alloc : fix allocation data of pre-allocated leafs

whisper : allocate encoder results in dedicated buffer
2025-06-24 17:15:19 +00:00 · 2024-03-16 15:47:14 +01:00 · 2024-03-16 16:15:12 +02:00
181 changed files with 19794 additions and 46062 deletions
--- a/.devops/main-cuda.Dockerfile
+++ b/.devops/main-cuda.Dockerfile
@ -28,8 +28,6 @@ COPY .. .
 RUN make

 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
-ENV CUDA_MAIN_VERSION=12.3
-ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 WORKDIR /app

 RUN apt-get update && \
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -15,10 +15,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -36,7 +36,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Dependencies
        run: |
@ -53,10 +53,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Build
-        uses: cross-platform-actions/action@v0.24.0
+        uses: cross-platform-actions/action@v0.15.0
        with:
          operating_system: freebsd
          version: '13.2'
@ -77,10 +77,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -105,10 +105,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -133,10 +133,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -152,21 +152,21 @@ jobs:

  ubuntu-22-cmake-sycl:
    runs-on: ubuntu-22.04
-
+    
    strategy:
      fail-fast: false
      matrix:
        dwhisper_sycl: [ON]
        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
+        dcmake_cxx_compiler: [icpx] 
        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]

    continue-on-error: true

    steps:
      - name: Clone
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+        
      - name: add oneAPI to apt
        shell: bash
        run: |
@ -189,8 +189,8 @@ jobs:

      - name: Clone
        id: checkout
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+            
      - name: Build
        id: cmake_build
        run: |
@ -202,21 +202,21 @@ jobs:

  ubuntu-22-cmake-sycl-fp16:
    runs-on: ubuntu-22.04
-
+    
    strategy:
      fail-fast: false
      matrix:
        dwhisper_sycl: [ON]
        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
+        dcmake_cxx_compiler: [icpx] 
        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]

    continue-on-error: true

    steps:
      - name: Clone
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+        
      - name: add oneAPI to apt
        shell: bash
        run: |
@ -239,8 +239,8 @@ jobs:

      - name: Clone
        id: checkout
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+            
      - name: Build
        id: cmake_build
        run: |
@ -249,65 +249,7 @@ jobs:
          cd build
          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
          cmake --build . --config Release -j $(nproc)
-
-  windows-msys2:
-    runs-on: windows-latest
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
-          - { sys: CLANG64, env: clang-x86_64, build: Release }
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Setup ${{ matrix.sys }}
-        uses: msys2/setup-msys2@v2
-        with:
-          update: true
-          msystem: ${{matrix.sys}}
-          install: >-
-            base-devel
-            mingw-w64-${{matrix.env}}-toolchain
-            mingw-w64-${{matrix.env}}-cmake
-            mingw-w64-${{matrix.env}}-SDL2
-            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 WHISPER_OPENBLAS=1 -j $(nproc)
-
-      - name: Build using CMake
-        shell: msys2 {0}
-        run: |
-            cmake -B build
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
-      - name: Clean after building using CMake
-        shell: msys2 {0}
-        run: |
-            rm -rf build
-
-      - name: Build using CMake w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            cmake -B build -DWHISPER_OPENBLAS=ON
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
+  
  windows:
    runs-on: windows-latest

@ -328,10 +270,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Fetch SDL2 and set SDL2_DIR
        if: matrix.sdl2 == 'ON'
@ -356,14 +298,14 @@ jobs:
        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}

      - name: Upload dll
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
        with:
          name: ${{ matrix.jnaPath }}_whisper.dll
          path: build/bin/${{ matrix.build }}/whisper.dll

      - name: Upload binaries
        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
          name: whisper-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}
@ -392,10 +334,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Fetch OpenBLAS
        if: matrix.blas == 'ON'
@ -453,13 +395,13 @@ jobs:

      - name: Upload binaries
        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}

  windows-cublas:
-    runs-on: windows-2019
+    runs-on: windows-latest

    strategy:
      matrix:
@ -476,14 +418,14 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Install CUDA Toolkit
        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
+        uses: Jimver/cuda-toolkit@v0.2.11
        with:
          cuda: '${{ matrix.cuda-toolkit }}'

@ -498,7 +440,7 @@ jobs:
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUDA=${{ matrix.cublas }}
+          -DWHISPER_CUBLAS=${{ matrix.cublas }}
          -DWHISPER_SDL2=${{ matrix.sdl2 }}

      - name: Build ${{ matrix.cuda-toolkit }}
@ -519,7 +461,7 @@ jobs:

      - name: Upload binaries
        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}
@ -533,10 +475,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Setup emsdk
-        uses: mymindstorm/setup-emsdk@v14
+        uses: mymindstorm/setup-emsdk@v12

      - name: Verify
        run: emcc -v
@ -555,7 +497,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Configure
        run: |
@ -573,24 +515,24 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
        with:
          path: whisper

      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
        with:
          repository: ggerganov/ggml
          path: ggml

      - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v3
        with:
          distribution: zulu
-          java-version: 21
+          java-version: 17

      - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
+        uses: android-actions/setup-android@v2

      - name: Build
        run: |
@ -608,19 +550,20 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: set up JDK 11
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v3
        with:
          java-version: '11'
          distribution: 'temurin'
          cache: gradle

      - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
+        uses: android-actions/setup-android@v2
        with:
-          cmdline-tools-version: 9.0
+          api-level: 30
+          build-tools-version: 30.0.3

      - name: Build
        run: |
@ -632,16 +575,15 @@ jobs:
    needs: [ 'windows' ]
    runs-on: windows-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3

      - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v1
        with:
-          distribution: zulu
-          java-version: 20
+          java-version: 17

      - name: Download Windows lib
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v3
        with:
          name: win32-x86-64_whisper.dll
          path: bindings/java/build/generated/resources/main/win32-x86-64
@ -654,7 +596,7 @@ jobs:
          ./gradlew build

      - name: Upload jar
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
        with:
          name: whispercpp.jar
          path: bindings/java/build/libs/whispercpp-*.jar
@ -676,7 +618,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Test quantize
        run: |
--- a/.github/workflows/examples.yml
+++ b/.github/workflows/examples.yml
@ -37,7 +37,7 @@ jobs:
        run: npm install

      - name: Compile addon.node
-        run: npx cmake-js compile -T addon.node -B Release
+        run: npx cmake-js compile -T whisper-addon -B Release

      - name: Download test model
        run: |
--- a/.gitignore
+++ b/.gitignore
@ -6,8 +6,6 @@
 .vs/
 .vscode/
 .DS_Store
-.vimspector.json
-/CMakeSettings.json

 build/
 build-coreml/
@ -60,4 +58,4 @@ benchmark_results.csv
 cmake-build-debug/
 .cxx/
 .gradle/
-local.properties
+local.properties
--- a/301
+++ b/301
@ -1,301 +0,0 @@
-# date: Tue Apr  9 20:27:03 EEST 2024
-# this file is auto-generated by scripts/gen-authors.sh
-
-0/0 <zero@imaskeleton.me>
-0cc4m <picard12@live.de>
-0xsourcecode <134374803+0xsourcecode@users.noreply.github.com>
-AT <manyoso@users.noreply.github.com>
-Aarni Koskela <akx@iki.fi>
-Aaron Pham <29749331+aarnphm@users.noreply.github.com>
-Aaron Taylor <aaron@exphat.com>
-Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
-Abitofevrything <54505189+abitofevrything@users.noreply.github.com>
-AfryMask <AfryMask@163.com>
-Ahmad Bilal <ahmad.bilal@empglabs.com>
-AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
-Akash Mahajan <akash7190@gmail.com>
-Akash Mahajan <akashmjn@stanford.edu>
-Al Hoang <3811822-hoanga@users.noreply.gitlab.com>
-Alan <unknown>
-Aleksander Andrzejewski <18704749+aleksanderandrzejewski@users.noreply.github.com>
-Alex Azarov <alex@azarov.by>
-Alex Bacart <13940752+alex-bacart@users.noreply.github.com>
-Alex Evgrashin <aevgrashin@yandex.ru>
-Alexandr Graschenkov <alexandr.graschenkov91@gmail.com>
-Alexandru Mariuti <alex@mariuti.com>
-Alexey Kharlamov <alexey@kharlamov.biz>
-Alfredo Montesinos <alfredo.montesinos@g.austincc.edu>
-Ali Alameh <ali.alameh@isae.edu.lb>
-Ananta Bastola <anantarajbastola@gmail.com>
-Andreu Huguet <andreuhuguet@gmail.com>
-Andrew Huynh <a5thuynh@gmail.com>
-Andrew S <andrews54757@gmail.com>
-Andy Maloney <asmaloney@gmail.com>
-Anton Kostin <masguit42@users.noreply.github.com>
-Artyom Mezin <psycho.fading@gmail.com>
-Asad Memon <asad.lionpk@gmail.com>
-Ashraful Islam <ashraful.meche@gmail.com>
-AsukaMinato <asukaminato@nyan.eu.org>
-AustinMroz <austinmroz@utexas.edu>
-Avik Sengupta <avik@sengupta.net>
-Bader-eddine Ouaich <49657842+baderouaich@users.noreply.github.com>
-Baffin Lee <baffinlee@gmail.com>
-Ben Nortier <bjnortier@gmail.com>
-Benjamin Heiniger <benjamin.heiniger@bluewin.ch>
-Bo-Yi Wu <appleboy.tw@gmail.com>
-Boris Bliznioukov <blib@mail.com>
-Borislav Stanimirov <b.stanimirov@abv.bg>
-Brad Murray <59848399+bradmurray-dt@users.noreply.github.com>
-Brian Murray <brian@bmurray.ca>
-CRD716 <crd716@gmail.com>
-Canis Lupus <Canis-UK@users.noreply.github.com>
-Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
-ChangSeok Oh <shivamidow@users.noreply.github.com>
-Chaoqun <27287694+OpenWaygate@users.noreply.github.com>
-Chia-Hsiang Cheng <88014292+garychia@users.noreply.github.com>
-Chidi Williams <williamschidi1@gmail.com>
-Christian <12550267+iceychris@users.noreply.github.com>
-Clifford Heath <clifford.heath@gmail.com>
-Colin <github@whoisc.cc>
-DGdev91 <DGdev91@users.noreply.github.com>
-Damian Czaja <trojan295@protonmail.com>
-Daniel Bevenius <daniel.bevenius@gmail.com>
-David <dnhkng@gmail.com>
-David Thorpe <djt@mutablelogic.com>
-Davidson Francis <davidsondfgl@gmail.com>
-Dener Stassun <denerstassun@gmail.com>
-Didzis Gosko <didzis@users.noreply.github.com>
-Digipom <admin@digipom.com>
-Dimo <dimo@ieee.org>
-Dody Suria Wijaya <dodysw@gmail.com>
-Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
-Duncan McConnell <ddmcconnell4@gmail.com>
-Egor Egorov <me@egorfine.com>
-Elkana Bardugo <ttv200@gmail.com>
-Emmanuel Schmidbauer <eschmidbauer@gmail.com>
-Engininja2 <139037756+Engininja2@users.noreply.github.com>
-Eric Swanson <eswanson@alloscomp.com>
-Eric Tendian <erictendian@gmail.com>
-Erik Scholz <Green-Sky@users.noreply.github.com>
-Evan Jones <evan.q.jones@gmail.com>
-Evan Martin <evan.martin@gmail.com>
-Eve <139727413+netrunnereve@users.noreply.github.com>
-Evgeny Kuznetsov <evgeny@kuznetsov.md>
-F1L1P <78918286+F1L1Pv2@users.noreply.github.com>
-Fangjun Kuang <csukuangfj@gmail.com>
-Felix <stenbackfelix@gmail.com>
-Finn Voorhees <finnvoorhees@gmail.com>
-FlippFuzz <41221030+FlippFuzz@users.noreply.github.com>
-Gang Chen <goncha@gmail.com>
-Gavin Cai <gavin1818@hotmail.com>
-George Hindle <george@georgehindle.com>
-Georgi Gerganov <ggerganov@gmail.com>
-GitAritron <103900385+GitAritron@users.noreply.github.com>
-GiviMAD <GiviMAD@users.noreply.github.com>
-Gleicon Moraes <gleicon@gmail.com>
-Gregor Jasny <gjasny@googlemail.com>
-Guillaume Wenzek <gwenzek@users.noreply.github.com>
-HY. Kelvin Lee <34256578+hykelvinlee42@users.noreply.github.com>
-Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
-Hang <bebound@gmail.com>
-Herman Semenov <GermanAizek@yandex.ru>
-Hrishikesh Barman <geekodour@users.noreply.github.com>
-Ian Bicking <ian@ianbicking.org>
-Ian Bull <irbull@eclipsesource.com>
-Ikko Ashimine <eltociear@gmail.com>
-InconsolableCellist <23345188+InconsolableCellist@users.noreply.github.com>
-Ismatulla Mansurov <47342870+sapoepsilon@users.noreply.github.com>
-Ivan Gorin <ivangorin21@gmail.com>
-JJ <103335846+computerscienceiscool@users.noreply.github.com>
-Jack Mousseau <jmousseau@users.noreply.github.com>
-JacobLinCool <jacoblincool@gmail.com>
-Jakub Ráček <blizzcz@gmail.com>
-Jared Van Bortel <jared@nomic.ai>
-Jay Binks <jaybinks@gmail.com>
-Jhen-Jie Hong <developer@jhen.me>
-Jhen-Jie Hong <iainst0409@gmail.com>
-JidongZhang-THU <1119708529@qq.com>
-Jo Liss <joliss42@gmail.com>
-Johan <jr.raffin@gmail.com>
-Johannes Gäßler <johannesg@5d6.de>
-John Balis <phobossystems@gmail.com>
-Jonathan Soo <jcsoo@agora.com>
-Jonno <1160532+razodactyl@users.noreply.github.com>
-Joonas Pihlajamaa <joonas.pihlajamaa@iki.fi>
-Jose <34888496+Jerry-Master@users.noreply.github.com>
-Josh Bleecher Snyder <josharian@gmail.com>
-Judd <foldl@users.noreply.github.com>
-Jumper775 <78500318+jumpers775@users.noreply.github.com>
-Justine Tunney <jtunney@gmail.com>
-KP Kaiser <kirk@zothcorp.com>
-Kamilake <exjang0@gmail.com>
-Kartik Saranathan <278928+Kartiku@users.noreply.github.com>
-Kasumi <90275229+kasumi-1@users.noreply.github.com>
-Kawrakow <48489457+ikawrakow@users.noreply.github.com>
-Kevin Brothaler <admin@digipom.com>
-Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
-Kreijstal <rainb@tfwno.gf>
-Kylin <56434533+KyL0N@users.noreply.github.com>
-LBlue <153975653+lbluep@users.noreply.github.com>
-Larry Battle <larry.battle.tech@gmail.com>
-Laytan Laats <laytanlaats@hotmail.com>
-Leo Moll <leo.moll@yeasoft.com>
-Lexevolution <31176843+Lexevolution@users.noreply.github.com>
-LittleLoli <26589867+WhichWho@users.noreply.github.com>
-Lucas Zanek <57494138+LucasZNK@users.noreply.github.com>
-Luis Herrera <herrera-luis@users.noreply.github.com>
-Lukas Rist <glaslos@gmail.com>
-M. A. Ali <73258591+MightyStud@users.noreply.github.com>
-M. Eren Akbiyik <erenakbiyik@gmail.com>
-Maciek <maciek.mab122@gmail.com>
-Marcin Mielniczuk <marmistrz.dev@zoho.eu>
-Martin Warnaar <martinwarnaar@gmail.com>
-Matheus de Sousa <23645013+keyehzy@users.noreply.github.com>
-Mathijs de Bruin <mathijs@mathijsfietst.nl>
-Matija Pevec <mightymatth@users.noreply.github.com>
-Maximiliano Levi <8160966+maxilevi@users.noreply.github.com>
-Meng, Hengyu <hengyu.meng@intel.com>
-Michael Podvitskiy <podvitskiymichael@gmail.com>
-Michael Rienstra <mrienstra@gmail.com>
-Mikhail Grigorev <sleuthhound@gmail.com>
-Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
-Mohit Agarwal <mohit@sdf.org>
-Murilo Santana <mvrilo@gmail.com>
-Neil Chudleigh <nchudleigh@users.noreply.github.com>
-Neo Zhang Jianyu <jianyu.zhang@intel.com>
-Neuman Vong <neuman.vong@gmail.com>
-Nicholas Albion <nalbion@yahoo.com>
-Niels Mayer <Niels.Mayer@gmail.com>
-Okabintaro <103938900+Okabintaro@users.noreply.github.com>
-Oleg Sidorov <me@whitebox.io>
-Oleg Sidorov <oleg@sidorov.nl>
-Ondrej Kokes <ondrej.kokes@gmail.com>
-Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
-Paul Tsochantaris <ptsochantaris@icloud.com>
-Philipp Zabel <philipp.zabel@gmail.com>
-Philippe Normand <phil@base-art.net>
-Przemysław Pawełczyk <przemoc@gmail.com>
-Qianhe Chen <54462604+chenqianhe@users.noreply.github.com>
-Radosław Gryta <radek.gryta@gmail.com>
-Reinforce-II <fate@eastal.com>
-Reinis Muiznieks <muiznieks.reinis@gmail.com>
-RelatedTitle <r3latedtitle@gmail.com>
-RhinoDevel <RhinoDevel@users.noreply.github.com>
-Rich Jones <miserlou@gmail.com>
-Robin <robin.xw@hotmail.com>
-Roddur Dasgupta <roddurd@gmail.com>
-Roland Rabien <figbug@gmail.com>
-Rotem Dan <rotemdan@gmail.com>
-Ryan Hitchman <hitchmanr@gmail.com>
-Ryan Metcalfe <107415876+RyanMetcalfeInt8@users.noreply.github.com>
-RyanChang <ftes90015@gmail.com>
-Sam <49637763+Onlyartist9@users.noreply.github.com>
-Sam Pullara <spullara@gmail.com>
-Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
-Sergio López <slp@sinrega.org>
-Siddharth Ramakrishnan <srr2141@columbia.edu>
-Simon Moisselin <simon.moisstoll@gmail.com>
-Sindre Sorhus <sindresorhus@gmail.com>
-Slava Primenko <primenko.s@gmail.com>
-Syahmi Azhar <prsyahmi@gmail.com>
-Syed Jafri <syedjafri97@gmail.com>
-Sơn Phan Trung <phantrungson17@gmail.com>
-Taisei Mima <bhbstar.me@gmail.com>
-Takeshi Inoue <inoue.takeshi@gmail.com>
-Tamotsu Takahashi <ttakah+github@gmail.com>
-Taras Glek <taras@thegp.com>
-Tauseef Mohiuddin <35351464+tauseefmohammed2@users.noreply.github.com>
-Thijs Raymakers <thijs@raymakers.nl>
-Thomas Fitzsimmons <fitzsim@fitzsim.org>
-Tiago Fassoni <tiagofassoni@users.noreply.github.com>
-Tienshiao Ma <tienshiao@tienshiao.org>
-Timothy Cronin <40186632+4imothy@users.noreply.github.com>
-Tobrun <tobrun.van.nuland@gmail.com>
-Todd <taf2@users.noreply.github.com>
-Tong Li <31761981+litongjava@users.noreply.github.com>
-Topping1 <78745143+Topping1@users.noreply.github.com>
-Travis Cline <travis.cline@gmail.com>
-UEXTM.com <84163508+uextm@users.noreply.github.com>
-Vadim Peretokin <vperetokin@hey.com>
-Valentin Gosu <1454649+valenting@users.noreply.github.com>
-Vulcan <93451215+trholding@users.noreply.github.com>
-WhiteOlivierus <36532695+WhiteOlivierus@users.noreply.github.com>
-Xiang (Kevin) Li <kevinli020508@gmail.com>
-Xiao-Yong Jin <jinxiaoyong@gmail.com>
-XiaotaoChen <chenxiaotao1234@gmail.com>
-Yajing Tang <phillis@google.com>
-Yang Shen <aplshenyang@gmail.com>
-Yunès <jean.baptiste.yunes@free.fr>
-ZaBlazzingZephyrus <119159668+blazingzephyr@users.noreply.github.com>
-Zigfrid Zvezdin <ziggerZZ@gmail.com>
-Zollner <24618122+Zolliner@users.noreply.github.com>
-ai-at-home <149282006+ai-at-home@users.noreply.github.com>
-alonfaraj <alonfaraj@gmail.com>
-andypayne <apayne@gmail.com>
-ardfork <134447697+ardfork@users.noreply.github.com>
-automaticcat <daogiatuank54@gmail.com>
-be-next <jerome.ramette@gmail.com>
-bert hubert <bert@hubertnet.nl>
-bmwl <brian.marshall@tolko.com>
-bobqianic <129547291+bobqianic@users.noreply.github.com>
-bocytko <bocytko+github@gmail.com>
-boolemancer <48014766+boolemancer@users.noreply.github.com>
-boolemancer <boolemancer@gmail.com>
-bradmit <151883577+bradmit@users.noreply.github.com>
-brunofaustino <b.fa.amorim@gmail.com>
-bssrdf <merlintiger@hotmail.com>
-byte-6174 <88070277+byte-6174@users.noreply.github.com>
-cdosoftei <ciprian.dosoftei@gmail.com>
-clach04 <Chris.Clark@actian.com>
-compilade <113953597+compilade@users.noreply.github.com>
-conradg <conradjgodfrey@gmail.com>
-ddpasa <112642920+ddpasa@users.noreply.github.com>
-denersc <denerstassun@gmail.com>
-dscripka <dscripka@users.noreply.github.com>
-duthils <duthils@duthils.net>
-ecneladis <ecneladis@users.noreply.github.com>
-faker <nspyia2002@gmail.com>
-fitzsim <fitzsim@fitzsim.org>
-fraxy-v <65565042+fraxy-v@users.noreply.github.com>
-genevera (she/her) <genevera@users.noreply.github.com>
-geniusnut <geniusnut@gmail.com>
-greeshmay <greeshmay@gmail.com>
-hydai <z54981220@gmail.com>
-iamthad <thadeus.j.fleming@gmail.com>
-james wolf <contractorwolf@hotmail.com>
-joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
-jorismertz <35079666+jorismertz@users.noreply.github.com>
-junkfood <69683722+JunkFood02@users.noreply.github.com>
-jwijffels <jwijffels@bnosac.be>
-kamranjon <kamranjon@gmail.com>
-katsu560 <katsu560oo-@docomo.ne.jp>
-kennethge <57784063+kenneth-ge@users.noreply.github.com>
-keyehzy <msamuel@aluno.puc-rio.br>
-leejet <leejet714@gmail.com>
-litong <31761981+litongjava@users.noreply.github.com>
-lnyan <lkwq007@gmail.com>
-m.bell <m.bell@techsmith.com>
-mkiol <mkiol@users.noreply.github.com>
-novag <7754358+novag@users.noreply.github.com>
-pajowu <pajowu@pajowu.de>
-polarmoon <90010972+polarmoon@users.noreply.github.com>
-rlapray <lapray.romain@gmail.com>
-sandrohanea <40202887+sandrohanea@users.noreply.github.com>
-semiformal-net <84111142+semiformal-net@users.noreply.github.com>
-shibukazu <61775791+shibukazu@users.noreply.github.com>
-shikokuchuo <53399081+shikokuchuo@users.noreply.github.com>
-slaren <slarengh@gmail.com>
-slashlib <slashlib@users.noreply.github.com>
-snadampal <87143774+snadampal@users.noreply.github.com>
-st-gr <38470677+st-gr@users.noreply.github.com>
-texmex76 <40733439+texmex76@users.noreply.github.com>
-thefinaldegree <thefinaldegree@gmail.com>
-trixirt <trix@redhat.com>
-ulatekh <ulatekh@yahoo.com>
-undef <undefdev@gmail.com>
-venkr <venkateshrameshkumar+1@gmail.com>
-vicalloy <zbirder@gmail.com>
-xdrudis <xavierdrudis@yahoo.es>
-zhouwg <6889919+zhouwg@users.noreply.github.com>
-布客飞龙 <562826179@qq.com>
-Артём Земляк <azemlyak@smart-consulting.ru>
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -1,9 +1,6 @@
 cmake_minimum_required (VERSION 3.5)

