server : fix default draft model parameters (#10586)

* server : force F16 KV cache for the draft model

ggml-ci

* server : fix draft params

ggml-ci

* server : various params fixes

ggml-ci

.clang-tidy

@@ -17,8 +17,10 @@ Checks: >
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,
     portability-*,
+    -portability-simd-intrinsics,
     misc-*,
     -misc-const-correctness,
     -misc-non-private-member-variables-in-classes,
     -misc-no-recursion,
+    -misc-use-anonymous-namespace,
 FormatStyle: none

.devops/llama-server.Dockerfile

@@ -3,22 +3,34 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-ENV LLAMA_CURL=1
 
-RUN make -j$(nproc) llama-server
+RUN \
+    # Build multiple versions of the CPU backend
+    scripts/build-cpu.sh avx         -DGGML_AVX=ON -DGGML_AVX2=OFF && \
+    scripts/build-cpu.sh avx2        -DGGML_AVX=ON -DGGML_AVX2=ON && \
+    scripts/build-cpu.sh avx512      -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON && \
+    scripts/build-cpu.sh amx         -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX_VNNI=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \
+    # Build llama-server
+    cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build --target llama-server -j $(nproc) && \
+    # Copy the built libraries to /app/lib
+    mkdir -p /app/lib && \
+    mv libggml-cpu* /app/lib/ && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
 
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/build/bin/llama-server /llama-server
+COPY --from=build /app/lib/ /
 
 ENV LC_ALL=C.utf8
 # Must be set to 0.0.0.0 so it can listen to requests from host machine

.github/pull_request_template.md

@@ -1,7 +1 @@
-
-
-- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
-- Self-reported review complexity:
-  - [ ] Low
-  - [ ] Medium
-  - [ ] High
+*Make sure to read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*

.github/workflows/build.yml

@@ -160,66 +160,6 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
           name: llama-bin-macos-x64.zip
 
-  ubuntu-focal-make:
-    runs-on: ubuntu-20.04
-    env:
-      LLAMA_NODE_AVAILABLE: true
-      LLAMA_PYTHON_AVAILABLE: true
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential gcc-8
-
-      - uses: actions/setup-node@v4
-        with:
-          node-version: "20"
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-
-      - name: Build
-        id: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
-        run: |
-          CC=gcc-8 make -j $(nproc)
-
-      - name: Test
-        id: make_test
-        run: |
-          CC=gcc-8 make tests -j $(nproc)
-          make test -j $(nproc)
-
-  ubuntu-focal-make-curl:
-    runs-on: ubuntu-20.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
-
-      - name: Build
-        id: make_build
-        env:
-          LLAMA_FATAL_WARNINGS: 1
-          LLAMA_CURL: 1
-        run: |
-          CC=gcc-8 make -j $(nproc)
-
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
 
@@ -517,36 +457,6 @@ jobs:
           cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON ..
           cmake --build . --config Release -j $(nproc)
 
-  # TODO: build with GGML_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
-  #       how to debug it.
-  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
-  macOS-latest-make:
-    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: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
-        run: |
-          GGML_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: make_test
-        run: |
-          GGML_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
-          GGML_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
-
   # TODO: build with GGML_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
   #       how to debug it.
   #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
@@ -642,33 +552,35 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-  macOS-latest-swift:
-    runs-on: macos-latest
-
-    strategy:
-      matrix:
-        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
-
-      - name: Build Swift Example
-        id: make_build_swift_example
-        run: |
-            make swift
+# TODO: tmp disabled. see for possible re-enable:
+#       https://github.com/ggerganov/llama.cpp/pull/10525
+#  macOS-latest-swift:
+#    runs-on: macos-latest
+#
+#    strategy:
+#      matrix:
+#        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        id: depends
+#        continue-on-error: true
+#        run: |
+#          brew update
+#
+#      - name: xcodebuild for swift package
+#        id: xcodebuild
+#        run: |
+#          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
+#
+#      - name: Build Swift Example
+#        id: make_build_swift_example
+#        run: |
+#            make swift
 
   windows-msys2:
     runs-on: windows-latest
@@ -695,21 +607,6 @@ jobs:
             mingw-w64-${{matrix.env}}-cmake
             mingw-w64-${{matrix.env}}-openblas
 
-      - name: Build using make
-        shell: msys2 {0}
-        run: |
-            make -j $(nproc)
-
-      - name: Clean after building using make
-        shell: msys2 {0}
-        run: |
-            make clean
-
-      - name: Build using make w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            make GGML_OPENBLAS=1 -j $(nproc)
-
       - name: Build using CMake
         shell: msys2 {0}
         run: |
@@ -904,6 +801,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+            fetch-depth: 0
 
       - name: Install Cuda Toolkit 11.7
         if: ${{ matrix.cuda == '11.7' }}
@@ -1119,6 +1018,11 @@ jobs:
         run: |
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2
+        with:
+          key: ${{ github.job }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -1139,6 +1043,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+            fetch-depth: 0
 
       - name: Install
         id: depends
@@ -1248,9 +1154,7 @@ jobs:
     runs-on: ubuntu-latest
 
     needs:
-      - ubuntu-focal-make
       - ubuntu-latest-cmake
-      - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-2019-cmake-cuda

AUTHORS

@@ -1,4 +1,4 @@
-# date: Wed Jun 26 19:36:34 EEST 2024
+# date: Thu Nov 28 20:46:15 EET 2024
 # this file is auto-generated by scripts/gen-authors.sh
 
