ci : enable git lfs for build.yml

This reverts commit 97e40df5d6.

.devops/cloud-v-pipeline

@@ -15,7 +15,7 @@ node('x86_runner1'){            // Running on x86 runner containing latest vecto
     stage('Running llama.cpp'){
         sh'''#!/bin/bash
             module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
-            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
             cat llama_log.txt                   # Printing results
         '''
     }

.devops/full-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -31,6 +31,6 @@ ENV LLAMA_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
 
-RUN make -j$(nproc)
+RUN make
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-rocm.Dockerfile

@@ -45,6 +45,6 @@ ENV LLAMA_CURL=1
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-RUN make -j$(nproc)
+RUN make
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -18,7 +18,7 @@ COPY . .
 ENV LLAMA_CURL=1
 
 
-RUN make -j$(nproc)
+RUN make
 
 ENV LC_ALL=C.utf8
 

.devops/llama-cpp-clblast.srpm.spec

@@ -36,9 +36,9 @@ make -j LLAMA_CLBLAST=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p llama-cli %{buildroot}%{_bindir}/llama-clblast-cli
-cp -p llama-server %{buildroot}%{_bindir}/llama-clblast-server
-cp -p llama-simple %{buildroot}%{_bindir}/llama-clblast-simple
+cp -p main %{buildroot}%{_bindir}/llamaclblast
+cp -p server %{buildroot}%{_bindir}/llamaclblastserver
+cp -p simple %{buildroot}%{_bindir}/llamaclblastsimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
 %{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamaclblast.service
@@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llama-clblast-server $LLAMA_ARGS
+ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,9 +67,9 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llama-clblast-cli
-%{_bindir}/llama-clblast-server
-%{_bindir}/llama-clblast-simple
+%{_bindir}/llamaclblast
+%{_bindir}/llamaclblastserver
+%{_bindir}/llamaclblastsimple
 /usr/lib/systemd/system/llamaclblast.service
 %config /etc/sysconfig/llama
 

.devops/llama-cpp-cuda.srpm.spec

@@ -36,9 +36,9 @@ make -j LLAMA_CUDA=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
-cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
-cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
+cp -p main %{buildroot}%{_bindir}/llamacppcuda
+cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
 %{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
@@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llama-cuda-server $LLAMA_ARGS
+ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,9 +67,9 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llama-cuda-cli
-%{_bindir}/llama-cuda-server
-%{_bindir}/llama-cuda-simple
+%{_bindir}/llamacppcuda
+%{_bindir}/llamacppcudaserver
+%{_bindir}/llamacppcudasimple
 /usr/lib/systemd/system/llamacuda.service
 %config /etc/sysconfig/llama
 

.devops/llama-cpp.srpm.spec

@@ -38,9 +38,9 @@ make -j
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
-cp -p llama-server %{buildroot}%{_bindir}/llama-server
-cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
+cp -p main %{buildroot}%{_bindir}/llama
+cp -p server %{buildroot}%{_bindir}/llamaserver
+cp -p simple %{buildroot}%{_bindir}/llamasimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
 %{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llama.service
@@ -51,7 +51,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llama-server $LLAMA_ARGS
+ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -69,9 +69,9 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llama-cli
-%{_bindir}/llama-server
-%{_bindir}/llama-simple
+%{_bindir}/llama
+%{_bindir}/llamaserver
+%{_bindir}/llamasimple
 /usr/lib/systemd/system/llama.service
 %config /etc/sysconfig/llama
 

.devops/main-cuda.Dockerfile

@@ -23,13 +23,10 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
 ENV LLAMA_CUDA=1
 
-RUN make -j$(nproc) llama-cli
+RUN make
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
-RUN apt-get update && \
-    apt-get install -y libgomp1
+COPY --from=build /app/main /main
 
-COPY --from=build /app/llama-cli /llama-cli
-
-ENTRYPOINT [ "/llama-cli" ]
+ENTRYPOINT [ "/main" ]

.devops/main-intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
@@ -15,12 +15,12 @@ RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
         export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
     fi && \
     cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
-    cmake --build build --config Release --target llama-cli
+    cmake --build build --config Release --target main
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
-COPY --from=build /app/build/bin/llama-cli /llama-cli
+COPY --from=build /app/build/bin/main /main
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-cli" ]
+ENTRYPOINT [ "/main" ]

.devops/main-rocm.Dockerfile

@@ -40,6 +40,6 @@ ENV LLAMA_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
 
-RUN make -j$(nproc) llama-cli
+RUN make
 
-ENTRYPOINT [ "/app/llama-cli" ]
+ENTRYPOINT [ "/app/main" ]

.devops/main-vulkan.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=jammy
 FROM ubuntu:$UBUNTU_VERSION as build
 
 # Install build tools
-RUN apt update && apt install -y git build-essential cmake wget libgomp1
+RUN apt update && apt install -y git build-essential cmake wget
 
 # Install Vulkan SDK
 RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
@@ -15,13 +15,13 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
 WORKDIR /app
 COPY . .
 RUN cmake -B build -DLLAMA_VULKAN=1 && \
-    cmake --build build --config Release --target llama-cli
+    cmake --build build --config Release --target main
 
 # Clean up
 WORKDIR /
-RUN cp /app/build/bin/llama-cli /llama-cli && \
+RUN cp /app/build/bin/main /main && \
     rm -rf /app
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-cli" ]
+ENTRYPOINT [ "/main" ]

.devops/main.Dockerfile

@@ -9,15 +9,12 @@ WORKDIR /app
 
 COPY . .
 
-RUN make -j$(nproc) llama-cli
+RUN make
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 
-RUN apt-get update && \
-    apt-get install -y libgomp1
-
-COPY --from=build /app/llama-cli /llama-cli
+COPY --from=build /app/main /main
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-cli" ]
+ENTRYPOINT [ "/main" ]

.devops/nix/apps.nix

@@ -6,11 +6,11 @@
         let
           inherit (config.packages) default;
           binaries = [
-            "llama-cli"
+            "llama"
             "llama-embedding"
             "llama-server"
-            "llama-quantize"
-            "llama-train-text-from-scratch"
+            "quantize"
+            "train-text-from-scratch"
           ];
           mkApp = name: {
             type = "app";

.devops/nix/package.nix

@@ -214,6 +214,7 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_CUDA" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
+        (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
         (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
@@ -226,23 +227,25 @@ effectiveStdenv.mkDerivation (
         )
       ]
       ++ optionals useRocm [
-        (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
-        (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
+        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
+        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
+
+        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
+        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
+        # and select the line that matches the current nixpkgs version of rocBLAS.
+        # Should likely use `rocmPackages.clr.gpuTargets`.
+        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
       ]
       ++ optionals useMetalKit [
         (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
         (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
       ];
 
-    # Environment variables needed for ROCm
-    env = optionals useRocm {
-      ROCM_PATH = "${rocmPackages.clr}";
-      HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
-    };
-
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''
+      mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
+      mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
       mkdir -p $out/include
       cp $src/llama.h $out/include/
     '';
@@ -292,7 +295,7 @@ effectiveStdenv.mkDerivation (
       license = lib.licenses.mit;
 
       # Accommodates `nix run` and `lib.getExe`
-      mainProgram = "llama-cli";
+      mainProgram = "llama";
 
       # These people might respond, on the best effort basis, if you ping them
       # in case of Nix-specific regressions or for reviewing Nix-specific PRs.

.devops/server-cuda.Dockerfile

@@ -25,13 +25,13 @@ ENV LLAMA_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
 
-RUN make -j$(nproc) llama-server
+RUN make
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/server /server
 
-ENTRYPOINT [ "/llama-server" ]
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
@@ -15,15 +15,15 @@ RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
         export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
     fi && \
     cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
-    cmake --build build --config Release --target llama-server
+    cmake --build build --config Release --target server
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-COPY --from=build /app/build/bin/llama-server /llama-server
+COPY --from=build /app/build/bin/server /server
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-server" ]
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -45,6 +45,6 @@ ENV LLAMA_CURL=1
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-RUN make -j$(nproc) llama-server
+RUN make
 
-ENTRYPOINT [ "/app/llama-server" ]
+ENTRYPOINT [ "/app/server" ]

.devops/server-vulkan.Dockerfile

@@ -19,13 +19,13 @@ RUN apt-get update && \
 WORKDIR /app
 COPY . .
 RUN cmake -B build -DLLAMA_VULKAN=1 -DLLAMA_CURL=1 && \
-    cmake --build build --config Release --target llama-server
+    cmake --build build --config Release --target server
 
 # Clean up
 WORKDIR /
-RUN cp /app/build/bin/llama-server /llama-server && \
+RUN cp /app/build/bin/server /server && \
     rm -rf /app
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-server" ]
+ENTRYPOINT [ "/server" ]

.devops/server.Dockerfile

@@ -11,15 +11,15 @@ COPY . .
 
 ENV LLAMA_CURL=1
 
-RUN make -j$(nproc) llama-server
+RUN make
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/server /server
 
 ENV LC_ALL=C.utf8
 
-ENTRYPOINT [ "/llama-server" ]
+ENTRYPOINT [ "/server" ]

.devops/tools.sh

@@ -8,13 +8,13 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert-hf-to-gguf.py "$@"
+    python3 ./convert.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./llama-quantize "$@"
+    ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./llama-cli "$@"
+    ./main "$@"
 elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
-    ./llama-finetune "$@"
+    ./finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -22,11 +22,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
             echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
         else
             echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..."
-            ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
+            ./quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./llama-server "$@"
+    ./server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "

.dockerignore

@@ -12,8 +12,8 @@ build*/
 
 models/*
 
-/llama-cli
-/llama-quantize
+/main
+/quantize
 
 arm_neon.h
 compile_commands.json

.editorconfig

@@ -26,6 +26,3 @@ indent_size = 2
 
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
-
-[examples/cvector-generator/*.txt]
-insert_final_newline = unset

.gitattributes

@@ -0,0 +1 @@
+models/ggml-vocab-*.gguf filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/01-bug-low.yml

@@ -1,50 +0,0 @@
-name: Low Severity Bugs
-description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
-title: "Bug: "
-labels: ["bug-unconfirmed", "low severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell

.github/ISSUE_TEMPLATE/02-bug-medium.yml

@@ -1,50 +0,0 @@
-name: Medium Severity Bug
-description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
-title: "Bug: "
-labels: ["bug-unconfirmed", "medium severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell

.github/ISSUE_TEMPLATE/03-bug-high.yml

@@ -1,50 +0,0 @@
-name: High Severity Bug
-description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
-title: "Bug: "
-labels: ["bug-unconfirmed", "high severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell

.github/ISSUE_TEMPLATE/04-bug-critical.yml

@@ -1,50 +0,0 @@
-name: Critical Severity Bug
-description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
-title: "Bug: "
-labels: ["bug-unconfirmed", "critical severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell

.github/ISSUE_TEMPLATE/05-enhancement.yml

@@ -1,51 +0,0 @@
-name: Enhancement
-description: Used to request enhancements for llama.cpp
-title: "Feature Request: "
-labels: ["enhancement"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
-
-  - type: checkboxes
-    id: prerequisites
-    attributes:
-      label: Prerequisites
-      description: Please confirm the following before submitting your enhancement request.
-      options:
-        - label: I am running the latest code. Mention the version if possible as well.
-          required: true
-        - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
-          required: true
-        - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
-          required: true
-        - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
-          required: true
-
-  - type: textarea
-    id: feature-description
-    attributes:
-      label: Feature Description
-      description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
-      placeholder: Detailed description of the enhancement
-    validations:
-      required: true
-
-  - type: textarea
-    id: motivation
-    attributes:
-      label: Motivation
-      description: Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
-      placeholder: Explanation of why this feature is needed and its benefits
-    validations:
-      required: true
-
-  - type: textarea
-    id: possible-implementation
-    attributes:
-      label: Possible Implementation
-      description: If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.
-      placeholder: Detailed description of potential implementation
-    validations:
-      required: false

.github/ISSUE_TEMPLATE/06-research.yml

@@ -1,52 +0,0 @@
-name: Research
-description: Track new technical research area
-title: "Research: "
-labels: ["research 🔬"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
-
-  - type: checkboxes
-    id: research-stage
-    attributes:
-      label: Research Stage
-      description: Track general state of this research ticket
-      options:
-        - label: Background Research (Let's try to avoid reinventing the wheel)
-        - label: Hypothesis Formed (How do you think this will work and it's effect?)
-        - label: Strategy / Implementation Forming
-        - label: Analysis of results
-        - label: Debrief / Documentation (So people in the future can learn from us)
-
-  - type: textarea
-    id: background
-    attributes:
-      label: Previous existing literature and research
-      description: Whats the current state of the art and whats the motivation for this research?
-
-  - type: textarea
-    id: hypothesis
-    attributes:
-      label: Hypothesis
-      description: How do you think this will work and it's effect?
-
-  - type: textarea
-    id: implementation
-    attributes:
-      label: Implementation
-      description: Got an approach? e.g. a PR ready to go?
-
-  - type: textarea
-    id: analysis
-    attributes:
-      label: Analysis
-      description: How does the proposed implementation behave?
-
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell

.github/ISSUE_TEMPLATE/07-refactor.yml

@@ -1,28 +0,0 @@
-name: Refactor (Maintainers)
-description: Used to track refactoring opportunities
-title: "Refactor: "
-labels: ["refactor"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
-        Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
-
-  - type: textarea
-    id: background-description
-    attributes:
-      label: Background Description
-      description: Please provide a detailed written description of the pain points you are trying to solve.
-      placeholder: Detailed description behind your motivation to request refactor
-    validations:
-      required: true
-
-  - type: textarea
-    id: possible-approaches
-    attributes:
-      label: Possible Refactor Approaches
-      description: If you have some idea of possible approaches to solve this problem. You may want to make it a todo list.
-      placeholder: Your idea of possible refactoring opportunity/approaches
-    validations:
-      required: false

.github/ISSUE_TEMPLATE/bug.md

@@ -0,0 +1,11 @@
+---
+name: Bug template
+about: Used to report bugs in llama.cpp
+labels: ["bug-unconfirmed"]
+assignees: ''
+
+---
+
+Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
+
+If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/ISSUE_TEMPLATE/config.yml

@@ -1,13 +0,0 @@
-blank_issues_enabled: true
-contact_links:
-  - name: Got an idea?
-    url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas
-    about: Pop it there. It may then become an enhancement ticket.
-  - name: Got a question?
-    url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a
-    about: Ask a question there!
-  - name: Want to contribute?
-    url: https://github.com/ggerganov/llama.cpp/wiki/contribute
-    about: Head to the contribution guide page of the wiki for areas you can help with
-
-

.github/ISSUE_TEMPLATE/enhancement.md

@@ -0,0 +1,28 @@
+---
+name: Enhancement template
+about: Used to request enhancements for llama.cpp
+labels: ["enhancement"]
+assignees: ''
+
+---
+
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Feature Description
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+
+# Motivation
+
+Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+
+# Possible Implementation
+
+If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.github/labeler.yml

@@ -1,90 +0,0 @@
-# https://github.com/actions/labeler
-Kompute:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml-kompute.h
-            - ggml-kompute.cpp
-            - README-kompute.md
-Apple Metal:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml-metal.h
-            - ggml-metal.cpp
-            - README-metal.md
-SYCL:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml-sycl.h
-            - ggml-sycl.cpp
-            - README-sycl.md
-Nvidia GPU:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml-cuda.h
-            - ggml-cuda/**
-Vulkan:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml_vk_generate_shaders.py
-            - ggml-vulkan*
-documentation:
-    - changed-files:
-        - any-glob-to-any-file:
-            - docs/**
-            - media/**
-testing:
-    - changed-files:
-        - any-glob-to-any-file:
-            - tests/**
-build:
-    - changed-files:
-        - any-glob-to-any-file:
-            - cmake/**
-            - CMakeLists.txt
-            - CMakePresets.json
-            - codecov.yml
-examples:
-    - changed-files:
-        - any-glob-to-any-file: examples/**
-devops:
-    - changed-files:
-        - any-glob-to-any-file:
-            - .devops/**
-            - .github/**
-            - ci/**
-python:
-    - changed-files:
-        - any-glob-to-any-file:
-            - "**/*.py"
-            - requirements/**
-            - gguf-py/**
-            - .flake8
-script:
-    - changed-files:
-        - any-glob-to-any-file:
-            - scripts/**
-android:
-    - changed-files:
-        - any-glob-to-any-file:
-            - examples/llama.android/**
-server:
-    - changed-files:
-        - any-glob-to-any-file:
-            - examples/server/**
-ggml:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml.c
-            - ggml.h
-            - ggml-*.c
-            - ggml-*.h
-            - ggml-cuda/**
-nix:
-    - changed-files:
-        - any-glob-to-any-file:
-            - "**/*.nix"
-            - .github/workflows/nix-*.yml
-            - .devops/nix/nixpkgs-instances.nix
-embedding:
-    - changed-files:
-        - any-glob-to-any-file: examples/embedding/

.github/pull_request_template.md

@@ -1,7 +0,0 @@
-
-
-- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
-- Self-reported review complexity:
-  - [ ] Low
-  - [ ] Medium
-  - [ ] High

.github/workflows/bench.yml

@@ -119,7 +119,7 @@ jobs:
               -DLLAMA_FATAL_WARNINGS=OFF \
               -DLLAMA_ALL_WARNINGS=OFF \
               -DCMAKE_BUILD_TYPE=Release;
-          cmake --build build --config Release -j $(nproc) --target llama-server
+          cmake --build build --config Release -j $(nproc) --target server
 
