_call));
+        }
     }
     return result;
 }
@@ -700,7 +732,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
     data.grammar_lazy = false;
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         builder.add_schema("root", schema);
-    }, grammar_options);
+    });
 
     auto tweaked_messages = common_chat_template::add_system(
         inputs.messages,
@@ -770,8 +802,11 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
             schema["maxItems"] = 1;
         }
         builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema));
-    }, grammar_options);
-    data.grammar_triggers.push_back({"[TOOL_CALLS]", /* .at_start = */ true});
+    });
+    data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"});
+    data.preserved_tokens = {
+        "[TOOL_CALLS]",
+    };
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
@@ -813,14 +848,18 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             schema["maxItems"] = 1;
         }
         builder.add_rule("root", "\"<|START_ACTION|>\" " + builder.add_schema("tool_calls", schema) + " \"<|END_ACTION|>\"");
-    }, grammar_options);
-    data.grammar_triggers.push_back({"<|START_ACTION|>", /* .at_start = */ false});
+    });
+    data.grammar_triggers.push_back({
+        COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+        "<|START_ACTION|>",
+    });
     data.preserved_tokens = {
+        "<|START_ACTION|>",
+        "<|END_ACTION|>",
         "<|START_RESPONSE|>",
         "<|END_RESPONSE|>",
         "<|START_THINKING|>",
         "<|END_THINKING|>",
-        "<|END_ACTION|>",
     };
     auto adjusted_messages = json::array();
     for (const auto & msg : inputs.messages) {
@@ -840,9 +879,9 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
     return data;
 }
 static common_chat_msg common_chat_parse_command_r7b(const std::string & input, bool extract_reasoning) {
-    static std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S\\n\\r]*?)<\\|END_THINKING\\|>)([\\s\\S\\n\\r]*)");
-    static std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S\\n\\r]*?)<\\|END_ACTION\\|>");
-    static std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S\\n\\r]*?)<\\|END_RESPONSE\\|>");
+    static const std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S]*?)<\\|END_THINKING\\|>)([\\s\\S]*)");
+    static const std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S]*?)<\\|END_ACTION\\|>");
+    static const std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S]*?)<\\|END_RESPONSE\\|>");
 
     std::smatch match;
 
@@ -945,23 +984,23 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com
                 builder.add_rule(
                     name + "-call",
                     "\"{\" space "
-                    "( \"\\\"type\\\":\" space \"\\\"function\\\",\" space )? "
-                    "\"\\\"name\\\": \\\"" + name + "\\\", \\\"parameters\\\": \" " +
-                        builder.add_schema(name + "-args", parameters) +
-                    " \"}\""));
-            data.grammar_triggers.push_back({"{\"name\": \"" + name + "\"", /* .at_start = */ true});
+                    "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
+                    "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
+                    "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
+                    "\"}\" space"));
+        });
+        // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
+        data.grammar_triggers.push_back({
+            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+            "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
         });
-        data.grammar_triggers.push_back({"{\"name\":", /* .at_start = */ true});
-        data.grammar_triggers.push_back({"{\n  \"name\":", /* .at_start = */ true});
-        data.grammar_triggers.push_back({"{\n    \"name\":", /* .at_start = */ true});
-        data.grammar_triggers.push_back({"{\"type\": \"function\"", /* .at_start = */ true});
-        data.grammar_triggers.push_back({"{\n  \"type\": \"function\"", /* .at_start = */ true});
-        data.grammar_triggers.push_back({"{\n    \"type\": \"function\"", /* .at_start = */ true});
         if (!builtin_tools.empty()) {
-            data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+            data.preserved_tokens.push_back("<|python_tag|>");
         }
+        // Allow a few empty lines on top of the usual constrained json schema space rule.
         builder.add_rule("root", string_join(tool_rules, " | "));
-    }, grammar_options);
+    });
     data.additional_stops.push_back("<|eom_id|>");
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
         {"tools_in_user_message", false},
@@ -974,33 +1013,33 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com
 }
 static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {
     // TODO: tighten & simplify the parser, don't accept leading text context.
-    static std::regex function_regex("\\{[\\s\\n\\r]*(?:\"type\"[\\s\\n\\r]*:[\\s\\n\\r]*\"function\"[\\s\\n\\r]*,[\\s\\n\\r]*|[\\s\\n\\r]*)\"name\"[\\s\\n\\r]*:[\\s\\n\\r]*\"([^\"]+)\"[\\s\\n\\r]*,[\\s\\n\\r]*\"parameters\": ");
-    static std::regex close_regex("\\}");
-    static std::regex builtin_call_regex("<\\|python_tag\\|>([^.(]+)\\.call\\((.*)\\)");
+    static const std::regex function_regex(
+        "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: ");
+    static const std::regex close_regex("\\}\\s*");
+    static const std::regex builtin_call_regex("<\\|python_tag\\|>\\s*([^.(]+)\\s*\\.\\s*call\\s*\\(\\s*([\\w]+)\\s*=\\s*([\\s\\S]*?)\\)");
 
     if (with_builtin_tools) {
         std::smatch match;
         if (std::regex_match(input, match, builtin_call_regex)) {
-            auto name = match[1].str();
-            auto raw_args = match[2].str();
+            try {
+                auto name = match[1].str();
+                auto arg_name = match[2].str();
+                auto arg_value_str = match[3].str();
+                auto arg_value = json::parse(arg_value_str);
 
-            // TODO: if/when builtin tools start accepting more than 1 argument, use parse_json for real parsing.
-            auto it_eq = raw_args.find('=');
-            auto arg_name = raw_args.substr(0, it_eq);
-            auto arg_value_str = raw_args.substr(it_eq + 1);
-            auto arg_value = json::parse(arg_value_str);
-
-            common_chat_msg msg;
-            msg.role = "assistant";
-            msg.content = match.prefix().str();
-            msg.tool_calls.push_back({
-                /* .name = */ name,
-                /* .arguments = */ (json {
-                    {arg_name, arg_value},
-                }).dump(),
-                /* .id = */ "",
-            });
-            return msg;
+                common_chat_msg msg;
+                msg.role = "assistant";
+                msg.tool_calls.push_back({
+                    /* .name = */ name,
+                    /* .arguments = */ (json {
+                        {arg_name, arg_value},
+                    }).dump(),
+                    /* .id = */ "",
+                });
+                return msg;
+            } catch (const std::exception & e) {
+                LOG_WRN("Failed to parse builtin tool call arguments (%s): %s", e.what(), input.c_str());
+            }
         }
     }
     return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex);
@@ -1017,10 +1056,10 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
                 std::string name = function.at("name");
                 auto parameters = function.at("parameters");
                 builder.resolve_refs(parameters);
-                auto args_rule = builder.add_schema(name + "-args", parameters);
                 tool_rules.push_back(builder.add_rule(name + "-call",
                     "\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n"
-                    "```json\\n\" " + args_rule + " \"```<｜tool▁call▁end｜>\""));
+                    "```json\\n\" " + builder.add_schema(name + "-args", parameters) + " "
+                    "\"```<｜tool▁call▁end｜>\""));
             });
             // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag,
             // so we accept common variants (then it's all constrained)
@@ -1029,18 +1068,20 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
                 "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " "
                 "\"<｜tool▁calls▁end｜>\""
                 " space");
-            data.grammar_triggers.push_back({"<｜tool▁calls▁begin｜>", /* .at_start = */ false});
-            data.grammar_triggers.push_back({"<｜tool_calls_begin｜>", /* .at_start = */ false});
-            data.grammar_triggers.push_back({"<｜tool calls begin｜>", /* .at_start = */ false});
-            data.grammar_triggers.push_back({"<｜tool\\_calls\\_begin｜>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool▁calls▁begin｜>"});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool_calls_begin｜>"});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool calls begin｜>"});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool\\_calls\\_begin｜>"});
             data.preserved_tokens = {
                 "<think>",
                 "</think>",
+                "<｜tool▁calls▁begin｜>",
+                "<｜tool▁call▁begin｜>",
                 "<｜tool▁sep｜>",
-                "<｜tool▁calls▁end｜",
                 "<｜tool▁call▁end｜>",
+                "<｜tool▁calls▁end｜",
             };
-        }, grammar_options);
+        });
     }
     auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
 
@@ -1065,34 +1106,42 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
     data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING : COMMON_CHAT_FORMAT_DEEPSEEK_R1;
     return data;
 }
-static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) {
-    static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n");
-    static std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
-    static std::regex reasoning_content_regex("((?:<think>)?([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");
-    static std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>");
-    common_chat_msg msg;
-    msg.role = "assistant";
+static common_chat_msg handle_think_tag_prelude(const std::string & input, bool extract_reasoning, const std::function<common_chat_msg(const std::string &)> & rest_parser) {
     std::smatch match;
+    static const std::regex reasoning_content_regex("((?:<think>)?([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");
     if (std::regex_match(input, match, reasoning_content_regex)) {
-        std::string rest;
+        auto rest = match[3].str();
+        auto msg = rest_parser(rest);
+        auto reasoning_content = string_strip(match[2].str());
         if (extract_reasoning) {
-            msg.reasoning_content = string_strip(match[2].str());
-        } else {
-            msg.content = match[1].str();
+            msg.reasoning_content = reasoning_content;
+        } else if (!reasoning_content.empty()) {
+            std::ostringstream content;
+            content << "<think>" << reasoning_content << "</think>" << msg.content;
+            msg.content = content.str();
         }
-        rest = match[3].str();
+        return msg;
+    }
+    return rest_parser(input);
+}
+static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) {
+    return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) {
+        static const std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n");
+        static const std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
+        static const std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>");
 
-        if (std::regex_search(rest, match, tool_calls_regex)) {
+        common_chat_msg msg;
+        msg.role = "assistant";
+        std::smatch match;
+        if (std::regex_search(input, match, tool_calls_regex)) {
             auto tool_calls = match[1].str();
             auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex);
             msg.tool_calls = std::move(msg2.tool_calls);
         } else {
-            msg.content += std::string(rest.begin() + rest.find_first_not_of(" \r\n"), rest.end());
+            msg.content = input;
         }
-    } else {
-        msg.content = input;
-    }
-    return msg;
+        return msg;
+    });
 }
 
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
@@ -1129,8 +1178,11 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
                 schema["maxItems"] = 1;
             }
             builder.add_rule("root", "\" functools\"? " + builder.add_schema("tool_calls", schema));
-        }, grammar_options);
-        data.grammar_triggers.push_back({" functools[", /* .at_start = */ false});
+        });
+        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, " functools["});
+        data.preserved_tokens = {
+            " functools[",
+        };
         data.format = COMMON_CHAT_FORMAT_FIREFUNCTION_V2;
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
@@ -1158,11 +1210,28 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
                 auto parameters = function.at("parameters");
                 builder.resolve_refs(parameters);
                 auto args_rule = builder.add_schema(name + "-args", parameters);
-                first_tool_rules.push_back(builder.add_rule(name + "-call", "\"" + name + "\\n\" " + args_rule));
+                first_tool_rules.push_back(builder.add_rule(name + "-call", "( \"assistant<|end_header_id|>\\n\" )? \"" + name + "\\n\" " + args_rule));
                 subsequent_tool_rules.push_back(builder.add_rule(name + "-call2", "\">>>" + name + "\\n\" " + args_rule));
-                data.grammar_triggers.push_back({name, /* .at_start = */ true});
-                data.grammar_triggers.push_back({">>>" + name, /* .at_start = */ false});
+                data.grammar_triggers.push_back({
+                    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+                    regex_escape(name + "\n"),
+                });
+                data.grammar_triggers.push_back({
+                    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+                    regex_escape("assistant<|end_header_id|>\n" + name + "\n"),
+                });
+                data.grammar_triggers.push_back({
+                    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+                    regex_escape(">>>" + name + "\n"),
+                });
+                data.grammar_triggers.push_back({
+                    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+                    ">>>assistant<|end_header_id|>\n" + name,
+                });
             });
+            data.preserved_tokens = {
+                "<|end_header_id|>",
+            };
             auto first_rule = first_tool_rules.empty() ? "" : builder.add_rule("first_tool_call", string_join(first_tool_rules, " | ")) + " space";
             if (inputs.parallel_tool_calls) {
                 auto subsequent_rule = builder.add_rule("subsequent_tool_call", string_join(subsequent_tool_rules, " | ")) + " space";
@@ -1171,34 +1240,20 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
                 builder.add_rule("root", first_rule);
             }
 
-        }, grammar_options);
+        });
     }
     return data;
 }
 
-static bool consume(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) {
-    auto expected_it = expected.begin();
-    auto tmp_it = it;
-    while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) {
-        ++tmp_it;
-        ++expected_it;
-    }
-    if (expected_it == expected.end()) {
-        it = tmp_it;
-        return true;
-    }
-    return false;
-}
-
 static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & input) {
-    static std::regex function_regex(R"((?:>>>)?(\w+)\n)");
-    static std::regex close_regex(R"($|(?=>>>))");
+    static const std::regex function_regex(R"((?:>>>)?(?:assistant<|end_header_id|>\n)?(\w+)\n)");
+    static const std::regex close_regex(R"($|(?=>>>))");
 
     std::string content;
     auto it = input.begin();
     const auto end = input.end();
 
-    if (consume(it, end, "all\n")) {
+    if (parse_literal(it, end, "all\n")) {
         std::smatch match;
         if (std::regex_search(it, end, match, function_regex)) {
             auto fun_it = match.prefix().second;
@@ -1213,7 +1268,7 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in
     }
     // TODO: tighten & simplify.
     try {
-        auto res = parse_json_tool_calls(std::string(it, end), std::nullopt, function_regex, close_regex);
+        auto res = parse_json_tool_calls(std::string(it, end), std::nullopt, function_regex, close_regex, /* allow_raw_python= */ true);
         res.content = content + res.content;
         return res;
     } catch (const std::exception & e) {
@@ -1266,12 +1321,13 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         });
         if (has_raw_python) {
             tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
-            data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+            data.preserved_tokens.push_back("<|python_tag|>");
         }
         auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
         builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
-        data.grammar_triggers.push_back({"<function=", /* .at_start = */ false});
-    }, grammar_options);
+        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
+    });
 
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     // TODO: if (has_raw_python)
@@ -1280,7 +1336,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
 }
 static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) {
     // This version of Functionary still supports the llama 3.1 tool call format for the python tool.
-    static std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)");
+    static const std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)");
     std::smatch match;
     if (std::regex_search(input, match, python_tag_regex)) {
         auto code = match[1].str();
@@ -1294,8 +1350,8 @@ static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::s
         });
         return msg;
     }
-    static std::regex function_regex(R"(<function=(\w+)>)");
-    static std::regex close_regex(R"(</function>)");
+    static const std::regex function_regex(R"(<function=(\w+)>)");
+    static const std::regex close_regex(R"(</function>)");
     // TODO: tighten & simplify.
     return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex);
 }
@@ -1306,6 +1362,7 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
     data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         std::vector<std::string> tool_rules;
+        std::vector<std::string> tool_call_alts;
         foreach_function(inputs.tools, [&](const json & tool) {
             const auto & function = tool.at("function");
             std::string name = function.at("name");
@@ -1319,68 +1376,187 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
                 }},
                 {"required", json::array({"name", "arguments"})},
             }));
+            tool_call_alts.push_back(builder.add_rule(
+                name + "-function-tag",
+                "\"<function\" ( \"=" + name + "\" | \" name=\\\"" + name + "\\\"\" ) \">\" space " +
+                builder.add_schema(name + "-args", parameters) + " "
+                "\"</function>\" space"));
+
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+                "<function=" + name + ">",
+            });
+            auto escaped_name = regex_escape(name);
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+                "<function\\s+name\\s*=\\s*\"" + escaped_name + "\"",
+            });
         });
-        auto tool_call = "\"<tool_call>\" space " + builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " \"</tool_call>\" space";
+        auto any_tool_call = builder.add_rule("any_tool_call", "( " + string_join(tool_rules, " | ") + " ) space");
+        std::vector<std::string> alt_tags {
+            any_tool_call,
+            "\"<tool_call>\" space "     + any_tool_call + " \"</tool_call>\"",
+            // The rest is just to accommodate common "good bad" outputs.
+            "\"<function_call>\" space " + any_tool_call + " \"</function_call>\"",
+            "\"<response>\"  space "     + any_tool_call + " \"</response>\"",
+            "\"<tools>\"     space "     + any_tool_call + " \"</tools>\"",
+            "\"<json>\"      space "     + any_tool_call + " \"</json>\"",
+            "\"<xml>\"      space "     + any_tool_call + " \"</xml>\"",
+            "\"<JSON>\"      space "     + any_tool_call + " \"</JSON>\"",
+        };
+        auto wrappable_tool_call = builder.add_rule("wrappable_tool_call", "( " + string_join(alt_tags, " | ") + " ) space");
+        tool_call_alts.push_back(wrappable_tool_call);
+        tool_call_alts.push_back(
+            "( \"```\\n\" | \"```json\\n\" | \"```xml\\n\" ) space " + wrappable_tool_call + " space \"```\" space ");
+        auto tool_call = builder.add_rule("tool_call", string_join(tool_call_alts, " | "));
         builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
-        data.grammar_triggers.push_back({"<tool_call>", /* .at_start = */ false});
-        data.preserved_tokens = { "</tool_call>" };
-    }, grammar_options);
+        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>"});
+        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"});
+        // Trigger on some common known "good bad" outputs (only from the start and with a json that's about a specific argument name to avoid false positives)
+        data.grammar_triggers.push_back({
+            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+            "(?:```(?:json|xml)?\n\\s*)?(?:<function_call>|<tools>|<xml><json>|<response>)?\\s*\\{\\s*\"", //name\"\\s*:\\s*\"" + escaped_name + "\"",
+        });
+        data.preserved_tokens = {
+            "<think>",
+            "</think>",
+            "<tool_call>",
+            "</tool_call>",
+            "<function",
+            "<tools>",
+            "</tools>",
+            "<response>",
+            "</response>",
+            "<function_call>",
+            "</function_call>",
+            "<json>",
+            "</json>",
+            "<JSON>",
+            "</JSON>",
+            "```",
+            "```json",
+            "```xml",
+        };
+    });
 
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
-    data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
+    data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING : COMMON_CHAT_FORMAT_HERMES_2_PRO;
     return data;
 }
-static common_chat_msg common_chat_parse_hermes_2_pro(const std::string & input) {
-    try {
-        std::regex start_pattern(R"([\n\s]*<tool_call>)");
-        std::regex middle_pattern(R"([\n\s]*</tool_call>[\n\s]*<tool_call>)");
-        std::regex end_pattern(R"([\n\s]*</tool_call>[\n\s]*$)");
+static common_chat_msg common_chat_parse_hermes_2_pro(const std::string& input, bool extract_reasoning) {
+    return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) {
+        static const std::regex open_regex(
+            "(?:"
+            "(```(?:xml|json)?\\n\\s*)?"         // match 1 (block_start)
+            "(<tool_call>"                   // match 2 (open_tag)
+            "|<function_call>"
+            "|<tool>"
+            "|<tools>"
+            "|<response>"
+            "|<json>"
+            "|<xml>"
+            "|<JSON>"
+            ")?"
+            "(\\s*\\{\\s*\"name\"\\s*:[\\s\\S]*)"    // match 3 (named tool call + rest)
+            ")"
+            "|"
+            "(?:<function=([^>]+)>"            // match 4 (function name)
+            "|<function name=\"([^\"]+)\">)" // match 5 (function name again)
+            "([\\s\\S]*)"                   // match 6 (function arguments + rest)})"
+        );
 
-        common_chat_msg msg;
-        msg.role = "assistant";
+        try {
+            common_chat_msg msg;
+            msg.role = "assistant";
 
-        auto end = input.end();
-        std::sregex_iterator rend;
-        std::sregex_iterator rit(input.begin(), end, start_pattern);
-        if (rit == rend) {
+            std::string::const_iterator it = input.begin();
+            const std::string::const_iterator end = input.end();
+            std::smatch match;
+
+            while (it != end) {
+                if (std::regex_search(it, end, match, open_regex)) {
+                    // Add content before the match
+                    msg.content += std::string(it, match[0].first);
+
+                    auto block_start = match[1].str();
+                    std::string block_end = block_start.empty() ? "" : "```";
+
+                    auto open_tag = match[2].str();
+                    std::string close_tag;
+
+                    if (match[3].matched) {
+                        close_tag = open_tag.empty() ? "" : "</" + open_tag.substr(1);
+                        auto json_it = match[3].first;
+                        json tool_call;
+                        if (parse_json(json_it, end, tool_call) && tool_call.contains("name") && tool_call.contains("arguments")) {
+
+                            msg.tool_calls.emplace_back(process_tool_call(tool_call));
+                            it = json_it;  // Move iterator past parsed JSON
+
+                            // Handle close tags
+                            consume_spaces(it, end);
+                            if (!close_tag.empty() && !parse_literal(it, end, close_tag)) {
+                                throw std::runtime_error("Failed to parse closing tag");
+                            }
+                            consume_spaces(it, end);
+                            if (!block_end.empty() && !parse_literal(it, end, block_end)) {
+                                throw std::runtime_error("Failed to parse block end");
+                            }
+                            consume_spaces(it, end);
+                        } else {
+                            // Not a valid tool call, treat as content
+                            msg.content += std::string(match[0].first, match[0].second);
+                            it = match[0].second;
+                        }
+                    } else {
+                        auto function_name = match[4].str();
+                        if (function_name.empty()) {
+                            function_name = match[5].str();
+                        }
+                        GGML_ASSERT(!function_name.empty());
+
+                        close_tag = "</function>";
+                        // Start parsing from after the opening tags
+                        auto json_it = match[6].first;
+                        json arguments;
+                        if (parse_json(json_it, end, arguments)) {
+                            msg.tool_calls.emplace_back(process_tool_call({
+                                {"name", function_name},
+                                {"arguments", arguments},
+                            }));
+                            it = json_it;  // Move iterator past parsed JSON
+
+                            // Handle close tags
+                            consume_spaces(it, end);
+                            if (!close_tag.empty() && !parse_literal(it, end, close_tag)) {
+                                throw std::runtime_error("Failed to parse closing tag");
+                            }
+                            consume_spaces(it, end);
+                            if (!block_end.empty() && !parse_literal(it, end, block_end)) {
+                                throw std::runtime_error("Failed to parse block end");
+                            }
+                            consume_spaces(it, end);
+                        } else {
+                            // Not a valid tool call, treat as content
+                            msg.content += std::string(match[0].first, match[0].second);
+                            it = match[0].second;
+                        }
+                    }
+                } else {
+                    // Add remaining content
+                    msg.content += std::string(it, end);
+                    break;
+                }
+            }
+            return msg;
+        } catch (const std::exception & e) {
+            LOG_ERR("Failed to parse hermes 2 pro input: %s\n", e.what());
+            common_chat_msg msg;
+            msg.role = "assistant";
             msg.content = input;
             return msg;
         }
-
-        msg.content = rit->prefix();
-
-        auto it = rit->suffix().first;
-        while (it != end) {
-            json call;
-            if (!parse_json(it, end, call)) {
-                throw std::runtime_error("Failed to parse json tool call");
-            }
-            const auto & arguments = call.at("arguments");
-            msg.tool_calls.push_back({
-                call.at("name"),
-                arguments.dump(),
-                // arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
-                /* id= */ "",
-            });
-            rit = {it, end, middle_pattern};
-            if (rit != rend) {
-                it = rit->suffix().first;
-            } else {
-                rit = {it, end, end_pattern};
-                if (rit == rend) {
-                    throw std::runtime_error("Malformed input, missing </tool_call>");
-                }
-                break;
-            }
-        }
-        return msg;
-    } catch (const std::exception & e) {
-        LOG_ERR("Failed to parse hermes 2 pro input: %s\n", e.what());
-        common_chat_msg msg;
-        msg.role = "assistant";
-        msg.content = input;
-        return msg;
-    }
+    });
 }
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
@@ -1445,6 +1621,11 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_command_r7b(tmpl, params);
     }
 
+    // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
+    if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
+        return common_chat_params_init_hermes_2_pro(tmpl, params);
+    }
+
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -1466,11 +1647,6 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_without_tools(tmpl, params);
     }
 