 0cc4m <picard12@live.de>
@@ -7,6 +7,7 @@
 2f38b454 <dxf@protonmail.com>
 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
 44670 <44670@users.noreply.github.com>
+65a <10104049+65a@users.noreply.github.com>
 AN Long <aisk@users.noreply.github.com>
 AT <manyoso@users.noreply.github.com>
 Aarni Koskela <akx@iki.fi>
@@ -19,20 +20,28 @@ Adithya Balaji <adithya.b94@gmail.com>
 AdithyanI <adithyan.i4internet@gmail.com>
 Adrian <smith.adriane@gmail.com>
 Adrian Hesketh <a-h@users.noreply.github.com>
+Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com>
 Ahmet Zeer <ahmed.zeer@std.yildiz.edu.tr>
 AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
+AidanBeltonS <aidan.belton@codeplay.com>
 Aisuko <urakiny@gmail.com>
+Akarshan Biswas <akarshan.biswas@gmail.com>
 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>
 Alex <awhill19@icloud.com>
 Alex Azarov <alex@azarov.by>
 Alex Azarov <alexander.azarov@mapbox.com>
 Alex Klinkhamer <from.github.com.917@grencez.dev>
 Alex Klinkhamer <git@grencez.dev>
 Alex Nguyen <tiendung@users.noreply.github.com>
+Alex O'Connell <35843486+acon96@users.noreply.github.com>
 Alex Petenchea <alex.petenchea@gmail.com>
 Alex Renda <alexrenda@users.noreply.github.com>
+Alex Tuddenham <61622354+AlexsCode@users.noreply.github.com>
 Alex von Gluck IV <kallisti5@unixzen.com>
 Alexey Parfenov <zxed@alkatrazstudio.net>
 Ali Chraghi <63465728+alichraghi@users.noreply.github.com>
@@ -45,18 +54,25 @@ AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
 Ananta Bastola <anantarajbastola@gmail.com>
 Anas Ahouzi <112881240+aahouzi@users.noreply.github.com>
 András Salamon <ott2@users.noreply.github.com>
+Andreas (Andi) Kunar <andreask@msn.com>
 Andrei <abetlen@gmail.com>
 Andrew Canis <andrew.canis@gmail.com>
 Andrew Downing <andrew2085@gmail.com>
 Andrew Duffy <a10y@users.noreply.github.com>
 Andrew Godfrey <AndrewGodfrey@users.noreply.github.com>
+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>
+Antonis Makropoulos <benuix@gmail.com>
 Arik Poznanski <arikpoz@users.noreply.github.com>
+Armen Kaleshian <kriation@users.noreply.github.com>
 Artem <guinmoon@gmail.com>
 Artem Zinnatullin <ceo@abstractny.gay>
 Artyom Lebedev <vagran.ast@gmail.com>
 Asbjørn Olling <asbjornolling@gmail.com>
 Ásgeir Bjarni Ingvarsson <asgeir@fundinn.org>
+Asghar Ghorbani <a-ghorbani@users.noreply.github.com>
 Ashish <1856117+ashishdatta@users.noreply.github.com>
 Ashok Gelal <401055+ashokgelal@users.noreply.github.com>
 Ashraful Islam <ashraful.meche@gmail.com>
@@ -76,12 +92,16 @@ Ben Williams <ben@719ben.com>
 Benjamin Findley <39356821+Kartoffelsaft@users.noreply.github.com>
 Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com>
 Bernat Vadell <hounter.caza@gmail.com>
+Bert Wagner <github@bertwagner.com>
 Bingan <70050083+binganao@users.noreply.github.com>
+Bjarke Viksøe <164612031+bviksoe@users.noreply.github.com>
 Bodo Graumann <mail@bodograumann.de>
 Bono Lv <lvscar@users.noreply.github.com>
 Borislav Stanimirov <b.stanimirov@abv.bg>
 Branden Butler <bwtbutler@hotmail.com>
+Brandon Squizzato <35474886+bsquizz@users.noreply.github.com>
 Brian <mofosyne@gmail.com>
+Brian Cunnie <brian.cunnie@gmail.com>
 Bruce MacDonald <brucewmacdonald@gmail.com>
 Bryan Honof <bryanhonof@gmail.com>
 CJ Pais <cj@cjpais.com>
@@ -90,32 +110,47 @@ Calvin Laurenson <calvin@laurenson.dev>
 Cameron <csteele@steelecameron.com>
 Cameron Kaiser <classilla@users.noreply.github.com>
 Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
+CarryFun <76023481+CarryFun@users.noreply.github.com>
+Carsten Kragelund Jørgensen <carsten@kragelund.me>
+CarterLi999 <664681047@qq.com>
 Casey Primozic <casey@cprimozic.net>
 Casey Primozic <me@ameo.link>
 CausalLM <148736309+CausalLM@users.noreply.github.com>
 Cebtenzzre <cebtenzzre@gmail.com>
 Chad Brewbaker <crb002@gmail.com>
+Changyeon Kim <cyzero.kim@samsung.com>
 Chao Jiang <jc19chaoj@zoho.com>
+Charles Xu <63788048+chaxu01@users.noreply.github.com>
+Charles Xu <charles.xu@arm.com>
+Chen Xi <xi2.chen@intel.com>
+Chen Xi <xixichen08@foxmail.com>
 Cheng Shao <terrorjack@type.dance>
+Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
 Chris Elrod <elrodc@gmail.com>
 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 Kögler <ck3d@gmx.de>
+Christian Köhnenkamp <cvk5@me.com>
 Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com>
 Clark Saben <76020733+csaben@users.noreply.github.com>
 Clint Herron <hanclinto@gmail.com>
+Conrad Kramer <conrad@conradkramer.com>
 CrispStrobe <154636388+CrispStrobe@users.noreply.github.com>
+Csaba Kecskemeti <csaba.kecskemeti@gmail.com>
 Cuong Trinh Manh <nguoithichkhampha@gmail.com>
 DAN™ <dranger003@gmail.com>
 Damian Stewart <d@damianstewart.com>
+Dan Johansson <164997844+eddnjjn@users.noreply.github.com>
+Dan Johansson <dan.johansson@arm.com>
 Dane Madsen <dane_madsen@hotmail.com>
 DaniAndTheWeb <57776841+DaniAndTheWeb@users.noreply.github.com>
 Daniel Bevenius <daniel.bevenius@gmail.com>
 Daniel Drake <drake@endlessos.org>
 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>
 DannyDaemonic <DannyDaemonic@gmail.com>
 Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
@@ -129,19 +164,28 @@ David Pflug <david@pflug.email>
 David Renshaw <dwrenshaw@gmail.com>
 David Sommers <12738+databyte@users.noreply.github.com>
 David Yang <davidyang6us@gmail.com>
+DavidKorczynski <david@adalogics.com>
 Dawid Potocki <github@dawidpotocki.com>
 Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
 Dean <Dean.Sinaean@gmail.com>
 Deins <deinsegle@gmail.com>
+Denis Spasyuk <34203011+dspasyuk@users.noreply.github.com>
+Derrick T. Woolworth <dwoolworth@gmail.com>
 Deven Mistry <31466137+deven367@users.noreply.github.com>
+Dibakar Gope <dibakar.gope@arm.com>
 Didzis Gosko <didzis@users.noreply.github.com>
+Diego Devesa <slarengh@gmail.com>
+Diogo Teles Sant'Anna <diogoteles@google.com>
 Djip007 <djip.perois@free.fr>
 Don Mahurin <dmahurin@users.noreply.github.com>
 DooWoong Lee (David) <manics99@naver.com>
 Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
+Dou Xinpeng <15529241576@163.com>
+Dou Xinpeng <81913537+Dou-Git@users.noreply.github.com>
 Douglas Hanley <thesecretaryofwar@gmail.com>
 Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
 Ebey Abraham <ebey97@gmail.com>
+Echo Nolan <echo@echonolan.net>
 Ed Lee <edilee@mozilla.com>
 Ed Lepedus <ed.lepedus@googlemail.com>
 Eddie-Wang <wangjinheng1120@163.com>
@@ -151,10 +195,13 @@ Elbios <141279586+Elbios@users.noreply.github.com>
 Elton Kola <eltonkola@gmail.com>
 Engininja2 <139037756+Engininja2@users.noreply.github.com>
 Equim <sayaka@ekyu.moe>
+Eric Curtin <ecurtin@redhat.com>
+Eric Curtin <ericcurtin17@gmail.com>
 Eric Sommerlade <es0m@users.noreply.github.com>
 Eric Zhang <34133756+EZForever@users.noreply.github.com>
 Erik Garrison <erik.garrison@gmail.com>
 Erik Scholz <Green-Sky@users.noreply.github.com>
+Esko Toivonen <eskot98@gmail.com>
 Ettore Di Giacinto <mudler@users.noreply.github.com>
 Evan Jones <evan.q.jones@gmail.com>
 Evan Miller <emmiller@gmail.com>
@@ -166,19 +213,26 @@ FK <sozforex@gmail.com>
 Fabian <cmdrf@users.noreply.github.com>
 Fabio R. Sluzala <Fabio3rs@users.noreply.github.com>
 Faez Shakil <faez.shakil@gmail.com>
+Faisal Zaghloul <faisal.zaghloul@gmail.com>
+Faisal Zaghloul <quic_fzaghlou@quicinc.com>
+Fan Shupei <dymarkfan@outlook.com>
 FantasyGmm <16450052+FantasyGmm@users.noreply.github.com>
+Farbod Bijary <110523279+farbodbj@users.noreply.github.com>
 Fattire <528174+fat-tire@users.noreply.github.com>
 Felix <stenbackfelix@gmail.com>
 Finn Voorhees <finnvoorhees@gmail.com>
 Firat <firatkiral@gmail.com>
+FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.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>
 Frank Mai <thxcode0824@gmail.com>
 FrankHB <frankhb1989@gmail.com>
+Frankie Robertson <frankier@users.noreply.github.com>
 Fred Douglas <43351173+fredlas@users.noreply.github.com>
 Frederik Vogel <Schaltfehler@users.noreply.github.com>
 Gabe Goodhart <gabe.l.hart@gmail.com>
+Gabe Goodhart <ghart@us.ibm.com>
 GainLee <perfecter.gen@gmail.com>
 Galunid <karolek1231456@gmail.com>
 Gary Linscott <glinscott@gmail.com>
@@ -187,11 +241,13 @@ Gavin Zhao <gavinzhaojw@protonmail.com>
 Genkagaku.GPT <hlhr202@163.com>
 Georgi Gerganov <ggerganov@gmail.com>
 Gilad S <giladgd@users.noreply.github.com>
+Gilad S. <7817232+giladgd@users.noreply.github.com>
 Giuseppe Scrivano <giuseppe@scrivano.org>
 GiviMAD <GiviMAD@users.noreply.github.com>
 Govlzkoy <gotope@users.noreply.github.com>
 Guillaume "Vermeille" Sanchez <Guillaume.V.Sanchez@gmail.com>
 Guillaume Wenzek <gwenzek@users.noreply.github.com>
+Guoliang Hua <32868157+nbcsm@users.noreply.github.com>
 Guoteng <32697156+SolenoidWGT@users.noreply.github.com>
 Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
 Haggai Nuchi <h.nuchi@gmail.com>
@@ -213,11 +269,14 @@ Hong Bo PENG <penghb@cn.ibm.com>
 Hongyu Ouyang <96765450+casavaca@users.noreply.github.com>
 Howard Su <howard0su@gmail.com>
 Hua Jiang <allenhjiang@outlook.com>
+Huang Qi <huangqi3@xiaomi.com>
 Huawei Lin <huaweilin.cs@gmail.com>
 Hugo Roussel <hugo.rous@gmail.com>
+Huifeng Ou <79071290+ho2103@users.noreply.github.com>
 Ian Bull <irbull@eclipsesource.com>
 Ian Bull <irbull@gmail.com>
 Ian Scrivener <github@zilogy.asia>
+Icecream95 <the.real.icecream95@gmail.com>
 Ido S <ido.pluto@gmail.com>
 IgnacioFDM <ignaciofdm@gmail.com>
 Igor Okulist <okigan@gmail.com>
@@ -226,11 +285,15 @@ Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
 Ionoclast Laboratories <brigham@ionoclast.com>
 Isaac McFadyen <isaac@imcf.me>
 IsaacDynamo <61521674+IsaacDynamo@users.noreply.github.com>
+Ivan <nekotekina@gmail.com>
+Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
 Ivan Komarov <Ivan.Komarov@dfyz.info>
 Ivan Stepanov <ivanstepanovftw@gmail.com>
 JH23X <165871467+JH23X@users.noreply.github.com>
+Jack Mousseau <jack@software.inc>
 Jack Mousseau <jmousseau@users.noreply.github.com>
 JackJollimore <130917767+JackJollimore@users.noreply.github.com>
+Jaeden Amero <jaeden@patater.com>
 Jaemin Son <woalsdnd@gmail.com>
 Jag Chadha <jagtesh@gmail.com>
 Jakub N <jakubniemczyk97@gmail.com>
@@ -243,10 +306,14 @@ Jannis Schönleber <joennlae@gmail.com>
 Jared Van Bortel <cebtenzzre@gmail.com>
 Jared Van Bortel <jared@nomic.ai>
 Jason McCartney <jmac@theroot.org>
+Jason Stillerman <jason.t.stillerman@gmail.com>
 Jean-Christophe Hoelt <hoelt@fovea.cc>
 Jean-Michaël Celerier <jeanmichael.celerier+github@gmail.com>
 Jed Fox <git@jedfox.com>
+Jeff Bolz <jbolz@nvidia.com>
+Jeffrey Morgan <jmorganca@gmail.com>
 Jeffrey Quesnelle <emozilla@nousresearch.com>
+Jeroen Mostert <jeroen.mostert@cm.com>
 Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
 Jeximo <jeximo@gmail.com>
 Jhen-Jie Hong <iainst0409@gmail.com>
@@ -258,6 +325,9 @@ Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
 Jiří Sejkora <Sejseloid@gmail.com>
 Joan Fontanals <jfontanalsmartinez@gmail.com>
 Joan Fontanals <joan.fontanals.martinez@jina.ai>
+João Dinis Ferreira <hello@joaof.eu>
+Joe Eli McIlvain <joe.eli.mac@gmail.com>
+Joe Todd <joe.todd@codeplay.com>
 Johan <JohanAR@users.noreply.github.com>
 Johannes Gäßler <johannesg@5d6.de>
 Johannes Rudolph <johannes.rudolph@gmail.com>
@@ -274,7 +344,9 @@ Joyce <joycebrum@google.com>
 Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
 Judd <foldl@users.noreply.github.com>
 Julius Arkenberg <arki05@users.noreply.github.com>
+Jun Hee Yoo <contact.jhyoo@gmail.com>
 Jun Jie <71215065+junnjiee16@users.noreply.github.com>
+Junil Kim <logyourself@gmail.com>
 Junyang Lin <justinlin930319@hotmail.com>
 Juraj Bednar <juraj@bednar.io>
 Justin Parker <jparkerweb@gmail.com>
@@ -292,12 +364,14 @@ Karthik Sethuraman <k.seth1993@gmail.com>
 Kasumi <90275229+kasumi-1@users.noreply.github.com>
 Kawrakow <48489457+ikawrakow@users.noreply.github.com>
 Keiichi Tabata <keiichi.tabata@outlook.com>
+Keke Han <hankeke303@163.com>
 Kenvix ⭐ <kenvixzure@live.com>
 Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
 Kevin Gibbons <bakkot@gmail.com>
 Kevin Ji <1146876+kevinji@users.noreply.github.com>
 Kevin Kwok <antimatter15@gmail.com>
 Kevin Lo <kevlo@kevlo.org>
+Kevin Wang <kevmo314@gmail.com>
 Kolen Cheung <ickc@users.noreply.github.com>
 Konstantin Herud <konstantin.herud@denkbares.com>
 Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
@@ -315,22 +389,29 @@ LeonEricsson <70749762+LeonEricsson@users.noreply.github.com>
 Leonardo Neumann <leonardo@neumann.dev.br>
 Li Tan <tanliboy@gmail.com>
 Linwei Wang <wanix1988@gmail.com>
+Liu Jia <109258120+Septa2112@users.noreply.github.com>
+Liu Jia <jia3.liu@intel.com>
 LoganDark <github@logandark.mozmail.com>
+Loïc Carrère <loic.carrere@gmail.com>
 LostRuins <39025047+LostRuins@users.noreply.github.com>
 Luciano <lucianostrika44@gmail.com>
 Luo Tian <lt@basecity.com>
 Lyle Dean <dean@lyle.dev>
+M-A <maruel@gmail.com>
 M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
+Ma Mingfei <mingfei.ma@intel.com>
 Maarten ter Huurne <maarten@treewalker.org>
 Mack Straight <eiz@users.noreply.github.com>
 Maël Kerbiriou <m431.kerbiriou@gmail.com>
 MaggotHATE <clay1326@gmail.com>
+Mahesh Madhav <67384846+heshpdx@users.noreply.github.com>
 Manuel <44313466+makuche@users.noreply.github.com>
 Marc Köhlbrugge <subscriptions@marckohlbrugge.com>
 Marco Matthies <71844+marcom@users.noreply.github.com>
 Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
 Marian Cepok <marian.cepok@gmail.com>
 Mark Fairbairn <thebaron88@gmail.com>
+Mark Zhuang <zhuangqiubin@gmail.com>
 Marko Tasic <mtasic85@gmail.com>
 Markus Tavenrath <mtavenrath@users.noreply.github.com>
 Martin Delille <martin@delille.org>
@@ -342,11 +423,15 @@ MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com>
 Mateusz Charytoniuk <mateusz.charytoniuk@protonmail.com>
 Matheus C. França <matheus-catarino@hotmail.com>
 Matheus Gabriel Alves Silva <matheusgasource@gmail.com>
+Mathieu Geli <mathieu.geli@gmail.com>
 Mathieu Nayrolles <MathieuNls@users.noreply.github.com>
+Mathijs Henquet <mathijs.henquet@gmail.com>
 Mathijs de Bruin <mathijs@mathijsfietst.nl>
 Matt Clayton <156335168+mattjcly@users.noreply.github.com>
 Matt Pulver <matt.pulver@heavy.ai>
+Matt Stephenson <mstephenson6@users.noreply.github.com>
 Matteo Boschini <12133566+mbosc@users.noreply.github.com>
+Matteo Mortari <matteo.mortari@gmail.com>
 Mattheus Chediak <shammcity00@gmail.com>
 Matthew Tejo <matthew.tejo@gmail.com>
 Matvey Soloviev <blackhole89@gmail.com>
@@ -356,8 +441,10 @@ Maxime <672982+maximegmd@users.noreply.github.com>
 Maximilian Winter <maximilian.winter.91@gmail.com>
 Meng Zhang <meng@tabbyml.com>
 Meng, Hengyu <hengyu.meng@intel.com>
+Mengqing Cao <cmq0113@163.com>
 Merrick Christensen <merrick.christensen@gmail.com>
 Michael Coppola <m18coppola@gmail.com>
+Michael Francis <edude03@gmail.com>
 Michael Hueschen <m@mhueschen.dev>
 Michael Kesper <mkesper@schokokeks.org>
 Michael Klimenko <mklimenko29@gmail.com>
@@ -365,41 +452,57 @@ Michael Podvitskiy <podvitskiymichael@gmail.com>
 Michael Potter <NanoTekGuy@Gmail.com>
 Michael de Gans <michael.john.degans@gmail.com>
 Michaël de Vries <vriesdemichael@gmail.com>
+Michał Tuszyński <srgtuszy@gmail.com>
 Mihai <mihai.chirculescu@yahoo.com>
 Mike <ytianhui2004@gmail.com>
 Mikko Juola <mikjuo@gmail.com>
 Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
+Minsoo Cheong <icycle0409@snu.ac.kr>
 Mirko185 <mirkosig@gmail.com>
 Mirror Azure <54669636+MirrorAzure@users.noreply.github.com>
+MistApproach <98988043+MistApproach@users.noreply.github.com>
 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>
+MorganRO8 <47795945+MorganRO8@users.noreply.github.com>
 Murilo Santana <mvrilo@gmail.com>
 Musab Gultekin <musabgultekin@users.noreply.github.com>
 Nam D. Tran <42194884+namtranase@users.noreply.github.com>
 Nathan Epstein <nate2@umbc.edu>
+Natsu <chino@hotococoa.moe>
 NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
 Nebula <infinitewormhole@gmail.com>
 Neo Zhang <14088817+arthw@users.noreply.github.com>
 Neo Zhang <zhang.jianyu@outlook.com>
 Neo Zhang Jianyu <jianyu.zhang@intel.com>
 Neuman Vong <neuman.vong@gmail.com>
+Nexes the Old <124105151+Nexesenex@users.noreply.github.com>
 Nexesenex <124105151+Nexesenex@users.noreply.github.com>
 Niall Coates <1349685+Niall-@users.noreply.github.com>
+Nicholai Tukanov <nicholaitukanov@gmail.com>
+Nico Bosshard <nico@bosshome.ch>
 Nicolai Weitkemper <kontakt@nicolaiweitkemper.de>
 Nicolás Pérez <nicolas_perez@brown.edu>
 Nigel Bosch <pnigelb@gmail.com>
 Niklas Korz <niklas@niklaskorz.de>
+NikolaiLyssogor <59844691+NikolaiLyssogor@users.noreply.github.com>
 Nikolas <127742645+nneubacher@users.noreply.github.com>
 Nindaleth <Nindaleth@users.noreply.github.com>
+OSecret <135510162+OLSecret@users.noreply.github.com>
 Oleksandr Nikitin <oleksandr@tvori.info>
 Oleksii Maryshchenko <oleksii.maryshchenko@gmail.com>
 Olivier Chafik <ochafik@users.noreply.github.com>
 Ondřej Čertík <ondrej@certik.us>
 Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
+PAB <pierreantoine.bannier@gmail.com>
+Pablo Duboue <pablo.duboue@gmail.com>
+Pascal Patry <ppatry@mtacitlabs.com>
 Patrice Ferlet <metal3d@gmail.com>
 Paul Tsochantaris <ptsochantaris@icloud.com>
+Pavel Zloi <github.com@drteam.rocks>
 Pavol Rusnak <pavol@rusnak.io>
+Paweł Wodnicki <151604+32bitmicro@users.noreply.github.com>
 Pedro Cuenca <pedro@huggingface.co>
 Peter Sugihara <peter@campsh.com>
 Phil H <5756783+phiharri@users.noreply.github.com>
@@ -407,10 +510,15 @@ Philip Taron <philip.taron@gmail.com>
 Phillip Kravtsov <phillip@kravtsov.net>
 Pierre Alexandre SCHEMBRI <pa.schembri@gmail.com>
 Pierrick Hymbert <pierrick.hymbert@gmail.com>
+Pieter Ouwerkerk <pieter.ouwerkerk@gmail.com>
+Plamen Minev <pacominev@gmail.com>
+Prashant Vithule <119530321+Vithulep@users.noreply.github.com>
 Przemysław Pawełczyk <przemoc@gmail.com>
 Qin Yue Chen <71813199+chenqiny@users.noreply.github.com>
 Qingyou Meng <meng.qingyou@gmail.com>
 Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
+R0CKSTAR <xiaodong.ye@mthreads.com>
+R0CKSTAR <yeahdongcn@gmail.com>
 RJ Adriaansen <adriaansen@eshcc.eur.nl>
 Radoslav Gerganov <rgerganov@gmail.com>
 Radosław Gryta <radek.gryta@gmail.com>
@@ -419,11 +527,13 @@ Raj Hammeer Singh Hada <hammeerraj@gmail.com>
 Ralph Soika <ralph.soika@imixs.com>
 Rand Xie <randxiexyy29@gmail.com>
 Randall Fitzgerald <randall@dasaku.net>
+Random Fly <renfei8@live.cn>
 Reinforce-II <fate@eastal.com>
 Ren Xuancheng <jklj077@users.noreply.github.com>
 Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com>
 RhinoDevel <RhinoDevel@users.noreply.github.com>
 Riceball LEE <snowyu.lee@gmail.com>
+Rich Dougherty <rich@rd.nz>
 Richard Kiss <him@richardkiss.com>
 Richard Roberson <richardr1126@gmail.com>
 Rick G <26732651+TheFlipbook@users.noreply.github.com>
@@ -439,21 +549,30 @@ Robey Holderith <robey@flaminglunchbox.net>
 Robyn <robyngraf@users.noreply.github.com>
 Roger Meier <r.meier@siemens.com>
 Roland <14355895+rbur0425@users.noreply.github.com>
+Romain Biessy <romain.biessy@codeplay.com>
 Romain D <90720+Artefact2@users.noreply.github.com>
 Romain Neutron <romain@neutron.io>
 Roman Parykin <donderom@gmail.com>
 Ron Evans <ron@hybridgroup.com>
 Ron Jailall <rojailal@gmail.com>
+Roni <sulpher@gmx.net>
 Ronny Brendel <ronnybrendel@gmail.com>
 Ronsor <ronsor@ronsor.pw>
 Rowan Hart <rowanbhart@gmail.com>
+Ruchira Hasaranga <ruchira66@gmail.com>
+Ruixin Huang <18860020911@163.com>
 Rune <43761327+Rune-AI@users.noreply.github.com>
+RunningLeon <maningsheng@sensetime.com>
+RunningLeon <mnsheng@yeah.net>
 Ryan Landay <rlanday@gmail.com>
 Ryder Wishart <ryderwishart@gmail.com>
 Ryuei <louixs@users.noreply.github.com>
 Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
+SRHMorris <69468379+SRHMorris@users.noreply.github.com>
+SXX <sxx1136965276@gmail.com>
 SakuraUmi <yukinon244@gmail.com>
 Salvador E. Tropea <stropea@inti.gob.ar>
+Salvatore Mesoraca <s.mesoraca16@gmail.com>
 Sam Spilsbury <smspillaz@gmail.com>
 Sami Farin <3876865+Safari77@users.noreply.github.com>
 Samuel Maynard <samwmaynard@gmail.com>
@@ -463,23 +582,29 @@ Sebastián A <sebastian.aedo29@gmail.com>
 SebastianApel <13675545+SebastianApel@users.noreply.github.com>
 Senemu <10880819+Senemu@users.noreply.github.com>
 Sergey Alirzaev <zl29ah@gmail.com>
+Sergio López <slp@redhat.com>
 Sergio López <slp@sinrega.org>
 Sertaç Özercan <852750+sozercan@users.noreply.github.com>
 SeungWon Jeong <65549245+redlion0929@users.noreply.github.com>
 ShadovvBeast <ShadovvBeast@gmail.com>
 Shakhar Dasgupta <shakhardasgupta@gmail.com>
+Shane A <shanea@allenai.org>
 Shangning Xu <32517059+xushangning@users.noreply.github.com>
+Shankar <gshankar.87@gmail.com>
+Shanshan Shen <467638484@qq.com>
 Shijie <821898965@qq.com>
 Shintarou Okada <kokuzen@gmail.com>
 Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
 Shouzheng Liu <lshzh.hi@gmail.com>
 Shuichi Tsutsumi <shuichi0526@gmail.com>
+Shupei Fan <dymarkfan@outlook.com>
 Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
 Simon Willison <swillison@gmail.com>
 Siwen Yu <yusiwen@gmail.com>
 Sky Yan <skyan83@gmail.com>
 Slaren <2141330+slaren@users.noreply.github.com>
 Slava Primenko <primenko.s@gmail.com>
+Small Grass Forest <zixuanxcl@gmail.com>
 SoftwareRenderer <138734813+SoftwareRenderer@users.noreply.github.com>
 Someone <sergei.kozlukov@aalto.fi>
 Someone Serge <sergei.kozlukov@aalto.fi>
@@ -491,12 +616,15 @@ Stefan Sydow <stefan@sydow.email>
 Steffen Röcker <sroecker@gmail.com>
 Stephan Walter <stephan@walter.name>
 Stephen Nichols <snichols@users.noreply.github.com>
+Steve Bonds <sbonds@gmail.com>
 Steve Grubb <ausearch.1@gmail.com>
 Steven Prichard <spprichard20@gmail.com>
 Steven Roussey <sroussey@gmail.com>
 Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
+StrangeBytesDev <141275258+StrangeBytesDev@users.noreply.github.com>
 Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
 SuperUserNameMan <yoann@terminajones.com>
+Sutou Kouhei <kou@cozmixng.org>
 Tai Duc Nguyen <taiducnguyen.drexel@gmail.com>
 Taikono-Himazin <kazu@po.harenet.ne.jp>
 Tameem <113388789+AhmadTameem@users.noreply.github.com>
@@ -507,7 +635,9 @@ Theia Vogel <theia@vgel.me>
 Thérence <13496987+Royalphax@users.noreply.github.com>
 Thibault Terrasson <thibault.terrasson@gmail.com>
 Thomas Klausner <wiz@gatalith.at>
+Thorsten Sommer <SommerEngineering@users.noreply.github.com>
 Tim Miller <drasticactions@users.noreply.github.com>
+Tim Wang <overocean@gmail.com>
 Timmy Knight <r2d2fish@gmail.com>
 Timothy Cronin <40186632+4imothy@users.noreply.github.com>
 Ting Lou <ting.lou@gmail.com>
@@ -517,24 +647,31 @@ Tom C <tom.corelis@gmail.com>
 Tom Jobbins <784313+TheBloke@users.noreply.github.com>
 Tomas <tom.tomas.36478119@gmail.com>
 Tomáš Pazdiora <tomas.pazdiora@gmail.com>
+Tony Wasserka <4840017+neobrain@users.noreply.github.com>
 Tristan Druyen <tristan@vault81.mozmail.com>
 Tristan Ross <rosscomputerguy@protonmail.com>
+Trivikram Kamat <16024985+trivikr@users.noreply.github.com>
 Tungsten842 <886724vf@anonaddy.me>
 Tungsten842 <quantmint@protonmail.com>
 Tushar <ditsuke@protonmail.com>
 UEXTM.com <84163508+uextm@users.noreply.github.com>
+Ujjawal Panchal <31011628+Ujjawal-K-Panchal@users.noreply.github.com>
 Ulrich Drepper <drepper@gmail.com>
 Uzo Nweke <uzoechi@gmail.com>
 Vaibhav Srivastav <vaibhavs10@gmail.com>
 Val Kharitonov <mail@kharvd.com>
 Valentin Konovalov <valle.ketsujin@gmail.com>
 Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com>
+Vali Malinoiu <0x4139@gmail.com>
 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>
 Vlad <spitfireage@gmail.com>
 Vladimir <bogdad@gmail.com>
 Vladimir Malyutin <first-leon@yandex.ru>
 Vladimir Zorin <vladimir@deviant.guru>
+VoidIsVoid <343750470@qq.com>
 Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com>
 WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com>
 Weird Constructor <weirdconstructor@gmail.com>
@@ -551,15 +688,22 @@ Xiang (Kevin) Li <kevinli020508@gmail.com>
 Xiao-Yong Jin <jinxiaoyong@gmail.com>
 XiaotaoChen <chenxiaotao1234@gmail.com>
 Xiaoyi Chen <cxychina@gmail.com>
+Xie Yanbo <xieyanbo@gmail.com>
 Xingchen Song(宋星辰) <xingchensong1996@163.com>
+Xinpeng Dou <81913537+Dou-Git@users.noreply.github.com>
 Xuan Son Nguyen <thichthat@gmail.com>
+Yaiko <elyaiko@hotmail.com>
 Yann Follet <131855179+YannFollet@users.noreply.github.com>
 Yaroslav <yaroslav.yashin@me.com>
 Yazan Agha-Schrader <mountaiin@icloud.com>
 Yiming Cui <conandiy@vip.qq.com>
 Yishuo Wang <MeouSker77@outlook.com>
+Yoshi Suhara <y.suhara@gmail.com>
+Yoshi Suhara <ysuhara@nvidia.com>
+Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
 Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com>
 Yui <dev@sleepyyui.com>
+Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
 Yusuf Kağan Hanoğlu <hanoglu@yahoo.com>
 Yuval Peled <31162840+Yuval-Peled@users.noreply.github.com>
 ZHAOKAI WANG <sanxianwei@163.com>
@@ -568,6 +712,8 @@ Zay <95888118+isaiahbjork@users.noreply.github.com>
 Zenix <zenixls2@gmail.com>
 Zhang Peiyuan <a1286225768@gmail.com>
 Zheng.Deng <32841220+dengzheng-cloud@users.noreply.github.com>
+Zhenwei Jin <109658203+kylo5aby@users.noreply.github.com>
+Zhiyuan Li <lizhiyuan@uniartisan.com>
 ZhouYuChen <zhouyuchen@naver.com>
 Ziad Ben Hadj-Alouane <zied.benhadjalouane@gmail.com>
 Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com>
@@ -581,6 +727,7 @@ alexpinel <93524949+alexpinel@users.noreply.github.com>
 alonfaraj <alonfaraj@gmail.com>
 alwqx <kenan3015@gmail.com>
 amd-lalithnc <lalithnc@amd.com>
+amritahs-ibm <amritahs@linux.vnet.ibm.com>
 andrijdavid <david@geek.mg>
 anon998 <131767832+anon998@users.noreply.github.com>
 anzz1 <anzz1@live.com>
@@ -588,14 +735,18 @@ apaz <aarpazdera@gmail.com>
 apcameron <37645737+apcameron@users.noreply.github.com>
 arch-btw <57669023+arch-btw@users.noreply.github.com>
 arcrank <arcrank@gmail.com>
+ardfork <134447697+ardfork@users.noreply.github.com>
 arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
 at8u <129688334+at8u@users.noreply.github.com>
 automaticcat <daogiatuank54@gmail.com>
+awatuna <23447591+awatuna@users.noreply.github.com>
+b4b4o <zwbao@foxmail.com>
 bandoti <141645996+bandoti@users.noreply.github.com>
 beiller <beiller@gmail.com>
 bhubbb <79117352+bhubbb@users.noreply.github.com>
 bmwl <brian.marshall@tolko.com>
 bobqianic <129547291+bobqianic@users.noreply.github.com>
+brucepro <git@brucepro.net>
 bryanSwk <93190252+bryanSwk@users.noreply.github.com>
 bsilvereagle <bsilvereagle@users.noreply.github.com>
 bssrdf <merlintiger@hotmail.com>
@@ -614,10 +765,14 @@ cpumaxx <163466046+cpumaxx@users.noreply.github.com>
 crasm <crasm@git.vczf.net>
 crasm <crasm@git.vczf.us>
 daboe01 <daboe01@googlemail.com>
+daghanerdonmez <44506702+daghanerdonmez@users.noreply.github.com>
+daminho <37615795+daminho@users.noreply.github.com>
 david raistrick <keen99@users.noreply.github.com>
 ddh0 <dylanhalladay02@icloud.com>
 ddpasa <112642920+ddpasa@users.noreply.github.com>
 deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
+devojony <61173062+devojony@users.noreply.github.com>
+ditsuke <ditsuke@protonmail.com>
 divinity76 <divinity76@gmail.com>
 dm4 <sunrisedm4@gmail.com>
 dotpy314 <33351922+dotpy314@users.noreply.github.com>
@@ -629,14 +784,18 @@ ebraminio <ebraminio@gmail.com>
 eiery <19350831+eiery@users.noreply.github.com>
 eric8607242 <e0928021388@gmail.com>
 fairydreaming <166155368+fairydreaming@users.noreply.github.com>
+fengerhu1 <2748250768@qq.com>
 fraxy-v <65565042+fraxy-v@users.noreply.github.com>
 github-actions[bot] <github-actions[bot]@users.noreply.github.com>
 gliptic <gliptic@users.noreply.github.com>
 goerch <jhr.walter@t-online.de>
 grahameth <96447521+grahameth@users.noreply.github.com>
+gtygo <gtydoit@gmail.com>
 gwjr <502526+gwjr@users.noreply.github.com>
 h-h-h-h <13482553+h-h-h-h@users.noreply.github.com>
 hankcs <cnhankmc@gmail.com>
+haopeng <657407891@qq.com>
+hipudding <huafengchun@gmail.com>
 hoangmit <hoangmit@users.noreply.github.com>
 hongbo.mo <352280764@qq.com>
 hopkins385 <98618192+hopkins385@users.noreply.github.com>
@@ -649,12 +808,14 @@ hxer7963 <hxer7963@gmail.com>
 hydai <z54981220@gmail.com>
 iSma <ismail.senhaji@gmail.com>
 iacore <74560659+iacore@users.noreply.github.com>
+icppWorld <124377669+icppWorld@users.noreply.github.com>
 igarnier <igarnier@protonmail.com>
 intelmatt <61025942+intelmatt@users.noreply.github.com>
 iohub <rickyang.pro@gmail.com>
 jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
 jaime-m-p <167997752+jaime-m-p@users.noreply.github.com>
 jameswu2014 <545426914@qq.com>
+jdomke <28772296+jdomke@users.noreply.github.com>
 jiez <373447296@qq.com>
 jneem <joeneeman@gmail.com>
 joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
@@ -677,28 +838,35 @@ klosax <131523366+klosax@users.noreply.github.com>
 kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
 kunnis <kunnis@users.noreply.github.com>
 kuronekosaiko <EvanChanJ@163.com>
+kustaaya <58045274+kustaaya@users.noreply.github.com>
 kuvaus <22169537+kuvaus@users.noreply.github.com>
 kwin1412 <42286931+kwin1412@users.noreply.github.com>
 l3utterfly <gc.pthzfoldr@gmail.com>
+laik <laik.lj@me.com>
 ldwang <ftgreat@163.com>
 le.chang <cljs118@126.com>
 leejet <leejet714@gmail.com>
+leo-pony <nengjunma@outlook.com>
 limitedAtonement <limitedAtonement@users.noreply.github.com>
 liuwei-git <14815172+liuwei-git@users.noreply.github.com>
 lon <114724657+longregen@users.noreply.github.com>
 loonerin <132926317+loonerin@users.noreply.github.com>
+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>
 makomk <makosoft@googlemail.com>
 manikbhandari <mbbhandarimanik2@gmail.com>
 maor-ps <154728172+maor-ps@users.noreply.github.com>
+matiaslin <45382001+matiaslin@users.noreply.github.com>
+matteo <matteogeniaccio@yahoo.it>
 mdrokz <mohammadmunshi@gmail.com>
 mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
 minarchist <minarchist@users.noreply.github.com>
 mj-shifu <77107165+mj-shifu@users.noreply.github.com>
 mmyjona <jonathan.gonse@gmail.com>
 momonga <115213907+mmnga@users.noreply.github.com>
+momonga <146910567+mmngays@users.noreply.github.com>
 moritzbrantner <31051084+moritzbrantner@users.noreply.github.com>
 mzcu <milos.cubrilo@gmail.com>
 nanahi <130121847+na-na-hi@users.noreply.github.com>
@@ -716,8 +884,10 @@ 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>
+pculliton <phillipculliton@gmail.com>
 pengxin99 <pengxin.yuan@intel.com>
 perserk <perserk@gmail.com>
+piDack <104877312+piDack@users.noreply.github.com>
 pmysl <piotr.myslinski@outlook.com>
 postmasters <namnguyen@google.com>
 pudepiedj <pudepiedj@gmail.com>
@@ -733,6 +903,7 @@ runfuture <runfuture@users.noreply.github.com>
 sandyiscool <sandyiscool@gmail.com>
 sasha0552 <admin@sasha0552.org>
 semidark <me@semidark.net>
+serhii-nakon <57632032+serhii-nakon@users.noreply.github.com>
 sharpHL <132747147+sharpHL@users.noreply.github.com>
 shibe2 <shibe@tuta.io>
 singularity <12184989+singularity-s0@users.noreply.github.com>
@@ -741,42 +912,55 @@ sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com>
 slaren <2141330+slaren@users.noreply.github.com>
 slaren <slarengh@gmail.com>
 snadampal <87143774+snadampal@users.noreply.github.com>
+standby24x7 <standby24x7@gmail.com>
 staviq <staviq@gmail.com>
 stduhpf <stephduh@live.fr>
 strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
 swittk <switt1995@gmail.com>
 takov751 <40316768+takov751@users.noreply.github.com>
 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>
+thewh1teagle <61390950+thewh1teagle@users.noreply.github.com>
 tjohnman <tjohnman@users.noreply.github.com>
+toyer <2042519524@qq.com>
 tslmy <tslmy@users.noreply.github.com>
 ubik2 <ubik2@users.noreply.github.com>
 uint256_t <konndennsa@gmail.com>
 uint256_t <maekawatoshiki1017@gmail.com>
 unbounded <haakon@likedan.net>
+uvos <devnull@uvos.xyz>
 valiray <133289098+valiray@users.noreply.github.com>
+vb <vaibhavs10@gmail.com>
 vik <vikhyatk@gmail.com>
 viric <viric@viric.name>
 vodkaslime <646329483@qq.com>
 vvhg1 <94630311+vvhg1@users.noreply.github.com>
 vxiiduu <73044267+vxiiduu@users.noreply.github.com>
+wangshuai09 <391746016@qq.com>
 wbpxre150 <100937007+wbpxre150@users.noreply.github.com>
 whoreson <139810751+whoreson@users.noreply.github.com>
 woachk <24752637+woachk@users.noreply.github.com>
 wonjun Jang <strutive07@gmail.com>
 woodx <124784234+woodx9@users.noreply.github.com>
+wwoodsTM <104587230+wwoodsTM@users.noreply.github.com>
 wzy <32936898+Freed-Wu@users.noreply.github.com>
 xaedes <xaedes@gmail.com>
 xaedes <xaedes@googlemail.com>
+xctan <axunlei@gmail.com>
 xloem <0xloem@gmail.com>
 yangli2 <yangli2@gmail.com>
 yuiseki <yuiseki@gmail.com>
+yuri@FreeBSD <yurivict@users.noreply.github.com>
 zakkor <edward.partenie@gmail.com>
 zhangkaihuo <zhangkaihuo@gmail.com>
+zhentaoyu <zhentao.yu@intel.com>
 zhouwg <6889919+zhouwg@users.noreply.github.com>
 zhouwg <zhouwg2000@gmail.com>
 zrm <trustiosity.zrm@gmail.com>
 Ștefan-Gabriel Muscalu <legraphista@users.noreply.github.com>
+杨朱 · Kiki <baofa.fan@daocloud.io>
 源文雨 <41315874+fumiama@users.noreply.github.com>
+蕭澧邦 <45505768+shou692199@users.noreply.github.com>
 Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>

CMakeLists.txt

@@ -96,10 +96,6 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_AMX)
-    set(GGML_AMX ON)
-endif()
-
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()

CODEOWNERS

@@ -0,0 +1,3 @@
+# collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
+
+ci/ @ggerganov

CONTRIBUTING.md

@@ -1,9 +1,10 @@
 # Pull requests (for contributors)
 
 - Test your changes:
-  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the `ggml` library
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
-- Optionally rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs
+  - Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
+  - If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
+  - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
@@ -12,6 +13,7 @@
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
 
 # Coding guidelines
 

Makefile

@@ -1,3 +1,7 @@
+ifndef LLAMA_MAKEFILE
+$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+endif
+
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	libllava.a \
@@ -251,11 +255,11 @@ endif
 # Compile flags
 #
 
-# keep standard at C11 and C++11
+# keep standard at C11 and C++17
 MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -DGGML_USE_CPU
 MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CXXFLAGS  = -std=c++17 -fPIC
+MK_NVCCFLAGS = -std=c++17
 
 ifdef LLAMA_NO_CCACHE
 GGML_NO_CCACHE := 1
@@ -575,9 +579,12 @@ endif
 
 ifndef GGML_NO_AMX
 	MK_CPPFLAGS += -DGGML_USE_AMX
-	OBJ_GGML_EXT += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
+	OBJ_GGML_EXT += ggml/src/ggml-cpu/amx/amx.o ggml/src/ggml-cpu/amx/mmq.o
 endif
 
+# only necessary for the CPU backend files
+MK_CPPFLAGS += -Iggml/src/ggml-cpu
+
 ifdef GGML_RPC
 	MK_CPPFLAGS  += -DGGML_USE_RPC
 	OBJ_GGML_EXT += ggml/src/ggml-rpc.o

Package.swift

@@ -28,13 +28,16 @@ var cSettings: [CSetting] =  [
     .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
     .unsafeFlags(["-fno-objc-arc"]),
     .headerSearchPath("ggml/src"),
+    .headerSearchPath("ggml/src/ggml-cpu"),
     // NOTE: NEW_LAPACK will required iOS version 16.4+
     // We should consider add this in the future when we drop support for iOS 14
     // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
     // .define("ACCELERATE_NEW_LAPACK"),
     // .define("ACCELERATE_LAPACK_ILP64")
+    .define("GGML_USE_CPU"),
 ]
 
+
 #if canImport(Darwin)
 sources.append("ggml/src/ggml-common.h")
 sources.append("ggml/src/ggml-metal/ggml-metal.m")
@@ -44,7 +47,6 @@ cSettings.append(
     contentsOf: [
         .define("GGML_USE_ACCELERATE"),
         .define("GGML_USE_METAL"),
-        .define("GGML_USE_CPU")
     ]
 )
 #endif
@@ -86,5 +88,5 @@ let package = Package(
             linkerSettings: linkerSettings
         )
     ],
-    cxxLanguageStandard: .cxx11
+    cxxLanguageStandard: .cxx17
 )

README.md

@@ -4,7 +4,6 @@
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
-[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
@@ -26,7 +25,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
-variety of hardware - locally and in the cloud.
+range of hardware - locally and in the cloud.
 