       - name: Download the dataset
         id: download_dataset

.github/workflows/build.yml

@@ -13,7 +13,7 @@ on:
     paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -33,6 +33,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -84,13 +85,14 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
-    runs-on: macos-12
+    runs-on: macos-latest
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -103,10 +105,12 @@ jobs:
         id: cmake_build
         run: |
           sysctl -a
+          mkdir build
+          cd build
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF -DLLAMA_CURL=ON
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: cmake_test
@@ -151,6 +155,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -186,6 +192,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -209,6 +217,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Dependencies
@@ -239,8 +248,8 @@ jobs:
           wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
           echo "Fetch llama2c model"
           wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
-          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
-          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
       - name: Determine tag name
         id: tag
@@ -269,73 +278,71 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
           name: llama-bin-ubuntu-x64.zip
 
-  ubuntu-latest-cmake-sanitizer:
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#        with:
+#          lfs: true
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+#
+#      - name: Test
+#        id: cmake_test
+#        run: |
+#          cd build
+#          ctest -L main --verbose --timeout 900
+
+  ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
 
     continue-on-error: true
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        mpi_library: [mpich, libopenmpi-dev]
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
-
-      - name: Build
-        id: cmake_build
-        if: ${{ matrix.sanitizer != 'THREAD' }}
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_OPENMP=OFF
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-  ubuntu-latest-cmake-rpc:
-    runs-on: ubuntu-latest
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential ${{ matrix.mpi_library }}
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake -DLLAMA_RPC=ON ..
+          cmake -DLLAMA_MPI=ON ..
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -351,6 +358,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -366,33 +375,6 @@ jobs:
           cmake -DLLAMA_VULKAN=ON ..
           cmake --build . --config Release -j $(nproc)
 
-  ubuntu-22-cmake-hip:
-    runs-on: ubuntu-22.04
-    container: rocm/dev-ubuntu-22.04:6.0.2
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
-
-      - name: Build with native CMake HIP support
-        id: cmake_build
-        run: |
-          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DLLAMA_HIPBLAS=ON
-          cmake --build build --config Release -j $(nproc)
-
-      - name: Build with legacy HIP support
-        id: cmake_build_legacy_hip
-        run: |
-          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DLLAMA_HIPBLAS=ON
-          cmake --build build2 --config Release -j $(nproc)
-
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -400,6 +382,8 @@ jobs:
 
     steps:
       - uses: actions/checkout@v2
+        with:
+          lfs: true
 
       - name: add oneAPI to apt
         shell: bash
@@ -424,6 +408,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build
         id: cmake_build
@@ -441,6 +427,8 @@ jobs:
 
     steps:
       - uses: actions/checkout@v2
+        with:
+          lfs: true
 
       - name: add oneAPI to apt
         shell: bash
@@ -465,6 +453,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build
         id: cmake_build
@@ -485,6 +475,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -516,6 +508,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -545,6 +539,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -574,6 +570,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -607,6 +605,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
+        with:
+          lfs: true
 
       - name: Dependencies
         id: depends
@@ -637,6 +637,8 @@ jobs:
     steps:
       - name: Clone
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Setup ${{ matrix.sys }}
         uses: msys2/setup-msys2@v2
@@ -682,53 +684,75 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
+      OPENCL_VERSION: 2023.04.17
+      CLBLAST_VERSION: 1.6.0
       SDE_VERSION: 9.33.0-2024-01-07
       VULKAN_VERSION: 1.3.261.1
 
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_RPC=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'noavx-x64'
+          - build: 'noavx'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx2-x64'
+          - build: 'avx2'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx-x64'
+          - build: 'avx'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx512-x64'
+          - build: 'avx512'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'openblas-x64'
+          - build: 'clblast'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+          - build: 'openblas'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'kompute-x64'
+          - build: 'kompute'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'vulkan-x64'
+          - build: 'vulkan'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'llvm-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'msvc-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'arm64'
+            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Clone Kompute submodule
         id: clone_kompute
-        if: ${{ matrix.build == 'kompute-x64' }}
+        if: ${{ matrix.build == 'kompute' }}
         run: |
           git submodule update --init kompute
 
+      - name: Download OpenCL SDK
+        id: get_opencl
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip"
+          mkdir $env:RUNNER_TEMP/opencl
+          tar.exe -xvf $env:RUNNER_TEMP/opencl.zip --strip-components=1 -C $env:RUNNER_TEMP/opencl
+
+      - name: Download CLBlast
+        id: get_clblast
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
+          curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/clblast.7z
+          rename-item $env:RUNNER_TEMP/CLBlast-${env:CLBLAST_VERSION}-windows-x64 clblast
+          foreach ($f in (gci -Recurse -Path "$env:RUNNER_TEMP/clblast" -Filter '*.cmake')) {
+            $txt = Get-Content -Path $f -Raw
+            $txt.Replace('C:/vcpkg/packages/opencl_x64-windows/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
+          }
+
       - name: Download OpenBLAS
         id: get_openblas
-        if: ${{ matrix.build == 'openblas-x64' }}
+        if: ${{ matrix.build == 'openblas' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
           curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
@@ -741,34 +765,38 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
-      - name: Install Ninja
-        id: install_ninja
-        run: |
-          choco install ninja
-
       - name: Build
         id: cmake_build
         run: |
-          cmake -S . -B build ${{ matrix.defines }}
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          mkdir build
+          cd build
+          cmake .. ${{ matrix.defines }}
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Add clblast.dll
+        id: add_clblast_dll
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release
+          cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas-x64' }}
+        if: ${{ matrix.build == 'openblas' }}
         run: |
           cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
           cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
 
       - name: Check AVX512F support
         id: check_avx512f
-        if: ${{ matrix.build == 'avx512-x64' }}
+        if: ${{ matrix.build == 'avx512' }}
         continue-on-error: true
         run: |
           cd build
@@ -782,14 +810,14 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
         run: |
           curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
           # for some weird reason windows tar doesn't like sde tar.xz
@@ -817,17 +845,17 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
-          name: llama-bin-win-${{ matrix.build }}.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
+          name: llama-bin-win-${{ matrix.build }}-x64.zip
 
   windows-latest-cmake-cuda:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -839,11 +867,11 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
-      - name: Install CUDA toolkit
+      - uses: Jimver/cuda-toolkit@v0.2.11
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
         with:
           cuda: ${{ matrix.cuda }}
           method: 'network'
@@ -905,14 +933,15 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
-      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Install
@@ -939,17 +968,6 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
-          echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -959,43 +977,14 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
-  windows-latest-cmake-hip:
-    runs-on: windows-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Install
-        id: depends
-        run: |
-          $ErrorActionPreference = "Stop"
-          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP SDK"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        id: verify
-        run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
-
-      - name: Build
-        id: cmake_build
-        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DLLAMA_HIPBLAS=ON
-          cmake --build build --config Release
-
   ios-xcode-build:
     runs-on: macos-latest
 
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
@@ -1006,6 +995,8 @@ jobs:
     steps:
       - name: Clone
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -1028,7 +1019,9 @@ jobs:
 #    runs-on: macos-12
 #    steps:
 #    - name: Clone
-#      uses: actions/checkout@v4
+#      uses: actions/checkout@#v4
+#        with:
+#          lfs: true
 #
 #    - name: Build
 #      uses: cross-platform-actions/action@v0.19.0
@@ -1038,7 +1031,7 @@ jobs:
 #        hypervisor: 'qemu'
 #        run: |
 #            sudo pkg update
-#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 clinfo clover opencl clblast openblas
 #            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
 
   release:
@@ -1061,6 +1054,7 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
+          lfs: true
           fetch-depth: 0
 
       - name: Determine tag name
@@ -1126,6 +1120,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1150,6 +1146,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1174,6 +1172,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |
@@ -1204,6 +1204,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1243,6 +1245,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1289,6 +1293,8 @@ jobs:
 #    steps:
 #      - name: Clone
 #        uses: actions/checkout@v4
+#        with:
+#          lfs: true
 #
 #      - name: Dependencies
 #        run: |

.github/workflows/code-coverage.yml

@@ -13,14 +13,16 @@ jobs:
   run:
     runs-on: ubuntu-20.04
     steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
       - name: Dependencies
         run: |
           sudo apt-get update
           sudo apt-get install build-essential gcc-8 lcov
 
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
       - name: Build
         run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
 

.github/workflows/docker.yml

@@ -30,20 +30,20 @@ jobs:
     strategy:
       matrix:
         config:
-          - { tag: "light", dockerfile: ".devops/llama-cli.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server", dockerfile: ".devops/llama-server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
           #                     have disabled them for now until the reason why
           #                     is understood.
-          - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

.github/workflows/labeler.yml

@@ -1,17 +0,0 @@
-name: "Pull Request Labeler"
-on:
-- pull_request_target
-
-jobs:
-  labeler:
-    permissions:
-      contents: read
-      pull-requests: write
-    runs-on: ubuntu-latest
-    steps:
-    - uses: actions/checkout@v4
-      with:
-        repository: "ggerganov/llama.cpp"
-    - uses: actions/labeler@v5
-      with:
-        configuration-path: '.github/labeler.yml'

.github/workflows/server.yml

@@ -16,9 +16,11 @@ on:
     branches:
       - master
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
-  pull_request:
+  pull_request_target:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+  schedule:
+    -  cron: '2 4 * * *'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
@@ -30,8 +32,10 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [RelWithDebInfo]
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
+        build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
@@ -96,10 +100,12 @@ jobs:
               -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server
+
 
       - name: Tests
         id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh
@@ -113,7 +119,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     steps:
       - name: Clone
@@ -136,7 +142,7 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
 
       - name: Python setup
         id: setup_python

.github/workflows/zig-build.yml

@@ -0,0 +1,29 @@
+name: Zig CI
+
+on:
+  pull_request:
+  push:
+    branches:
+      - master
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    strategy:
+      fail-fast: false
+      matrix:
+        runs-on: [ubuntu-latest, macos-latest, windows-latest]
+    runs-on: ${{ matrix.runs-on }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+          fetch-depth: 0
+      - uses: goto-bus-stop/setup-zig@v2
+        with:
+          version: 0.11.0
+      - name: Build Summary
+        run: zig build --summary all -freference-trace

.gitignore

@@ -34,11 +34,9 @@ ggml-metal-embed.metal
 lcov-report/
 gcovr-report/
 
-tags
 build*
 !build.zig
 cmake-build-*
-android-ndk-*
 out/
 tmp/
 
@@ -46,9 +44,48 @@ models/*
 models-mnt
 
 /Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
+/embd-input-test
+/embedding
+/eval-callback
+/gguf
+/gguf-llama-simple
+/gguf-split
+/gritlm
+/imatrix
+/infill
 /libllama.so
-/llama-*
-llama-batched-swift
+/llama-bench
+/llava-cli
+/lookahead
+/lookup
+/lookup-create
+/lookup-merge
+/lookup-stats
+/main
+/metal
+/passkey
+/perplexity
+/q8dot
+/quantize
+/quantize-stats
+/result
+/save-load-state
+/server
+/simple
+/batched
+/batched-bench
+/export-lora
+/finetune
+/retrieval
+/speculative
+/parallel
+/train-text-from-scratch
+/tokenize
+/vdot
 /common/build-info.cpp
 arm_neon.h
 compile_commands.json
@@ -68,7 +105,6 @@ examples/jeopardy/results.txt
 examples/server/*.html.hpp
 examples/server/*.js.hpp
 examples/server/*.mjs.hpp
-examples/server/*.css.hpp
 
 poetry.lock
 poetry.toml

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
@@ -39,12 +39,8 @@ endif()
 
 if (APPLE)
     set(LLAMA_METAL_DEFAULT ON)
-    set(LLAMA_BLAS_DEFAULT ON)
-    set(LLAMA_BLAS_VENDOR_DEFAULT "Apple")
 else()
     set(LLAMA_METAL_DEFAULT OFF)
-    set(LLAMA_BLAS_DEFAULT OFF)
-    set(LLAMA_BLAS_VENDOR_DEFAULT "Generic")
 endif()
 
 set(LLAMA_LLAMAFILE_DEFAULT ON)
@@ -76,13 +72,11 @@ else()
     set(INS_ENB ON)
 endif()
 
-option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
 option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
 option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
 option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
-option(LLAMA_AVX512_BF16                     "llama: enable AVX512-BF16"                        OFF)
 option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
@@ -95,10 +89,9 @@ endif()
 
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
-option(LLAMA_BLAS                            "llama: use BLAS"                                  ${LLAMA_BLAS_DEFAULT})
-set(LLAMA_BLAS_VENDOR ${LLAMA_BLAS_VENDOR_DEFAULT} CACHE STRING
-                                             "llama: BLAS library vendor")
+option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 option(LLAMA_LLAMAFILE                       "llama: use llamafile SGEMM"                       ${LLAMA_LLAMAFILE_DEFAULT})
+set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
 option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
@@ -111,15 +104,14 @@ set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
 option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
 option(LLAMA_CUDA_NO_VMM                     "llama: do not try to use CUDA VMM"                OFF)
-option(LLAMA_CUDA_FA_ALL_QUANTS              "llama: compile all quants for FlashAttention"     OFF)
 
 option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
+option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 option(LLAMA_VULKAN_CHECK_RESULTS            "llama: run Vulkan op checks"                      OFF)
 option(LLAMA_VULKAN_DEBUG                    "llama: enable Vulkan debug output"                OFF)
-option(LLAMA_VULKAN_MEMORY_DEBUG             "llama: enable Vulkan memory debug output"         OFF)
 option(LLAMA_VULKAN_VALIDATE                 "llama: enable Vulkan validation"                  OFF)
 option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"                          OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
@@ -130,8 +122,8 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                              "llama: metal minimum macOS version")
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
-option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
-option(LLAMA_OPENMP                          "llama: use OpenMP"                                ON)
+option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
+option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -141,8 +133,6 @@ set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeli
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
-option(LLAMA_LASX                            "llama: enable lasx"                               ON)
-option(LLAMA_LSX                             "llama: enable lsx"                                ON)
 
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
@@ -302,24 +292,13 @@ if (LLAMA_METAL)
         )
 endif()
 
-if (LLAMA_OPENMP)
-    find_package(OpenMP)
-    if (OpenMP_FOUND)
-        message(STATUS "OpenMP found")
-        add_compile_definitions(GGML_USE_OPENMP)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
-    else()
-        message(WARNING "OpenMP not found")
-    endif()
-endif()
-
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
     endif()
-    #if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
-    #    set(BLA_SIZEOF_INTEGER 8)
-    #endif()
+    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+        set(BLA_SIZEOF_INTEGER 8)
+    endif()
 
     set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
@@ -327,7 +306,7 @@ if (LLAMA_BLAS)
     if (BLAS_FOUND)
         message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
-        if (("${BLAS_INCLUDE_DIRS}" STREQUAL "") AND NOT (${LLAMA_BLAS_VENDOR} MATCHES "Apple"))
+        if ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
             # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
             # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
             find_package(PkgConfig REQUIRED)
@@ -380,15 +359,12 @@ if (LLAMA_BLAS)
 
         add_compile_options(${BLAS_LINKER_FLAGS})
 
-        add_compile_definitions(GGML_USE_BLAS)
+        add_compile_definitions(GGML_USE_OPENBLAS)
 
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
 
-        set(GGML_HEADERS_BLAS ggml-blas.h)
-        set(GGML_SOURCES_BLAS ggml-blas.cpp)
-
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
     else()
@@ -405,45 +381,30 @@ if (LLAMA_LLAMAFILE)
     set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
 endif()
 
+if (LLAMA_QKK_64)
+    add_compile_definitions(GGML_QKK_64)
+endif()
+
 if (LLAMA_CUBLAS)
     message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
     set(LLAMA_CUDA ON)
 endif()
 
 if (LLAMA_CUDA)
-    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+    cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
     if (CUDAToolkit_FOUND)
         message(STATUS "CUDA found")
 
-        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-            # 52 == lowest CUDA 12 standard
-            # 60 == f16 CUDA intrinsics
-            # 61 == integer CUDA intrinsics
-            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-            if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
-                set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
-                #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
-            endif()
-        endif()
-        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
-
         enable_language(CUDA)
 
         set(GGML_HEADERS_CUDA ggml-cuda.h)
 
         file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
         list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
-        file(GLOB SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB SRCS "ggml-cuda/template-instances/mmq*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
 
         add_compile_definitions(GGML_USE_CUDA)
-        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -466,22 +427,10 @@ if (LLAMA_CUDA)
         if (LLAMA_CUDA_NO_PEER_COPY)
             add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
         endif()
-        if (LLAMA_CUDA_FA_ALL_QUANTS)
-            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-        else()
-            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
-                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
                 set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
                 set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -495,20 +444,69 @@ if (LLAMA_CUDA)
         else()
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
         endif()
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # 52 == lowest CUDA 12 standard
+        # 60 == f16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
     else()
         message(WARNING "CUDA not found")
     endif()
 endif()
 
-if (LLAMA_RPC)
-    add_compile_definitions(GGML_USE_RPC)
+if (LLAMA_MPI)
+    cmake_minimum_required(VERSION 3.10)
+    find_package(MPI)
+    if (MPI_C_FOUND)
+        message(STATUS "MPI found")
 