-    // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
-    if (src.find("<tool_call>") != std::string::npos) {
-        return common_chat_params_init_hermes_2_pro(tmpl, params);
-    }
-
     // Functionary v3.1 (w/ tools)
     if (src.find("<|start_header_id|>") != std::string::npos
         && src.find("<function=") != std::string::npos) {
@@ -1588,7 +1764,9 @@ common_chat_msg common_chat_parse(const std::string & input, common_chat_format
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:
             return common_chat_parse_functionary_v3_1_llama_3_1(input);
         case COMMON_CHAT_FORMAT_HERMES_2_PRO:
-            return common_chat_parse_hermes_2_pro(input);
+            return common_chat_parse_hermes_2_pro(input, /* extract_reasoning= */ false);
+        case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING:
+            return common_chat_parse_hermes_2_pro(input, /* extract_reasoning= */ true);
         case COMMON_CHAT_FORMAT_FIREFUNCTION_V2:
             return common_chat_parse_firefunction_v2(input);
         case COMMON_CHAT_FORMAT_COMMAND_R7B:

common/chat.h

@@ -53,6 +53,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
+    COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
     COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
 

common/common.cpp

@@ -7,10 +7,6 @@
 
 #include "common.h"
 #include "log.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
-#include "json-schema-to-grammar.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -52,47 +48,11 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #endif
-#if defined(LLAMA_USE_CURL)
-#include <curl/curl.h>
-#include <curl/easy.h>
-#include <future>
-#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if defined(LLAMA_USE_CURL)
-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#else
-#include <sys/syslimits.h>
-#endif
-#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
-//
-// CURL utils
-//
-
-using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
-
-// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
-struct curl_slist_ptr {
-    struct curl_slist * ptr = nullptr;
-    ~curl_slist_ptr() {
-        if (ptr) {
-            curl_slist_free_all(ptr);
-        }
-    }
-};
-#endif // LLAMA_USE_CURL
-
-using json = nlohmann::ordered_json;
-
 //
 // CPU utils
 //
@@ -483,6 +443,11 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+std::string regex_escape(const std::string & s) {
+    static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
+    return std::regex_replace(s, special_chars, "\\$0");
+}
+
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
     std::ostringstream result;
     for (size_t i = 0; i < values.size(); ++i) {
@@ -896,22 +861,14 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
 
-    llama_model * model = nullptr;
-
-    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
-    } else if (!params.model_url.empty()) {
-        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
-    } else {
-        model = llama_model_load_from_file(params.model.c_str(), mparams);
-    }
-
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -946,13 +903,13 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
@@ -1029,6 +986,8 @@ struct common_init_result common_init_from_params(common_params & params) {
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
+        llama_set_warmup(lctx, true);
+
         std::vector<llama_token> tmp;
         llama_token bos = llama_vocab_bos(vocab);
         llama_token eos = llama_vocab_eos(vocab);
@@ -1056,9 +1015,10 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
+        llama_set_warmup(lctx, false);
     }
 
     iparams.model.reset(model);
@@ -1082,15 +1042,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (!params.devices.empty()) {
         mparams.devices = params.devices.data();
     }
+
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1098,6 +1061,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
+    if (params.tensor_buft_overrides.empty()) {
+        mparams.tensor_buft_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
+    }
+
     return mparams;
 }
 
@@ -1157,451 +1127,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     return tpp;
 }
 
-#ifdef LLAMA_USE_CURL
-
-#define CURL_MAX_RETRY 3
-#define CURL_RETRY_DELAY_SECONDS 2
-
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
-    int remaining_attempts = max_attempts;
-
-    while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
-
-        CURLcode res = curl_easy_perform(curl);
-        if (res == CURLE_OK) {
-            return true;
-        }
-
-        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
-        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
-
-        remaining_attempts--;
-        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-    }
-
-    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
-
-    return false;
-}
-
-static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
-    // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    if (!curl) {
-        LOG_ERR("%s: error initializing libcurl\n", __func__);
-        return false;
-    }
-
-    bool force_download = false;
-
-    // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-
-    // Check if hf-token or bearer-token was specified
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-    }
-
-#if defined(_WIN32)
-    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
-    //   operating system. Currently implemented under MS-Windows.
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
-
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
-    std::string etag;
-    std::string last_modified;
-
-    if (file_exists) {
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
-            }
-        }
-    } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
-    }
-
-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
-    };
-
-    common_load_model_from_url_headers headers;
-
-    {
-        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
-
-            static std::regex header_regex("([^:]+): (.*)\r\n");
-            static std::regex etag_regex("ETag", std::regex_constants::icase);
-            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-            std::string header(buffer, n_items);
-            std::smatch match;
-            if (std::regex_match(header, match, header_regex)) {
-                const std::string & key = match[1];
-                const std::string & value = match[2];
-                if (std::regex_match(key, match, etag_regex)) {
-                    headers->etag = value;
-                } else if (std::regex_match(key, match, last_modified_regex)) {
-                    headers->last_modified = value;
-                }
-            }
-            return n_items;
-        };
-
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-        }
-    }
-
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
-        if (!etag.empty() && etag != headers.etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
-            should_download = true;
-        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
-            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
-            should_download = true;
-        }
-    }
-    if (should_download) {
-        std::string path_temporary = path + ".downloadInProgress";
-        if (file_exists) {
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
-            }
-        }
-
-        // Set the output file
-
-        struct FILE_deleter {
-            void operator()(FILE * f) const {
-                fclose(f);
-            }
-        };
-
-        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
-        if (!outfile) {
-            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
-            return false;
-        }
-
-        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
-        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
-            return fwrite(data, size, nmemb, (FILE *)fd);
-        };
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
-
-        //  display download progress
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
-
-        // helper function to hide password in URL
-        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
-            std::size_t protocol_pos = url.find("://");
-            if (protocol_pos == std::string::npos) {
-                return url;  // Malformed URL
-            }
-
-            std::size_t at_pos = url.find('@', protocol_pos + 3);
-            if (at_pos == std::string::npos) {
-                return url;  // No password in URL
-            }
-
-            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
-        };
-
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code < 200 || http_code >= 400) {
-            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-            return false;
-        }
-
-        // Causes file to be closed explicitly here before we rename it.
-        outfile.reset();
-
-        // Write the updated JSON metadata file.
-        metadata.update({
-            {"url", url},
-            {"etag", headers.etag},
-            {"lastModified", headers.last_modified}
-        });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
-
-        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
-        }
-    }
-
-    return true;
-}
-
-struct llama_model * common_load_model_from_url(
-        const std::string & model_url,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // Basic validation of the model_url
-    if (model_url.empty()) {
-        LOG_ERR("%s: invalid model_url\n", __func__);
-        return NULL;
-    }
-
-    if (!common_download_file(model_url, local_path, hf_token)) {
-        return NULL;
-    }
-
-    // check for additional GGUFs split to download
-    int n_split = 0;
-    {
-        struct gguf_init_params gguf_params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ NULL,
-        };
-        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
-        if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
-            return NULL;
-        }
-
-        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-        if (key_n_split >= 0) {
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-        }
-
-        gguf_free(ctx_gguf);
-    }
-
-    if (n_split > 1) {
-        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-
-        // Verify the first split file format
-        // and extract split URL and PATH prefixes
-        {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
-                return NULL;
-            }
-
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
-                return NULL;
-            }
-        }
-
-        // Prepare download in parallel
-        std::vector<std::future<bool>> futures_download;
-        for (int idx = 1; idx < n_split; idx++) {
-            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
-                char split_path[PATH_MAX] = {0};
-                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
-
-                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
-
-                return common_download_file(split_url, split_path, hf_token);
-            }, idx));
-        }
-
-        // Wait for all downloads to complete
-        for (auto & f : futures_download) {
-            if (!f.get()) {
-                return NULL;
-            }
-        }
-    }
-
-    return llama_model_load_from_file(local_path.c_str(), params);
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & repo,
-        const std::string & remote_path,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // construct hugging face model url:
-    //
-    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
-    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
-    //
-    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //
-
-    std::string model_url = "https://huggingface.co/";
-    model_url += repo;
-    model_url += "/resolve/main/";
-    model_url += remote_path;
-
-    return common_load_model_from_url(model_url, local_path, hf_token, params);
-}
-
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
-
-    // fetch model info from Hugging Face Hub API
-    json model_info;
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
-        model_info = json::parse(res_str);
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
-    } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
-    }
-
-    // check response
-    if (!model_info.contains("ggufFile")) {
-        throw std::runtime_error("error: model does not have ggufFile");
-    }
-    json & gguf_file = model_info.at("ggufFile");
-    if (!gguf_file.contains("rfilename")) {
-        throw std::runtime_error("error: ggufFile does not have rfilename");
-    }
-
-    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
-}
-
-#else
-
-struct llama_model * common_load_model_from_url(
-        const std::string & /*model_url*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
-    return nullptr;
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & /*repo*/,
-        const std::string & /*remote_path*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return nullptr;
-}
-
-std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return std::make_pair("", "");
-}
-
-#endif // LLAMA_USE_CURL
-
 //
 // Batch utils
 //
@@ -2025,4 +1550,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
-

common/common.h

@@ -110,9 +110,17 @@ enum common_conversation_mode {
     COMMON_CONVERSATION_MODE_AUTO     = 2,
 };
 
+enum common_grammar_trigger_type {
+    COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+};
+
 struct common_grammar_trigger {
-    std::string word;
-    bool at_start;
+    common_grammar_trigger_type type;
+    std::string value;
+    llama_token token = LLAMA_TOKEN_NULL;
 };
 
 // sampling parameters
@@ -163,8 +171,7 @@ struct common_params_sampling {
 
     std::string                         grammar; // optional BNF-like grammar to constrain sampling
     bool                                grammar_lazy = false;
-    std::vector<common_grammar_trigger> grammar_trigger_words;  // optional trigger words to trigger lazy grammar
-    std::vector<llama_token>            grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens.
+    std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
     std::vector<llama_logit_bias> logit_bias; // logit biases to apply
@@ -173,6 +180,13 @@ struct common_params_sampling {
     std::string print() const;
 };
 
+struct common_params_model {
+    std::string path    = ""; // model local path                                           // NOLINT
+    std::string url     = ""; // model url to download                                      // NOLINT
+    std::string hf_repo = ""; // HF repo                                                    // NOLINT
+    std::string hf_file = ""; // HF file                                                    // NOLINT
+};
+
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
@@ -186,19 +200,13 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model = "";     // draft model for speculative decoding                      // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    struct common_params_model model;
 };
 
 struct common_params_vocoder {
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
+    struct common_params_model model;
 
-    std::string model     = ""; // model path                                                // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    std::string speaker_file = ""; // speaker file path                                      // NOLINT
 
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
@@ -254,13 +262,12 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    std::string model                = ""; // model path                                                    // NOLINT
+    struct common_params_model model;
+
     std::string model_alias          = ""; // model alias                                                   // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
+    std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
@@ -272,6 +279,7 @@ struct common_params {
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
     std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
@@ -325,13 +333,15 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
 
+    bool single_turn       = false; // single turn chat conversation
+
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    struct common_params_model mmproj;
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -391,8 +401,6 @@ struct common_params {
     int32_t i_pos  = -1;  // position of the passkey in the junk text
 
     // imatrix params
-    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
-
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
@@ -404,16 +412,16 @@ struct common_params {
     int n_pca_batch = 100;
     int n_pca_iterations = 1000;
     dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_outfile       = "control_vector.gguf";
     std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 
-    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
-
     // batched-bench params
     bool batched_bench_output_jsonl = false;
+
+    // common params
+    std::string out_file; // output filename for all example programs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -453,6 +461,8 @@ std::string string_repeat(const std::string & str, size_t n);
 
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
 
+std::string regex_escape(const std::string & s);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
@@ -530,23 +540,6 @@ struct llama_model_params     common_model_params_to_llama  (      common_params
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
-struct llama_model * common_load_model_from_url(
-    const std::string & model_url,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-struct llama_model * common_load_model_from_hf(
-    const std::string & repo,
-    const std::string & remote_path,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-std::pair<std::string, std::string> common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & hf_token);
-
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 

common/json-schema-to-grammar.cpp

@@ -264,7 +264,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     throw std::runtime_error("At least one of min_value or max_value must be set");
 }
 
-const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
+const std::string SPACE_RULE = "| \" \" | \"\\n\"{1,2} [ \\t]{0,20}";
 
 struct BuiltinRule {
     std::string content;
@@ -764,11 +764,10 @@ private:
 public:
     SchemaConverter(
         const std::function<json(const std::string &)> & fetch_json,
-        bool dotall,
-        bool compact_spaces)
+        bool dotall)
           : _fetch_json(fetch_json), _dotall(dotall)
     {
-        _rules["space"] = compact_spaces ? "\" \"?" : SPACE_RULE;
+        _rules["space"] = SPACE_RULE;
     }
 
     void resolve_refs(json & schema, const std::string & url) {
@@ -1007,7 +1006,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 }
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
-    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall, options.compact_spaces);
+    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall);
     common_grammar_builder builder {
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);

common/json-schema-to-grammar.h

@@ -16,7 +16,6 @@ struct common_grammar_builder {
 
 struct common_grammar_options {
     bool dotall = false;
-    bool compact_spaces = false;
 };
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

common/llguidance.cpp

@@ -11,25 +11,24 @@ struct llama_sampler_llg {
     std::string         grammar_kind;
     std::string         grammar_data;
     LlgTokenizer *      tokenizer;
-    LlgConstraint *     grammar;
-    LlgMaskResult       llg_res;
-    bool                has_llg_res;
+    LlgMatcher *        grammar;
 };
 
-static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                             const char * grammar_data) {
+static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                          const char * grammar_data) {
     LlgConstraintInit cinit;
     llg_constraint_init_set_defaults(&cinit, tokenizer);
     const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
     if (log_level && *log_level) {
         cinit.log_stderr_level = atoi(log_level);
     }
-    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
-    if (llg_get_error(c)) {
-        LOG_ERR("llg error: %s\n", llg_get_error(c));
-        llg_free_constraint(c);
+    auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data);
+    if (llg_matcher_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_matcher_get_error(c));
+        llg_free_matcher(c);
         return nullptr;
     }
+
     return c;
 }
 
@@ -40,39 +39,29 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
 static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        LlgCommitResult res;
-        llg_commit_token(ctx->grammar, token, &res);
-        ctx->has_llg_res = false;
+        llg_matcher_consume_token(ctx->grammar, token);
     }
 }
 
 static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        if (!ctx->has_llg_res) {
-            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
-                ctx->has_llg_res = true;
+        const uint32_t * mask = llg_matcher_get_mask(ctx->grammar);
+        if (mask == nullptr) {
+            if (llg_matcher_compute_mask(ctx->grammar) == 0) {
+                mask = llg_matcher_get_mask(ctx->grammar);
             } else {
-                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
-                llg_free_constraint(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
+                llg_free_matcher(ctx->grammar);
                 ctx->grammar = nullptr;
+                return;
             }
         }
-        if (ctx->has_llg_res) {
-            if (ctx->llg_res.is_stop) {
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
-            } else {
-                const uint32_t * mask = ctx->llg_res.sample_mask;
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    auto token = cur_p->data[i].id;
-                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            auto token = cur_p->data[i].id;
+            if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                cur_p->data[i].logit = -INFINITY;
             }
         }
     }
@@ -80,14 +69,9 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array
 
 static void llama_sampler_llg_reset(llama_sampler * smpl) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
-    if (!ctx->grammar) {
-        return;
+    if (ctx->grammar) {
+        llg_matcher_reset(ctx->grammar);
     }
-
-    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
-    llg_free_constraint(ctx->grammar);
-    ctx->grammar     = grammar_new;
-    ctx->has_llg_res = false;
 }
 
 static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
@@ -102,7 +86,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
         if (ctx->grammar) {
             result_ctx->grammar_kind = ctx->grammar_kind;
             result_ctx->grammar_data = ctx->grammar_data;
-            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
+            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
             result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
         }
     }
@@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_llg *) smpl->ctx;
 
     if (ctx->grammar) {
-        llg_free_constraint(ctx->grammar);
+        llg_free_matcher(ctx->grammar);
         llg_free_tokenizer(ctx->tokenizer);
     }
 
@@ -239,9 +223,11 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ grammar_data,
             /* .tokenizer    = */ tokenizer,
             /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
+        if (ctx->grammar) {
+            GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 ==
+                        llg_matcher_get_mask_byte_size(ctx->grammar));
+        }
     } else {
         *ctx = {
             /* .vocab        = */ vocab,
@@ -249,15 +235,12 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ {},
             /* .tokenizer    = */ nullptr,
             /* .grammar      = */ nullptr,
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
     }
 
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx
-    );
+        /* .ctx   = */ ctx);
 }
 
 #else

common/minja/minja.hpp

@@ -240,7 +240,7 @@ public:
       auto index = key.get<int>();
       return array_->at(index < 0 ? array_->size() + index : index);
     } else if (object_) {
-      if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+      if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
       auto it = object_->find(key.primitive_);
       if (it == object_->end()) return Value();
       return it->second;
@@ -249,7 +249,7 @@ public:
   }
   void set(const Value& key, const Value& value) {
     if (!object_) throw std::runtime_error("Value is not an object: " + dump());
-    if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     (*object_)[key.primitive_] = value;
   }
   Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
@@ -1378,13 +1378,27 @@ struct ArgumentsExpression {
     }
 };
 
-static std::string strip(const std::string & s) {
-  auto start = s.find_first_not_of(" \t\n\r");
+static std::string strip(const std::string & s, const std::string & chars = "", bool left = true, bool right = true) {
+  auto charset = chars.empty() ? " \t\n\r" : chars;
+  auto start = left ? s.find_first_not_of(charset) : 0;
   if (start == std::string::npos) return "";
-  auto end = s.find_last_not_of(" \t\n\r");
+  auto end = right ? s.find_last_not_of(charset) : s.size() - 1;
   return s.substr(start, end - start + 1);
 }
 
+static std::vector<std::string> split(const std::string & s, const std::string & sep) {
+  std::vector<std::string> result;
+  size_t start = 0;
+  size_t end = s.find(sep);
+  while (end != std::string::npos) {
+    result.push_back(s.substr(start, end - start));
+    start = end + sep.length();
+    end = s.find(sep, start);
+  }
+  result.push_back(s.substr(start));
+  return result;
+}
+
 static std::string capitalize(const std::string & s) {
   if (s.empty()) return s;
   auto result = s;
@@ -1467,8 +1481,26 @@ public:
         } else if (obj.is_string()) {
           auto str = obj.get<std::string>();
           if (method->get_name() == "strip") {
-            vargs.expectArgs("strip method", {0, 0}, {0, 0});
-            return Value(strip(str));
+            vargs.expectArgs("strip method", {0, 1}, {0, 0});
+            auto chars = vargs.args.empty() ? "" : vargs.args[0].get<std::string>();
+            return Value(strip(str, chars));
+          } else if (method->get_name() == "lstrip") {
+            vargs.expectArgs("lstrip method", {0, 1}, {0, 0});
+            auto chars = vargs.args.empty() ? "" : vargs.args[0].get<std::string>();
+            return Value(strip(str, chars, /* left= */ true, /* right= */ false));
+          } else if (method->get_name() == "rstrip") {
+            vargs.expectArgs("rstrip method", {0, 1}, {0, 0});
+            auto chars = vargs.args.empty() ? "" : vargs.args[0].get<std::string>();
+            return Value(strip(str, chars, /* left= */ false, /* right= */ true));
+          } else if (method->get_name() == "split") {
+            vargs.expectArgs("split method", {1, 1}, {0, 0});
+            auto sep = vargs.args[0].get<std::string>();
+            auto parts = split(str, sep);
+            Value result = Value::array();
+            for (const auto& part : parts) {
+              result.push_back(Value(part));
+            }
+            return result;
           } else if (method->get_name() == "capitalize") {
             vargs.expectArgs("capitalize method", {0, 0}, {0, 0});
             return Value(capitalize(str));
@@ -2574,14 +2606,18 @@ inline std::shared_ptr<Context> Context::builtins() {
     auto & text = args.at("text");
     return text.is_null() ? text : Value(strip(text.get<std::string>()));
   }));
-  globals.set("lower", simple_function("lower", { "text" }, [](const std::shared_ptr<Context> &, Value & args) {
-    auto text = args.at("text");
-    if (text.is_null()) return text;
-    std::string res;
-    auto str = text.get<std::string>();
-    std::transform(str.begin(), str.end(), std::back_inserter(res), ::tolower);
-    return Value(res);
-  }));
+  auto char_transform_function = [](const std::string & name, const std::function<char(char)> & fn) {
+    return simple_function(name, { "text" }, [=](const std::shared_ptr<Context> &, Value & args) {
+      auto text = args.at("text");
+      if (text.is_null()) return text;
+      std::string res;
+      auto str = text.get<std::string>();
+      std::transform(str.begin(), str.end(), std::back_inserter(res), fn);
+      return Value(res);
+    });
+  };
+  globals.set("lower", char_transform_function("lower", ::tolower));
+  globals.set("upper", char_transform_function("upper", ::toupper));
   globals.set("default", Value::callable([=](const std::shared_ptr<Context> &, ArgumentsValue & args) {
     args.expectArgs("default", {2, 3}, {0, 1});
     auto & value = args.args[0];

common/ngram-cache.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <fstream>
 #include <thread>
+#include <algorithm>
 
 void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
                               std::vector<llama_token> & inp, int nnew, bool print_progress) {

common/sampling.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <unordered_map>
+#include <algorithm>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -159,17 +160,57 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
-        std::vector<const char *> trigger_words;
-        trigger_words.reserve(params.grammar_trigger_words.size());
-        for (const auto & str : params.grammar_trigger_words) {
-            trigger_words.push_back(str.word.c_str());
+        std::vector<std::string> patterns_at_start;
+        std::vector<std::string> patterns_anywhere;
+        std::vector<llama_token> trigger_tokens;
+        for (const auto & trigger : params.grammar_triggers) {
+            switch (trigger.type) {
+                case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
+                {
+                    const auto & word = trigger.value;
+                    patterns_anywhere.push_back(regex_escape(word));
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
+                {
+                    const auto & pattern = trigger.value;
+                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
+                {
+                    const auto token = trigger.token;
+                    trigger_tokens.push_back(token);
+                    break;
+                }
+                default:
+                    GGML_ASSERT(false && "unknown trigger type");
+            }
+        }
+
+        std::vector<std::string> trigger_patterns;
+        if (!patterns_at_start.empty()) {
+            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
+        }
+        if (!patterns_anywhere.empty()) {
+            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
+        }
+
+        std::vector<const char *> trigger_patterns_c;
+        trigger_patterns_c.reserve(trigger_patterns.size());
+        for (const auto & regex : trigger_patterns) {
+            trigger_patterns_c.push_back(regex.c_str());
         }
 
         grmr = params.grammar_lazy
-             ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
-                                               trigger_words.data(), trigger_words.size(),
-                                               params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size())
+             ? llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                                                        trigger_patterns_c.data(), trigger_patterns_c.size(),
+                                                        trigger_tokens.data(), trigger_tokens.size())
              :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
     }
 
     auto * result = new common_sampler {

common/speculative.cpp

@@ -5,6 +5,7 @@
 #include "sampling.h"
 
 #include <cstring>
+#include <algorithm>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -172,7 +173,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         prompt.clear();
     } else {
@@ -191,14 +192,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -180,7 +180,8 @@ class Model:
             extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
             missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
             if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                 f"Missing tensors: {missing}")
             else:
                 raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
                                  f"Missing tensors: {missing}\n"
@@ -528,6 +529,8 @@ class Model:
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
 