-# Allow for the creation of solution folders.
-set_property(GLOBAL PROPERTY USE_FOLDERS ON)
-
-project(whisper.cpp VERSION 1.6.2)
+project(whisper.cpp VERSION 1.5.4)
 set(SOVERSION 1)

 # Add path to modules
@ -59,17 +56,10 @@ option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDA

 option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)

-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    option(WHISPER_FFMPEG                 "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
-endif()
-
-option(WHISPER_NO_AVX                 "whisper: disable AVX"         OFF)
-option(WHISPER_NO_AVX2                "whisper: disable AVX2"        OFF)
-option(WHISPER_NO_AVX512              "whisper: disable AVX512"      ON)
-option(WHISPER_NO_AVX512_VBMI         "whisper: disable AVX512-VBMI" ON)
-option(WHISPER_NO_AVX512_VNNI         "whisper: disable AVX512-VNNI" ON)
-option(WHISPER_NO_FMA                 "whisper: disable FMA"         OFF)
-option(WHISPER_NO_F16C                "whisper: disable F16c"        OFF)
+option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)

 option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)

@ -84,12 +74,9 @@ else()
    option(WHISPER_BLAS                  "whisper: use BLAS libraries"                        OFF)
    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor"                       Generic)
    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"                           OFF)
-    option(WHISPER_OPENBLAS_INTERFACE64  "whisper: use OpenBLAS w/ 64-bit interface"          OFF)
-    option(WHISPER_CUDA                  "whisper: support for CUDA"                          OFF)
-    option(WHISPER_CUBLAS                "whisper: support for CUDA (deprecated)"             OFF)
+    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"                        OFF)
    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               OFF)
-    option(WHISPER_MKL                   "whisper: use Intel Math Kernel Library (MKL)"       OFF)
    option(WHISPER_SYCL                  "whisper: use SYCL"                                  OFF)
    option(WHISPER_SYCL_F16              "whisper: use 16 bit floats for sycl calculations"   OFF)
 endif()
@ -129,26 +116,6 @@ else()
    set(CMAKE_CXX_STANDARD 11)
 endif()

-if (WHISPER_FFMPEG)
-    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
-    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
-    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
-    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
-    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
-    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
-    find_package(FFmpeg REQUIRED)
-    if (NOT ${FFMPEG_FOUND})
-        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
-    endif()
-    message(STATUS "Found ffmpeg libs: ${FFMPEG_LIBRARIES}")
-    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
-    message(STATUS "ffmpeg definitions: ${FFMPEG_DEFINITIONS}")
-    message(STATUS "Found avformat ${AVFORMAT_VERSION}")
-    include_directories(${FFMPEG_INCLUDE_DIRS})
-    add_compile_definitions(WHISPER_FFMPEG)
-    set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS} ${FFMPEG_LIBRARIES})
-endif()
-
 # on APPLE
 if (APPLE)
    # include Accelerate framework
@ -189,37 +156,27 @@ if (APPLE)

        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)

-        # copy ggml-common.h and ggml-metal.metal to bin directory
-        configure_file(ggml-common.h    bin/ggml-common.h    COPYONLY)
+        # copy ggml-metal.metal to bin directory
        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)

        if (WHISPER_METAL_EMBED_LIBRARY)
            enable_language(ASM)
            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)

-            set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-            set(COMMON_HEADER   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
+            set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")

            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-            set(EMBED_METALLIB_SOURCE "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-combined.metal")
-
-            add_custom_command(
-                OUTPUT ${EMBED_METALLIB_SOURCE}
-                COMMAND sed -e "/^#include \\\"ggml-common.h\\\"/r ${COMMON_HEADER}" -e "/^#include \\\"ggml-common.h\\\"/d" ${METALLIB_SOURCE} > ${EMBED_METALLIB_SOURCE}
-                DEPENDS ${METALLIB_SOURCE} ${COMMON_HEADER}
-                COMMENT "Generating combined Metal library for embedding"
-            )

            add_custom_command(
                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".incbin \\\"${EMBED_METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-                DEPENDS ${EMBED_METALLIB_SOURCE}
+                DEPENDS ${METALLIB_SOURCE}
                COMMENT "Generate assembly for embedded Metal library"
            )

@ -248,82 +205,30 @@ endif()
 if (WHISPER_OPENBLAS)
    set(WHISPER_BLAS_VENDOR "OpenBLAS")
    set(WHISPER_BLAS ON)
-    # BLA_PKGCONFIG_BLAS is supported since CMake 3.25.
-    # FindBLAS.cmake pkg-config logic seems incomplete, because when
-    # BLA_SIZEOF_INTEGER is 8, then it should search for blas64 instead of blas.
-    # blas.pc/blas64.pc are not always provided, so let's be more specific
-    # and go with openblas.pc/openblas64.pc if WHISPER_OPENBLAS is on.
-    if (WHISPER_OPENBLAS_INTERFACE64)
-        set(WHISPER_BLAS_LIB "openblas64")
-    else ()
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
-    set(BLA_PKGCONFIG_BLAS ${WHISPER_BLAS_LIB})
-    # OpenBLAS prebuilt libraries for Windows do not have "64" suffix in filename.
-    # (But .pc file has "64" suffix in filename for USE_64BITINT=1 Windows build.)
-    if (MSVC)
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
 endif()

 if (WHISPER_BLAS)
-    if (NOT "$ENV{OPENBLAS_PATH}" STREQUAL "")
-        if (WHISPER_STATIC)
-            set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_STATIC_LIBRARY_PREFIX})
-            set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_STATIC_LIBRARY_SUFFIX})
+    if (WIN32)
+        if(DEFINED ENV{OPENBLAS_PATH})
+            set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories($ENV{OPENBLAS_PATH}/include)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
        else ()
-            if (CMAKE_IMPORT_LIBRARY_SUFFIX)
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_IMPORT_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_IMPORT_LIBRARY_SUFFIX})
-            else ()
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_SHARED_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_SHARED_LIBRARY_SUFFIX})
-            endif ()
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
        endif ()
-        # OpenBLAS prebuilt libraries hardcode "lib" prefix in filename even on Windows
-        if (WHISPER_OPENBLAS)
-            set(WHISPER_BLAS_LIB_PREFIX "lib")
-        endif ()
-        message(STATUS "BLAS compatible library path provided")
-        set(BLAS_LIBRARIES "$ENV{OPENBLAS_PATH}/lib/${WHISPER_BLAS_LIB_PREFIX}${WHISPER_BLAS_LIB}${WHISPER_BLAS_LIB_SUFFIX}")
-        message(STATUS "Libraries ${BLAS_LIBRARIES}")
-        set(BLAS_INCLUDE_DIRS "$ENV{OPENBLAS_PATH}/include")
-        message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
-        if (NOT EXISTS "${BLAS_LIBRARIES}")
-            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH misdefined.")
-        endif ()
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-        include_directories(${BLAS_INCLUDE_DIRS})
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
    else ()
-        if (WHISPER_STATIC)
-            # FindBLAS.cmake pkg-config logic seems incomplete, because when
-            # BLA_STATIC is on, then it should use pkg_check_modules_static
-            # instead of pkg_check_modules.
-            # Some manual variable overriding may be necessary if you don't
-            # achieve desired results.
-            set(BLA_STATIC 1)
-        endif ()
+        set(BLA_STATIC 1)
        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-        if (WHISPER_OPENBLAS_INTERFACE64)
-            set(BLA_SIZEOF_INTEGER 8)
-        else ()
-            set(BLA_SIZEOF_INTEGER 4)
-        endif()
+        set(BLA_SIZEOF_INTEGER 8)
        set(BLA_PREFER_PKGCONFIG 1)
        find_package(BLAS)

        if(BLAS_FOUND)
            message(STATUS "BLAS compatible library found")
            message(STATUS "Libraries ${BLAS_LIBRARIES}")
-            if (NOT DEFINED BLAS_INCLUDE_DIRS)
-                if (PKGC_BLAS_FOUND)
-                    set(BLAS_INCLUDE_DIRS "${PKGC_BLAS_INCLUDE_DIRS}")
-                else ()
-                    find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas)
-                endif()
-            endif()
-            message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
+            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
            include_directories(${BLAS_INCLUDE_DIRS})
            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
@ -333,19 +238,7 @@ if (WHISPER_BLAS)
    endif ()
 endif ()

-if (WHISPER_MKL)
-    find_package(MKL CONFIG REQUIRED PATHS $ENV{MKLROOT})
-    message(STATUS "Imported oneMKL targets: ${MKL_IMPORTED_TARGETS}")
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_BLAS_USE_MKL)
-endif()
-
 if (WHISPER_CUBLAS)
-    message(WARNING "WHISPER_CUBLAS is deprecated and will be removed in the future.\nUse WHISPER_CUDA instead")
-    set(WHISPER_CUDA ON)
-endif()
-
-if (WHISPER_CUDA)
    cmake_minimum_required(VERSION 3.17)

    find_package(CUDAToolkit)
@ -355,21 +248,19 @@ if (WHISPER_CUDA)

        enable_language(CUDA)

-        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.h)
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.cu)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)

-        add_compile_definitions(GGML_USE_CUDA)
+        add_compile_definitions(GGML_USE_CUBLAS)

        if (WHISPER_STATIC)
            if (WIN32)
                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt CUDA::cufft)
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
            else ()
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static CUDA::cufft_static)
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
            endif()
        else()
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt CUDA::cufft)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
        endif()

        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
@ -394,18 +285,16 @@ if (WHISPER_HIPBLAS)

    if (${hipblas_FOUND} AND ${hip_FOUND})
        message(STATUS "HIP and hipBLAS found")
-        set(GGML_HEADERS_ROCM "ggml-cuda.h")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)

-        file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
-
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
        if (WHISPER_STATIC)
            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
        endif()
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
    else()
        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
    endif()
@ -502,30 +391,16 @@ else()
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
-        if(NOT WHISPER_NO_AVX512)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX512")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX512")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX512")
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (NOT WHISPER_NO_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
-            endif()
-            if (NOT WHISPER_NO_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
-            endif()
-        elseif(NOT WHISPER_NO_AVX2)
+        if(NOT WHISPER_NO_AVX2)
            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-        elseif(NOT WHISPER_NO_AVX)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
+        else()
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
+            endif()
        endif()
    else()
        if (EMSCRIPTEN)
@ -538,15 +413,6 @@ else()
            if(NOT WHISPER_NO_AVX2)
                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
            endif()
-            if(NOT WHISPER_NO_AVX512)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw")
-                if(NOT WHISPER_NO_AVX512_VBMI)
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vbmi")
-                endif()
-                if(NOT WHISPER_NO_AVX512_VNNI)
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vnni")
-                endif()
-            endif()
            if(NOT WHISPER_NO_FMA)
                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
            endif()
@ -633,7 +499,6 @@ if (WHISPER_COREML)
    set_target_properties(${TARGET} PROPERTIES
        COMPILE_FLAGS "-fobjc-arc"
        )
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()

 if (WHISPER_OPENVINO)
@ -652,7 +517,6 @@ if (WHISPER_OPENVINO)
    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)

    target_link_libraries(${TARGET} PRIVATE openvino::runtime)
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()

 #
@ -673,19 +537,12 @@ add_library(${TARGET}
    ${GGML_SOURCES_METAL}
    ${GGML_SOURCES_CUDA}
    ${GGML_SOURCES_OPENCL}
-    ${GGML_SOURCES_SYCL}        ${GGML_HEADERS_SYCL}
-    ${GGML_SOURCES_ROCM}        ${GGML_HEADERS_ROCM}
+    ${GGML_SOURCES_SYCL}
+    ${GGML_HEADERS_SYCL}
    whisper.h
    whisper.cpp
    )

-if (WHISPER_CUDA)
-    target_sources(${TARGET} PRIVATE whisper-mel-cuda.cu)
-endif()
-
-include_directories (
-    .
-)
 # Set the version numbers
 set_target_properties(whisper PROPERTIES
    VERSION ${PROJECT_VERSION}
@ -706,10 +563,6 @@ if (WHISPER_OPENVINO)
    target_link_libraries(${TARGET} PRIVATE whisper.openvino)
 endif()

-if (WHISPER_MKL)
-    target_link_libraries(${TARGET} PUBLIC MKL::MKL)
-endif()
-
 if (MSVC)
    target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})

@ -762,7 +615,6 @@ target_compile_definitions(${TARGET} PUBLIC
    )

 set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")

 include(GNUInstallDirs)

--- a/2
+++ b/2
@ -1,6 +1,6 @@
 MIT License

-Copyright (c) 2023-2024 The ggml authors
+Copyright (c) 2023 Georgi Gerganov

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
--- a/160
+++ b/160
@ -18,17 +18,6 @@ ifndef NVCC_VERSION
 	endif
 endif

-# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
-# of non-gcc compilers don't have to provide g++ alias or wrapper.
-DEFCC  := cc
-DEFCXX := c++
-ifeq ($(origin CC),default)
-CC  := $(DEFCC)
-endif
-ifeq ($(origin CXX),default)
-CXX := $(DEFCXX)
-endif
-
 CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)

@ -142,68 +131,41 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysinfo -cpu
 	endif

-	# x86 ISA extensions (chronological order)
 	ifdef CPUINFO_CMD
-		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
-		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
 		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
-		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
-		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
-		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
-		AVX512F_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512F')
-		AVX512VBMI_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512VBMI')
-		AVX512VNNI_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX512_VNNI|AVX512VNNI')
-
-		# AVX-512 has many subsets, so let's make it easy to disable them all
-		ifneq ($(filter-out 0,$(WHISPER_NO_AVX512)),)
-			AVX512F_M :=
-			AVX512VBMI_M :=
-			AVX512VNNI_M :=
-		endif
-
-		ifneq (,$(SSE3_M))
-			CFLAGS   += -msse3
-			CXXFLAGS += -msse3
-		endif
-
-		ifneq (,$(SSSE3_M))
-			CFLAGS   += -mssse3
-			CXXFLAGS += -mssse3
-		endif
-
 		ifneq (,$(AVX_M))
 			CFLAGS   += -mavx
 			CXXFLAGS += -mavx
 		endif

-		ifneq (,$(F16C_M))
-			CFLAGS   += -mf16c
-			CXXFLAGS += -mf16c
-		endif
-
-		ifneq (,$(FMA_M))
-			CFLAGS   += -mfma
-			CXXFLAGS += -mfma
-		endif
-
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
 		ifneq (,$(AVX2_M))
 			CFLAGS   += -mavx2
 			CXXFLAGS += -mavx2
 		endif

-		ifneq (,$(AVX512F_M))
-			CFLAGS   += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
-			CXXFLAGS += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
 		endif

-		ifneq (,$(AVX512VBMI_M))
-			CFLAGS   += -mavx512vbmi
-			CXXFLAGS += -mavx512vbmi
+		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
+		ifneq (,$(F16C_M))
+			CFLAGS   += -mf16c
+			CXXFLAGS += -mf16c
 		endif

-		ifneq (,$(AVX512VNNI_M))
-			CFLAGS   += -mavx512vnni
-			CXXFLAGS += -mavx512vnni
+		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
+		ifneq (,$(SSE3_M))
+			CFLAGS   += -msse3
+			CXXFLAGS += -msse3
+		endif
+
+		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
+		ifneq (,$(SSSE3_M))
+			CFLAGS   += -mssse3
+			CXXFLAGS += -mssse3
 		endif
 	endif
 endif
@ -248,76 +210,41 @@ ifndef WHISPER_NO_METAL
 	endif
 endif

-ifneq ($(filter-out 0,$(WHISPER_OPENBLAS)),) # OpenBLAS
-	WHISPER_OPENBLAS_INTERFACE64 ?= 0 # use 32-bit interface by default
-	ifneq ($(filter-out 0,$(WHISPER_OPENBLAS_INTERFACE64)),)
-		WHISPER_BLAS_LIB := openblas64
-	else
-		WHISPER_BLAS_LIB := openblas
-	endif
-	ifneq ($(OPENBLAS_PATH),)
-		WHISPER_BLAS_CFLAGS  := -I$(OPENBLAS_PATH)/include
-		WHISPER_BLAS_LDFLAGS := -L$(OPENBLAS_PATH)/lib -l$(WHISPER_BLAS_LIB)
-	else
-		WHISPER_BLAS_LIB_PC_EXISTS := $(shell pkg-config --exists $(WHISPER_BLAS_LIB) && echo 1)
-		ifneq ($(filter-out 0,$(WHISPER_BLAS_LIB_PC_EXISTS)),)
-			WHISPER_BLAS_CFLAGS  := $(shell pkg-config --cflags $(WHISPER_BLAS_LIB))
-			WHISPER_BLAS_LDFLAGS := $(shell pkg-config --libs   $(WHISPER_BLAS_LIB))
-		else
-			WHISPER_BLAS_CFLAGS  := -I/usr/include/openblas
-			WHISPER_BLAS_LDFLAGS := -l$(WHISPER_BLAS_LIB)
-		endif
-	endif
-	CFLAGS  += $(WHISPER_BLAS_CFLAGS) -DGGML_USE_OPENBLAS
-	LDFLAGS += $(WHISPER_BLAS_LDFLAGS)
+ifdef WHISPER_OPENBLAS
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
+	LDFLAGS += -lopenblas
 endif

 ifdef WHISPER_CUBLAS
-# WHISPER_CUBLAS is deprecated and will be removed in the future
-	WHISPER_CUDA := 1
-endif
-
-ifdef WHISPER_CUDA
 	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
 		CUDA_ARCH_FLAG ?= native
 	else
 		CUDA_ARCH_FLAG ?= all
 	endif

-	CFLAGS      += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	CXXFLAGS    += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lcufft -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
-	WHISPER_OBJ += ggml-cuda.o whisper-mel-cuda.o
-	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)

-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -c $< -o $@
-
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif

-whisper-mel-cuda.o: whisper-mel-cuda.cu whisper.h ggml.h ggml-backend.h whisper-mel.hpp whisper-mel-cuda.hpp
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
-
 ifdef WHISPER_HIPBLAS
 	ROCM_PATH   ?= /opt/rocm
 	HIPCC       ?= $(ROCM_PATH)/bin/hipcc
 	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
-	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
 	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	LDFLAGS     += -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
 	WHISPER_OBJ += ggml-cuda.o
-	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))

-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 endif

@ -382,13 +309,6 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )

-ifdef WHISPER_CUBLAS
-$(info !!!!)
-$(info WHISPER_CUBLAS is deprecated and will be removed in the future. Use WHISPER_CUDA instead.)
-$(info !!!!)
-$(info )
-endif
-
 #
 # Build library
 #
@ -407,7 +327,7 @@ ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h

 WHISPER_OBJ += ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o

-whisper.o: whisper.cpp whisper.h whisper-mel.hpp ggml.h ggml-cuda.h
+whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@

 ifndef WHISPER_COREML
@ -431,19 +351,17 @@ WHISPER_OBJ += ggml-metal.o
 ifdef WHISPER_METAL_EMBED_LIBRARY
 CFLAGS += -DGGML_METAL_EMBED_LIBRARY

-ggml-metal-embed.o: ggml-metal.metal ggml-common.h
+ggml-metal-embed.o: ggml-metal.metal
 	@echo "Embedding Metal library"
 	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	$(eval TEMP_METALLIB=$(shell mktemp))
-	@sed "/^#include \"ggml-common.h\"/{r ggml-common.h"$$'\n'"d;}" ggml-metal.metal > $(TEMP_METALLIB)
 	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
 	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
 	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$(TEMP_METALLIB)\"" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
 	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
 	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
 	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f $(TEMP_ASSEMBLY) $(TEMP_METALLIB)
+	@rm -f ${TEMP_ASSEMBLY}

 WHISPER_OBJ += ggml-metal-embed.o
 endif
@ -464,7 +382,7 @@ clean:

 CC_SDL=`sdl2-config --cflags --libs`

-SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.cpp
+SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp

 main: examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ)
@ -483,8 +401,8 @@ server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
 stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)

-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)

 lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
@ -492,8 +410,8 @@ lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)

-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)

 #
 # Audio samples
--- a/README.md
+++ b/README.md
@ -4,10 +4,9 @@

 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)

-Stable: [v1.6.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)

 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:

@ -415,11 +414,11 @@ For more information about the Core ML implementation please refer to PR [#1037]
 With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads

-Now build `whisper.cpp` with CUDA support:
+Now build `whisper.cpp` with cuBLAS support:

 ```
 make clean
-WHISPER_CUDA=1 make -j
+WHISPER_CUBLAS=1 make -j
 ```

 ## OpenCL GPU support via CLBlast
@ -456,21 +455,6 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```

-## BLAS CPU support via Intel MKL
-
-Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
-First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
-
-Now build `whisper.cpp` with Intel MKL BLAS support:
-
-```
-source /opt/intel/oneapi/setvars.sh
-mkdir build
-cd build
-cmake -DWHISPER_MKL=ON ..
-WHISPER_MKL=1 make -j
-```
-
 ## Docker

 ### Prerequisites
@ -503,16 +487,6 @@ docker run -it --rm \
  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
 ```

-## Installing with Conan
-
-You can install pre-built binaries for whisper.cpp or build it from source using [Conan](https://conan.io/). Use the following command:
-
-```
-conan install --requires="whisper-cpp/[*]" --build=missing
-```
-
-For detailed instructions on how to use Conan, please refer to the [Conan documentation](https://docs.conan.io/2/).
-
 ## Limitations

 - Inference only
@ -721,7 +695,7 @@ The [main](examples/main) example provides support for output of karaoke-style m
 currently pronounced word is highlighted. Use the `-wts` argument and run the generated bash script.
 This requires to have `ffmpeg` installed.

-Here are a few _"typical"_ examples:
+Here are a few *"typical"* examples:

 ```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
@ -755,10 +729,10 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a

 ## Video comparison of different models

-Use the [scripts/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/scripts/bench-wts.sh) script to generate a video in the following format:
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:

 ```bash
-./scripts/bench-wts.sh samples/jfk.wav
+./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```

@ -779,7 +753,7 @@ Additionally a script to run whisper.cpp with different models and audio files i
 You can run it with the following command, by default it will run against any standard model in the models folder.