-    if (WIN32)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ws2_32)
+        set(GGML_HEADERS_MPI ggml-mpi.h)
+        set(GGML_SOURCES_MPI ggml-mpi.c)
+
+        add_compile_definitions(GGML_USE_MPI)
+        add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
+
+        if (NOT MSVC)
+            add_compile_options(-Wno-cast-qual)
+        endif()
+
+        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
+
+        # Even if you're only using the C header, C++ programs may bring in MPI
+        # C++ functions, so more linkage is needed
+        if (MPI_CXX_FOUND)
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}     ${MPI_CXX_LIBRARIES})
+        endif()
+    else()
+        message(WARNING "MPI not found")
     endif()
+endif()
 
-    set(GGML_HEADERS_RPC ggml-rpc.h)
-    set(GGML_SOURCES_RPC ggml-rpc.cpp)
+if (LLAMA_CLBLAST)
+    find_package(CLBlast)
+    if (CLBlast_FOUND)
+        message(STATUS "CLBlast found")
+
+        set(GGML_HEADERS_OPENCL ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp)
+
+        add_compile_definitions(GGML_USE_CLBLAST)
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} clblast)
+    else()
+        message(WARNING "CLBlast not found")
+    endif()
 endif()
 
 if (LLAMA_VULKAN)
@@ -521,12 +519,6 @@ if (LLAMA_VULKAN)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
-        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
-        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
-        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
-        endif()
-
         if (LLAMA_VULKAN_CHECK_RESULTS)
             add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()
@@ -535,10 +527,6 @@ if (LLAMA_VULKAN)
             add_compile_definitions(GGML_VULKAN_DEBUG)
         endif()
 
-        if (LLAMA_VULKAN_MEMORY_DEBUG)
-            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
-        endif()
-
         if (LLAMA_VULKAN_VALIDATE)
             add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
@@ -554,42 +542,16 @@ if (LLAMA_VULKAN)
 endif()
 
 if (LLAMA_HIPBLAS)
-    if (NOT EXISTS $ENV{ROCM_PATH})
-        if (NOT EXISTS /opt/rocm)
-            set(ROCM_PATH /usr)
-        else()
-            set(ROCM_PATH /opt/rocm)
-        endif()
-    else()
-        set(ROCM_PATH $ENV{ROCM_PATH})
-    endif()
-    list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
-    list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 
-    # CMake on Windows doesn't support the HIP language yet
-    if(WIN32)
-        set(CXX_IS_HIPCC TRUE)
-    else()
-        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
 
-    if(CXX_IS_HIPCC)
-        if(LINUX)
-            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-            endif()
-
-            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
-                    " Prefer setting the HIP compiler directly. See README for details.")
-        endif()
-    else()
-        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
-        if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
-        endif()
-        cmake_minimum_required(VERSION 3.21)
-        enable_language(HIP)
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
+
     find_package(hip     REQUIRED)
     find_package(hipblas REQUIRED)
     find_package(rocblas REQUIRED)
@@ -600,10 +562,6 @@ if (LLAMA_HIPBLAS)
 
     file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
     list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-    file(GLOB SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    file(GLOB SRCS "ggml-cuda/template-instances/mmq*.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
 
     add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
 
@@ -623,35 +581,17 @@ if (LLAMA_HIPBLAS)
         add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
     endif()
 
-    if (LLAMA_CUDA_FA_ALL_QUANTS)
-        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-    else()
-        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    endif()
-
     add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
     add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
     add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
-    if (CXX_IS_HIPCC)
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device)
-    else()
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
-    endif()
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
 
     if (LLAMA_STATIC)
         message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
     endif()
 
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -674,10 +614,6 @@ if (LLAMA_SYCL)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
 
-    if (LLAMA_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
-    endif()
-
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
 
@@ -689,8 +625,7 @@ if (LLAMA_SYCL)
     endif()
 
     set(GGML_HEADERS_SYCL ggml-sycl.h)
-    file(GLOB GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
-    list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
+    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
@@ -794,7 +729,6 @@ if (LLAMA_KOMPUTE)
             kompute-shaders/op_mul_mat_q4_0.comp
             kompute-shaders/op_mul_mat_q4_1.comp
             kompute-shaders/op_mul_mat_q6_k.comp
-            kompute-shaders/op_getrows_f32.comp
             kompute-shaders/op_getrows_f16.comp
             kompute-shaders/op_getrows_q4_0.comp
             kompute-shaders/op_getrows_q4_1.comp
@@ -827,7 +761,6 @@ if (LLAMA_KOMPUTE)
             shaderop_mul_mat_q4_0.h
             shaderop_mul_mat_q4_1.h
             shaderop_mul_mat_q6_k.h
-            shaderop_getrows_f32.h
             shaderop_getrows_f16.h
             shaderop_getrows_q4_0.h
             shaderop_getrows_q4_1.h
@@ -1061,11 +994,6 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         if (GGML_COMPILER_SUPPORT_DOTPROD)
             add_compile_definitions(__ARM_FEATURE_DOTPROD)
         endif ()
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
-        endif ()
-
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
         if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
             add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
@@ -1094,9 +1022,6 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
-        if (LLAMA_SVE)
-            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
-        endif()
     endif()
 elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
         (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
@@ -1121,10 +1046,6 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
-            if (LLAMA_AVX512_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
-            endif()
         elseif (LLAMA_AVX2)
             list(APPEND ARCH_FLAGS /arch:AVX2)
         elseif (LLAMA_AVX)
@@ -1156,9 +1077,6 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         if (LLAMA_AVX512_VNNI)
             list(APPEND ARCH_FLAGS -mavx512vnni)
         endif()
-        if (LLAMA_AVX512_BF16)
-            list(APPEND ARCH_FLAGS -mavx512bf16)
-        endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
@@ -1168,17 +1086,6 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
         list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
         #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-    message(STATUS "loongarch64 detected")
-
-    list(APPEND ARCH_FLAGS -march=loongarch64)
-    if (LLAMA_LASX)
-        list(APPEND ARCH_FLAGS -mlasx)
-    endif()
-    if (LLAMA_LSX)
-        list(APPEND ARCH_FLAGS -mlsx)
-    endif()
-
 else()
     message(STATUS "Unknown architecture")
 endif()
@@ -1265,14 +1172,14 @@ add_library(ggml OBJECT
             ggml-quants.c
             ggml-quants.h
             ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
+            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
             ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
             ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
             ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_BLAS}      ${GGML_HEADERS_BLAS}
             ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
             )
 
@@ -1353,9 +1260,8 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}"
-        "${GGML_HEADERS_METAL}"
-        "${GGML_HEADERS_EXTRA}")
+        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml PUBLIC_HEADER)
@@ -1364,7 +1270,18 @@ set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 install(
-    FILES convert-hf-to-gguf.py
+    FILES convert.py
+    PERMISSIONS
+        OWNER_READ
+        OWNER_WRITE
+        OWNER_EXECUTE
+        GROUP_READ
+        GROUP_EXECUTE
+        WORLD_READ
+        WORLD_EXECUTE
+    DESTINATION ${CMAKE_INSTALL_BINDIR})
+install(
+    FILES convert-lora-to-ggml.py
     PERMISSIONS
         OWNER_READ
         OWNER_WRITE
@@ -1391,13 +1308,6 @@ if (LLAMA_METAL)
     endif()
 endif()
 
-configure_file(cmake/llama.pc.in
-        "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
-        @ONLY)
-
-install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
-        DESTINATION lib/pkgconfig)
-
 #
 # programs, examples and tests
 #

CMakePresets.json

@@ -1,49 +0,0 @@
-{
-  "version": 4,
-  "configurePresets": [
-    {
-        "name":  "base",
-        "hidden": true,
-        "generator":   "Ninja",
-        "binaryDir":   "${sourceDir}/build-${presetName}",
-        "cacheVariables": {
-            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
-            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
-        }
-    },
-
-    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
-
-    {
-        "name": "arm64-windows-msvc", "hidden": true,
-        "architecture": { "value": "arm64",       "strategy": "external" },
-        "toolset":      { "value": "host=x86_64", "strategy": "external" },
-        "cacheVariables": {
-            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
-        }
-    },
-
-    {
-        "name": "arm64-windows-llvm", "hidden": true,
-        "architecture": { "value": "arm64",       "strategy": "external" },
-        "toolset":      { "value": "host=x86_64", "strategy": "external" },
-        "cacheVariables": {
-            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
-        }
-    },
-
-    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
-    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "release" ] },
-    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "release", "static" ] },
-
-    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
-    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] },
-
-    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
-    { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
-    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
-  ]
-}

CONTRIBUTING.md

@@ -1,14 +0,0 @@
-# Contributing Guidelines
-
-## Checklist
-
-* Make sure your PR follows the [coding guidelines](https://github.com/ggerganov/llama.cpp/blob/master/README.md#coding-guidelines)
-* Test your changes using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
-* Execute [the full CI locally on your machine](ci/README.md) before publishing
-
-## PR formatting
-
-* Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
-    - The PR template has a series of review complexity checkboxes `[ ]` that you can mark as `[X]` for your conveience. Refer to [About task lists](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) for more information.
-* If the pull request only contains documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times.
-* When squashing multiple commits on merge, use the following format for your commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : Fix typo in utils.py (#1234)`

Makefile

@@ -1,45 +1,8 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	libllava.a \
-	llama-baby-llama \
-	llama-batched \
-	llama-batched-bench \
-	llama-bench \
-	llama-benchmark-matmult \
-	llama-cli \
-	llama-convert-llama2c-to-ggml \
-	llama-embedding \
-	llama-eval-callback \
-	llama-export-lora \
-	llama-finetune \
-	llama-gbnf-validator \
-	llama-gguf \
-	llama-gguf-split \
-	llama-gritlm \
-	llama-imatrix \
-	llama-infill \
-	llama-llava-cli \
-	llama-lookahead \
-	llama-lookup \
-	llama-lookup-create \
-	llama-lookup-merge \
-	llama-lookup-stats \
-	llama-parallel \
-	llama-passkey \
-	llama-perplexity \
-	llama-q8dot \
-	llama-quantize \
-	llama-quantize-stats \
-	llama-retrieval \
-	llama-save-load-state \
-	llama-server \
-	llama-simple \
-	llama-speculative \
-	llama-tokenize \
-	llama-train-text-from-scratch \
-	llama-vdot \
-	llama-cvector-generator \
-	tests/test-c.o
+	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search  \
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -94,8 +57,6 @@ ifeq ($(UNAME_S),Darwin)
 		LLAMA_METAL := 1
 	endif
 
-	LLAMA_NO_OPENMP := 1
-
 	ifneq ($(UNAME_P),arm)
 		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
 		ifeq ($(SYSCTL_M),1)
@@ -106,10 +67,6 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
-ifdef LLAMA_RPC
-	BUILD_TARGETS += rpc-server
-endif
-
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -178,16 +135,12 @@ MK_NVCCFLAGS = -std=c++11
 ifdef LLAMA_FAST
 MK_CFLAGS     += -Ofast
 HOST_CXXFLAGS += -Ofast
-ifndef LLAMA_DEBUG
 MK_NVCCFLAGS  += -O3
-endif # LLAMA_DEBUG
 else
 MK_CFLAGS     += -O3
 MK_CXXFLAGS   += -O3
-ifndef LLAMA_DEBUG
 MK_NVCCFLAGS  += -O3
-endif # LLAMA_DEBUG
-endif # LLAMA_FAST
+endif
 
 ifndef LLAMA_NO_CCACHE
 CCACHE := $(shell which ccache)
@@ -248,10 +201,9 @@ ifdef LLAMA_SCHED_MAX_COPIES
 endif
 
 ifdef LLAMA_DEBUG
-	MK_CFLAGS    += -O0 -g
-	MK_CXXFLAGS  += -O0 -g
-	MK_LDFLAGS   += -g
-	MK_NVCCFLAGS += -O0 -g
+	MK_CFLAGS   += -O0 -g
+	MK_CXXFLAGS += -O0 -g
+	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
 		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
@@ -427,90 +379,64 @@ ifneq ($(filter ppc64le%,$(UNAME_M)),)
 	CUDA_POWER_ARCH = 1
 endif
 
-ifneq ($(filter loongarch64%,$(UNAME_M)),)
-	MK_CFLAGS   += -mlasx
-	MK_CXXFLAGS += -mlasx
-endif
-
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
+ifdef LLAMA_QKK_64
+	MK_CPPFLAGS += -DGGML_QKK_64
+endif
+
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
 		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
 		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
 		MK_LDFLAGS  += -framework Accelerate
-		OBJS        += ggml-blas.o
 	endif
 endif # LLAMA_NO_ACCELERATE
 
-ifndef LLAMA_NO_OPENMP
-	MK_CPPFLAGS += -DGGML_USE_OPENMP
-	MK_CFLAGS   += -fopenmp
-	MK_CXXFLAGS += -fopenmp
-endif # LLAMA_NO_OPENMP
+ifdef LLAMA_MPI
+	MK_CPPFLAGS += -DGGML_USE_MPI
+	MK_CFLAGS   += -Wno-cast-qual
+	MK_CXXFLAGS += -Wno-cast-qual
+	OBJS        += ggml-mpi.o
+endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
-	OBJS        += ggml-blas.o
 endif # LLAMA_OPENBLAS
 
-ifdef LLAMA_OPENBLAS64
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
-	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
-	OBJS        += ggml-blas.o
-endif # LLAMA_OPENBLAS64
-
-ifdef LLAMA_BLIS
-	MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS  += -lblis -L/usr/local/lib
-	OBJS        += ggml-blas.o
-endif # LLAMA_BLIS
-
 ifndef LLAMA_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
 	OBJS        += sgemm.o
 endif
 
-ifdef LLAMA_RPC
-	MK_CPPFLAGS   += -DGGML_USE_RPC
-	OBJS          += ggml-rpc.o
-endif # LLAMA_RPC
+ifdef LLAMA_BLIS
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
+endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
 # LLAMA_CUBLAS is deprecated and will be removed in the future
 	LLAMA_CUDA := 1
 endif
 
-OBJS_CUDA_TEMP_INST      = $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-wmma*.cu))
-OBJS_CUDA_TEMP_INST     += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/mmq*.cu))
-ifdef LLAMA_CUDA_FA_ALL_QUANTS
-	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*.cu))
-else
-	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
-	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
-	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
-endif # LLAMA_CUDA_FA_ALL_QUANTS
-
 ifdef LLAMA_CUDA
 	ifneq ('', '$(wildcard /opt/cuda)')
 		CUDA_PATH ?= /opt/cuda
 	else
 		CUDA_PATH ?= /usr/local/cuda
 	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
-	OBJS         += $(OBJS_CUDA_TEMP_INST)
 	MK_NVCCFLAGS += -use_fast_math
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -521,9 +447,6 @@ endif # JETSON_EOL_MODULE_DETECT
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG
-ifdef LLAMA_CUDA_DEBUG
-	MK_NVCCFLAGS += --device-debug
-endif # LLAMA_CUDA_DEBUG
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
 else
@@ -573,10 +496,7 @@ ifdef LLAMA_CUDA_NO_PEER_COPY
 endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
-endif # LLAMA_CUDA_CCBIN
-ifdef LLAMA_CUDA_FA_ALL_QUANTS
-	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
-endif # LLAMA_CUDA_FA_ALL_QUANTS
+endif
 
 ifdef JETSON_EOL_MODULE_DETECT
 define NVCC_COMPILE
@@ -588,13 +508,30 @@ define NVCC_COMPILE
 endef # NVCC_COMPILE
 endif # JETSON_EOL_MODULE_DETECT
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml.h ggml-common.h ggml-cuda/common.cuh
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
 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)
 	$(NVCC_COMPILE)
 endif # LLAMA_CUDA
 
+ifdef LLAMA_CLBLAST
+	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
+	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
+
+	# Mac provides OpenCL as a framework
+	ifeq ($(UNAME_S),Darwin)
+		MK_LDFLAGS += -lclblast -framework OpenCL
+	else
+		MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
+	endif
+	OBJS    += ggml-opencl.o
+
+ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_CLBLAST
+
 ifdef LLAMA_VULKAN
 	MK_CPPFLAGS  += -DGGML_USE_VULKAN
 	MK_LDFLAGS += -lvulkan
@@ -608,10 +545,6 @@ ifdef LLAMA_VULKAN_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
 endif
 
-ifdef LLAMA_VULKAN_MEMORY_DEBUG
-	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
-endif
-
 ifdef LLAMA_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -627,10 +560,10 @@ endif # LLAMA_VULKAN
 ifdef LLAMA_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
-		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
 	else
 		ROCM_PATH	?= /opt/rocm
-		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
@@ -641,9 +574,8 @@ ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
-	MK_LDFLAGS  += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
-	HIPFLAGS    += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
@@ -655,12 +587,11 @@ ifdef LLAMA_CUDA_NO_PEER_COPY
 endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
 	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
-	OBJS        += $(OBJS_CUDA_TEMP_INST)
 
 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)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml.h ggml-common.h ggml-cuda/common.cuh
+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 $@ $<
 
 endif # LLAMA_HIPBLAS
@@ -698,26 +629,16 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+ifdef LLAMA_MPI
+ggml-mpi.o: ggml-mpi.c ggml-mpi.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_MPI
 
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif
 
-ifdef LLAMA_RPC
-ggml-rpc.o: ggml-rpc.cpp ggml-rpc.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-rpc-server.o: examples/rpc/rpc-server.cpp ggml-rpc.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-rpc-server: rpc-server.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-endif # LLAMA_RPC
-
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -791,18 +712,20 @@ ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-ggml-blas.o: ggml-blas.cpp ggml-blas.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 unicode.o: unicode.cpp unicode.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 unicode-data.o: unicode-data.cpp unicode-data.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
+
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
@@ -831,9 +754,8 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 	rm -vrf ggml-cuda/*.o
-	rm -vrf ggml-cuda/template-instances/*.o
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -847,62 +769,62 @@ clean:
 # Helper function that replaces .c, .cpp, and .cu file endings with .o:
 GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
 
-llama-cli: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo
-	@echo '====  Run ./llama-cli -h for help.  ===='
+	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-llama-infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-quantize: examples/quantize/quantize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+quantize: examples/quantize/quantize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/colorthemes.css.hpp examples/server/style.css.hpp examples/server/theme-beeninorder.css.hpp examples/server/theme-ketivah.css.hpp examples/server/theme-mangotango.css.hpp examples/server/theme-playground.css.hpp examples/server/theme-polarnight.css.hpp examples/server/theme-snowstorm.css.hpp examples/server/index.html.hpp examples/server/index-new.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/system-prompts.js.hpp examples/server/prompt-formats.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
@@ -915,27 +837,23 @@ examples/server/%.hpp: examples/server/public/% Makefile
 		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
 	) > $@
 
-llama-gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
+convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -946,61 +864,59 @@ llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS)
 libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
-llama-llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
 	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
 
-llama-baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-lookup-create: examples/lookup/lookup-create.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
+
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-lookup-merge: examples/lookup/lookup-merge.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookup-stats: examples/lookup/lookup-stats.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -1026,20 +942,20 @@ build-info.o: common/build-info.cpp
 
 tests: $(TEST_TARGETS)
 
-llama-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-run-benchmark-matmult: llama-benchmark-matmult
+run-benchmark-matmult: benchmark-matmult
 	./$@
 
 .PHONY: run-benchmark-matmult swift
 
-llama-vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
+vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
+q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 

README-sycl.md

@@ -29,7 +29,7 @@ The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based o
 
 When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend.
 