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
@@ -537,13 +540,13 @@ class Model:
                 if token in added_vocab:
                     # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
+                    if not added_tokens_decoder[i].normalized:
                         previous_token = token
                         token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
-                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
                         # NOTE: this was added for Gemma.
@@ -702,6 +705,15 @@ class Model:
         if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
             # ref: https://huggingface.co/Xenova/gpt-4o
             res = "gpt-4o"
+        if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f":
+            # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k
+            res = "superbpe"
+        if chkhsh == "1994ffd01900cfb37395608534236ecd63f2bd5995d6cb1004dda1af50240f15":
+            # ref: https://huggingface.co/trillionlabs/Trillion-7B-preview
+            res = "trillion"
+        if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224":
+            # ref: https://huggingface.co/inclusionAI/Ling-lite
+            res = "bailingmoe"
 
         if res is None:
             logger.warning("\n")
@@ -861,6 +873,9 @@ class Model:
                 for token_id, token_data in added_tokens_decoder.items():
                     token_id = int(token_id)
                     token: str = token_data["content"]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
                     if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         if tokens[token_id] != token.encode("utf-8"):
                             logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}')
@@ -905,6 +920,40 @@ class Model:
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_rwkv_world(self):
+        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
+        vocab_size = self.hparams.get("vocab_size", 65536)
+
+        tokens: list[bytes] = ['<s>'.encode("utf-8")]
+        toktypes: list[int] = [gguf.TokenType.CONTROL]
+
+        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
+            lines = f.readlines()
+            for line in lines:
+                parts = line.split(' ')
+                assert len(parts) >= 3
+                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
+                token = token.encode("utf-8") if isinstance(token, str) else token
+                assert isinstance(token, bytes)
+                assert len(token) == token_len
+                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
+                tokens.append(token_text.encode("utf-8"))
+                toktypes.append(gguf.TokenType.NORMAL)
+        remainder = vocab_size - len(tokens)
+        assert remainder >= 0
+        for i in range(len(tokens), vocab_size):
+            tokens.append(f"[PAD{i}]".encode("utf-8"))
+            toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("rwkv")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int):
         tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf"
         logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
@@ -1062,13 +1111,6 @@ class BloomModel(Model):
 
         tensors.append((self.map_tensor_name(name), data_torch))
 
-        if name == "word_embeddings.weight":
-            assert self.tensor_names is not None
-
-            # TODO: tie them at runtime, don't duplicate in the model file
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -1710,6 +1752,25 @@ class LlamaModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("Mistral3ForConditionalGeneration")
+class Mistral3Model(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        name = name.replace("language_model.", "")
+        if "multi_modal_projector" in name or "vision_tower" in name:
+            return []
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("DeciLMForCausalLM")
 class DeciModel(Model):
     model_arch = gguf.MODEL_ARCH.DECI
@@ -2214,7 +2275,7 @@ class Qwen2Model(Model):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
-@Model.register("Qwen2VLForConditionalGeneration")
+@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2VL
 
@@ -2367,10 +2428,6 @@ class GPT2Model(Model):
 
         tensors.append((new_name, data_torch))
 
-        # note: GPT2 output is tied to (same as) wte in original model
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -2700,21 +2757,26 @@ class CodeShellModel(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
+    _has_tok_embd = False
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
         new_name = self.map_tensor_name(name)
 
-        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+        # assuming token_embd.weight is seen before output.weight
+        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
+            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
+                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
+                self.tensor_names.remove("transformer.wte.weight")
+        elif new_name == tok_embd_name:
+            self._has_tok_embd = True
 
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            assert self.tensor_names is not None
-
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                # copy tok_embd.weight to output.weight
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
-        return tensors
+        return [(new_name, data_torch)]
 
 
 @Model.register("InternLM2ForCausalLM")
@@ -3322,6 +3384,83 @@ class Gemma2Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
+class Gemma3Model(Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA3
+    has_vision: bool = False
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+        if "vision_config" in hparams:
+            logger.info("Has vision encoder, but it will be ignored")
+            self.has_vision = True
+
+    def write(self):
+        super().write()
+        if self.has_vision:
+            logger.info("NOTE: this script only convert the language model to GGUF")
+            logger.info("      for the vision model, please use gemma3_convert_encoder_to_gguf.py")
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        # some default values are not specified in the hparams
+        self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072))
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8))
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6))
+        self.gguf_writer.add_key_length(hparams.get("head_dim", 256))
+        self.gguf_writer.add_value_length(hparams.get("head_dim", 256))
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
+        # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3
+        assert hparams.get("attn_logit_softcapping") is None
+        assert hparams.get("final_logit_softcapping") is None
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
+        if hparams.get("rope_scaling") is not None:
+            assert hparams["rope_scaling"]["rope_type"] == "linear"
+            # important: this rope_scaling is only applied for global layers, and not used by 1B model
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+        elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
+                or name.startswith("multimodal_projector.") or name.startswith("vision_model."): # this is for old HF model, should be removed later
+            # ignore vision tensors
+            return []
+
+        # remove OOV (out-of-vocabulary) rows in token_embd
+        if "embed_tokens.weight" in name:
+            vocab = self._create_vocab_sentencepiece()
+            tokens = vocab[0]
+            data_torch = data_torch[:len(tokens)]
+
+        # ref code in Gemma3RMSNorm
+        # output = output * (1.0 + self.weight.float())
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("Starcoder2ForCausalLM")
 class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
@@ -3332,38 +3471,7 @@ class Rwkv6Model(Model):
     model_arch = gguf.MODEL_ARCH.RWKV6
 
     def set_vocab(self):
-        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
-        vocab_size = self.hparams.get("vocab_size", 65536)
-
-        tokens: list[bytes] = ['<s>'.encode("utf-8")]
-        toktypes: list[int] = [gguf.TokenType.CONTROL]
-
-        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
-            lines = f.readlines()
-            for line in lines:
-                parts = line.split(' ')
-                assert len(parts) >= 3
-                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
-                token = token.encode("utf-8") if isinstance(token, str) else token
-                assert isinstance(token, bytes)
-                assert len(token) == token_len
-                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
-                tokens.append(token_text.encode("utf-8"))
-                toktypes.append(gguf.TokenType.NORMAL)
-        remainder = vocab_size - len(tokens)
-        assert remainder >= 0
-        for i in range(len(tokens), vocab_size):
-            tokens.append(f"[PAD{i}]".encode("utf-8"))
-            toktypes.append(gguf.TokenType.UNUSED)
-
-        self.gguf_writer.add_tokenizer_model("rwkv")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
-        # hack: Add '\n\n' as the EOT token to make it chat normally
-        special_vocab._set_special_token("eot", 261)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        self._set_vocab_rwkv_world()
 
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
@@ -3449,8 +3557,8 @@ class RWKV6Qwen2Model(Rwkv6Model):
         head_size = hidden_size // num_attention_heads
         rms_norm_eps = self.hparams["rms_norm_eps"]
         intermediate_size = self.hparams["intermediate_size"]
-        time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
-        time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
+        time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32)
+        time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64)
 
         # RWKV isn't context limited
         self.gguf_writer.add_context_length(1048576)
@@ -3485,6 +3593,168 @@ class RWKV6Qwen2Model(Rwkv6Model):
             yield (new_name, data)
 
 
+@Model.register("Rwkv7ForCausalLM", "RWKV7ForCausalLM")
+class Rwkv7Model(Model):
+    model_arch = gguf.MODEL_ARCH.RWKV7
+
+    def set_vocab(self):
+        self._set_vocab_rwkv_world()
+
+    def calc_lora_rank(self, hidden_size, exponent, multiplier):
+        return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        try:
+            head_size = self.hparams["head_size"]
+            layer_norm_eps = self.hparams["layer_norm_epsilon"]
+        except KeyError:
+            head_size = self.hparams["head_dim"]
+            layer_norm_eps = self.hparams["norm_eps"]
+        hidden_size = self.hparams["hidden_size"]
+        intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else (hidden_size * 4)
+
+        # ICLR: In-Context-Learning-Rate
+        try:
+            lora_rank_decay = self.hparams["lora_rank_decay"] if self.hparams["lora_rank_decay"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["lora_rank_iclr"] if self.hparams["lora_rank_iclr"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["lora_rank_value_residual_mix"] if self.hparams["lora_rank_value_residual_mix"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["lora_rank_gate"] if self.hparams["lora_rank_gate"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+        except KeyError:
+            lora_rank_decay = self.hparams["decay_low_rank_dim"] if self.hparams["decay_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["a_low_rank_dim"] if self.hparams["a_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["v_low_rank_dim"] if self.hparams["v_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["gate_low_rank_dim"] if self.hparams["gate_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    lerp_weights: dict[int, dict[str, Tensor]] = {}
+    lora_needs_transpose: bool = True
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # unify tensor names here to make life easier
+        name = name.replace("blocks", "layers").replace("ffn", "feed_forward")
+        name = name.replace("self_attn", "attention").replace("attn", "attention")
+        name = name.replace("time_mixer.", "")
+        # lora layer names in fla-hub's impl
+        if "_lora.lora" in name:
+            self.lora_needs_transpose = False
+        name = name.replace("_lora.lora.0.weight", "1.weight")
+        name = name.replace("_lora.lora.2.weight", "2.weight")
+        name = name.replace("_lora.lora.2.bias", "0.weight")
+
+        name = name.replace("feed_forward_norm", "ln2")
+        name = name.replace("g_norm", "ln_x")
+
+        if "attention.v" in name and "value" not in self.map_tensor_name(name) and bid == 0:
+            # some models have dummy v0/v1/v2 on first layer while others don't
+            # ignore them all since they are not used
+            return
+
+        wkv_has_gate = self.hparams.get("wkv_has_gate", True)
+        lerp_list = ["r", "w", "k", "v", "a", "g"] if wkv_has_gate else ["r", "w", "k", "v", "a"]
+
+        if bid is not None and "attention.x_" in name:
+            if "attention.x_x" in name:
+                # already concatenated
+                new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                data = data_torch.reshape(len(lerp_list), 1, 1, -1)
+                yield (new_name, data)
+            else:
+                try:
+                    self.lerp_weights[bid][name] = data_torch
+                except KeyError:
+                    self.lerp_weights[bid] = {name: data_torch}
+                if all(f"model.layers.{bid}.attention.x_{i}" in self.lerp_weights[bid].keys() for i in lerp_list):
+                    new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                    data = torch.stack([self.lerp_weights[bid][f"model.layers.{bid}.attention.x_{i}"] for i in lerp_list], dim=0)
+                    yield (new_name, data)
+            return
+        else:
+            data_torch = data_torch.squeeze()
+            new_name = self.map_tensor_name(name)
+
+            if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
+                new_name += ".weight"
+
+            if self.lora_needs_transpose and any(
+                new_name.endswith(t) for t in [
+                    "time_mix_w1.weight", "time_mix_w2.weight",
+                    "time_mix_a1.weight", "time_mix_a2.weight",
+                    "time_mix_v1.weight", "time_mix_v2.weight",
+                    "time_mix_g1.weight", "time_mix_g2.weight",
+                ]
+            ):
+                data_torch = data_torch.transpose(0, 1)
+
+            if 'r_k' in new_name:
+                data_torch = data_torch.flatten()
+
+            if bid == 0 and "time_mix_a" in new_name:
+                # dummy v0/v1/v2 on first layer
+                # easist way to make llama happy
+                yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch)
+
+            yield (new_name, data_torch)
+
+
+@Model.register("RwkvHybridForCausalLM")
+class ARwkv7Model(Rwkv7Model):
+    model_arch = gguf.MODEL_ARCH.ARWKV7
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        hidden_size = self.hparams["hidden_size"]
+        head_size = self.hparams["head_size"]
+        rms_norm_eps = self.hparams["rms_norm_eps"]
+        intermediate_size = self.hparams["intermediate_size"]
+        wkv_has_gate = self.hparams["wkv_has_gate"]
+        assert self.hparams["wkv_version"] == 7
+
+        # ICLR: In-Context-Learning-Rate
+        lora_rank_decay = 64
+        lora_rank_iclr = 64
+        lora_rank_value_residual_mix = 32
+        lora_rank_gate = 128 if wkv_has_gate else 0
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_token_shift_count(1)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+
 @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
@@ -3539,8 +3809,6 @@ class MambaModel(Model):
     _tok_embd = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
         tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
 
@@ -3550,6 +3818,10 @@ class MambaModel(Model):
             logger.debug("A_log --> A ==> " + new_name)
             data_torch = -torch.exp(data_torch)
 
+        # [4 1 8192 1] -> [4 8192 1 1]
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
         # assuming token_embd.weight is seen before output.weight
         if self._tok_embd is not None and new_name == output_name:
             if torch.equal(self._tok_embd, data_torch):
@@ -4153,6 +4425,29 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("PLMForCausalLM")
+class PLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLM
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+
 @Model.register("T5WithLMHeadModel")
 @Model.register("T5ForConditionalGeneration")
 @Model.register("MT5ForConditionalGeneration")
@@ -4841,6 +5136,105 @@ class GraniteMoeModel(GraniteModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@Model.register("BailingMoeForCausalLM")
+class BailingMoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
+
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_weights_scale(1.0)
+        self.gguf_writer.add_expert_count(hparams["num_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
+        if n_head_kv is not None and n_head != n_head_kv:
+            n_head = n_head_kv
+        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_embd = self.hparams["hidden_size"]
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
+
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+
+        if name.endswith("attention.dense.weight"):
+            return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)]
+        elif name.endswith("query_key_value.weight"):
+            q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2)
+
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+            ]
+        elif name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            tensors: list[tuple[str, Tensor]] = []
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+
+            return tensors
+
+        new_name = self.map_tensor_name(name)
+
+        if new_name == output_name and self.hparams.get("norm_head"):
+            data_torch = data_torch.float()
+            data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("ChameleonForConditionalGeneration")
 @Model.register("ChameleonForCausalLM")  # obsolete
 class ChameleonModel(Model):
@@ -5094,7 +5488,7 @@ def main() -> None:
             logger.error(f"Model {model_architecture} is not supported")
             sys.exit(1)
 
-        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+        model_instance = model_class(dir_model, output_type, fname_out,
                                      is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
                                      eager=args.no_lazy,
                                      metadata_override=args.metadata, model_name=args.model_name,

convert_hf_to_gguf_update.py

@@ -110,6 +110,9 @@ models = [
     {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
+    {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
+    {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
+    {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
 ]
 
 

docs/backend/CUDA-FEDORA.md

@@ -14,9 +14,7 @@ In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox
 - [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
 - [Installing Essential Development Tools](#installing-essential-development-tools)
 - [Adding the CUDA Repository](#adding-the-cuda-repository)
-- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
-- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
-- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
+- [Installing Nvidia Driver Libraries](#installing-nvidia-driver-libraries)
 - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
 - [Configuring the Environment](#configuring-the-environment)
 - [Verifying the Installation](#verifying-the-installation)
@@ -67,7 +65,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for
    sudo dnf distro-sync
    ```
 
-2. **Install the Default Text Editor (Optional):**
+2. **Install **Vim** the default text editor (Optional):**
 
    ```bash
    sudo dnf install vim-default-editor --allowerasing
@@ -97,36 +95,48 @@ After adding the repository, synchronize the package manager again:
 sudo dnf distro-sync
 ```
 
-## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+## Installing Nvidia Driver Libraries
 
-We need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go).
+First, we need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go):
 
 ```bash
 ls -la /usr/lib64/libcuda.so.1
 ```
 
+### If *`libcuda.so.1`* is missing:
+
+```
+ls: cannot access '/usr/lib64/libcuda.so.1': No such file or directory
+```
+
 **Explanation:**
+The host dose not supply the CUDA drivers, **install them now:**
 
-- `nvidia-driver-libs` and `nvidia-driver-cuda-libs` contains necessary NVIDIA driver libraries required by CUDA,
-  on hosts with NVIDIA drivers installed the Fedora Container will supply the host libraries.
-
-### Install Nvidia Driver Libraries on Guest (if `libcuda.so.1` was NOT found).
+#### Install the Nvidia Driver Libraries on Guest:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found).
+### If *`libcuda.so.1`* exists:
+```
+lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07
+```
 
-If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
+**Explanation:**
+The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:**
 
-#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies)
+#### Update the Toolbox RPM Database to include the Host-Supplied Libraries:
+
+Note: we do not actually install the libraries, we just update the DB so that the guest system knows they are supplied by the host.
+
+##### 1. Download `nvidia-` parts that are supplied by the host RPM's (with dependencies)
 
 ```bash
-sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-#### 2. Update the RPM database to assume the installation of these packages.
+##### 2. Update the RPM database to assume the installation of these packages.
 
 ```bash
 sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
@@ -134,23 +144,26 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
 
 **Note:**
 
-- The `--justdb` option only updates the RPM database, without touching the filesystem.
+- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere.
 
-#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+##### Check that the RPM Database has been correctly updated:
+
+**Note:** This is the same command as in the *"Install the Nvidia Driver Libraries on Guest"* for if *`libcuda.so.1`* was missing.
 
-After manually installing the dependencies, run:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-You should receive a message indicating the package is already installed:
+*(this time it will not install anything, as the database things that these packages are already installed)*
 
 ```
 Updating and loading repositories:
 Repositories loaded.
-Package "nvidia-driver-libs-3:570.86.10-1.fc41.x86_64" is already installed.
-Package "nvidia-driver-cuda-libs-3:570.86.10-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-persistenced-3:570.124.06-1.fc41.x86_64" is already installed.
 
 Nothing to do.
 ```
@@ -207,9 +220,9 @@ You should see output similar to:
 ```
 nvcc: NVIDIA (R) Cuda compiler driver
 Copyright (c) 2005-2025 NVIDIA Corporation
-Built on Wed_Jan_15_19:20:09_PST_2025
-Cuda compilation tools, release 12.8, V12.8.61
-Build cuda_12.8.r12.8/compiler.35404655_0
+Built on Fri_Feb_21_20:23:50_PST_2025
+Cuda compilation tools, release 12.8, V12.8.93
+Build cuda_12.8.r12.8/compiler.35583870_0
 ```
 
 This output confirms that the CUDA compiler is accessible and indicates the installed version.

docs/backend/OPENCL.md

@@ -145,8 +145,13 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
 * Clang 19
 * Ninja
 * Visual Studio 2022
+* Powershell 7
 
-Powershell is used for the following instructions.
+Visual Studio provides necessary headers and libraries although it is not directly used for building.
+Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
+
+Powershell 7 is used for the following commands.
+If an older version of Powershell is used, these commands may not work as they are.
 
 ### I. Setup Environment
 
@@ -196,10 +201,9 @@ ninja
 
 ## Known Issues
 
-- Qwen2.5 0.5B model produces gibberish output with Adreno kernels.
+- Currently OpenCL backend does not work on Adreno 6xx GPUs.
 
 ## TODO
 
-- Fix Qwen2.5 0.5B
 - Optimization for Q6_K
 - Support and optimization for Q4_K

docs/backend/SYCL.md

@@ -20,7 +20,7 @@
 **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
-- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
@@ -227,30 +227,19 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 **oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
 
-
-**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
+**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
 
 ```sh
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
-cmake --build buildWithCublas --config Release
+git clone https://github.com/oneapi-src/oneDNN.git
+cd oneDNN
+cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake --build build-nvidia --config Release
 ```
 
 - **Adding support to AMD GPUs**
 
 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
 
-**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
-
-```sh
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-# Find your HIPTARGET with rocminfo, under the key 'Name:'
-cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
-cmake --build buildWithrocBLAS --config Release
-```
-
 3. **Verify installation and environment**
 
 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -313,37 +302,39 @@ cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -
 cmake --build build --config Release -j -v
 ```
 
+It is possible to come across some precision issues when running tests that stem from using faster
+instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS`
+as `-cl-fp32-correctly-rounded-divide-sqrt`
+
 #### Nvidia GPU
 
-```sh
-# Export relevant ENV variables
-export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
-export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
+By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
 
+```sh
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 # Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # build all binary
 cmake --build build --config Release -j -v
 ```
 
+It is possible to come across some precision issues when running tests that stem from using faster
+instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler.
+
 #### AMD GPU
 
-```sh
-# Export relevant ENV variables
-export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
-export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
+The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
+By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
 
+```sh
 # Build LLAMA with rocBLAS acceleration through SYCL
 
 ## AMD
@@ -484,6 +475,12 @@ b. Enable oneAPI running environment:
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
+- if you are using Powershell, enable the runtime environment with the following:
+
+```
+cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'
+```
+
 c. Verify installation
 
 In the oneAPI command line, run the following to print the available SYCL devices:
@@ -514,13 +511,13 @@ You could download the release package for Windows directly, which including bin
 
 Choose one of following methods to build from source code.
 
-1. Script
+#### 1. Script
 
 ```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-2. CMake
+#### 2. CMake
 
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
@@ -549,13 +546,84 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-3. Visual Studio
+#### 3. Visual Studio
 
-You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+You have two options to use Visual Studio to build llama.cpp:
+- As CMake Project using CMake presets.
+- Creating a Visual Studio solution to handle the project.
+
+**Note**:
+
+All following commands are executed in PowerShell.
+
+##### - Open as a CMake Project
+
+You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets:
+
+- `x64-windows-sycl-release`
+
+- `x64-windows-sycl-debug`
 
 *Notes:*
+- For a minimal experimental setup, you can build only the inference executable using:
 
-- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
+    ```Powershell
+    cmake --build build --config Release -j --target llama-cli
+    ```
+
+##### - Generating a Visual Studio Solution
+
+You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file.
+
+If you want to use the Intel C++ Compiler for the entire `llama.cpp` project, run the following command:
+
+```Powershell
+cmake -B build -G "Visual Studio 17 2022" -T "Intel C++ Compiler 2025" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release
+```
+
+If you prefer to use the Intel C++ Compiler only for `ggml-sycl`, ensure that `ggml` and its backend libraries are built as shared libraries ( i.e. `-DBUILD_SHARED_LIBRARIES=ON`, this is default behaviour):
+
+```Powershell
+cmake -B build -G "Visual Studio 17 2022" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release \
+      -DSYCL_INCLUDE_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\include" \
+      -DSYCL_LIBRARY_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\lib"
+```
+
+If successful the build files have been written to: *path/to/llama.cpp/build*
+Open the project file **build/llama.cpp.sln** with Visual Studio.
+
+Once the Visual Studio solution is created, follow these steps:
+
+1. Open the solution in Visual Studio.
+
+2. Right-click on `ggml-sycl` and select **Properties**.
+
+3. In the left column, expand **C/C++** and select **DPC++**.
+
+4. In the right panel, find **Enable SYCL Offload** and set it to `Yes`.
+
+5. Apply the changes and save.
+
+
+*Navigation Path:*
+
+```
+Properties -> C/C++ -> DPC++ -> Enable SYCL Offload (Yes)
+```
+
+Now, you can build `llama.cpp` with the SYCL backend as a Visual Studio project.
+To do it from menu: `Build -> Build Solution`.
+Once it is completed, final results will be in **build/Release/bin**
+
+*Additional Note*
+
+- You can avoid specifying `SYCL_INCLUDE_DIR` and `SYCL_LIBRARY_DIR` in the CMake command by setting the environment variables:
+
+    - `SYCL_INCLUDE_DIR_HINT`
+
+    - `SYCL_LIBRARY_DIR_HINT`
+
+- Above instruction has been tested with Visual Studio 17 Community edition and oneAPI 2025.0. We expect them to work also with future version if the instructions are adapted accordingly.
 
 ### III. Run the inference
 
@@ -660,8 +728,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
-| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)          | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
+| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
@@ -671,6 +740,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 

docs/build.md

@@ -132,12 +132,14 @@ You may find the official downloads here: [NVIDIA developer site](https://develo
 
 
 #### Compile and run inside a Fedora Toolbox Container
-We also have a [guide](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
+We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
 
 **Recommended for:**
-
-- ***Particularly*** *convenient* for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
-- Toolbox is installed by default: [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde).
+- ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
+  - (there are no supported CUDA packages for these systems)
+- ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads).
+  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system)
+- ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean.
 - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download)
 
 
@@ -189,7 +191,7 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
@@ -197,29 +199,54 @@ The following compilation options are also available to tweak performance:
 
 ## MUSA
 
-This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
+This provides GPU acceleration using a Moore Threads GPU. Make sure to have the [MUSA SDK](https://developer.mthreads.com/musa/musa-sdk) installed.
 