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
@@ -36,14 +35,17 @@ variety of hardware - locally and in the cloud.
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
-Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
-improved significantly thanks to many contributions. It is the main playground for developing new features for the
-[ggml](https://github.com/ggerganov/ggml) library.
+The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
 
-**Supported models:**
+<details>
+<summary>Models</summary>
 
 Typically finetunes of the base models below are supported as well.
 
+Instructions for adding support for new models: [HOWTO-add-model.md](docs/development/HOWTO-add-model.md)
+
+#### Text-only
+
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
 - [x] LLaMA 3 🦙🦙🦙
@@ -97,9 +99,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 
-(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
-
-**Multimodal models:**
+#### Multimodal
 
 - [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
 - [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
@@ -111,7 +111,10 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
-**Bindings:**
+</details>
+
+<details>
+<summary>Bindings</summary>
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
@@ -138,316 +141,314 @@ Typically finetunes of the base models below are supported as well.
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
 - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 
-**UI:**
+</details>
 
-Unless otherwise noted these projects are open-source with permissive licensing:
-
-- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
-- [iohub/collama](https://github.com/iohub/coLLaMA)
-- [janhq/jan](https://github.com/janhq/jan) (AGPL)
-- [nat/openplayground](https://github.com/nat/openplayground)
-- [Faraday](https://faraday.dev/) (proprietary)
-- [LMStudio](https://lmstudio.ai/) (proprietary)
-- [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
-- [ramalama](https://github.com/containers/ramalama) (MIT)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
-- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
-- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
-- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
-- [ollama/ollama](https://github.com/ollama/ollama)
-- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
-- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
-- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
-- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
-- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
-- [RAGNA Desktop](https://ragna.app/) (proprietary)
-- [RecurseChat](https://recurse.chat/) (proprietary)
-- [semperai/amica](https://github.com/semperai/amica)
-- [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Msty](https://msty.app) (proprietary)
-- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
-- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file)(Apachev2.0 or later)
-- [Dot](https://github.com/alexpinel/Dot) (GPL)
-- [MindMac](https://mindmac.app) (proprietary)
-- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
-- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
-- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
-- [AIKit](https://github.com/sozercan/aikit) (MIT)
-- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
-- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
-- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
-- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
+<details>
+<summary>UIs</summary>
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
-**Tools:**
+- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
+- [Dot](https://github.com/alexpinel/Dot) (GPL)
+- [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)
+- [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)
+- [LARS](https://github.com/abgulati/LARS) (AGPL)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
+- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
+- [LMStudio](https://lmstudio.ai/) (proprietary)
+- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
+- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
+- [MindMac](https://mindmac.app) (proprietary)
+- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
+- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
+- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile) (Apache-2.0)
+- [nat/openplayground](https://github.com/nat/openplayground) (MIT)
+- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all) (MIT)
+- [ollama/ollama](https://github.com/ollama/ollama) (MIT)
+- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
+- [PocketPal AI](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
+- [psugihara/FreeChat](https://github.com/psugihara/FreeChat) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) (MIT)
+- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
+- [semperai/amica](https://github.com/semperai/amica) (MIT)
+- [withcatai/catai](https://github.com/withcatai/catai) (MIT)
+
+</details>
+
+<details>
+<summary>Tools</summary>
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
 - [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
-- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
+- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
 
-**Infrastructure:**
+</details>
+
+<details>
+<summary>Infrastructure</summary>
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 
-**Games:**
+</details>
+
+<details>
+<summary>Games</summary>
+
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
 
-## Demo
-
-<details>
-<summary>Typical run using LLaMA v2 13B on M2 Ultra</summary>
-
-```
-$ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
-I llama.cpp build info:
-I UNAME_S:  Darwin
-I UNAME_P:  arm
-I UNAME_M:  arm64
-I CFLAGS:   -I.            -O3 -std=c11   -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -pthread -DGGML_USE_K_QUANTS -DGGML_USE_ACCELERATE
-I CXXFLAGS: -I. -I./common -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -DGGML_USE_K_QUANTS
-I LDFLAGS:   -framework Accelerate
-I CC:       Apple clang version 14.0.3 (clang-1403.0.22.14.1)
-I CXX:      Apple clang version 14.0.3 (clang-1403.0.22.14.1)
-
-make: Nothing to be done for `default'.
-main: build = 1041 (cf658ad)
-main: seed  = 1692823051
-llama_model_loader: loaded meta data with 16 key-value pairs and 363 tensors from models/llama-13b-v2/ggml-model-q4_0.gguf (version GGUF V1 (latest))
-llama_model_loader: - type  f32:   81 tensors
-llama_model_loader: - type q4_0:  281 tensors
-llama_model_loader: - type q6_K:    1 tensors
-llm_load_print_meta: format         = GGUF V1 (latest)
-llm_load_print_meta: arch           = llama
-llm_load_print_meta: vocab type     = SPM
-llm_load_print_meta: n_vocab        = 32000
-llm_load_print_meta: n_merges       = 0
-llm_load_print_meta: n_ctx_train    = 4096
-llm_load_print_meta: n_ctx          = 512
-llm_load_print_meta: n_embd         = 5120
-llm_load_print_meta: n_head         = 40
-llm_load_print_meta: n_head_kv      = 40
-llm_load_print_meta: n_layer        = 40
-llm_load_print_meta: n_rot          = 128
-llm_load_print_meta: n_gqa          = 1
-llm_load_print_meta: f_norm_eps     = 1.0e-05
-llm_load_print_meta: f_norm_rms_eps = 1.0e-05
-llm_load_print_meta: n_ff           = 13824
-llm_load_print_meta: freq_base      = 10000.0
-llm_load_print_meta: freq_scale     = 1
-llm_load_print_meta: model type     = 13B
-llm_load_print_meta: model ftype    = mostly Q4_0
-llm_load_print_meta: model size     = 13.02 B
-llm_load_print_meta: general.name   = LLaMA v2
-llm_load_print_meta: BOS token = 1 '<s>'
-llm_load_print_meta: EOS token = 2 '</s>'
-llm_load_print_meta: UNK token = 0 '<unk>'
-llm_load_print_meta: LF token  = 13 '<0x0A>'
-llm_load_tensors: ggml ctx size =    0.11 MB
-llm_load_tensors: mem required  = 7024.01 MB (+  400.00 MB per state)
-...................................................................................................
-llama_new_context_with_model: kv self size  =  400.00 MB
-llama_new_context_with_model: compute buffer total size =   75.41 MB
-
-system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
-sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
-generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0
-
-
- Building a website can be done in 10 simple steps:
-Step 1: Find the right website platform.
-Step 2: Choose your domain name and hosting plan.
-Step 3: Design your website layout.
-Step 4: Write your website content and add images.
-Step 5: Install security features to protect your site from hackers or spammers
-Step 6: Test your website on multiple browsers, mobile devices, operating systems etc…
-Step 7: Test it again with people who are not related to you personally – friends or family members will work just fine!
-Step 8: Start marketing and promoting the website via social media channels or paid ads
-Step 9: Analyze how many visitors have come to your site so far, what type of people visit more often than others (e.g., men vs women) etc…
-Step 10: Continue to improve upon all aspects mentioned above by following trends in web design and staying up-to-date on new technologies that can enhance user experience even further!
-How does a Website Work?
-A website works by having pages, which are made of HTML code. This code tells your computer how to display the content on each page you visit – whether it’s an image or text file (like PDFs). In order for someone else’s browser not only be able but also want those same results when accessing any given URL; some additional steps need taken by way of programming scripts that will add functionality such as making links clickable!
-The most common type is called static HTML pages because they remain unchanged over time unless modified manually (either through editing files directly or using an interface such as WordPress). They are usually served up via HTTP protocols – this means anyone can access them without having any special privileges like being part of a group who is allowed into restricted areas online; however, there may still exist some limitations depending upon where one lives geographically speaking.
-How to
-llama_print_timings:        load time =   576.45 ms
-llama_print_timings:      sample time =   283.10 ms /   400 runs   (    0.71 ms per token,  1412.91 tokens per second)
-llama_print_timings: prompt eval time =   599.83 ms /    19 tokens (   31.57 ms per token,    31.68 tokens per second)
-llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms per token,    16.28 tokens per second)
-llama_print_timings:       total time = 25431.49 ms
-```
-
 </details>
 
-<details>
-<summary>Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook</summary>
-
-And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
-
-https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4
-
-</details>
-
-## Usage
-
-Here are the end-to-end binary build and model conversion steps for most supported models.
-
-### Basic usage
-
-Firstly, you need to get the binary. There are different methods that you can follow:
-- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md)
-- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md)
-- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md)
-- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases)
-
-You can run a basic completion using this command:
-
-```bash
-llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128
-
-# Output:
-# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
-```
-
-See [this page](./examples/main/README.md) for a full list of parameters.
-
-### Conversation mode
-
-If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
-
-```bash
-llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
-
-# Output:
-# > hi, who are you?
-# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
-#
-# > what is 1+1?
-# Easy peasy! The answer to 1+1 is... 2!
-```
-
-By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-
-```bash
-./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
-```
-
-You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
-
-```bash
-./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
-```
-
-### Web server
-
-[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
-
-Example usage:
-
-```bash
-./llama-server -m your_model.gguf --port 8080
-
-# Basic web UI can be accessed via browser: http://localhost:8080
-# Chat completion endpoint: http://localhost:8080/v1/chat/completions
-```
-
-### Interactive mode
-
-> [!NOTE]
-> If you prefer basic usage, please consider using conversation mode instead of interactive mode
-
-In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
-
-Here is an example of a few-shot interaction, invoked with the command
-
-```bash
-# default arguments using a 7B model
-./examples/chat.sh
-
-# advanced chat with a 13B model
-./examples/chat-13B.sh
-
-# custom arguments using a 13B model
-./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
-```
-
-Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `llama-cli` example program.
-
-![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
-
-### Persistent Interaction
-
-The prompt, user inputs, and model generations can be saved and resumed across calls to `./llama-cli` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
-
-```bash
-# Start a new chat
-PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
-
-# Resume that chat
-PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
-
-# Start a different chat with the same prompt/model
-PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/another ./examples/chat-persistent.sh
-
-# Different prompt cache for different prompt/model
-PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
-    CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
-```
-
-### Constrained output with grammars
-
-`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
-
-```bash
-./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
-```
-
-The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
-
-For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
-
-## Build
-
-Please refer to [Build llama.cpp locally](./docs/build.md)
-
 ## Supported backends
 
 | Backend | Target devices |
 | --- | --- |
-| [Metal](./docs/build.md#metal-build) | Apple Silicon |
-| [BLAS](./docs/build.md#blas-build) | All |
-| [BLIS](./docs/backend/BLIS.md) | All |
-| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](./docs/build.md#musa) | Moore Threads MTT GPU |
-| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
-| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
-| [Vulkan](./docs/build.md#vulkan) | GPU |
-| [CANN](./docs/build.md#cann) | Ascend NPU |
+| [Metal](docs/build.md#metal-build) | Apple Silicon |
+| [BLAS](docs/build.md#blas-build) | All |
+| [BLIS](docs/backend/BLIS.md) | All |
+| [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
+| [MUSA](docs/build.md#musa) | Moore Threads MTT GPU |
+| [CUDA](docs/build.md#cuda) | Nvidia GPU |
+| [hipBLAS](docs/build.md#hipblas) | AMD GPU |
+| [Vulkan](docs/build.md#vulkan) | GPU |
+| [CANN](docs/build.md#cann) | Ascend NPU |
 
-## Tools
+## Building the project
 
-### Prepare and Quantize
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
 
-> [!NOTE]
-> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
+- Clone this repository and build locally, see [how to build](docs/build.md)
+- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
+- Use a Docker image, see [documentation for Docker](docs/docker.md)
+- Download pre-built binaries from [releases](https://github.com/ggerganov/llama.cpp/releases)
 
-To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
+## Obtaining and quantizing models
 
-Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
-It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
+The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
 
-To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
+- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
+- [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-### Perplexity (measuring model quality)
+After downloading a model, use the CLI tools to run it locally - see below.
 
-You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
-For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
+`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
+
+The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
+
+- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
+- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggerganov/llama.cpp/discussions/10123)
+- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggerganov/llama.cpp/discussions/9268)
+- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggerganov/llama.cpp/discussions/9669)
+
+To learn more about model quantization, [read this documentation](examples/quantize/README.md)
+
+## [`llama-cli`](examples/main)
+
+#### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
+
+- <details open>
+    <summary>Run simple text completion</summary>
+
+    ```bash
+    llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128
+
+    # I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
+    ```
+
+    </details>
+
+- <details>
+    <summary>Run in conversation mode</summary>
+
+    ```bash
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv
+
+    # > hi, who are you?
+    # Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
+    #
+    # > what is 1+1?
+    # Easy peasy! The answer to 1+1 is... 2!
+    ```
+
+    </details>
+
+- <details>
+    <summary>Run with custom chat template</summary>
+
+    ```bash
+    # use the "chatml" template
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
+
+    # use a custom template
+    llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
+    ```
+
+    [Supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+
+    </details>
+
+- <details>
+    <summary>Constrain the output with a custom grammar</summary>
+
+    ```bash
+    llama-cli -m model.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+
+    # {"appointmentTime": "8pm", "appointmentDetails": "schedule a a call"}
+    ```
+
+    The [grammars/](grammars/) folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](grammars/README.md).
+
+    For authoring more complex JSON grammars, check out https://grammar.intrinsiclabs.ai/
+
+    </details>
+
+
+## [`llama-server`](examples/server)
+
+#### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.
+
+- <details open>
+    <summary>Start a local HTTP server with default configuration on port 8080</summary>
+
+    ```bash
+    llama-server -m model.gguf --port 8080
+
+    # Basic web UI can be accessed via browser: http://localhost:8080
+    # Chat completion endpoint: http://localhost:8080/v1/chat/completions
+    ```
+
+    </details>
+
+- <details>
+    <summary>Support multiple-users and parallel decoding</summary>
+
+    ```bash
+    # up to 4 concurrent requests, each with 4096 max context
+    llama-server -m model.gguf -c 16384 -np 4
+    ```
+
+    </details>
+
+- <details>
+    <summary>Enable speculative decoding</summary>
+
+    ```bash
+    # the draft.gguf model should be a small variant of the target model.gguf
+    llama-server -m model.gguf -md draft.gguf
+    ```
+
+    </details>
+
+- <details>
+    <summary>Serve an embedding model</summary>
+
+    ```bash
+    # use the /embedding endpoint
+    llama-server -m model.gguf --embedding --pooling cls -ub 8192
+    ```
+
+    </details>
+
+- <details>
+    <summary>Serve a reranking model</summary>
+
+    ```bash
+    # use the /reranking endpoint
+    llama-server -m model.gguf --reranking
+    ```
+
+    </details>
+
+- <details>
+    <summary>Constrain all outputs with a grammar</summary>
+
+    ```bash
+    # custom grammar
+    llama-server -m model.gguf --grammar-file grammar.gbnf
+
+    # JSON
+    llama-server -m model.gguf --grammar-file grammars/json.gbnf
+    ```
+
+    </details>
+
+
+## [`llama-perplexity`](examples/perplexity)
+
+#### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.
+
+- <details open>
+    <summary>Measure the perplexity over a text file</summary>
+
+    ```bash
+    llama-perplexity -m model.gguf -f file.txt
+
+    # [1]15.2701,[2]5.4007,[3]5.3073,[4]6.2965,[5]5.8940,[6]5.6096,[7]5.7942,[8]4.9297, ...
+    # Final estimate: PPL = 5.4007 +/- 0.67339
+    ```
+
+    </details>
+
+- <details>
+    <summary>Measure KL divergence</summary>
+
+    ```bash
+    # TODO
+    ```
+
+    </details>
+
+[^1]: [examples/perplexity/README.md](examples/perplexity/README.md)
+[^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
+
+## [`llama-bench`](example/bench)
+
+#### Benchmark the performance of the inference for various parameters.
+
+- <details open>
+    <summary>Run default benchmark</summary>
+
+    ```bash
+    llama-bench -m model.gguf
+
+    # Output:
+    # | model               |       size |     params | backend    | threads |          test |                  t/s |
+    # | ------------------- | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
+    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         pp512 |      5765.41 ± 20.55 |
+    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         tg128 |        197.71 ± 0.81 |
+    #
+    # build: 3e0ba0e60 (4229)
+    ```
+
+    </details>
+
+
+## [`llama-simple`](examples/simple)
+
+#### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
+
+- <details>
+    <summary>Basic text completion</summary>
+
+    ```bash
+    llama-simple -m model.gguf
+
+    # Hello my name is Kaitlyn and I am a 16 year old girl. I am a junior in high school and I am currently taking a class called "The Art of
+    ```
+
+    </details>
 
-To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md)
 
 ## Contributing
 
@@ -462,20 +463,19 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 
 ## Other documentation
 
-- [main (cli)](./examples/main/README.md)
-- [server](./examples/server/README.md)
-- [jeopardy](./examples/jeopardy/README.md)
-- [GBNF grammars](./grammars/README.md)
+- [main (cli)](examples/main/README.md)
+- [server](examples/server/README.md)
+- [GBNF grammars](grammars/README.md)
 
-**Development documentation**
+#### Development documentation
 
-- [How to build](./docs/build.md)
-- [Running on Docker](./docs/docker.md)
-- [Build on Android](./docs/android.md)
-- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
+- [How to build](docs/build.md)
+- [Running on Docker](docs/docker.md)
+- [Build on Android](docs/android.md)
+- [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
 
-**Seminal papers and background on the models**
+#### Seminal papers and background on the models
 
 If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
 - LLaMA:
@@ -486,3 +486,6 @@ If your issue is with model generation quality, then please at least scan the fo
 - GPT-3.5 / InstructGPT / ChatGPT:
     - [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)
+
+#### References
+

ci/run.sh

@@ -815,7 +815,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
     # Create a fresh python3 venv and enter it
-    python3 -m venv "$MNT/venv"
+    if ! python3 -m venv "$MNT/venv"; then
+        echo "Error: Failed to create Python virtual environment at $MNT/venv."
+        exit 1
+    fi
     source "$MNT/venv/bin/activate"
 
     pip install -r ${SRC}/requirements.txt --disable-pip-version-check

common/CMakeLists.txt

@@ -88,5 +88,5 @@ if (LLAMA_CURL)
 endif ()
 
 target_include_directories(${TARGET} PUBLIC .)
-target_compile_features   (${TARGET} PUBLIC cxx_std_11)
+target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)

common/arg.cpp

@@ -348,6 +348,18 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
     return true;
 }
 
+static std::string list_builtin_chat_templates() {
+    std::vector<const char *> supported_tmpl;
+    int32_t res = llama_chat_builtin_templates(nullptr, 0);
+    supported_tmpl.resize(res);
+    res = llama_chat_builtin_templates(supported_tmpl.data(), supported_tmpl.size());
+    std::ostringstream msg;
+    for (auto & tmpl : supported_tmpl) {
+        msg << tmpl << (&tmpl == &supported_tmpl.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     // load dynamic backends
     ggml_backend_load_all();
@@ -1370,8 +1382,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.n_gpu_layers = value;
             if (!llama_supports_gpu_offload()) {
-                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers option will be ignored\n");
-                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+                fprintf(stderr, "warning: no usable GPU found, --gpu-layers option will be ignored\n");
+                fprintf(stderr, "warning: one possible reason is that llama.cpp was compiled without GPU support\n");
+                fprintf(stderr, "warning: consult docs/build.md for compilation instructions\n");
             }
         }
     ).set_env("LLAMA_ARG_N_GPU_LAYERS"));
@@ -1813,9 +1826,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
-        "set custom jinja chat template (default: template taken from model's metadata)\n"
-        "if suffix/prefix are specified, template will be disabled\n"
-        "only commonly used templates are accepted:\nhttps://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template",
+        string_format(
+            "set custom jinja chat template (default: template taken from model's metadata)\n"
+            "if suffix/prefix are specified, template will be disabled\n"
+            "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
+        ),
         [](common_params & params, const std::string & value) {
             if (!common_chat_verify_template(value)) {
                 throw std::runtime_error(string_format(
@@ -2104,8 +2119,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.speculative.n_gpu_layers = value;
             if (!llama_supports_gpu_offload()) {
-                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
-                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+                fprintf(stderr, "warning: no usable GPU found, --gpu-layers-draft option will be ignored\n");
+                fprintf(stderr, "warning: one possible reason is that llama.cpp was compiled without GPU support\n");
+                fprintf(stderr, "warning: consult docs/build.md for compilation instructions\n");
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));

common/common.cpp

@@ -652,7 +652,17 @@ bool fs_validate_filename(const std::string & filename) {
 
     std::u32string filename_utf32;
     try {
+#if defined(__clang__)
+        // disable C++17 deprecation warning for std::codecvt_utf8
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
         filename_utf32 = converter.from_bytes(filename);
 
         // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,

common/common.h

@@ -133,6 +133,7 @@ struct common_params_sampling {
     bool    penalize_nl        = false; // consider newlines as a repeatable token
     bool    ignore_eos         = false;
     bool    no_perf            = false; // disable performance metrics
+    bool    timing_per_token   = false;
 
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 

docs/android.md

@@ -23,10 +23,10 @@ $ curl -L {model-url} -o ~/{model}.gguf
 Then, if you are not already in the repo directory, `cd` into `llama.cpp` and:
 
 ```
-$ ./build/bin/llama-simple -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
+$ ./build/bin/llama-cli -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
 ```
 
-Here, we show `llama-simple`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
+Here, we show `llama-cli`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
 
 To see what it might look like visually, here's an old demo of an interactive session running on a Pixel 5 phone:
 

docs/backend/BLIS.md

@@ -27,13 +27,6 @@ We recommend using openmp since it's easier to modify the cores being used.
 