-It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
+It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, CLBlast etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
 
 ## News
 
@@ -54,10 +54,10 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 
 ## OS
 
-| OS      | Status  | Verified                                       |
-|---------|---------|------------------------------------------------|
-| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39, Arch Linux |
-| Windows | Support | Windows 11                                     |
+| OS      | Status  | Verified                           |
+|---------|---------|------------------------------------|
+| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39 |
+| Windows | Support | Windows 11                         |
 
 
 ## Hardware
@@ -70,14 +70,14 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
+| Intel Arc Series              | Support | Arc 770, 730M                         |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
 | Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
 - **Memory**
-  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
+  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/main`.
 
   - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
 
@@ -99,14 +99,14 @@ The docker build option is currently limited to *intel GPU* targets.
 ### Build image
 ```sh
 # Using FP16
-docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/llama-cli-intel.Dockerfile .
+docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
 ```
 
 *Notes*:
 
 To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="LLAMA_SYCL_F16=ON"` argument from the previous command.
 
-You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
+You can also use the `.devops/server-intel.Dockerfile`, which builds the *"server"* alternative.
 
 ### Run container
 
@@ -275,7 +275,7 @@ source /opt/intel/oneapi/setvars.sh
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
-./build/bin/llama-ls-sycl-device
+./build/bin/ls-sycl-device
 ```
 A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
 ```
@@ -313,7 +313,7 @@ Examples:
 - Use device 0:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
 or run by script:
 
@@ -324,7 +324,7 @@ or run by script:
 - Use multiple devices:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
 Otherwise, you can run the script:
@@ -427,7 +427,7 @@ Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former in
 
 *Notes:*
 
-- By default, calling `make` will build all target binary files. In case of a minimal experimental setup, the user can build the inference executable only through `make llama-cli`.
+- By default, calling `make` will build all target binary files. In case of a minimal experimental setup, the user can build the inference executable only through `make main`.
 
 ### III. Run the inference
 
@@ -488,13 +488,13 @@ Examples:
 - Use device 0:
 
 ```
-build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
 ```
 
 - Use multiple devices:
 
 ```
-build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
 Otherwise, run the following wrapper script:
 

README.md

@@ -3,16 +3,11 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
-[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-> [!IMPORTANT]
-[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
-
 ### Recent API changes
 
 - [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
@@ -25,8 +20,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
-- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
+- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021**
 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
@@ -56,6 +50,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
         <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
+        <li><a href="#instruct-mode">Instruct mode</a></li>
         <li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
         <li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
         <li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
@@ -79,7 +74,7 @@ variety of hardware - locally and in the cloud.
 - AVX, AVX2 and AVX512 support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
-- Vulkan and SYCL backend support
+- Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
 Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
@@ -112,6 +107,7 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
+- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
 - [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
@@ -132,7 +128,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
-- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
@@ -146,14 +141,11 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
-- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
 **HTTP server**
 
 [llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
-[simplechat](./examples/server/public_simplechat) is a simple chat client, which can be used to chat with the model exposed using above web server (use --path to point to simplechat), from a local web browser.
-
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
@@ -184,7 +176,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
-- [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -195,7 +186,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
-- [RAGNA Desktop](https://ragna.app/) (proprietary)
 - [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
@@ -208,20 +198,15 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
-- [AIKit](https://github.com/sozercan/aikit) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
-**Tools:**
-
-- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
-
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
 ```
-$ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
+$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
 I UNAME_P:  arm
@@ -315,7 +300,7 @@ cd llama.cpp
 
 ### Build
 
-In order to build llama.cpp you have four different options.
+In order to build llama.cpp you have three different options.
 
 - Using `make`:
   - On Linux or MacOS:
@@ -324,6 +309,8 @@ In order to build llama.cpp you have four different options.
       make
       ```
 
+      **Note**: for `Debug` builds, run `make LLAMA_DEBUG=1`
+
   - On Windows:
 
     1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
@@ -335,38 +322,40 @@ In order to build llama.cpp you have four different options.
         make
         ```
 
-  - Notes:
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, run `make LLAMA_DEBUG=1`
-
 - Using `CMake`:
 
-  ```bash
-  cmake -B build
-  cmake --build build --config Release
-  ```
+    ```bash
+    cmake -B build
+    cmake --build build --config Release
+    ```
 
-  **Notes**:
+    **Note**: for `Debug` builds, there are two cases:
 
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, there are two cases:
-
-      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+    - Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
 
       ```bash
       cmake -B build -DCMAKE_BUILD_TYPE=Debug
       cmake --build build
       ```
 
-      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
+    - Multi-config generators (`-G` param set to Visual Studio, XCode...):
 
       ```bash
       cmake -B build -G "Xcode"
       cmake --build build --config Debug
       ```
 
+- Using `Zig` (version 0.11 or later):
+
+    Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C,
+    it's also possible to cross compile for other operating systems and architectures:
+
+    ```bash
+    zig build -Doptimize=ReleaseFast -Dtarget=x86_64-windows-gnu -Dcpu=x86_64+avx2+fma+f16c
+    ```
+
+    The `zig targets` command will give you valid options to use.
+
 -   Using `gmake` (FreeBSD):
 
     1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
@@ -374,18 +363,15 @@ In order to build llama.cpp you have four different options.
     3. Install compilation dependencies.
 
         ```bash
-        sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
+        sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
+            opencl clblast openblas
 
         gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
         ```
 
-### Homebrew
-
-On Mac and Linux, the homebrew package manager can be used via
-```
-brew install llama.cpp
-```
-The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+    **Notes:** With this packages you can build llama.cpp with OPENBLAS and
+    CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read
+    the instructions for use and activate this options in this document below.
 
 ### Metal Build
 
@@ -395,9 +381,48 @@ To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or th
 When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
 argument.
 
+### MPI Build
+
+MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
+
+First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
+
+Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
+
+- Using `make`:
+
+  ```bash
+  make CC=mpicc CXX=mpicxx LLAMA_MPI=1
+  ```
+
+- Using `CMake`:
+
+  ```bash
+  cmake -S . -B build -DLLAMA_MPI=ON
+  ```
+
+Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
+
+Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
+
+Here is an example hostfile:
+
+```
+192.168.0.1:2
+malvolio.local:1
+```
+
+The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
+
+Finally, you're ready to run a computation using `mpirun`:
+
+```bash
+mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
+```
+
 ### BLAS Build
 
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
 
 - #### Accelerate Framework:
 
@@ -485,12 +510,10 @@ Building the program with BLAS support may lead to some performance improvements
   |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
   | LLAMA_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
-  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                               |
-  | LLAMA_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. |                                                                                                                                         |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
   | LLAMA_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
   | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
   | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-  | LLAMA_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
 - #### hipBLAS
 
@@ -504,28 +527,13 @@ Building the program with BLAS support may lead to some performance improvements
     ```
   - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
     ```bash
-    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
+        cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
         && cmake --build build --config Release -- -j 16
     ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
+    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
     However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
-    Note that if you get the following error:
-    ```
-    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
-    ```
-    Try searching for a directory under `HIP_PATH` that contains the file
-    `oclc_abi_version_400.bc`. Then, add the following to the start of the
-    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
-    like:
-    ```bash
-    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
-    HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-        && cmake --build build -- -j 16
-    ```
-
   - Using `make` (example for target gfx1030, build with 16 CPU threads):
     ```bash
     make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
@@ -534,8 +542,10 @@ Building the program with BLAS support may lead to some performance improvements
   - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
     ```bash
     set PATH=%HIP_PATH%\bin;%PATH%
-    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
-    cmake --build build
+    mkdir build
+    cd build
+    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ..
+    cmake --build .
     ```
     Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
     Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
@@ -551,6 +561,111 @@ Building the program with BLAS support may lead to some performance improvements
   | LLAMA_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
 
+- #### CLBlast
+
+  OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
+
+  You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
+    - For Ubuntu, Debian, and Fedora the packages `opencl-headers`, `ocl-icd` may be needed.
+
+    - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
+
+    - <details>
+        <summary>Installing the OpenCL SDK from source</summary>
+
+        ```sh
+        git clone --recurse-submodules https://github.com/KhronosGroup/OpenCL-SDK.git
+        cd OpenCL-SDK
+        cmake -B build -DBUILD_DOCS=OFF \
+          -DBUILD_EXAMPLES=OFF \
+          -DBUILD_TESTING=OFF \
+          -DOPENCL_SDK_BUILD_SAMPLES=OFF \
+          -DOPENCL_SDK_TEST_SAMPLES=OFF
+        cmake --build build
+        cmake --install build --prefix /some/path
+        ```
+      </details>
+
+  ##### Installing CLBlast
+
+  Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
+
+  Linux packaging:
+  Fedora Linux:
+  ```bash
+  sudo dnf install clblast
+  ```
+
+  Alternatively, they may be built from source.
+
+  - <details>
+    <summary>Windows:</summary>
+
+      ```cmd
+      set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64"
+      git clone https://github.com/CNugteren/CLBlast.git
+      cd CLBlast
+      cmake -B build -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64
+      cmake --build build --config Release
+      cmake --install build --prefix C:/CLBlast
+      ```
+
+      (note: `--config Release` at build time is the default and only relevant for Visual Studio builds - or multi-config Ninja builds)
+
+  - <details>
+    <summary>Unix:</summary>
+
+      ```sh
+      git clone https://github.com/CNugteren/CLBlast.git
+      cd CLBlast
+      cmake -B build -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF
+      cmake --build build --config Release
+      cmake --install build --prefix /some/path
+      ```
+
+      Where `/some/path` is where the built library will be installed (default is `/usr/local`).
+    </details>
+
+  ##### Building Llama with CLBlast
+
+  - Build with make:
+    ```sh
+    make LLAMA_CLBLAST=1
+    ```
+  - CMake (Unix):
+    ```sh
+    cmake -B build -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
+    cmake --build build --config Release
+    ```
+  - CMake (Windows):
+    ```cmd
+    set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast"
+    git clone https://github.com/ggerganov/llama.cpp
+    cd llama.cpp
+    cmake -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64
+    cmake --build build --config Release
+    cmake --install build --prefix C:/LlamaCPP
+    ```
+
+  ##### Running Llama with CLBlast
+
+  The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
+
+  To select the correct platform (driver) and device (GPU), you can use the environment variables `GGML_OPENCL_PLATFORM` and `GGML_OPENCL_DEVICE`.
+  The selection can be a number (starting from 0) or a text string to search:
+
+  ```sh
+  GGML_OPENCL_PLATFORM=1 ./main ...
+  GGML_OPENCL_DEVICE=2 ./main ...
+  GGML_OPENCL_PLATFORM=Intel ./main ...
+  GGML_OPENCL_PLATFORM=AMD GGML_OPENCL_DEVICE=1 ./main ...
+  ```
+
+  The default behavior is to find the first GPU device, but when it is an integrated GPU on a laptop, for instance, the selectors are useful.
+  Using the variables it is possible to select a CPU-based driver as well, if so desired.
+
+  You can get a list of platforms and devices from the `clinfo -l` command, etc.
+
 - #### Vulkan
 
   **With docker**:
@@ -559,7 +674,7 @@ Building the program with BLAS support may lead to some performance improvements
 
   ```sh
   # Build the image
-  docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
+  docker build -t llama-cpp-vulkan -f .devops/main-vulkan.Dockerfile .
 
   # Then, use it:
   docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
@@ -580,9 +695,7 @@ Building the program with BLAS support may lead to some performance improvements
   vulkaninfo
   ```
 
-  Alternatively your package manager might be able to provide the appropriate libraries.
-  For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-  For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
+  Alternatively your package manager might be able to provide the appropiate libraries. For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
 
   Then, build llama.cpp using the cmake command below:
 
@@ -590,7 +703,7 @@ Building the program with BLAS support may lead to some performance improvements
   cmake -B build -DLLAMA_VULKAN=1
   cmake --build build --config Release
   # Test the output binary (with "-ngl 33" to offload all layers to GPU)
-  ./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+  ./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
 
   # You should see in the output, ggml_vulkan detected your GPU. For example:
   # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
@@ -598,13 +711,9 @@ Building the program with BLAS support may lead to some performance improvements
 
 ### Prepare and Quantize
 
-> [!NOTE]
-> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
-
 To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
-Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
-It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
+Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
 
 ```bash
 # obtain the official LLaMA model weights and place them in ./models
@@ -621,20 +730,23 @@ ls ./models
 python3 -m pip install -r requirements.txt
 
 # convert the model to ggml FP16 format
-python3 convert-hf-to-gguf.py models/mymodel/
+python3 convert.py models/mymodel/
+
+# [Optional] for models using BPE tokenizers
+python convert.py models/mymodel/ --vocab-type bpe
 
 # quantize the model to 4-bits (using Q4_K_M method)
-./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
+./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
 
 # update the gguf filetype to current version if older version is now unsupported
-./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+./quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
 ```
 
 ### Run the quantized model
 
 ```bash
 # start inference on a gguf model
-./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
+./main -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
 ```
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
@@ -709,7 +821,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 #### How to run
 
 1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-2. Run `./llama-perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
+2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```
 perplexity : calculating perplexity over 655 chunks
@@ -733,16 +845,16 @@ Here is an example of a few-shot interaction, invoked with the command
 ./examples/chat-13B.sh
 
 # custom arguments using a 13B model
-./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
+./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
 ```
 
-Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `llama-cli` example program.
+Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `main` example program.
 
 ![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
 
 ### Persistent Interaction
 
-The prompt, user inputs, and model generations can be saved and resumed across calls to `./llama-cli` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
+The prompt, user inputs, and model generations can be saved and resumed across calls to `./main` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
 
 ```bash
 # Start a new chat
@@ -764,13 +876,41 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
 `llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
 
 ```bash
-./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
 ```
 
 The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
 
 For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
 
+### Instruct mode
+
+1. First, download and place the `ggml` model into the `./models` folder
+2. Run the `main` tool like this:
+
+```
+./examples/alpaca.sh
+```
+
+Sample run:
+
+```
+== Running in interactive mode. ==
+ - Press Ctrl+C to interject at any time.
+ - Press Return to return control to LLaMA.
+ - If you want to submit another line, end your input in '\'.
+
+ Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+> How many letters are there in the English alphabet?
+There 26 letters in the English Alphabet
+> What is the most common way of transportation in Amsterdam?
+The majority (54%) are using public transit. This includes buses, trams and metros with over 100 lines throughout the city which make it very accessible for tourists to navigate around town as well as locals who commute by tram or metro on a daily basis
+> List 5 words that start with "ca".
+cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
+>
+```
+
 ### Obtaining and using the Facebook LLaMA 2 model
 
 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
@@ -843,7 +983,7 @@ $mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/ho
 Now, you can start chatting:
 ```
 $cd /data/data/com.termux/files/home/bin
-$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
 ```
 
 Here's a demo of an interactive session running on Pixel 5 phone:
@@ -910,8 +1050,8 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
-docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -961,7 +1101,7 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m
 