-- Using `CMake`:
+#### Download directly from Moore Threads
 
-  ```bash
-  cmake -B build -DGGML_MUSA=ON
-  cmake --build build --config Release
+You may find the official downloads here: [Moore Threads developer site](https://developer.mthreads.com/sdk/download/musa).
+
+### Compilation
+
+```bash
+cmake -B build -DGGML_MUSA=ON
+cmake --build build --config Release
+```
+
+#### Override Compute Capability Specifications
+
+By default, all supported compute capabilities are enabled. To customize this behavior, you can specify the `MUSA_ARCHITECTURES` option in the CMake command:
+
+```bash
+cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
+cmake --build build --config Release
+```
+
+This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.
+
+#### Compilation options
+
+Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
+
+- For static builds, add `-DBUILD_SHARED_LIBS=OFF` and `-DCMAKE_POSITION_INDEPENDENT_CODE=ON`:
   ```
-
-  For static build:
-
-  ```bash
   cmake -B build -DGGML_MUSA=ON \
     -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
   cmake --build build --config Release
   ```
 
-The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
+### Runtime MUSA environmental variables
+
+You may set the [musa environmental variables](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) at runtime.
+
+```bash
+# Use `MUSA_VISIBLE_DEVICES` to hide the first compute device.
+MUSA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf
+```
+
+### Unified Memory
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
 
-Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
-
 ## HIP
 
 This provides GPU acceleration on HIP-supported AMD GPUs.
@@ -235,6 +262,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
   However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
+  To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
+
+  The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager.
+
+  As an alternative, you can manually install the library by cloning it from the official [GitHub repository](https://github.com/ROCm/rocWMMA), checkout the corresponding version tag (e.g. `rocm-6.2.4`) and set `-DCMAKE_CXX_FLAGS="-I<path/to/rocwmma>/library/include/"` in CMake. This also works under Windows despite not officially supported by AMD.
+
   Note that if you get the following error:
   ```
   clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
@@ -403,6 +436,116 @@ llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
 
 For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
 
+## Arm® KleidiAI™
+KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
+
+To enable KleidiAI, go to the llama.cpp directory and build using CMake
+```bash
+cmake -B build -DGGML_CPU_KLEIDIAI=ON
+cmake --build build --config Release
+```
+You can verify that KleidiAI is being used by running
+```bash
+./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
+```
+If KleidiAI is enabled, the ouput will contain a line similar to:
+```
+load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
+```
+KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
+
+Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
+
+## OpenCL
+
+This provides GPU acceleration through OpenCL on recent Adreno GPU.
+More information about OpenCL backend can be found in [OPENCL.md](./backend/OPENCL.md) for more information.
+
+### Android
+
+Assume NDK is available in `$ANDROID_NDK`. First, install OpenCL headers and ICD loader library if not available,
+
+```sh
+mkdir -p ~/dev/llm
+cd ~/dev/llm
+
+git clone https://github.com/KhronosGroup/OpenCL-Headers && \
+cd OpenCL-Headers && \
+cp -r CL $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+
+cd ~/dev/llm
+
+git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \
+cd OpenCL-ICD-Loader && \
+mkdir build_ndk && cd build_ndk && \
+cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+  -DOPENCL_ICD_LOADER_HEADERS_DIR=$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=24 \
+  -DANDROID_STL=c++_shared && \
+ninja && \
+cp libOpenCL.so $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+```
+
+Then build llama.cpp with OpenCL enabled,
+
+```sh
+cd ~/dev/llm
+
+git clone https://github.com/ggml-org/llama.cpp && \
+cd llama.cpp && \
+mkdir build-android && cd build-android
+
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DBUILD_SHARED_LIBS=OFF \
+  -DGGML_OPENCL=ON
+
+ninja
+```
+
+### Windows Arm64
+
+First, install OpenCL headers and ICD loader library if not available,
+
+```powershell
+mkdir -p ~/dev/llm
+
+cd ~/dev/llm
+git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers
+mkdir build && cd build
+cmake .. -G Ninja `
+  -DBUILD_TESTING=OFF `
+  -DOPENCL_HEADERS_BUILD_TESTING=OFF `
+  -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
+  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
+cmake --build . --target install
+
+cd ~/dev/llm
+git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader
+mkdir build && cd build
+cmake .. -G Ninja `
+  -DCMAKE_BUILD_TYPE=Release `
+  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
+  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
+cmake --build . --target install
+```
+
+Then build llama.cpp with OpenCL enabled,
+
+```powershell
+cmake .. -G Ninja `
+  -DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" `
+  -DCMAKE_BUILD_TYPE=Release `
+  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
+  -DBUILD_SHARED_LIBS=OFF `
+  -DGGML_OPENCL=ON
+ninja
+```
+
 ## Android
 
 To read documentation for how to build on Android, [click here](./android.md)

docs/function-calling.md

@@ -287,30 +287,32 @@ Here are some models known to work (w/ chat template override when needed):
 
 llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
 llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
-llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
 llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
 
-# Native support for DeepSeek R1 works best w/ our own template (official template buggy)
+# Native support for DeepSeek R1 works best w/ our template override (official template is buggy, although we do work around it)
 
 llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
---chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
 
 llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
---chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
 
 # Native support requires the right template for these GGUFs:
 
+llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
+    --chat-template-file models/templates/meetkai-functionary-medium-v3.2.jinja
+
 llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \
---chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use )
+    --chat-template-file models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja
 
 llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \
---chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use )
+    --chat-template-file models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja
 
 llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \
---chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use )
+    --chat-template-file models/templates/fireworks-ai-llama-3-firefunction-v2.jinja
 
 llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \
---chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use )
+    --chat-template-file models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja
 
 # Generic format support
 llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0
@@ -318,6 +320,8 @@ llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0
 llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K
 ```
 
+To get the official template from original HuggingFace repos, you can use [scripts/get_chat_template.py](../scripts/get_chat_template.py) (see examples invocations in [models/templates/README.md](../models/templates/README.md))
+
 > [!TIP]
 > If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills)
 

docs/install.md

@@ -9,6 +9,13 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
+## MacPorts
+
+```sh
+sudo port install llama.cpp
+```
+see also: https://ports.macports.org/port/llama.cpp/details/
+
 ## Nix
 
 On Mac and Linux, the Nix package manager can be used via

examples/batched-bench/batched-bench.cpp

@@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@@ -132,7 +132,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_clear(ctx);
+                llama_kv_self_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +141,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/batched/batched.cpp

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

examples/cvector-generator/cvector-generator.cpp

@@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
@@ -394,6 +394,8 @@ static int prepare_entries(common_params & params, train_context & ctx_train) {
 int main(int argc, char ** argv) {
     common_params params;
 
+    params.out_file = "control_vector.gguf";
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
@@ -498,7 +500,7 @@ int main(int argc, char ** argv) {
     }
 
     // write output vectors to gguf
-    export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
+    export_gguf(ctx_train.v_final, params.out_file, model_hint);
 
     llama_backend_free();
 

examples/embedding/embedding.cpp

@@ -4,6 +4,7 @@
 #include "llama.h"
 
 #include <ctime>
+#include <algorithm>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -37,7 +38,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const struct llama_model * model = llama_get_model(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/export-lora/export-lora.cpp

@@ -413,20 +413,22 @@ static void print_usage(int, char ** argv) {
 int main(int argc, char ** argv) {
     common_params params;
 
+    params.out_file = "ggml-lora-merged-f16.gguf";
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
         return 1;
     }
 
     g_verbose = (params.verbosity > 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
+        lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());
         exit(EXIT_FAILURE);
     }
 
-    printf("done, output file is %s\n", params.lora_outfile.c_str());
+    printf("done, output file is %s\n", params.out_file.c_str());
 
     return 0;
 }

examples/gguf-split/gguf-split.cpp

@@ -408,8 +408,6 @@ static void gguf_merge(const split_params & split_params) {
         exit(EXIT_FAILURE);
     }
 
-    std::ofstream fout(split_params.output.c_str(), std::ios::binary);
-    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
 
     auto * ctx_out = gguf_init_empty();
 
@@ -453,7 +451,6 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
-                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -466,7 +463,6 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
-                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -479,7 +475,6 @@ static void gguf_merge(const split_params & split_params) {
                 gguf_free(ctx_gguf);
                 ggml_free(ctx_meta);
                 gguf_free(ctx_out);
-                fout.close();
                 exit(EXIT_FAILURE);
             }
 
@@ -500,9 +495,11 @@ static void gguf_merge(const split_params & split_params) {
 
         fprintf(stderr, "\033[3Ddone\n");
     }
-
-    // placeholder for the meta data
-    {
+    std::ofstream fout;
+    if (!split_params.dry_run) {
+        fout.open(split_params.output.c_str(), std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+        // placeholder for the meta data
         auto meta_size = gguf_get_meta_size(ctx_out);
         ::zeros(fout, meta_size);
     }
@@ -518,7 +515,9 @@ static void gguf_merge(const split_params & split_params) {
                 ggml_free(ctx_metas[i]);
             }
             gguf_free(ctx_out);
-            fout.close();
+            if (!split_params.dry_run) {
+                fout.close();
+            }
             exit(EXIT_FAILURE);
         }
         fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
@@ -540,10 +539,11 @@ static void gguf_merge(const split_params & split_params) {
             auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
             f_input.seekg(offset);
             f_input.read((char *)read_data.data(), n_bytes);
-
-            // write tensor data + padding
-            fout.write((const char *)read_data.data(), n_bytes);
-            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            if (!split_params.dry_run) {
+                // write tensor data + padding
+                fout.write((const char *)read_data.data(), n_bytes);
+                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            }
         }
 
         gguf_free(ctx_gguf);
@@ -552,16 +552,15 @@ static void gguf_merge(const split_params & split_params) {
         fprintf(stderr, "\033[3Ddone\n");
     }
 
-    {
+    if (!split_params.dry_run) {
         // go back to beginning of file and write the updated metadata
         fout.seekp(0);
         std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
         gguf_get_meta_data(ctx_out, data.data());
         fout.write((const char *)data.data(), data.size());
-
         fout.close();
-        gguf_free(ctx_out);
     }
+    gguf_free(ctx_out);
 
     fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
             __func__, split_params.output.c_str(), n_split, total_tensors);

examples/gritlm/gritlm.cpp

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
@@ -168,7 +168,7 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
 
     // create generation context
     llama_context * ctx = llama_init_from_model(model, cparams);

examples/imatrix/imatrix.cpp

@@ -206,9 +206,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
 void IMatrixCollector::save_imatrix(int ncall) const {
     auto fname = m_params.out_file;
-    if (fname.empty()) {
-        fname = "imatrix.dat";
-    }
 
     if (ncall > 0) {
         fname += ".at_";
@@ -498,7 +495,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -583,6 +580,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
 int main(int argc, char ** argv) {
     common_params params;
 
+    params.out_file = "imatrix.dat" ;
+
     params.n_ctx = 512;
     params.logits_all = true;
     params.escape = false;

examples/infill/infill.cpp

@@ -332,8 +332,8 @@ int main(int argc, char ** argv) {
                 LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json_schema_to_grammar.py

@@ -195,7 +195,7 @@ class BuiltinRule:
         self.deps = deps or []
 
 # Constraining spaces to prevent model "running away".
-SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}'
+SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}'
 
 PRIMITIVE_RULES = {
     'boolean'      : BuiltinRule('("true" | "false") space', []),

examples/llama-bench/llama-bench.cpp

@@ -1578,7 +1578,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // cool off before the test
         if (params.delay) {
@@ -1618,7 +1618,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_self_clear(ctx);
 
             uint64_t t_start = get_time_ns();
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -361,7 +361,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     const auto tokens_list = common_tokenize(context, text, true, parse_special);
 
     auto n_ctx = llama_n_ctx(context);
-    auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
+    auto n_kv_req = tokens_list.size() + n_len;
 
     LOGi("n_len = %d, n_ctx = %d, n_kv_req = %d", n_len, n_ctx, n_kv_req);
 
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
+    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.swiftui/README.md

@@ -5,6 +5,21 @@ point for more advanced projects.
 
 For usage instructions and performance stats, check the following discussion: https://github.com/ggml-org/llama.cpp/discussions/4508
 
+
+### Building
+First llama.cpp need to be built and a XCFramework needs to be created. This can be done by running
+the following script from the llama.cpp project root:
+```console
+$ ./build-xcframework.sh
+```
+Open `llama.swiftui.xcodeproj` project in Xcode and you should be able to build and run the app on
+a simulator or a real device.
+
+To use the framework with a different project, the XCFramework can be added to the project by
+adding `build-apple/llama.xcframework` by dragging and dropping it into the project navigator, or
+by manually selecting the framework in the "Frameworks, Libraries, and Embedded Content" section
+of the project settings.
+
 ![image](https://github.com/ggml-org/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
 
 Video demonstration:

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_cache_clear(context)
+        llama_kv_self_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -7,7 +7,6 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
-		1809696D2D05A39F00400EE8 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = 1809696C2D05A39F00400EE8 /* llama */; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
 		79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
@@ -18,9 +17,25 @@
 		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
 		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
 		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
+		DD84C9FD2D747FED007778EC /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = DD84C9FC2D747FED007778EC /* llama.xcframework */; };
+		DD84C9FE2D747FED007778EC /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = DD84C9FC2D747FED007778EC /* llama.xcframework */; settings = {ATTRIBUTES = (CodeSignOnCopy, RemoveHeadersOnCopy, ); }; };
 		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
 /* End PBXBuildFile section */
 
+/* Begin PBXCopyFilesBuildPhase section */
+		DD84C9FF2D747FED007778EC /* Embed Frameworks */ = {
+			isa = PBXCopyFilesBuildPhase;
+			buildActionMask = 2147483647;
+			dstPath = "";
+			dstSubfolderSpec = 10;
+			files = (
+				DD84C9FE2D747FED007778EC /* llama.xcframework in Embed Frameworks */,
+			);
+			name = "Embed Frameworks";
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXCopyFilesBuildPhase section */
+
 /* Begin PBXFileReference section */
 		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
 		79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InputButton.swift; sourceTree = "<group>"; };
@@ -33,6 +48,7 @@
 		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
 		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
 		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+		DD84C9FC2D747FED007778EC /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; name = llama.xcframework; path = "../../build-apple/llama.xcframework"; sourceTree = "<group>"; };
 		DF2D2FE72B4A59BE00FCB72D /* llama.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = llama.cpp; path = ../..; sourceTree = "<group>"; };
 		F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LoadCustomButton.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */
@@ -42,9 +58,9 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				1809696D2D05A39F00400EE8 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
+				DD84C9FD2D747FED007778EC /* llama.xcframework in Frameworks */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -86,6 +102,7 @@
 		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
+				DD84C9FC2D747FED007778EC /* llama.xcframework */,
 				549479CA2AC9E16000E0F78B /* Metal.framework */,
 				8A39BE092AC7601000BFEB40 /* Accelerate.framework */,
 			);
@@ -144,6 +161,7 @@
 				8A1C836F2AC328BD0096AF73 /* Sources */,
 				8A1C83702AC328BD0096AF73 /* Frameworks */,
 				8A1C83712AC328BD0096AF73 /* Resources */,
+				DD84C9FF2D747FED007778EC /* Embed Frameworks */,
 			);
 			buildRules = (
 			);
@@ -151,7 +169,6 @@
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
-				1809696C2D05A39F00400EE8 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -427,13 +444,6 @@
 			defaultConfigurationName = Release;
 		};
 /* End XCConfigurationList section */
-
-/* Begin XCSwiftPackageProductDependency section */
-		1809696C2D05A39F00400EE8 /* llama */ = {
-			isa = XCSwiftPackageProductDependency;
-			productName = llama;
-		};
-/* End XCSwiftPackageProductDependency section */
 	};
 	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
 }

examples/llava/CMakeLists.txt

@@ -51,6 +51,13 @@ install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
+set(TARGET llama-gemma3-cli)
+add_executable(${TARGET} gemma3-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
 set(TARGET llama-llava-clip-quantize-cli)
 add_executable(${TARGET} clip-quantize-cli.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-clip-quantize-cli)

examples/llava/README-gemma3.md

@@ -0,0 +1,50 @@
+# Gemma 3 vision
+
+> [!IMPORTANT]
+>
+> This is very experimental, only used for demo purpose.
+
+## Quick started
+
+You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)'s Hugging Face account
+
+```bash
+# build
+cmake -B build
+cmake --build build --target llama-gemma3-cli
+
+# alternatively, install from brew (MacOS)
+brew install llama.cpp
+
+# run it
+llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# note: 1B model does not support vision
+```
+
+## How to get mmproj.gguf?
+
+```bash
+cd gemma-3-4b-it
+python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py .
+
+# output file is mmproj.gguf
+```
+
+## How to run it?
+
+What you need:
+- The text model GGUF, can be converted using `convert_hf_to_gguf.py`
+- The mmproj file from step above
+- An image file
+
+```bash
+# build
+cmake -B build
+cmake --build build --target llama-gemma3-cli
+
+# run it
+./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+```

examples/llava/README-minicpmo2.6.md

@@ -5,13 +5,25 @@ Currently, this readme only supports minicpm-omni's image capabilities, and we w
 
 Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch model from huggingface to "MiniCPM-o-2_6" folder.
 
+
+### Build llama.cpp
+Readme modification time: 20250206
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
 Clone llama.cpp:
 ```bash
-git clone git@github.com:OpenBMB/llama.cpp.git
+git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
-git checkout minicpm-omni
 ```
 
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
 ### Usage of MiniCPM-o 2.6
 
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
@@ -22,25 +34,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 # quantize int4 version
-./llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
+./build/bin/llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
 ```
 
-Build llama.cpp using `CMake`:
-https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md
-
-```bash
-cmake -B build
-cmake --build build --config Release
-```
 
 Inference on Linux or Mac
-```
+```bash
 # run f16 version
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
 # run quantized int4 version
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
-
-# or run in interactive mode
-./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/README-minicpmv2.5.md

@@ -4,13 +4,26 @@
 
 Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder.
 
+
+### Build llama.cpp
+Readme modification time: 20250206
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
 Clone llama.cpp:
 ```bash
 git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 
-### Usage
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-Llama3-V 2.5
 
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
@@ -20,80 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version
-./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+./build/bin/llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
 ```
 
-Build for Linux or Mac
-
-```bash
-make
-make llama-minicpmv-cli
-```
 
 Inference on Linux or Mac
-```
+```bash
 # run f16 version
-./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
 # run quantized int4 version
-./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
-
-# or run in interactive mode
-./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
-```
-
-### Android
-
-#### Build on Android device using Termux
-We found that build on Android device would bring better runtime performance, so we recommend to build on device.
-
-[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
-
-Install tools in Termux:
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
-
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
-
-#### Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-
-```bash
-mkdir build-android
-cd build-android
-export NDK=/your_ndk_path
-cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
-$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/README-minicpmv2.6.md

@@ -4,13 +4,25 @@
 
 Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder.
 
+
+### Build llama.cpp
+Readme modification time: 20250206
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
 Clone llama.cpp:
 ```bash
-git clone git@github.com:OpenBMB/llama.cpp.git
+git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
-git checkout minicpmv-main
 ```
 
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
 ### Usage of MiniCPM-V 2.6
 
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
@@ -21,87 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 # quantize int4 version
-./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
+./build/bin/llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
 ```
 