 ### llama.cpp compilation
 
-Makefile:
-
-```bash
-make GGML_BLIS=1 -j
-# make GGML_BLIS=1 llama-benchmark-matmult
-```
-
 CMake:
 
 ```bash

docs/backend/CANN.md

@@ -23,6 +23,8 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 
 ## News
 
+- 2024.11
+  - Support F16 and F32 data type model for Ascend 310P NPU.
 - 2024.8
   - Support `Q4_0` and `Q8_0` data type for Ascend NPU.
 - 2024.7
@@ -40,9 +42,11 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 ### Ascend NPU
 
 **Verified devices**
+
 | Ascend NPU                    | Status  |
 |:-----------------------------:|:-------:|
 | Atlas 300T A2                 | Support |
+| Atlas 300I Duo                | Support |
 
 *Notes:*
 

docs/build.md

@@ -7,124 +7,63 @@ git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
 ```
 
-In order to build llama.cpp you have four different options.
+The following sections describe how to build with different backends and options.
 
-- Using `make`:
-  - On Linux or MacOS:
+## CPU Build
 
-      ```bash
-      make
-      ```
+Build llama.cpp using `CMake`:
 
-  - On Windows (x86/x64 only, arm64 requires cmake):
+```bash
+cmake -B build
+cmake --build build --config Release
+```
 
-    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-    2. Extract `w64devkit` on your pc.
-    3. Run `w64devkit.exe`.
-    4. Use the `cd` command to reach the `llama.cpp` folder.
-    5. From here you can run:
-        ```bash
-        make
-        ```
+**Notes**:
 
-  - Notes:
-    - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, run `make LLAMA_DEBUG=1`
+- For faster compilation, add the `-j` argument to run multiple jobs in parallel, or use a generator that does this automatically such as Ninja. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
+- For faster repeated compilation, install [ccache](https://ccache.dev/)
+- For debug builds, there are two cases:
 
-- Using `CMake`:
+    1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
 
-  ```bash
-  cmake -B build
-  cmake --build build --config Release
-  ```
+    ```bash
+    cmake -B build -DCMAKE_BUILD_TYPE=Debug
+    cmake --build build
+    ```
 
-  **Notes**:
+    2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
 
-    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, there are two cases:
+    ```bash
+    cmake -B build -G "Xcode"
+    cmake --build build --config Debug
+    ```
 
-      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+    For more details and a list of supported generators, see the [CMake documentation](https://cmake.org/cmake/help/latest/manual/cmake-generators.7.html).
 
-      ```bash
-      cmake -B build -DCMAKE_BUILD_TYPE=Debug
-      cmake --build build
-      ```
-
-      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
-
-      ```bash
-      cmake -B build -G "Xcode"
-      cmake --build build --config Debug
-      ```
-    - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
-      - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
-        - Tab Workload: Desktop-development with C++
-        - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
-      - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
-      - For Windows on ARM (arm64, WoA) build with:
-        ```bash
-        cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
-        cmake --build build-arm64-windows-llvm-release
-        ```
-        Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
-
--   Using `gmake` (FreeBSD):
-
-    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
-    2. Add your user to **video** group
-    3. Install compilation dependencies.
-
-        ```bash
-        sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
-
-        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
-        ```
-
-## Metal Build
-
-On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
-To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
-
-When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
-argument.
+- Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
+    - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
+    - Tab Workload: Desktop-development with C++
+    - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
+    - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
+    - For Windows on ARM (arm64, WoA) build with:
+    ```bash
+    cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
+    cmake --build build-arm64-windows-llvm-release
+    ```
+    Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
 
 ## BLAS Build
 
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Using BLAS doesn't affect the generation performance. There are currently several different BLAS implementations available for build and use:
 
-### Accelerate Framework:
+### Accelerate Framework
 
 This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
 
-### OpenBLAS:
+### OpenBLAS
 
 This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
 
-- Using `make`:
-  - On Linux:
-    ```bash
-    make GGML_OPENBLAS=1
-    ```
-
-  - On Windows:
-
-    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-    2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
-    3. Extract `w64devkit` on your pc.
-    4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
-    5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
-    6. Run `w64devkit.exe`.
-    7. Use the `cd` command to reach the `llama.cpp` folder.
-    8. From here you can run:
-
-        ```bash
-        make GGML_OPENBLAS=1
-        ```
-
 - Using `CMake` on Linux:
 
     ```bash
@@ -136,14 +75,6 @@ This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS i
 
 Check [BLIS.md](./backend/BLIS.md) for more information.
 
-### SYCL
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
-
-llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
-
-For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
-
 ### Intel oneMKL
 
 Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
@@ -161,16 +92,29 @@ Building through oneAPI compilers will make avx_vnni instruction set available f
 
 Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
 
-### CUDA
+### Other BLAS libraries
 
-This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+Any other BLAS library can be used by setting the `GGML_BLAS_VENDOR` option. See the [CMake documentation](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) for a list of supported vendors.
 
-For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
+## Metal Build
+
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `-DGGML_METAL=OFF` cmake option.
+
+When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers 0` command-line argument.
+
+## SYCL
+
+SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
+
+## CUDA
+
+This provides GPU acceleration using an NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from the [NVIDIA developer site](https://developer.nvidia.com/cuda-downloads).
 
-- Using `make`:
-  ```bash
-  make GGML_CUDA=1
-  ```
 - Using `CMake`:
 
   ```bash
@@ -192,14 +136,10 @@ The following compilation options are also available to tweak performance:
 | 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.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
-### MUSA
+## 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).
 
-- Using `make`:
-  ```bash
-  make GGML_MUSA=1
-  ```
 - Using `CMake`:
 
   ```bash
@@ -213,16 +153,12 @@ The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enab
 
 Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
 
-### hipBLAS
+## HIP
 
-This provides BLAS acceleration on HIP-supported AMD GPUs.
+This provides GPU acceleration on HIP-supported AMD GPUs.
 Make sure to have ROCm installed.
 You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
 
-- Using `make`:
-  ```bash
-  make GGML_HIP=1
-  ```
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
@@ -247,11 +183,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
       && cmake --build build -- -j 16
   ```
 
-- Using `make` (example for target gfx1030, build with 16 CPU threads):
-  ```bash
-  make -j16 GGML_HIP=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
-  ```
-
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
   ```bash
   set PATH=%HIP_PATH%\bin;%PATH%
@@ -265,11 +196,11 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
 
-### Vulkan
+## Vulkan
 
 **Windows**
 
-#### w64devkit
+### w64devkit
 
 Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
@@ -289,9 +220,14 @@ Libs: -lvulkan-1
 EOF
 
 ```
-Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
 
-#### Git Bash MINGW64
+Switch into the `llama.cpp` directory and build using CMake.
+```sh
+cmake -B build -DGGML_VULKAN=ON
+cmake --build build --config Release
+```
+
+### Git Bash MINGW64
 
 Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
 
@@ -310,20 +246,21 @@ cmake --build build --config Release
 
 Now you can load the model in conversation mode using `Vulkan`
 
-```
-build/bin/release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
+```sh
+build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
 ```
 
-#### MSYS2
+### MSYS2
 Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
-  ```sh
-  pacman -S git \
-      mingw-w64-ucrt-x86_64-gcc \
-      mingw-w64-ucrt-x86_64-cmake \
-      mingw-w64-ucrt-x86_64-vulkan-devel \
-      mingw-w64-ucrt-x86_64-shaderc
-  ```
-Switch into `llama.cpp` directory and build using CMake.
+```sh
+pacman -S git \
+    mingw-w64-ucrt-x86_64-gcc \
+    mingw-w64-ucrt-x86_64-cmake \
+    mingw-w64-ucrt-x86_64-vulkan-devel \
+    mingw-w64-ucrt-x86_64-shaderc
+```
+
+Switch into the `llama.cpp` directory and build using CMake.
 ```sh
 cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
@@ -372,7 +309,7 @@ cmake --build build --config Release
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
 ```
 
-### CANN
+## CANN
 This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
 
 For more information about Ascend NPU in [Ascend Community](https://www.hiascend.com/en/).
@@ -387,22 +324,26 @@ cmake --build build --config release
 
 You can test with:
 
-`./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
-
-If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
 ```bash
-llm_load_tensors:       CANN buffer size = 13313.00 MiB
+./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32
+```
+
+If the following info is output on screen, you are using `llama.cpp` with the CANN backend:
+```bash
+llm_load_tensors:       CANN model buffer size = 13313.00 MiB
 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).
 
-### Android
+## Android
 
 To read documentation for how to build on Android, [click here](./android.md)
 
-### Arm CPU optimized mulmat kernels
+## Notes about GPU-accelerated backends
 
-Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
+The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.
 
-To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).
+In most cases, it is possible to build and use multiple backends at the same time. For example, you can build llama.cpp with both CUDA and Vulkan support by using the `-DGGML_CUDA=ON -DGGML_VULKAN=ON` options with CMake. At runtime, you can specify which backend devices to use with the `--device` option. To see a list of available devices, use the `--list-devices` option.
+
+Backends can be built as dynamic libraries that can be loaded dynamically at runtime. This allows you to use the same llama.cpp binary on different machines with different GPUs. To enable this feature, use the `GGML_BACKEND_DL` option when building.

examples/base-translate.sh

@@ -1,61 +0,0 @@
-#!/bin/bash
-#
-# Few-shot translation example.
-# Requires a base model (i.e. no fine-tuned or instruct models).
-#
-# Usage:
-#
-#   cd llama.cpp
-#   make -j
-#
-#   ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
-#
-
-if [ $# -lt 2 ]; then
-  echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
-  exit 1
-fi
-
-eargs=""
-if [ $# -gt 2 ]; then
-  eargs="${@:3}"
-fi
-
-ftmp="__llama.cpp_example_tmp__.txt"
-trap "rm -f $ftmp" EXIT
-
-echo "Translate from English to French:
-
-===
-
-sea otter, peppermint, plush girafe:
-
-sea otter => loutre de mer
-peppermint => menthe poivrée
-plush girafe => girafe peluche
-
-===
-
-violin
-
-violin => violon
-
-===
-
-phone, computer, mouse, keyboard:
-
-phone => téléphone
-computer => ordinateur
-mouse => souris
-keyboard => clavier
-
-===
-" > $ftmp
-
-echo "$2
-" >> $ftmp
-
-model=$1
-
-# generate the most likely continuation until the string "===" is found
-./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

examples/batched-bench/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-batched-bench)
 add_executable(${TARGET} batched-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/batched/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-batched)
 add_executable(${TARGET} batched.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/convert-llama2c-to-ggml/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-convert-llama2c-to-ggml)
 add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/convert-llama2c-to-ggml/README.md

@@ -2,11 +2,8 @@
 
 This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
 
-To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository:
+To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository.
 
-`$ make -j`
-
-After successful compilation, following usage options are available:
 ```
 usage: ./llama-convert-llama2c-to-ggml [options]
 

examples/cvector-generator/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-cvector-generator)
 add_executable(${TARGET} cvector-generator.cpp pca.hpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/embedding/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-embedding)
 add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/eval-callback/CMakeLists.txt

@@ -2,7 +2,7 @@ set(TARGET llama-eval-callback)
 add_executable(${TARGET} eval-callback.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TEST_TARGET test-eval-callback)
 add_test(NAME ${TEST_TARGET}

examples/export-lora/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-export-lora)
 add_executable(${TARGET} export-lora.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gbnf-validator/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-gbnf-validator)
 add_executable(${TARGET} gbnf-validator.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gen-docs/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-gen-docs)
 add_executable(${TARGET} gen-docs.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gguf-hash/CMakeLists.txt

@@ -19,4 +19,4 @@ add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
 target_link_libraries(${TARGET} PRIVATE sha256)
 
 target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gguf-split/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-gguf-split)
 add_executable(${TARGET} gguf-split.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gguf/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-gguf)
 add_executable(${TARGET} gguf.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/gritlm/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-gritlm)
 add_executable(${TARGET} gritlm.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/imatrix/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-imatrix)
 add_executable(${TARGET} imatrix.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/imatrix/README.md

@@ -25,8 +25,6 @@ For faster computation, make sure to use GPU offloading via the `-ngl` argument
 ## Example
 
 ```bash
-GGML_CUDA=1 make -j
-
 # generate importance matrix (imatrix.dat)
 ./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
 

examples/imatrix/imatrix.cpp

@@ -637,10 +637,19 @@ int main(int argc, char ** argv) {
         LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
-    if (!compute_imatrix(ctx, params)) {
-        return 1;
+    if (params.prompt.empty()) {
+        if (params.in_files.empty()) {
+            LOG_ERR("Error: No prompt provided and no precomputed matrices (--in-file) to combine.\n");
+            return 1;
+        }
+        LOG_INF("No prompt provided; combining precomputed matrices only.\n");
+    } else {
+        if (!compute_imatrix(ctx, params)) {
+            return 1;
+        }
     }
 
+
     g_collector.save_imatrix();
 
     LOG("\n");

examples/infill/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-infill)
 add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llama-bench/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-bench)
 add_executable(${TARGET} llama-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llava/CMakeLists.txt

@@ -11,7 +11,7 @@ target_include_directories(llava PUBLIC .)
 target_include_directories(llava PUBLIC ../..)
 target_include_directories(llava PUBLIC ../../common)
 
-target_compile_features(llava PRIVATE cxx_std_11)
+target_compile_features(llava PRIVATE cxx_std_17)
 
 add_library(llava_static STATIC $<TARGET_OBJECTS:llava>)
 if (BUILD_SHARED_LIBS)
@@ -35,11 +35,11 @@ add_executable(${TARGET} llava-cli.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-minicpmv-cli)
 add_executable(${TARGET} minicpmv-cli.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/llava/clip.cpp

@@ -40,10 +40,17 @@
 #include <cinttypes>
 #include <limits>
 
-#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#define LOG_DBG(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#if defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...)
+#   define LOG_WRN(...)
+#   define LOG_ERR(...)
+#   define LOG_DBG(...)
+#else // defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#endif // defined(LLAVA_LOG_OFF)
 
 //#define CLIP_DEBUG_FUNCTIONS
 

examples/llava/llava.cpp

@@ -11,13 +11,17 @@
 #include <limits>
 #include <vector>
 
-#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
-#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
-
-#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#if defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...)
+#   define LOG_WRN(...)
+#   define LOG_ERR(...)
+#   define LOG_DBG(...)
+#else // defined(LLAVA_LOG_OFF)
+#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#endif // defined(LLAVA_LOG_OFF)
 
 // RGB uint8 image
 struct clip_image_u8 {
@@ -498,10 +502,16 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     errno = 0;
     size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
     if (ferror(file)) {
-        die_fmt("read error: %s", strerror(errno));
+        LOG_ERR("read error: %s", strerror(errno));
+        free(buffer);
+        fclose(file);
+        return false;
     }
     if (ret != (size_t) fileSize) {
-        die("unexpectedly reached end of file");
+        LOG_ERR("unexpectedly reached end of file");
+        free(buffer);
+        fclose(file);
+        return false;
     }
     fclose(file); // Close the file
 

examples/lookahead/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-lookahead)
 add_executable(${TARGET} lookahead.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/lookup/CMakeLists.txt

@@ -2,22 +2,22 @@ set(TARGET llama-lookup)
 add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-create)
 add_executable(${TARGET} lookup-create.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-merge)
 add_executable(${TARGET} lookup-merge.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-stats)
 add_executable(${TARGET} lookup-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/main-cmake-pkg/CMakeLists.txt

@@ -29,4 +29,4 @@ add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
 target_include_directories(${TARGET} PRIVATE ${_common_path})
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/main/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-cli)
 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/parallel/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-parallel)
 add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/passkey/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-passkey)
 add_executable(${TARGET} passkey.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/perplexity/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-perplexity)
 add_executable(${TARGET} perplexity.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/quantize-stats/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/quantize/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/retrieval/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-retrieval)
 add_executable(${TARGET} retrieval.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/run/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-run)
 add_executable(${TARGET} run.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/save-load-state/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-save-load-state)
 add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/server/CMakeLists.txt

@@ -50,4 +50,4 @@ if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
 
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/server/README.md

@@ -69,6 +69,8 @@ The project is under active development, and we are [looking for feedback and co
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
+| `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
+| `--list-devices` | print list of available devices and exit |
 | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
@@ -158,9 +160,16 @@ The project is under active development, and we are [looking for feedback and co
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted:<br/>https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>list of built-in templates:<br/>chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
+| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16) |
+| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9) |
+| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model) |
+| `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model |
+| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
@@ -188,12 +197,6 @@ services:
 
 `llama-server` is built alongside everything else from the root of the project
 
-- Using `make`:
-
-  ```bash
-  make llama-server
-  ```
-
 - Using `CMake`:
 
   ```bash
@@ -207,15 +210,6 @@ services:
 
 `llama-server` can also be built with SSL support using OpenSSL 3
 
-- Using `make`:
-
-  ```bash
-  # NOTE: For non-system openssl, use the following:
-  #   CXXFLAGS="-I /path/to/openssl/include"
-  #   LDFLAGS="-L /path/to/openssl/lib"
-  make LLAMA_SERVER_SSL=true llama-server
-  ```
-
 - Using `CMake`:
 