 ```bash
-python3 scripts/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
 ```

 It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
@ -819,7 +793,6 @@ For more details, see the conversion script [models/convert-pt-to-ggml.py](model
  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
 - [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
-  - [AIWintermuteAI/whispercpp](https://github.com/AIWintermuteAI/whispercpp) (Updated fork of aarnphm/whispercpp)
  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
 - [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
 - [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
--- a/bindings/go/examples/go-whisper/flags.go
+++ b/bindings/go/examples/go-whisper/flags.go
@ -68,6 +68,10 @@ func (flags *Flags) GetOut() string {
 	return strings.ToLower(flags.Lookup("out").Value.String())
 }

+func (flags *Flags) IsSpeedup() bool {
+	return flags.Lookup("speedup").Value.String() == "true"
+}
+
 func (flags *Flags) IsTokens() bool {
 	return flags.Lookup("tokens").Value.String() == "true"
 }
@ -107,6 +111,10 @@ func (flags *Flags) SetParams(context whisper.Context) error {
 		fmt.Fprintf(flags.Output(), "Setting duration to %v\n", duration)
 		context.SetDuration(duration)
 	}
+	if flags.IsSpeedup() {
+		fmt.Fprintf(flags.Output(), "Setting speedup to true\n")
+		context.SetSpeedup(true)
+	}
 	if threads := flags.GetThreads(); threads != 0 {
 		fmt.Fprintf(flags.Output(), "Setting threads to %d\n", threads)
 		context.SetThreads(threads)
@ -138,6 +146,7 @@ func registerFlags(flag *Flags) {
 	flag.Duration("offset", 0, "Time offset")
 	flag.Duration("duration", 0, "Duration of audio to process")
 	flag.Uint("threads", 0, "Number of threads to use")
+	flag.Bool("speedup", false, "Enable speedup")
 	flag.Uint("max-len", 0, "Maximum segment length in characters")
 	flag.Uint("max-tokens", 0, "Maximum tokens per segment")
 	flag.Float64("word-thold", 0, "Maximum segment score")
--- a/bindings/go/params.go
+++ b/bindings/go/params.go
@ -47,6 +47,10 @@ func (p *Params) SetPrintTimestamps(v bool) {
 	p.print_timestamps = toBool(v)
 }

+func (p *Params) SetSpeedup(v bool) {
+	p.speed_up = toBool(v)
+}
+
 // Set language id
 func (p *Params) SetLanguage(lang int) error {
 	if lang == -1 {
@ -173,6 +177,9 @@ func (p *Params) String() string {
 	if p.token_timestamps {
 		str += " token_timestamps"
 	}
+	if p.speed_up {
+		str += " speed_up"
+	}

 	return str + ">"
 }
--- a/bindings/go/pkg/whisper/context.go
+++ b/bindings/go/pkg/whisper/context.go
@ -76,6 +76,11 @@ func (context *context) SetTranslate(v bool) {
 	context.params.SetTranslate(v)
 }

+// Set speedup flag
+func (context *context) SetSpeedup(v bool) {
+	context.params.SetSpeedup(v)
+}
+
 func (context *context) SetSplitOnWord(v bool) {
 	context.params.SetSplitOnWord(v)
 }
--- a/bindings/go/pkg/whisper/interface.go
+++ b/bindings/go/pkg/whisper/interface.go
@ -41,6 +41,7 @@ type Context interface {
 	SetOffset(time.Duration)        // Set offset
 	SetDuration(time.Duration)      // Set duration
 	SetThreads(uint)                // Set number of threads to use
+	SetSpeedup(bool)                // Set speedup flag
 	SetSplitOnWord(bool)            // Set split on word flag
 	SetTokenThreshold(float32)      // Set timestamp token probability threshold
 	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
--- a/bindings/ios
+++ b/bindings/ios
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
@ -20,7 +20,7 @@ public interface WhisperCppJnaLibrary extends Library {
     * @return Whisper context on success, null on failure
     */
    Pointer whisper_init_from_file(String path_model);
-
+    
    /**
     * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
     * Because this function allocates memory for the params, the caller must call either:
@ -304,6 +304,14 @@ public interface WhisperCppJnaLibrary extends Library {
    /** Language id associated with the provided state */
    int whisper_full_lang_id_from_state(Pointer state);

+    /**
+     * Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
+     * The resulting spectrogram is stored inside the default state of the provided whisper context.
+     * @return 0 on success
+     */
+    int whisper_pcm_to_mel_phase_vocoder(Pointer ctx, final float[] samples, int n_samples, int n_threads);
+
+    int whisper_pcm_to_mel_phase_vocoder_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);

    /** Get the start time of the specified segment. */
    long whisper_full_get_segment_t0(Pointer ctx, int i_segment);
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java
@ -129,6 +129,14 @@ public class WhisperFullParams extends Structure {
    /** Maximum tokens per segment (0, default = no limit) */
    public int max_tokens;

+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public CBool speed_up;
+
+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public void speedUp(boolean enable) {
+        speed_up = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
    /** Overwrite the audio context size (0 = use default). */
    public int audio_ctx;

@ -140,9 +148,6 @@ public class WhisperFullParams extends Structure {
        tdrz_enable = enable ? CBool.TRUE : CBool.FALSE;
    }

-    /** Regular expression matching tokens to suppress. */
-    public String suppress_regex;
-
    /** Tokens to provide to the whisper decoder as an initial prompt.
     * These are prepended to any existing text context from a previous call. */
    public String initial_prompt;
@ -313,8 +318,8 @@ public class WhisperFullParams extends Structure {
        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
                "no_context", "single_segment", "no_timestamps",
                "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
-                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "audio_ctx",
-                "tdrz_enable", "suppress_regex", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "tdrz_enable", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
                "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
                "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",
                "new_segment_callback", "new_segment_callback_user_data",
--- a/bindings/javascript/package.json
+++ b/bindings/javascript/package.json
@ -1,6 +1,6 @@
 {
  "name": "whisper.cpp",
-  "version": "1.6.2",
+  "version": "1.5.4",
  "description": "Whisper speech recognition",
  "main": "whisper.js",
  "scripts": {
--- a/bindings/ruby/Rakefile
+++ b/bindings/ruby/Rakefile
@ -1,12 +0,0 @@
-require 'rake/clean'
-  require 'rubygems/package'
-
-desc 'Build gem'
-task :package do
-  spec_source = File.read File.join(File.dirname(__FILE__),'whispercpp.gemspec')
-  spec = nil
-  # see: http://gist.github.com/16215
-  Thread.new { spec = eval("#{spec_source}") }.join
-  spec.validate
-  Gem::Package.build(spec)
-end
--- a/bindings/ruby/ext/extconf.rb
+++ b/bindings/ruby/ext/extconf.rb
@ -1,7 +1,6 @@
 require 'mkmf'
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper-mel.hpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
@ -10,7 +9,6 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} ."
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-common.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
--- a/bindings/ruby/ext/ggml-backend-impl.h
+++ b/bindings/ruby/ext/ggml-backend-impl.h
@ -12,63 +12,31 @@ extern "C" {
    // Backend buffer
    //

-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
    typedef void * ggml_backend_buffer_context_t;

    struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
    };

    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
        ggml_backend_buffer_context_t context;
+
        size_t size;
-        enum ggml_backend_buffer_usage usage;
    };

-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
            struct ggml_backend_buffer_i           iface,
                   ggml_backend_buffer_context_t   context,
                   size_t                          size);

-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
    //
    // Backend
    //
@ -76,66 +44,44 @@ extern "C" {
    typedef void * ggml_backend_context_t;

    struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+        const char * (*get_name)(ggml_backend_t backend);

-        void (*GGML_CALL free)(ggml_backend_t backend);
+        void (*free)(ggml_backend_t backend);

        // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);

-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);

-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);

-        // compute graph with a plan (not used currently)
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);

        // compute graph with a plan
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);

        // check if the backend supports an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
    };

    struct ggml_backend {
-        ggml_guid_t guid;
-
        struct ggml_backend_i iface;
+
        ggml_backend_context_t context;
    };

-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
 #ifdef  __cplusplus
 }
 #endif
--- a/bindings/ruby/ext/ggml-backend.c
+++ b/bindings/ruby/ext/ggml-backend.c
--- a/bindings/ruby/ext/ggml-backend.h
+++ b/bindings/ruby/ext/ggml-backend.h
@ -7,123 +7,69 @@
 extern "C" {
 #endif

-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
    //
    // Backend buffer
    //

-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;

-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-    };
-
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+    // backend buffer functions
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);

    //
    // Backend
    //

-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+    struct ggml_backend;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
    GGML_API void         ggml_backend_free(ggml_backend_t backend);

-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);

-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);

-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);

-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);

-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);

-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
-
    //
    // CPU backend
    //

    GGML_API ggml_backend_t ggml_backend_cpu_init(void);

-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);

    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);

-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);

    //
    // Backend scheduler
@ -137,96 +83,53 @@ extern "C" {
    /*
      Example usage:

-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
+        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph

-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed

-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
+        // in build_graph:
+        build_graph(...) {
+            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
+            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
+            ggml_allocr_alloc(alloc_cpu, tensor);

-        // manually assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
+        }

-        ggml_backend_sched_reserve(sched, reserve_graph);
+        // allocate backend buffers from measure graph
+        ggml_backend_sched_init_measure(sched, measure_graph);
+
+        // the scheduler is now ready to compute graphs

        // compute
        graph = build_graph(sched);
        ggml_backend_sched_graph_compute(sched, graph);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
    */

    struct ggml_backend_sched;
    typedef struct ggml_backend_sched * ggml_backend_sched_t;

-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+
+    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);

    // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);

-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);

-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);

+    // Allocate a graph on the backend scheduler
+    GGML_API void ggml_backend_sched_graph_compute(
+            ggml_backend_sched_t sched,
+            struct ggml_cgraph * graph);

 #ifdef  __cplusplus
 }
--- a/bindings/ruby/ext/ggml-common.h
+++ b/bindings/ruby/ext/ggml-common.h
--- a/bindings/ruby/ext/ggml-cuda.h
+++ b/bindings/ruby/ext/ggml-cuda.h
@ -1,43 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-impl.h
+++ b/bindings/ruby/ext/ggml-impl.h
@ -5,7 +5,6 @@
 // GGML internal header

 #include <assert.h>
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
@ -19,7 +18,6 @@ extern "C" {
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef __cplusplus
 #ifndef static_assert
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
 #define static_assert(cond, msg) _Static_assert(cond, msg)
@ -27,7 +25,6 @@ extern "C" {
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
 #endif
-#endif

 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
@ -37,18 +34,17 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
 #endif

+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
@ -60,30 +56,14 @@ extern "C" {
 //
 #include <arm_neon.h>

-typedef __fp16 ggml_fp16_internal_t;
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)

-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)

 #else

-typedef uint16_t ggml_fp16_internal_t;
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@ -237,7 +217,8 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    uint16_t s;
    memcpy(&s, &f, sizeof(uint16_t));
@ -245,23 +226,19 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }

 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #endif

 #define GGML_HASHTABLE_FULL ((size_t)-1)
 #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)

-struct ggml_hash_set ggml_hash_set_new(size_t size);
-
 bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);

 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);

-// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);

 // return index, asserts if table is full
--- a/bindings/ruby/ext/ggml-kompute.h
+++ b/bindings/ruby/ext/ggml-kompute.h
@ -1,46 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-#ifdef __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-metal.h
+++ b/bindings/ruby/ext/ggml-metal.h
@ -1,66 +0,0 @@
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
-
-#ifdef __cplusplus
-}
-#endif
-
--- a/bindings/ruby/ext/ggml-opencl.h
+++ b/bindings/ruby/ext/ggml-opencl.h
@ -1,36 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-GGML_API void ggml_cl_init(void);
-
-GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
-GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-// GGML_API void * ggml_cl_host_malloc(size_t size);
-// GGML_API void   ggml_cl_host_free(void * ptr);
-
-GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-// backend API
-
-// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
-
-// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
-// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-quants.c
+++ b/bindings/ruby/ext/ggml-quants.c
--- a/bindings/ruby/ext/ggml-quants.h
+++ b/bindings/ruby/ext/ggml-quants.h
@ -1,133 +1,224 @@
 #pragma once

-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
+#include "ggml-impl.h"

 // GGML internal header

-#ifdef __cplusplus
-extern "C" {
+#include <stdint.h>
+#include <stddef.h>
+
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
 #endif

+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
 // Quantization
-void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);

-void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);

-void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);

-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);

 // Dequantization
-void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);

-void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);

 // Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
-
-#ifdef __cplusplus
-}
-#endif
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
--- a/bindings/ruby/ext/ggml-sycl.h
+++ b/bindings/ruby/ext/ggml-sycl.h
@ -1,49 +0,0 @@
-//
-//  MIT license
-//  Copyright (C) 2024 Intel Corporation
-//  SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
-
-// devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
-
-GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
-
-// SYCL doesn't support registering host memory, keep here for reference
-// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-vulkan.h
+++ b/bindings/ruby/ext/ggml-vulkan.h
@ -1,29 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_VK_NAME "Vulkan"
-#define GGML_VK_MAX_DEVICES 16
-
-GGML_API void ggml_vk_instance_init(void);
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
-
-GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ruby_whisper.cpp
+++ b/bindings/ruby/ext/ruby_whisper.cpp
@ -311,6 +311,12 @@ static VALUE ruby_whisper_params_get_split_on_word(VALUE self) {
 static VALUE ruby_whisper_params_set_split_on_word(VALUE self, VALUE value) {
  BOOL_PARAMS_SETTER(self, split_on_word, value)
 }
+static VALUE ruby_whisper_params_get_speed_up(VALUE self) {
+  BOOL_PARAMS_GETTER(self, speed_up)
+}
+static VALUE ruby_whisper_params_set_speed_up(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, speed_up, value)
+}
 static VALUE ruby_whisper_params_get_diarize(VALUE self) {
  ruby_whisper_params *rwp;
  Data_Get_Struct(self, ruby_whisper_params, rwp);
@ -402,6 +408,8 @@ void Init_whisper() {
  rb_define_method(cParams, "token_timestamps=", ruby_whisper_params_set_token_timestamps, 1);
  rb_define_method(cParams, "split_on_word", ruby_whisper_params_get_split_on_word, 0);
  rb_define_method(cParams, "split_on_word=", ruby_whisper_params_set_split_on_word, 1);
+  rb_define_method(cParams, "speed_up", ruby_whisper_params_get_speed_up, 0);
+  rb_define_method(cParams, "speed_up=", ruby_whisper_params_set_speed_up, 1);
  rb_define_method(cParams, "diarize", ruby_whisper_params_get_diarize, 0);
  rb_define_method(cParams, "diarize=", ruby_whisper_params_set_diarize, 1);

--- a/bindings/ruby/tests/test_whisper.rb
+++ b/bindings/ruby/tests/test_whisper.rb
@ -117,6 +117,13 @@ class TestWhisper < Test::Unit::TestCase
    assert !@params.split_on_word
  end

+  def test_speed_up
+    @params.speed_up = true
+    assert @params.speed_up
+    @params.speed_up = false
+    assert !@params.speed_up
+  end
+
  def test_whisper
    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
    params  = Whisper::Params.new
--- a/bindings/ruby/whispercpp.gemspec
+++ b/bindings/ruby/whispercpp.gemspec
@ -1,28 +0,0 @@
-Gem::Specification.new do |s|
-  s.name    = "whispercpp"
-  s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
-  s.version = '1.3.0'
-  s.date    = '2024-05-14'
-  s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
-  s.email   = 'todd.fisher@gmail.com'
-  s.extra_rdoc_files = ['LICENSE', 'README.md']
-  
-  s.files = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"]
-
-  #### Load-time details
-  s.require_paths = ['lib','ext']
-  s.summary = %q{Ruby whisper.cpp bindings}
-  s.test_files = ["tests/test_whisper.rb"]
-  
-  s.extensions << 'ext/extconf.rb'
-  
-
-  #### Documentation and testing.
-  s.homepage = 'https://github.com/ggerganov/whisper.cpp'
-  s.rdoc_options = ['--main', '../../README.md']
-
-  
-    s.platform = Gem::Platform::RUBY
-  
-  s.licenses = ['MIT']
-end
--- a/cmake/FindFFmpeg.cmake
+++ b/cmake/FindFFmpeg.cmake
@ -1,163 +0,0 @@
-# From
-# https://github.com/snikulov/cmake-modules/blob/master/FindFFmpeg.cmake
-#
-# vim: ts=2 sw=2
-# - Try to find the required ffmpeg components(default: AVFORMAT, AVUTIL, AVCODEC)
-#
-# Once done this will define
-#  FFMPEG_FOUND         - System has the all required components.
-#  FFMPEG_INCLUDE_DIRS  - Include directory necessary for using the required components headers.
-#  FFMPEG_LIBRARIES     - Link these to use the required ffmpeg components.
-#  FFMPEG_DEFINITIONS   - Compiler switches required for using the required ffmpeg components.
-#
-# For each of the components it will additionally set.
-#   - AVCODEC
-#   - AVDEVICE
-#   - AVFORMAT
-#   - AVFILTER
-#   - AVUTIL
-#   - POSTPROC
-#   - SWSCALE
-# the following variables will be defined
-#  <component>_FOUND        - System has <component>
-#  <component>_INCLUDE_DIRS - Include directory necessary for using the <component> headers
-#  <component>_LIBRARIES    - Link these to use <component>
-#  <component>_DEFINITIONS  - Compiler switches required for using <component>
-#  <component>_VERSION      - The components version
-#
-# Copyright (c) 2006, Matthias Kretz, <kretz@kde.org>
-# Copyright (c) 2008, Alexander Neundorf, <neundorf@kde.org>
-# Copyright (c) 2011, Michael Jansen, <kde@michael-jansen.biz>
-#
-# Redistribution and use is allowed according to the terms of the BSD license.
-# For details see the accompanying COPYING-CMAKE-SCRIPTS file.
-
-include(FindPackageHandleStandardArgs)
-
-# The default components were taken from a survey over other FindFFMPEG.cmake files
-if (NOT FFmpeg_FIND_COMPONENTS)
-  set(FFmpeg_FIND_COMPONENTS AVFORMAT AVCODEC AVUTIL SWRESAMPLE) 
-endif()
-
-#
-### Macro: set_component_found
-#
-# Marks the given component as found if both *_LIBRARIES AND *_INCLUDE_DIRS is present.
-#
-macro(set_component_found _component )
-  if (${_component}_LIBRARIES AND ${_component}_INCLUDE_DIRS)
-    message(DEBUG "  - ${_component} found.")
-    set(${_component}_FOUND TRUE)
-  else ()
-  message(DEBUG "  - ${_component} not found.")
-  endif ()
-endmacro()
-
-#
-### Macro: find_component
-#
-# Checks for the given component by invoking pkgconfig and then looking up the libraries and
-# include directories.
-#
-macro(find_component _component _pkgconfig _library _header)
-
-  if (NOT WIN32)
-     # use pkg-config to get the directories and then use these values
-     # in the FIND_PATH() and FIND_LIBRARY() calls
-     find_package(PkgConfig)
-     if (PKG_CONFIG_FOUND)
-       pkg_check_modules(PC_${_component} ${_pkgconfig})
-       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDEDIR}")
-       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDE_DIRS}")
-       message(STATUS "${PC_${_component}_CFLAGS}")
-     endif ()
-  endif (NOT WIN32)
-
-
-  find_path(${_component}_INCLUDE_DIRS ${_header}
-    HINTS
-      ${PC_${_component}_INCLUDEDIR}
-      ${PC_${_component}_INCLUDE_DIRS}
-    PATH_SUFFIXES
-      ffmpeg
-  )
-
-  # CMake's default is to search first for shared libraries and then for static libraries.
-  # Todo later: add option to prefer static libs over dynamic:
-  find_library(${_component}_LIBRARIES NAMES ${_library} lib${_library}.a  
-      HINTS
-      ${PC_${_component}_LIBDIR}
-      ${PC_${_component}_LIBRARY_DIRS}
-  )
-
-  set(${_component}_DEFINITIONS  ${PC_${_component}_CFLAGS_OTHER} CACHE STRING "The ${_component} CFLAGS.")
-  set(${_component}_VERSION      ${PC_${_component}_VERSION}      CACHE STRING "The ${_component} version number.")
-
-  set_component_found(${_component})
-
-  mark_as_advanced(
-    ${_component}_INCLUDE_DIRS
-    ${_component}_LIBRARIES
-    ${_component}_DEFINITIONS
-    ${_component}_VERSION)
-
-endmacro()
-
-
-# Check for cached results. If there are skip the costly part.
-if (NOT FFMPEG_LIBRARIES)
-
-  # Check for all possible component.
-  find_component(AVCODEC    libavcodec    avcodec  libavcodec/avcodec.h)
-  find_component(AVFORMAT   libavformat   avformat libavformat/avformat.h)
-  find_component(AVDEVICE   libavdevice   avdevice libavdevice/avdevice.h)
-  #find_component(AVRESAMPLE libavresample avresample libavresample/avresample.h) # old name for swresample
-  find_component(AVUTIL     libavutil     avutil   libavutil/avutil.h)
-  find_component(AVFILTER   libavfilter   avfilter libavfilter/avfilter.h)
-  find_component(SWSCALE    libswscale    swscale  libswscale/swscale.h)
-  find_component(POSTPROC   libpostproc   postproc libpostproc/postprocess.h)
-  find_component(SWRESAMPLE libswresample swresample libswresample/swresample.h)
-
-  # Check if the required components were found and add their stuff to the FFMPEG_* vars.
-  foreach (_component ${FFmpeg_FIND_COMPONENTS})
-    if (${_component}_FOUND)
-      # message(STATUS "Required component ${_component} present.")
-      set(FFMPEG_LIBRARIES   ${FFMPEG_LIBRARIES}   ${${_component}_LIBRARIES})
-      set(FFMPEG_DEFINITIONS ${FFMPEG_DEFINITIONS} ${${_component}_DEFINITIONS})
-      list(APPEND FFMPEG_INCLUDE_DIRS ${${_component}_INCLUDE_DIRS})
-    else ()
-      # message(STATUS "Required component ${_component} missing.")
-    endif ()
-  endforeach ()
-
-  # Build the include path with duplicates removed.
-  if (FFMPEG_INCLUDE_DIRS)
-    list(REMOVE_DUPLICATES FFMPEG_INCLUDE_DIRS)
-  endif ()
-
-  # cache the vars.
-  set(FFMPEG_INCLUDE_DIRS ${FFMPEG_INCLUDE_DIRS} CACHE STRING "The FFmpeg include directories." FORCE)
-  set(FFMPEG_LIBRARIES    ${FFMPEG_LIBRARIES}    CACHE STRING "The FFmpeg libraries." FORCE)
-  set(FFMPEG_DEFINITIONS  ${FFMPEG_DEFINITIONS}  CACHE STRING "The FFmpeg cflags." FORCE)
-
-  mark_as_advanced(FFMPEG_INCLUDE_DIRS
-                   FFMPEG_LIBRARIES
-                   FFMPEG_DEFINITIONS)
-
-endif ()
-
-# Now set the noncached _FOUND vars for the components.
-# whisper.cpp does not need SWSCALE
-foreach (_component AVCODEC AVDEVICE AVFORMAT AVRESAMPLE AVUTIL POSTPROCESS)
-  set_component_found(${_component})
-endforeach ()
-
-# Compile the list of required vars
-set(_FFmpeg_REQUIRED_VARS FFMPEG_LIBRARIES FFMPEG_INCLUDE_DIRS)
-foreach (_component ${FFmpeg_FIND_COMPONENTS})
-  list(APPEND _FFmpeg_REQUIRED_VARS ${_component}_LIBRARIES ${_component}_INCLUDE_DIRS)
-endforeach ()
-
-# Give a nice error message if some of the required vars are missing.
-find_package_handle_standard_args(FFmpeg DEFAULT_MSG ${_FFmpeg_REQUIRED_VARS})
-
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@ -22,18 +22,12 @@ endif()

 set(TARGET common)

-if (WHISPER_FFMPEG)
-    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
-endif()
-
 add_library(${TARGET} STATIC
    common.h
    common.cpp
    common-ggml.h
    common-ggml.cpp
-    grammar-parser.h
    grammar-parser.cpp
-    ${COMMON_SOURCES_FFMPEG}
    )

 include(DefaultTargetOptions)
@ -41,7 +35,6 @@ include(DefaultTargetOptions)
 target_link_libraries(${TARGET} PRIVATE whisper)

 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")

 if (WHISPER_SDL2)
    # common-sdl
@ -59,12 +52,10 @@ if (WHISPER_SDL2)
    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})

    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()

 # add json lib
-add_library(json_cpp INTERFACE)
-target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
+add_library(json_cpp INTERFACE json.hpp)

 # examples

@ -72,50 +63,25 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})

 if (EMSCRIPTEN)
    add_subdirectory(whisper.wasm)
-    set_target_properties(libmain PROPERTIES FOLDER "libs")
    add_subdirectory(stream.wasm)
-    set_target_properties(libstream PROPERTIES FOLDER "libs")
    add_subdirectory(command.wasm)
-    set_target_properties(libcommand PROPERTIES FOLDER "libs")
    add_subdirectory(talk.wasm)
-    set_target_properties(libtalk PROPERTIES FOLDER "libs")
    add_subdirectory(bench.wasm)
-    set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
    add_subdirectory(addon.node)
-    set_target_properties(addon.node PROPERTIES FOLDER "examples")
 else()
    add_subdirectory(main)
-    set_target_properties(main PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    add_subdirectory(stream)
-    set_target_properties(stream PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
    add_subdirectory(server)
-    set_target_properties(server PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    add_subdirectory(command)
-    set_target_properties(command PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
    add_subdirectory(bench)
-    set_target_properties(bench PROPERTIES FOLDER "examples")
    add_subdirectory(quantize)
-    set_target_properties(quantize PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    add_subdirectory(talk)
-    set_target_properties(talk PROPERTIES FOLDER "examples")
    add_subdirectory(talk-llama)
-    set_target_properties(talk-llama PROPERTIES FOLDER "examples")
    add_subdirectory(lsp)
-    set_target_properties(lsp PROPERTIES FOLDER "examples")
    if (LLAMA_SYCL)
        add_subdirectory(sycl)
-        set_target_properties(sycl PROPERTIES FOLDER "examples")
    endif()
-endif (WHISPER_SDL2)
 endif()

-if (WHISPER_SDL2)
-    add_subdirectory(wchess)
-    set_target_properties(wchess PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
+add_subdirectory(wchess)
--- a/examples/addon.node/CMakeLists.txt
+++ b/examples/addon.node/CMakeLists.txt
@ -1,4 +1,4 @@
-set(TARGET addon.node)
+set(TARGET whisper-addon)

 # Base settings
 #==================================================================
--- a/examples/addon.node/README.md
+++ b/examples/addon.node/README.md
@ -14,14 +14,14 @@ npm install
 Make sure it is in the project root directory and compiled with make-js.