 ### Docs
 
-- [main (cli)](./examples/main/README.md)
+- [main](./examples/main/README.md)
 - [server](./examples/server/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [BLIS](./docs/BLIS.md)

build.zig

@@ -0,0 +1,172 @@
+// Compatible with Zig Version 0.11.0
+const std = @import("std");
+const ArrayList = std.ArrayList;
+const Compile = std.Build.Step.Compile;
+const ConfigHeader = std.Build.Step.ConfigHeader;
+const Mode = std.builtin.Mode;
+const CrossTarget = std.zig.CrossTarget;
+
+const Maker = struct {
+    builder: *std.build.Builder,
+    target: CrossTarget,
+    optimize: Mode,
+    enable_lto: bool,
+
+    include_dirs: ArrayList([]const u8),
+    cflags: ArrayList([]const u8),
+    cxxflags: ArrayList([]const u8),
+    objs: ArrayList(*Compile),
+
+    fn addInclude(m: *Maker, dir: []const u8) !void {
+        try m.include_dirs.append(dir);
+    }
+    fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
+        try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
+    }
+    fn addCFlag(m: *Maker, flag: []const u8) !void {
+        try m.cflags.append(flag);
+    }
+    fn addCxxFlag(m: *Maker, flag: []const u8) !void {
+        try m.cxxflags.append(flag);
+    }
+    fn addFlag(m: *Maker, flag: []const u8) !void {
+        try m.addCFlag(flag);
+        try m.addCxxFlag(flag);
+    }
+
+    fn init(builder: *std.build.Builder) !Maker {
+        const target = builder.standardTargetOptions(.{});
+        const zig_version = @import("builtin").zig_version_string;
+        const commit_hash = try std.ChildProcess.exec(
+            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
+        );
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
+            \\int LLAMA_BUILD_NUMBER = {};
+            \\char const *LLAMA_COMMIT = "{s}";
+            \\char const *LLAMA_COMPILER = "Zig {s}";
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
+            \\
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
+        var m = Maker{
+            .builder = builder,
+            .target = target,
+            .optimize = builder.standardOptimizeOption(.{}),
+            .enable_lto = false,
+            .include_dirs = ArrayList([]const u8).init(builder.allocator),
+            .cflags = ArrayList([]const u8).init(builder.allocator),
+            .cxxflags = ArrayList([]const u8).init(builder.allocator),
+            .objs = ArrayList(*Compile).init(builder.allocator),
+        };
+
+        try m.addCFlag("-std=c11");
+        try m.addCxxFlag("-std=c++11");
+        try m.addProjectInclude(&.{});
+        try m.addProjectInclude(&.{"common"});
+        return m;
+    }
+
+    fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
+        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
+        if (o.target.getAbi() != .msvc)
+            o.defineCMacro("_GNU_SOURCE", null);
+
+        if (std.mem.endsWith(u8, src, ".c")) {
+            o.addCSourceFiles(&.{src}, m.cflags.items);
+            o.linkLibC();
+        } else {
+            o.addCSourceFiles(&.{src}, m.cxxflags.items);
+            if (o.target.getAbi() == .msvc) {
+                o.linkLibC(); // need winsdk + crt
+            } else {
+                // linkLibCpp already add (libc++ + libunwind + libc)
+                o.linkLibCpp();
+            }
+        }
+        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
+        o.want_lto = m.enable_lto;
+        return o;
+    }
+
+    fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
+        const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
+        e.addCSourceFiles(&.{src}, m.cxxflags.items);
+        for (deps) |d| e.addObject(d);
+        for (m.objs.items) |o| e.addObject(o);
+        for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
+
+        // https://github.com/ziglang/zig/issues/15448
+        if (e.target.getAbi() == .msvc) {
+            e.linkLibC(); // need winsdk + crt
+        } else {
+            // linkLibCpp already add (libc++ + libunwind + libc)
+            e.linkLibCpp();
+        }
+        m.builder.installArtifact(e);
+        e.want_lto = m.enable_lto;
+        return e;
+    }
+};
+
+pub fn build(b: *std.build.Builder) !void {
+    var make = try Maker.init(b);
+    make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
+
+    const ggml = make.obj("ggml", "ggml.c");
+    const sgemm = make.obj("sgemm", "sgemm.cpp");
+    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
+    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
+    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
+    const unicode = make.obj("unicode", "unicode.cpp");
+    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
+    const llama = make.obj("llama", "llama.cpp");
+    const buildinfo = make.obj("common", "common/build-info.cpp");
+    const common = make.obj("common", "common/common.cpp");
+    const console = make.obj("console", "common/console.cpp");
+    const sampling = make.obj("sampling", "common/sampling.cpp");
+    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
+    const train = make.obj("train", "common/train.cpp");
+    const clip = make.obj("clip", "examples/llava/clip.cpp");
+    const llava = make.obj("llava", "examples/llava/llava.cpp");
+
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
+    if (server.target.isWindows()) {
+        server.linkSystemLibrary("ws2_32");
+    }
+
+    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
+    for (server_assets) |asset| {
+        const input_path = b.fmt("examples/server/public/{s}", .{asset});
+        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
+
+        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
+
+        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
+        defer b.allocator.free(input);
+
+        var buf = std.ArrayList(u8).init(b.allocator);
+        defer buf.deinit();
+
+        for (input) |byte| {
+            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
+        }
+
+        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
+        defer b.allocator.free(name);
+        std.mem.replaceScalar(u8, name, '.', '_');
+
+        try std.fs.cwd().writeFile(output_path, b.fmt(
+            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
+            .{ name, buf.items, name, input.len },
+        ));
+
+        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
+    }
+}

ci/run.sh

@@ -202,15 +202,12 @@ function gg_sum_test_scripts_release {
 }
 
 function gg_get_model {
-    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
-    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
-    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-    if [[ -s $gguf_0 ]]; then
-        echo -n "$gguf_0"
-    elif [[ -s $gguf_1 ]]; then
-        echo -n "$gguf_1"
-    elif [[ -s $gguf_2 ]]; then
-        echo -n "$gguf_2"
+    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
+    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_3b ]]; then
+        echo -n "$gguf_3b"
+    elif [[ -s $gguf_7b ]]; then
+        echo -n "$gguf_7b"
     else
         echo >&2 "No model found. Can't run gg_run_ctest_with_model."
         exit 1
@@ -259,35 +256,33 @@ function gg_sum_ctest_with_model_release {
     gg_printf '```\n'
 }
 
-# open_llama_7b_v2
-# requires: GG_BUILD_CUDA
+# open_llama_3b_v2
 
-function gg_run_open_llama_7b_v2 {
+function gg_run_open_llama_3b_v2 {
     cd ${SRC}
 
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
     gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
-    path_models="../models-mnt/open-llama/7B-v2"
+    path_models="../models-mnt/open-llama/3B-v2"
     path_wiki="../models-mnt/wikitext/wikitext-2-raw"
 
     rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert.py ${path_models}
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -301,49 +296,47 @@ function gg_run_open_llama_7b_v2 {
     model_q5_k="${path_models}/ggml-model-q5_k.gguf"
     model_q6_k="${path_models}/ggml-model-q6_k.gguf"
 
-    wiki_test="${path_wiki}/wiki.test.raw"
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -372,148 +365,58 @@ function gg_run_open_llama_7b_v2 {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
-    set +e
-}
-
-function gg_sum_open_llama_7b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 7B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
-# pythia_1.4b
-
-function gg_run_pythia_1_4b {
-    cd ${SRC}
-
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/pythia/1.4B"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/llama-cli --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/llama-save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
+    # lora
+    function compare_ppl {
         qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
 
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
             return 20
         fi
 
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
         return 0
     }
 
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    path_lora="../models-mnt/open-llama/3B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
 
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     set +e
 }
 
-function gg_sum_pythia_1_4b {
+function gg_sum_open_llama_3b_v2 {
     gg_printf '### %s\n\n' "${ci}"
 
-    gg_printf 'Pythia 1.4B:\n'
+    gg_printf 'OpenLLaMA 3B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -526,24 +429,32 @@ function gg_sum_pythia_1_4b {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
-# pythia_2_8b
+# open_llama_7b_v2
 # requires: GG_BUILD_CUDA
 
-function gg_run_pythia_2_8b {
+function gg_run_open_llama_7b_v2 {
     cd ${SRC}
 
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
     gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
-    path_models="../models-mnt/pythia/2.8B"
+    path_models="../models-mnt/open-llama/7B-v2"
     path_wiki="../models-mnt/wikitext/wikitext-2-raw"
 
     rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
@@ -553,7 +464,7 @@ function gg_run_pythia_2_8b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert.py ${path_models}
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -569,47 +480,47 @@ function gg_run_pythia_2_8b {
 
     wiki_test="${path_wiki}/wiki.test.raw"
 
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -630,7 +541,7 @@ function gg_run_pythia_2_8b {
     check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
@@ -638,16 +549,59 @@ function gg_run_pythia_2_8b {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/7B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # currently not supported by the CUDA backend
+    # q8_0
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
     set +e
 }
 
-function gg_sum_pythia_2_8b {
+function gg_sum_open_llama_7b_v2 {
     gg_printf '### %s\n\n' "${ci}"
 
-    gg_printf 'Pythia 2.8B:\n'
+    gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -660,6 +614,11 @@ function gg_sum_pythia_2_8b {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # bge-small
@@ -688,15 +647,15 @@ function gg_run_embd_bge_small {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert-hf-to-gguf.py ${path_models}
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
 
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
@@ -742,10 +701,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
-            test $ret -eq 0 && gg_run pythia_1_4b
+            test $ret -eq 0 && gg_run open_llama_3b_v2
         else
-            test $ret -eq 0 && gg_run pythia_2_8b
-            #test $ret -eq 0 && gg_run open_llama_7b_v2
+            test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
         test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release

cmake/arm64-windows-llvm.cmake

@@ -1,16 +0,0 @@
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR arm64 )
-
-set( target arm64-pc-windows-msvc )
-
-set( CMAKE_C_COMPILER    clang )
-set( CMAKE_CXX_COMPILER  clang++ )
-
-set( CMAKE_C_COMPILER_TARGET   ${target} )
-set( CMAKE_CXX_COMPILER_TARGET ${target} )
-
-set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
-set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
-
-set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
-set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

cmake/arm64-windows-msvc.cmake

@@ -1,6 +0,0 @@
-set( CMAKE_SYSTEM_NAME Windows )
-set( CMAKE_SYSTEM_PROCESSOR arm64 )
-
-set( target arm64-pc-windows-msvc )
-set( CMAKE_C_COMPILER_TARGET   ${target} )
-set( CMAKE_CXX_COMPILER_TARGET ${target} )

cmake/llama.pc.in

@@ -1,10 +0,0 @@
-prefix=@CMAKE_INSTALL_PREFIX@
-exec_prefix=${prefix}
-libdir=${exec_prefix}/lib
-includedir=${prefix}/include
-
-Name: llama
-Description: Port of Facebook's LLaMA model in C/C++
-Version: @PROJECT_VERSION@
-Libs: -L${libdir} -lllama
-Cflags: -I${includedir}

common/CMakeLists.txt

@@ -84,4 +84,4 @@ endif ()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)

common/common.h

@@ -27,7 +27,7 @@
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
 #define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
@@ -35,18 +35,14 @@
 
 // build info
 extern int LLAMA_BUILD_NUMBER;
-extern char const * LLAMA_COMMIT;
-extern char const * LLAMA_COMPILER;
-extern char const * LLAMA_BUILD_TARGET;
+extern char const *LLAMA_COMMIT;
+extern char const *LLAMA_COMPILER;
+extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
-//
-// CPU utils
-//
-
-int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
+int get_math_cpu_count();
+int32_t get_num_physical_cores();
 
 //
 // CLI argument parsing
@@ -55,35 +51,36 @@ int32_t cpu_get_num_math();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = cpu_get_num_math();
-    int32_t n_threads_draft       =    -1;
-    int32_t n_threads_batch       =    -1; // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft =    -1;
-    int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
-    int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                =     0; // number of tokens to keep from initial prompt
-    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
-    int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel            =     1; // number of parallel sequences to decode
-    int32_t n_sequences           =     1; // number of sequences to decode
-    float   p_split               =  0.1f; // speculative decoding split probability
-    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
-    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
-    int32_t n_beams               =     0; // if non-zero then use beam search of given width.
-    int32_t grp_attn_n            =     1; // group-attention factor
-    int32_t grp_attn_w            =   512; // group-attention width
-    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base        =  0.0f; // RoPE base frequency
-    float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
+    int32_t n_threads             = get_math_cpu_count();
+    int32_t n_threads_draft       = -1;
+    int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_threads_batch_draft = -1;
+    int32_t n_predict             = -1;    // new tokens to predict
+    int32_t n_ctx                 = 512;   // context size
+    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
+    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
+    int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel            = 1;     // number of parallel sequences to decode
+    int32_t n_sequences           = 1;     // number of sequences to decode
+    float   p_split               = 0.1f;  // speculative decoding split probability
+    int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
+    int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
+    int32_t grp_attn_n            = 1;     // group-attention factor
+    int32_t grp_attn_w            = 512;   // group-attention width
+    int32_t n_print               = -1;    // print token count every n tokens (-1 = disabled)
+    float   rope_freq_base        = 0.0f;  // RoPE base frequency
+    float   rope_freq_scale       = 0.0f;  // RoPE frequency scaling factor
     float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
+    float   yarn_attn_factor      = 1.0f;  // YaRN magnitude scaling factor
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
-    int32_t yarn_orig_ctx         =     0; // YaRN original context length
+    float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
+    int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
@@ -91,32 +88,29 @@ struct gpt_params {
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
-    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
 
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model                = ""; // model path
-    std::string model_draft          = ""; // draft model for speculative decoding
+    std::string model                = "";  // model path
+    std::string model_draft          = "";  // draft model for speculative decoding
     std::string model_alias          = "unknown"; // model alias
-    std::string model_url            = ""; // model url to download
-    std::string hf_repo              = ""; // HF repo
-    std::string hf_file              = ""; // HF file
+    std::string model_url            = "";  // model url to download
+    std::string hf_repo              = "";  // HF repo
+    std::string hf_file              = "";  // HF file
     std::string prompt               = "";
-    std::string prompt_file          = ""; // store the external prompt file name
-    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state
-    std::string input_prefix         = ""; // string to prefix user inputs with
-    std::string input_suffix         = ""; // string to suffix user inputs with
-    std::string logdir               = ""; // directory in which to save YAML log files
+    std::string prompt_file          = "";  // store the external prompt file name
+    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix         = "";  // string to prefix user inputs with
+    std::string input_suffix         = "";  // string to suffix user inputs with
+    std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
+    std::string logdir               = "";  // directory in which to save YAML log files
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
-    std::string logits_file          = ""; // file for saving *all* logits
-    std::string rpc_servers          = ""; // comma separated list of RPC servers
+    std::string logits_file          = "";  // file for saving *all* logits
 
-    std::vector<std::string> in_files;   // all input files
-    std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
     // TODO: avoid tuple, use struct
@@ -125,36 +119,34 @@ struct gpt_params {
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
 
-    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
-    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
-                                     //                                       (which is more convenient to use for plotting)
-                                     //
-    bool   hellaswag        = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
-    size_t hellaswag_tasks  = 400;   // number of tasks to use when computing the HellaSwag score
+    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
+    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
+                                    //                                       (which is more convenient to use for plotting)
+                                    //
+    bool   hellaswag       = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool   winogrande       = false; // compute Winogrande score over random tasks from datafile supplied in prompt
-    size_t winogrande_tasks = 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
+    bool   winogrande      = false; // compute Winogrande score over random tasks from datafile supplied in prompt
+    size_t winogrande_tasks= 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
 
-    bool   multiple_choice  = false;  // compute TruthfulQA score over random tasks from datafile supplied in prompt
-    size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
+    bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
+    size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
 
-    bool   kl_divergence    = false; // compute KL divergence
+    bool   kl_divergence   = false; // compute KL divergence
 
-    bool usage             = false; // print usage
+    bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
-    bool special           = false; // enable special token output
     bool interactive       = false; // interactive mode
-    bool interactive_first = false; // wait for user input immediately
-    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
+    bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
     bool embedding         = false; // get only sentence embedding
-    bool escape            = true;  // escape "\n", "\r", "\t", "\'", "\"", and "\\"
+    bool escape            = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
+    bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
@@ -162,6 +154,7 @@ struct gpt_params {
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
+    bool instruct          = false; // instruction mode (used for Alpaca models)
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
@@ -179,114 +172,37 @@ struct gpt_params {
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector
     std::vector<std::string> image; // path to image file(s)
-
-    // server params
-    int32_t port           = 8080;         // server listens on this network port
-    int32_t timeout_read   = 600;          // http read timeout in seconds
-    int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
-
-    std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";
-    std::string chat_template = "";
-    std::string system_prompt = "";
-
-    std::vector<std::string> api_keys;
-
-    std::string ssl_file_key  = "";
-    std::string ssl_file_cert = "";
-
-    bool endpoint_slots   = true;
-    bool endpoint_metrics = false;
-
-    bool log_json = false;
-
-    std::string slot_save_path;
-
-    float slot_prompt_similarity = 0.5f;
-
-    // batched-bench params
-    bool is_pp_shared = false;
-
-    std::vector<int32_t> n_pp;
-    std::vector<int32_t> n_tg;
-    std::vector<int32_t> n_pl;
-
-    // retrieval params
-    std::vector<std::string> context_files; // context files to embed
-
-    int32_t chunk_size = 64; // chunk size for context embedding
-
-    std::string chunk_separator = "\n"; // chunk separator for context embedding
-
-    // passkey params
-    int32_t n_junk = 250; // number of times to repeat the junk text
-    int32_t i_pos  = -1;  // position of the passkey in the junk text
-
-    // imatrix params
-    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
-
-    int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
-    int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
-    int32_t i_chunk     =  0; // start processing from this chunk
-
-    bool process_output = false; // collect data for the output tensor
-    bool compute_ppl    = true;  // whether to compute perplexity
-
-    // cvector-generator params
-    int n_completions = 64;
-    int n_pca_batch = 20;
-    int n_pca_iterations = 1000;
-    std::string cvector_outfile          = "control_vector.gguf";
-    std::string cvector_completions_file = "examples/cvector-generator/completions.txt";
-    std::string cvector_positive_file    = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file    = "examples/cvector-generator/negative.txt";
 };
 
 void gpt_params_handle_model_default(gpt_params & params);
 
-bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
-bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
-bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
-void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
+bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 
-std::string gpt_params_get_system_info(const gpt_params & params);
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+
+std::string get_system_info(const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+void process_escapes(std::string& input);
+
+bool validate_file_name(const std::string & filename);
 
 //
 // String utils
 //
 
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
-
 std::string string_strip(const std::string & str);
-std::string string_get_sortable_timestamp();
-
-template<class T>
-static std::vector<T> string_split(const std::string & str, char delim) {
-    std::vector<T> values;
-    std::istringstream str_stream(str);
-    std::string token;
-    while (std::getline(str_stream, token, delim)) {
-        T value;
-        std::istringstream token_stream(token);
-        token_stream >> value;
-        values.push_back(value);
-    }
-    return values;
-}
-
-bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
-void string_process_escapes(std::string & input);
-
-//
-// Filesystem utils
-//
-
-bool fs_validate_filename(const std::string & filename);
-bool fs_create_directory_with_parents(const std::string & path);
-
-std::string fs_get_cache_directory();
-std::string fs_get_cache_file(const std::string & filename);
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //
 // Model utils
@@ -358,21 +274,28 @@ std::string llama_detokenize_bpe(
 bool llama_should_add_bos_token(const llama_model * model);
 
 //
-// Chat template utils
+// YAML utils
 //
 
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool llama_chat_verify_template(const std::string & tmpl);
+bool create_directory_with_parents(const std::string & path);
+void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
+void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
+void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
+std::string get_sortable_timestamp();
+
+void dump_non_result_info_yaml(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
 
 //
 // KV cache utils
 //
 
 // Dump the KV cache view with the number of sequences per cell.
-void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
-void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
 
 //
 // Embedding utils
@@ -406,20 +329,6 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
 //
 // Split utils
 //
-
 static const char * const LLM_KV_SPLIT_NO            = "split.no";
 static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
-
-//
-// YAML utils
-//
-
-void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
-void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
-void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
-
-void yaml_dump_non_result_info(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-

common/grammar-parser.cpp

@@ -26,7 +26,7 @@ namespace grammar_parser {
 
     static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
-        auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
         return result.first->second;
     }
 
@@ -46,12 +46,8 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
-    static bool is_digit_char(char c) {
-        return '0' <= c && c <= '9';
-    }
-
     static bool is_word_char(char c) {
-        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
     }
 
     static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
@@ -103,17 +99,6 @@ namespace grammar_parser {
         return pos;
     }
 
-    static const char * parse_int(const char * src) {
-        const char * pos = src;
-        while (is_digit_char(*pos)) {
-            pos++;
-        }
-        if (pos == src) {
-            throw std::runtime_error(std::string("expecting integer at ") + src);
-        }
-        return pos;
-    }
-
     static std::pair<uint32_t, const char *> parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
@@ -152,68 +137,11 @@ namespace grammar_parser {
             bool                                 is_nested) {
         size_t last_sym_start = out_elements.size();
         const char * pos = src;
-
-        auto handle_repetitions = [&](int min_times, int max_times) {
-
-            if (last_sym_start == out_elements.size()) {
-                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
-            }
-
-            // apply transformation to previous symbol (last_sym_start to end) according to
-            // the following rewrite rules:
-            // S{m,n} --> S S S (m times) S'(n-m)
-            //            S'(x)   ::= S S'(x-1) |
-            //            (... n-m definitions of these S' rules ...)
-            //            S'(1)   ::= S |
-            // S{m,} -->  S S S (m times) S'
-            //            S'     ::= S S' |
-            // S*     --> S{0,}
-            //        --> S'     ::= S S' |
-            // S+     --> S{1,}
-            //        --> S S'
-            //            S'     ::= S S' |
-            // S?     --> S{0,1}
-            //        --> S'
-            //            S'     ::= S |
-
-            std::vector<llama_grammar_element> previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
-            if (min_times == 0) {
-                out_elements.resize(last_sym_start);
-            } else {
-                // Repeat the previous elements (min_times - 1) times
-                for (int i = 1; i < min_times; i++) {
-                    out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
-                }
-            }
-
-            uint32_t last_rec_rule_id = 0;
-            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
-
-            std::vector<llama_grammar_element> rec_rule(previous_elements);
-            for (int i = 0; i < n_opt; i++) {
-                rec_rule.resize(previous_elements.size());
-                uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
-                if (i > 0 || max_times < 0) {
-                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
-                }
-                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
-                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
-                add_rule(state, rec_rule_id, rec_rule);
-                last_rec_rule_id = rec_rule_id;
-            }
-            if (n_opt > 0) {
-                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
-            }
-        };
-
         while (*pos) {
             if (*pos == '"') { // literal string
                 pos++;
                 last_sym_start = out_elements.size();
                 while (*pos != '"') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
                     auto char_pair = parse_char(pos);
                          pos       = char_pair.second;
                     out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
@@ -228,9 +156,6 @@ namespace grammar_parser {
                 }
                 last_sym_start = out_elements.size();
                 while (*pos != ']') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
                     auto char_pair = parse_char(pos);
                          pos       = char_pair.second;
                     enum llama_gretype type = last_sym_start < out_elements.size()
@@ -239,9 +164,6 @@ namespace grammar_parser {
 
                     out_elements.push_back({type, char_pair.first});
                     if (pos[0] == '-' && pos[1] != ']') {
-                        if (!pos[1]) {
-                            throw std::runtime_error("unexpected end of input");
-                        }
                         auto endchar_pair = parse_char(pos + 1);
                              pos          = endchar_pair.second;
                         out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
@@ -266,51 +188,40 @@ namespace grammar_parser {
                     throw std::runtime_error(std::string("expecting ')' at ") + pos);
                 }
                 pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '.') { // any char
-                last_sym_start = out_elements.size();
-                out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, -1);
-            } else if (*pos == '+') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(1, -1);
-            } else if (*pos == '?') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, 1);
-            } else if (*pos == '{') {
-                pos = parse_space(pos + 1, is_nested);
-
-                if (!is_digit_char(*pos)) {
-                    throw std::runtime_error(std::string("expecting an int at ") + pos);
+            } 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);
                 }
-                const char * int_end = parse_int(pos);
-                int min_times = std::stoul(std::string(pos, int_end - pos));
-                pos = parse_space(int_end, is_nested);
 
-                int max_times = -1;
-
-                if (*pos == '}') {
-                    max_times = min_times;
-                    pos = parse_space(pos + 1, is_nested);
-                } else if (*pos == ',') {
-                    pos = parse_space(pos + 1, is_nested);
-
-                    if (is_digit_char(*pos)) {
-                        const char * int_end = parse_int(pos);
-                        max_times = std::stoul(std::string(pos, int_end - pos));
-                        pos = parse_space(int_end, is_nested);
-                    }
-
-                    if (*pos != '}') {
-                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
-                    }
-                    pos = parse_space(pos + 1, is_nested);
-                } else {
-                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<llama_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
                 }
-                handle_repetitions(min_times, max_times);
+                // mark start of alternate def
+                sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
             } else {
                 break;
             }
@@ -405,7 +316,6 @@ namespace grammar_parser {
             case LLAMA_GRETYPE_CHAR_NOT:       return true;
             case LLAMA_GRETYPE_CHAR_ALT:       return true;
             case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
-            case LLAMA_GRETYPE_CHAR_ANY:       return true;
             default:                           return false;
         }
     }
@@ -420,7 +330,6 @@ namespace grammar_parser {
                 case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
                 case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
                 case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
-                case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
             }
             switch (elem.type) {
                 case LLAMA_GRETYPE_END:
@@ -432,7 +341,6 @@ namespace grammar_parser {
                 case LLAMA_GRETYPE_CHAR_NOT:
                 case LLAMA_GRETYPE_CHAR_RNG_UPPER:
                 case LLAMA_GRETYPE_CHAR_ALT:
-                case LLAMA_GRETYPE_CHAR_ANY:
                     fprintf(file, "(\"");
                     print_grammar_char(file, elem.value);
                     fprintf(file, "\") ");
@@ -490,15 +398,11 @@ namespace grammar_parser {
                     }
                     print_grammar_char(file, elem.value);
                     break;
-                case LLAMA_GRETYPE_CHAR_ANY:
-                    fprintf(file, ".");
-                    break;
             }
             if (is_char_element(elem)) {
                 switch (rule[i + 1].type) {
                     case LLAMA_GRETYPE_CHAR_ALT:
                     case LLAMA_GRETYPE_CHAR_RNG_UPPER:
-                    case LLAMA_GRETYPE_CHAR_ANY:
                         break;
                     default:
                         fprintf(file, "] ");

common/json-schema-to-grammar.cpp

@@ -16,55 +16,92 @@ static std::string join(Iterator begin, Iterator end, const std::string & separa
 
 static std::string repeat(const std::string & str, size_t n);
 
-static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
-    auto has_max = max_items != std::numeric_limits<int>::max();
-
-    if (min_items == 0 && max_items == 1) {
-        return item_rule + "?";
-    }
-
+static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
     if (separator_rule.empty()) {
-        if (min_items == 1 && !has_max) {
+        if (min_items == 0 && max_items == 1) {
+            return item_rule + "?";
+        } else if (min_items == 1 && max_items == std::numeric_limits<int>::max()) {
             return item_rule + "+";
-        } else if (min_items == 0 && !has_max) {
-            return item_rule + "*";
-        } else {
-            return item_rule + "{" + std::to_string(min_items) + "," + (has_max ? std::to_string(max_items) : "") + "}";
         }
     }
 
-    auto result = item_rule + " " + build_repetition("(" + separator_rule + " " + item_rule + ")", min_items == 0 ? 0 : min_items - 1, has_max ? max_items - 1 : max_items);
-    if (min_items == 0) {
-        result = "(" + result + ")?";
+    std::string result;
+    if (min_items > 0) {
+        if (item_rule_is_literal && separator_rule.empty()) {
+            result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
+        } else {
+            std::vector<std::string> items(min_items, item_rule);
+            result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
+        }
     }
+
+    std::function<std::string(int, bool)> opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
+        auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
+
+        if (up_to_n == 0) {
+            return "";
+        } else if (up_to_n == 1) {
+            return "(" + content + ")?";
+        } else if (!separator_rule.empty() && !prefix_with_sep) {
+            return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
+        } else {
+            std::string res = repeat("(" + content + " ", up_to_n);
+            // strip trailing space
+            res = res.substr(0, res.length() - 1);
+            res += repeat(")?", up_to_n);
+            return res;
+        }
+    };
+
+    if (min_items > 0 && max_items != min_items) {
+        result += " ";
+    }
+
+    if (max_items != std::numeric_limits<int>::max()) {
+        result += opt_repetitions(max_items - min_items, min_items > 0);
+    } else {
+        std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
+        if (min_items == 0 && !separator_rule.empty()) {
+            result = "(" + item_rule + " " + item_operator + "*)?";
+        } else {
+            result += item_operator + "*";
+        }
+    }
+
     return result;
 }
 
-const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
+const std::string SPACE_RULE = "\" \"?";
 
 struct BuiltinRule {
     std::string content;
     std::vector<std::string> deps;
 };
 
+const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
+
 std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
     {"boolean", {"(\"true\" | \"false\") space", {}}},
-    {"decimal-part", {"[0-9]{1,16}", {}}},
-    {"integral-part", {"[0] | [1-9] [0-9]{0,15}", {}}},
+    {"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
+    {"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
     {"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
     {"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
     {"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
     {"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
     {"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
-    {"uuid", {"\"\\\"\" [0-9a-fA-F]{8} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{12} \"\\\"\" space", {}}},
-    {"char",   {"[^\"\\\\\\x7F\\x00-\\x1F] | [\\\\] ([\"\\\\bfnrt] | \"u\" [0-9a-fA-F]{4})", {}}},
+    {"uuid", {"\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space", {}}},
+    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
     {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
     {"null", {"\"null\" space", {}}},
 };
 
 std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
-    {"date", {"[0-9]{4} \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
-    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9]{3} )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
+    {"date", {"[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
+    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
     {"date-time", {"date \"T\" time", {"date", "time"}}},
     {"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
     {"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
@@ -235,7 +272,7 @@ private:
                     if (literal.empty()) {
                         return false;
                     }
-                    ret.emplace_back(literal, true);
+                    ret.push_back(std::make_pair(literal, true));
                     literal.clear();
                     return true;
                 };
@@ -261,7 +298,7 @@ private:
             while (i < length) {
                 char c = sub_pattern[i];
                 if (c == '.') {
-                    seq.emplace_back(get_dot(), false);
+                    seq.push_back(std::make_pair(get_dot(), false));
                     i++;
                 } else if (c == '(') {
                     i++;
@@ -270,7 +307,7 @@ private:
                             _warnings.push_back("Unsupported pattern syntax");
                         }
                     }
-                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
+                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
                 } else if (c == ')') {
                     i++;
                     if (start > 0 && sub_pattern[start - 1] != '(') {
@@ -294,9 +331,9 @@ private:
                     }
                     square_brackets += ']';
                     i++;
-                    seq.emplace_back(square_brackets, false);
+                    seq.push_back(std::make_pair(square_brackets, false));
                 } else if (c == '|') {
-                    seq.emplace_back("|", false);
+                    seq.push_back(std::make_pair("|", false));
                     i++;
                 } else if (c == '*' || c == '+' || c == '?') {
                     seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
@@ -348,7 +385,8 @@ private:
                         sub_is_literal ? "\"" + sub + "\"" : sub,
                         min_times,
                         max_times,
-                        ""
+                        "",
+                        sub_is_literal
                     );
                     seq.back().second = false;
                 } else {
@@ -379,7 +417,7 @@ private:
                         }
                     }
                     if (!literal.empty()) {
-                        seq.emplace_back(literal, true);
+                        seq.push_back(std::make_pair(literal, true));
                     }
                 }
             }

common/json-schema-to-grammar.h

@@ -1,8 +1,4 @@
 #pragma once
-
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);

common/log.h

@@ -211,7 +211,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_FLF_FMT "%s"
@@ -224,7 +224,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_TEE_FLF_FMT "%s"
@@ -294,7 +294,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Main LOG macro.
 //  behaves like printf, and supports arguments the exact same way.
 //
-#if !defined(_MSC_VER) || defined(__clang__)
+#ifndef _MSC_VER
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
     #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
@@ -308,14 +308,14 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Secondary target can be changed just like LOG_TARGET
 //  by defining LOG_TEE_TARGET
 //
-#if !defined(_MSC_VER) || defined(__clang__)
+#ifndef _MSC_VER
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
     #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
-#if !defined(_MSC_VER) || defined(__clang__)
+#ifndef _MSC_VER
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else

common/sampling.cpp

@@ -35,7 +35,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
-    result->n_valid = 0;
+    result->n_considered = 0;
 
     llama_sampling_set_rng_seed(result, params.seed);
 
@@ -66,7 +66,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
     ctx->cur.clear();
-    ctx->n_valid = 0;
+    ctx->n_considered = 0;
 }
 
 void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
@@ -125,7 +125,7 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
         for (auto sampler_type : params.samplers_sequence) {
-            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
             if (!sampler_type_name.empty()) {
                 result += "-> " + sampler_type_name + " ";
             }
@@ -137,87 +137,6 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     return result;
 }
 
-std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
-    switch (sampler_type) {
-        case llama_sampler_type::TOP_K:       return "top_k";
-        case llama_sampler_type::TFS_Z:       return "tfs_z";
-        case llama_sampler_type::TYPICAL_P:   return "typical_p";
-        case llama_sampler_type::TOP_P:       return "top_p";
-        case llama_sampler_type::MIN_P:       return "min_p";
-        case llama_sampler_type::TEMPERATURE: return "temperature";
-        default : return "";
-    }
-}
-
-std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
-    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
-        {"top_k",       llama_sampler_type::TOP_K},
-        {"top_p",       llama_sampler_type::TOP_P},
-        {"typical_p",   llama_sampler_type::TYPICAL_P},
-        {"min_p",       llama_sampler_type::MIN_P},
-        {"tfs_z",       llama_sampler_type::TFS_Z},
-        {"temperature", llama_sampler_type::TEMPERATURE}
-    };
-
-    // since samplers names are written multiple ways
-    // make it ready for both system names and input names
-    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
-        {"top-k",       llama_sampler_type::TOP_K},
-        {"top-p",       llama_sampler_type::TOP_P},
-        {"nucleus",     llama_sampler_type::TOP_P},
-        {"typical-p",   llama_sampler_type::TYPICAL_P},
-        {"typical",     llama_sampler_type::TYPICAL_P},
-        {"min-p",       llama_sampler_type::MIN_P},
-        {"tfs-z",       llama_sampler_type::TFS_Z},
-        {"tfs",         llama_sampler_type::TFS_Z},
-        {"temp",        llama_sampler_type::TEMPERATURE}
-    };
-
-    std::vector<llama_sampler_type> sampler_types;
-    sampler_types.reserve(names.size());
-    for (const auto & name : names)
-    {
-        auto sampler_item = sampler_canonical_name_map.find(name);
-        if (sampler_item != sampler_canonical_name_map.end())
-        {
-            sampler_types.push_back(sampler_item->second);
-        }
-        else
-        {
-            if (allow_alt_names)
-            {
-                sampler_item = sampler_alt_name_map.find(name);
-                if (sampler_item != sampler_alt_name_map.end())
-                {
-                    sampler_types.push_back(sampler_item->second);
-                }
-            }
-        }
-    }
-    return sampler_types;
-}
-
-std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
-    std::unordered_map<char, llama_sampler_type> sampler_name_map {
-        {'k', llama_sampler_type::TOP_K},
-        {'p', llama_sampler_type::TOP_P},
-        {'y', llama_sampler_type::TYPICAL_P},
-        {'m', llama_sampler_type::MIN_P},
-        {'f', llama_sampler_type::TFS_Z},
-        {'t', llama_sampler_type::TEMPERATURE}
-    };
-
-    std::vector<llama_sampler_type> sampler_types;
-    sampler_types.reserve(names_string.size());
-    for (const auto & c : names_string) {
-        const auto sampler_item = sampler_name_map.find(c);
-        if (sampler_item != sampler_name_map.end()) {
-            sampler_types.push_back(sampler_item->second);
-        }
-    }
-    return sampler_types;
-}
-
 // no reasons to expose this function in header
 static void sampler_queue(
                    struct llama_context * ctx_main,
@@ -260,7 +179,7 @@ static llama_token llama_sampling_sample_impl(
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
                   const int idx,
-                  bool is_resampling) {
+                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
     const llama_sampling_params & params = ctx_sampling->params;
 
     const float   temp            = params.temp;
@@ -269,8 +188,8 @@ static llama_token llama_sampling_sample_impl(
     const float   mirostat_eta    = params.mirostat_eta;
 
     std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
-    if (ctx_sampling->grammar != NULL && !is_resampling) {
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
+    if (!is_resampling) {
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
@@ -333,11 +252,11 @@ static llama_token llama_sampling_sample_impl(
             // Restore logits from the copy
             std::copy(original_logits.begin(), original_logits.end(), logits);
 