   ```bash
@@ -416,6 +410,8 @@ node index.js
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 
+    `timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
+
 **Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.

examples/server/server.cpp

@@ -177,6 +177,8 @@ struct server_slot {
     bool stopped_word   = false;
     bool stopped_limit  = false;
 
+    bool timings_per_token = false;
+
     bool oaicompat = false;
 
     std::string oaicompat_model;
@@ -694,8 +696,9 @@ struct server_context {
 
             params_dft.devices      = params_base.speculative.devices;
             params_dft.model        = params_base.speculative.model;
-            params_dft.n_ctx        = params_base.speculative.n_ctx;
+            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;
 
             common_init_result llama_init_dft = common_init_from_params(params_dft);
 
@@ -715,8 +718,14 @@ struct server_context {
                 return false;
             }
 
-            cparams_dft = common_context_params_to_llama(params_base);
-            cparams_dft.n_batch = llama_n_ctx(llama_init_dft.context);
+            const int n_ctx_dft = llama_n_ctx(llama_init_dft.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_free(llama_init_dft.context);
@@ -882,6 +891,8 @@ struct server_context {
             slot.oaicompat_model = "";
         }
 
+        slot.timings_per_token       = json_value(data, "timings_per_token",  false);
+
         slot.params.stream           = json_value(data, "stream",             false);
         slot.params.cache_prompt     = json_value(data, "cache_prompt",       true);
         slot.params.n_predict        = json_value(data, "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
@@ -1279,6 +1290,7 @@ struct server_context {
             {"speculative.n_max",         slot.params.speculative.n_max},
             {"speculative.n_min",         slot.params.speculative.n_min},
             {"speculative.p_min",         slot.params.speculative.p_min},
+            {"timings_per_token",         slot.timings_per_token},
         };
     }
 
@@ -1336,6 +1348,10 @@ struct server_context {
             res.data["model"] = slot.oaicompat_model;
         }
 
+        if (slot.timings_per_token) {
+            res.data["timings"] = slot.get_formated_timings();
+        }
+
         queue_results.send(res);
     }
 
@@ -2274,12 +2290,17 @@ struct server_context {
                 common_sampler_accept(slot.smpl, id, true);
 
                 slot.n_decoded += 1;
+
+                const int64_t t_current = ggml_time_us();
+
                 if (slot.n_decoded == 1) {
-                    slot.t_start_generation = ggml_time_us();
+                    slot.t_start_generation = t_current;
                     slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
                     metrics.on_prompt_eval(slot);
                 }
 
+                slot.t_token_generation = (t_current - slot.t_start_generation) / 1e3;
+
                 completion_token_output result;
                 result.tok = id;
 
@@ -2308,6 +2329,10 @@ struct server_context {
                     continue;
                 }
 
+                if (slot.state != SLOT_STATE_GENERATING) {
+                    continue;
+                }
+
                 llama_token id = slot.sampled;
 
                 struct common_speculative_params params_spec;
@@ -3347,8 +3372,18 @@ int main(int argc, char ** argv) {
         llama_backend_free();
     };
 
-    // bind HTTP listen port, run the HTTP server in a thread
-    if (!svr->bind_to_port(params.hostname, params.port)) {
+    // 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;
+        }
+    } else {
+        was_bound = svr->bind_to_port(params.hostname, params.port);
+    }
+
+    if (!was_bound) {
         //LOG_ERROR("couldn't bind HTTP server socket", {
         //    {"hostname", params.hostname},
         //    {"port", params.port},
@@ -3357,6 +3392,8 @@ int main(int argc, char ** argv) {
         clean_up();
         return 1;
     }
+
+    // run the HTTP server in a thread
     std::thread t([&]() { svr->listen_after_bind(); });
     svr->wait_until_ready();
 

examples/server/tests/requirements.txt

@@ -2,6 +2,6 @@ aiohttp~=3.9.3
 pytest~=8.3.3
 huggingface_hub~=0.23.2
 numpy~=1.26.4
-openai~=1.30.3
+openai~=1.55.3
 prometheus-client~=0.20.0
 requests~=2.32.3

examples/server/tests/unit/test_basic.py

@@ -32,3 +32,17 @@ def test_server_models():
     assert res.status_code == 200
     assert len(res.body["data"]) == 1
     assert res.body["data"][0]["id"] == server.model_alias
+
+def test_load_split_model():
+    global server
+    server.model_hf_repo = "ggml-org/models"
+    server.model_hf_file = "tinyllamas/split/stories15M-q8_0-00001-of-00003.gguf"
+    server.model_alias = "tinyllama-split"
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": 16,
+        "prompt": "Hello",
+        "temperature": 0.0,
+    })
+    assert res.status_code == 200
+    assert match_regex("(little|girl)+", res.body["content"])

examples/server/tests/unit/test_chat_completion.py

@@ -127,3 +127,39 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int
         assert res.status_code != 200
         assert "error" in res.body
 
+
+@pytest.mark.parametrize("messages", [
+    None,
+    "string",
+    [123],
+    [{}],
+    [{"role": 123}],
+    [{"role": "system", "content": 123}],
+    # [{"content": "hello"}], # TODO: should not be a valid case
+    [{"role": "system", "content": "test"}, {}],
+])
+def test_invalid_chat_completion_req(messages):
+    global server
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "messages": messages,
+    })
+    assert res.status_code == 400 or res.status_code == 500
+    assert "error" in res.body
+
+
+def test_chat_completion_with_timings_per_token():
+    global server
+    server.start()
+    res = server.make_stream_request("POST", "/chat/completions", data={
+        "max_tokens": 10,
+        "messages": [{"role": "user", "content": "test"}],
+        "stream": True,
+        "timings_per_token": True,
+    })
+    for data in res:
+        assert "timings" in data
+        assert "prompt_per_second" in data["timings"]
+        assert "predicted_per_second" in data["timings"]
+        assert "predicted_n" in data["timings"]
+        assert data["timings"]["predicted_n"] <= 10

examples/server/tests/unit/test_infill.py

@@ -8,6 +8,7 @@ def create_server():
     global server
     server = ServerPreset.tinyllama_infill()
 
+
 def test_infill_without_input_extra():
     global server
     server.start()
@@ -19,6 +20,7 @@ def test_infill_without_input_extra():
     assert res.status_code == 200
     assert match_regex("(One|day|she|saw|big|scary|bird)+", res.body["content"])
 
+
 def test_infill_with_input_extra():
     global server
     server.start()
@@ -33,3 +35,23 @@ def test_infill_with_input_extra():
     })
     assert res.status_code == 200
     assert match_regex("(cuts|Jimmy|mom|came|into|the|room)+", res.body["content"])
+
+
+@pytest.mark.parametrize("input_extra", [
+    {},
+    {"filename": "ok"},
+    {"filename": 123},
+    {"filename": 123, "text": "abc"},
+    {"filename": 123, "text": 456},
+])
+def test_invalid_input_extra_req(input_extra):
+    global server
+    server.start()
+    res = server.make_request("POST", "/infill", data={
+        "prompt": "Complete this",
+        "input_extra": [input_extra],
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
+        "input_suffix": "}\n",
+    })
+    assert res.status_code == 400
+    assert "error" in res.body

examples/server/tests/unit/test_rerank.py

@@ -36,3 +36,20 @@ def test_rerank():
     assert most_relevant["relevance_score"] > least_relevant["relevance_score"]
     assert most_relevant["index"] == 2
     assert least_relevant["index"] == 3
+
+
+@pytest.mark.parametrize("documents", [
+    [],
+    None,
+    123,
+    [1, 2, 3],
+])
+def test_invalid_rerank_req(documents):
+    global server
+    server.start()
+    res = server.make_request("POST", "/rerank", data={
+        "query": "Machine learning is",
+        "documents": documents,
+    })
+    assert res.status_code == 400
+    assert "error" in res.body

examples/server/tests/unit/test_speculative.py

@@ -0,0 +1,103 @@
+import pytest
+from utils import *
+
+# We use a F16 MOE gguf as main model, and q4_0 as draft model
+
+server = ServerPreset.stories15m_moe()
+
+MODEL_DRAFT_FILE_URL = "https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories15M-q4_0.gguf"
+
+def create_server():
+    global server
+    server = ServerPreset.stories15m_moe()
+    # download draft model file if needed
+    file_name = MODEL_DRAFT_FILE_URL.split('/').pop()
+    model_draft_file = f'../../../{file_name}'
+    if not os.path.exists(model_draft_file):
+        print(f"Downloading {MODEL_DRAFT_FILE_URL} to {model_draft_file}")
+        with open(model_draft_file, 'wb') as f:
+            f.write(requests.get(MODEL_DRAFT_FILE_URL).content)
+        print(f"Done downloading draft model file")
+    # set default values
+    server.model_draft = model_draft_file
+    server.draft_min = 4
+    server.draft_max = 8
+
+
+@pytest.fixture(scope="module", autouse=True)
+def fixture_create_server():
+    return create_server()
+
+
+def test_with_and_without_draft():
+    global server
+    server.model_draft = None  # disable draft model
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "temperature": 0.0,
+        "top_k": 1,
+    })
+    assert res.status_code == 200
+    content_no_draft = res.body["content"]
+    server.stop()
+
+    # create new server with draft model
+    create_server()
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "temperature": 0.0,
+        "top_k": 1,
+    })
+    assert res.status_code == 200
+    content_draft = res.body["content"]
+
+    assert content_no_draft == content_draft
+
+
+def test_different_draft_min_draft_max():
+    global server
+    test_values = [
+        (1, 2),
+        (1, 4),
+        (4, 8),
+        (4, 12),
+        (8, 16),
+    ]
+    last_content = None
+    for draft_min, draft_max in test_values:
+        server.stop()
+        server.draft_min = draft_min
+        server.draft_max = draft_max
+        server.start()
+        res = server.make_request("POST", "/completion", data={
+            "prompt": "I believe the meaning of life is",
+            "temperature": 0.0,
+            "top_k": 1,
+        })
+        assert res.status_code == 200
+        if last_content is not None:
+            assert last_content == res.body["content"]
+        last_content = res.body["content"]
+
+
+@pytest.mark.parametrize("n_slots,n_requests", [
+    (1, 2),
+    (2, 2),
+])
+def test_multi_requests_parallel(n_slots: int, n_requests: int):
+    global server
+    server.n_slots = n_slots
+    server.start()
+    tasks = []
+    for _ in range(n_requests):
+        tasks.append((server.make_request, ("POST", "/completion", {
+            "prompt": "I believe the meaning of life is",
+            "temperature": 0.0,
+            "top_k": 1,
+        })))
+    results = parallel_function_calls(tasks)
+    for res in results:
+        assert res.status_code == 200
+        assert match_regex("(wise|kind|owl|answer)+", res.body["content"])

examples/server/tests/utils.py

@@ -8,7 +8,6 @@ import os
 import re
 import json
 import sys
-import threading
 import requests
 import time
 from concurrent.futures import ThreadPoolExecutor, as_completed
@@ -47,6 +46,7 @@ class ServerProcess:
     model_alias: str | None = None
     model_url: str | None = None
     model_file: str | None = None
+    model_draft: str | None = None
     n_threads: int | None = None
     n_gpu_layer: int | None = None
     n_batch: int | None = None
@@ -69,6 +69,8 @@ class ServerProcess:
     response_format: str | None = None
     lora_files: List[str] | None = None
     disable_ctx_shift: int | None = False
+    draft_min: int | None = None
+    draft_max: int | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -103,6 +105,8 @@ class ServerProcess:
             server_args.extend(["--model", self.model_file])
         if self.model_url:
             server_args.extend(["--model-url", self.model_url])
+        if self.model_draft:
+            server_args.extend(["--model-draft", self.model_draft])
         if self.model_hf_repo:
             server_args.extend(["--hf-repo", self.model_hf_repo])
         if self.model_hf_file:
@@ -148,6 +152,10 @@ class ServerProcess:
             server_args.extend(["--no-context-shift"])
         if self.api_key:
             server_args.extend(["--api-key", self.api_key])
+        if self.draft_max:
+            server_args.extend(["--draft-max", self.draft_max])
+        if self.draft_min:
+            server_args.extend(["--draft-min", self.draft_min])
 
         args = [str(arg) for arg in [server_path, *server_args]]
         print(f"bench: starting server with: {' '.join(args)}")
@@ -161,26 +169,12 @@ class ServerProcess:
         self.process = subprocess.Popen(
             [str(arg) for arg in [server_path, *server_args]],
             creationflags=flags,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
+            stdout=sys.stdout,
+            stderr=sys.stdout,
             env={**os.environ, "LLAMA_CACHE": "tmp"},
         )
         server_instances.add(self)
 
-        def server_log(in_stream, out_stream):
-            for line in iter(in_stream.readline, b""):
-                print(line.decode("utf-8"), end="", file=out_stream)
-
-        thread_stdout = threading.Thread(
-            target=server_log, args=(self.process.stdout, sys.stdout), daemon=True
-        )
-        thread_stdout.start()
-
-        thread_stderr = threading.Thread(
-            target=server_log, args=(self.process.stderr, sys.stderr), daemon=True
-        )
-        thread_stderr.start()
-
         print(f"server pid={self.process.pid}, pytest pid={os.getpid()}")
 
         # wait for server to start
@@ -200,7 +194,8 @@ class ServerProcess:
         raise TimeoutError(f"Server did not start within {timeout_seconds} seconds")
 
     def stop(self) -> None:
-        server_instances.remove(self)
+        if self in server_instances:
+            server_instances.remove(self)
         if self.process:
             print(f"Stopping server with pid={self.process.pid}")
             self.process.kill()

examples/server/utils.hpp

@@ -650,6 +650,10 @@ static json format_final_response_oaicompat(const json & request, const json & r
         res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
     }
 
+    if (result.contains("timings")) {
+        res.push_back({"timings", json_value(result, "timings", json::object())});
+    }
+
     return res;
 }
 
@@ -740,6 +744,11 @@ static std::vector<json> format_partial_response_oaicompat(const json & result,
         {"model",   modelname},
         {"object",  "chat.completion.chunk"}
     };
+
+    if (result.contains("timings")) {
+        ret.push_back({"timings", json_value(result, "timings", json::object())});
+    }
+
     if (!finish_reason.empty()) {
         int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
         int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);

examples/simple-chat/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-simple-chat)
 add_executable(${TARGET} simple-chat.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/simple/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-simple)
 add_executable(${TARGET} simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/simple/README.md

@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is"
 
 ...
 

examples/speculative-simple/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-speculative-simple)
 add_executable(${TARGET} speculative-simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/speculative/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-speculative)
 add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/tokenize/CMakeLists.txt

@@ -2,4 +2,4 @@ set(TARGET llama-tokenize)
 add_executable(${TARGET} tokenize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

ggml/CMakeLists.txt

@@ -96,6 +96,7 @@ option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 
 option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX_VNNI    "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
 option(GGML_AVX512      "ggml: enable AVX512"           OFF)
 option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
@@ -161,7 +162,6 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
-option(GGML_AMX                             "ggml: use AMX"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING

ggml/include/ggml-amx.h

@@ -1,25 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-// buffer_type API
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
-
-GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
-
-// backend API
-GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
-
-GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
-
-#ifdef  __cplusplus
-}
-#endif

ggml/include/ggml-cpu.h

@@ -91,6 +91,7 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_neon       (void);
     GGML_BACKEND_API int ggml_cpu_has_arm_fma    (void);
     GGML_BACKEND_API int ggml_cpu_has_fp16_va    (void);
+    GGML_BACKEND_API int ggml_cpu_has_dotprod    (void);
     GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
     GGML_BACKEND_API int ggml_cpu_has_sve        (void);
     GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes

ggml/include/ggml.h

@@ -389,6 +389,9 @@ extern "C" {
         GGML_TYPE_Q4_0_8_8 = 33,
         GGML_TYPE_TQ1_0   = 34,
         GGML_TYPE_TQ2_0   = 35,
+        GGML_TYPE_IQ4_NL_4_4 = 36,
+        // GGML_TYPE_IQ4_NL_4_8 = 37,
+        // GGML_TYPE_IQ4_NL_8_8 = 38,
         GGML_TYPE_COUNT,
     };
 

ggml/src/CMakeLists.txt

@@ -261,21 +261,15 @@ function(ggml_add_backend backend)
     if (${backend_id})
         string(TOLOWER "ggml-${backend}" backend_target)
         add_subdirectory(${backend_target})
-        # check again in case the backend disabled itself
-        # note that this should NOT be the normal behavior, in case of errors the backend should fail the build
-        # however, currently it is necessary for AMX, since it is enabled by default on llama.cpp
-        if (${backend_id})
-            message(STATUS "Including ${backend} backend")
-            if (NOT GGML_BACKEND_DL)
-                string(TOUPPER "GGML_USE_${backend}" backend_use)
-                target_compile_definitions(ggml PUBLIC ${backend_use})
-            endif()
+        message(STATUS "Including ${backend} backend")
+        if (NOT GGML_BACKEND_DL)
+            string(TOUPPER "GGML_USE_${backend}" backend_use)
+            target_compile_definitions(ggml PUBLIC ${backend_use})
         endif()
     endif()
 endfunction()
 
 ggml_add_backend(CPU)
-ggml_add_backend(AMX)
 ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
@@ -289,7 +283,7 @@ ggml_add_backend(Vulkan)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
-    target_compile_features   (${target} PRIVATE c_std_11) # don't bump
+    target_compile_features   (${target} PRIVATE c_std_11 cxx_std_17) # don't bump
 endforeach()
 
 target_link_libraries(ggml-base PRIVATE Threads::Threads)

ggml/src/ggml-amx/CMakeLists.txt

@@ -1,105 +0,0 @@
-if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
-        CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
-    message(STATUS "Using AMX")
-
-    file(GLOB   GGML_HEADERS_AMX "*.h")
-    list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
-
-    file(GLOB   GGML_SOURCES_AMX "*.cpp")
-
-    ggml_add_backend_library(ggml-amx
-                             ${GGML_HEADERS_AMX}
-                             ${GGML_SOURCES_AMX}
-                            )
-
-    # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
-    # TODO: integrate AMX backend into the CPU backend
-    if (MSVC)
-        # instruction set detection for MSVC only
-        if (GGML_NATIVE)
-            # TODO: improve, should not reference files from the parent folder
-            include(../ggml-cpu/cmake/FindSIMD.cmake)
-        endif ()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS /arch:AVX512)
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (GGML_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
-            endif()
-            if (GGML_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
-            endif()
-            if (GGML_AVX512_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
-            endif()
-            if (GGML_AMX_TILE)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
-            endif()
-            if (GGML_AMX_INT8)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
-            endif()
-            if (GGML_AMX_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
-            endif()
-        elseif (GGML_AVX2)
-            list(APPEND ARCH_FLAGS /arch:AVX2)
-        elseif (GGML_AVX)
-            list(APPEND ARCH_FLAGS /arch:AVX)
-        endif()
-    else()
-        if (GGML_NATIVE)
-            list(APPEND ARCH_FLAGS -march=native)
-        endif()
-        if (GGML_F16C)
-            list(APPEND ARCH_FLAGS -mf16c)
-        endif()
-        if (GGML_FMA)
-            list(APPEND ARCH_FLAGS -mfma)
-        endif()
-        if (GGML_AVX)
-            list(APPEND ARCH_FLAGS -mavx)
-        endif()
-        if (GGML_AVX2)
-            list(APPEND ARCH_FLAGS -mavx2)
-        endif()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f)
-            list(APPEND ARCH_FLAGS -mavx512dq)
-            list(APPEND ARCH_FLAGS -mavx512bw)
-        endif()
-        if (GGML_AVX512_VBMI)
-            list(APPEND ARCH_FLAGS -mavx512vbmi)
-        endif()
-        if (GGML_AVX512_VNNI)
-            list(APPEND ARCH_FLAGS -mavx512vnni)
-        endif()
-        if (GGML_AVX512_BF16)
-            list(APPEND ARCH_FLAGS -mavx512bf16)
-        endif()
-        if (GGML_AMX_TILE)
-            list(APPEND ARCH_FLAGS -mamx-tile)
-        endif()
-        if (GGML_AMX_INT8)
-            list(APPEND ARCH_FLAGS -mamx-int8)
-        endif()
-        if (GGML_AMX_BF16)
-            list(APPEND ARCH_FLAGS -mamx-bf16)
-        endif()
-    endif()
-
-    target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
-else()
-    set(GGML_AMX OFF PARENT_SCOPE)
-    message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
-endif()

ggml/src/ggml-amx/ggml-amx.cpp

@@ -1,449 +0,0 @@
-#include "ggml-amx.h"
-#include "ggml-amx/common.h"
-#include "ggml-amx/mmq.h"
-#include "ggml-backend-impl.h"
-#include "ggml-impl.h"
-
-#if defined(__gnu_linux__)
-#include <sys/syscall.h>
-#include <unistd.h>
-#endif
-
-#include <cstdlib>
-#include <cstring>
-#include <memory>
-
-#if defined(__AMX_INT8__)
-
-// AMX buffer interface
-static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-}
-
-static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)(buffer->context);
-}
-
-static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    memset((char *)tensor->data + offset, value, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    if (qtype_has_amx_kernels(tensor->type)) {
-        ggml_backend_amx_convert_weight(tensor, data, offset, size);
-    } else {
-        memcpy((char *)tensor->data + offset, data, size);
-    }
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        if (qtype_has_amx_kernels(src->type)) {
-            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
-        } else {
-            memcpy(dst->data, src->data, ggml_nbytes(src));
-        }
-        return true;
-    }
-    return false;
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    memset(buffer->context, value, buffer->size);
-}
-
-static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
-    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
-    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
-    /* .clear           = */ ggml_backend_amx_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "AMX";
-
-    GGML_UNUSED(buft);
-}
-
-static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
-    if (data == NULL) {
-        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
-        return NULL;
-    }
-
-    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
-}
-
-static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return TENSOR_ALIGNMENT;
-
-    GGML_UNUSED(buft);
-}
-
-static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
-    return ggml_backend_amx_get_alloc_size(tensor);
-
-    GGML_UNUSED(buft);
-}
-
-static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return false;
-
-    GGML_UNUSED(buft);
-}
-
-ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
-            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
-        },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_buffer_type_amx;
-}
-
-// backend interface
-
-static const char * ggml_backend_amx_name(ggml_backend_t backend) {
-    return "AMX";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_amx_free(ggml_backend_t backend) {
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
-    delete ctx;
-    delete backend;
-}
-
-static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
-
-        switch (node->op) {
-        case GGML_OP_MUL_MAT:
-            ggml_backend_amx_mul_mat(ctx, node);
-            break;
-
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            break;
-
-        default:
-            fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-            GGML_ASSERT(false);
-        }
-    }
-
-    return GGML_STATUS_SUCCESS;
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i ggml_backend_amx_i = {
-    /* .get_name                = */ ggml_backend_amx_name,
-    /* .free                    = */ ggml_backend_amx_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_amx_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_amx_guid() {
-    static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
-    return &guid;
-}
-
-#define ARCH_GET_XCOMP_PERM     0x1022
-#define ARCH_REQ_XCOMP_PERM     0x1023
-#define XFEATURE_XTILECFG       17
-#define XFEATURE_XTILEDATA      18
-
-static bool ggml_amx_init() {
-#if defined(__gnu_linux__)
-    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
-        fprintf(stderr, "AMX is not ready to be used!\n");
-        return false;
-    }
-    return true;
-#elif defined(_WIN32)
-    return true;
-#endif
-}
-
-ggml_backend_t ggml_backend_amx_init() {
-
-    // invoke a Linux system call to request access to AMX features
-    ggml_amx_init();
-
-    // backend context
-    ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
-
-    // ggml amx backend
-    ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_amx_guid(),
-        /* .interface = */ ggml_backend_amx_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-    return backend;
-}
-
-bool ggml_backend_is_amx(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
-}
-
-void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_amx(backend_amx));
-
-    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
-    ctx->n_threads = n_threads;
-}
-
-// device interface
-
-static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
-    return "AMX";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
-    return "Intel Advanced Matrix Extensions";
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
-    GGML_UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_amx_device_get_name(dev);
-    props->description = ggml_backend_amx_device_get_description(dev);
-    props->type        = ggml_backend_amx_device_get_type(dev);
-    ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
-
-    // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_amx_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_amx_buffer_type();
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-
-    // handle only 2d gemm for now
-    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
-        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
-    };
-
-    switch (op->op) {
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            return true;
-
-        case GGML_OP_MUL_MAT: {
-            const struct ggml_tensor * src0 = op->src[0];
-            const struct ggml_tensor * src1 = op->src[1];
-
-            const enum ggml_type type = src0->type;
-            const int64_t ne0 = op->ne[0];
-
-            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
-            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
-            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
-
-            bool can_use_amx =
-                is_contiguous_2d(src0) &&       // src0 must be contiguous
-                is_contiguous_2d(src1) &&       // src1 must be contiguous
-                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
-                has_amx_kernels &&              // with amx kernel impls
-                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
-
-            return can_use_amx;
-        }
-        default:
-            return false;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
-    /* .get_name             = */ ggml_backend_amx_device_get_name,
-    /* .get_description      = */ ggml_backend_amx_device_get_description,
-    /* .get_memory           = */ ggml_backend_amx_device_get_memory,
-    /* .get_type             = */ ggml_backend_amx_device_get_type,
-    /* .get_props            = */ ggml_backend_amx_device_get_props,
-    /* .init_backend         = */ ggml_backend_amx_device_init,
-    /* .get_buffer_type      = */ ggml_backend_amx_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_amx_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_amx_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// backend reg interface
-
-static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
-    return "AMX";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_device ggml_backend_amx_device = {
-        /* .iface   = */ ggml_backend_amx_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ nullptr,
-    };
-
-    return &ggml_backend_amx_device;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(index);
-}
-
-static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_amx_set_n_threads;
-    }
-    return NULL;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(name);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
-    /* .get_name         = */ ggml_backend_amx_reg_get_name,
-    /* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_amx_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_amx_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_amx_reg(void) {
-    static struct ggml_backend_reg ggml_backend_amx_reg = {
-        /* .api_version = */ GGML_BACKEND_API_VERSION,
-        /* .iface       = */ ggml_backend_amx_reg_i,
-        /* .context     = */ NULL,
-    };
-
-    return &ggml_backend_amx_reg;
-}
-
-#else // if defined(__AMX_INT8__)
-
-ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
-    return nullptr;
-}
-
-bool ggml_backend_is_amx(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
-    return false;
-}
-
-ggml_backend_t ggml_backend_amx_init(void) {
-    fprintf(stderr, "GGML is not compiled with AMX support!\n");
-    return nullptr;
-}
-
-void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
-    fprintf(stderr, "GGML is not compiled with AMX support!\n");
-
-    GGML_UNUSED(backend_amx);
-    GGML_UNUSED(n_threads);
-}
-