 ```shell
-npx cmake-js compile -T addon.node -B Release
+npx cmake-js compile -T whisper-addon -B Release
 ```

 For Electron addon and cmake-js options, you can see [cmake-js](https://github.com/cmake-js/cmake-js) and make very few configuration changes.

 > Such as appointing special cmake path:
 > ```shell
-> npx cmake-js compile -c 'xxx/cmake' -T addon.node -B Release
+> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -B Release
 > ```

 ## Run
--- a/examples/addon.node/test/whisper.spec.js
+++ b/examples/addon.node/test/whisper.spec.js
@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
  __dirname,
-  "../../../build/Release/addon.node"
+  "../../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");

@ -12,12 +12,6 @@ const whisperParamsMock = {
  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
  use_gpu: true,
-  flash_attn: false,
-  no_prints: true,
-  comma_in_time: false,
-  translate: true,
-  no_timestamps: false,
-  audio_ctx: 0,
 };

 describe("Run whisper.node", () => {
--- a/examples/addon.node/addon.cpp
+++ b/examples/addon.node/addon.cpp
@ -19,12 +19,12 @@ struct whisper_params {
    int32_t max_len      = 0;
    int32_t best_of      = 5;
    int32_t beam_size    = -1;
-    int32_t audio_ctx    = 0;

    float word_thold    = 0.01f;
    float entropy_thold = 2.4f;
    float logprob_thold = -1.0f;

+    bool speed_up       = false;
    bool translate      = false;
    bool diarize        = false;
    bool output_txt     = false;
@ -36,10 +36,7 @@ struct whisper_params {
    bool print_colors   = false;
    bool print_progress = false;
    bool no_timestamps  = false;
-    bool no_prints      = false;
    bool use_gpu        = true;
-    bool flash_attn     = false;
-    bool comma_in_time  = true;

    std::string language = "en";
    std::string prompt;
@ -47,8 +44,6 @@ struct whisper_params {

    std::vector<std::string> fname_inp = {};
    std::vector<std::string> fname_out = {};
-
-    std::vector<float> pcmf32 = {}; // mono-channel F32 PCM
 };

 struct whisper_print_user_data {
@ -125,15 +120,9 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
    }
 }

-void cb_log_disable(enum ggml_log_level, const char *, void *) {}
-
 int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
-    if (params.no_prints) {
-        whisper_log_set(cb_log_disable, NULL);
-    }
-
-    if (params.fname_inp.empty() && params.pcmf32.empty()) {
-        fprintf(stderr, "error: no input files or audio buffer specified\n");
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
        return 2;
    }

@ -146,7 +135,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {

    struct whisper_context_params cparams = whisper_context_default_params();
    cparams.use_gpu = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

    if (ctx == nullptr) {
@ -154,14 +142,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        return 3;
    }

-    // if params.pcmf32 is provided, set params.fname_inp to "buffer"
-    // this is simpler than further modifications in the code
-    if (!params.pcmf32.empty()) {
-        fprintf(stderr, "info: using audio buffer as input\n");
-        params.fname_inp.clear();
-        params.fname_inp.emplace_back("buffer");
-    }
-
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
        const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -169,25 +149,20 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        std::vector<float> pcmf32; // mono-channel F32 PCM
        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM

-        // read the input audio file if params.pcmf32 is not provided
-        if (params.pcmf32.empty()) {
-            if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-                fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
-                continue;
-            }
-        } else {
-            pcmf32 = params.pcmf32;
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
        }

        // print system information
-        if (!params.no_prints) {
+        {
            fprintf(stderr, "\n");
            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
        }

        // print some info about the processing
-        if (!params.no_prints) {
+        {
            fprintf(stderr, "\n");
            if (!whisper_is_multilingual(ctx)) {
                if (params.language != "en" || params.translate) {
@ -196,13 +171,12 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                }
            }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d, audio_ctx = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
                    __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                    params.n_threads, params.n_processors,
                    params.language.c_str(),
                    params.translate ? "translate" : "transcribe",
-                    params.no_timestamps ? 0 : 1,
-                    params.audio_ctx);
+                    params.no_timestamps ? 0 : 1);

            fprintf(stderr, "\n");
        }
@ -229,15 +203,14 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
            wparams.entropy_thold    = params.entropy_thold;
            wparams.logprob_thold    = params.logprob_thold;
            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
-            wparams.audio_ctx        = params.audio_ctx;
+
+            wparams.speed_up         = params.speed_up;

            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;

            wparams.initial_prompt   = params.prompt.c_str();

-            wparams.no_timestamps    = params.no_timestamps;
-
            whisper_print_user_data user_data = { &params, &pcmf32s };

            // this callback is called on each new segment
@ -273,8 +246,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

-        result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
-        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
+        result[i].emplace_back(to_timestamp(t0, true));
+        result[i].emplace_back(to_timestamp(t1, true));
        result[i].emplace_back(text);
    }

@ -325,33 +298,11 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
  std::string model = whisper_params.Get("model").As<Napi::String>();
  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
-  bool flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
-  bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
-  bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
-  int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
-  bool comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
-
-  Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
-  std::vector<float> pcmf32_vec;
-  if (pcmf32Value.IsTypedArray()) {
-    Napi::Float32Array pcmf32 = pcmf32Value.As<Napi::Float32Array>();
-    size_t length = pcmf32.ElementLength();
-    pcmf32_vec.reserve(length);
-    for (size_t i = 0; i < length; i++) {
-      pcmf32_vec.push_back(pcmf32[i]);
-    }
-  }

  params.language = language;
  params.model = model;
  params.fname_inp.emplace_back(input);
  params.use_gpu = use_gpu;
-  params.flash_attn = flash_attn;
-  params.no_prints = no_prints;
-  params.no_timestamps = no_timestamps;
-  params.audio_ctx = audio_ctx;
-  params.pcmf32 = pcmf32_vec;
-  params.comma_in_time = comma_in_time;

  Napi::Function callback = info[1].As<Napi::Function>();
  Worker* worker = new Worker(callback, params);
--- a/examples/addon.node/index.js
+++ b/examples/addon.node/index.js
@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
  __dirname,
-  "../../build/Release/addon.node"
+  "../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");

@ -10,27 +10,15 @@ const whisperAsync = promisify(whisper);
 const whisperParams = {
  language: "en",
  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
-  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  fname_inp: "../../samples/jfk.wav",
  use_gpu: true,
-  flash_attn: false,
-  no_prints: true,
-  comma_in_time: false,
-  translate: true,
-  no_timestamps: false,
-  audio_ctx: 0,
 };

 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
  arguments.reduce((pre, item) => {
    if (item.startsWith("--")) {
-      const [key, value] = item.slice(2).split("=");
-      if (key === "audio_ctx") {
-        whisperParams[key] = parseInt(value);
-      } else {
-        whisperParams[key] = value;
-      }
-      return pre;
+      return [...pre, item.slice(2).split("=")];
    }
    return pre;
  }, [])
@ -45,6 +33,5 @@ for (const key in params) {
 console.log("whisperParams =", whisperParams);

 whisperAsync(whisperParams).then((result) => {
-  console.log();
-  console.log(result);
+  console.log(`Result from whisper: ${result}`);
 });
--- a/examples/addon.node/package.json
+++ b/examples/addon.node/package.json
@ -1,5 +1,5 @@
 {
-  "name": "addon.node",
+  "name": "whisper-addon",
  "version": "0.0.0",
  "description": "",
  "main": "index.js",
--- a/examples/bench/bench.cpp
+++ b/examples/bench/bench.cpp
@ -12,8 +12,7 @@ struct whisper_params {

    std::string model = "models/ggml-base.en.bin";

-    bool use_gpu    = true;
-    bool flash_attn = false;
+    bool use_gpu = true;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@ -26,11 +25,10 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
            whisper_print_usage(argc, argv, params);
            exit(0);
        }
-        else if (arg == "-t"  || arg == "--threads")    { params.n_threads  = std::stoi(argv[++i]); }
-        else if (arg == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
-        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(argv[++i]); }
-        else if (arg == "-ng" || arg == "--no-gpu")     { params.use_gpu    = false; }
-        else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
+        else if (arg == "-t"  || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")    { params.what      = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")  { params.use_gpu   = false; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -51,7 +49,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
@ -62,9 +59,7 @@ int whisper_bench_full(const whisper_params & params) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

--- a/examples/command/README.md
+++ b/examples/command/README.md
@ -37,13 +37,9 @@ https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9
 The `command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:

 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/command/command.cpp
+++ b/examples/command/command.cpp
@ -38,12 +38,12 @@ struct whisper_params {

    grammar_parser::parse_state grammar_parsed;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -52,9 +52,6 @@ struct whisper_params {
    std::string prompt;
    std::string context;
    std::string grammar;
-
-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@ -75,11 +72,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
@ -88,7 +85,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
-        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -113,11 +109,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@ -126,7 +122,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
-    fprintf(stderr, "  --suppress-regex REGEX       [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
    fprintf(stderr, "\n");
 }

@ -162,6 +157,7 @@ std::string transcribe(
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx = params.audio_ctx;
+    wparams.speed_up  = params.speed_up;

    wparams.temperature     = 0.4f;
    wparams.temperature_inc = 1.0f;
@ -171,8 +167,6 @@ std::string transcribe(

    wparams.initial_prompt = params.context.data();

-    wparams.suppress_regex = params.suppress_regex.c_str();
-
    const auto & grammar_parsed = params.grammar_parsed;
    auto grammar_rules = grammar_parsed.c_rules();

@ -367,6 +361,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
            wparams.n_threads        = params.n_threads;

            wparams.audio_ctx        = params.audio_ctx;
+            wparams.speed_up         = params.speed_up;

            wparams.prompt_tokens    = k_tokens.data();
            wparams.prompt_n_tokens  = k_tokens.size();
@ -694,9 +689,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

--- a/examples/common-ggml.cpp
+++ b/examples/common-ggml.cpp
@ -70,8 +70,6 @@ bool ggml_common_quantize_0(
        case GGML_FTYPE_MOSTLY_IQ1_S:
        case GGML_FTYPE_MOSTLY_IQ4_NL:
        case GGML_FTYPE_MOSTLY_IQ4_XS:
-        case GGML_FTYPE_MOSTLY_IQ1_M:
-        case GGML_FTYPE_MOSTLY_BF16:
                {
                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                    return false;
@ -195,8 +193,6 @@ bool ggml_common_quantize_0(
                case GGML_TYPE_I8:
                case GGML_TYPE_I16:
                case GGML_TYPE_I32:
-                case GGML_TYPE_I64:
-                case GGML_TYPE_F64:
                case GGML_TYPE_Q8_1:
                case GGML_TYPE_Q8_K:
                case GGML_TYPE_IQ2_XXS:
@ -207,8 +203,6 @@ bool ggml_common_quantize_0(
                case GGML_TYPE_IQ1_S:
                case GGML_TYPE_IQ4_NL:
                case GGML_TYPE_IQ4_XS:
-                case GGML_TYPE_IQ1_M:
-                case GGML_TYPE_BF16:
                case GGML_TYPE_COUNT:
                    {
                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
--- a/examples/common.cpp
+++ b/examples/common.cpp
@ -19,16 +19,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

-#ifdef _WIN32
-#include <fcntl.h>
-#include <io.h>
-#endif
-
-#ifdef WHISPER_FFMPEG
-// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
-extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
-#endif
-
 // Function to check if the next argument exists
 std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
    if (i + 1 < argc && argv[i + 1][0] != '-') {
@ -642,14 +632,10 @@ bool is_wav_buffer(const std::string buf) {

 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
    drwav wav;
-    std::vector<uint8_t> wav_data; // used for pipe input from stdin or ffmpeg decoding output
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin

    if (fname == "-") {
        {
-            #ifdef _WIN32
-            _setmode(_fileno(stdin), _O_BINARY);
-            #endif
-
            uint8_t buf[1024];
            while (true)
            {
@ -675,42 +661,27 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
        }
    }
    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
-#if defined(WHISPER_FFMPEG)
-        if (ffmpeg_decode_audio(fname, wav_data) != 0) {
-            fprintf(stderr, "error: failed to ffmpeg decode '%s' \n", fname.c_str());
-            return false;
-        }
-        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-            fprintf(stderr, "error: failed to read wav data as wav \n");
-            return false;
-        }
-#else
        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
        return false;
-#endif
    }

    if (wav.channels != 1 && wav.channels != 2) {
        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

    if (stereo && wav.channels != 2) {
        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
-        drwav_uninit(&wav);
        return false;
    }

    if (wav.bitsPerSample != 16) {
        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

--- a/examples/common.h
+++ b/examples/common.h
@ -185,7 +185,7 @@ private:
    // It is assumed that PCM data is normalized to a range from -1 to 1
    bool write_audio(const float * data, size_t length) {
        for (size_t i = 0; i < length; ++i) {
-            const int16_t intSample = int16_t(data[i] * 32767);
+            const int16_t intSample = data[i] * 32767;
            file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
            dataSize += sizeof(int16_t);
        }
--- a/examples/ffmpeg-transcode.cpp
+++ b/examples/ffmpeg-transcode.cpp
@ -1,350 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-
-/*
- * transcode.c - convert audio file to WAVE
- *
- * Copyright (C) 2019		Andrew Clayton <andrew@digital-domain.net>
- * Copyright (C) 2024       William Tambellini <william.tambellini@gmail.com>
- */
-
-// Just for conveninent C++ API
-#include <vector>
-#include <string>
-
-// C
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <stdbool.h>
-#include <stdint.h>
-#include <sys/types.h>
-#include <sys/stat.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <sys/mman.h>
-
-extern "C" {
-#include <libavutil/opt.h>
-#include <libavcodec/avcodec.h>
-#include <libavformat/avformat.h>
-#include <libswresample/swresample.h>
-}
-
-typedef uint64_t u64;
-typedef int64_t  s64;
-typedef uint32_t u32;
-typedef int32_t  s32;
-typedef uint16_t u16;
-typedef int16_t  s16;
-typedef uint8_t   u8;
-typedef int8_t    s8;
-
-#define WAVE_SAMPLE_RATE	16000
-#define AVIO_CTX_BUF_SZ		 4096
-
-static const char* ffmpegLog = getenv("FFMPEG_LOG");
-// Todo: add __FILE__ __LINE__
-#define LOG(...) \
-  do { if (ffmpegLog) fprintf(stderr, __VA_ARGS__); } while(0) // C99
-
-/*
- * WAVE file header based on definition from
- * https://gist.github.com/Jon-Schneider/8b7c53d27a7a13346a643dac9c19d34f
- *
- * We must ensure this structure doesn't have any holes or
- * padding so we can just map it straight to the WAVE data.
- */
-struct wave_hdr {
-	/* RIFF Header: "RIFF" */
-	char riff_header[4];
-	/* size of audio data + sizeof(struct wave_hdr) - 8 */
-	int wav_size;
-	/* "WAVE" */
-	char wav_header[4];
-
-	/* Format Header */
-	/* "fmt " (includes trailing space) */
-	char fmt_header[4];
-	/* Should be 16 for PCM */
-	int fmt_chunk_size;
-	/* Should be 1 for PCM. 3 for IEEE Float */
-	s16 audio_format;
-	s16 num_channels;
-	int sample_rate;
-	/*
-	 * Number of bytes per second
-	 * sample_rate * num_channels * bit_depth/8
-	 */
-	int byte_rate;
-	/* num_channels * bytes per sample */
-	s16 sample_alignment;
-	/* bits per sample */
-	s16 bit_depth;
-
-	/* Data Header */
-	/* "data" */
-	char data_header[4];
-	/*
-	 * size of audio
-	 * number of samples * num_channels * bit_depth/8
-	 */
-	int data_bytes;
-} __attribute__((__packed__));
-
-struct audio_buffer {
-	u8 *ptr;
-	int size; /* size left in the buffer */
-};
-
-static void set_wave_hdr(wave_hdr& wh, size_t size) {
-    memcpy(&wh.riff_header, "RIFF", 4);
-    wh.wav_size = size + sizeof(struct wave_hdr) - 8;
-    memcpy(&wh.wav_header, "WAVE", 4);
-    memcpy(&wh.fmt_header, "fmt ", 4);
-    wh.fmt_chunk_size = 16;
-    wh.audio_format = 1;
-    wh.num_channels = 1;
-    wh.sample_rate = WAVE_SAMPLE_RATE;
-    wh.sample_alignment = 2;
-    wh.bit_depth = 16;
-    wh.byte_rate = wh.sample_rate * wh.sample_alignment;
-    memcpy(&wh.data_header, "data", 4);
-    wh.data_bytes = size;
-}
-
-static void write_wave_hdr(int fd, size_t size) {
-	struct wave_hdr wh;
-    set_wave_hdr(wh, size);
-	write(fd, &wh, sizeof(struct wave_hdr));
-}
-
-static int map_file(int fd, u8 **ptr, size_t *size)
-{
-	struct stat sb;
-
-	fstat(fd, &sb);
-	*size = sb.st_size;
-
-    *ptr = (u8*)mmap(NULL, *size, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0);
-	if (*ptr == MAP_FAILED) {
-		perror("mmap");
-		return -1;
-	}
-
-	return 0;
-}
-
-static int read_packet(void *opaque, u8 *buf, int buf_size)
-{
-    struct audio_buffer *audio_buf = (audio_buffer*)opaque;
-
-	buf_size = FFMIN(buf_size, audio_buf->size);
-
-	/* copy internal buffer data to buf */
-	memcpy(buf, audio_buf->ptr, buf_size);
-	audio_buf->ptr += buf_size;
-	audio_buf->size -= buf_size;
-
-	return buf_size;
-}
-
-static void convert_frame(struct SwrContext *swr, AVCodecContext *codec,
-			  AVFrame *frame, s16 **data, int *size, bool flush)
-{
-	int nr_samples;
-	s64 delay;
-	u8 *buffer;
-
-	delay = swr_get_delay(swr, codec->sample_rate);
-	nr_samples = av_rescale_rnd(delay + frame->nb_samples,
-				    WAVE_SAMPLE_RATE, codec->sample_rate,
-				    AV_ROUND_UP);
-	av_samples_alloc(&buffer, NULL, 1, nr_samples, AV_SAMPLE_FMT_S16, 0);
-
-	/*
-	 * !flush is used to check if we are flushing any remaining
-	 * conversion buffers...
-	 */
-	nr_samples = swr_convert(swr, &buffer, nr_samples,
-				 !flush ? (const u8 **)frame->data : NULL,
-				 !flush ? frame->nb_samples : 0);
-
-    *data = (s16*)realloc(*data, (*size + nr_samples) * sizeof(s16));
-	memcpy(*data + *size, buffer, nr_samples * sizeof(s16));
-	*size += nr_samples;
-	av_freep(&buffer);
-}
-
-static bool is_audio_stream(const AVStream *stream)
-{
-	if (stream->codecpar->codec_type == AVMEDIA_TYPE_AUDIO)
-		return true;
-
-	return false;
-}
-
-// Return non zero on error, 0 on success
-// audio_buffer: input memory
-// data: decoded output audio data (wav file)
-// size: size of output data
-static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
-{
-    LOG("decode_audio: input size: %d\n", audio_buf->size);
-	AVFormatContext *fmt_ctx;
-	AVIOContext *avio_ctx;
-	AVStream *stream;
-	AVCodecContext *codec;
-	AVPacket packet;
-	AVFrame *frame;
-	struct SwrContext *swr;
-	u8 *avio_ctx_buffer;
-	unsigned int i;
-	int stream_index = -1;
-	int err;
-    const size_t errbuffsize = 1024;
-    char errbuff[errbuffsize];
-
-    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
-
-    fmt_ctx = avformat_alloc_context();
-    avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
-    LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);
-    avio_ctx = avio_alloc_context(avio_ctx_buffer, AVIO_CTX_BUF_SZ, 0, audio_buf, &read_packet, NULL, NULL);
-	fmt_ctx->pb = avio_ctx;
-
-    // open the input stream and read header
-	err = avformat_open_input(&fmt_ctx, NULL, NULL, NULL);
-	if (err) {
-        LOG("Could not read audio buffer: %d: %s\n", err, av_make_error_string(errbuff, errbuffsize, err));
-        return err;
-	}
-
-	err = avformat_find_stream_info(fmt_ctx, NULL);
-	if (err < 0) {
-        LOG("Could not retrieve stream info from audio buffer: %d\n", err);
-        return err;
-	}
-
-	for (i = 0; i < fmt_ctx->nb_streams; i++) {
-		if (is_audio_stream(fmt_ctx->streams[i])) {
-			stream_index = i;
-			break;
-		}
-	}
-
-	if (stream_index == -1) {
-        LOG("Could not retrieve audio stream from buffer\n");
-		return -1;
-	}
-
-	stream = fmt_ctx->streams[stream_index];
-	codec = avcodec_alloc_context3(
-			avcodec_find_decoder(stream->codecpar->codec_id));
-	avcodec_parameters_to_context(codec, stream->codecpar);
-	err = avcodec_open2(codec, avcodec_find_decoder(codec->codec_id),
-							NULL);
-	if (err) {
-        LOG("Failed to open decoder for stream #%d in audio buffer\n", stream_index);
-        return err;
-	}
-
-	/* prepare resampler */
-	swr = swr_alloc();
-
-	av_opt_set_int(swr, "in_channel_count", codec->channels, 0);
-	av_opt_set_int(swr, "out_channel_count", 1, 0);
-	av_opt_set_int(swr, "in_channel_layout", codec->channel_layout, 0);
-	av_opt_set_int(swr, "out_channel_layout", AV_CH_LAYOUT_MONO, 0);
-	av_opt_set_int(swr, "in_sample_rate", codec->sample_rate, 0);
-	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
-	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
-	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
-
-	swr_init(swr);
-	if (!swr_is_initialized(swr)) {
-        LOG("Resampler has not been properly initialized\n");
-		return -1;
-	}
-
-	av_init_packet(&packet);
-	frame = av_frame_alloc();
-	if (!frame) {
-        LOG("Error allocating the frame\n");
-		return -1;
-	}
-
-	/* iterate through frames */
-	*data = NULL;
-	*size = 0;
-	while (av_read_frame(fmt_ctx, &packet) >= 0) {
-		avcodec_send_packet(codec, &packet);
-
-		err = avcodec_receive_frame(codec, frame);
-		if (err == AVERROR(EAGAIN))
-			continue;
-
-		convert_frame(swr, codec, frame, data, size, false);
-	}
-	/* Flush any remaining conversion buffers... */
-	convert_frame(swr, codec, frame, data, size, true);
-
-	av_frame_free(&frame);
-	swr_free(&swr);
-    //avio_context_free(); // todo?
-	avcodec_close(codec);
-	avformat_close_input(&fmt_ctx);
-	avformat_free_context(fmt_ctx);
-
-	if (avio_ctx) {
-		av_freep(&avio_ctx->buffer);
-		av_freep(&avio_ctx);
-	}
-
-	return 0;
-}
-
-// in mem decoding/conversion/resampling:
-// ifname: input file path
-// owav_data: in mem wav file. Can be forwarded as it to whisper/drwav
-// return 0 on success
-int ffmpeg_decode_audio(const std::string &ifname, std::vector<uint8_t>& owav_data) {
-    LOG("ffmpeg_decode_audio: %s\n", ifname.c_str());
-    int ifd = open(ifname.c_str(), O_RDONLY);
-    if (ifd == -1) {
-        fprintf(stderr, "Couldn't open input file %s\n", ifname.c_str());
-        return -1;
-    }
-    u8 *ibuf = NULL;
-    size_t ibuf_size;
-    int err = map_file(ifd, &ibuf, &ibuf_size);
-    if (err) {
-        LOG("Couldn't map input file %s\n", ifname.c_str());
-        return err;
-    }
-    LOG("Mapped input file: %x size: %d\n", ibuf, ibuf_size);