-Build for Linux or Mac
-
-```bash
-make
-make llama-minicpmv-cli
-```
 
 Inference on Linux or Mac
-```
+```bash
 # run f16 version
-./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
 # run quantized int4 version
-./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
-
-# or run in interactive mode
-./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
-```
-
-### Video
-Install FFmpeg
-```
-brew install ffmpeg
-brew install pkg-config
-```
-
-### Android
-
-#### Build on Android device using Termux
-We found that build on Android device would bring better runtime performance, so we recommend to build on device.
-
-[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
-
-Install tools in Termux:
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
-
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
-
-#### Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-
-```bash
-mkdir build-android
-cd build-android
-export NDK=/your_ndk_path
-cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
-$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
 ```

examples/llava/clip.cpp

@@ -4,31 +4,12 @@
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
 #include "ggml.h"
+#include "ggml-cpp.h"
 #include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 #include "gguf.h"
 
-//#ifdef GGML_USE_CUDA
-//#include "ggml-cuda.h"
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//#include "ggml-sycl.h"
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//#include "ggml-metal.h"
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//#include "ggml-cann.h"
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//#include "ggml-vulkan.h"
-//#endif
-
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -155,6 +136,8 @@ static std::string format(const char * fmt, ...) {
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
+#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
 #define TN_MINICPMV_QUERY "resampler.query"
@@ -181,6 +164,7 @@ enum projector_type {
     PROJECTOR_TYPE_RESAMPLER,
     PROJECTOR_TYPE_GLM_EDGE,
     PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -191,6 +175,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_RESAMPLER, "resampler"},
     { PROJECTOR_TYPE_GLM_EDGE, "adapter"},
     { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3, "gemma3"},
 };
 
 
@@ -317,7 +302,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
             return kv.first;
         }
     }
-    return PROJECTOR_TYPE_UNKNOWN;
+    throw std::runtime_error(format("Unknown projector type: %s", name.c_str()));
 }
 
 #ifdef CLIP_DEBUG_FUNCTIONS
@@ -574,6 +559,10 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_ln_kv_b;
     struct ggml_tensor * mm_model_ln_post_w;
     struct ggml_tensor * mm_model_ln_post_b;
+
+    // gemma3
+    struct ggml_tensor * mm_input_proj_w;
+    struct ggml_tensor * mm_soft_emb_norm_w;
 };
 
 struct clip_ctx {
@@ -588,7 +577,7 @@ struct clip_ctx {
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
 
-    int32_t max_feature_layer;
+    int32_t max_feature_layer; // unused in newer models like gemma3
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
@@ -600,21 +589,209 @@ struct clip_ctx {
     bool has_post_norm = false;
     bool has_patch_bias = false;
 
-    struct gguf_context * ctx_gguf;
-    struct ggml_context * ctx_data;
+    struct gguf_context * ctx_gguf = nullptr;
+    struct ggml_context * ctx_data = nullptr;
 
     std::vector<uint8_t> buf_compute_meta;
 
-    // memory buffers to evaluate the model
-    ggml_backend_buffer_t params_buffer  = NULL;
+    std::vector<ggml_backend_t> backend_ptrs;
+    std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    ggml_backend_t backend       = NULL;
-    ggml_gallocr_t compute_alloc = NULL;
+    ggml_backend_t backend     = nullptr;
+    ggml_backend_t backend_cpu = nullptr;
+    ggml_backend_buffer_t buf  = nullptr;
 
-    struct clip_image_size * load_image_size;
+    ggml_backend_sched_ptr sched;
+
+    struct clip_image_size * load_image_size = nullptr;
+
+    clip_ctx(clip_context_params & ctx_params) {
+        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+        backend     = ctx_params.use_gpu
+                        ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
+                        : nullptr;
+
+        if (backend) {
+            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));
+            backend_ptrs.push_back(backend);
+            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend));
+        } else {
+            backend = backend_cpu;
+            LOG_INF("%s: CLIP using CPU backend\n", __func__);
+        }
+
+        backend_ptrs.push_back(backend_cpu);
+        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
+
+        sched.reset(
+            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false)
+        );
+    }
+
+    ~clip_ctx() {
+        ggml_free(ctx_data);
+        gguf_free(ctx_gguf);
+        ggml_backend_buffer_free(buf);
+        ggml_backend_free(backend);
+        if (backend_cpu != backend) {
+            ggml_backend_free(backend_cpu);
+        }
+    }
 };
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+
+    const int patch_size           = hparams.patch_size;
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int hidden_size          = hparams.hidden_size;
+    const int n_head               = hparams.n_head;
+    const int d_head               = hidden_size / n_head;
+    const int n_layer              = hparams.n_layer;
+    const float eps                = hparams.eps;
+
+    GGML_ASSERT(imgs->size == 1); // batch_size == 1
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // input raw
+    struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+    inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size);
+    inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+    inp = ggml_add(ctx0, inp, model.patch_bias);
+
+    // position embeddings
+    struct ggml_tensor * embeddings = ggml_add(ctx0, inp, model.position_embeddings);
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
+
+        // layernorm1
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), model.layers[il].ln_1_b);
+        }
+
+        // self-attention
+        {
+
+            struct ggml_tensor * Q =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
+
+            Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches);
+            Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+
+            struct ggml_tensor * K =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
+
+            K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches);
+            K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+
+            struct ggml_tensor * V =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
+
+            V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches);
+            V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head));
+            KQ = ggml_soft_max_inplace(ctx0, KQ);
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+            KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head);
+            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches);
+        }
+
+        // attention output
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, embeddings);
+
+        embeddings = cur; // embeddings = residual, cur = hidden_states
+
+        // layernorm2
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
+
+        // siglip uses gelu
+        cur = ggml_gelu(ctx0, cur);
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
+
+        // residual 2
+        cur = ggml_add(ctx0, embeddings, cur);
+
+        embeddings = cur;
+    }
+
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        const int batch_size = 1;
+        const int mm_tokens_per_image = 256; // default value for gemma3
+        const int tokens_per_side = sqrt(mm_tokens_per_image);
+        const int patches_per_image = sqrt(num_patches);
+        const int kernel_size = patches_per_image / tokens_per_side;
+
+        embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings));
+        embeddings = ggml_reshape_4d(ctx0, embeddings, patches_per_image, patches_per_image, hidden_size, batch_size);
+
+        // doing a pool2d to reduce the number of output tokens to 256
+        embeddings = ggml_pool_2d(ctx0, embeddings, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, embeddings->ne[0] * embeddings->ne[0], hidden_size, batch_size);
+        embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings));
+
+        // apply norm before projection
+        embeddings = ggml_rms_norm(ctx0, embeddings, eps);
+        embeddings = ggml_mul(ctx0, embeddings, model.mm_soft_emb_norm_w);
+
+        // apply projection
+        embeddings = ggml_mul_mat(ctx0,
+            ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)),
+            embeddings);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
         LOG_ERR("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -1160,7 +1337,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         } else {
             GGML_ABORT("fatel error");
         }
-    } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+    }
+    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
 
         embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
@@ -1182,8 +1360,25 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     return gf;
 }
 
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return clip_image_build_graph_siglip(ctx, imgs);
+    } else {
+        // TODO: we should have one build_* function per model
+        return clip_image_build_graph_legacy(ctx, imgs, load_image_size, is_inf);
+    }
+}
+
 // read and create ggml_context containing the tensors and their data
 struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+    return clip_init(fname, clip_context_params{
+        /* use_gpu */   true,
+        /* verbosity */ verbosity,
+    });
+}
+
+struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) {
+    int verbosity = ctx_params.verbosity;
     struct ggml_context * meta = NULL;
 
     struct gguf_init_params params = {
@@ -1201,14 +1396,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         const int n_kv = gguf_get_n_kv(ctx);
         const int ftype = get_u32(ctx, KEY_FTYPE);
         const std::string ftype_str = get_ftype(ftype);
-        const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION);
-        const std::string description = gguf_get_val_str(ctx, idx_desc);
         const int idx_name = gguf_find_key(ctx, KEY_NAME);
         if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
             const std::string name = gguf_get_val_str(ctx, idx_name);
             LOG_INF("%s: model name:   %s\n", __func__, name.c_str());
         }
-        LOG_INF("%s: description:  %s\n", __func__, description.c_str());
+        const int idx_desc = gguf_find_key(ctx, KEY_DESCRIPTION);
+        if (idx_desc != -1) { // ditto
+            const std::string description = gguf_get_val_str(ctx, idx_desc);
+            LOG_INF("%s: description:  %s\n", __func__, description.c_str());
+        }
         LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
         LOG_INF("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
         LOG_INF("%s: n_tensors:    %d\n", __func__, n_tensors);
@@ -1277,7 +1474,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx{};
+    clip_ctx * new_clip = new clip_ctx(ctx_params);
 
     // update projector type
     {
@@ -1296,36 +1493,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-//#ifdef GGML_USE_CUDA
-//    new_clip->backend = ggml_backend_cuda_init(0);
-//    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//    new_clip->backend = ggml_backend_metal_init();
-//    LOG_INF("%s: CLIP using Metal backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//    new_clip->backend = ggml_backend_cann_init(0);
-//    LOG_INF("%s: CLIP using CANN backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//    new_clip->backend = ggml_backend_vk_init(0);
-//    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//    new_clip->backend = ggml_backend_sycl_init(0);
-//    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
-//#endif
-
-    if (!new_clip->backend) {
-        new_clip->backend = ggml_backend_cpu_init();
-        LOG_INF("%s: CLIP using CPU backend\n", __func__);
-    }
-
     // model size and capabilities
     {
         int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC);
@@ -1363,8 +1530,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         GGML_ASSERT(new_clip->has_vision_encoder);
         GGML_ASSERT(!new_clip->has_text_encoder);
 
-        idx = get_key_idx(ctx, KEY_USE_GELU);
-        new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+        try {
+            idx = get_key_idx(ctx, KEY_USE_GELU);
+            new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+        } catch (std::runtime_error & /*e*/) {
+            new_clip->use_gelu = false;
+        }
 
         try {
             idx = get_key_idx(ctx, KEY_USE_SILU);
@@ -1378,6 +1549,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             LOG_INF("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
             LOG_INF("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
+            LOG_INF("%s: minicpmv_version:  %d\n", __func__, new_clip->minicpmv_version);
             LOG_INF("%s: glm_projector:  %d\n", __func__, new_clip->has_glm_projector);
             LOG_INF("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_INF("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
@@ -1420,7 +1592,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         // alloc memory and offload data
-        new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend);
+        ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(new_clip->backend);
+        new_clip->buf = ggml_backend_alloc_ctx_tensors_from_buft(new_clip->ctx_data, buft);
+        ggml_backend_buffer_set_usage(new_clip->buf, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
@@ -1433,7 +1607,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 return nullptr;
             }
             int num_bytes = ggml_nbytes(cur);
-            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+            if (ggml_backend_buft_is_host(buft)) {
                 // for the CPU and Metal backend, we can read directly into the tensor
                 fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
             } else {
@@ -1569,11 +1743,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         try {
-            vision_model.patch_embeddings_0    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+        } catch(const std::exception& /*e*/) {
+            vision_model.patch_embeddings_0 = nullptr;
+        }
+
+        try {
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& /*e*/) {
-            LOG_ERR("%s: failed to load vision model tensors\n", __func__);
+            vision_model.position_embeddings = nullptr;
         }
+
         try {
             vision_model.patch_embeddings_1    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1);
         } catch(const std::exception& /*e*/) {
@@ -1684,6 +1864,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
             vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_GEMMA3) {
+            vision_model.mm_input_proj_w    = get_tensor(new_clip->ctx_data, TN_MM_INP_PROJ);
+            vision_model.mm_soft_emb_norm_w = get_tensor(new_clip->ctx_data, TN_MM_SOFT_EMB_N);
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1719,14 +1903,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     // measure mem requirement and allocate
     {
         new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
-        new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
         batch.data = nullptr;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
-        ggml_gallocr_reserve(new_clip->compute_alloc, gf);
-        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
-        LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+        ggml_backend_sched_reserve(new_clip->sched.get(), gf);
+        for (size_t i = 0; i < new_clip->backend_ptrs.size(); ++i) {
+            ggml_backend_t backend = new_clip->backend_ptrs[i];
+            ggml_backend_buffer_type_t buft = new_clip->backend_buft[i];
+            size_t size = ggml_backend_sched_get_buffer_size(new_clip->sched.get(), backend);
+            if (size > 1) {
+                LOG_INF("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
+                        ggml_backend_buft_name(buft),
+                        size / 1024.0 / 1024.0);
+            }
+        }
     }
 
     return new_clip;
@@ -2218,7 +2409,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         return true;
     }
 
-    if (ctx->has_glm_projector) {
+    if (ctx->has_glm_projector || ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
         res_imgs->size = 1;
         res_imgs->data = new clip_image_f32[res_imgs->size];
         clip_image_u8 resized_image;
@@ -2407,12 +2598,6 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
 }
 
 void clip_free(clip_ctx * ctx) {
-    ggml_free(ctx->ctx_data);
-    gguf_free(ctx->ctx_gguf);
-
-    ggml_backend_buffer_free(ctx->params_buffer);
-    ggml_backend_free(ctx->backend);
-    ggml_gallocr_free(ctx->compute_alloc);
     delete ctx;
 }
 
@@ -2608,8 +2793,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     }
 
     // build the inference graph
+    ggml_backend_sched_reset(ctx->sched.get());
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
-    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+    ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);
 
     // set inputs
     const auto & model = ctx->vision_model;
@@ -2748,6 +2934,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
         }
+        else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+            // do nothing
+        }
         else {
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
 
@@ -2774,11 +2963,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         }
     }
 
-    if (ggml_backend_is_cpu(ctx->backend)) {
-        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
-    }
+    ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
 
-    ggml_backend_graph_compute(ctx->backend, gf);
+    auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
+    if (status != GGML_STATUS_SUCCESS) {
+        LOG_ERR("%s: ggml_backend_sched_graph_compute failed with error %d\n", __func__, status);
+        return false;
+    }
 
     // the last node is the embedding tensor
     struct ggml_tensor * embeddings = ggml_graph_node(gf, -1);
@@ -2800,7 +2991,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     assert(itype < GGML_TYPE_COUNT);
     ggml_type type = static_cast<ggml_type>(itype);
 
-    auto * ctx_clip = clip_model_load(fname_inp, 2);
+    auto * ctx_clip = clip_init(fname_inp, clip_context_params{
+        /* use_gpu */   false,
+        /* verbosity */ 2,
+    });
 
     const auto & ctx_src = ctx_clip->ctx_gguf;
     const auto & ctx_data = ctx_clip->ctx_data;
@@ -2958,6 +3152,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         return ctx->vision_model.mm_1_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return ctx->vision_model.mm_input_proj_w->ne[0];
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));

examples/llava/clip.h

@@ -39,8 +39,15 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+struct clip_context_params {
+    bool use_gpu;
+    int verbosity;
+};
+
+// deprecated, use clip_init
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
+CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params);
 
 CLIP_API void clip_free(struct clip_ctx * ctx);
 

examples/llava/convert_image_encoder_to_gguf.py

@@ -89,6 +89,7 @@ def bytes_to_unicode():
 ap = argparse.ArgumentParser()
 ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
 ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument('--bigendian', action="store_true", default=False, help="Model is executed on big-endian machine")
 ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
@@ -191,7 +192,7 @@ output_dir = args.output_dir if args.output_dir is not None else dir_model
 os.makedirs(output_dir, exist_ok=True)
 output_prefix = os.path.basename(output_dir).replace("ggml_", "")
 fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
-fout = GGUFWriter(path=fname_out, arch="clip")
+fout = GGUFWriter(path=fname_out, arch="clip", endianess=GGUFEndian.LITTLE if not args.bigendian else GGUFEndian.BIG)
 
 fout.add_bool("clip.has_text_encoder", has_text_encoder)
 fout.add_bool("clip.has_vision_encoder", has_vision_encoder)

examples/llava/gemma3-cli.cpp

@@ -0,0 +1,341 @@
+#include "arg.h"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "stb_image.h"
+#include "llama.h"
+#include "ggml.h"
+#include "console.h"
+
+#include <vector>
+#include <limits.h>
+#include <inttypes.h>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+static bool g_is_generating = false;
+
+/**
+ * Please note that this is NOT a production-ready stuff.
+ * It is a playground for trying Gemma 3 vision capabilities.
+ * For contributors: please keep this code simple and easy to understand.
+ */
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG(
+        "Experimental CLI for using Gemma 3 vision model\n\n"
+        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
+        "  -m and --mmproj are required\n"
+        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
+        argv[0]
+    );
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (g_is_generating) {
+            g_is_generating = false;
+        } else {
+            console::cleanup();
+            LOG("\nInterrupted by user\n");
+            _exit(130);
+        }
+    }
+}
+#endif
+
+struct gemma3_context {
+    struct clip_ctx    * ctx_clip = NULL;
+    common_init_result   llama_init;
+
+    llama_model       * model;
+    llama_context     * lctx;
+    const llama_vocab * vocab;
+    llama_batch         batch;
+
+    int n_threads    = 1;
+    llama_pos n_past = 0;
+
+    gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) {
+        model = llama_init.model.get();
+        lctx = llama_init.context.get();
+        vocab = llama_model_get_vocab(model);
+        n_threads = params.cpuparams.n_threads;
+        batch = llama_batch_init(params.n_batch, 0, 1);
+        init_clip_model(params);
+    }
+
+    void init_clip_model(common_params & params) {
+        const char * clip_path = params.mmproj.path.c_str();
+        ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
+    }
+
+    ~gemma3_context() {
+        clip_free(ctx_clip);
+    }
+};
+
+struct decode_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
+static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
+    llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
+    common_batch_clear(ctx.batch);
+    for (llama_token & t : tokens) {
+        common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false);
+    }
+    if (logits_last) {
+        ctx.batch.logits[ctx.batch.n_tokens - 1] = true;
+    }
+    // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
+    if (llama_decode(ctx.lctx, ctx.batch)) {
+        LOG_ERR("Failed to decode text\n");
+        return 1;
+    }
+    return 0;
+}
+
+static int eval_image(gemma3_context & ctx, std::string & fname) {
+    std::vector<float> image_embd_v;
+    int n_embd = llama_model_n_embd(ctx.model);
+    int n_tokens = 256;
+    image_embd_v.resize(n_tokens * n_embd);
+
+    bool ok;
+    struct clip_image_u8 * img_u8 = clip_image_u8_init();
+    ok = clip_image_load_from_file(fname.c_str(), img_u8);
+    if (!ok) {
+        LOG_ERR("Unable to load image %s\n", fname.c_str());
+        clip_image_u8_free(img_u8);
+        return 2; // non-fatal error
+    }
+
+    clip_image_f32_batch batch_f32;
+    ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32);
+    if (!ok) {
+        LOG_ERR("Unable to preprocess image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+
+    int64_t t0 = ggml_time_ms();
+    LOG("Encoding image %s\n", fname.c_str());
+    ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data());
+    if (!ok) {
+        LOG_ERR("Unable to encode image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+    LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+
+    clip_image_f32_batch_free(&batch_f32);
+    clip_image_u8_free(img_u8);
+
+    // decode image embeddings
+    int64_t t1 = ggml_time_ms();
+    eval_text(ctx, "<start_of_image>");
+    llama_set_causal_attn(ctx.lctx, false);
+    decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0);
+    if (llama_decode(ctx.lctx, batch_img.batch)) {
+        LOG_ERR("failed to decode image\n");
+        return 1;
+    }
+    ctx.n_past += n_tokens;
+    llama_set_causal_attn(ctx.lctx, true);
+    eval_text(ctx, "<end_of_image>");
+    LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
+    return 0;
+}
+
+static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
+    for (int i = 0; i < n_predict; i++) {
+        if (i > n_predict || !g_is_generating) {
+            printf("\n");
+            break;
+        }
+
+        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        common_sampler_accept(smpl, token_id, true);
+
+        if (llama_vocab_is_eog(ctx.vocab, token_id)) {
+            printf("\n");
+            break; // end of generation
+        }
+
+        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        fflush(stdout);
+
+        // eval the token
+        common_batch_clear(ctx.batch);
+        common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
+        if (llama_decode(ctx.lctx, ctx.batch)) {
+            LOG_ERR("failed to decode token\n");
+            return 1;
+        }
+    }
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+    params.sampling.temp = 0.2; // lower temp by default for better quality
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.path.empty()) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    gemma3_context ctx(params);
+    printf("%s: %s\n", __func__, params.model.path.c_str());
+
+    bool is_single_turn = !params.prompt.empty() && !params.image.empty();
+
+    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
+    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
+
+    // ctrl+C handling
+    {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    }
+
+    if (eval_text(ctx, "<bos>")) {
+        return 1;
+    }
+
+    if (is_single_turn) {
+        g_is_generating = true;
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+        for (auto & fname : params.image) {
+            if (eval_image(ctx, fname)) {
+                return 1;
+            }
+        }
+        if (eval_text(ctx, params.prompt + "<end_of_turn><start_of_turn>model\n", true)) {
+            return 1;
+        }
+        if (generate_response(ctx, smpl, n_predict)) {
+            return 1;
+        }
+
+    } else {
+        LOG("\n Running in chat mode, available commands:");
+        LOG("\n   /image <path>    load an image");
+        LOG("\n   /clear           clear the chat history");
+        LOG("\n   /quit or /exit   exit the program");
+        LOG("\n");
+
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+
+        while (true) {
+            g_is_generating = false;
+            LOG("\n> ");
+            console::set_display(console::user_input);
+            std::string line;
+            console::readline(line, false);
+            console::set_display(console::reset);
+            line = string_strip(line);
+            if (line.empty()) {
+                continue;
+            }
+            if (line == "/quit" || line == "/exit") {
+                break;
+            }
+            if (line == "/clear") {
+                ctx.n_past = 0;
+                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                LOG("Chat history cleared\n\n");
+                continue;
+            }
+            g_is_generating = true;
+            if (line.find("/image") == 0) {
+                std::string image = line.substr(7);
+                int res = eval_image(ctx, image);
+                if (res == 2) {
+                    continue; // image not found
+                }
+                if (res) {
+                    return 1;
+                }
+                continue;
+            }
+            if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
+                return 1;
+            }
+            if (generate_response(ctx, smpl, n_predict)) {
+                return 1;
+            }
+            if (eval_text(ctx, "<end_of_turn><start_of_turn>user\n")) {
+                return 1;
+            }
+        }
+    }
+
+    return 0;
+}

examples/llava/gemma3_convert_encoder_to_gguf.py

@@ -0,0 +1,307 @@
+import gguf
+import argparse
+import logging
+import sys
+import torch
+import json
+import os
+import numpy as np
+from typing import cast, ContextManager, Any, Iterator
+from pathlib import Path
+from torch import Tensor
+
+logger = logging.getLogger("gemma3-mmproj")
+
+
+# (copied from convert_hf_to_gguf.py)
+# tree of lazy tensors
+class LazyTorchTensor(gguf.LazyBase):
+    _tensor_type = torch.Tensor
+    # to keep the type-checker happy
+    dtype: torch.dtype
+    shape: torch.Size
+
+    # only used when converting a torch.Tensor to a np.ndarray
+    _dtype_map: dict[torch.dtype, type] = {
+        torch.float16: np.float16,
+        torch.float32: np.float32,
+    }
+
+    # used for safetensors slices
+    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
+    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
+    _dtype_str_map: dict[str, torch.dtype] = {
+        "F64": torch.float64,
+        "F32": torch.float32,
+        "BF16": torch.bfloat16,
+        "F16": torch.float16,
+        # "U64": torch.uint64,
+        "I64": torch.int64,
+        # "U32": torch.uint32,
+        "I32": torch.int32,
+        # "U16": torch.uint16,
+        "I16": torch.int16,
+        "U8": torch.uint8,
+        "I8": torch.int8,
+        "BOOL": torch.bool,
+        "F8_E4M3": torch.float8_e4m3fn,
+        "F8_E5M2": torch.float8_e5m2,
+    }
+
+    def numpy(self) -> gguf.LazyNumpyTensor:
+        dtype = self._dtype_map[self.dtype]
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
+            args=(self,),
+            func=(lambda s: s.numpy())
+        )
+
+    @classmethod
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
+
+    @classmethod
+    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
+        dtype = cls._dtype_str_map[st_slice.get_dtype()]
+        shape: tuple[int, ...] = tuple(st_slice.get_shape())
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
+        return cast(torch.Tensor, lazy)
+
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.Tensor.numpy:
+            return args[0].numpy()
+
+        return cls._wrap_fn(func)(*args, **kwargs)
+
+
+class Gemma3VisionTower:
+    hparams: dict
+    gguf_writer: gguf.GGUFWriter
+    fname_out: Path
+    ftype: gguf.LlamaFileType
+
+    @staticmethod
+    def load_hparams(dir_model: Path):
+        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
+        part_names: list[str] = []
+        for filename in os.listdir(dir_model):
+            if filename.startswith(prefix) and filename.endswith(suffix):
+                part_names.append(filename)
+        part_names.sort()
+        return part_names
+
+    def __init__(self,
+                 dir_model: Path,
+                 fname_out: Path,
+                 ftype: gguf.LlamaFileType,
+                 is_big_endian: bool,):
+        hparams = Gemma3VisionTower.load_hparams(dir_model)
+        self.hparams = hparams
+        self.fname_out = fname_out
+        self.ftype = ftype
+        endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess)
+
+        text_config = hparams["text_config"]
+        vision_config = hparams["vision_config"]
+
+        assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration"
+        assert text_config is not None
+        assert vision_config is not None
+
+        self.gguf_writer.add_string ("clip.projector_type",              "gemma3")
+        self.gguf_writer.add_bool   ("clip.has_text_encoder",            False)
+        self.gguf_writer.add_bool   ("clip.has_vision_encoder",          True)
+        self.gguf_writer.add_bool   ("clip.has_llava_projector",         False) # legacy
+        self.gguf_writer.add_uint32 ("clip.vision.image_size",           vision_config["image_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.patch_size",           vision_config["patch_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.embedding_length",     vision_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length",  vision_config["intermediate_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.projection_dim",       text_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.block_count",          vision_config["num_hidden_layers"])
+        self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"])
+        self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6))
+        # default values taken from HF tranformers code
+        self.gguf_writer.add_array  ("clip.vision.image_mean", [0.5, 0.5, 0.5])
+        self.gguf_writer.add_array  ("clip.vision.image_std",  [0.5, 0.5, 0.5])