-ggml_backend_reg_t ggml_backend_amx_reg(void) {
-    return nullptr;
-}
-
-#endif
-
-GGML_BACKEND_DL_IMPL(ggml_backend_amx_reg)

ggml/src/ggml-backend-impl.h

@@ -211,27 +211,45 @@ extern "C" {
     GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Add backend dynamic loading support to the backend
-    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
 
-    #ifdef GGML_BACKEND_DL
-        #ifdef __cplusplus
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
-                extern "C" {                                                     \
-                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                }                                                                \
-                ggml_backend_reg_t ggml_backend_init(void) {                     \
-                    return reg_fn();                                             \
-                }
-        #else
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                             \
-                GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                ggml_backend_reg_t ggml_backend_init(void) {                 \
-                    return reg_fn();                                         \
-                }
-        #endif
-    #else
-    #    define GGML_BACKEND_DL_IMPL(reg_fn)
-    #endif
+    // Initialize the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+    // Optional: obtain a score for the backend based on the system configuration
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
+    typedef int                (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+#    ifdef __cplusplus
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            extern "C" {                                                 \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+            }                                                            \
+            ggml_backend_reg_t ggml_backend_init(void) {                 \
+                return reg_fn();                                         \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
+            extern "C" {                                   \
+            GGML_BACKEND_API int ggml_backend_score(void); \
+            }                                              \
+            int ggml_backend_score(void) {                 \
+                return score_fn();                         \
+            }
+#    else
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
+            ggml_backend_reg_t                  ggml_backend_init(void) { \
+                return reg_fn();                                          \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
+            GGML_BACKEND_API int ggml_backend_score(void);  \
+            int                  ggml_backend_score(void) { \
+                return score_fn();                          \
+            }
+#    endif
+#else
+#    define GGML_BACKEND_DL_IMPL(reg_fn)
+#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend-reg.cpp

@@ -2,8 +2,13 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include <algorithm>
+#include <codecvt>
 #include <cstring>
+#include <filesystem>
+#include <locale>
+#include <memory>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 #ifdef _WIN32
@@ -49,10 +54,6 @@
 #include "ggml-rpc.h"
 #endif
 
-#ifdef GGML_USE_AMX
-#  include "ggml-amx.h"
-#endif
-
 #ifdef GGML_USE_CANN
 #include "ggml-cann.h"
 #endif
@@ -61,9 +62,71 @@
 #include "ggml-kompute.h"
 #endif
 
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static dl_handle * dl_load_library(const std::wstring & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static dl_handle * dl_load_library(const std::string & path) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return dl_load_library(converter.from_bytes(path));
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static void * dl_load_library(const std::string & path) {
+    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
 struct ggml_backend_reg_entry {
     ggml_backend_reg_t reg;
-    void * handle;
+    dl_handle_ptr handle;
 };
 
 struct ggml_backend_registry {
@@ -92,9 +155,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_AMX
-        register_backend(ggml_backend_amx_reg());
-#endif
 #ifdef GGML_USE_KOMPUTE
         register_backend(ggml_backend_kompute_reg());
 #endif
@@ -104,13 +164,16 @@ struct ggml_backend_registry {
     }
 
     ~ggml_backend_registry() {
-        while (!backends.empty()) {
-            // use silent since the log system may have been destroyed at this point
-            unload_backend(backends.back().reg, true);
+        // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
+        // since backend threads may still be running and accessing resources from the dynamic library
+        for (auto & entry : backends) {
+            if (entry.handle) {
+                entry.handle.release(); // NOLINT
+            }
         }
     }
 
-    void register_backend(ggml_backend_reg_t reg, void * handle = nullptr) {
+    void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -119,7 +182,7 @@ struct ggml_backend_registry {
         GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
 #endif
-        backends.push_back({ reg, handle });
+        backends.push_back({ reg, std::move(handle) });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
             register_device(ggml_backend_reg_dev_get(reg, i));
         }
@@ -133,79 +196,53 @@ struct ggml_backend_registry {
     }
 
     ggml_backend_reg_t load_backend(const char * path, bool silent) {
-#ifdef _WIN32
-        // suppress error dialogs for missing DLLs
-        DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
-        SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
-
-        HMODULE handle = LoadLibraryA(path);
-
+        dl_handle_ptr handle { dl_load_library(path) };
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s: %lu\n", __func__, path, GetLastError());
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
             }
-            SetErrorMode(old_mode);
             return nullptr;
         }
 
-        ggml_backend_init_t backend_init = (ggml_backend_init_t) GetProcAddress(handle, "ggml_backend_init");
-
-        SetErrorMode(old_mode);
-
-        if (!backend_init) {
+        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+        if (score_fn && score_fn() == 0) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %lu\n", __func__, path, GetLastError());
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
             }
-            FreeLibrary(handle);
             return nullptr;
         }
-#else
-        void * handle = dlopen(path, RTLD_NOW | RTLD_LOCAL);
 
-        if (!handle) {
+        auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
+        if (!backend_init_fn) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path, dlerror());
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
             }
             return nullptr;
         }
 
-        auto * backend_init = (ggml_backend_init_t) dlsym(handle, "ggml_backend_init");
-
-        if (!backend_init) {
-            if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %s\n", __func__, path, dlerror());
-            }
-            dlclose(handle);
-            return nullptr;
-        }
-#endif
-        ggml_backend_reg_t reg = backend_init();
-
+        ggml_backend_reg_t reg = backend_init_fn();
         if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
             if (!silent) {
                 if (!reg) {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
                 } else {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                                   __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
                 }
             }
-#ifdef _WIN32
-            FreeLibrary(handle);
-#else
-            dlclose(handle);
-#endif
             return nullptr;
         }
 
         GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
-        register_backend(reg, handle);
+
+        register_backend(reg, std::move(handle));
+
         return reg;
     }
 
     void unload_backend(ggml_backend_reg_t reg, bool silent) {
         auto it = std::find_if(backends.begin(), backends.end(),
-                                [reg](ggml_backend_reg_entry entry) { return entry.reg == reg; });
+                               [reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
 
         if (it == backends.end()) {
             if (!silent) {
@@ -224,15 +261,6 @@ struct ggml_backend_registry {
                             [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
             devices.end());
 
-        // unload library
-        if (it->handle) {
-#ifdef _WIN32
-            FreeLibrary((HMODULE) it->handle);
-#else
-            dlclose(it->handle);
-#endif
-        }
-
         // remove backend
         backends.erase(it);
     }
@@ -348,12 +376,7 @@ void ggml_backend_unload(ggml_backend_reg_t reg) {
     get_reg().unload_backend(reg, true);
 }
 
-void ggml_backend_load_all() {
-    std::vector<std::string> search_prefix;
-
-    // add the executable directory to the search path
-    // FIXME: this is convenient for development, but it should probably be disabled in production
-
+static std::string get_executable_path() {
 #if defined(__APPLE__)
     // get executable path
     std::vector<char> path;
@@ -371,7 +394,7 @@ void ggml_backend_load_all() {
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    search_prefix.push_back(base_path + "/");
+    return base_path + "/";
 #elif defined(__linux__)
     std::string base_path = ".";
     std::vector<char> path(1024);
@@ -393,38 +416,104 @@ void ggml_backend_load_all() {
         path.resize(path.size() * 2);
     }
 
-    search_prefix.push_back(base_path + "/");
+    return base_path + "/";
+#elif defined(_WIN32)
+    std::vector<char> path(MAX_PATH);
+    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    if (len == 0) {
+        return "";
+    }
+    std::string base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "\\";
 #endif
+}
 
-    auto & reg = get_reg();
-
-    auto try_load = [&](const std::string & name) {
-        std::string os_name;
+static std::string backend_filename_prefix() {
 #ifdef _WIN32
-        os_name = "ggml-" + name + ".dll";
+    return "ggml-";
 #else
-        os_name = "libggml-" + name + ".so";
+    return "libggml-";
 #endif
-        if (reg.load_backend(os_name.c_str(), true)) {
-            return;
+}
+
+static std::string backend_filename_suffix() {
+#ifdef _WIN32
+    return ".dll";
+#else
+    return ".so";
+#endif
+}
+
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) {
+    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
+     // TODO: search system paths
+    std::vector<std::string> search_paths = { "./", get_executable_path() };
+    std::string file_prefix = backend_filename_prefix() + name + "-";
+
+    int best_score = 0;
+    std::string best_path;
+
+    namespace fs = std::filesystem;
+    for (const auto & search_path : search_paths) {
+        if (!fs::exists(search_path)) {
+            continue;
         }
-        for (const auto & prefix : search_prefix) {
-            if (reg.load_backend((prefix + os_name).c_str(), true)) {
-                return;
+        for (const auto & entry : fs::directory_iterator(search_path)) {
+            if (entry.is_regular_file()) {
+                std::string filename = entry.path().filename().string();
+                std::string ext = entry.path().extension().string();
+                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
+                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    if (!handle && !silent) {
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                    }
+                    if (handle) {
+                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().string();
+                            }
+                        }
+                    }
+                }
             }
         }
-    };
+    }
 
-    try_load("amx");
-    try_load("blas");
-    try_load("cann");
-    try_load("cuda");
-    try_load("hip");
-    try_load("kompute");
-    try_load("metal");
-    try_load("rpc");
-    try_load("sycl");
-    try_load("vulkan");
-    try_load("musa");
-    try_load("cpu");
+    if (best_score == 0) {
+        // try to load the base backend
+        for (const auto & search_path : search_paths) {
+            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
+            if (fs::exists(path)) {
+                return get_reg().load_backend(path.c_str(), silent);
+            }
+        }
+        return nullptr;
+    }
+
+    return get_reg().load_backend(best_path.c_str(), silent);
+}
+
+void ggml_backend_load_all() {
+    ggml_backend_load_best("blas", true);
+    ggml_backend_load_best("cann", true);
+    ggml_backend_load_best("cuda", true);
+    ggml_backend_load_best("hip", true);
+    ggml_backend_load_best("kompute", true);
+    ggml_backend_load_best("metal", true);
+    ggml_backend_load_best("rpc", true);
+    ggml_backend_load_best("sycl", true);
+    ggml_backend_load_best("vulkan", true);
+    ggml_backend_load_best("musa", true);
+    ggml_backend_load_best("cpu", true);
 }

ggml/src/ggml-backend.cpp

@@ -742,7 +742,8 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
+        ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+        GGML_ABORT("pre-allocated tensor (%s) in a buffer (%s) that cannot run the operation (%s)", tensor->name, ggml_backend_buffer_name(buffer), ggml_op_name(tensor->op));
     }
 
     // graph input

ggml/src/ggml-cann/CMakeLists.txt

@@ -22,13 +22,14 @@ if(NOT SOC_TYPE)
     detect_ascend_soc_type(SOC_VERSION)
     set(SOC_TYPE "${SOC_VERSION}")
     message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
-else()
-    string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 endif()
 
-# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND310P.
+string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
+
+# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND_310P.
 string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
 set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
+string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
 
 if (CANN_INSTALL_DIR)
     # Only Support Linux.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -21,22 +21,23 @@
  */
 
 #include "aclnn_ops.h"
-#include "ggml-impl.h"
 
+#include <aclnnop/aclnn_addcdiv.h>
 #include <aclnnop/aclnn_avgpool2d.h>
+#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_cast.h>
 #include <aclnnop/aclnn_constant_pad_nd.h>
 #include <aclnnop/aclnn_copy.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_div.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_group_norm.h>
 #include <aclnnop/aclnn_index_fill_tensor.h>
 #include <aclnnop/aclnn_layer_norm.h>
-#include <aclnnop/aclnn_mm.h>
-#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_matmul.h>
 #include <aclnnop/aclnn_max_pool.h>
+#include <aclnnop/aclnn_mm.h>
 #include <aclnnop/aclnn_permute.h>
 #include <aclnnop/aclnn_pow_tensor_tensor.h>
 #include <aclnnop/aclnn_reduce_sum.h>
@@ -56,6 +57,7 @@
 #include <exception>
 #include <vector>
 
+#include "ggml-impl.h"
 #include "kernels/ascendc_kernels.h"
 
 #define GGML_COMMON_DECL_C
@@ -1103,9 +1105,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
 }
 
 /**
- * @brief Creates an ACL tensor initialized with ones using a provided buffer.
+ * @brief Creates an ACL tensor initialized with value using a provided buffer.
  *
- * This function initializes a tensor with ones using the specified buffer and
+ * This function initializes a tensor with value using the specified buffer and
  * tensor parameters.
  *
  * @param ctx The context for the CANN backend operations.
@@ -1118,12 +1120,12 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
  * @param type_size The size of each element in the tensor data type.
  * @param value The value to be used for initializing the tensor (default
  * is 1.0).
- * @return An ACL tensor initialized with ones.
+ * @return An ACL tensor initialized with value.
  */
-static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer,
-                             size_t n_bytes, int64_t* ne, int64_t dims,
-                             aclDataType type, size_t type_size,
-                             float value = 1.0f) {
+static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
+                               size_t n_bytes, int64_t* ne, int64_t dims,
+                               aclDataType type, size_t type_size,
+                               float value = 1.0f) {
     aclTensor* acl_tensor =
         aclnn_zero(ctx, buffer, n_bytes, ne, dims, type, type_size);
     float alpha_host = 1.0f;
@@ -1165,7 +1167,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
-    aclTensor* acl_gamma = aclnn_ones(
+    aclTensor* acl_gamma = aclnn_values(
         ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
         ggml_cann_type_mapping(src->type), ggml_element_size(src));
 
@@ -1209,9 +1211,9 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
     aclTensor* mask_tensor =
-        aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne,
-                   GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src), value);
+        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
+                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
+                     ggml_element_size(src), value);
 
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -1768,6 +1770,92 @@ static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
     ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs element-wise division of tensor1 by tensor2 , multiplies the
+ result by the scalar value and adds it to self .
+ *
+ * Performs element-wise division of tensor1 by tensor2,
+ * multiplies the result by the scalar value and adds it to self .
+ * The operation is defined as:
+ * \f[
+ *     \text{out}_i = \text{selft}_i + \text{value} \times
+ \frac{\text{tensor1}_i}{\text{tensor2}_i}
+ * \f]
+
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_self The source tensor on which the addcdiv function will be
+ applied.
+ * @param tensor1 Numerator tensor.
+ * @param tensor2 Denominator tensor.
+ * @param value The value to be used for coefficient.
+ */
+static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx,
+                                  aclTensor* acl_self, aclTensor* tensor1,
+                                  aclTensor* tensor2, float value) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+    aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize(
+        acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+}
+
+/**
+ * @brief Matrix division, optionally in-place.
+ *
+ * This function division each element of the source tensor `acl_src` by the
+ * tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * If `inplace` is true, `acl_dst` will not be used and the operation is
+ * performed in-place on `acl_src`. The operation is defined as: \f[
+ *     \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src Numerator tensor..
+ * @param acl_other Denominator tensor.
+ * @param acl_dst The destination tensor where the result will be stored if
+ * `inplace` is false.
+ * @param inplace Flag indicating whether to perform the operation in-place on
+ * `acl_src`.
+ */
+static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                             aclTensor* acl_other, aclTensor* acl_dst,
+                             bool inplace) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    if (inplace) {
+        ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other,
+                                                  &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+    } else {
+        ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst,
+                                           &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(
+            aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    }
+}
+
 void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
                                   ggml_tensor* dst) {
     const ggml_tensor* src = dst->src[0];
@@ -2311,12 +2399,13 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                ctx.stream()));
 
     switch (src0->type) {
-        case GGML_TYPE_F32:
-        {
+        case GGML_TYPE_F32: {
 #ifdef ASCEND_310P
-             // Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            // Special operation for get_row_f32 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
             if ((src0->ne[0] % 8) != 0) {
-                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
+                                 src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
                 ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
             }
 #endif
@@ -2329,12 +2418,15 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->nb);
             break;
         }
-        case GGML_TYPE_F16:
-        {
+        case GGML_TYPE_F16: {
 #ifdef ASCEND_310P
-             // Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            // Special operation for get_row_f16 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
             if ((src0->ne[0] % 16) != 0) {
-                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16
+                size_t dst_len =
+                    src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
+                    ggml_type_size(
+                        GGML_TYPE_F32);  // out is also f32, even input is f16
                 ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
             }
 #endif
@@ -2459,8 +2551,9 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
  * @param acl_dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
-                             aclTensor* acl_weight, aclTensor* acl_dst) {
+static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
     int8_t cube_math_type = 2;
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -2475,8 +2568,7 @@ static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_inpu
         workspaceAddr = workspace_allocator.get();
     }
 
-    ACL_CHECK(
-        aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
 /**
@@ -2496,8 +2588,9 @@ static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_inpu
  * @param acl_dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
-                             aclTensor* acl_weight, aclTensor* acl_dst) {
+static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
     int8_t cube_math_type = 2;
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -2548,31 +2641,27 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
 
     aclTensor* acl_input_tensor =
         ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
-    int64_t transpose_ne[] = {
-        bcast_weight_ne[1], bcast_weight_ne[0],
-        bcast_weight_ne[2], bcast_weight_ne[3],
-        bcast_weight_ne[4], bcast_weight_ne[5]
-    };
-    size_t transpose_nb[] = {
-        bcast_weight_nb[1], bcast_weight_nb[0],
-        bcast_weight_nb[2], bcast_weight_nb[3],
-        bcast_weight_nb[4], bcast_weight_nb[5]
-    };
+    int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
+                              bcast_weight_ne[2], bcast_weight_ne[3],
+                              bcast_weight_ne[4], bcast_weight_ne[5]};
+    size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
+                             bcast_weight_nb[2], bcast_weight_nb[3],
+                             bcast_weight_nb[4], bcast_weight_nb[5]};
     aclTensor* acl_weight_tensor =
         ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
     aclTensor* acl_dst =
         ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
 
     switch (n_dims) {
-    case 2:
-        aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
-        break;
-    case 3:
-        aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
-        break;
-    default:
-        aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
-        break;
+        case 2:
+            aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        case 3:
+            aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        default:
+            aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
     }
 
     ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
@@ -2594,8 +2683,8 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * multiplication will be stored.
  */
 static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst,
-                                   const enum ggml_type type) {
+                                    ggml_tensor* dst,
+                                    const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
@@ -2617,14 +2706,15 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
     // scale stored at the end of weight. Also need transpose.
     size_t scale_elem_size = sizeof(uint16_t);
-    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size, scale_elem_size};
+    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
+                         scale_elem_size};
     size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
     char* scale_offset = (char*)src0->data + weight_size;
 
     // input
     size_t input_elem_size = sizeof(uint16_t);
     int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
-    size_t input_nb[] = {input_elem_size,  input_ne[0] * input_elem_size};
+    size_t input_nb[] = {input_elem_size, input_ne[0] * input_elem_size};
     size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size;
     ggml_cann_pool_alloc input_alloctor(ctx.pool());
     void* input_buffer = src1->data;
@@ -2632,7 +2722,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
     // case in
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        input_buffer = input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
+        input_buffer =
+            input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
 
         int64_t* input_cast_ne = src1->ne;
         size_t input_cast_nb[GGML_MAX_DIMS];
@@ -2642,9 +2733,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
         }
 
         aclTensor* acl_input_tensor = ggml_cann_create_tensor(
-            input_buffer,
-            ACL_FLOAT16,
-            input_elem_size, input_cast_ne, input_cast_nb, GGML_MAX_DIMS);
+            input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
+            input_cast_nb, GGML_MAX_DIMS);
         aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
 
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@@ -2655,7 +2745,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
     size_t output_elem_size = sizeof(uint16_t);
     size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
     ggml_cann_pool_alloc output_allocator(ctx.pool());
-    void* output_buffer = output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
+    void* output_buffer =
+        output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
     size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
 
     // aclnn
@@ -2679,7 +2770,9 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
             // first split
             int64_t weight_ne_offset = 0;
-            int64_t weight_ne[2] = {max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size, src0->ne[0]};
+            int64_t weight_ne[2] = {
+                max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size,
+                src0->ne[0]};
             int64_t scale_ne_offset = 0;
             int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
             int64_t output_ne_offset = 0;
@@ -2687,24 +2780,21 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 (char*)src0->data + batch0 * weight_stride,
-                ggml_cann_type_mapping(type),
-                weight_elem_size, weight_ne, weight_nb, 2,
-                ACL_FORMAT_ND, weight_ne_offset);
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
-                scale_offset + batch0 * scale_stride,
-                ACL_FLOAT16,
-                scale_elem_size, scale_ne, scale_nb, 2,
-                ACL_FORMAT_ND, scale_ne_offset);
+                scale_offset + batch0 * scale_stride, ACL_FLOAT16,
+                scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                scale_ne_offset);
             aclTensor* acl_output_tensor = ggml_cann_create_tensor(
-                (char*)output_buffer + batch1 * output_stride,
-                ACL_FLOAT16,
-                output_elem_size, output_ne, output_nb, 2,
-                ACL_FORMAT_ND, output_ne_offset);
+                (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
+                output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                output_ne_offset);
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
-                acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
-                nullptr, nullptr, nullptr, nullptr, QK8_0,
-                acl_output_tensor, &workspaceSize, &executor));
+                acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
+                nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
+                &workspaceSize, &executor));
             if (workspaceAddr == nullptr) {
                 workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
@@ -2717,28 +2807,29 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
             // other splits
             for (int64_t split = 1; split < split_size; split++) {
-                weight_ne_offset += weight_elem_size * weight_ne[0] * weight_ne[1];
-                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1] ? src0->ne[1] - (max_elem_size * split) : max_elem_size;
+                weight_ne_offset +=
+                    weight_elem_size * weight_ne[0] * weight_ne[1];
+                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1]
+                                   ? src0->ne[1] - (max_elem_size * split)
+                                   : max_elem_size;
                 scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
                 scale_ne[0] = weight_ne[0];
-                output_ne_offset += output_elem_size * output_ne[0] * output_ne[1];
+                output_ne_offset +=
+                    output_elem_size * output_ne[0] * output_ne[1];
                 output_ne[0] = weight_ne[0];
 
                 acl_weight_tensor = ggml_cann_create_tensor(
                     (char*)src0->data + batch0 * weight_stride,
-                    ggml_cann_type_mapping(type),
-                    weight_elem_size, weight_ne, weight_nb, 2,
-                    ACL_FORMAT_ND, weight_ne_offset);
+                    ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                    weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
                 acl_scale_tensor = ggml_cann_create_tensor(
-                    scale_offset + batch0 * scale_stride,
-                    ACL_FLOAT16,
-                    scale_elem_size, scale_ne, scale_nb, 2,
-                    ACL_FORMAT_ND, scale_ne_offset);
+                    scale_offset + batch0 * scale_stride, ACL_FLOAT16,
+                    scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                    scale_ne_offset);