-    struct audio_buffer inaudio_buf;
-    inaudio_buf.ptr = ibuf;
-    inaudio_buf.size = ibuf_size;
-
-    s16 *odata=NULL;
-    int osize=0;
-
-    err = decode_audio(&inaudio_buf, &odata, &osize);
-    LOG("decode_audio returned %d \n", err);
-    if (err != 0) {
-        LOG("decode_audio failed\n");
-        return err;
-    }
-    LOG("decode_audio output size: %d\n", osize);
-
-    wave_hdr wh;
-    const size_t outdatasize = osize * sizeof(s16);
-    set_wave_hdr(wh, outdatasize);
-    owav_data.resize(sizeof(wave_hdr) + outdatasize);
-    // header:
-    memcpy(owav_data.data(), &wh, sizeof(wave_hdr));
-    // the data:
-    memcpy(owav_data.data() + sizeof(wave_hdr), odata, osize* sizeof(s16));
-
-    return 0;
-}
--- a/examples/grammar-parser.cpp
+++ b/examples/grammar-parser.cpp
@ -190,7 +190,7 @@ namespace grammar_parser {
                pos = parse_space(pos + 1, is_nested);
            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
                }

                // apply transformation to previous symbol (last_sym_start to end) according to
--- a/examples/helpers.js
+++ b/examples/helpers.js
@ -34,6 +34,9 @@ async function fetchRemote(url, cbProgress, cbPrint) {
        url,
        {
            method: 'GET',
+            headers: {
+                'Content-Type': 'application/octet-stream',
+            },
        }
    );

--- a/examples/lsp/lsp.cpp
+++ b/examples/lsp/lsp.cpp
@ -26,11 +26,11 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -69,11 +69,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else {
@ -100,11 +100,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "\n");
@ -181,6 +181,7 @@ json unguided_transcription(struct whisper_context * ctx, audio_async &audio, js
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
    wparams.suppress_non_speech_tokens = true;
    // run the transformer and a single decoding pass
    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
@ -219,6 +220,7 @@ json guided_transcription(struct whisper_context * ctx, audio_async &audio, cons
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    // TODO: Do some time testing. Does an overly long prompt slow down processing?
    // Set up command sets/precompute prompts
@ -434,10 +436,7 @@ int main(int argc, char ** argv) {

    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
+    cparams.use_gpu = params.use_gpu;
    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
    // init audio

--- a/examples/main/CMakeLists.txt
+++ b/examples/main/CMakeLists.txt
@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)

 include(DefaultTargetOptions)

-target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@ -1,12 +1,10 @@
 #include "common.h"

 #include "whisper.h"
-#include "grammar-parser.h"

 #include <cmath>
 #include <fstream>
 #include <cstdio>
-#include <regex>
 #include <string>
 #include <thread>
 #include <vector>
@ -28,25 +26,23 @@ void replace_all(std::string & s, const std::string & search, const std::string

 // command-line parameters
 struct whisper_params {
-    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors  = 1;
-    int32_t offset_t_ms   = 0;
-    int32_t offset_n      = 0;
-    int32_t duration_ms   = 0;
-    int32_t progress_step = 5;
-    int32_t max_context   = -1;
-    int32_t max_len       = 0;
-    int32_t best_of       = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
-    int32_t beam_size     = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
-    int32_t audio_ctx     = 0;
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
+    int32_t max_context  = -1;
+    int32_t max_len      =  0;
+    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t audio_ctx   = 0;

-    float word_thold      =  0.01f;
-    float entropy_thold   =  2.40f;
-    float logprob_thold   = -1.00f;
-    float grammar_penalty = 100.0f;
-    float temperature     = 0.0f;
-    float temperature_inc = 0.2f;
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;

+    bool speed_up        = false;
    bool debug_mode      = false;
    bool translate       = false;
    bool detect_language = false;
@ -69,29 +65,19 @@ struct whisper_params {
    bool no_timestamps   = false;
    bool log_score       = false;
    bool use_gpu         = true;
-    bool flash_attn      = false;

    std::string language  = "en";
    std::string prompt;
    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
    std::string model     = "models/ggml-base.en.bin";
-    std::string grammar;
-    std::string grammar_rule;

    // [TDRZ] speaker turn string
    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line

-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
-
    std::string openvino_encode_device = "CPU";

-    std::string dtw = "";
-
    std::vector<std::string> fname_inp = {};
    std::vector<std::string> fname_out = {};
-
-    grammar_parser::parse_state grammar_parsed;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@ -131,12 +117,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-context")   { params.audio_ctx       = std::stoi(argv[++i]); }
        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
-        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@ -164,14 +149,8 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
-        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = argv[++i]; }
-        else if (                  arg == "--grammar")         { params.grammar         = argv[++i]; }
-        else if (                  arg == "--grammar-rule")    { params.grammar_rule    = argv[++i]; }
-        else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -202,8 +181,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    fprintf(stderr, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
-    fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
+    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@ -226,18 +204,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt (max n_text_ctx/2 tokens)\n",       params.prompt.c_str());
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
    fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
-    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
    fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
-    fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
-    fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
-    fprintf(stderr, "  --grammar-rule RULE            [%-7s] top-level GBNF grammar rule name\n",               params.grammar_rule.c_str());
-    fprintf(stderr, "  --grammar-penalty N            [%-7.1f] scales down logits of nongrammar tokens\n",      params.grammar_penalty);
    fprintf(stderr, "\n");
 }

@ -477,38 +449,6 @@ char *escape_double_quotes_and_backslashes(const char *str) {
    return escaped;
 }

-// double quote should be escaped by another double quote. (rfc4180)
-char *escape_double_quotes_in_csv(const char *str) {
-    if (str == NULL) {
-        return NULL;
-    }
-
-    size_t escaped_length = strlen(str) + 1;
-
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped_length++;
-        }
-    }
-
-    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
-    if (escaped == NULL) {
-        return NULL;
-    }
-
-    size_t pos = 0;
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped[pos++] = '"';
-        }
-        escaped[pos++] = str[i];
-    }
-
-    // no need to set zero due to calloc() being used prior
-
-    return escaped;
-}
-
 bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
@ -530,7 +470,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
        const char * text = whisper_full_get_segment_text(ctx, i);
        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-        char * text_escaped = escape_double_quotes_in_csv(text);
+        char * text_escaped = escape_double_quotes_and_backslashes(text);

        //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
        fout << 10 * t0 << "," << 10 * t1 << ",";
@ -709,8 +649,7 @@ bool output_json(
                                    times_o(token.t0, token.t1, false);
                                }
                                value_i("id", token.id, false);
-                                value_f("p", token.p, false);
-                                value_f("t_dtw", token.t_dtw, true);
+                                value_f("p", token.p, true);
                            end_obj(j == (n - 1));
                        }
                        end_arr(!params.diarize && !params.tinydiarize);
@ -905,35 +844,6 @@ void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
 int main(int argc, char ** argv) {
    whisper_params params;

-    // If the only argument starts with "@", read arguments line-by-line
-    // from the given file.
-    std::vector<std::string> vec_args;
-    if (argc == 2 && argv != nullptr && argv[1] != nullptr && argv[1][0] == '@') {
-        // Save the name of the executable.
-        vec_args.push_back(argv[0]);
-
-        // Open the response file.
-        char const * rspfile = argv[1] + sizeof(char);
-        std::ifstream fin(rspfile);
-        if (fin.is_open() == false) {
-            fprintf(stderr, "error: response file '%s' not found\n", rspfile);
-            return 1;
-        }
-
-        // Read the entire response file.
-        std::string line;
-        while (std::getline(fin, line)) {
-            vec_args.push_back(line);
-        }
-
-        // Use the contents of the response file as the command-line arguments.
-        argc = static_cast<int>(vec_args.size());
-        argv = static_cast<char **>(alloca(argc * sizeof (char *)));
-        for (int i = 0; i < argc; ++i) {
-            argv[i] = const_cast<char *>(vec_args[i].c_str());
-        }
-    }
-
    if (whisper_params_parse(argc, argv, params) == false) {
        whisper_print_usage(argc, argv, params);
        return 1;
@ -977,31 +887,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
-    if (!params.dtw.empty()) {
-        cparams.dtw_token_timestamps = true;
-        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
-
-        if (params.dtw == "tiny")      cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
-        if (params.dtw == "tiny.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
-        if (params.dtw == "base")      cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
-        if (params.dtw == "base.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
-        if (params.dtw == "small")     cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
-        if (params.dtw == "small.en")  cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
-        if (params.dtw == "medium")    cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
-        if (params.dtw == "medium.en") cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
-        if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
-        if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
-        if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
-
-        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
-            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
-            return 3;
-        }
-    }
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -1013,29 +899,6 @@ int main(int argc, char ** argv) {
    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);

-    if (!params.grammar.empty()) {
-        auto & grammar = params.grammar_parsed;
-        if (is_file_exist(params.grammar.c_str())) {
-            // read grammar from file
-            std::ifstream ifs(params.grammar.c_str());
-            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
-            grammar = grammar_parser::parse(txt.c_str());
-        } else {
-            // read grammar from string
-            grammar = grammar_parser::parse(params.grammar.c_str());
-        }
-
-        // will be empty (default) if there are parse errors
-        if (grammar.rules.empty()) {
-            fprintf(stderr, "error: failed to parse grammar \"%s\"\n", params.grammar.c_str());
-            return 4;
-        } else {
-            fprintf(stderr, "%s: grammar:\n", __func__);
-            grammar_parser::print_grammar(stderr, grammar);
-            fprintf(stderr, "\n");
-        }
-    }
-
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -1082,8 +945,7 @@ int main(int argc, char ** argv) {
        {
            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);

-            const bool use_grammar = (!params.grammar_parsed.rules.empty() && !params.grammar_rule.empty());
-            wparams.strategy = (params.beam_size > 1 || use_grammar) ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;

            wparams.print_realtime   = false;
            wparams.print_progress   = params.print_progress;
@ -1103,20 +965,17 @@ int main(int argc, char ** argv) {
            wparams.split_on_word    = params.split_on_word;
            wparams.audio_ctx        = params.audio_ctx;

+            wparams.speed_up         = params.speed_up;
            wparams.debug_mode       = params.debug_mode;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]

-            wparams.suppress_regex   = params.suppress_regex.empty() ? nullptr : params.suppress_regex.c_str();
-
            wparams.initial_prompt   = params.prompt.c_str();

            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;

-            wparams.temperature_inc  = params.no_fallback ? 0.0f : params.temperature_inc;
-            wparams.temperature      = params.temperature;
-
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
            wparams.entropy_thold    = params.entropy_thold;
            wparams.logprob_thold    = params.logprob_thold;

@ -1124,20 +983,6 @@ int main(int argc, char ** argv) {

            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };

-            const auto & grammar_parsed = params.grammar_parsed;
-            auto grammar_rules = grammar_parsed.c_rules();
-
-            if (use_grammar) {
-                if (grammar_parsed.symbol_ids.find(params.grammar_rule) == grammar_parsed.symbol_ids.end()) {
-                    fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, params.grammar_rule.c_str());
-                } else {
-                    wparams.grammar_rules = grammar_rules.data();
-                    wparams.n_grammar_rules = grammar_rules.size();
-                    wparams.i_start_rule = grammar_parsed.symbol_ids.at(params.grammar_rule);
-                    wparams.grammar_penalty = params.grammar_penalty;
-                }
-            }
-
            // this callback is called on each new segment
            if (!wparams.print_realtime) {
                wparams.new_segment_callback           = whisper_print_segment_callback;
@ -1234,9 +1079,7 @@ int main(int argc, char ** argv) {
        }
    }

-    if (!params.no_prints) {
-        whisper_print_timings(ctx);
-    }
+    whisper_print_timings(ctx);
    whisper_free(ctx);

    return 0;
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@ -61,6 +61,7 @@ struct whisper_params {
    float temperature     =  0.00f;
    float temperature_inc =  0.20f;

+    bool speed_up        = false;
    bool debug_mode      = false;
    bool translate       = false;
    bool detect_language = false;
@ -74,7 +75,6 @@ struct whisper_params {
    bool print_progress  = false;
    bool no_timestamps   = false;
    bool use_gpu         = true;
-    bool flash_attn      = false;

    std::string language        = "en";
    std::string prompt          = "";
@ -87,8 +87,6 @@ struct whisper_params {
    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line

    std::string openvino_encode_device = "CPU";
-
-    std::string dtw = "";
 };

 void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
@ -111,6 +109,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@ -127,7 +126,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
    // server params
-    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n", params.dtw.c_str());
    fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
    fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
    fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
@ -153,10 +151,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-context")   { params.audio_ctx       = std::stoi(argv[++i]); }
        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@ -174,9 +173,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
        // server params
        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
@ -501,53 +498,7 @@ int main(int argc, char ** argv) {
    }
    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
-    if (!params.dtw.empty()) {
-        cparams.dtw_token_timestamps = true;
-        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
-
-        if (params.dtw == "tiny") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
-        }
-        if (params.dtw == "tiny.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
-        }
-        if (params.dtw == "base") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
-        }
-        if (params.dtw == "base.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
-        }
-        if (params.dtw == "small") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
-        }
-        if (params.dtw == "small.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
-        }
-        if (params.dtw == "medium") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
-        }
-        if (params.dtw == "medium.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
-        }
-        if (params.dtw == "large.v1") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
-        }
-        if (params.dtw == "large.v2") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
-        }
-        if (params.dtw == "large.v3") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
-        }
-
-        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
-            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
-            return 3;
-        }
-    }
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -765,6 +716,7 @@ int main(int argc, char ** argv) {
            wparams.split_on_word    = params.split_on_word;
            wparams.audio_ctx        = params.audio_ctx;

+            wparams.speed_up         = params.speed_up;
            wparams.debug_mode       = params.debug_mode;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
@ -832,7 +784,7 @@ int main(int argc, char ** argv) {
        if (params.response_format == text_format)
        {
            std::string results = output_str(ctx, params, pcmf32s);
-            res.set_content(results.c_str(), "text/html; charset=utf-8");
+            res.set_content(results.c_str(), "text/html");
        }
        else if (params.response_format == srt_format)
        {
@ -913,7 +865,6 @@ int main(int argc, char ** argv) {
                    if (!params.no_timestamps) {
                        word["start"] = token.t0 * 0.01;
                        word["end"] = token.t1 * 0.01;
-                        word["t_dtw"] = token.t_dtw;
                    }
                    word["probability"] = token.p;
                    total_logprob += token.plog;
@ -943,7 +894,7 @@ int main(int argc, char ** argv) {
                            "application/json");
        }

-        // reset params to their defaults
+        // reset params to thier defaults
        params = default_params;
    });
    svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
--- a/examples/stream/README.md
+++ b/examples/stream/README.md
@ -30,13 +30,9 @@ a transcription block that is suitable for parsing.
 The `stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:

 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/stream/stream.cpp
+++ b/examples/stream/stream.cpp
@ -27,6 +27,7 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool no_fallback   = false;
    bool print_special = false;
@ -35,7 +36,6 @@ struct whisper_params {
    bool tinydiarize   = false;
    bool save_audio    = false; // save audio to wav file
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -61,6 +61,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
        else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
        else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
        else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
@ -71,7 +72,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-tdrz" || arg == "--tinydiarize")   { params.tinydiarize   = true; }
        else if (arg == "-sa"   || arg == "--save-audio")    { params.save_audio    = true; }
        else if (arg == "-ng"   || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"   || arg == "--flash-attn")    { params.flash_attn    = true; }

        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -98,6 +98,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
@ -108,7 +109,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
    fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention during inference\n",               params.flash_attn ? "true" : "false");
    fprintf(stderr, "\n");
 }

@ -153,9 +153,7 @@ int main(int argc, char ** argv) {
    }

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -311,6 +309,7 @@ int main(int argc, char ** argv) {
            wparams.n_threads        = params.n_threads;

            wparams.audio_ctx        = params.audio_ctx;
+            wparams.speed_up         = params.speed_up;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]

--- a/examples/talk-llama/CMakeLists.txt
+++ b/examples/talk-llama/CMakeLists.txt
@ -1,7 +1,7 @@
 if (WHISPER_SDL2)
    # talk-llama
    set(TARGET talk-llama)
-    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp unicode-data.cpp)
+    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp)
    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})

    if (WHISPER_CLBLAST)
--- a/examples/talk-llama/README.md
+++ b/examples/talk-llama/README.md
@ -15,13 +15,9 @@ https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-
 The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:

 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@ -37,13 +37,9 @@

 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
-#define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'

 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 6
-
-#define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#define LLAMA_SESSION_VERSION 4

 #ifdef __cplusplus
 extern "C" {
@ -64,26 +60,9 @@ extern "C" {

    enum llama_vocab_type {
        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-    };
-
-    // pre-tokenization types
-    enum llama_vocab_pre_type {
-        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
-        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
-        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
-        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
-        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
-        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
-        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
    };

    // note: these values should be synchronized with ggml_rope
@ -138,8 +117,6 @@ extern "C" {
        LLAMA_FTYPE_MOSTLY_IQ2_S         = 28, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors

        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
    };
@ -177,7 +154,7 @@ extern "C" {
        bool sorted;
    } llama_token_data_array;

-    typedef bool (*llama_progress_callback)(float progress, void * user_data);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);

    // Input data for llama_decode
    // A llama_batch object can contain input about one or many sequences
@ -213,19 +190,15 @@ extern "C" {
        LLAMA_KV_OVERRIDE_TYPE_INT,
        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
        LLAMA_KV_OVERRIDE_TYPE_BOOL,
-        LLAMA_KV_OVERRIDE_TYPE_STR,
    };

    struct llama_model_kv_override {
-        enum llama_model_kv_override_type tag;
-
        char key[128];
-
+        enum llama_model_kv_override_type tag;
        union {
-            int64_t val_i64;
-            double  val_f64;
-            bool    val_bool;
-            char    val_str[128];
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
        };
    };

@ -254,10 +227,9 @@ extern "C" {
        const struct llama_model_kv_override * kv_overrides;

        // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
    };

    struct llama_context_params {
@ -293,7 +265,6 @@ extern "C" {
        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
        bool embeddings;  // if true, extract embeddings (together with logits)
        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention

        // Abort callback
        // if it returns true, execution of llama_decode() will be aborted
@ -304,17 +275,13 @@ extern "C" {

    // model quantization parameters
    typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // itoken embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;      // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // quantize all tensors to the default type
+        void * imatrix;              // pointer to importance matrix data
    } llama_model_quantize_params;

    // grammar types
@ -415,15 +382,12 @@ extern "C" {
    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
    LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);

-    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
-
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model   * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model   * model);
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);

    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);

    // Get the model's RoPE frequency scaling factor
    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
@ -471,24 +435,10 @@ extern "C" {
    // Returns 0 on success
    LLAMA_API int32_t llama_model_apply_lora_from_file(
            const struct llama_model * model,
-                          const char * path_lora,
-                               float   scale,
-                          const char * path_base_model,
-                             int32_t   n_threads);
-
-    // Apply a loaded control vector to a llama_context, or if data is NULL, clear
-    // the currently loaded vector.
-    // n_embd should be the size of a single layer's control, and data should point
-    // to an n_embd x n_layers buffer starting from layer 1.
-    // il_start and il_end are the layer range the vector should apply to (both inclusive)
-    // See llama_control_vector_load in common to load a control vector.
-    LLAMA_API int32_t llama_control_vector_apply(
-            struct llama_context * lctx,
-                     const float * data,
-                          size_t   len,
-                         int32_t   n_embd,
-                         int32_t   il_start,
-                         int32_t   il_end);
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                         int32_t   n_threads);

    //
    // KV cache
@ -549,12 +499,11 @@ extern "C" {
    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);

-    // Clear the KV cache - both cell info is erased and KV data is zeroed
+    // Clear the KV cache
    LLAMA_API void llama_kv_cache_clear(
            struct llama_context * ctx);

    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
    // seq_id < 0 : match any sequence
    // p0 < 0     : [0,  p1]
    // p1 < 0     : [p0, inf)
@ -626,92 +575,34 @@ extern "C" {

    // Returns the maximum size in bytes of the state (rng, logits, embedding
    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
-        "use llama_state_get_size instead");
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);

    // Copies the state to the specified destination address.
    // Destination needs to have allocated enough memory.
    // Returns the number of bytes copied
-    LLAMA_API size_t llama_state_get_data(
+    LLAMA_API size_t llama_copy_state_data(
            struct llama_context * ctx,
                         uint8_t * dst);
-    LLAMA_API DEPRECATED(size_t llama_copy_state_data(
-            struct llama_context * ctx,
-                         uint8_t * dst),
-        "use llama_state_get_data instead");

    // Set the state reading from the specified address
    // Returns the number of bytes read
-    LLAMA_API size_t llama_state_set_data(
+    LLAMA_API size_t llama_set_state_data(
            struct llama_context * ctx,
                   const uint8_t * src);
-    LLAMA_API DEPRECATED(size_t llama_set_state_data(
-            struct llama_context * ctx,
-                   const uint8_t * src),
-        "use llama_state_set_data instead");

    // Save/load session file
-    LLAMA_API bool llama_state_load_file(
+    LLAMA_API bool llama_load_session_file(
            struct llama_context * ctx,
                      const char * path_session,
                     llama_token * tokens_out,
                          size_t   n_token_capacity,
                          size_t * n_token_count_out);
-    LLAMA_API DEPRECATED(bool llama_load_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out),
-        "use llama_state_load_file instead");

-    LLAMA_API bool llama_state_save_file(
+    LLAMA_API bool llama_save_session_file(
            struct llama_context * ctx,
                      const char * path_session,
               const llama_token * tokens,
                          size_t   n_token_count);
-    LLAMA_API DEPRECATED(bool llama_save_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-               const llama_token * tokens,
-                          size_t   n_token_count),
-        "use llama_state_save_file instead");
-
-    // Get the exact size needed to copy the KV cache of a single sequence
-    LLAMA_API size_t llama_state_seq_get_size(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // Copy the KV cache of a single sequence into the specified buffer
-    LLAMA_API size_t llama_state_seq_get_data(
-            struct llama_context * ctx,
-                         uint8_t * dst,
-                    llama_seq_id   seq_id);
-
-    // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
-    // Returns:
-    //  - Positive: Ok
-    //  - Zero: Failed to load
-    LLAMA_API size_t llama_state_seq_set_data(
-            struct llama_context * ctx,
-                   const uint8_t * src,
-                    llama_seq_id   dest_seq_id);
-
-    LLAMA_API size_t llama_state_seq_save_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   seq_id,
-               const llama_token * tokens,
-                          size_t   n_token_count);
-
-    LLAMA_API size_t llama_state_seq_load_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   dest_seq_id,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out);

    //
    // Decoding
@ -768,31 +659,23 @@ extern "C" {
    LLAMA_API void llama_synchronize(struct llama_context * ctx);

    // Token logits obtained from the last call to llama_decode()
-    // The logits for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // Rows: number of tokens for which llama_batch.logits[i] != 0
+    // The logits for the last token are stored in the last row
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
    // Cols: n_vocab
    LLAMA_API float * llama_get_logits(struct llama_context * ctx);

-    // Logits for the ith token. For positive indices, Equivalent to:
-    // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
-    // Negative indicies can be used to access logits in reverse order, -1 is the last logit.
-    // returns NULL for invalid ids.
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);

-    // Get all output token embeddings.
-    // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,
-    // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // shape: [n_outputs*n_embd]
-    // Otherwise, returns NULL.
+    // Get all output token embeddings
+    // shape: [n_tokens*n_embd] (1-dimensional)
    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);

-    // Get the embeddings for the ith token. For positive indices, Equivalent to:
-    // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
-    // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
+    // Get the embeddings for the ith token
+    // llama_get_embeddings(ctx) + i*n_embd
    // shape: [n_embd] (1-dimensional)
-    // returns NULL for invalid ids.
    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);

    // Get the embeddings for a sequence id
@ -810,14 +693,9 @@ extern "C" {

    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);

-    // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
-    LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
-
    // Special tokens
    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
-    LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
-    LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line

    // Returns -1 if unknown, 1 for true or 0 for false.
@ -826,7 +704,7 @@ extern "C" {
    // Returns -1 if unknown, 1 for true or 0 for false.
    LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);

-    // Codellama infill tokens
+    // codellama infill tokens
    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
@ -840,28 +718,26 @@ extern "C" {
    /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
    /// @return Returns the number of tokens on success, no more than n_tokens_max
    /// @return Returns a negative number on failure - the number of tokens that would have been returned
-    /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
-    ///                      as plaintext. Does not insert a leading space.
+    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
+    ///                Does not insert a leading space.
    LLAMA_API int32_t llama_tokenize(
        const struct llama_model * model,
                      const char * text,
                         int32_t   text_len,
                     llama_token * tokens,
                         int32_t   n_tokens_max,
-                            bool   add_special,
-                            bool   parse_special);
+                            bool   add_bos,
+                            bool   special);

    // Token Id -> Piece.
    // Uses the vocabulary in the provided context.
    // Does not write null terminator to the buffer.
    // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
-    // @param special If true, special tokens are rendered in the output.
    LLAMA_API int32_t llama_token_to_piece(
              const struct llama_model * model,
                           llama_token   token,
                                  char * buf,
-                               int32_t   length,
-                                  bool   special);
+                               int32_t   length);

    /// Apply chat template. Inspired by hf apply_chat_template() on python.
    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
@ -1014,7 +890,7 @@ extern "C" {
            struct llama_context * ctx,
          llama_token_data_array * candidates);

-    /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
+    /// @details Randomly selects a token from the candidates based on their probabilities.
    LLAMA_API llama_token llama_sample_token(
            struct llama_context * ctx,
          llama_token_data_array * candidates);
@ -1069,16 +945,6 @@ extern "C" {
                                int32_t   n_past,
                                int32_t   n_predict);

-    /// @details Build a split GGUF final path for this chunk.
-    ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
-    //  Returns the split_path length.
-    LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
-
-    /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
-    ///          llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
-    //  Returns the split_prefix length.
-    LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
-
    // Performance information
    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);

@ -1101,49 +967,15 @@ extern "C" {
 // Internal API to be implemented by llama.cpp and used by tests/benchmarks only
 #ifdef LLAMA_API_INTERNAL

-#include <random>
-#include <string>
 #include <vector>
+#include <string>

 struct ggml_tensor;

-struct llama_partial_utf8 {
-    uint32_t value;    // bit value so far (unshifted)
-    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
-};
-
-struct llama_grammar {
-    const std::vector<std::vector<llama_grammar_element>>   rules;
-    std::vector<std::vector<const llama_grammar_element *>> stacks;
-
-    // buffer for partially generated UTF-8 sequence from accepted tokens
-    llama_partial_utf8                                      partial_utf8;
-};
-
-struct llama_grammar_candidate {
-    size_t               index;
-    const uint32_t     * code_points;
-    llama_partial_utf8   partial_utf8;
-};
-
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
    struct llama_context * ctx
 );

-void llama_grammar_accept(
-        const std::vector<std::vector<llama_grammar_element>>         & rules,
-        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
-        const uint32_t                                                  chr,
-        std::vector<std::vector<const llama_grammar_element *>>       & new_stacks);
-
-std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        const std::string & src,
-        llama_partial_utf8   partial_start);
-
-// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
-// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
-llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
-
 #endif // LLAMA_API_INTERNAL

 #endif // LLAMA_H
--- a/examples/talk-llama/talk-llama.cpp
+++ b/examples/talk-llama/talk-llama.cpp
@ -35,10 +35,10 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s

 std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
    std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
    if (n_tokens < 0) {
        result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
        GGML_ASSERT(check == -n_tokens);
    } else {
        result.resize(n_tokens);
@ -59,13 +59,13 @@ struct whisper_params {
    float vad_thold  = 0.6f;
    float freq_thold = 100.0f;

+    bool speed_up       = false;
    bool translate      = false;
    bool print_special  = false;
    bool print_energy   = false;
    bool no_timestamps  = true;
    bool verbose_prompt = false;
    bool use_gpu        = true;
-    bool flash_attn     = false;

    std::string person      = "Georgi";
    std::string bot_name    = "LLaMA";
@ -99,12 +99,12 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ngl" || arg == "--n-gpu-layers")   { params.n_gpu_layers   = std::stoi(argv[++i]); }
        else if (arg == "-vth" || arg == "--vad-thold")      { params.vad_thold      = std::stof(argv[++i]); }
        else if (arg == "-fth" || arg == "--freq-thold")     { params.freq_thold     = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")       { params.speed_up       = true; }
        else if (arg == "-tr"  || arg == "--translate")      { params.translate      = true; }
        else if (arg == "-ps"  || arg == "--print-special")  { params.print_special  = true; }
        else if (arg == "-pe"  || arg == "--print-energy")   { params.print_energy   = true; }
        else if (arg == "-vp"  || arg == "--verbose-prompt") { params.verbose_prompt = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")         { params.use_gpu        = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")     { params.flash_attn     = true; }
        else if (arg == "-p"   || arg == "--person")         { params.person         = argv[++i]; }
        else if (arg == "-bn"   || arg == "--bot-name")      { params.bot_name       = argv[++i]; }
        else if (arg == "--session")                         { params.path_session   = argv[++i]; }
@ -123,6 +123,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
            }
        }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -147,12 +148,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ngl N,   --n-gpu-layers N [%-7d] number of layers to store in VRAM\n",           params.n_gpu_layers);
    fprintf(stderr, "  -vth N,   --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,      --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -vp,      --verbose-prompt [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
    fprintf(stderr, "  -ng,      --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -p NAME,  --person NAME    [%-7s] person name (for prompt selection)\n",          params.person.c_str());
    fprintf(stderr, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
    fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
@ -202,6 +203,7 @@ std::string transcribe(
    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
        return "";
@ -283,15 +285,9 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
-    if (!ctx_wsp) {
-        fprintf(stderr, "No whisper.cpp model specified. Please provide using -mw <modelfile>\n");
-        return 1;
-    }

    // llama init

@ -305,10 +301,6 @@ int main(int argc, char ** argv) {
    }

    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
-    if (!model_llama) {
-        fprintf(stderr, "No llama.cpp model specified. Please provide using -ml <modelfile>\n");
-        return 1;
-    }

    llama_context_params lcparams = llama_context_default_params();

@ -316,7 +308,6 @@ int main(int argc, char ** argv) {
    lcparams.n_ctx      = 2048;
    lcparams.seed       = 1;
    lcparams.n_threads  = params.n_threads;
-    lcparams.flash_attn = params.flash_attn;

    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);

--- a/examples/talk-llama/unicode-data.cpp
+++ b/examples/talk-llama/unicode-data.cpp
--- a/examples/talk-llama/unicode-data.h
+++ b/examples/talk-llama/unicode-data.h
@ -1,17 +0,0 @@
-#pragma once
-
-#include <cstdint>
-#include <map>
-#include <utility>
-#include <vector>
-
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
-extern const std::multimap<uint32_t, uint32_t>          unicode_map_nfd;
-extern const std::map<char32_t, char32_t>               unicode_map_lowercase;
--- a/examples/talk-llama/unicode.cpp
+++ b/examples/talk-llama/unicode.cpp
--- a/examples/talk-llama/unicode.h
+++ b/examples/talk-llama/unicode.h
@ -5,9 +5,9 @@
 #include <vector>

 #define CODEPOINT_TYPE_UNIDENTIFIED 0
-#define CODEPOINT_TYPE_NUMBER       1
+#define CODEPOINT_TYPE_DIGIT        1
 #define CODEPOINT_TYPE_LETTER       2
-#define CODEPOINT_TYPE_SEPARATOR    3
+#define CODEPOINT_TYPE_WHITESPACE   3
 #define CODEPOINT_TYPE_ACCENT_MARK  4
 #define CODEPOINT_TYPE_PUNCTUATION  5
 #define CODEPOINT_TYPE_SYMBOL       6
@ -21,11 +21,6 @@ std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & c
 int unicode_cpt_type(uint32_t cp);
 int unicode_cpt_type(const std::string & utf8);

-bool unicode_cpt_is_whitespace(uint32_t cp);
-
 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);

-char32_t unicode_tolower(char32_t cp);
-
-std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);
--- a/examples/talk/README.md
+++ b/examples/talk/README.md
@ -11,13 +11,9 @@ Web version: [examples/talk.wasm](/examples/talk.wasm)
 The `talk` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:

 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/talk/talk.cpp
+++ b/examples/talk/talk.cpp
@ -26,12 +26,12 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string person    = "Santa";
    std::string language  = "en";
@ -59,11 +59,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
@ -94,11 +94,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
@ -129,6 +129,7 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
        return "";
@ -187,9 +188,7 @@ int main(int argc, char ** argv) {

    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);

--- a/examples/wchess/CMakeLists.txt
+++ b/examples/wchess/CMakeLists.txt
@ -1,10 +1,9 @@
+set(CMAKE_CXX_STANDARD 11)
+
 add_subdirectory(libwchess)
-set_target_properties(wchess-core PROPERTIES FOLDER "libs")

 if (EMSCRIPTEN)
    add_subdirectory(wchess.wasm)
-    set_target_properties(wchess.wasm PROPERTIES FOLDER "libs")
 else()
    add_subdirectory(wchess.cmd)
-    set_target_properties(wchess PROPERTIES FOLDER "libs")
 endif()
--- a/examples/wchess/wchess.cmd/wchess.cmd.cpp
+++ b/examples/wchess/wchess.cmd/wchess.cmd.cpp
@ -26,12 +26,12 @@ struct whisper_params {

    float grammar_penalty = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -56,11 +56,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention during decoding\n",             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@ -87,11 +87,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
@ -183,9 +183,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
    if (!ctx) {
--- a/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
+++ b/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
@ -145,7 +145,7 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
            val start = System.currentTimeMillis()
            val text = whisperContext?.transcribeData(data)
            val elapsed = System.currentTimeMillis() - start
-            printMessage("Done ($elapsed ms): \n$text\n")
+            printMessage("Done ($elapsed ms): $text\n")
        } catch (e: Exception) {
            Log.w(LOG_TAG, e)
            printMessage("${e.localizedMessage}\n")
--- a/examples/whisper.android/lib/src/main/java/com/whispercpp/whisper/LibWhisper.kt
+++ b/examples/whisper.android/lib/src/main/java/com/whispercpp/whisper/LibWhisper.kt
@ -16,7 +16,7 @@ class WhisperContext private constructor(private var ptr: Long) {
        Executors.newSingleThreadExecutor().asCoroutineDispatcher()
    )

-    suspend fun transcribeData(data: FloatArray, printTimestamp: Boolean = true): String = withContext(scope.coroutineContext) {
+    suspend fun transcribeData(data: FloatArray): String = withContext(scope.coroutineContext) {
        require(ptr != 0L)
        val numThreads = WhisperCpuConfig.preferredThreadCount
        Log.d(LOG_TAG, "Selecting $numThreads threads")
@ -24,13 +24,7 @@ class WhisperContext private constructor(private var ptr: Long) {
        val textCount = WhisperLib.getTextSegmentCount(ptr)
        return@withContext buildString {
            for (i in 0 until textCount) {
-                if (printTimestamp) {
-                    val textTimestamp = "[${toTimestamp(WhisperLib.getTextSegmentT0(ptr, i))} --> ${toTimestamp(WhisperLib.getTextSegmentT1(ptr, i))}]"
-                    val textSegment = WhisperLib.getTextSegment(ptr, i)
-                    append("$textTimestamp: $textSegment\n")
-                } else {
-                    append(WhisperLib.getTextSegment(ptr, i))
-                }
+                append(WhisperLib.getTextSegment(ptr, i))
            }
        }
    }
@ -137,29 +131,12 @@ private class WhisperLib {
        external fun fullTranscribe(contextPtr: Long, numThreads: Int, audioData: FloatArray)
        external fun getTextSegmentCount(contextPtr: Long): Int
        external fun getTextSegment(contextPtr: Long, index: Int): String
-        external fun getTextSegmentT0(contextPtr: Long, index: Int): Long
-        external fun getTextSegmentT1(contextPtr: Long, index: Int): Long
        external fun getSystemInfo(): String
        external fun benchMemcpy(nthread: Int): String
        external fun benchGgmlMulMat(nthread: Int): String
    }
 }

-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-private fun toTimestamp(t: Long, comma: Boolean = false): String {
-    var msec = t * 10
-    val hr = msec / (1000 * 60 * 60)
-    msec -= hr * (1000 * 60 * 60)
-    val min = msec / (1000 * 60)
-    msec -= min * (1000 * 60)
-    val sec = msec / 1000
-    msec -= sec * 1000
-
-    val delimiter = if (comma) "," else "."
-    return String.format("%02d:%02d:%02d%s%03d", hr, min, sec, delimiter, msec)
-}
-
 private fun isArmEabiV7a(): Boolean {
    return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
--- a/examples/whisper.android/lib/src/main/jni/whisper/jni.c
+++ b/examples/whisper.android/lib/src/main/jni/whisper/jni.c
@ -212,22 +212,6 @@ Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegment(
    return string;
 }

-JNIEXPORT jlong JNICALL
-Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegmentT0(
-        JNIEnv *env, jobject thiz, jlong context_ptr, jint index) {
-    UNUSED(thiz);
-    struct whisper_context *context = (struct whisper_context *) context_ptr;
-    return whisper_full_get_segment_t0(context, index);
-}
-
-JNIEXPORT jlong JNICALL
-Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegmentT1(
-        JNIEnv *env, jobject thiz, jlong context_ptr, jint index) {
-    UNUSED(thiz);
-    struct whisper_context *context = (struct whisper_context *) context_ptr;
-    return whisper_full_get_segment_t1(context, index);
-}
-
 JNIEXPORT jstring JNICALL
 Java_com_whispercpp_whisper_WhisperLib_00024Companion_getSystemInfo(
        JNIEnv *env, jobject thiz
--- a/examples/whisper.nvim/whisper.nvim
+++ b/examples/whisper.nvim/whisper.nvim
@ -45,6 +45,6 @@ if [ ! -f ./models/ggml-${model}.bin ] ; then
 fi

 # fine-tune the parameters according to your machine specs
-./stream -t 8 -m models/ggml-${model}.bin --step 350 --length 10000 -f /tmp/whisper.nvim 2> /dev/null