-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
         }
     }
 
-    ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;
+    ctx_sampling->n_considered = cur_p.size;
 
     return id;
 }
@@ -366,8 +285,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    if (ctx_sampling->grammar != NULL && !apply_grammar) {
-        GGML_ASSERT(original_logits != NULL);
+    if (apply_grammar && original_logits != NULL) {
         // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
         *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
     }
@@ -424,7 +342,7 @@ llama_token llama_sampling_sample(
                   struct llama_context * ctx_cfg,
                   const int idx) {
     // Call the implementation function with is_resampling set to false by default
-    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
 llama_token_data_array llama_sampling_prepare(

common/sampling.h

@@ -81,7 +81,7 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
-    size_t n_valid; // Number of correct top tokens with correct probabilities.
+    size_t n_considered;
 
     std::mt19937 rng;
 };
@@ -116,11 +116,6 @@ std::string llama_sampling_print(const llama_sampling_params & params);
 // Print sampling order into a string
 std::string llama_sampling_order_print(const llama_sampling_params & params);
 
-std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
-
-std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
-
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call

common/train.cpp

@@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() {
 
     params.custom_n_ctx = false;
 
-    params.use_flash              = false;
+    params.use_flash              = true;
     params.use_checkpointing      = true;
 
     params.sample_start           = "";
@@ -1380,7 +1380,7 @@ bool consume_common_train_arg(
 
 void finish_processing_train_args(struct train_params_common * params) {
     if (params->escape) {
-        string_process_escapes(params->sample_start);
+        process_escapes(params->sample_start);
     }
 }
 

convert-hf-to-gguf-update.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python3
-# -*- coding: utf-8 -*-
 
 # This script downloads the tokenizer models of the specified models from Huggingface and
 # generates the get_vocab_base_pre() function for convert-hf-to-gguf.py
@@ -21,13 +20,11 @@
 # - Update llama.cpp with the new pre-tokenizer if necessary
 #
 # TODO: generate tokenizer tests for llama.cpp
+# TODO: automate the update of convert-hf-to-gguf.py
 #
 
 import logging
 import os
-import pathlib
-import re
-
 import requests
 import sys
 import json
@@ -38,7 +35,6 @@ from transformers import AutoTokenizer
 
 logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger("convert-hf-to-gguf-update")
-sess = requests.Session()
 
 
 class TOKENIZER_TYPE(IntEnum):
@@ -53,10 +49,6 @@ chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍
 
 if len(sys.argv) == 2:
     token = sys.argv[1]
-    if not token.startswith("hf_"):
-        logger.info("Huggingface token seems invalid")
-        logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
-        sys.exit(1)
 else:
     logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
     sys.exit(1)
@@ -73,58 +65,68 @@ models = [
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
-    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
     {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
-    {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
-    {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
-    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
-    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
-    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
-    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
-    {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
-    {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
 ]
 
+# make directory "models/tokenizers" if it doesn't exist
+if not os.path.exists("models/tokenizers"):
+    os.makedirs("models/tokenizers")
+
 
 def download_file_with_auth(url, token, save_path):
     headers = {"Authorization": f"Bearer {token}"}
-    response = sess.get(url, headers=headers)
-    response.raise_for_status()
-    os.makedirs(os.path.dirname(save_path), exist_ok=True)
-    with open(save_path, 'wb') as f:
-        f.write(response.content)
-    logger.info(f"File {save_path} downloaded successfully")
+    response = requests.get(url, headers=headers)
+    if response.status_code == 200:
+        with open(save_path, 'wb') as f:
+            f.write(response.content)
+        logger.info(f"File {save_path} downloaded successfully")
+    else:
+        logger.info(f"Failed to download file. Status code: {response.status_code}")
 
 
-def download_model(model):
+# download the tokenizer models
+for model in models:
     name = model["name"]
     repo = model["repo"]
     tokt = model["tokt"]
 
-    os.makedirs(f"models/tokenizers/{name}", exist_ok=True)
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        os.makedirs(f"models/tokenizers/{name}")
+    else:
+        logger.info(f"Directory models/tokenizers/{name} already exists - skipping")
+        continue
+
+    logger.info(f"Downloading {name} to models/tokenizers/{name}")
+
+    url = f"{repo}/raw/main/config.json"
+    save_path = f"models/tokenizers/{name}/config.json"
+    download_file_with_auth(url, token, save_path)
+
+    url = f"{repo}/raw/main/tokenizer.json"
+    save_path = f"models/tokenizers/{name}/tokenizer.json"
+    download_file_with_auth(url, token, save_path)
+
+    # if downloaded file is less than 1KB, we likely need to download an LFS instead
+    if os.path.getsize(save_path) < 1024:
+        # remove the file
+        os.remove(save_path)
+        url = f"{repo}/resolve/main/tokenizer.json"
+        save_path = f"models/tokenizers/{name}/tokenizer.json"
+        download_file_with_auth(url, token, save_path)
 
-    files = ["config.json", "tokenizer.json", "tokenizer_config.json"]
     if tokt == TOKENIZER_TYPE.SPM:
-        files.append("tokenizer.model")
-
-    for file in files:
-        save_path = f"models/tokenizers/{name}/{file}"
-        if os.path.isfile(save_path):
-            logger.info(f"{name}: File {save_path} already exists - skipping")
-            continue
-        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
-
-
-for model in models:
-    try:
-        download_model(model)
-    except Exception as e:
-        logger.error(f"Failed to download model {model['name']}. Error: {e}")
+        url = f"{repo}/resolve/main/tokenizer.model"
+        save_path = f"models/tokenizers/{name}/tokenizer.model"
+        download_file_with_auth(url, token, save_path)
 
+    url = f"{repo}/raw/main/tokenizer_config.json"
+    save_path = f"models/tokenizers/{name}/tokenizer_config.json"
+    download_file_with_auth(url, token, save_path)
 
 # generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function:
+# TODO: auto-update convert-hf-to-gguf.py with the generated function
 
 src_ifs = ""
 for model in models:
@@ -134,17 +136,8 @@ for model in models:
     if tokt == TOKENIZER_TYPE.SPM:
         continue
 
-    # Skip if the tokenizer folder does not exist or there are other download issues previously
-    if not os.path.exists(f"models/tokenizers/{name}"):
-        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
-        continue
-
     # create the tokenizer
-    try:
-        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
-        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
-        continue  # Skip to the next model if the tokenizer can't be loaded
+    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
 
     chktok = tokenizer.encode(chktxt)
     chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -158,12 +151,8 @@ for model in models:
     # print the "pre_tokenizer" content from the tokenizer.json
     with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
         cfg = json.load(f)
-        normalizer = cfg["normalizer"]
-        logger.info("normalizer: " + json.dumps(normalizer, indent=4))
         pre_tokenizer = cfg["pre_tokenizer"]
         logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
-        if "ignore_merges" in cfg["model"]:
-            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
 
     logger.info("")
 
@@ -213,18 +202,11 @@ src_func = f"""
         return res
 """
 
-convert_py_pth = pathlib.Path("convert-hf-to-gguf.py")
-convert_py = convert_py_pth.read_text()
-convert_py = re.sub(
-    r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
-    lambda m: m.group(1) + src_func + m.group(3),
-    convert_py,
-    flags=re.DOTALL | re.MULTILINE,
-)
+print(src_func) # noqa: NP100
 
-convert_py_pth.write_text(convert_py)
-
-logger.info("+++ convert-hf-to-gguf.py was updated")
+logger.info("\n")
+logger.info("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!")
+logger.info("\n")
 
 # generate tests for each tokenizer model
 
@@ -271,7 +253,6 @@ tests = [
     "3333333",
     "33333333",
     "333333333",
-    # "Cửa Việt", # llama-bpe fails on this
     chktxt,
 ]
 
@@ -292,17 +273,8 @@ for model in models:
     name = model["name"]
     tokt = model["tokt"]
 
-    # Skip if the tokenizer folder does not exist or there are other download issues previously
-    if not os.path.exists(f"models/tokenizers/{name}"):
-        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
-        continue
-
     # create the tokenizer
-    try:
-        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
-        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
-        continue  # Skip this model and continue with the next one in the loop
+    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
 
     with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
         for text in tests:

convert-lora-to-ggml.py

@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import logging
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any, BinaryIO, Sequence
+
+import numpy as np
+import torch
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+logging.basicConfig(level=logging.DEBUG)
+logger = logging.getLogger("lora-to-gguf")
+
+NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
+
+
+def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
+    fout.write(b"ggla"[::-1])  # magic (ggml lora)
+    fout.write(struct.pack("i", 1))  # file version
+    fout.write(struct.pack("i", params["r"]))
+    # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
+    # but some models ship a float value instead
+    # let's convert to int, but fail if lossless conversion is not possible
+    assert (
+        int(params["lora_alpha"]) == params["lora_alpha"]
+    ), "cannot convert float to int losslessly"
+    fout.write(struct.pack("i", int(params["lora_alpha"])))
+
+
+def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
+    sname = name.encode("utf-8")
+    fout.write(
+        struct.pack(
+            "iii",
+            len(shape),
+            len(sname),
+            NUMPY_TYPE_TO_FTYPE[data_type.name],
+        )
+    )
+    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
+    fout.write(sname)
+    fout.seek((fout.tell() + 31) & -32)
+
+
+if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
+        logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
+        logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
+        sys.exit(1)
+
+    input_json = os.path.join(sys.argv[1], "adapter_config.json")
+    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
+    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+
+    if os.path.exists(input_model):
+        model = torch.load(input_model, map_location="cpu")
+    else:
+        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
+        # lazy import load_file only if lora is in safetensors format.
+        from safetensors.torch import load_file
+        model = load_file(input_model, device="cpu")
+
+    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+
+    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
+        logger.error(f"Error: unsupported architecture {arch_name}")
+        sys.exit(1)
+
+    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
+    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
+
+    with open(input_json, "r") as f:
+        params = json.load(f)
+
+    if params["peft_type"] != "LORA":
+        logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+        sys.exit(1)
+
+    if params["fan_in_fan_out"] is True:
+        logger.error("Error: param fan_in_fan_out is not supported")
+        sys.exit(1)
+
+    if params["bias"] is not None and params["bias"] != "none":
+        logger.error("Error: param bias is not supported")
+        sys.exit(1)
+
+    # TODO: these seem to be layers that have been trained but without lora.
+    # doesn't seem widely used but eventually should be supported
+    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
+        logger.error("Error: param modules_to_save is not supported")
+        sys.exit(1)
+
+    with open(output_path, "wb") as fout:
+        fout.truncate()
+
+        write_file_header(fout, params)
+        for k, v in model.items():
+            orig_k = k
+            if k.endswith(".default.weight"):
+                k = k.replace(".default.weight", ".weight")
+            if k in ["llama_proj.weight", "llama_proj.bias"]:
+                continue
+            if k.endswith("lora_A.weight"):
+                if v.dtype != torch.float16 and v.dtype != torch.float32:
+                    v = v.float()
+                v = v.T
+            else:
+                v = v.float()
+
+            t = v.detach().numpy()
+
+            prefix = "base_model.model."
+            if k.startswith(prefix):
+                k = k[len(prefix) :]
+
+            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
+            if k.endswith(lora_suffixes):
+                suffix = k[-len(lora_suffixes[0]):]
+                k = k[: -len(lora_suffixes[0])]
+            else:
+                logger.error(f"Error: unrecognized tensor name {orig_k}")
+                sys.exit(1)
+
+            tname = name_map.get_name(k)
+            if tname is None:
+                logger.error(f"Error: could not map tensor name {orig_k}")
+                logger.error(" Note: the arch parameter must be specified if the model is not llama")
+                sys.exit(1)
+
+            if suffix == ".lora_A.weight":
+                tname += ".weight.loraA"
+            elif suffix == ".lora_B.weight":
+                tname += ".weight.loraB"
+            else:
+                assert False
+
+            logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+            write_tensor_header(fout, tname, t.shape, t.dtype)
+            t.tofile(fout)
+
+    logger.info(f"Converted {input_json} and {input_model} to {output_path}")

convert-persimmon-to-gguf.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import logging
+import argparse
+import os
+import sys
+from pathlib import Path
+from pprint import pprint
+
+import torch
+from sentencepiece import SentencePieceProcessor
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+logger = logging.getLogger("persimmon-to-gguf")
+
+
+def _flatten_dict(dct, tensors, prefix=None):
+    assert isinstance(dct, dict)
+    for key in dct.keys():
+        new_prefix = prefix + '.' + key if prefix is not None else key
+        if isinstance(dct[key], torch.Tensor):
+            tensors[new_prefix] = dct[key]
+        elif isinstance(dct[key], dict):
+            _flatten_dict(dct[key], tensors, new_prefix)
+        else:
+            raise ValueError(type(dct[key]))
+    return None
+
+
+def _get_sentencepiece_tokenizer_info(dir_model: Path):
+    tokenizer_path = dir_model / 'adept_vocab.model'
+    logger.info('getting sentencepiece tokenizer from', tokenizer_path)
+    tokenizer = SentencePieceProcessor(str(tokenizer_path))
+    logger.info('adding tokens')
+    tokens: list[bytes] = []
+    scores: list[float] = []
+    toktypes: list[int] = []
+
+    for i in range(tokenizer.vocab_size()):
+        text: bytes
+        score: float
+
+        piece = tokenizer.id_to_piece(i)
+        text = piece.encode("utf-8")
+        score = tokenizer.get_score(i)
+
+        toktype = 1
+        if tokenizer.is_unknown(i):
+            toktype = 2
+        if tokenizer.is_control(i):
+            toktype = 3
+        if tokenizer.is_unused(i):
+            toktype = 5
+        if tokenizer.is_byte(i):
+            toktype = 6
+
+        tokens.append(text)
+        scores.append(score)
+        toktypes.append(toktype)
+        pass
+    return tokens, scores, toktypes
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
+    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
+    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
+    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
+    parser.add_argument("--adept-inference-dir", type=str,  help="path to adept-inference code directory")
+    parser.add_argument("--verbose",  action="store_true",  help="increase output verbosity")
+    args = parser.parse_args()
+    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+    sys.path.append(str(args.adept_inference_dir))
+    persimmon_model = torch.load(args.ckpt_path)
+    hparams = persimmon_model['args']
+    pprint(hparams)
+    tensors: dict[str, torch.Tensor] = {}
+    _flatten_dict(persimmon_model['model'], tensors, None)
+
+    arch = gguf.MODEL_ARCH.PERSIMMON
+    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
+
+    block_count = hparams.num_layers
+    head_count = hparams.num_attention_heads
+    head_count_kv = head_count
+    ctx_length = hparams.seq_length
+    hidden_size = hparams.hidden_size
+
+    gguf_writer.add_name('persimmon-8b-chat')
+    gguf_writer.add_context_length(ctx_length)
+    gguf_writer.add_embedding_length(hidden_size)
+    gguf_writer.add_block_count(block_count)
+    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
+    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
+    gguf_writer.add_head_count(head_count)
+    gguf_writer.add_head_count_kv(head_count_kv)
+    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
+    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
+
+    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
+    gguf_writer.add_tokenizer_model('llama')
+    gguf_writer.add_tokenizer_pre('default')
+    gguf_writer.add_token_list(tokens)
+    gguf_writer.add_token_scores(scores)
+    gguf_writer.add_token_types(toktypes)
+    gguf_writer.add_bos_token_id(71013)
+    gguf_writer.add_eos_token_id(71013)
+
+    tensor_map = gguf.get_tensor_name_map(arch, block_count)
+    logger.info(tensor_map)
+    for name in tensors.keys():
+        data_torch = tensors[name]
+        if name.endswith(".self_attention.rotary_emb.inv_freq"):
+            continue
+        old_dtype = data_torch.dtype
+        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
+        data = data_torch.to(torch.float32).squeeze().numpy()
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            raise ValueError(f"Can not map tensor '{name}'")
+
+        n_dims = len(data.shape)
+        logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}")
+        gguf_writer.add_tensor(new_name, data)
+    logger.info("gguf: write header")
+    gguf_writer.write_header_to_file()
+    logger.info("gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    logger.info("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+    gguf_writer.close()
+
+    logger.info(f"gguf: model successfully exported to '{args.outfile}'")
+
+
+if __name__ == '__main__':
+    main()

convert.py

@@ -24,16 +24,14 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
+from sentencepiece import SentencePieceProcessor
 
 if 'NO_LOCAL_GGUF' not in os.environ:
-    # use .parent.parent since we are in "examples" directory
-    sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py'))
-
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
-from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab
 
 if TYPE_CHECKING:
     from typing_extensions import Self, TypeAlias
@@ -176,7 +174,7 @@ class Params:
     rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
     f_rope_scale: float | None = None
-    n_ctx_orig: int | None = None
+    n_orig_ctx: int | None = None
     rope_finetuned: bool | None = None
 
     ftype: GGMLFileType | None = None
@@ -226,7 +224,7 @@ class Params:
         with open(config_path) as f:
             config = json.load(f)
 