                 acl_output_tensor = ggml_cann_create_tensor(
-                    (char*)output_buffer + batch1 * output_stride,
-                    ACL_FLOAT16,
-                    output_elem_size, output_ne, output_nb, 2,
-                    ACL_FORMAT_ND, output_ne_offset);
+                    (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
+                    output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                    output_ne_offset);
 
                 ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                     acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
@@ -2766,11 +2857,11 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
         }
 
         aclTensor* acl_output_tensor = ggml_cann_create_tensor(
-            output_buffer,
-            ACL_FLOAT16,
-            output_elem_size, output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
+            output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne,
+            output_cast_nb, GGML_MAX_DIMS);
         aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
+        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor,
+                   ggml_cann_type_mapping(dst->type));
 
         ACL_CHECK(aclDestroyTensor(acl_output_tensor));
         ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
@@ -2873,12 +2964,14 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                              aclTensor* acl_cos_repeat_tensor,
                              aclTensor* acl_sin_repeat_tensor,
-                             float theta_scale, bool is_neox) {
+                             float theta_scale, float freq_scale,
+                             float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
+    ggml_tensor* src2 = dst->src[2];  // freq_factors
 
     // arange, [0,1,...,ne0/2]
     int64_t arange_length = src0->ne[0] / 2;
@@ -2907,11 +3000,26 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
                                                arange_length * sizeof(float_t));
     void* theta_scale_buffer = theta_scale_allocator.get();
-    aclTensor* acl_theta_scale_tensor = aclnn_ones(
+    aclTensor* acl_theta_scale_tensor = aclnn_values(
         ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne,
         GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale);
     aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor);
 
+    // freq_scale
+    if (freq_scale != 1) {
+        aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
+    }
+
+    // freq_factors
+    if (src2) {
+        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+            src2->data, ggml_cann_type_mapping(src2->type),
+            ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS);
+        aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor,
+                         nullptr, true);
+        ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor));
+    }
+
     // position
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
@@ -2975,6 +3083,12 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor);
 
+    // attn_factor
+    if (attn_factor != 1) {
+        aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
+        aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
+    }
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -3038,19 +3152,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    // TODO: with freq_factors
-    GGML_ASSERT(src2 == NULL);
-    // TODO: attn_factor != 1
-    GGML_ASSERT(attn_factor == 1);
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
-    // TODO: freq_scale != 1
-    GGML_ASSERT(freq_scale == 1);
-    // TODO: type == GGML_TYPE_F16
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -3081,7 +3187,217 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                     theta_scale, is_neox);
+                     theta_scale, freq_scale, attn_factor, is_neox);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+#ifdef ASCEND_310P
+    // Special ROPE operation for 310P
+
+    // roll input
+    void* input_roll_buffer;
+    aclTensor* acl_minus_one_tensor;
+    void* minus_one_scale_buffer = nullptr;
+    ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
+    ggml_cann_pool_alloc minus_one_scale_allocator(
+        ctx.pool(), sizeof(float_t) * src0->ne[0]);
+    if (!is_neox) {
+        // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
+        input_roll_buffer = roll_allocator.get();
+        int64_t input_roll_ne[4] = {2, src0->ne[1] * (src0->ne[0] / 2),
+                                    src0->ne[2], src0->ne[3]};
+        size_t input_roll_nb[GGML_MAX_DIMS];
+        input_roll_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_roll_nb[i] = input_roll_nb[i - 1] * input_roll_ne[i - 1];
+        }
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
+            src0->data, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+
+        int64_t shifts[] = {1};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+
+        // init [-1, 1, -1, 1, ...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_values(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        int64_t dim = 3;
+        int64_t* index = new int64_t[src0->ne[0]];
+        for (int i = 0; i < src0->ne[0]; i++) {
+            index[i] = i / 2 * 2;
+        }
+        int64_t index_num = src0->ne[0];
+        float value = -1;
+        aclnn_index_fill_tensor(ctx, acl_minus_one_tensor, dim, index,
+                                index_num, value);
+    } else {
+        // roll input: [q0,q1,q2,...] ->
+        // [q_half,q_half+1,...,q_end,q0,q1,...q_half-1]
+        input_roll_buffer = roll_allocator.get();
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(src0);
+
+        int64_t shifts[] = {src0->ne[0] / 2};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+        // init [-1, -1, -1, 1, 1，1，...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_values(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        // -1 * first half
+        int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
+        size_t first_half_nb[GGML_MAX_DIMS];
+        first_half_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
+        }
+        aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
+            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
+            first_half_nb, GGML_MAX_DIMS);
+        bool inplace = true;
+        float scale = -1;
+        aclnn_muls(ctx, acl_first_half_tensor, scale, nullptr, inplace);
+        ACL_CHECK(aclDestroyTensor(acl_first_half_tensor));
+    }
+
+    // TODO: n_dims < ne0
+    GGML_ASSERT(n_dims == src0->ne[0]);
+
+    // input * scale
+    ggml_cann_pool_alloc roll_mul_scale_allocator(ctx.pool(),
+                                                  ggml_nbytes(src0));
+    void* input_roll_mul_scale_buffer = roll_mul_scale_allocator.get();
+    size_t input_nb[GGML_MAX_DIMS];
+    input_nb[0] = ggml_type_size(src0->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        input_nb[i] = input_nb[i - 1] * src0->ne[i - 1];
+    }
+    aclTensor* acl_input_roll_mul_scale_tensor = ggml_cann_create_tensor(
+        input_roll_mul_scale_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+    aclTensor* acl_input_roll_reshape_tensor = ggml_cann_create_tensor(
+        input_roll_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+
+    aclnn_mul(ctx, acl_input_roll_reshape_tensor, acl_minus_one_tensor,
+              acl_input_roll_mul_scale_tensor);
+
+    // output
+    void* output_fp32_buffer;
+    if (src0->type == GGML_TYPE_F32) {
+        aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor);
+        aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
+                          acl_sin_reshape_tensor);
+        aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst);
+        // TODO: ne0 != n_dims in mode2
+    } else if (src0->type == GGML_TYPE_F16) {
+        size_t input_fp32_nb[GGML_MAX_DIMS];
+        input_fp32_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
+        }
+        ggml_cann_pool_alloc fp32_allocator1(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer1 = fp32_allocator1.get();
+        aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
+            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        ggml_cann_pool_alloc fp32_allocator2(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer2 = fp32_allocator2.get();
+        aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
+            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+
+        ggml_cann_pool_alloc fp32_allocator(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        output_fp32_buffer = fp32_allocator.get();
+        aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
+            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1);
+        aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
+                  input_fp32_tensor2);
+        aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
+                  output_fp32_tensor);
+        aclnn_cast(ctx, output_fp32_tensor, acl_dst, ACL_FLOAT16);
+
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
+        ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_src));
+    }
+    return;
+#endif
+
+    // src0 == GGML_TYPE_F16
+    // TODO: optimization this `if` code
+    if (src0->type == GGML_TYPE_F16) {
+        ggml_cann_pool_alloc sin_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        ggml_cann_pool_alloc cos_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        void* sin_final_buffer = sin_final_allocator.get();
+        void* cos_final_buffer = cos_final_allocator.get();
+
+        int64_t sin_final_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
+        size_t sin_final_nb[GGML_MAX_DIMS];
+        sin_final_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            sin_final_nb[i] = sin_final_nb[i - 1] * sin_final_ne[i - 1];
+        }
+        aclTensor* acl_sin_final_tensor = ggml_cann_create_tensor(
+            sin_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_cos_final_tensor = ggml_cann_create_tensor(
+            cos_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+
+        aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        acl_sin_reshape_tensor = acl_sin_final_tensor;
+        acl_cos_reshape_tensor = acl_cos_final_tensor;
+    }
 
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -3093,10 +3409,9 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         acl_mode = 1;
     }
 
-    aclTensor* acl_x = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
-        acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, acl_dst, &workspaceSize, &executor));
+        acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
+        acl_dst, &workspaceSize, &executor));
     if (workspaceSize > 0) {
         ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
         workspaceAddr = workspace_allocator.get();
@@ -3105,7 +3420,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
                                            executor, ctx.stream()));
 
-    ACL_CHECK(aclDestroyTensor(acl_x));
+    ACL_CHECK(aclDestroyTensor(acl_src));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));

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

@@ -1738,13 +1738,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
-            float * freq_scale = (float*)((int32_t*)op->op_params + 6);
             float * ext_factor = (float*)((int32_t*)op->op_params + 7);
-            float * attn_factor = (float*)((int32_t*)op->op_params + 8);
-            // TODO: with freq_factors
-            if (op->src[2] != NULL) {
-                return false;
-            }
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
@@ -1753,21 +1747,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (*ext_factor != 0) {
                 return false;
             }
-            // TODO: freq_scale != 1
-            if (*freq_scale != 1) {
-                return false;
-            }
-            // TODO: attn_factor != 1
-            if (*attn_factor != 1) {
-                return false;
-            }
-            //TODO: type == GGML_TYPE_F16
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                    return true;
-                default:
-                    return false;
-            }
+            return true;
         }
         case GGML_OP_UPSCALE: {
             // aclnnUpsampleNearest2dGetWorkspaceSize not support

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

@@ -25,6 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-message(STATUS "CANN: compile ascend kernels witch SOC_VERSION:${SOC_VERSION}.")
+message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
 ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

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

@@ -20,7 +20,6 @@ class DupByRows {
         // Input has four dims.
         int64_t op_block_num = GetBlockNum();
         int64_t op_block_idx = GetBlockIdx();
-        assert(op_block_idx < SUPPORTED_MAX_DIM && op_block_idx >= 0, "Invalid block index:%d, max is:%d\n", op_block_idx, SUPPORTED_MAX_DIM);
 
         // param
         num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];

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

@@ -2,6 +2,15 @@
 
 // optimize me. Use template to avoid copy code.
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support 4bit get row
+    extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
+        GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+        GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
+        GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support 4bit get row.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 
@@ -110,12 +119,9 @@ class GET_ROW_Q4_0 {
         LocalTensor<float> output_local = output_queue.AllocTensor<float>();
 
         // TODO: cast more data to speed up.
-#ifdef ASCEND_310P
-        // TODO: 310P support quantification
-#else
         Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
         Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
-#endif
+
         // Only mul need compile by group.
         half scale = scale_gm.GetValue(scale_offset);
 
@@ -194,3 +200,5 @@ extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
             indices_nb_ub, output_ne_ub, output_nb_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

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

@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -206,3 +214,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

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

@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support f32->8bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -204,3 +212,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

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

@@ -1,6 +1,21 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support float->4bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->4bit quantization.\n");
+    }
+
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->4bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define Group_Size 32
@@ -276,3 +291,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

ggml/src/ggml-common.h

@@ -418,6 +418,12 @@ typedef struct {
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
 
+typedef struct {
+    ggml_half d[4];        // deltas for 4 iq4_nl blocks
+    uint8_t qs[QK4_NL * 2];// nibbles / quants for 4 iq4_nl blocks
+} block_iq4_nlx4;
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,12 +1,20 @@
-ggml_add_backend_library(ggml-cpu
-                         ggml-cpu.c
-                         ggml-cpu.cpp
-                         ggml-cpu-aarch64.c
-                         ggml-cpu-aarch64.h
-                         ggml-cpu-quants.c
-                         ggml-cpu-quants.h
-                        )
+ggml_add_backend_library(ggml-cpu)
 
+list (APPEND GGML_CPU_SOURCES
+    ggml-cpu.c
+    ggml-cpu.cpp
+    ggml-cpu-aarch64.c
+    ggml-cpu-aarch64.h
+    ggml-cpu-quants.c
+    ggml-cpu-quants.h
+    amx/amx.cpp
+    amx/amx.h
+    amx/mmq.cpp
+    amx/mmq.h
+    ggml-cpu-impl.h
+    )
+
+target_compile_features(ggml-cpu PRIVATE c_std_11 cxx_std_17)
 target_include_directories(ggml-cpu PRIVATE .)
 
 if (APPLE AND GGML_ACCELERATE)
@@ -14,9 +22,9 @@ if (APPLE AND GGML_ACCELERATE)
     if (ACCELERATE_FRAMEWORK)
         message(STATUS "Accelerate framework found")
 
-        add_compile_definitions(GGML_USE_ACCELERATE)
-        add_compile_definitions(ACCELERATE_NEW_LAPACK)
-        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+        target_compile_definitions(ggml-cpu PRIVATE GGML_USE_ACCELERATE)
+        target_compile_definitions(ggml-cpu PRIVATE ACCELERATE_NEW_LAPACK)
+        target_compile_definitions(ggml-cpu PRIVATE ACCELERATE_LAPACK_ILP64)
 
         target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
     else()
@@ -29,15 +37,9 @@ if (GGML_OPENMP)
     if (OpenMP_FOUND)
         message(STATUS "OpenMP found")
 
-        add_compile_definitions(GGML_USE_OPENMP)
+        target_compile_definitions(ggml-cpu PRIVATE GGML_USE_OPENMP)
 
         target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
-
-        # FIXME: should be replaced with a compiler id check
-        #if (GGML_MUSA)
-        #    list(APPEND GGML_CPU_EXTRA_INCLUDES     "/usr/lib/llvm-14/lib/clang/14.0.0/include")
-        #    list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
-        #endif()
     else()
         message(WARNING "OpenMP not found")
     endif()
@@ -46,11 +48,11 @@ endif()
 if (GGML_LLAMAFILE)
     message(STATUS "Using llamafile")
 
-    add_compile_definitions(GGML_USE_LLAMAFILE)
+    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_LLAMAFILE)
 
-    target_sources(ggml-cpu PRIVATE
-                    llamafile/sgemm.cpp
-                    llamafile/sgemm.h)
+    list(APPEND GGML_CPU_SOURCES
+                llamafile/sgemm.cpp
+                llamafile/sgemm.h)
 endif()
 
 if (GGML_CPU_HBM)
@@ -58,7 +60,7 @@ if (GGML_CPU_HBM)
 
     message(STATUS "Using memkind for CPU HBM")
 
-    add_compile_definitions(GGML_USE_CPU_HBM)
+    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_CPU_HBM)
 
     target_link_libraries(ggml-cpu PUBLIC memkind)
 endif()
@@ -72,27 +74,33 @@ if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
     message(STATUS "ARM detected")
 
     if (MSVC)
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
-        add_compile_definitions(__ARM_NEON)
-        add_compile_definitions(__ARM_FEATURE_FMA)
+        list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
+        list(APPEND ARCH_DEFINITIONS __ARM_NEON)
+        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
 
         set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
         string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
 
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
         if (GGML_COMPILER_SUPPORT_DOTPROD)
-            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+
+            message(STATUS "ARM feature DOTPROD enabled")
         endif ()
 
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
 
         if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+
+            message(STATUS "ARM feature MATMUL_INT8 enabled")
         endif ()
 
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
         if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+
+            message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
         endif ()
 
         set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
@@ -112,18 +120,24 @@ if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
                 check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
                 if (GGML_COMPILER_SUPPORT_DOTPROD)
                     set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
-                    add_compile_definitions(__ARM_FEATURE_DOTPROD)
+                    list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+
+                    message(STATUS "ARM feature DOTPROD enabled")
                 endif ()
 
                 set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
 
                 set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
+                set(CMAKE_REQUIRED_FLAGS     "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
+
                 check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
                 if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
                     set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
-                    add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+                    list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+
+                    message(STATUS "ARM feature MATMUL_INT8 enabled")
                 endif ()
+
                 set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
 
                 list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
@@ -163,7 +177,6 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
     if (MSVC)
         # instruction set detection for MSVC only
         if (GGML_NATIVE)
-            # TODO: improve, should not reference files from the parent folder
             include(cmake/FindSIMD.cmake)
         endif ()
         if (GGML_AVX512)
@@ -173,43 +186,43 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
             # macros corresponding to the extensions.
             # Do it manually.
             if (GGML_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+                list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
                 if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
                     list(APPEND ARCH_FLAGS -mavx512vbmi)
                 endif()
             endif()
             if (GGML_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+                list(APPEND ARCH_DEFINITIONS __AVX512VNNI__)
                 if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
                     list(APPEND ARCH_FLAGS -mavx512vnni)
                 endif()
             endif()
             if (GGML_AVX512_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+                list(APPEND ARCH_DEFINITIONS __AVX512BF16__)
                 if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
                     list(APPEND ARCH_FLAGS -mavx512bf16)
                 endif()
             endif()
             if (GGML_AMX_TILE)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+                list(APPEND ARCH_DEFINITIONS __AMX_TILE__)
             endif()
             if (GGML_AMX_INT8)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+                list(APPEND ARCH_DEFINITIONS __AMX_INT8__)
             endif()
             if (GGML_AMX_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+                list(APPEND ARCH_DEFINITIONS __AMX_BF16__)
             endif()
         elseif (GGML_AVX2)
             list(APPEND ARCH_FLAGS /arch:AVX2)
         elseif (GGML_AVX)
             list(APPEND ARCH_FLAGS /arch:AVX)
         endif()
+        if (GGML_AVX_VNNI)
+            list(APPEND ARCH_DEFINITIONS __AVXVNNI__)
+            if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                list(APPEND ARCH_FLAGS -mavxvnni)
+            endif()
+        endif()
     else()
         if (GGML_NATIVE)
             list(APPEND ARCH_FLAGS -march=native)
@@ -226,6 +239,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         if (GGML_AVX2)
             list(APPEND ARCH_FLAGS -mavx2)
         endif()
+        if (GGML_AVX_VNNI)
+            list(APPEND ARCH_FLAGS -mavxvnni)
+        endif()
         if (GGML_AVX512)
             list(APPEND ARCH_FLAGS -mavx512f)
             list(APPEND ARCH_FLAGS -mavx512dq)
@@ -264,7 +280,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
     else()
         list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+        # TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
     message(STATUS "loongarch64 detected")
@@ -287,11 +303,16 @@ endif()
 
 if (GGML_CPU_AARCH64)
     message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
-    add_compile_definitions(GGML_USE_CPU_AARCH64)
+    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_CPU_AARCH64)
 endif()
 
-target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
-target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+target_sources(ggml-cpu PRIVATE ${GGML_CPU_SOURCES})
+set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_OPTIONS     "${ARCH_FLAGS}")
+set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_DEFINITIONS "${ARCH_DEFINITIONS}")
+
+# the feature detection code must be compiled without any architecture flags
+target_sources(ggml-cpu PRIVATE cpu-feats-x86.cpp)
+# target_sources(ggml-cpu PRIVATE cpu-feats-arm.cpp) # TODO: ARM feature detection
 
 if (EMSCRIPTEN)
     set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")

ggml/src/ggml-cpu/amx/amx.cpp

@@ -0,0 +1,196 @@
+#include "amx.h"
+#include "common.h"
+#include "mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+
+// AMX buffer interface
+static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)(buffer->context);
+}
+
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *)tensor->data + offset, data, size);
+    }
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_nbytes(dst));
+        } else {
+            memcpy(dst->data, src->data, ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_amx_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    return ggml_backend_amx_get_alloc_size(tensor);
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    if (!ggml_amx_init()) {
+        return NULL;
+    }
+
+    return &ggml_backend_buffer_type_amx;
+}
+
+bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
+}
+
+bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op) {
+    // handle only 2d gemm for now
+    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+    };
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT: {
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const enum ggml_type type = src0->type;
+            const int64_t ne0 = op->ne[0];
+
+            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
+            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
+            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
+
+            bool can_use_amx =
+                is_contiguous_2d(src0) &&       // src0 must be contiguous
+                is_contiguous_2d(src1) &&       // src1 must be contiguous
+                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
+                has_amx_kernels &&              // with amx kernel impls
+                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
+
+            return can_use_amx;
+        }
+        default:
+            return false;
+    }
+}
+
+#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)

ggml/src/ggml-cpu/amx/amx.h

@@ -0,0 +1,20 @@
+#include "ggml-backend.h"
+#include "ggml-cpu-impl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft);
+bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op);
+void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
+
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/amx/common.h

@@ -1,8 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-// hack until AMX is moved into the CPU backend
-#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
+#include "ggml-cpu-impl.h"
 
 #include <algorithm>
 #include <memory>
@@ -74,16 +73,23 @@ inline void parallel_for(int nth, int n, const func_t& f) {
 #endif
 }
 
+template <typename func_t>
+inline void parallel_for_ggml(const ggml_compute_params * params, int n, const func_t & f) {
+    int tbegin, tend;
+    balance211(n, params->nth, params->ith, tbegin, tend);
+    f(tbegin, tend);
+}
+
 // quantized types that have AMX support
 inline bool qtype_has_amx_kernels(const enum ggml_type type) {
     // TODO: fix padding for vnni format
     return (type == GGML_TYPE_Q4_0) ||
-        (type == GGML_TYPE_Q4_1);
-        //(type == GGML_TYPE_Q8_0) ||
-        //(type == GGML_TYPE_Q4_K) ||
-        //(type == GGML_TYPE_Q5_K) ||
-        //(type == GGML_TYPE_Q6_K) ||
-        //(type == GGML_TYPE_IQ4_XS);
+        (type == GGML_TYPE_Q4_1) ||
+        (type == GGML_TYPE_Q8_0) ||
+        (type == GGML_TYPE_Q4_K) ||
+        (type == GGML_TYPE_Q5_K) ||
+        (type == GGML_TYPE_Q6_K) ||
+        (type == GGML_TYPE_IQ4_XS);
 }
 
 // ggml backend context

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -4,8 +4,11 @@
 #pragma GCC diagnostic ignored "-Wunused-local-typedefs"
 #endif
 
+#include "amx.h"
 #include "mmq.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-cpu-quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>
@@ -33,7 +36,7 @@
 #define ALWAYS_INLINE inline
 #endif
 
-#if defined(__AMX_INT8__)
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
 
 namespace {
 
@@ -496,13 +499,12 @@ inline void from_float(const float * x, char * vy, int64_t k);
 
 template <>
 inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
-    // FIXME: using unoptimized reference impl until moved to CPU backend
-    quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
+    quantize_row_q8_0(x, (block_q8_0 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
+    quantize_row_q8_1(x, (block_q8_1 *)vy, k);
 }
 
 template <>
@@ -950,7 +952,7 @@ template<typename TB, typename packed_B_t = packed_B_type<TB>>
 void unpack_B(packed_B_t * RESTRICT tile, const void * RESTRICT packed_B) {
     GGML_UNUSED(tile);
     GGML_UNUSED(packed_B);
-};
+}
 
 template <>
 void unpack_B<block_q4_0>(int8_t * RESTRICT tile, const void * RESTRICT packed_B) {
@@ -1338,21 +1340,19 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
         __m512 vb[COLS];
         __m512 vc[ROWS * COLS];
 
-        auto loadc = [&](int idx) {
+        auto loadc = [&](auto idx) {
             vc[idx] = _mm512_setzero_ps();
         };
         Unroll<ROWS * COLS>{}(loadc);
 
-        auto compute = [&](int idx, int k) {
-            // TODO: use `constexpr` here to get rid of interger div
-            // when upgraded to C++17
-            const int row = idx / COLS;
-            const int col = idx % COLS;
+        auto compute = [&](auto idx, auto k) {
+            constexpr int row = idx / COLS;
+            constexpr int col = idx % COLS;
 
-            if (col == 0) {
+            if constexpr (col == 0) {
                 va = _mm512_loadu_ps(A + row * K + k);
             }
-            if (row == 0) {
+            if constexpr (row == 0) {
                 vb[col] =  _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k)));
             }
             vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]);
@@ -1362,9 +1362,9 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
             Unroll<ROWS * COLS>{}(compute, k);
         }
 
-        auto storec = [&](int idx) {
-            const int row = idx / COLS;
-            const int col = idx % COLS;
+        auto storec = [&](auto idx) {
+            constexpr int row = idx / COLS;
+            constexpr int col = idx % COLS;
             C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]);
         };
         Unroll<ROWS * COLS>{}(storec);