+./stream -t 8 -m models/ggml-base.en.bin --step 350 --length 10000 -f /tmp/whisper.nvim 2> /dev/null

 exit 0
--- a/scripts/bench-all.sh
+++ b/scripts/bench-all.sh
@ -2,7 +2,7 @@

 # Helper script to run the bench tool on all models and print the results in share-able format

-printf "Usage: ./bench.sh [n_threads] [encoder-only] [flash-attn]\n"
+printf "Usage: ./bench.sh [n_threads] [encoder-only]\n"

 if [ -z "$1" ]; then
    n_threads=4
@ -11,19 +11,12 @@ else
 fi

 encoder_only=0
-if [ -z "$2" ] || [ "$2" -eq 0 ]; then
+if [ -z "$2" ]; then
    encoder_only=0
 else
    encoder_only=$2
 fi

-fattn=""
-if [ -z "$3" ] || [ "$3" -eq 0 ]; then
-    fattn=""
-else
-    fattn="-fa"
-fi
-
 models=(                                                                                                    \
      "tiny"     "tiny-q4_0"     "tiny-q4_1"     "tiny-q5_0"     "tiny-q5_1"     "tiny-q8_0"                \
      "base"     "base-q4_0"     "base-q4_1"     "base-q5_0"     "base-q5_1"     "base-q8_0"                \
@ -51,19 +44,13 @@ if [ "$encoder_only" -eq 0 ]; then
    printf "\n"
 fi

-if [ "$fattn" == "-fa" ]; then
-    fattn_i=1
-else
-    fattn_i=0
-fi
-
-printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "FA" "Enc." "Dec." "Bch5" "PP" "Commit"
-printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"
+printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "Bch5" "PP" "Commit"
+printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"

 for model in "${models[@]}"; do
    # actual run
    # store stderr output in a variable in order to parse it later
-    output=$(./bench -m ./models/ggml-$model.bin -t $n_threads $fattn 2>&1)
+    output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
    ret=$?

    # parse the output:
@ -108,6 +95,6 @@ for model in "${models[@]}"; do
    commit=$(git rev-parse --short HEAD)

    if [ $ret -eq 0 ]; then
-        printf "| <todo> | <todo> | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$fattn_i" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
+        printf "| <todo> | <todo> | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
    fi
 done
--- a/scripts/bench-wts.sh
+++ b/scripts/bench-wts.sh
--- a/scripts/bench.py
+++ b/scripts/bench.py
--- a/scripts/convert-all.sh
+++ b/scripts/convert-all.sh
--- a/scripts/deploy-wasm.sh
+++ b/scripts/deploy-wasm.sh
@ -4,7 +4,7 @@
 # Run from the build directory:
 #
 # cd build-em
-# ../scripts/deploy-wasm.sh
+# ../extra/deploy-wasm.sh
 #

 # check if emcmake is available
--- a/scripts/quantize-all.sh
+++ b/scripts/quantize-all.sh
--- a/scripts/sha-all.sh
+++ b/scripts/sha-all.sh
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@ -5,7 +5,7 @@
 # Usage:
 #
 #   $ cd /path/to/whisper.cpp
-#   $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2...
+#   $ ./extra/sync-ggml-am.sh -skip hash0,hash1,hash2...
 #

 set -e
@ -21,7 +21,7 @@ if [ ! -d $SRC_GGML ]; then
    exit 1
 fi

-lc=$(cat $SRC_WHISPER/scripts/sync-ggml.last)
+lc=$(cat $SRC_WHISPER/extra/sync-ggml.last)
 echo "Syncing ggml changes since commit $lc"

 to_skip=""
@ -60,19 +60,14 @@ while read c; do
        src/ggml*.m \
        src/ggml*.metal \
        src/ggml*.cu \
-        src/ggml-cuda/* \
        examples/common.h \
        examples/common.cpp \
        examples/common-ggml.h \
        examples/common-ggml.cpp \
-        examples/whisper/grammar-parser.h \
-        examples/whisper/grammar-parser.cpp \
        examples/whisper/whisper.h \
        examples/whisper/whisper.cpp \
        examples/whisper/main.cpp \
        examples/whisper/quantize.cpp \
-        LICENSE \
-        scripts/gen-authors.sh \
        >> $SRC_WHISPER/ggml-src.patch
 done < $SRC_WHISPER/ggml-commits

@ -103,7 +98,6 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
    # src/ggml-backend-impl.h     -> ggml-backend-impl.h
    # src/ggml-backend.c          -> ggml-backend.c
    # src/ggml-common.h           -> ggml-common.h
-    # src/ggml-cuda/*             -> ggml-cuda/
    # src/ggml-cuda.cu            -> ggml-cuda.cu
    # src/ggml-cuda.h             -> ggml-cuda.h
    # src/ggml-impl.h             -> ggml-impl.h
@ -117,8 +111,6 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
    # src/ggml-opencl.h           -> ggml-opencl.h
    # src/ggml-quants.c           -> ggml-quants.c
    # src/ggml-quants.h           -> ggml-quants.h
-    # src/ggml-rpc.cpp            -> ggml-rpc.cpp
-    # src/ggml-rpc.h              -> ggml-rpc.h
    # src/ggml-sycl.cpp           -> ggml-sycl.cpp
    # src/ggml-sycl.h             -> ggml-sycl.h
    # src/ggml-vulkan.cpp         -> ggml-vulkan.cpp
@ -127,20 +119,15 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
    # include/ggml/ggml-alloc.h   -> ggml-alloc.h
    # include/ggml/ggml-backend.h -> ggml-backend.h
    #
-    # examples/common.h                   -> examples/common.h
-    # examples/common.cpp                 -> examples/common.cpp
-    # examples/common-ggml.h              -> examples/common-ggml.h
-    # examples/common-ggml.cpp            -> examples/common-ggml.cpp
-    # examples/whisper/grammar-parser.h   -> examples/grammar-parser.h
-    # examples/whisper/grammar-parser.cpp -> examples/grammar-parser.cpp
+    # examples/common.h           -> examples/common.h
+    # examples/common.cpp         -> examples/common.cpp
+    # examples/common-ggml.h      -> examples/common-ggml.h
+    # examples/common-ggml.cpp    -> examples/common-ggml.cpp
    #
    # examples/whisper/whisper.h    -> whisper.h
    # examples/whisper/whisper.cpp  -> whisper.cpp
    # examples/whisper/main.cpp     -> examples/main/main.cpp
    # examples/whisper/quantize.cpp -> examples/quantize/quantize.cpp
-    #
-    # LICENSE                     -> LICENSE
-    # ggml/scripts/gen-authors.sh -> scripts/gen-authors.sh

    cat ggml-src.patch | sed \
        -e 's/src\/ggml\.c/ggml.c/g' \
@ -148,7 +135,6 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
        -e 's/src\/ggml-backend-impl\.h/ggml-backend-impl.h/g' \
        -e 's/src\/ggml-backend\.c/ggml-backend.c/g' \
        -e 's/src\/ggml-common\.h/ggml-common.h/g' \
-        -e 's/src\/ggml-cuda\//ggml-cuda\//g' \
        -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \
        -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \
        -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \
@ -162,8 +148,6 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
        -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
        -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \
        -e 's/src\/ggml-quants\.h/ggml-quants.h/g' \
-        -e 's/src\/ggml-rpc\.cpp/ggml-rpc.cpp/g' \
-        -e 's/src\/ggml-rpc\.h/ggml-rpc.h/g' \
        -e 's/src\/ggml-sycl\.cpp/ggml-sycl.cpp/g' \
        -e 's/src\/ggml-sycl\.h/ggml-sycl.h/g' \
        -e 's/src\/ggml-vulkan\.cpp/ggml-vulkan.cpp/g' \
@ -175,14 +159,10 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
        -e 's/examples\/common\.cpp/examples\/common.cpp/g' \
        -e 's/examples\/common-ggml\.h/examples\/common-ggml.h/g' \
        -e 's/examples\/common-ggml\.cpp/examples\/common-ggml.cpp/g' \
-        -e 's/examples\/whisper\/grammar-parser\.h/examples\/grammar-parser.h/g' \
-        -e 's/examples\/whisper\/grammar-parser\.cpp/examples\/grammar-parser.cpp/g' \
        -e 's/examples\/whisper\/whisper\.h/whisper.h/g' \
        -e 's/examples\/whisper\/whisper\.cpp/whisper.cpp/g' \
        -e 's/examples\/whisper\/main\.cpp/examples\/main\/main.cpp/g' \
        -e 's/examples\/whisper\/quantize\.cpp/examples\/quantize\/quantize.cpp/g' \
-        -e 's/LICENSE/LICENSE/g' \
-        -e 's/ggml\/scripts\/gen-authors\.sh/scripts\/gen-authors.sh/g' \
        > ggml-src.patch.tmp
    mv ggml-src.patch.tmp ggml-src.patch

@ -193,7 +173,7 @@ fi

 # update last commit
 cd $SRC_GGML
-git log -1 --format=%H > $SRC_WHISPER/scripts/sync-ggml.last
+git log -1 --format=%H > $SRC_WHISPER/extra/sync-ggml.last

 echo "Done"

--- a/extra/sync-ggml.last
+++ b/extra/sync-ggml.last
@ -0,0 +1 @@
+7652115c79fa0ffedb58a28c09cd2871403d5a20
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@ -6,7 +6,6 @@ cp -rpv ../ggml/src/ggml-alloc.c        ./ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c      ./ggml-backend.c
 cp -rpv ../ggml/src/ggml-common.h       ./ggml-common.h
-cp -rpv ../ggml/src/ggml-cuda/*         ./ggml-cuda/
 cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml-cuda.cu
 cp -rpv ../ggml/src/ggml-cuda.h         ./ggml-cuda.h
 cp -rpv ../ggml/src/ggml-kompute.cpp    ./ggml-kompute.cpp
@ -20,8 +19,6 @@ cp -rpv ../ggml/src/ggml-opencl.cpp     ./ggml-opencl.cpp
 cp -rpv ../ggml/src/ggml-opencl.h       ./ggml-opencl.h
 cp -rpv ../ggml/src/ggml-quants.c       ./ggml-quants.c
 cp -rpv ../ggml/src/ggml-quants.h       ./ggml-quants.h
-cp -rpv ../ggml/src/ggml-rpc.cpp        ./ggml-rpc.cpp
-cp -rpv ../ggml/src/ggml-rpc.h          ./ggml-rpc.h
 cp -rpv ../ggml/src/ggml-sycl.cpp       ./ggml-sycl.cpp
 cp -rpv ../ggml/src/ggml-sycl.h         ./ggml-sycl.h
 cp -rpv ../ggml/src/ggml-vulkan.cpp     ./ggml-vulkan.cpp
@ -31,17 +28,13 @@ cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
 cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h

-cp -rpv ../ggml/examples/common.h                   ./examples/common.h
-cp -rpv ../ggml/examples/common.cpp                 ./examples/common.cpp
-cp -rpv ../ggml/examples/common-ggml.h              ./examples/common-ggml.h
-cp -rpv ../ggml/examples/common-ggml.cpp            ./examples/common-ggml.cpp
-cp -rpv ../ggml/examples/whisper/grammar-parser.h   ./examples/grammar-parser.h
-cp -rpv ../ggml/examples/whisper/grammar-parser.cpp ./examples/grammar-parser.cpp
+cp -rpv ../ggml/examples/common.h        ./examples/common.h
+cp -rpv ../ggml/examples/common.cpp      ./examples/common.cpp
+cp -rpv ../ggml/examples/common-ggml.h   ./examples/common-ggml.h
+cp -rpv ../ggml/examples/common-ggml.cpp ./examples/common-ggml.cpp

 cp -rpv ../ggml/examples/whisper/whisper.h    ./whisper.h
 cp -rpv ../ggml/examples/whisper/whisper.cpp  ./whisper.cpp
 cp -rpv ../ggml/examples/whisper/main.cpp     ./examples/main/main.cpp
 cp -rpv ../ggml/examples/whisper/quantize.cpp ./examples/quantize/quantize.cpp

-cp -rpv ../LICENSE                     ./LICENSE
-cp -rpv ../ggml/scripts/gen-authors.sh ./scripts/gen-authors.sh
--- a/extra/sync-llama.sh
+++ b/extra/sync-llama.sh
@ -0,0 +1,6 @@
+#!/bin/bash
+
+cp -rpv ../llama.cpp/llama.h     ./examples/talk-llama/llama.h
+cp -rpv ../llama.cpp/llama.cpp   ./examples/talk-llama/llama.cpp
+cp -rpv ../llama.cpp/unicode.h   ./examples/talk-llama/unicode.h
+cp -rpv ../llama.cpp/unicode.cpp ./examples/talk-llama/unicode.cpp
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@ -371,16 +371,16 @@ struct ggml_gallocr {
 };

 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
    GGML_ASSERT(galloc != NULL);

-    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
+    galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
    GGML_ASSERT(galloc->bufts != NULL);

-    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
+    galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
    GGML_ASSERT(galloc->buffers != NULL);

-    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+    galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
    GGML_ASSERT(galloc->buf_tallocs != NULL);

    for (int i = 0; i < n_bufs; i++) {
@ -548,11 +548,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];

-        // TODO: better way to add external dependencies
-        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
-        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
-        // itself is never used and should not be considered a dependency
-        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+        if (ggml_is_view(node)) {
            struct ggml_tensor * view_src = node->view_src;
            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
        }
@ -569,8 +565,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr

            ggml_gallocr_hash_get(galloc, src)->n_children += 1;

-            // allocate explicit inputs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
+            // allocate explicit inputs and leafs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
            }
        }
@ -646,8 +642,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
        free(galloc->hash_set.keys);
        free(galloc->hash_values);
        galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
-        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
+        galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
+        galloc->hash_values   = calloc(sizeof(struct hash_node), hash_size);
        GGML_ASSERT(galloc->hash_set.keys != NULL);
        GGML_ASSERT(galloc->hash_values != NULL);
    } else {
@ -667,7 +663,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
    // set the node_allocs from the hash table
    if (galloc->n_nodes < graph->n_nodes) {
        free(galloc->node_allocs);
-        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
        GGML_ASSERT(galloc->node_allocs != NULL);
    }
    galloc->n_nodes = graph->n_nodes;
@ -697,7 +693,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
    }
    if (galloc->n_leafs < graph->n_leafs) {
        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
        GGML_ASSERT(galloc->leaf_allocs != NULL);
    }
    galloc->n_leafs = graph->n_leafs;
--- a/ggml-backend-impl.h
+++ b/ggml-backend-impl.h
@ -103,11 +103,6 @@ extern "C" {
        // check if the backend supports an operation
        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);

-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
        // (optional) event synchronization
        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
--- a/ggml-backend.c
+++ b/ggml-backend.c
@ -278,7 +278,7 @@ enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_
    return err;
 }

-enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    return backend->iface.graph_compute(backend, cgraph);
 }

@ -286,13 +286,6 @@ bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor *
    return backend->iface.supports_op(backend, op);
 }

-bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    if (backend->iface.offload_op != NULL) {
-        return backend->iface.offload_op(backend, op);
-    }
-    return false;
-}
-
 // backend copy

 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@ -420,7 +413,7 @@ GGML_CALL static void ggml_backend_registry_init(void) {
    ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);

    // add forward decls here to avoid including the backend headers
-#ifdef GGML_USE_CUDA
+#ifdef GGML_USE_CUBLAS
    extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
    ggml_backend_cuda_reg_devices();
 #endif
@ -768,10 +761,6 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg

    if (cpu_plan->cplan.work_size > 0) {
        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-        if (cpu_plan->cplan.work_data == NULL) {
-            free(cpu_plan);
-            return NULL;
-        }
    }

    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
@ -822,11 +811,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
    switch (op->op) {
        case GGML_OP_CPY:
-            return
-                op->type != GGML_TYPE_IQ2_XXS &&
-                op->type != GGML_TYPE_IQ2_XS  &&
-                op->type != GGML_TYPE_IQ1_S   &&
-                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
        case GGML_OP_MUL_MAT:
            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
        default:
@ -849,7 +834,6 @@ static struct ggml_backend_i cpu_backend_i = {
    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
    /* .supports_op             = */ ggml_backend_cpu_supports_op,
-    /* .offload_op              = */ NULL,
    /* .event_new               = */ NULL,
    /* .event_free              = */ NULL,
    /* .event_record            = */ NULL,
@ -1015,11 +999,11 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif

 #ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 2048
+#define GGML_SCHED_MAX_SPLITS 256
 #endif

 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
-#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#define GGML_SCHED_MAX_SPLIT_INPUTS 16
 #endif

 #ifndef GGML_SCHED_MAX_COPIES
@ -1059,9 +1043,8 @@ struct ggml_backend_sched {
    struct ggml_cgraph * graph;

    // graph splits
-    struct ggml_backend_sched_split * splits;
+    struct ggml_backend_sched_split splits[GGML_SCHED_MAX_SPLITS];
    int n_splits;
-    int splits_capacity;

    // pipeline parallelism support
    int n_copies;
@ -1131,48 +1114,40 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
    // TODO: use supports_op to check if the backend supports the op

    // assign pre-allocated nodes to their backend
-    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor);
-    if (cur_backend_id != -1) {
+    // dst
+    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor);
+    if (cur_backend != -1) {
        SET_CAUSE(tensor, "1.dst");
-        return cur_backend_id;
+        return cur_backend;
    }

    // view_src
    if (tensor->view_src != NULL) {
-        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
-        if (cur_backend_id != -1) {
+        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
+        if (cur_backend != -1) {
            SET_CAUSE(tensor, "1.vsrc");
-            return cur_backend_id;
+            return cur_backend;
        }
    }

-    // graph input
+    // input
    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
-        cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
+        cur_backend = sched->n_backends - 1; // last backend (assumed CPU)
        SET_CAUSE(tensor, "1.inp");
-        return cur_backend_id;
+        return cur_backend;
    }

    // assign nodes that use weights to the backend of the weights
-    // operations with weights are preferably run on the same backend as the weights
    for (int i = 0; i < GGML_MAX_SRC; i++) {
        const struct ggml_tensor * src = tensor->src[i];
        if (src == NULL) {
            continue;
        }
        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src);
-            // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1) {
-                for (int b = 0; b < src_backend_id; b++) {
-                    if (ggml_backend_offload_op(sched->backends[b], tensor)) {
-                        SET_CAUSE(tensor, "1.off");
-                        return b;
-                    }
-                }
-            }
+            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src);
+            // operations with weights are always run on the same backend as the weights
            SET_CAUSE(tensor, "1.wgt%d", i);
-            return src_backend_id;
+            return src_backend;
        }
    }

@ -1182,9 +1157,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 static char * fmt_size(size_t size) {
    static char buffer[128];
    if (size >= 1024*1024) {
-        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
+        sprintf(buffer, "%zuM", size/1024/1024);
    } else {
-        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
+        sprintf(buffer, "%zuK", size/1024);
    }
    return buffer;
 }
@ -1252,31 +1227,28 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
    // pass 1: assign backends to ops with pre-allocated inputs
    for (int i = 0; i < graph->n_leafs; i++) {
        struct ggml_tensor * leaf = graph->leafs[i];
-        int * leaf_backend_id = &tensor_backend_id(leaf);
-        if (*leaf_backend_id != -1) {
+        if (tensor_backend_id(leaf) != -1) {
            // do not overwrite user assignments
            continue;
        }
-        *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
+        tensor_backend_id(leaf) = ggml_backend_sched_backend_id_from_cur(sched, leaf);
    }

    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
-        int * node_backend_id = &tensor_backend_id(node);
-        if (*node_backend_id != -1) {
+        if (tensor_backend_id(node) != -1) {
            // do not overwrite user assignments
            continue;
        }
-        *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+        tensor_backend_id(node) = ggml_backend_sched_backend_id_from_cur(sched, node);
        // src
        for (int j = 0; j < GGML_MAX_SRC; j++) {
            struct ggml_tensor * src = node->src[j];
            if (src == NULL) {
                continue;
            }
-            int * src_backend_id = &tensor_backend_id(src);
-            if (*src_backend_id == -1) {
-                *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+            if (tensor_backend_id(src) == -1) {
+                tensor_backend_id(src) = ggml_backend_sched_backend_id_from_cur(sched, src);
            }
        }
    }
@ -1298,20 +1270,21 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (ggml_is_view_op(node->op)) {
                continue;
            }
-            int * node_backend_id = &tensor_backend_id(node);
-            if (*node_backend_id != -1) {
-                if (*node_backend_id == sched->n_backends - 1) {
+            int tensor_backend_id = tensor_backend_id(node);
+            if (tensor_backend_id != -1) {
+                if (tensor_backend_id == sched->n_backends - 1) {
                    // skip cpu (lowest prio backend)
                    cur_backend_id = -1;
                } else {
-                    cur_backend_id = *node_backend_id;
+                    cur_backend_id = tensor_backend_id;
                }
            } else {
-                *node_backend_id = cur_backend_id;
+                tensor_backend_id(node) = cur_backend_id;
                SET_CAUSE(node, "2.2");
            }
        }
    }
+
    // pass 2.1 expand gpu up
    {
        int cur_backend_id = -1;
@ -1320,20 +1293,22 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (ggml_is_view_op(node->op)) {
                continue;
            }
-            int * node_backend_id = &tensor_backend_id(node);
-            if (*node_backend_id != -1) {
-                if (*node_backend_id == sched->n_backends - 1) {
+            int tensor_backend_id = tensor_backend_id(node);
+            if (tensor_backend_id != -1) {
+                if (tensor_backend_id == sched->n_backends - 1) {
                    // skip cpu (lowest prio backend)
                    cur_backend_id = -1;
                } else {
-                    cur_backend_id = *node_backend_id;
+                    cur_backend_id = tensor_backend_id;
                }
            } else {
-                *node_backend_id = cur_backend_id;
+                tensor_backend_id(node) = cur_backend_id;
                SET_CAUSE(node, "2.1");
            }
        }
    }
+
+
    // pass 2.4 expand rest down
    {
        int cur_backend_id = -1;
@ -1342,16 +1317,16 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (ggml_is_view_op(node->op)) {
                continue;
            }
-            int * node_backend_id = &tensor_backend_id(node);
-            if (*node_backend_id != -1) {
-                cur_backend_id = *node_backend_id;
+            int tensor_backend_id = tensor_backend_id(node);
+            if (tensor_backend_id != -1) {
+                cur_backend_id = tensor_backend_id;
            } else {
-                *node_backend_id = cur_backend_id;
+                tensor_backend_id(node) = cur_backend_id;
                SET_CAUSE(node, "2.4");
            }
        }
    }
-    // pass 2.3 expand rest up
+        // pass 2.3 expand rest up
    {
        int cur_backend_id = -1;
        for (int i = graph->n_nodes - 1; i >= 0; i--) {
@ -1359,11 +1334,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (ggml_is_view_op(node->op)) {
                continue;
            }
-            int * node_backend_id = &tensor_backend_id(node);
-            if (*node_backend_id != -1) {
-                cur_backend_id = *node_backend_id;
+            int tensor_backend_id = tensor_backend_id(node);
+            if (tensor_backend_id != -1) {
+                cur_backend_id = tensor_backend_id;
            } else {
-                *node_backend_id = cur_backend_id;
+                tensor_backend_id(node) = cur_backend_id;
                SET_CAUSE(node, "2.3");
            }
        }
@ -1376,9 +1351,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
    // pass 3: assign backends to remaining src from dst and view_src
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
-        int * cur_backend_id = &tensor_backend_id(node);
-        if (node->view_src != NULL && *cur_backend_id == -1) {
-            *cur_backend_id = tensor_backend_id(node->view_src);