+        self.gguf_writer.add_bool   ("clip.use_gelu", True)
+
+        # load tensors
+        for name, data_torch in self.get_tensors(dir_model):
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+            self.add_tensor(name, data_torch)
+
+    def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]:
+        part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors")
+        tensor_names_from_parts: set[str] = set()
+        for part_name in part_names:
+            logger.info(f"gguf: loading model part '{part_name}'")
+            from safetensors import safe_open
+            ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu"))
+            with ctx as model_part:
+                tensor_names_from_parts.update(model_part.keys())
+
+                for name in model_part.keys():
+                    data = model_part.get_slice(name)
+                    data = LazyTorchTensor.from_safetensors_slice(data)
+                    yield name, data
+
+    def add_tensor(self, name: str, data_torch: Tensor):
+        is_1d = len(data_torch.shape) == 1
+        is_embd = ".embeddings." in name
+        old_dtype = data_torch.dtype
+        can_quantize = not is_1d and not is_embd
+        data_qtype = gguf.GGMLQuantizationType.F32
+
+        # this is to support old checkpoint
+        # TODO: remove this when we have the final model
+        name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.")
+        name = name.replace("multimodal_projector.", "multi_modal_projector.")
+
+        # filter only vision tensors
+        if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."):
+            return
+        # prefix
+        name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.")
+        name = name.replace("vision_tower.vision_model.", "v.")
+        # projector and input embd
+        name = name.replace(".embeddings.patch_embedding.", ".patch_embd.")
+        name = name.replace(".embeddings.position_embedding.", ".position_embd.")
+        name = name.replace(
+            "multi_modal_projector.mm_input_projection_weight",
+            "mm.input_projection.weight"
+        )
+        name = name.replace(
+            "multi_modal_projector.mm_soft_emb_norm.weight",
+            "mm.soft_emb_norm.weight"
+        )
+        name = name.replace("post_layernorm.", "post_ln.")
+        # each block
+        name = name.replace(".self_attn.k_proj.", ".attn_k.")
+        name = name.replace(".self_attn.v_proj.", ".attn_v.")
+        name = name.replace(".self_attn.q_proj.", ".attn_q.")
+        name = name.replace(".self_attn.out_proj.", ".attn_out.")
+        name = name.replace(".layer_norm1.", ".ln1.")
+        name = name.replace(".layer_norm2.", ".ln2.")
+        name = name.replace(".mlp.fc1.", ".ffn_down.")
+        name = name.replace(".mlp.fc2.", ".ffn_up.")
+
+        if can_quantize:
+            if self.ftype == gguf.LlamaFileType.ALL_F32:
+                data_qtype = gguf.GGMLQuantizationType.F32
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                data_qtype = gguf.GGMLQuantizationType.F16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                data_qtype = gguf.GGMLQuantizationType.BF16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                data_qtype = gguf.GGMLQuantizationType.Q8_0
+            else:
+                raise ValueError(f"Unsupported file type: {self.ftype}")
+
+        # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
+        # the other norm values are part of SigLIP model, and they are already correct
+        # ref code: Gemma3RMSNorm
+        if "soft_emb_norm.weight" in name:
+            logger.info(f"Correcting norm value for '{name}'")
+            data_torch = data_torch + 1
+
+        data = data_torch.numpy()
+
+        try:
+            data = gguf.quants.quantize(data, data_qtype)
+        except Exception as e:
+            logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16")
+            data_qtype = gguf.GGMLQuantizationType.F16
+            data = gguf.quants.quantize(data, data_qtype)
+
+        # reverse shape to make it similar to the internal ggml dimension order
+        shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}"
+        logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
+
+        self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
+
+    def write(self):
+        self.gguf_writer.write_header_to_file(path=self.fname_out)
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.write_tensors_to_file(progress=True)
+        self.gguf_writer.close()
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert Gemma 3 vision tower safetensors to GGUF format",)
+    parser.add_argument(
+        "--outfile", type=Path, default="mmproj.gguf",
+        help="path to write to",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
+        help="output format",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file",
+        nargs="?",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+
+    args = parser.parse_args()
+    if args.model is None:
+        parser.error("the following arguments are required: model")
+    return args
+
+
+def main() -> None:
+    args = parse_args()
+
+    if args.verbose:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    dir_model = args.model
+
+    if not dir_model.is_dir():
+        logger.error(f'Error: {args.model} is not a directory')
+        sys.exit(1)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+    }
+
+    logger.info(f"Loading model: {dir_model.name}")
+
+    with torch.inference_mode():
+        gemma3_vision_tower = Gemma3VisionTower(
+            dir_model=dir_model,
+            fname_out=args.outfile,
+            ftype=ftype_map[args.outtype],
+            is_big_endian=args.bigendian,
+        )
+        gemma3_vision_tower.write()
+
+
+if __name__ == '__main__':
+    main()
+

examples/llava/llava-cli.cpp

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

examples/llava/minicpmv-cli.cpp

@@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) {
 
     llama_model_params model_params = common_model_params_to_llama(*params);
 
-    llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
@@ -80,13 +80,17 @@ static void llava_free(struct llava_context * ctx_llava) {
 }
 
 static struct clip_ctx * clip_init_context(common_params * params) {
-    const char * clip_path = params->mmproj.c_str();
+    const char * clip_path = params->mmproj.path.c_str();
 
     auto prompt = params->prompt;
     if (prompt.empty()) {
         prompt = "describe the image in detail.";
     }
-    auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+    struct clip_context_params clip_params = {
+        /* use_gpu */   params->n_gpu_layers != 0,
+        /* verbosity */ params->verbosity,
+    };
+    auto * ctx_clip = clip_init(clip_path, clip_params);
     return ctx_clip;
 }
 
@@ -148,19 +152,34 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
     eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
     if (num_image_embeds > 1) {
-        size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
-        eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
-        for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
-            for (size_t j = 0; j < num_image_embeds_col; ++j) {
-                eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
-                process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
-                eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
-                if (j == num_image_embeds_col - 1) {
-                    eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+        if (has_minicpmv_projector == 2) {
+            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+            eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+                for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                    eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                    if (j == num_image_embeds_col - 1) {
+                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                    }
+                }
+            }
+            eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+        }
+        else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
+            size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+            for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+                for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                    eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+                    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                    eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+                    if (j == num_image_embeds_col - 1) {
+                        eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                    }
                 }
             }
         }
-        eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
     }
     LOG_INF("%s: image token past: %d\n", __func__, n_past);
 }
@@ -271,7 +290,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.mmproj.empty() || (params.image.empty())) {
+    if (params.mmproj.path.empty() || (params.image.empty())) {
         show_additional_info(argc, argv);
         return 1;
     }

examples/llava/minicpmv-convert-image-encoder-to-gguf.py

@@ -597,7 +597,6 @@ elif args.minicpmv_projector is not None:
     fname_middle = "mmproj-"
     has_text_encoder = False
     has_minicpmv_projector = True
-    minicpmv_version = 4
 elif args.vision_only:
     fname_middle = "vision-"
     has_text_encoder = False

examples/llava/qwen2vl-cli.cpp

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

examples/lookahead/lookahead.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <algorithm>
 
 struct ngram_data {
     bool active = false;
@@ -95,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -437,17 +438,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_cache_seq_keep(ctx, seq_id_best);
-            llama_kv_cache_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_cache_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_kv_self_seq_keep(ctx, seq_id_best);
+            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -192,7 +192,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/main/README.md

@@ -27,12 +27,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 ##### Input prompt (One-and-done)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 ##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -44,12 +56,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 
 ##### Input prompt (One-and-done)
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 #### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -77,6 +101,8 @@ The `llama-cli` program provides several ways to interact with the LLaMA models
 
 -   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
 -   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
+-   `--system-prompt PROMPT`: Provide a system prompt (will otherwise use the default one in the chat template (if provided)).
+-   `--system-prompt-file FNAME`: Provide a file containing a system prompt.
 -   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
 
 ## Interaction
@@ -89,7 +115,10 @@ In interactive mode, users can participate in text generation by injecting their
 
 -   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
 -   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
--   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false)
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default or provided chat template) (default: true if chat template found)
+-   `-no-cnv`:  Disable conversation mode (default: false)
+-   `-st, --single-turn`:  Only process a single conversation turn (user input) and then exit.
+-   `--jinja`:  Enable jinja chat template parser, will use the model's built-in template or a user-provided one (default: false)
 -   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
 
 By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
@@ -125,6 +154,8 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t
 
  Example usage: `--chat-template gemma`
 
+`--chat-template-file FNAME`:  Load a custom jinja chat template from an external file, useful if the model contains outdated or incompatible template, some examples can be found in models/templates. Up-to-date chat templates can be downloaded from Hugging Face using scripts/get_chat_template.py
+
 ## Context Management
 
 During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.

examples/main/main.cpp

@@ -31,8 +31,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const char * DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant";
-
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
 static common_sampler          ** g_smpl;
@@ -47,8 +45,8 @@ static void print_usage(int argc, char ** argv) {
     (void) argc;
 
     LOG("\nexample usage:\n");
-    LOG("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
-    LOG("\n  chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
+    LOG("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128 -no-cnv\n", argv[0]);
+    LOG("\n  chat (conversation): %s -m your_model.gguf -sys \"You are a helpful assistant\"\n", argv[0]);
     LOG("\n");
 }
 
@@ -219,6 +217,10 @@ int main(int argc, char ** argv) {
     // print chat template example in conversation mode
     if (params.conversation_mode) {
         if (params.enable_chat_template) {
+            if (!params.prompt.empty() && params.system_prompt.empty()) {
+                LOG_WRN("*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?\n");
+            }
+
             LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str());
         } else {
             LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
@@ -263,6 +265,7 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd_inp;
 
+    bool waiting_for_first_input = false;
     auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
         common_chat_msg new_msg;
         new_msg.role = role;
@@ -273,13 +276,34 @@ int main(int argc, char ** argv) {
         return formatted;
     };
 
+    std::string prompt;
     {
-        auto prompt = (params.conversation_mode && params.enable_chat_template)
-            // format the system prompt in conversation mode (fallback to default if empty)
-            ? chat_add_and_format("system", params.prompt.empty() ? DEFAULT_SYSTEM_MESSAGE : params.prompt)
+        if (params.conversation_mode && params.enable_chat_template) {
+            if (!params.system_prompt.empty()) {
+                // format the system prompt (will use template default if empty)
+                chat_add_and_format("system", params.system_prompt);
+            }
+
+            if (!params.prompt.empty()) {
+                // format and append the user prompt
+                chat_add_and_format("user", params.prompt);
+            } else {
+                waiting_for_first_input = true;
+            }
+
+            if (!params.system_prompt.empty() || !params.prompt.empty()) {
+                common_chat_templates_inputs inputs;
+                inputs.messages = chat_msgs;
+                inputs.add_generation_prompt = !params.prompt.empty();
+
+                prompt = common_chat_templates_apply(chat_templates.get(), inputs).prompt;
+            }
+        } else {
             // otherwise use the prompt as is
-            : params.prompt;
-        if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
+            prompt = params.prompt;
+        }
+
+        if (params.interactive_first || !prompt.empty() || session_tokens.empty()) {
             LOG_DBG("tokenize the prompt\n");
             embd_inp = common_tokenize(ctx, prompt, true, true);
         } else {
@@ -292,7 +316,7 @@ int main(int argc, char ** argv) {
     }
 
     // Should not run without any tokens
-    if (embd_inp.empty()) {
+    if (!waiting_for_first_input && embd_inp.empty()) {
         if (add_bos) {
             embd_inp.push_back(llama_vocab_bos(vocab));
             LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
@@ -330,7 +354,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -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",
@@ -352,7 +376,12 @@ int main(int argc, char ** argv) {
     }
 
     if (params.conversation_mode) {
-        params.interactive_first = true;
+        if (params.single_turn && !params.prompt.empty()) {
+            params.interactive = false;
+            params.interactive_first = false;
+        } else {
+            params.interactive_first = true;
+        }
     }
 
     // enable interactive mode if interactive start is specified
@@ -476,8 +505,8 @@ int main(int argc, char ** argv) {
         LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
 #endif
         LOG_INF(       "%s", control_message);
-        if (params.conversation_mode && params.enable_chat_template && params.prompt.empty()) {
-            LOG_INF(   " - Using default system message. To change it, set a different value via -p PROMPT or -f FILE argument.\n");
+        if (params.conversation_mode && params.enable_chat_template && params.system_prompt.empty()) {
+            LOG_INF(   " - Not using system message. To change it, set a different value via -sys PROMPT\n");
         }
         LOG_INF("\n");
 
@@ -573,8 +602,8 @@ int main(int argc, char ** argv) {
                     LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_self_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -597,9 +626,9 @@ int main(int argc, char ** argv) {
                     LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_self_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 
@@ -773,7 +802,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of generation tokens in interactive mode
-            if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
+            if (!waiting_for_first_input && llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
                 LOG_DBG("found an EOG token\n");
 
                 if (params.interactive) {
@@ -793,12 +822,17 @@ int main(int argc, char ** argv) {
             }
 
             // if current token is not EOG, we add it to current assistant message
-            if (params.conversation_mode) {
+            if (params.conversation_mode && !waiting_for_first_input) {
                 const auto id = common_sampler_last(smpl);
                 assistant_ss << common_token_to_piece(ctx, id, false);
+
+                if (!prompt.empty()) {
+                    prompt.clear();
+                    is_interacting = false;
+                }
             }
 
-            if (n_past > 0 && is_interacting) {
+            if ((n_past > 0 || waiting_for_first_input) && is_interacting) {
                 LOG_DBG("waiting for user input\n");
 
                 if (params.conversation_mode) {
@@ -888,11 +922,17 @@ int main(int argc, char ** argv) {
                 input_echo = false; // do not echo this again
             }
 
-            if (n_past > 0) {
+            if (n_past > 0 || waiting_for_first_input) {
                 if (is_interacting) {
                     common_sampler_reset(smpl);
                 }
                 is_interacting = false;
+
+                if (waiting_for_first_input && params.single_turn) {
+                    params.interactive = false;
+                    params.interactive_first = false;
+                }
+                waiting_for_first_input = false;
             }
         }
 

examples/parallel/parallel.cpp

@@ -12,6 +12,7 @@
 #include <string>
 #include <vector>
 #include <ctime>
+#include <algorithm>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {
@@ -105,6 +106,8 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    params.n_predict = 128;
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
@@ -201,7 +204,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -233,9 +236,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
+                llama_kv_self_seq_rm(ctx, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -371,8 +374,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
+                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 
@@ -404,7 +407,7 @@ int main(int argc, char ** argv) {
         params.prompt_file = "used built-in defaults";
     }
     LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
-    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
+    LOG_INF("Model and path used:  \033[32m%s\033[0m\n\n", params.model.path.c_str());
 
     LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
     LOG_INF("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);

examples/passkey/passkey.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <algorithm>
 
 static void print_usage(int, char ** argv) {
     LOG("\nexample usage:\n");
@@ -63,7 +64,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = common_model_params_to_llama(params);
 
-    llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n" , __func__);
@@ -132,11 +133,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_update  (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -166,12 +167,12 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+      //llama_kv_self_defrag (ctx);
+        llama_kv_self_update (ctx);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -197,12 +198,12 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+          //llama_kv_self_defrag (ctx);
+            llama_kv_self_update (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1203,7 +1203,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1575,7 +1575,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1765,7 +1765,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

examples/quantize-stats/quantize-stats.cpp

@@ -1,6 +1,6 @@
 #include "ggml.h"
 #include "llama.h"
-#include "llama-context.h"
+#include "llama-model.h"
 #include "common.h"
 
 #include <algorithm>
@@ -328,7 +328,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const auto & tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(model);
 
     // check layer tensors
     int included_layers = 0;

examples/quantize/quantize.cpp

@@ -8,6 +8,7 @@
 #include <unordered_map>
 #include <fstream>
 #include <cmath>
+#include <cctype>
 
 struct quant_option {
     std::string name;

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/rpc/CMakeLists.txt

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

examples/rpc/README.md

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

examples/rpc/rpc-server.cpp

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

examples/run/linenoise.cpp/linenoise.h

@@ -47,27 +47,27 @@ extern "C" {
 #include <stddef.h> /* For size_t. */
 #include <stdlib.h>
 
-extern const char *linenoiseEditMore;
+extern const char * linenoiseEditMore;
 
 /* The linenoiseState structure represents the state during line editing.
  * We pass this state to functions implementing specific editing
  * functionalities. */
 struct linenoiseState {
-    int in_completion;  /* The user pressed TAB and we are now in completion
+    int          in_completion;  /* The user pressed TAB and we are now in completion
                          * mode, so input is handled by completeLine(). */
-    size_t completion_idx; /* Index of next completion to propose. */
-    int ifd;            /* Terminal stdin file descriptor. */
-    int ofd;            /* Terminal stdout file descriptor. */
-    char *buf;          /* Edited line buffer. */
-    size_t buflen;      /* Edited line buffer size. */
-    const char *prompt; /* Prompt to display. */
-    size_t plen;        /* Prompt length. */
-    size_t pos;         /* Current cursor position. */
-    size_t oldpos;      /* Previous refresh cursor position. */
-    size_t len;         /* Current edited line length. */
-    size_t cols;        /* Number of columns in terminal. */
-    size_t oldrows;     /* Rows used by last refrehsed line (multiline mode) */
-    int history_index;  /* The history index we are currently editing. */
+    size_t       completion_idx; /* Index of next completion to propose. */
+    int          ifd;            /* Terminal stdin file descriptor. */
+    int          ofd;            /* Terminal stdout file descriptor. */
+    char *       buf;            /* Edited line buffer. */
+    size_t       buflen;         /* Edited line buffer size. */
+    const char * prompt;         /* Prompt to display. */
+    size_t       plen;           /* Prompt length. */
+    size_t       pos;            /* Current cursor position. */
+    size_t       oldcolpos;      /* Previous refresh cursor column position. */
+    size_t       len;            /* Current edited line length. */
+    size_t       cols;           /* Number of columns in terminal. */
+    size_t       oldrows;        /* Rows used by last refreshed line (multiline mode) */
+    int          history_index;  /* The history index we are currently editing. */
 };
 
 struct linenoiseCompletions {
@@ -89,19 +89,20 @@ struct linenoiseCompletions {
 };
 
 /* Non blocking API. */
-int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, char *buf, size_t buflen, const char *prompt);
-const char *linenoiseEditFeed(struct linenoiseState *l);
-void linenoiseEditStop(struct linenoiseState *l);
-void linenoiseHide(struct linenoiseState *l);
-void linenoiseShow(struct linenoiseState *l);
+int          linenoiseEditStart(struct linenoiseState * l, int stdin_fd, int stdout_fd, char * buf, size_t buflen,
+                                const char * prompt);
+const char * linenoiseEditFeed(struct linenoiseState * l);
+void         linenoiseEditStop(struct linenoiseState * l);
+void         linenoiseHide(struct linenoiseState * l);
+void         linenoiseShow(struct linenoiseState * l);
 
 /* Blocking API. */
-const char *linenoise(const char *prompt);
-void linenoiseFree(void *ptr);
+const char * linenoise(const char * prompt);
+void         linenoiseFree(void * ptr);
 
 /* Completion API. */
 typedef void(linenoiseCompletionCallback)(const char *, linenoiseCompletions *);
-typedef const char*(linenoiseHintsCallback)(const char *, int *color, int *bold);
+typedef const char *(linenoiseHintsCallback) (const char *, int * color, int * bold);
 typedef void(linenoiseFreeHintsCallback)(const char *);
 void linenoiseSetCompletionCallback(linenoiseCompletionCallback *);
 void linenoiseSetHintsCallback(linenoiseHintsCallback *);
@@ -109,10 +110,10 @@ void linenoiseSetFreeHintsCallback(linenoiseFreeHintsCallback *);
 void linenoiseAddCompletion(linenoiseCompletions *, const char *);
 
 /* History API. */
-int linenoiseHistoryAdd(const char *line);
+int linenoiseHistoryAdd(const char * line);
 int linenoiseHistorySetMaxLen(int len);
-int linenoiseHistorySave(const char *filename);
-int linenoiseHistoryLoad(const char *filename);
+int linenoiseHistorySave(const char * filename);
+int linenoiseHistoryLoad(const char * filename);
 
 /* Other utilities. */
 void linenoiseClearScreen(void);
@@ -121,6 +122,14 @@ void linenoisePrintKeyCodes(void);
 void linenoiseMaskModeEnable(void);
 void linenoiseMaskModeDisable(void);
 
+/* Encoding functions. */
+typedef size_t(linenoisePrevCharLen)(const char * buf, size_t buf_len, size_t pos, size_t * col_len);
+typedef size_t(linenoiseNextCharLen)(const char * buf, size_t buf_len, size_t pos, size_t * col_len);
+typedef size_t(linenoiseReadCode)(int fd, char * buf, size_t buf_len, int * c);
+
+void linenoiseSetEncodingFunctions(linenoisePrevCharLen * prevCharLenFunc, linenoiseNextCharLen * nextCharLenFunc,
+                                   linenoiseReadCode * readCodeFunc);
+
 #ifdef __cplusplus
 }
 #endif

examples/run/run.cpp

@@ -38,24 +38,6 @@
 }
 #endif
 
-GGML_ATTRIBUTE_FORMAT(1, 2)
-static std::string fmt(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    const int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::string buf;
-    buf.resize(size);
-    const int size2 = vsnprintf(const_cast<char *>(buf.data()), buf.size() + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-
-    return buf;
-}
-
 GGML_ATTRIBUTE_FORMAT(1, 2)
 static int printe(const char * fmt, ...) {
     va_list args;
@@ -79,6 +61,7 @@ class Opt {
         ctx_params           = llama_context_default_params();
         model_params         = llama_model_default_params();
         context_size_default = ctx_params.n_batch;
+        n_threads_default    = ctx_params.n_threads;
         ngl_default          = model_params.n_gpu_layers;
         common_params_sampling sampling;
         temperature_default = sampling.temp;
@@ -104,6 +87,7 @@ class Opt {
 
         ctx_params.n_batch        = context_size >= 0 ? context_size : context_size_default;
         ctx_params.n_ctx          = ctx_params.n_batch;
+        ctx_params.n_threads = ctx_params.n_threads_batch = n_threads >= 0 ? n_threads : n_threads_default;
         model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default;
         temperature               = temperature >= 0 ? temperature : temperature_default;
 
@@ -116,12 +100,12 @@ class Opt {
     std::string chat_template_file;
     std::string          user;
     bool                 use_jinja   = false;
-    int                  context_size = -1, ngl = -1;
+    int                  context_size = -1, ngl = -1, n_threads = -1;
     float                temperature = -1;
     bool                 verbose     = false;
 
   private:
-    int   context_size_default = -1, ngl_default = -1;
+    int   context_size_default = -1, ngl_default = -1, n_threads_default = -1;
     float temperature_default = -1;
     bool  help                = false;
 
@@ -159,53 +143,94 @@ class Opt {
         return 0;
     }
 
+    int parse_options_with_value(int argc, const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
+            if (handle_option_with_value(argc, argv, i, context_size) == 1) {
+                return 1;
+            }
+        } else if (options_parsing &&
+                   (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
+            if (handle_option_with_value(argc, argv, i, ngl) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && (strcmp(argv[i], "-t") == 0 || strcmp(argv[i], "--threads") == 0)) {
+            if (handle_option_with_value(argc, argv, i, n_threads) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
+            if (handle_option_with_value(argc, argv, i, temperature) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0) {
+            if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
+                return 1;
+            }
+            use_jinja = true;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_options(const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
+            verbose = true;
+        } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
+            use_jinja = true;
+        } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
+            help = true;
+            return 0;
+        } else if (options_parsing && strcmp(argv[i], "--") == 0) {
+            options_parsing = false;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_positional_args(const char ** argv, int & i, int & positional_args_i) {
+        if (positional_args_i == 0) {
+            if (!argv[i][0] || argv[i][0] == '-') {
+                return 1;
+            }
+
+            ++positional_args_i;
+            model_ = argv[i];
+        } else if (positional_args_i == 1) {
+            ++positional_args_i;
+            user = argv[i];
+        } else {
+            user += " " + std::string(argv[i]);
+        }
+
+        return 0;
+    }
+
     int parse(int argc, const char ** argv) {
         bool options_parsing   = true;
         for (int i = 1, positional_args_i = 0; i < argc; ++i) {
-            if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
-                if (handle_option_with_value(argc, argv, i, context_size) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
-                if (handle_option_with_value(argc, argv, i, ngl) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
-                if (handle_option_with_value(argc, argv, i, temperature) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
-                verbose = true;
-            } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
-                use_jinja = true;
-            } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0){
-                if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
-                    return 1;
-                }
-                use_jinja = true;
-            } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
-                help = true;
-                return 0;
-            } else if (options_parsing && strcmp(argv[i], "--") == 0) {
-                options_parsing = false;
-            } else if (positional_args_i == 0) {
-                if (!argv[i][0] || argv[i][0] == '-') {
-                    return 1;
-                }
+            int ret = parse_options_with_value(argc, argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
 
-                ++positional_args_i;
-                model_ = argv[i];
-            } else if (positional_args_i == 1) {
-                ++positional_args_i;
-                user = argv[i];
-            } else {
-                user += " " + std::string(argv[i]);
+            ret = parse_options(argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