@@ -1427,14 +1427,14 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         const __m512i off = _mm512_set1_epi8(8);
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a and compute compensation
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 vcomp = _mm512_setzero_si512();
                 for (int k = 0; k < 8; ++k) {
@@ -1466,7 +1466,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1490,14 +1490,14 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
@@ -1531,7 +1531,7 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1562,14 +1562,14 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
         //
         const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a and add offset 128
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
@@ -1602,7 +1602,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1634,7 +1634,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
@@ -1648,9 +1648,9 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
         //     int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8
         //     from {16,  8} to {4, 32}
         //
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
                     va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
                 }
@@ -1702,7 +1702,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1735,15 +1735,15 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLO
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
         // Q5_K and Q4_K shares the same vnni formats, refer to notes above.
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
                     va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
                 }
@@ -1808,7 +1808,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1841,13 +1841,13 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q6_K, float, BLOCK_M, BLOCK_N, BLO
         const __m512i m32s = _mm512_set1_epi32(32);
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
-            if (col == 0) {
+        auto compute = [&](auto col, auto i) {
+            if constexpr (col == 0) {
                 // load a
                 va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
                 va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
@@ -1959,13 +1959,13 @@ struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, B
         const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
         const __m512i values256 = _mm512_add_epi8(values128, off);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
-            if (col == 0) {
+        auto compute = [&](auto col, auto i) {
+            if constexpr (col == 0) {
                 // load a
                 va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
                 va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
@@ -2015,7 +2015,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, B
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -2327,9 +2327,7 @@ size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) {
 
 // pack weight to vnni format
 void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-
-    size_t alloc_size = ggml_backend_amx_get_alloc_size(tensor);
-    GGML_ASSERT(alloc_size == size);
+    GGML_ASSERT(offset == 0 && size == ggml_nbytes(tensor)); // only full tensor conversion is supported for now
 
     const enum ggml_type TYPE = tensor->type;
 
@@ -2348,6 +2346,29 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d
     });
 }
 
+size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
+    struct ggml_tensor * src0 = dst->src[0];
+
+    const enum ggml_type TYPE = src0->type;
+
+    const bool is_floating_type = TYPE == GGML_TYPE_F16;
+    if (is_floating_type) {
+        return 0;
+    }
+
+    const int M = dst->ne[1];
+    const int K = src0->ne[0];
+
+    size_t desired_wsize = 0;
+
+    GGML_DISPATCH_QTYPES(TYPE, [&] {
+        const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
+        desired_wsize = M * row_size_A;
+    });
+
+    return desired_wsize;
+}
+
 // NB: mixed dtype gemm with Advanced Matrix Extensions (Intel AMX)
 //
 // src0: weight in shape of {N, K}, quantized
@@ -2356,14 +2377,12 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d
 //
 // the function performs: dst = src1 @ src0.T
 //
-void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst) {
+void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_tensor * dst) {
     struct ggml_tensor * src0 = dst->src[0];
     struct ggml_tensor * src1 = dst->src[1];
 
     const enum ggml_type TYPE = src0->type;
 
-    const int n_threads = ctx->n_threads;
-
     // f16 only has avx512 kernels for now,
     // amx kernels will be added once 6th gen xeon is released.
     const bool is_floating_type = TYPE == GGML_TYPE_F16;
@@ -2379,7 +2398,7 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
         const int MB = div_up(M, BLOCK_M);
         const int NB = div_up(N, BLOCK_N);
 
-        parallel_for(n_threads, MB * NB, [&](int begin, int end) {
+        parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
             GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] {
                 for (int i = begin; i < end; ++i) {
                     int mb = i / NB;
@@ -2412,27 +2431,29 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
     }
 
     // pointer to work space, used convert A from float to quantized type
-    void * wdata = nullptr;
+    void * wdata = params->wdata;
 
     //TODO: performance improvement: merge quant A
-    GGML_DISPATCH_QTYPES(TYPE, [&] {
-        const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
-        const size_t desired_wsize = M * row_size_A;
-        if (ctx->work_size < desired_wsize) {
-            ctx->work_data.reset(new char[desired_wsize]);
-            ctx->work_size = desired_wsize;
-        }
-        wdata = ctx->work_data.get();
+    if (params->ith == 0) {
+        GGML_DISPATCH_QTYPES(TYPE, [&] {
+            const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
+            const size_t desired_wsize = M * row_size_A;
+            if (params->wsize < desired_wsize) {
+                GGML_ABORT("insufficient work space size");
+            }
 
-        // Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size
-        // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
-        GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
+            // Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size
+            // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
+            GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
 
-        const float * A_data = static_cast<const float *>(src1->data);
-        for (int m = 0; m < M; ++m) {
-            from_float<vec_dot_type>(A_data + m * K, (char *)wdata + m * row_size_A, K);
-        }
-    });
+            const float * A_data = static_cast<const float *>(src1->data);
+            for (int m = 0; m < M; ++m) {
+                from_float<vec_dot_type>(A_data + m * K, (char *)wdata + m * row_size_A, K);
+            }
+        });
+    }
+
+    ggml_barrier(params->threadpool);
 
     if (M == 1) {
         // MB = 1 and handle 8 tiles in each block
@@ -2440,7 +2461,7 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
         constexpr int BLOCK_N = TILE_N * kTilesN;
         const int NB = div_up(N, BLOCK_N);
 
-        parallel_for(n_threads, NB, [&](int begin, int end) {
+        parallel_for_ggml(params, NB, [&](int begin, int end) {
             GGML_DISPATCH_QTYPES(TYPE, [&] {
                 const int KB = K / blck_size;
                 const int TILE_SIZE = get_tile_size<type>();
@@ -2470,7 +2491,7 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
     const int MB = div_up(M, BLOCK_M);
     const int NB = div_up(N, BLOCK_N);
 
-    parallel_for(n_threads, MB * NB, [&](int begin, int end) {
+    parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
         // init tile config for each thread
         ggml_tile_config_init();
 
@@ -2498,13 +2519,4 @@ void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor
     });
 }
 
-#else // if defined(__AMX_INT8__)
-
-void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst) {
-    fprintf(stderr, "GGML is not compiled with AMX support!\n");
-
-    GGML_UNUSED(ctx);
-    GGML_UNUSED(dst);
-}
-
-#endif // if defined(__AMX_INT8__)
+#endif // if defined(__AMX_INT8__) && defined(__AVX512VNNI__)

ggml/src/ggml-cpu/amx/mmq.h

@@ -1,6 +1,5 @@
 #pragma once
 #include "common.h"
-#include <stdint.h>
 
 #ifdef __cplusplus
 extern "C" {
@@ -10,7 +9,7 @@ size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
 
 void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 
-void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst);
+void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -0,0 +1,298 @@
+#include "ggml-cpu.h"
+#include "ggml-backend-impl.h"
+
+#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include <cstring>
+#include <vector>
+#include <bitset>
+#include <array>
+#include <string>
+
+struct cpuid_x86 {
+    bool SSE3(void) { return f_1_ecx[0]; }
+    bool PCLMULQDQ(void) { return f_1_ecx[1]; }
+    bool MONITOR(void) { return f_1_ecx[3]; }
+    bool SSSE3(void) { return f_1_ecx[9]; }
+    bool FMA(void) { return f_1_ecx[12]; }
+    bool CMPXCHG16B(void) { return f_1_ecx[13]; }
+    bool SSE41(void) { return f_1_ecx[19]; }
+    bool SSE42(void) { return f_1_ecx[20]; }
+    bool MOVBE(void) { return f_1_ecx[22]; }
+    bool POPCNT(void) { return f_1_ecx[23]; }
+    bool AES(void) { return f_1_ecx[25]; }
+    bool XSAVE(void) { return f_1_ecx[26]; }
+    bool OSXSAVE(void) { return f_1_ecx[27]; }
+    bool AVX(void) { return f_1_ecx[28]; }
+    bool F16C(void) { return f_1_ecx[29]; }
+    bool RDRAND(void) { return f_1_ecx[30]; }
+
+    bool MSR(void) { return f_1_edx[5]; }
+    bool CX8(void) { return f_1_edx[8]; }
+    bool SEP(void) { return f_1_edx[11]; }
+    bool CMOV(void) { return f_1_edx[15]; }
+    bool CLFSH(void) { return f_1_edx[19]; }
+    bool MMX(void) { return f_1_edx[23]; }
+    bool FXSR(void) { return f_1_edx[24]; }
+    bool SSE(void) { return f_1_edx[25]; }
+    bool SSE2(void) { return f_1_edx[26]; }
+
+    bool FSGSBASE(void) { return f_7_ebx[0]; }
+    bool BMI1(void) { return f_7_ebx[3]; }
+    bool HLE(void) { return is_intel && f_7_ebx[4]; }
+    bool AVX2(void) { return f_7_ebx[5]; }
+    bool BMI2(void) { return f_7_ebx[8]; }
+    bool ERMS(void) { return f_7_ebx[9]; }
+    bool INVPCID(void) { return f_7_ebx[10]; }
+    bool RTM(void) { return is_intel && f_7_ebx[11]; }
+    bool AVX512F(void) { return f_7_ebx[16]; }
+    bool RDSEED(void) { return f_7_ebx[18]; }
+    bool ADX(void) { return f_7_ebx[19]; }
+    bool AVX512PF(void) { return f_7_ebx[26]; }
+    bool AVX512ER(void) { return f_7_ebx[27]; }
+    bool AVX512CD(void) { return f_7_ebx[28]; }
+    bool SHA(void) { return f_7_ebx[29]; }
+
+    bool PREFETCHWT1(void) { return f_7_ecx[0]; }
+
+    bool LAHF(void) { return f_81_ecx[0]; }
+    bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
+    bool ABM(void) { return is_amd && f_81_ecx[5]; }
+    bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
+    bool XOP(void) { return is_amd && f_81_ecx[11]; }
+    bool TBM(void) { return is_amd && f_81_ecx[21]; }
+
+    bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
+    bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
+    bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
+    bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
+    bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
+
+    bool AVX512_VBMI(void) { return f_7_ecx[1]; }
+    bool AVX512_VNNI(void) { return f_7_ecx[11]; }
+    bool AVX512_FP16(void) { return f_7_edx[23]; }
+    bool AVX512_BF16(void) { return f_7_1_eax[5]; }
+    bool AVX_VNNI(void) { return f_7_1_eax[4]; }
+
+    bool AMX_TILE(void) { return f_7_edx[24]; }
+    bool AMX_INT8(void) { return f_7_edx[25]; }
+    bool AMX_FP16(void) { return f_7_1_eax[21]; }
+    bool AMX_BF16(void) { return f_7_edx[22]; }
+
+#ifdef _MSC_VER
+    static void cpuid(int cpu_info[4], int eax) {
+        __cpuid(cpu_info, eax);
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __cpuidex(cpu_info, eax, ecx);
+    }
+#else
+    static void cpuid(int cpu_info[4], int eax) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(0));
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(ecx));
+    }
+#endif
+
+    cpuid_x86() {
+        std::array<int, 4> cpui;
+        std::vector<std::array<int, 4>> data;
+
+        // calling __cpuid with 0x0 as the function_id argument
+        // gets the number of the highest valid function ID.
+        cpuid(cpui.data(), 0);
+        int n_ids = cpui[0];
+
+        for (int i = 0; i <= n_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            data.push_back(cpui);
+        }
+
+        // capture vendor string
+        char vendor[0x20] = {};
+        *reinterpret_cast<int *>(vendor)     = data[0][1];
+        *reinterpret_cast<int *>(vendor + 4) = data[0][3];
+        *reinterpret_cast<int *>(vendor + 8) = data[0][2];
+        this->vendor = vendor;
+        if (this->vendor == "GenuineIntel") {
+            is_intel = true;
+        } else if (this->vendor == "AuthenticAMD") {
+            is_amd = true;
+        }
+
+        // load bitset with flags for function 0x00000001
+        if (n_ids >= 1) {
+            f_1_ecx = data[1][2];
+            f_1_edx = data[1][3];
+        }
+
+        // load bitset with flags for function 0x00000007
+        if (n_ids >= 7) {
+            f_7_ebx = data[7][1];
+            f_7_ecx = data[7][2];
+            f_7_edx = data[7][3];
+            cpuidex(cpui.data(), 7, 1);
+            f_7_1_eax = cpui[0];
+        }
+
+        // calling __cpuid with 0x80000000 as the function_id argument
+        // gets the number of the highest valid extended ID.
+        cpuid(cpui.data(), 0x80000000);
+        unsigned int n_ex_ids = cpui[0];
+
+        std::vector<std::array<int, 4>> ext_data;
+        for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            ext_data.push_back(cpui);
+        }
+
+        // load bitset with flags for function 0x80000001
+        if (n_ex_ids >= 0x80000001) {
+            f_81_ecx = ext_data[1][2];
+            f_81_edx = ext_data[1][3];
+        }
+
+        // interpret CPU brand string if reported
+        char brand[0x40] = {};
+        if (n_ex_ids >= 0x80000004) {
+            std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
+            std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
+            std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
+            this->brand = brand;
+        }
+    }
+
+    bool is_intel = false;
+    bool is_amd = false;
+    std::string vendor;
+    std::string brand;
+    std::bitset<32> f_1_ecx;
+    std::bitset<32> f_1_edx;
+    std::bitset<32> f_7_ebx;
+    std::bitset<32> f_7_ecx;
+    std::bitset<32> f_7_edx;
+    std::bitset<32> f_7_1_eax;
+    std::bitset<32> f_81_ecx;
+    std::bitset<32> f_81_edx;
+};
+
+#if 0
+void test_x86_is() {
+    cpuid_x86 is;
+    printf("CPU Vendor: %s\n", is.vendor.c_str());
+    printf("Brand: %s\n", is.brand.c_str());
+    printf("is_intel: %d\n", is.is_intel);
+    printf("is_amd: %d\n", is.is_amd);
+    printf("sse3: %d\n", is.SSE3());
+    printf("pclmulqdq: %d\n", is.PCLMULQDQ());
+    printf("ssse3: %d\n", is.SSSE3());
+    printf("fma: %d\n", is.FMA());
+    printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
+    printf("sse41: %d\n", is.SSE41());
+    printf("sse42: %d\n", is.SSE42());
+    printf("movbe: %d\n", is.MOVBE());
+    printf("popcnt: %d\n", is.POPCNT());
+    printf("aes: %d\n", is.AES());
+    printf("xsave: %d\n", is.XSAVE());
+    printf("osxsave: %d\n", is.OSXSAVE());
+    printf("avx: %d\n", is.AVX());
+    printf("f16c: %d\n", is.F16C());
+    printf("rdrand: %d\n", is.RDRAND());
+    printf("msr: %d\n", is.MSR());
+    printf("cx8: %d\n", is.CX8());
+    printf("sep: %d\n", is.SEP());
+    printf("cmov: %d\n", is.CMOV());
+    printf("clflush: %d\n", is.CLFSH());
+    printf("mmx: %d\n", is.MMX());
+    printf("fxsr: %d\n", is.FXSR());
+    printf("sse: %d\n", is.SSE());
+    printf("sse2: %d\n", is.SSE2());
+    printf("fsgsbase: %d\n", is.FSGSBASE());
+    printf("bmi1: %d\n", is.BMI1());
+    printf("hle: %d\n", is.HLE());
+    printf("avx2: %d\n", is.AVX2());
+    printf("bmi2: %d\n", is.BMI2());
+    printf("erms: %d\n", is.ERMS());
+    printf("invpcid: %d\n", is.INVPCID());
+    printf("rtm: %d\n", is.RTM());
+    printf("avx512f: %d\n", is.AVX512F());
+    printf("rdseed: %d\n", is.RDSEED());
+    printf("adx: %d\n", is.ADX());
+    printf("avx512pf: %d\n", is.AVX512PF());
+    printf("avx512er: %d\n", is.AVX512ER());
+    printf("avx512cd: %d\n", is.AVX512CD());
+    printf("sha: %d\n", is.SHA());
+    printf("prefetchwt1: %d\n", is.PREFETCHWT1());
+    printf("lahf: %d\n", is.LAHF());
+    printf("lzcnt: %d\n", is.LZCNT());
+    printf("abm: %d\n", is.ABM());
+    printf("sse4a: %d\n", is.SSE4a());
+    printf("xop: %d\n", is.XOP());
+    printf("tbm: %d\n", is.TBM());
+    printf("syscall: %d\n", is.SYSCALL());
+    printf("mmxext: %d\n", is.MMXEXT());
+    printf("rdtscp: %d\n", is.RDTSCP());
+    printf("3dnowext: %d\n", is._3DNOWEXT());
+    printf("3dnow: %d\n", is._3DNOW());
+    printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
+    printf("avx512_vnni: %d\n", is.AVX512_VNNI());
+    printf("avx512_fp16: %d\n", is.AVX512_FP16());
+    printf("avx512_bf16: %d\n", is.AVX512_BF16());
+    printf("amx_tile: %d\n", is.AMX_TILE());
+    printf("amx_int8: %d\n", is.AMX_INT8());
+    printf("amx_fp16: %d\n", is.AMX_FP16());
+    printf("amx_bf16: %d\n", is.AMX_BF16());
+}
+#endif
+
+static int ggml_backend_cpu_x86_score() {
+    // FIXME: this does not check for OS support
+
+    cpuid_x86 is;
+    // if the CPU backend was built with any features not supported by the current CPU, it cannot be used
+    if (ggml_cpu_has_fma() && !is.FMA()) { return 0; }
+    if (ggml_cpu_has_f16c() && !is.F16C()) { return 0; }
+    if (ggml_cpu_has_ssse3() && !is.SSSE3()) { return 0; }
+    if (ggml_cpu_has_sse3() && !is.SSE3()) { return 0; }
+    if (ggml_cpu_has_avx() && !is.AVX()) { return 0; }
+    if (ggml_cpu_has_avx_vnni() && !is.AVX_VNNI()) { return 0; }
+    if (ggml_cpu_has_avx2() && !is.AVX2()) { return 0; }
+    if (ggml_cpu_has_avx512() && !is.AVX512F()) { return 0; }
+    if (ggml_cpu_has_avx512_vbmi() && !is.AVX512_VBMI()) { return 0; }
+    if (ggml_cpu_has_avx512_bf16() && !is.AVX512_BF16()) { return 0; }
+    if (ggml_cpu_has_avx512_vnni() && !is.AVX512_VNNI()) { return 0; }
+    if (ggml_cpu_has_amx_int8() && !is.AMX_INT8()) { return 0; }
+
+    // calculate a backend score based on the supported features
+    // more important features have a higher weight
+    int score = 0;
+    score +=  ggml_cpu_has_fma        () * 1;
+    score +=  ggml_cpu_has_f16c       () * 1<<1;
+    score +=  ggml_cpu_has_ssse3      () * 1<<2;
+    score +=  ggml_cpu_has_sse3       () * 1<<3;
+    score +=  ggml_cpu_has_avx_vnni   () * 1<<4;
+    score +=  ggml_cpu_has_avx        () * 1<<5;
+    score +=  ggml_cpu_has_avx2       () * 1<<6;
+    score +=  ggml_cpu_has_avx512     () * 1<<7;
+    // score +=  ggml_cpu_has_avx512_vbmi() * 1<<8; // not used
+    score +=  ggml_cpu_has_avx512_bf16() * 1<<9;
+    score +=  ggml_cpu_has_avx512_vnni() * 1<<10;
+    score +=  ggml_cpu_has_amx_int8   () * 1<<11;
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
+
+#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))

ggml/src/ggml-cpu/ggml-cpu-aarch64.c

@@ -1,7 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
-// SPDX-License-Identifier: MIT
-//
-
 #define GGML_COMMON_IMPL_C
 #include "ggml-common.h"
 
@@ -132,7 +128,7 @@ static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) {
 }
 
 static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__)
     const __m512i zero = _mm512_setzero_si512();
     return _mm512_dpbusd_epi32(zero, ax, sy);
 #else
@@ -187,6 +183,8 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y)
 }
 #endif
 
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
 static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -527,67 +525,47 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    if (ggml_cpu_has_neon()) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *)vx;
 
-        __asm__ __volatile__(
-            "movi v31.16b, #0x4\n"
-            "movi v30.16b, #0xf0\n"
-            "add %x[b_ptr], %x[b_ptr], #0x8\n"
-            "1:"  // Column loop
-            "add x22, %x[a_ptr], #0x2\n"
-            "movi v29.16b, #0x0\n"
-            "mov x21, %x[nb]\n"
-            "2:"  // Block loop
-            "ldr q28, [%x[b_ptr], #0x0]\n"
-            "ldr q27, [x22, #0x0]\n"
-            "movi v26.4s, #0x0\n"
-            "sub x20, x22, #0x2\n"
-            "ldr q25, [x22, #0x10]\n"
-            "ldr q24, [%x[b_ptr], #0x10]\n"
-            "sub x21, x21, #0x1\n"
-            "add x22, x22, #0x22\n"
-            "ldr q23, [%x[b_ptr], #0x20]\n"
-            "ldr q22, [%x[b_ptr], #0x30]\n"
-            "ld1r { v21.8h }, [x20]\n"
-            "ldr q20, [%x[b_ptr], #-0x8]\n"
-            "sshl v16.16b, v28.16b, v31.16b\n"
-            "and v28.16b, v28.16b, v30.16b\n"
-            "sshl v19.16b, v24.16b, v31.16b\n"
-            "and v24.16b, v24.16b, v30.16b\n"
-            "add %x[b_ptr], %x[b_ptr], #0x48\n"
-            "sshl v18.16b, v23.16b, v31.16b\n"
-            "and v23.16b, v23.16b, v30.16b\n"
-            ".inst 0x4f9be21a  // sdot v26.4s, v16.16b, v27.4b[0]\n"
-            "sshl v17.16b, v22.16b, v31.16b\n"
-            "and v22.16b, v22.16b, v30.16b\n"
-            "fcvtl v21.4s, v21.4h\n"
-            "fcvtl v16.4s, v20.4h\n"
-            ".inst 0x4f99e39a  // sdot v26.4s, v28.16b, v25.4b[0]\n"
-            "fmul v16.4s, v16.4s, v21.4s\n"
-            ".inst 0x4fbbe27a  // sdot v26.4s, v19.16b, v27.4b[1]\n"
-            ".inst 0x4fb9e31a  // sdot v26.4s, v24.16b, v25.4b[1]\n"
-            ".inst 0x4f9bea5a  // sdot v26.4s, v18.16b, v27.4b[2]\n"
-            ".inst 0x4f99eafa  // sdot v26.4s, v23.16b, v25.4b[2]\n"
-            ".inst 0x4fbbea3a  // sdot v26.4s, v17.16b, v27.4b[3]\n"
-            ".inst 0x4fb9eada  // sdot v26.4s, v22.16b, v25.4b[3]\n"
-            "scvtf v26.4s, v26.4s, #0x4\n"
-            "fmla v29.4s, v26.4s, v16.4s\n"
-            "cbnz x21, 2b\n"
-            "sub %x[nc], %x[nc], #0x4\n"
-            "str q29, [%x[res_ptr], #0x0]\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "cbnz %x[nc], 1b\n"
-            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
-            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
-            : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
-            );
+        for (int c = 0; c < nc; c += ncols_interleaved) {
+            const block_q8_0 * a_ptr = (const block_q8_0 *)vy;
+            float32x4_t acc = vdupq_n_f32(0);
+            for (int b = 0; b < nb; b++) {
+                int8x16_t b0 = vld1q_s8((const int8_t *)b_ptr->qs);
+                int8x16_t b1 = vld1q_s8((const int8_t *)b_ptr->qs + 16);
+                int8x16_t b2 = vld1q_s8((const int8_t *)b_ptr->qs + 32);
+                int8x16_t b3 = vld1q_s8((const int8_t *)b_ptr->qs + 48);
+                float16x4_t bd = vld1_f16((const __fp16 *)b_ptr->d);
+
+                int8x16_t a0 = vld1q_s8(a_ptr->qs);
+                int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
+                float16x4_t ad = vld1_dup_f16((const __fp16 *)&a_ptr->d);
+
+                int32x4_t ret = vdupq_n_s32(0);
+
+                ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0);
+                ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1);
+                ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2);
+                ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3);
+
+                ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0);
+                ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1);
+                ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2);
+                ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3);
+
+                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                                vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+                a_ptr++;
+                b_ptr++;
+            }
+            vst1q_f32(s, acc);
+            s += ncols_interleaved;
+        }
         return;
     }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
     float sumf[4];
     int sumi;
 
@@ -996,6 +974,102 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float * res_ptr = s;
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            float32x4_t sumf = vdupq_n_f32(0);
+            for (int l = 0; l < nb; l++) {
+                uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
+                uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
+                uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
+                uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
+
+                int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
+                int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
+                int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
+                int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
+                int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
+                int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
+                int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
+                int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
+
+                int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
+                int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
+
+                int32x4_t sumi = vdupq_n_s32(0);
+                sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
+                sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
+                sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
+                sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
+                sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
+                sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
+                sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
+                sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
+
+                float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+                float32x4_t d = a_d * b_d;
+
+                sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
+            }
+
+            vst1q_f32(res_ptr + x * 4, sumf);
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1017,7 +1091,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    if (ggml_cpu_has_neon()) {
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -3386,6 +3460,117 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+                float32x4_t sumf[4];
+                for (int m = 0; m < 4; m++) {
+                    sumf[m] = vdupq_n_f32(0);
+                }
+
+                for (int l = 0; l < nb; l++) {
+                    float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
+                    float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+
+                    int32x4_t sumi_0 = vdupq_n_s32(0);
+                    int32x4_t sumi_1 = vdupq_n_s32(0);
+                    int32x4_t sumi_2 = vdupq_n_s32(0);
+                    int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                    for (int k = 0; k < 4; k++) {
+                        int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
+                        int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
+
+                        uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
+                        int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
+                        int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
+
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
+                    }
+
+                    sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                    sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                    sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                    sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+                }
+
+                for (int m = 0; m < 4; m++) {
+                    vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+                }
+            }
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
 // FIXME: this code is duplicated from ggml-aarch64.c
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
@@ -3518,6 +3703,70 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block,
     GGML_UNUSED(data_size);
 }
 
+static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+
+    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
+    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    block_iq4_nl dst_tmp[4];
+    int nrow = t->ne[1]; // Number of rows
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 // Prepare for optimized kernels if applicable
 void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * restrict data, size_t data_size) {
     if (cur->type == repack_type) {
@@ -3525,20 +3774,30 @@ void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_
         return;
     }
 
-    GGML_ASSERT(cur->type == GGML_TYPE_Q4_0);
-
-    switch (repack_type) {
-        case GGML_TYPE_Q4_0_8_8:
-            repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size);
-            break;
-        case GGML_TYPE_Q4_0_4_8:
-            repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size);
-            break;
-        case GGML_TYPE_Q4_0_4_4:
-            repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size);
-            break;
-        default:
-            GGML_ABORT("Unsupported type");
+    if (cur->type == GGML_TYPE_Q4_0) {
+        switch (repack_type) {
+            case GGML_TYPE_Q4_0_8_8:
+                repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size);
+                break;
+            case GGML_TYPE_Q4_0_4_8:
+                repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size);
+                break;
+            case GGML_TYPE_Q4_0_4_4:
+                repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size);
+                break;
+            default:
+                GGML_ABORT("Unsupported type");
+        }
+    } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        switch (repack_type) {
+            case GGML_TYPE_IQ4_NL_4_4:
+                repack_iq4_nl_to_iq4_nl_4_bl(cur, 4, data, data_size);
+                break;
+            default:
+                GGML_ABORT("Unsupported type");
+        }
+    } else {
+        GGML_ABORT("Unsupported type");
     }
 }
 
@@ -3551,9 +3810,13 @@ enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * c
         if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
             return GGML_TYPE_Q4_0_4_8;
         }
-        if (ggml_cpu_has_neon()) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             return GGML_TYPE_Q4_0_4_4;
         }
+    } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            return GGML_TYPE_IQ4_NL_4_4;
+        }
     }
 
     return cur->type;