+        int cur_backend_id = tensor_backend_id(node);
+        if (node->view_src != NULL && cur_backend_id == -1) {
+            cur_backend_id = tensor_backend_id(node) = tensor_backend_id(node->view_src);
            SET_CAUSE(node, "3.vsrc");
        }
        for (int j = 0; j < GGML_MAX_SRC; j++) {
@ -1386,14 +1361,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (src == NULL) {
                continue;
            }
-            int * src_backend_id = &tensor_backend_id(src);
-            if (*src_backend_id == -1) {
+            int src_backend_id = tensor_backend_id(src);
+            if (src_backend_id == -1) {
                if (src->view_src != NULL) {
                    // views are always on the same backend as the source
-                    *src_backend_id = tensor_backend_id(src->view_src);
+                    tensor_backend_id(src) = tensor_backend_id(src->view_src);
                    SET_CAUSE(src, "3.vsrc");
                } else {
-                    *src_backend_id = *cur_backend_id;
+                    tensor_backend_id(src) = cur_backend_id;
                    SET_CAUSE(src, "3.cur");
                }
            }
@ -1405,20 +1380,19 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg

    // pass 4: split graph, find tensors that need to be copied
    {
-        int i_split = 0;
-        struct ggml_backend_sched_split * split = &sched->splits[0];
+        int cur_split = 0;
        // find the backend of the first split, skipping view ops
        for (int i = 0; i < graph->n_nodes; i++) {
            struct ggml_tensor * node = graph->nodes[i];
            if (!ggml_is_view_op(node->op)) {
-                split->backend_id = tensor_backend_id(node);
+                sched->splits[0].backend_id = tensor_backend_id(node);
                break;
            }
        }
-        split->i_start = 0;
-        split->n_inputs = 0;
-        memset(split->inputs, 0, sizeof(split->inputs)); //HACK
-        int cur_backend_id = split->backend_id;
+        sched->splits[0].i_start = 0;
+        sched->splits[0].n_inputs = 0;
+        memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
+        int cur_backend_id = sched->splits[0].backend_id;
        for (int i = 0; i < graph->n_nodes; i++) {
            struct ggml_tensor * node = graph->nodes[i];

@ -1426,54 +1400,18 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                continue;
            }

-            const int node_backend_id = tensor_backend_id(node);
+            int tensor_backend_id = tensor_backend_id(node);

-            GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now
+            GGML_ASSERT(tensor_backend_id != -1); // all nodes should be assigned by now

-            // check if we should start a new split based on the sources of the current node
-            bool need_new_split = false;
-            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
-                for (int j = 0; j < GGML_MAX_SRC; j++) {
-                    struct ggml_tensor * src = node->src[j];
-                    if (src == NULL) {
-                        continue;
-                    }
-                    // check if a weight is on a different backend
-                    // by starting a new split, the memory of the previously offloaded weights can be reused
-                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-                        int src_backend_id = tensor_backend_id(src);
-                        if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
-                            need_new_split = true;
-                            break;
-                        }
-                    }
-                    // check if the split has too many inputs
-                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
-                        const size_t id = hash_id(src);
-                        int src_backend_id = sched->tensor_backend_id[id];
-                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL) {
-                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
-                            need_new_split = true;
-                            break;
-                        }
-                    }
-                }
-            }
-
-            if (node_backend_id != cur_backend_id || need_new_split) {
-                split->i_end = i;
-                i_split++;
-                if (i_split >= sched->splits_capacity) {
-                    sched->splits_capacity *= 2;
-                    sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
-                    GGML_ASSERT(sched->splits != NULL);
-                }
-                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
-                split = &sched->splits[i_split];
-                split->backend_id = node_backend_id;
-                split->i_start = i;
-                split->n_inputs = 0;
-                cur_backend_id = node_backend_id;
+            if (tensor_backend_id != cur_backend_id) {
+                sched->splits[cur_split].i_end = i;
+                cur_split++;
+                GGML_ASSERT(cur_split < GGML_SCHED_MAX_SPLITS);
+                sched->splits[cur_split].backend_id = tensor_backend_id;
+                sched->splits[cur_split].i_start = i;
+                sched->splits[cur_split].n_inputs = 0;
+                cur_backend_id = tensor_backend_id;
            }

            // find inputs that are not on the same backend
@ -1483,10 +1421,10 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                    continue;
                }

-                const int src_backend_id = tensor_backend_id(src);
+                int src_backend_id = tensor_backend_id(src);
                assert(src_backend_id != -1); // all inputs should be assigned by now

-                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1)  {
+                if (src->flags & GGML_TENSOR_FLAG_INPUT)  {
                    size_t id = hash_id(src);
                    if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
                        ggml_backend_t backend = sched->backends[src_backend_id];
@ -1503,6 +1441,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                            }
                            sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = src_backend_id;
                            SET_CAUSE(tensor_copy, "4.cpy");
                        }
                        int n_graph_inputs = sched->n_graph_inputs++;
@ -1511,9 +1450,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                    }
                }

-                if (src_backend_id != node_backend_id) {
+                if (src_backend_id != tensor_backend_id) {
                    // create a copy of the input in the split's backend
-                    const size_t id = hash_id(src);
+                    size_t id = hash_id(src);
                    if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
                        ggml_backend_t backend = sched->backends[cur_backend_id];
                        for (int c = 0; c < sched->n_copies; c++) {
@ -1524,42 +1463,76 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                            }
                            sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = cur_backend_id;
                            SET_CAUSE(tensor_copy, "4.cpy");
                        }
-                        int n_inputs = split->n_inputs++;
+                        int n_inputs = sched->splits[cur_split].n_inputs++;
                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
-                        split->inputs[n_inputs] = src;
+                        sched->splits[cur_split].inputs[n_inputs] = src;
                    }
                    node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
                }
            }
        }
-        split->i_end = graph->n_nodes;
-        sched->n_splits = i_split + 1;
+        sched->splits[cur_split].i_end = graph->n_nodes;
+        sched->n_splits = cur_split + 1;
    }
 #ifdef DEBUG_PASS4
    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif

+#ifndef NDEBUG
+    // sanity check: all sources should have the same backend as the node
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+        if (tensor_backend == NULL) {
+            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
+        }
+        if (node->view_src != NULL && tensor_backend != ggml_backend_sched_get_tensor_backend(sched, node->view_src)) {
+            fprintf(stderr, "!!!!!!! %s has backend %s, view_src %s has backend %s\n",
+                node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
+                node->view_src->name, ggml_backend_sched_get_tensor_backend(sched, node->view_src) ?
+                    ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, node->view_src)) : "NULL");
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+            if (src_backend != tensor_backend /* && src_backend != NULL */) {
+                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
+                    node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
+                    j, src->name, src_backend ? ggml_backend_name(src_backend) : "NULL");
+            }
+            if (src->view_src != NULL && src_backend != ggml_backend_sched_get_tensor_backend(sched, src->view_src)) {
+                fprintf(stderr, "!!!!!!! [src] %s has backend %s, view_src %s has backend %s\n",
+                    src->name, src_backend ? ggml_backend_name(src_backend) : "NULL",
+                    src->view_src->name, ggml_backend_sched_get_tensor_backend(sched, src->view_src) ?
+                        ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, src->view_src)) : "NULL");
+            }
+        }
+    }
+    fflush(stderr);
+#endif
+
    // create copies of the graph for each split
    // TODO: avoid this copy
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false);
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS, false);
    for (int i = 0; i < sched->n_splits; i++) {
        struct ggml_backend_sched_split * split = &sched->splits[i];
        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);

        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
        for (int j = 0; j < split->n_inputs; j++) {
-            assert(graph_copy->size > (graph_copy->n_nodes + 1));
-
            struct ggml_tensor * input = split->inputs[j];
-            const size_t input_id = hash_id(input);
-            struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy];
+            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id][sched->cur_copy];

            // add a dependency to the input source so that it is not freed before the copy is done
            struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
            input_dep->src[0] = input;
-            sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id];
+            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(input);
            graph_copy->nodes[graph_copy->n_nodes++] = input_dep;

            // add a dependency to the input copy so that it is allocated at the start of the split
@ -1568,7 +1541,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
        }

        for (int j = split->i_start; j < split->i_end; j++) {
-            assert(graph_copy->size > graph_copy->n_nodes);
            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
        }
@ -1653,12 +1625,13 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                }
                ggml_backend_tensor_copy(input, input_cpy);
            } else {
-                // wait for the split backend to finish using the input before overwriting it
                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
                } else {
                    ggml_backend_synchronize(split_backend);
+                    ggml_backend_synchronize(input_backend);
                }
+
                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
            }
        }
@ -1725,24 +1698,20 @@ ggml_backend_sched_t ggml_backend_sched_new(
    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU

-    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
+    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);

    // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
-    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
-
-    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
-    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
+    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
+    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
+    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), graph_size);
+    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), graph_size);

    sched->n_backends = n_backends;

    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;

-    const int initial_splits_capacity = 16;
-    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
-    sched->splits_capacity = initial_splits_capacity;
+    GGML_ASSERT(sched->n_copies <= GGML_SCHED_MAX_COPIES);

    for (int b = 0; b < n_backends; b++) {
        sched->backends[b] = backends[b];
@ -1773,7 +1742,6 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
    }
    ggml_gallocr_free(sched->galloc);
    ggml_free(sched->ctx);
-    free(sched->splits);
    free(sched->hash_set.keys);
    free(sched->tensor_backend_id);
    free(sched->tensor_copies);
@ -1784,20 +1752,16 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {

 void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
    // reset state for the next run
-    if (!sched->is_reset) {
-        size_t hash_size = sched->hash_set.size;
-        memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
-        memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
-        memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
+    size_t hash_size = sched->hash_set.size;
+    memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
+    memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
+    memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);

-        sched->is_reset = true;
-    }
+    sched->is_reset = true;
    sched->is_alloc = false;
 }

 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes);
-
    ggml_backend_sched_split_graph(sched, measure_graph);

    // TODO: extract this to a separate function
@ -1812,7 +1776,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 }

 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes);
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);

    ggml_backend_sched_split_graph(sched, graph);

@ -1974,10 +1938,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
    struct ggml_hash_set hash_set = {
        /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
+        /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
    };
-    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
-    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
+    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
+    bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);

    struct ggml_init_params params = {
        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
--- a/ggml-backend.h
+++ b/ggml-backend.h
@ -70,11 +70,11 @@ extern "C" {
    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);

-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
@ -137,7 +137,7 @@ extern "C" {
    /*
      Example usage:

-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be asigned
        // preferrably to run on the same backend as the buffer
        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);

--- a/ggml-common.h
+++ b/ggml-common.h
@ -377,27 +377,6 @@ typedef struct {
 } block_iq1_s;
 static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");

-// 1.75 bpw
-typedef struct {
-    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
-    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
-    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
-} block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
-
-// Used by IQ1_M quants
-typedef union {
-    ggml_half f16;
-    uint16_t  u16;
-} iq1m_scale_t;
-
 // Non-linear quants
 #define QK4_NL 32
 typedef struct {
@ -447,11 +426,10 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_

 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_SYCL)
-
 #include <cstdint>

-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
+#define GGML_TABLE_BEGIN(type, name, size) static dpct::global_memory<const type, 1> name(sycl::range<1>(size), {
+#define GGML_TABLE_END() });

 #define GGML_COMMON_IMPL
 #endif
@ -1072,7 +1050,6 @@ GGML_TABLE_END()

 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
-#define IQ1M_DELTA 0.125f
 #if defined(GGML_COMMON_IMPL_C)
 GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@ -17,17 +17,29 @@ extern "C" {

 #define GGML_CUDA_MAX_DEVICES       16

+// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
+GGML_API GGML_CALL void   ggml_init_cublas(void);
+
+// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
+GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
+
+GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
+GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
+
+GGML_API GGML_CALL bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API GGML_CALL bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+
+GGML_API GGML_CALL int    ggml_cuda_get_device_count(void);
+GGML_API GGML_CALL void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
+
 // backend API
 GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);

 GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);

-// device buffer
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

@ -35,9 +47,6 @@ GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);

-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
 #ifdef  __cplusplus
 }
 #endif
--- a/ggml-cuda/acc.cu
+++ b/ggml-cuda/acc.cu
@ -1,47 +0,0 @@
-#include "acc.cuh"
-
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
-    if (i >= ne) {
-        return;
-    }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
-    }
-}
-
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
-}
-
-void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
-
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
-
-    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
-}
--- a/ggml-cuda/acc.cuh
+++ b/ggml-cuda/acc.cuh
@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_ACC_BLOCK_SIZE 256
-
-void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/arange.cu
+++ b/ggml-cuda/arange.cu
@ -1,34 +0,0 @@
-#include "arange.cuh"
-
-static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
-    // blockIDx.x: idx of ne0 / BLOCK_SIZE
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    dst[nidx] = start + step * nidx;
-}
-
-static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
-    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
-}
-
-void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    float start;
-    float stop;
-    float step;
-    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
-    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
-    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
-
-    int64_t steps = (int64_t)ceil((stop - start) / step);
-    GGML_ASSERT(ggml_nelements(dst) == steps);
-
-    arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
-}
--- a/ggml-cuda/arange.cuh
+++ b/ggml-cuda/arange.cuh
@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_ARANGE_BLOCK_SIZE 256
-
-void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/argsort.cu
+++ b/ggml-cuda/argsort.cu
@ -1,103 +0,0 @@
-#include "argsort.cuh"
-
-template<typename T>
-static inline __device__ void ggml_cuda_swap(T & a, T & b) {
-    T tmp = a;
-    a = b;
-    b = tmp;
-}
-
-template<ggml_sort_order order>
-static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
-    // bitonic sort
-    int col = threadIdx.x;
-    int row = blockIdx.y;
-
-    if (col >= ncols_pad) {
-        return;
-    }
-
-    const float * x_row = x + row * ncols;
-    extern __shared__ int dst_row[];
-
-    // initialize indices
-    dst_row[col] = col;
-
-    __syncthreads();
-
-    for (int k = 2; k <= ncols_pad; k *= 2) {
-        for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= ncols ||
-                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (dst_row[ixj] >= ncols ||
-                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
-                    }
-                }
-            }
-            __syncthreads();
-        }
-    }
-
-    // copy the result to dst without the padding
-    if (col < ncols) {
-        dst[row * ncols + col] = dst_row[col];
-    }
-}
-
-static int next_power_of_2(int x) {
-    int n = 1;
-    while (n < x) {
-        n *= 2;
-    }
-    return n;
-}
-
-static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
-    // bitonic sort requires ncols to be power of 2
-    const int ncols_pad = next_power_of_2(ncols);
-
-    const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
-    const size_t shared_mem = ncols_pad * sizeof(int);
-
-    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
-
-    if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-    } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-    } else {
-        GGML_ASSERT(false);
-    }
-}
-
-void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_I32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-
-    const int64_t ncols = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
-
-    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
-}
--- a/ggml-cuda/argsort.cuh
+++ b/ggml-cuda/argsort.cuh
@ -1,3 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/binbcast.cu
+++ b/ggml-cuda/binbcast.cu
@ -1,280 +0,0 @@
-#include "binbcast.cuh"
-
-static __device__ __forceinline__ float op_repeat(const float a, const float b) {
-    return b;
-    GGML_UNUSED(a);
-}
-
-static __device__ __forceinline__ float op_add(const float a, const float b) {
-    return a + b;
-}
-
-static __device__ __forceinline__ float op_mul(const float a, const float b) {
-    return a * b;
-}
-
-static __device__ __forceinline__ float op_div(const float a, const float b) {
-    return a / b;
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13) {
-    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
-    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
-    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-    }
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13) {
-
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-}
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_cuda {
-    template<typename src0_t, typename src1_t, typename dst_t>
-    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
-            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
-            cudaStream_t stream) {
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        int nr0 = ne10/ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
-                }
-            }
-        }
-
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            //int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
-            //int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
-            //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
-            //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s00 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            dim3 block_dims;
-            block_dims.x = std::min<unsigned int>(hne0, block_size);
-            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
-            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
-
-            dim3 block_nums(
-                (hne0 + block_dims.x - 1) / block_dims.x,
-                (ne1 + block_dims.y - 1) / block_dims.y,
-                (ne2*ne3 + block_dims.z - 1) / block_dims.z
-            );
-
-            if (block_nums.z > 65535) {
-                // this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
-                    /* s10, */ s11, s12, s13);
-            } else {
-                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
-                    /* s10, */ s11, s12, s13);
-            }
-        }
-    }
-};
-
-template<class op>
-static void ggml_cuda_op_bin_bcast(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
-    }
-}
-
-void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
slaren	66df44b0b7	alloc : fix allocation data of pre-allocated leafs	2024-03-16 15:47:14 +01:00
Georgi Gerganov	c6c94de43a	whisper : allocate encoder results in dedicated buffer	2024-03-16 16:15:12 +02:00