+
+            if (parse_positional_args(argv, i, positional_args_i)) {
+                return 1;
             }
         }
 
-        if (model_.empty()){
+        if (model_.empty()) {
             return 1;
         }
 
@@ -232,6 +257,8 @@ class Opt {
             "      Number of GPU layers (default: %d)\n"
             "  --temp <value>\n"
             "      Temperature (default: %.1f)\n"
+            "  -t, --threads <value>\n"
+            "      Number of threads to use during generation (default: %d)\n"
             "  -v, --verbose, --log-verbose\n"
             "      Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n"
             "  -h, --help\n"
@@ -260,7 +287,7 @@ class Opt {
             "  llama-run file://some-file3.gguf\n"
             "  llama-run --ngl 999 some-file4.gguf\n"
             "  llama-run --ngl 999 some-file5.gguf Hello World\n",
-            context_size_default, ngl_default, temperature_default);
+            context_size_default, ngl_default, temperature_default, n_threads_default);
     }
 };
 
@@ -480,11 +507,11 @@ class HttpClient {
         int secs = static_cast<int>(seconds) % 60;
 
         if (hrs > 0) {
-            return fmt("%dh %02dm %02ds", hrs, mins, secs);
+            return string_format("%dh %02dm %02ds", hrs, mins, secs);
         } else if (mins > 0) {
-            return fmt("%dm %02ds", mins, secs);
+            return string_format("%dm %02ds", mins, secs);
         } else {
-            return fmt("%ds", secs);
+            return string_format("%ds", secs);
         }
     }
 
@@ -499,7 +526,7 @@ class HttpClient {
             }
         }
 
-        return fmt("%.2f %s", dbl_size, suffix[i]);
+        return string_format("%.2f %s", dbl_size, suffix[i]);
     }
 
     static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
@@ -533,7 +560,9 @@ class HttpClient {
         return (now_downloaded_plus_file_size * 100) / total_to_download;
     }
 
-    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
+    static std::string generate_progress_prefix(curl_off_t percentage) {
+        return string_format("%3ld%% |", static_cast<long int>(percentage));
+    }
 
     static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
         const auto                          now             = std::chrono::steady_clock::now();
@@ -544,9 +573,9 @@ class HttpClient {
     static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download,
                                                 double speed, double estimated_time) {
         const int width = 10;
-        return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width,
-                   human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width,
-                   human_readable_time(estimated_time).c_str());
+        return string_format("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(),
+                             width, human_readable_size(total_to_download).c_str(), width,
+                             human_readable_size(speed).c_str(), width, human_readable_time(estimated_time).c_str());
     }
 
     static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) {
@@ -891,7 +920,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;
 
     const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
     prompt_tokens.resize(n_prompt_tokens);
@@ -907,7 +936,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
     const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");

examples/save-load-state/save-load-state.cpp

@@ -15,7 +15,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    print_build_info();
+    common_init();
 
     if (params.n_predict < 0) {
         params.n_predict = 16;
@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_cache_clear(ctx3);
+        llama_kv_self_clear(ctx3);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/server/public_legacy/json-schema-to-grammar.mjs

@@ -1,5 +1,5 @@
 // WARNING: This file was ported from json_schema_to_grammar.py, please fix bugs / add features there first.
-const SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}';
+const SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}';
 
 function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
   if (minItems === 0 && maxItems === 1) {

examples/server/server.cpp

@@ -131,9 +131,10 @@ struct slot_params {
             lora.push_back({{"id", i}, {"scale", this->lora[i].scale}});
         }
 
-        std::vector<std::string> grammar_trigger_words;
-        for (const auto & trigger : sampling.grammar_trigger_words) {
-            grammar_trigger_words.push_back(trigger.word);
+        auto grammar_triggers = json::array();
+        for (const auto & trigger : sampling.grammar_triggers) {
+            server_grammar_trigger ct(std::move(trigger));
+            grammar_triggers.push_back(ct.to_json());
         }
 
         return json {
@@ -170,8 +171,8 @@ struct slot_params {
             {"n_probs",                   sampling.n_probs},
             {"min_keep",                  sampling.min_keep},
             {"grammar",                   sampling.grammar},
-            {"grammar_trigger_words",     grammar_trigger_words},
-            {"grammar_trigger_tokens",    sampling.grammar_trigger_tokens},
+            {"grammar_lazy",              sampling.grammar_lazy},
+            {"grammar_triggers",          grammar_triggers},
             {"preserved_tokens",          sampling.preserved_tokens},
             {"chat_format",               common_chat_format_name(oaicompat_chat_format)},
             {"samplers",                  samplers},
@@ -356,24 +357,6 @@ struct server_task {
         }
 
         {
-            const auto grammar_triggers = data.find("grammar_triggers");
-            if (grammar_triggers != data.end()) {
-                for (const auto & t : *grammar_triggers) {
-                    common_grammar_trigger trigger;
-                    trigger.word = t.at("word");
-                    trigger.at_start = t.at("at_start");
-
-                    auto ids = common_tokenize(vocab, trigger.word, /* add_special= */ false, /* parse_special= */ true);
-                    if (ids.size() == 1) {
-                        SRV_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str());
-                        params.sampling.grammar_trigger_tokens.push_back(ids[0]);
-                        params.sampling.preserved_tokens.insert(ids[0]);
-                        continue;
-                    }
-                    SRV_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str());
-                    params.sampling.grammar_trigger_words.push_back(trigger);
-                }
-            }
             const auto preserved_tokens = data.find("preserved_tokens");
             if (preserved_tokens != data.end()) {
                 for (const auto & t : *preserved_tokens) {
@@ -383,12 +366,39 @@ struct server_task {
                         params.sampling.preserved_tokens.insert(ids[0]);
                     } else {
                         // This may happen when using a tool call style meant for a model with special tokens to preserve on a model without said tokens.
-                        SRV_WRN("Not preserved because more than 1 token (wrong chat template override?): %s\n", t.get<std::string>().c_str());
+                        SRV_DBG("Not preserved because more than 1 token: %s\n", t.get<std::string>().c_str());
                     }
                 }
             }
-            if (params.sampling.grammar_lazy) {
-                GGML_ASSERT(params.sampling.grammar_trigger_tokens.size() > 0 || params.sampling.grammar_trigger_words.size() > 0);
+            const auto grammar_triggers = data.find("grammar_triggers");
+            if (grammar_triggers != data.end()) {
+                for (const auto & t : *grammar_triggers) {
+                    server_grammar_trigger ct(t);
+                    if (ct.value.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
+                        const auto & word = ct.value.value;
+                        auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true);
+                        if (ids.size() == 1) {
+                            auto token = ids[0];
+                            if (std::find(params.sampling.preserved_tokens.begin(), params.sampling.preserved_tokens.end(), (llama_token) token) == params.sampling.preserved_tokens.end()) {
+                                throw std::runtime_error("Grammar trigger word should be marked as preserved token: " + word);
+                            }
+                            SRV_DBG("Grammar trigger token: %d (`%s`)\n", token, word.c_str());
+                            common_grammar_trigger trigger;
+                            trigger.type = COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN;
+                            trigger.value = word;
+                            trigger.token = token;
+                            params.sampling.grammar_triggers.push_back(std::move(trigger));
+                        } else {
+                            SRV_DBG("Grammar trigger word: `%s`\n", word.c_str());
+                            params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word});
+                        }
+                    } else {
+                        params.sampling.grammar_triggers.push_back(std::move(ct.value));
+                    }
+                }
+            }
+            if (params.sampling.grammar_lazy && params.sampling.grammar_triggers.empty()) {
+                throw std::runtime_error("Error: no triggers set for lazy grammar!");
             }
         }
 
@@ -480,8 +490,12 @@ struct result_timings {
     double predicted_per_token_ms;
     double predicted_per_second;
 
+    // Optional speculative metrics - only included when > 0
+    int32_t draft_n = 0;
+    int32_t draft_n_accepted = 0;
+
     json to_json() const {
-        return {
+        json base = {
             {"prompt_n",               prompt_n},
             {"prompt_ms",              prompt_ms},
             {"prompt_per_token_ms",    prompt_per_token_ms},
@@ -492,6 +506,13 @@ struct result_timings {
             {"predicted_per_token_ms", predicted_per_token_ms},
             {"predicted_per_second",   predicted_per_second},
         };
+
+        if (draft_n > 0) {
+            base["draft_n"] = draft_n;
+            base["draft_n_accepted"] = draft_n_accepted;
+        }
+
+        return base;
     }
 };
 
@@ -742,7 +763,10 @@ struct server_task_result_cmpl_final : server_task_result {
                         {"name", tc.name},
                         {"arguments", tc.arguments},
                     }},
-                    {"id", tc.id},
+                    // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo).
+                    // We only generate a random id for the ones that don't generate one by themselves
+                    // (they also won't get to see it as their template likely doesn't use it, so it's all for the client)
+                    {"id", tc.id.empty() ? gen_tool_call_id() : tc.id},
                 });
             }
             message["tool_calls"] = tool_calls;
@@ -818,6 +842,11 @@ struct server_task_result_cmpl_final : server_task_result {
             ret.push_back({"timings", timings.to_json()});
         }
 
+        // extra fields for debugging purposes
+        if (verbose) {
+            ret["__verbose"] = to_json_non_oaicompat();
+        }
+
         return ret;
     }
 };
@@ -1282,6 +1311,10 @@ struct server_slot {
 
     std::function<void(int)> callback_on_release;
 
+    // Speculative decoding stats
+    int32_t n_draft_total = 0;      // Total draft tokens generated
+    int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+
     void reset() {
         SLT_DBG(*this, "%s", "\n");
 
@@ -1298,13 +1331,17 @@ struct server_slot {
 
         generated_tokens.clear();
         generated_token_probs.clear();
+
+        // clear speculative decoding stats
+        n_draft_total = 0;
+        n_draft_accepted = 0;
     }
 
     bool is_non_causal() const {
         return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK;
     }
 
-    bool can_batch_with(server_slot & other_slot) {
+    bool can_batch_with(server_slot & other_slot) const {
         return is_non_causal() == other_slot.is_non_causal()
             && are_lora_equal(lora, other_slot.lora);
     }
@@ -1364,6 +1401,12 @@ struct server_slot {
         timings.predicted_per_token_ms = t_token_generation / n_decoded;
         timings.predicted_per_second = 1e3 / t_token_generation * n_decoded;
 
+        // Add speculative metrics
+        if (n_draft_total > 0) {
+            timings.draft_n = n_draft_total;
+            timings.draft_n_accepted = n_draft_accepted;
+        }
+
         return timings;
     }
 
@@ -1411,6 +1454,15 @@ struct server_slot {
                 t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                 t_token_generation, n_decoded, t_gen, n_gen_second,
                 t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
+
+        if (n_draft_total > 0) {
+            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            SLT_INF(*this,
+                    "\n"
+                    "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total
+            );
+        }
     }
 
     json to_json() const {
@@ -1825,7 +1877,7 @@ struct server_context {
     }
 
     bool load_model(const common_params & params) {
-        SRV_INF("loading model '%s'\n", params.model.c_str());
+        SRV_INF("loading model '%s'\n", params.model.path.c_str());
 
         params_base = params;
 
@@ -1835,7 +1887,7 @@ struct server_context {
         ctx   = llama_init.context.get();
 
         if (model == nullptr) {
-            SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
+            SRV_ERR("failed to load model, '%s'\n", params_base.model.path.c_str());
             return false;
         }
 
@@ -1846,31 +1898,32 @@ struct server_context {
         add_bos_token = llama_vocab_get_add_bos(vocab);
         has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
 
-        if (!params_base.speculative.model.empty() || !params_base.speculative.hf_repo.empty()) {
-            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
+        if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
 
             auto params_dft = params_base;
 
             params_dft.devices      = params_base.speculative.devices;
-            params_dft.hf_file      = params_base.speculative.hf_file;
-            params_dft.hf_repo      = params_base.speculative.hf_repo;
             params_dft.model        = params_base.speculative.model;
-            params_dft.model_url    = params_base.speculative.model_url;
             params_dft.n_ctx        = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx;
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
 
+            // force F16 KV cache for the draft model for extra performance
+            params_dft.cache_type_k = GGML_TYPE_F16;
+            params_dft.cache_type_v = GGML_TYPE_F16;
+
             llama_init_dft = common_init_from_params(params_dft);
 
             model_dft = llama_init_dft.model.get();
 
             if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
+                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
                 return false;
             }
 
             if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
-                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
+                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
 
                 return false;
             }
@@ -1880,10 +1933,6 @@ struct server_context {
             cparams_dft = common_context_params_to_llama(params_dft);
             cparams_dft.n_batch = n_ctx_dft;
 
-            // force F16 KV cache for the draft model for extra performance
-            cparams_dft.type_k = GGML_TYPE_F16;
-            cparams_dft.type_v = GGML_TYPE_F16;
-
             // the context is not needed - we will create one for each slot
             llama_init_dft.context.reset();
         }
@@ -1892,6 +1941,7 @@ struct server_context {
         try {
             common_chat_format_example(chat_templates.get(), params.use_jinja);
         } catch (const std::exception & e) {
+            SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what());
             SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
             chat_templates = common_chat_templates_init(model, "chatml");
         }
@@ -2027,6 +2077,18 @@ struct server_context {
         return ret;
     }
 
+    bool can_be_detokenized(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
+        const llama_model * model = llama_get_model(ctx);
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const int32_t n_vocab = llama_vocab_n_tokens(vocab);
+        for (const auto & token : tokens) {
+            if (token < 0 || token >= n_vocab) {
+                return false;
+            }
+        }
+        return true;
+    }
+
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot.reset();
         slot.id_task       = task.id;
@@ -2041,11 +2103,16 @@ struct server_context {
             slot.lora = task.params.lora;
         }
 
+        bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
+        if (!can_detokenize) {
+            send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST);
+            return false;
+        }
         SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
             // Might be better to reject the request with a 400 ?
-            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.params.n_predict, slot.n_predict);
+            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d\n", slot.params.n_predict, slot.n_predict);
             slot.params.n_predict = slot.n_predict;
         }
 
@@ -2083,7 +2150,7 @@ struct server_context {
         SRV_DBG("%s", "clearing KV cache\n");
 
         // clear the entire KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         clean_kv_cache = false;
     }
 
@@ -2148,14 +2215,6 @@ struct server_context {
         }
 
         if (slot.has_new_line) {
-            // if we have already seen a new line, we stop after a certain time limit
-            if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) {
-                slot.stop           = STOP_TYPE_LIMIT;
-                slot.has_next_token = false;
-
-                SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms);
-            }
-
             // require that each new line has a whitespace prefix (i.e. indentation) of at least slot.params.n_indent
             if (slot.params.n_indent > 0) {
                 // check the current indentation
@@ -2194,6 +2253,14 @@ struct server_context {
         // check if there is a new line in the generated text
         if (result.text_to_send.find('\n') != std::string::npos) {
             slot.has_new_line = true;
+
+            // if we have seen a new line, we stop after a certain time limit, but only upon another new line
+            if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) {
+                slot.stop           = STOP_TYPE_LIMIT;
+                slot.has_next_token = false;
+
+                SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms);
+            }
         }
 
         // if context shift is disabled, we stop when it reaches the context limit
@@ -2625,8 +2692,8 @@ struct server_context {
                     res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                     res->t_start             = metrics.t_start;
 
-                    res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx);
-                    res->kv_cache_used_cells   = llama_get_kv_cache_used_cells(ctx);
+                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
+                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);
 
                     res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                     res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
@@ -2742,7 +2809,7 @@ struct server_context {
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
+                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
                     slot->cache_tokens.clear();
 
                     auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2810,8 +2877,8 @@ struct server_context {
 
                 SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
+                llama_kv_self_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
+                llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 if (slot.params.cache_prompt) {
                     for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@@ -3002,8 +3069,8 @@ struct server_context {
 
                                             const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
 
-                                            llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
-                                            llama_kv_cache_seq_add(ctx, slot.id, head_c, -1,     kv_shift);
+                                            llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
+                                            llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@@ -3041,9 +3108,9 @@ struct server_context {
                     }
 
                     // keep only the common part
-                    if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
                         // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
+                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
 
                         // there is no common part left
                         slot.n_past = 0;
@@ -3255,6 +3322,9 @@ struct server_context {
 
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
+                // keep track of total number of tokens generated in the draft
+                slot.n_draft_total += draft.size();
+
                 // ignore small drafts
                 if (slot.params.speculative.n_min > (int) draft.size()) {
                     SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
@@ -3280,10 +3350,13 @@ struct server_context {
                 slot.n_past    += ids.size();
                 slot.n_decoded += ids.size();
 
+                // update how many tokens out of draft was accepted
+                slot.n_draft_accepted += ids.size() - 1;
+
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
 
-                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
 
                 for (size_t i = 0; i < ids.size(); ++i) {
                     completion_token_output result;
@@ -3790,7 +3863,7 @@ int main(int argc, char ** argv) {
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
-            { "model_path",                  ctx_server.params_base.model },
+            { "model_path",                  ctx_server.params_base.model.path },
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -4056,7 +4129,7 @@ int main(int argc, char ** argv) {
             {"object", "list"},
             {"data", {
                 {
-                    {"id",       params.model_alias.empty() ? params.model : params.model_alias},
+                    {"id",       params.model_alias.empty() ? params.model.path : params.model_alias},
                     {"object",   "model"},
                     {"created",  std::time(0)},
                     {"owned_by", "llamacpp"},
@@ -4424,15 +4497,24 @@ int main(int argc, char ** argv) {
         llama_backend_free();
     };
 
-    // bind HTTP listen port
     bool was_bound = false;
-    if (params.port == 0) {
-        int bound_port = svr->bind_to_any_port(params.hostname);
-        if ((was_bound = (bound_port >= 0))) {
-            params.port = bound_port;
-        }
+    if (string_ends_with(std::string(params.hostname), ".sock")) {
+        LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
+        svr->set_address_family(AF_UNIX);
+        // bind_to_port requires a second arg, any value other than 0 should
+        // simply get ignored
+        was_bound = svr->bind_to_port(params.hostname, 8080);
     } else {
-        was_bound = svr->bind_to_port(params.hostname, params.port);
+        LOG_INF("%s: binding port with default address family\n", __func__);
+        // bind HTTP listen port
+        if (params.port == 0) {
+            int bound_port = svr->bind_to_any_port(params.hostname);
+            if ((was_bound = (bound_port >= 0))) {
+                params.port = bound_port;
+            }
+        } else {
+            was_bound = svr->bind_to_port(params.hostname, params.port);
+        }
     }
 
     if (!was_bound) {

examples/server/tests/unit/test_chat_completion.py

@@ -144,6 +144,7 @@ def test_apply_chat_template():
 @pytest.mark.parametrize("response_format,n_predicted,re_content", [
     ({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
     ({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
+    ({"type": "json_schema", "json_schema": {"schema": {"const": "foooooo"}}}, 10, "\"foooooo\""),
     ({"type": "json_object"}, 10, "(\\{|John)+"),
     ({"type": "sound"}, 0, None),
     # invalid response format (expected to fail)

examples/server/tests/unit/test_tool_call.py

@@ -1,4 +1,12 @@
+#!/usr/bin/env python
 import pytest
+
+# ensure grandparent path is in sys.path
+from pathlib import Path
+import sys
+path = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(path))
+
 from utils import *
 
 server: ServerProcess
@@ -66,15 +74,8 @@ WEATHER_TOOL = {
 }
 
 
-def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, argument_key: str | None):
-    global server
-    n_predict = 512
-    # server = ServerPreset.stories15m_moe()
-    server.jinja = True
-    server.n_predict = n_predict
-    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
-    server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
+def do_test_completion_with_required_tool_tiny(server: ServerProcess, tool: dict, argument_key: str | None, n_predict, **kwargs):
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
             {"role": "system", "content": "You are a coding assistant."},
@@ -83,16 +84,15 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a
         "tool_choice": "required",
         "tools": [tool],
         "parallel_tool_calls": False,
-        "temperature": 0.0,
-        "top_k": 1,
-        "top_p": 1.0,
+        **kwargs,
     })
     assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
     choice = res.body["choices"][0]
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
-    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
+    assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}'
+    assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}'
     expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
     assert expected_function_name == tool_call["function"]["name"]
     actual_arguments = tool_call["function"]["arguments"]
@@ -108,7 +108,14 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a
     ("meta-llama-Llama-3.3-70B-Instruct",             PYTHON_TOOL,          "code"),
 ])
 def test_completion_with_required_tool_tiny_fast(template_name: str, tool: dict, argument_key: str | None):
-    do_test_completion_with_required_tool_tiny(template_name, tool, argument_key)
+    global server
+    n_predict = 512
+    # server = ServerPreset.stories15m_moe()
+    server.jinja = True
+    server.n_predict = n_predict
+    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    do_test_completion_with_required_tool_tiny(server, tool, argument_key, n_predict, temperature=0.0, top_k=1, top_p=1.0)
 
 
 @pytest.mark.slow
@@ -130,10 +137,17 @@ def test_completion_with_required_tool_tiny_fast(template_name: str, tool: dict,
     ("deepseek-ai-DeepSeek-R1-Distill-Llama-8B",      TEST_TOOL,            "success"),
     ("deepseek-ai-DeepSeek-R1-Distill-Llama-8B",      PYTHON_TOOL,          "code"),
     ("fireworks-ai-llama-3-firefunction-v2",          TEST_TOOL,            "success"),
-    ("fireworks-ai-llama-3-firefunction-v2",          PYTHON_TOOL,          "code"),
+    # ("fireworks-ai-llama-3-firefunction-v2",          PYTHON_TOOL,          "code"),
 ])
 def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict, argument_key: str | None):
-    do_test_completion_with_required_tool_tiny(template_name, tool, argument_key)
+    global server
+    n_predict = 512
+    # server = ServerPreset.stories15m_moe()
+    server.jinja = True
+    server.n_predict = n_predict
+    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    do_test_completion_with_required_tool_tiny(server, tool, argument_key, n_predict)
 
 
 @pytest.mark.slow
@@ -142,25 +156,33 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
     (PYTHON_TOOL,  "code",     "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
     (PYTHON_TOOL,  "code",     "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
 
-    # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
     (TEST_TOOL,    "success",  "bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
     (PYTHON_TOOL,  "code",     "bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
+    (PYTHON_TOOL,  "code",     "bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
     (PYTHON_TOOL,  "code",     "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
     (PYTHON_TOOL,  "code",     "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      "chatml"),
 
+    (TEST_TOOL,    "success",  "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M",      None),
+    (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M",      None),
+    (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M",      "chatml"),
+
+    (TEST_TOOL,    "success",  "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M",      None),
+    (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M",      None),
+    (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
+
     (TEST_TOOL,    "success",  "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
     (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
     (PYTHON_TOOL,  "code",     "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
     (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
-    # (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
+    (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
     (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
-    # (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   "chatml"),
+    (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
     (PYTHON_TOOL,  "code",     "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
@@ -176,10 +198,10 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
 
     (TEST_TOOL,    "success",  "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama/Llama-3.2-3B-Instruct", None)),
     (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama/Llama-3.2-3B-Instruct", None)),
-    # (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
-    # TODO: fix these
-    # (TEST_TOOL,    "success",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    # (PYTHON_TOOL,  "code",     "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
+
+    (TEST_TOOL,    "success",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (PYTHON_TOOL,  "code",     "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 ])
 def test_completion_with_required_tool_real_model(tool: dict, argument_key: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
@@ -197,7 +219,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
     elif isinstance(template_override, str):
         server.chat_template = template_override
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
             {"role": "system", "content": "You are a coding assistant."},
@@ -215,7 +237,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
-    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
+    # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}'
     expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
     assert expected_function_name == tool_call["function"]["name"]
     actual_arguments = tool_call["function"]["arguments"]
@@ -225,13 +247,8 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
         assert argument_key in actual_arguments, f"tool arguments: {json.dumps(actual_arguments)}, expected: {argument_key}"
 
 
-def do_test_completion_without_tool_call(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
-    global server
-    server.jinja = True
-    server.n_predict = n_predict
-    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
-    server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
+def do_test_completion_without_tool_call(server: ServerProcess, n_predict: int, tools: list[dict], tool_choice: str | None, **kwargs):
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
             {"role": "system", "content": "You are a coding assistant."},
@@ -239,9 +256,7 @@ def do_test_completion_without_tool_call(template_name: str, n_predict: int, too
         ],
         "tools": tools if tools else None,
         "tool_choice": tool_choice,
-        "temperature": 0.0,
-        "top_k": 1,
-        "top_p": 1.0,
+        **kwargs,
     }, timeout=TIMEOUT_HTTP_REQUEST)
     assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
     choice = res.body["choices"][0]
@@ -254,7 +269,12 @@ def do_test_completion_without_tool_call(template_name: str, n_predict: int, too
     ("meta-llama-Llama-3.3-70B-Instruct",         128, [PYTHON_TOOL], 'none'),
 ])
 def test_completion_without_tool_call_fast(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
-    do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice)
+    global server
+    server.jinja = True
+    server.n_predict = n_predict
+    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    do_test_completion_without_tool_call(server, n_predict, tools, tool_choice)
 
 
 @pytest.mark.slow
@@ -270,7 +290,12 @@ def test_completion_without_tool_call_fast(template_name: str, n_predict: int, t
     ("meta-llama-Llama-3.2-3B-Instruct",              256, [PYTHON_TOOL], 'none'),
 ])
 def test_completion_without_tool_call_slow(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
-    do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice)
+    global server
+    server.jinja = True
+    server.n_predict = n_predict
+    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    do_test_completion_without_tool_call(server, n_predict, tools, tool_choice)
 
 
 @pytest.mark.slow
@@ -281,6 +306,12 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
     ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
     ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      "chatml"),
 
+    ("bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M",      None),
+    ("bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M",      "chatml"),
+
+    ("bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M",      None),
+    ("bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
+
     ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
     ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        "chatml"),
 
@@ -324,48 +355,53 @@ def test_weather(hf_repo: str, template_override: str | Tuple[str, str | None] |
     elif isinstance(template_override, str):
         server.chat_template = template_override
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
-        "max_tokens": n_predict,
+    do_test_weather(server, max_tokens=n_predict)
+
+
+def do_test_weather(server: ServerProcess, **kwargs):
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "messages": [
             {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."},
             {"role": "user", "content": "What is the weather in Istanbul?"},
         ],
         "tools": [WEATHER_TOOL],
+        **kwargs,
     }, timeout=TIMEOUT_HTTP_REQUEST)
     assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
     choice = res.body["choices"][0]
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
-    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
-    assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"]
+    # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}'
+    assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"], f'Expected weather tool call, got {tool_call["function"]["name"]}'
+    assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}'
     actual_arguments = json.loads(tool_call["function"]["arguments"])
     assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}"
     location = actual_arguments["location"]
     assert isinstance(location, str), f"Expected location to be a string, got {type(location)}: {json.dumps(location)}"
-    assert re.match('^Istanbul(, (TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}'
+    assert re.match('^Istanbul(( |, ?)(TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}'
 
 
 @pytest.mark.slow
 @pytest.mark.parametrize("result_override,n_predict,hf_repo,template_override", [
     (None,                                           128,  "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       "chatml"),
-    (None,                                           128,  "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",         None),
+    (None,                                           128,  "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", None),
+    (None,                                           128,  "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", "chatml"),
     (None,                                           128,  "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",         "chatml"),
     (None,                                           128,  "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",     ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
     (None,                                           128,  "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",       ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
     (None,                                           128,  "bartowski/functionary-small-v3.2-GGUF:Q8_0",        ("meetkai/functionary-medium-v3.2", None)),
-    (None,                                           128,  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M",  None),
     (None,                                           128,  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M",  None),
     (None,                                           128,  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M",  "chatml"),
     (None,                                           128,  "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)",            8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
 
     # TODO: fix these (wrong results, either didn't respect decimal instruction or got wrong value)
-    ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)",            8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    # ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)",            8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+    # (None,                                           128,  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M",  None),
+    # ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)",            8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 ])
 def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
-    # n_predict = 512
     server.n_slots = 1
     server.jinja = True
     server.n_ctx = 8192 * 2
@@ -379,10 +415,14 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
     elif isinstance(template_override, str):
         server.chat_template = template_override
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
+    do_test_calc_result(server, result_override, n_predict)
+
+
+def do_test_calc_result(server: ServerProcess, result_override: str | None, n_predict: int, **kwargs):
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
-            {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things, and provide very concise answers. Do not explain your reasoning to the user. Provide any numerical values back to the user with at most two decimals."},
+            {"role": "system", "content": "You are a tools-calling assistant. You express numerical values with at most two decimals."},
             {"role": "user", "content": "What's the y coordinate of a point on the unit sphere at angle 30 degrees?"},
             {
                 "role": "assistant",
@@ -423,7 +463,8 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
                     }
                 }
             }
-        ]
+        ],
+        **kwargs,
     }, timeout=TIMEOUT_HTTP_REQUEST)
     assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
     choice = res.body["choices"][0]
@@ -434,19 +475,19 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
     if result_override is not None:
         assert re.match(result_override, content), f'Expected {result_override}, got {content}'
     else:
-        assert re.match('^[\\s\\S]*?The (y[ -])?coordinate [\\s\\S]*?is (approximately )?0\\.56\\b|^0\\.56$', content), \
+        assert re.match('^[\\s\\S]*?((That\'s|\\bis) (approximately )?)?\\b0\\.(5\\b|56\\b|556)', content), \
             f'Expected something like "The y coordinate is 0.56.", got {content}'
 
 
 @pytest.mark.slow
 @pytest.mark.parametrize("n_predict,reasoning_format,expect_content,expect_reasoning_content,hf_repo,template_override", [
-    (128, 'deepseek',  "^The sum of 102 and 7 is 109.*",                        None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
-    (128,  None,        "^The sum of 102 and 7 is 109.*",                       None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (128, 'deepseek',  "^The sum of 102 and 7 is 109[\\s\\S]*",                        None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (128,  None,        "^The sum of 102 and 7 is 109[\\s\\S]*",                       None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
 
-    (1024, 'deepseek',  "To find the sum of.*",                                 "I need to calculate the sum of 102 and 7.*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    (1024, 'none',      "^I need[\\s\\S]*?</think>\n?To find.*",                None,                                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (1024, 'deepseek',  "To find the sum of[\\s\\S]*",                                 "I need to calculate the sum of 102 and 7[\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (1024, 'none',      "^(<think>\\s*)?I need[\\s\\S]*?</think>\\s*To find[\\s\\S]*",                None,                                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 
-    (1024, 'deepseek',  "To find the sum of.*",                                 "First, I [\\s\\S]*",                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+    (1024, 'deepseek',  "To find the sum of[\\s\\S]*",                                 "First, I [\\s\\S]*",                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
 ])
 def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none'] | None, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
@@ -464,7 +505,7 @@ def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none']
     elif isinstance(template_override, str):
         server.chat_template = template_override
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
             {"role": "user", "content": "What's the sum of 102 and 7?"},
@@ -476,7 +517,7 @@ def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none']
 
     content = choice["message"].get("content")
     if expect_content is None:
-        assert content is None, f'Expected no content in {choice["message"]}'
+        assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}'
     else:
         assert re.match(expect_content, content), f'Expected {expect_content}, got {content}'
 
@@ -488,46 +529,46 @@ def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none']
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [
-    (None,                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    # (None,                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", "chatml"),
+@pytest.mark.parametrize("hf_repo,template_override", [
+    ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 
-    (None,                 "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
-    (None,                 "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      "chatml"),
+    ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
+    ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      "chatml"),
 
-    (None,                 "bartowski/functionary-small-v3.2-GGUF:Q8_0",       ("meetkai-functionary-medium-v3.2", None)),
-    (None,                 "bartowski/functionary-small-v3.2-GGUF:Q8_0",       "chatml"),
+    ("bartowski/functionary-small-v3.2-GGUF:Q8_0",       ("meetkai-functionary-medium-v3.2", None)),
+    ("bartowski/functionary-small-v3.2-GGUF:Q8_0",       "chatml"),
 
-    ('{"code":"print("}',  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
-    (None,                 "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
+    # ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
+    ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
 
-    (None,                 "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
-    (None,                 "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
+    ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
+    ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      None),
 
-    ('{"code":"print("}',  "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
-    (None,                 "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
+    ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
+    ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      None),
 
-    (None,                 "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
-    (None,                 "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        "chatml"),
+    ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
+    ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        "chatml"),
 
-    (None,                 "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",    ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
-    (None,                 "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",    "chatml"),
+    ("bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",    ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
+    ("bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",    "chatml"),
 
-    (None,                 "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",      ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")),
-    (None,                 "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",      "chatml"),
+    ("bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",      ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")),
+    ("bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",      "chatml"),
 
-    (None,                 "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
-    (None,                 "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
+    ("bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
+    ("bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
 
-    # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
-    (None,                 "bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
+    ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
+    ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              "chatml"),
 ])
-def test_hello_world(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
+def test_hello_world(hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
+    n_predict = 512 # High because of DeepSeek R1
     server.n_slots = 1
     server.jinja = True
     server.n_ctx = 8192
-    server.n_predict = 512 # High because of DeepSeek R1
+    server.n_predict = n_predict
     server.model_hf_repo = hf_repo
     server.model_hf_file = None
     if isinstance(template_override, tuple):
@@ -537,31 +578,29 @@ def test_hello_world(expected_arguments_override: str | None, hf_repo: str, temp
     elif isinstance(template_override, str):
         server.chat_template = template_override
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
-    res = server.make_request("POST", "/chat/completions", data={
-        "max_tokens": 256,
+
+    do_test_hello_world(server, max_tokens=n_predict)
+
+
+def do_test_hello_world(server: ServerProcess, **kwargs):
+    res = server.make_request("POST", "/v1/chat/completions", data={
         "messages": [
-            {"role": "system", "content": "You are a coding assistant."},
+            {"role": "system", "content": "You are a tool-calling agent."},
             {"role": "user", "content": "say hello world with python"},
         ],
         "tools": [PYTHON_TOOL],
-        # Note: without these greedy params, Functionary v3.2 writes `def hello_world():\n    print("Hello, World!")\nhello_world()` which is correct but a pain to test.
-        "temperature": 0.0,
-        "top_k": 1,
-        "top_p": 1.0,
+        **kwargs,
     }, timeout=TIMEOUT_HTTP_REQUEST)
     assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
     choice = res.body["choices"][0]
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
-    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
+    # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}'
     assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"]
-    actual_arguments = tool_call["function"]["arguments"]
-    if expected_arguments_override is not None:
-        assert actual_arguments == expected_arguments_override
-    else:
-        actual_arguments = json.loads(actual_arguments)
-        assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}"
-        code = actual_arguments["code"]
-        assert isinstance(code, str), f"Expected code to be a string, got {type(code)}: {json.dumps(code)}"
-        assert re.match(r'''print\(("[Hh]ello,? [Ww]orld!?"|'[Hh]ello,? [Ww]orld!?')\)''', code), f'Expected hello world, got {code}'
+    assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}'
+    actual_arguments = json.loads(tool_call["function"]["arguments"])
+    assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}"
+    code = actual_arguments["code"]
+    assert isinstance(code, str), f"Expected code to be a string, got {type(code)}: {json.dumps(code)}"
+    assert re.match(r'''print\(("[Hh]ello,? [Ww]orld!?"|'[Hh]ello,? [Ww]orld!?')\)''', code), f'Expected hello world, got {code}'

examples/server/tests/utils.py

@@ -26,7 +26,10 @@ from re import RegexFlag
 import wget
 
 
-DEFAULT_HTTP_TIMEOUT = 12 if "LLAMA_SANITIZE" not in os.environ else 30
+DEFAULT_HTTP_TIMEOUT = 12
+
+if "LLAMA_SANITIZE" in os.environ or "GITHUB_ACTION" in os.environ:
+    DEFAULT_HTTP_TIMEOUT = 30
 
 
 class ServerResponse:
@@ -64,6 +67,9 @@ class ServerProcess:
     id_slot: int | None = None
     cache_prompt: bool | None = None
     n_slots: int | None = None
+    ctk: str | None = None
+    ctv: str | None = None
+    fa: bool | None = None
     server_continuous_batching: bool | None = False
     server_embeddings: bool | None = False
     server_reranking: bool | None = False
@@ -81,6 +87,7 @@ class ServerProcess:
     reasoning_format: Literal['deepseek', 'none'] | None = None
     chat_template: str | None = None
     chat_template_file: str | None = None
+    server_path: str | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -94,7 +101,9 @@ class ServerProcess:
             self.server_port = int(os.environ["PORT"])
 
     def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
-        if "LLAMA_SERVER_BIN_PATH" in os.environ:
+        if self.server_path is not None:
+            server_path = self.server_path
+        elif "LLAMA_SERVER_BIN_PATH" in os.environ:
             server_path = os.environ["LLAMA_SERVER_BIN_PATH"]
         elif os.name == "nt":
             server_path = "../../../build/bin/Release/llama-server.exe"
@@ -148,6 +157,12 @@ class ServerProcess:
             server_args.extend(["--ctx-size", self.n_ctx])
         if self.n_slots:
             server_args.extend(["--parallel", self.n_slots])
+        if self.ctk:
+            server_args.extend(["-ctk", self.ctk])
+        if self.ctv:
+            server_args.extend(["-ctv", self.ctv])
+        if self.fa is not None:
+            server_args.append("-fa")
         if self.n_predict:
             server_args.extend(["--n-predict", self.n_predict])
         if self.slot_save_path:
@@ -181,7 +196,7 @@ class ServerProcess:
             server_args.extend(["--chat-template-file", self.chat_template_file])
 
         args = [str(arg) for arg in [server_path, *server_args]]
-        print(f"bench: starting server with: {' '.join(args)}")
+        print(f"tests: starting server with: {' '.join(args)}")
 
         flags = 0
         if "nt" == os.name:
@@ -212,6 +227,10 @@ class ServerProcess:
                     return  # server is ready
             except Exception as e:
                 pass
+            # Check if process died
+            if self.process.poll() is not None:
+                raise RuntimeError(f"Server process died with return code {self.process.returncode}")
+
             print(f"Waiting for server to start...")
             time.sleep(0.5)
         raise TimeoutError(f"Server did not start within {timeout_seconds} seconds")
@@ -283,7 +302,7 @@ class ServerPreset:
         server.model_hf_repo = "ggml-org/models"
         server.model_hf_file = "tinyllamas/stories260K.gguf"
         server.model_alias = "tinyllama-2"
-        server.n_ctx = 256
+        server.n_ctx = 512
         server.n_batch = 32
         server.n_slots = 2
         server.n_predict = 64

examples/server/utils.hpp

@@ -58,6 +58,32 @@ static T json_value(const json & body, const std::string & key, const T & defaul
 
 const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
 
+// thin wrapper around common_grammar_trigger with (de)serialization functions
+struct server_grammar_trigger {
+    common_grammar_trigger value;
+
+    server_grammar_trigger() = default;
+    server_grammar_trigger(const common_grammar_trigger & value) : value(value) {}
+    server_grammar_trigger(const json & in) {
+        value.type = (common_grammar_trigger_type) in.at("type").get<int>();
+        value.value = in.at("value").get<std::string>();
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            value.token = (llama_token) in.at("token").get<int>();
+        }
+    }
+
+    json to_json() const {
+        json out {
+            {"type", (int) value.type},
+            {"value", value.value},
+        };
+        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
+            out["token"] = (int) value.token;
+        }
+        return out;
+    }
+};
+
 //
 // tokenizer and input processing utils
 //
@@ -435,6 +461,10 @@ static std::string gen_chatcmplid() {
     return "chatcmpl-" + random_string();
 }
 
+static std::string gen_tool_call_id() {
+    return random_string();
+}
+
 //
 // other common utils
 //
@@ -590,8 +620,8 @@ static json oaicompat_completion_params_parse(
         if (response_type == "json_object") {
             json_schema = json_value(response_format, "schema", json::object());
         } else if (response_type == "json_schema") {
-            json json_schema = json_value(response_format, "json_schema", json::object());
-            json_schema = json_value(json_schema, "schema", json::object());
+            auto schema_wrapper = json_value(response_format, "json_schema", json::object());
+            json_schema = json_value(schema_wrapper, "schema", json::object());
         } else if (!response_type.empty() && response_type != "text") {
             throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
         }
@@ -607,6 +637,7 @@ static json oaicompat_completion_params_parse(
     inputs.use_jinja             = use_jinja;
     inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
     inputs.extract_reasoning     = reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
     if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
         throw std::runtime_error("Cannot use custom grammar constraints with tools.");
     }
@@ -616,14 +647,14 @@ static json oaicompat_completion_params_parse(
 
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
     llama_params["prompt"]           = chat_params.prompt;
-    llama_params["grammar"]          = chat_params.grammar;
+    if (!chat_params.grammar.empty()) {
+        llama_params["grammar"] = chat_params.grammar;
+    }
     llama_params["grammar_lazy"]     = chat_params.grammar_lazy;
     auto grammar_triggers = json::array();
     for (const auto & trigger : chat_params.grammar_triggers) {
-        grammar_triggers.push_back({
-            {"word", trigger.word},
-            {"at_start", trigger.at_start},
-        });
+        server_grammar_trigger ct(trigger);
+        grammar_triggers.push_back(ct.to_json());
     }
     llama_params["grammar_triggers"] = grammar_triggers;
     llama_params["preserved_tokens"] = chat_params.preserved_tokens;

examples/server/webui/package.json

@@ -13,9 +13,11 @@
   "dependencies": {
     "@heroicons/react": "^2.2.0",
     "@sec-ant/readable-stream": "^0.6.0",
+    "@tailwindcss/postcss": "^4.1.1",
+    "@tailwindcss/vite": "^4.1.1",
     "@vscode/markdown-it-katex": "^1.1.1",
     "autoprefixer": "^10.4.20",
-    "daisyui": "^4.12.14",
+    "daisyui": "^5.0.12",
     "dexie": "^4.0.11",
     "highlight.js": "^11.10.0",
     "katex": "^0.16.15",
@@ -29,7 +31,7 @@
     "remark-breaks": "^4.0.0",
     "remark-gfm": "^4.0.0",
     "remark-math": "^6.0.0",
-    "tailwindcss": "^3.4.15",
+    "tailwindcss": "^4.1.1",
     "textlinestream": "^1.1.1",
     "vite-plugin-singlefile": "^2.0.3"
   },

examples/server/webui/postcss.config.js

@@ -1,6 +1,5 @@
 export default {
   plugins: {
-    tailwindcss: {},
-    autoprefixer: {},
+    "@tailwindcss/postcss": {},
   },
 }

examples/server/webui/src/App.tsx

@@ -28,7 +28,7 @@ function AppLayout() {
     <>
       <Sidebar />
       <div
-        className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto"
+        className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto bg-base-100"
         id="main-scroll"
       >
         <Header />

examples/server/webui/src/Config.ts

@@ -1,4 +1,4 @@
-import daisyuiThemes from 'daisyui/src/theming/themes';
+import daisyuiThemes from 'daisyui/theme/object';
 import { isNumeric } from './utils/misc';
 
 export const isDev = import.meta.env.MODE === 'development';

examples/server/webui/src/components/ChatScreen.tsx

@@ -2,7 +2,7 @@ import { useEffect, useMemo, useRef, useState } from 'react';
 import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
 import ChatMessage from './ChatMessage';
 import { CanvasType, Message, PendingMessage } from '../utils/types';
-import { classNames, throttle } from '../utils/misc';
+import { classNames, cleanCurrentUrl, throttle } from '../utils/misc';
 import CanvasPyInterpreter from './CanvasPyInterpreter';
 import StorageUtils from '../utils/storage';
 import { useVSCodeContext } from '../utils/llama-vscode';
@@ -18,6 +18,24 @@ export interface MessageDisplay {
   isPending?: boolean;
 }
 
+/**
+ * If the current URL contains "?m=...", prefill the message input with the value.
+ * If the current URL contains "?q=...", prefill and SEND the message.
+ */
+const prefilledMsg = {
+  content() {
+    const url = new URL(window.location.href);
+    return url.searchParams.get('m') ?? url.searchParams.get('q') ?? '';
+  },
+  shouldSend() {
+    const url = new URL(window.location.href);
+    return url.searchParams.has('q');
+  },
+  clear() {
+    cleanCurrentUrl(['m', 'q']);
+  },
+};
+
 function getListMessageDisplay(
   msgs: Readonly<Message[]>,
   leafNodeId: Message['id']
@@ -81,13 +99,9 @@ export default function ChatScreen() {
     canvasData,
     replaceMessageAndGenerate,
   } = useAppContext();
-  const [inputMsg, setInputMsg] = useState('');
-  const inputRef = useRef<HTMLTextAreaElement>(null);
+  const textarea = useOptimizedTextarea(prefilledMsg.content());
 
-  const { extraContext, clearExtraContext } = useVSCodeContext(
-    inputRef,
-    setInputMsg
-  );
+  const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
   // TODO: improve this when we have "upload file" feature
   const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
 
@@ -117,9 +131,10 @@ export default function ChatScreen() {
   };
 
   const sendNewMessage = async () => {
-    if (inputMsg.trim().length === 0 || isGenerating(currConvId ?? '')) return;
-    const lastInpMsg = inputMsg;
-    setInputMsg('');
+    const lastInpMsg = textarea.value();
+    if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? ''))
+      return;
+    textarea.setValue('');
     scrollToBottom(false);
     setCurrNodeId(-1);
     // get the last message node
@@ -128,13 +143,13 @@ export default function ChatScreen() {
       !(await sendMessage(
         currConvId,
         lastMsgNodeId,
-        inputMsg,
+        lastInpMsg,
         currExtra,
         onChunk
       ))
     ) {
       // restore the input message if failed
-      setInputMsg(lastInpMsg);
+      textarea.setValue(lastInpMsg);
     }
     // OK
     clearExtraContext();
@@ -172,6 +187,19 @@ export default function ChatScreen() {
 
   const hasCanvas = !!canvasData;
 
+  useEffect(() => {
+    if (prefilledMsg.shouldSend()) {
+      // send the prefilled message if needed
+      sendNewMessage();
+    } else {
+      // otherwise, focus on the input
+      textarea.focus();
+    }
+    prefilledMsg.clear();
+    // no need to keep track of sendNewMessage
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [textarea.ref]);
+
   // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg)
   const pendingMsgDisplay: MessageDisplay[] =
     pendingMsg && messages.at(-1)?.msg.id !== pendingMsg.id
@@ -224,9 +252,7 @@ export default function ChatScreen() {
           <textarea
             className="textarea textarea-bordered w-full"
             placeholder="Type a message (Shift+Enter to add a new line)"
-            ref={inputRef}
-            value={inputMsg}
-            onChange={(e) => setInputMsg(e.target.value)}
+            ref={textarea.ref}
             onKeyDown={(e) => {
               if (e.nativeEvent.isComposing || e.keyCode === 229) return;
               if (e.key === 'Enter' && e.shiftKey) return;
@@ -246,11 +272,7 @@ export default function ChatScreen() {
               Stop
             </button>
           ) : (
-            <button
-              className="btn btn-primary ml-2"
-              onClick={sendNewMessage}
-              disabled={inputMsg.trim().length === 0}
-            >
+            <button className="btn btn-primary ml-2" onClick={sendNewMessage}>
               Send
             </button>
           )}
@@ -264,3 +286,43 @@ export default function ChatScreen() {
     </div>
   );
 }
+
+export interface OptimizedTextareaValue {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  useEffect(() => {
+    if (textareaRef.current && savedInitValue) {
+      textareaRef.current.value = savedInitValue;
+      setSavedInitValue('');
+    }
+  }, [textareaRef, savedInitValue, setSavedInitValue]);
+
+  return {
+    value: () => {
+      return textareaRef.current?.value ?? savedInitValue;
+    },
+    setValue: (value: string) => {
+      if (textareaRef.current) {
+        textareaRef.current.value = value;
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        // focus and move the cursor to the end
+        textareaRef.current.focus();
+        textareaRef.current.selectionStart = textareaRef.current.value.length;
+      }
+    },
+    ref: textareaRef,
+  };
+}

examples/server/webui/src/components/Header.tsx

@@ -2,7 +2,7 @@ import { useEffect, useState } from 'react';
 import StorageUtils from '../utils/storage';
 import { useAppContext } from '../utils/app.context';
 import { classNames } from '../utils/misc';
-import daisyuiThemes from 'daisyui/src/theming/themes';
+import daisyuiThemes from 'daisyui/theme/object';
 import { THEMES } from '../Config';
 import { useNavigate } from 'react-router';
 
@@ -20,7 +20,6 @@ export default function Header() {
     document.body.setAttribute('data-theme', selectedTheme);
     document.body.setAttribute(
       'data-color-scheme',
-      // @ts-expect-error daisyuiThemes complains about index type, but it should work
       daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto'
     );
   }, [selectedTheme]);

examples/server/webui/src/components/SettingDialog.tsx

@@ -148,13 +148,13 @@ const SETTING_SECTIONS: SettingSection[] = [
     fields: [
       {
         type: SettingInputType.CHECKBOX,
-        label: 'Expand though process by default for generating message',
+        label: 'Expand thought process by default when generating messages',
         key: 'showThoughtInProgress',
       },
       {
         type: SettingInputType.CHECKBOX,
         label:
-          'Exclude thought process when sending request to API (Recommended for DeepSeek-R1)',
+          'Exclude thought process when sending requests to API (Recommended for DeepSeek-R1)',
         key: 'excludeThoughtOnReq',
       },
     ],
@@ -247,7 +247,7 @@ const SETTING_SECTIONS: SettingSection[] = [
               This feature uses{' '}
               <OpenInNewTab href="https://pyodide.org">pyodide</OpenInNewTab>,
               downloaded from CDN. To use this feature, ask the LLM to generate
-              python code inside a markdown code block. You will see a "Run"
+              Python code inside a Markdown code block. You will see a "Run"
               button on the code block, near the "Copy" button.
             </small>
           </>
@@ -274,7 +274,7 @@ export default function SettingDialog({
   );
 
   const resetConfig = () => {
-    if (window.confirm('Are you sure to reset all settings?')) {
+    if (window.confirm('Are you sure you want to reset all settings?')) {
       setLocalConfig(CONFIG_DEFAULT);
     }
   };
@@ -296,9 +296,9 @@ export default function SettingDialog({
           return;
         }
       } else if (mustBeNumeric) {
-        const trimedValue = value.toString().trim();
-        const numVal = Number(trimedValue);
-        if (isNaN(numVal) || !isNumeric(numVal) || trimedValue.length === 0) {
+        const trimmedValue = value.toString().trim();
+        const numVal = Number(trimmedValue);
+        if (isNaN(numVal) || !isNumeric(numVal) || trimmedValue.length === 0) {
           alert(`Value for ${key} must be numeric`);
           return;
         }