-        rope_scaling_type = f_rope_scale = n_ctx_orig = rope_finetuned = None
+        rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
 
         if rope_scaling is not None and (typ := rope_scaling.get("type")):
@@ -236,7 +234,7 @@ class Params:
                 rope_scaling_type = gguf.RopeScalingType.LINEAR
             elif typ == "yarn":
                 rope_scaling_type = gguf.RopeScalingType.YARN
-                n_ctx_orig = rope_scaling['original_max_position_embeddings']
+                n_orig_ctx = rope_scaling['original_max_position_embeddings']
                 rope_finetuned = rope_scaling['finetuned']
             else:
                 raise NotImplementedError(f'Unknown rope scaling type: {typ}')
@@ -272,7 +270,7 @@ class Params:
             f_rope_freq_base  = config.get("rope_theta"),
             rope_scaling_type = rope_scaling_type,
             f_rope_scale      = f_rope_scale,
-            n_ctx_orig        = n_ctx_orig,
+            n_orig_ctx        = n_orig_ctx,
             rope_finetuned    = rope_finetuned,
         )
 
@@ -286,7 +284,6 @@ class Params:
         n_experts      = None
         n_experts_used = None
         f_rope_freq_base = None
-        n_ff = None
 
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
         if config.get("moe"):
@@ -311,8 +308,6 @@ class Params:
             n_experts_used = config["moe"]["num_experts_per_tok"]
             f_rope_freq_base = 1e6
 
-        assert n_ff is not None
-
         return Params(
             n_vocab          = model["tok_embeddings.weight"].shape[0],
             n_embd           = config["dim"],
@@ -346,40 +341,302 @@ class Params:
         return params
 
 
-@dataclass
-class Metadata:
-    name: Optional[str] = None
-    author: Optional[str] = None
-    version: Optional[str] = None
-    url: Optional[str] = None
-    description: Optional[str] = None
-    licence: Optional[str] = None
-    source_url: Optional[str] = None
-    source_hf_repo: Optional[str] = None
+#
+# vocab
+#
 
-    @staticmethod
-    def load(metadata_path: Path) -> Metadata:
-        if metadata_path is None or not metadata_path.exists():
-            return Metadata()
+@runtime_checkable
+class BaseVocab(Protocol):
+    tokenizer_model: ClassVar[str]
+    name: ClassVar[str]
 
-        with open(metadata_path, 'r') as file:
-            data = json.load(file)
 
-        # Create a new Metadata instance
-        metadata = Metadata()
+class NoVocab(BaseVocab):
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
 
-        # Assigning values to Metadata attributes if they exist in the JSON file
-        # This is based on LLM_KV_NAMES mapping in llama.cpp
-        metadata.name = data.get("general.name")
-        metadata.author = data.get("general.author")
-        metadata.version = data.get("general.version")
-        metadata.url = data.get("general.url")
-        metadata.description = data.get("general.description")
-        metadata.license = data.get("general.license")
-        metadata.source_url = data.get("general.source.url")
-        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
 
-        return metadata
+
+@runtime_checkable
+class Vocab(BaseVocab, Protocol):
+    vocab_size: int
+    added_tokens_dict: dict[str, int]
+    added_tokens_list: list[str]
+    fname_tokenizer: Path
+
+    def __init__(self, base_path: Path): ...
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
+
+
+class BpeVocab(Vocab):
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+
+        if (fname_tokenizer := base_path / 'vocab.json').exists():
+            # "slow" tokenizer
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                self.vocab = json.load(f)
+
+            try:
+                # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        else:
+            # "fast" tokenizer
+            fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+
+            # if this fails, FileNotFoundError propagates to caller
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+
+            tokenizer_model: dict[str, Any] = tokenizer_json['model']
+            if (
+                tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
+                or tokenizer_json['decoder']['type'] != 'ByteLevel'
+            ):
+                raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
+
+            self.vocab = tokenizer_model["vocab"]
+
+            if (added := tokenizer_json.get('added_tokens')) is not None:
+                # Added tokens here can be duplicates of the main vocabulary.
+                added_tokens = {item['content']: item['id']
+                                for item in added
+                                if item['content'] not in self.vocab}
+
+        vocab_size   = len(self.vocab)
+        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids   = sorted(added_tokens.values())
+        if expected_ids != actual_ids:
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
+                             f"{vocab_size} - {expected_end_id}; got {actual_ids}")
+
+        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_dict    = added_tokens
+        self.added_tokens_list    = [text for (text, idx) in items]
+        self.vocab_size_base      = vocab_size
+        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer      = fname_tokenizer
+
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
+
+        for i, _ in enumerate(self.vocab):
+            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.bpe_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+class SentencePieceVocab(Vocab):
+    tokenizer_model = "llama"
+    name = "spm"
+
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+        if (fname_tokenizer := base_path / 'tokenizer.model').exists():
+            # normal location
+            try:
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
+            # not found in alternate location either
+            raise FileNotFoundError('Cannot find tokenizer.model')
+
+        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        vocab_size = self.sentencepiece_tokenizer.vocab_size()
+
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
+        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
+        actual_new_ids   = sorted(new_tokens.keys())
+
+        if expected_new_ids != actual_new_ids:
+            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
+
+        # Token pieces that were added to the base vocabulary.
+        self.added_tokens_dict  = added_tokens
+        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
+        self.vocab_size_base    = vocab_size
+        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer    = fname_tokenizer
+
+    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        tokenizer = self.sentencepiece_tokenizer
+        for i in range(tokenizer.vocab_size()):
+            piece = tokenizer.id_to_piece(i)
+            text         = piece.encode("utf-8")
+            score: float = tokenizer.get_score(i)
+
+            toktype = gguf.TokenType.NORMAL
+            if tokenizer.is_unknown(i):
+                toktype = gguf.TokenType.UNKNOWN
+            if tokenizer.is_control(i):
+                toktype = gguf.TokenType.CONTROL
+
+            # NOTE: I think added_tokens are user defined.
+            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
+            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
+
+            if tokenizer.is_unused(i):
+                toktype = gguf.TokenType.UNUSED
+            if tokenizer.is_byte(i):
+                toktype = gguf.TokenType.BYTE
+
+            yield text, score, toktype
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.sentencepiece_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+class LlamaHfVocab(Vocab):
+    tokenizer_model = "llama"
+    name = "hfft"
+
+    def __init__(self, base_path: Path):
+        fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+        # if this fails, FileNotFoundError propagates to caller
+        with open(fname_tokenizer, encoding='utf-8') as f:
+            tokenizer_json = json.load(f)
+
+        # pre-check so we know if we need transformers
+        tokenizer_model: dict[str, Any] = tokenizer_json['model']
+        is_llama3 = (
+            tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
+            and not tokenizer_model.get('byte_fallback', True)
+        )
+        if is_llama3:
+            raise TypeError('Llama 3 must be converted with BpeVocab')
+
+        if not is_llama3 and (
+            tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
+            or tokenizer_json['decoder']['type'] != 'Sequence'
+        ):
+            raise FileNotFoundError('Cannot find Llama BPE tokenizer')
+
+        try:
+            from transformers import AutoTokenizer
+        except ImportError as e:
+            raise ImportError(
+                "To use LlamaHfVocab, please install the `transformers` package. "
+                "You can install it with `pip install transformers`."
+            ) from e
+
+        # Allow the tokenizer to default to slow or fast versions.
+        # Explicitly set tokenizer to use local paths.
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            base_path,
+            cache_dir=base_path,
+            local_files_only=True,
+        )
+        assert self.tokenizer.is_fast  # assume tokenizer.json is used
+
+        # Initialize lists and dictionaries for added tokens
+        self.added_tokens_list = []
+        self.added_tokens_dict = dict()
+        self.added_tokens_ids  = set()
+
+        # Process added tokens
+        for tok, tokidx in sorted(
+            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
+        ):
+            # Only consider added tokens that are not in the base vocabulary
+            if tokidx >= self.tokenizer.vocab_size:
+                self.added_tokens_list.append(tok)
+                self.added_tokens_dict[tok] = tokidx
+                self.added_tokens_ids.add(tokidx)
+
+        # Store special tokens and their IDs
+        self.specials = {
+            tok: self.tokenizer.get_vocab()[tok]
+            for tok in self.tokenizer.all_special_tokens
+        }
+        self.special_ids = set(self.tokenizer.all_special_ids)
+
+        # Set vocabulary sizes
+        self.vocab_size_base = self.tokenizer.vocab_size
+        self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
+
+        self.fname_tokenizer = fname_tokenizer
+
+    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        reverse_vocab = {
+            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
+        }
+
+        for token_id in range(self.vocab_size_base):
+            # Skip processing added tokens here
+            if token_id in self.added_tokens_ids:
+                continue
+
+            # Convert token text to bytes
+            token_text = reverse_vocab[token_id].encode("utf-8")
+
+            # Yield token text, score, and type
+            yield token_text, self.get_token_score(token_id), self.get_token_type(
+                token_id, token_text, self.special_ids  # Reuse already stored special IDs
+            )
+
+    def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType:
+        # Special case for byte tokens
+        if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+            return gguf.TokenType.BYTE
+
+        # Determine token type based on whether it's a special token
+        return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
+
+    def get_token_score(self, token_id: int) -> float:
+        # Placeholder for actual logic to determine the token's score
+        # This needs to be implemented based on specific requirements
+        return -1000.0  # Default score
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            if text in self.specials:
+                toktype = self.get_token_type(self.specials[text], b'', self.special_ids)
+                score = self.get_token_score(self.specials[text])
+            else:
+                toktype = gguf.TokenType.USER_DEFINED
+                score = -1000.0
+
+            yield text.encode("utf-8"), score, toktype
+
+    def has_newline_token(self):
+        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.hf_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 #
@@ -649,7 +906,7 @@ class LazyUnpickler(pickle.Unpickler):
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: dict[tuple[str, str], type[LazyTensor] | LazyStorageKind] = {
+    CLASSES = {
         # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
@@ -805,42 +1062,21 @@ class OutputFile:
     def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
-    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
-        # Metadata About The Model And Its Provenence
-        name = "LLaMA"
-        if metadata is not None and metadata.name is not None:
-            name = metadata.name
-        elif params.path_model is not None:
-            name = params.path_model.name
-        elif params.n_ctx == 4096:
-            # Heuristic detection of LLaMA v2 model
-            name = "LLaMA v2"
-
-        self.gguf.add_name(name)
-
-        if metadata is not None:
-            if metadata.author is not None:
-                self.gguf.add_author(metadata.author)
-            if metadata.version is not None:
-                self.gguf.add_version(metadata.version)
-            if metadata.url is not None:
-                self.gguf.add_url(metadata.url)
-            if metadata.description is not None:
-                self.gguf.add_description(metadata.description)
-            if metadata.licence is not None:
-                self.gguf.add_licence(metadata.licence)
-            if metadata.source_url is not None:
-                self.gguf.add_source_url(metadata.source_url)
-            if metadata.source_hf_repo is not None:
-                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
-
     def add_meta_arch(self, params: Params) -> None:
-        # Metadata About The Neural Architecture Itself
-        self.gguf.add_vocab_size(params.n_vocab)
-        self.gguf.add_context_length(params.n_ctx)
-        self.gguf.add_embedding_length(params.n_embd)
-        self.gguf.add_block_count(params.n_layer)
-        self.gguf.add_feed_forward_length(params.n_ff)
+        name = "LLaMA"
+
+        # TODO: better logic to determine model name
+        if params.n_ctx == 4096:
+            name = "LLaMA v2"
+        elif params.path_model is not None:
+            name = str(params.path_model.parent).split('/')[-1]
+
+        self.gguf.add_name                (name)
+        self.gguf.add_vocab_size          (params.n_vocab)
+        self.gguf.add_context_length      (params.n_ctx)
+        self.gguf.add_embedding_length    (params.n_embd)
+        self.gguf.add_block_count         (params.n_layer)
+        self.gguf.add_feed_forward_length (params.n_ff)
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
@@ -864,8 +1100,8 @@ class OutputFile:
             self.gguf.add_rope_scaling_type(params.rope_scaling_type)
             self.gguf.add_rope_scaling_factor(params.f_rope_scale)
 
-        if params.n_ctx_orig is not None:
-            self.gguf.add_rope_scaling_orig_ctx_len(params.n_ctx_orig)
+        if params.n_orig_ctx is not None:
+            self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx)
 
         if params.rope_finetuned is not None:
             self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
@@ -943,14 +1179,13 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
-        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
         of.add_meta_special_vocab(svocab)
@@ -977,14 +1212,12 @@ class OutputFile:
         fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
-        metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
-        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         if isinstance(vocab, Vocab):
             of.add_meta_vocab(vocab)
@@ -1020,37 +1253,6 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
     raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
-def model_parameter_count(model: LazyModel) -> int:
-    total_model_parameters = 0
-    for i, (name, lazy_tensor) in enumerate(model.items()):
-        sum_weights_in_tensor = 1
-        for dim in lazy_tensor.shape:
-            sum_weights_in_tensor *= dim
-        total_model_parameters += sum_weights_in_tensor
-    return total_model_parameters
-
-
-def model_parameter_count_rounded_notation(model_params_count: int) -> str:
-    if model_params_count > 1e12 :
-        # Trillions Of Parameters
-        scaled_model_params = model_params_count * 1e-12
-        scale_suffix = "T"
-    elif model_params_count > 1e9 :
-        # Billions Of Parameters
-        scaled_model_params = model_params_count * 1e-9
-        scale_suffix = "B"
-    elif model_params_count > 1e6 :
-        # Millions Of Parameters
-        scaled_model_params = model_params_count * 1e-6
-        scale_suffix = "M"
-    else:
-        # Thousands Of Parameters
-        scaled_model_params = model_params_count * 1e-3
-        scale_suffix = "K"
-
-    return f"{round(scaled_model_params)}{scale_suffix}"
-
-
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
     return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
             for (name, tensor) in model.items()}
@@ -1230,35 +1432,13 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
-    quantization = {
-        GGMLFileType.AllF32:    "F32",
-        GGMLFileType.MostlyF16: "F16",
-        GGMLFileType.MostlyQ8_0: "Q8_0",
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
+    namestr = {
+        GGMLFileType.AllF32:    "f32",
+        GGMLFileType.MostlyF16: "f16",
+        GGMLFileType.MostlyQ8_0:"q8_0",
     }[file_type]
-
-    parameters = model_parameter_count_rounded_notation(model_params_count)
-
-    expert_count = ""
-    if params.n_experts is not None:
-        expert_count = f"{params.n_experts}x"
-
-    version = ""
-    if metadata is not None and metadata.version is not None:
-        version = f"-{metadata.version}"
-
-    name = "ggml-model"
-    if metadata is not None and metadata.name is not None:
-        name = metadata.name
-    elif params.path_model is not None:
-        name = params.path_model.name
-
-    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
-
-
-def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
-    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
-    ret = model_paths[0].parent / f"{default_filename}.gguf"
+    ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
     if ret in model_paths:
         logger.error(
             f"Error: Default output path ({ret}) would overwrite the input. "
@@ -1296,30 +1476,17 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
     parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
-    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
-    parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
 
     args = parser.parse_args(args_in)
 
     if args.verbose:
         logging.basicConfig(level=logging.DEBUG)
-    elif args.dump_single or args.dump or args.get_outfile:
+    elif args.dump_single or args.dump:
         # Avoid printing anything besides the dump output
         logging.basicConfig(level=logging.WARNING)
     else:
         logging.basicConfig(level=logging.INFO)
 
-    metadata = Metadata.load(args.metadata)
-
-    if args.get_outfile:
-        model_plus = load_some_model(args.model)
-        params = Params.load(model_plus)
-        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
-        model_params_count = model_parameter_count(model_plus.model)
-        ftype   = pick_output_type(model, args.outtype)
-        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
-        return
-
     if args.no_vocab and args.vocab_only:
         raise ValueError("--vocab-only does not make sense with --no-vocab")
 
@@ -1333,9 +1500,6 @@ def main(args_in: list[str] | None = None) -> None:
     else:
         model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
 
-    model_params_count = model_parameter_count(model_plus.model)
-    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
-
     if args.dump:
         do_dump_model(model_plus)
         return
@@ -1344,27 +1508,25 @@ def main(args_in: list[str] | None = None) -> None:
     if args.big_endian:
         endianess = gguf.GGUFEndian.BIG
 
-    params = None
-    if args.pad_vocab or not args.vocab_only:
-        params = Params.load(model_plus)
-        if params.n_ctx == -1:
-            if args.ctx is None:
-                msg = """\
-                    The model doesn't have a context size, and you didn't specify one with --ctx
-                    Please specify one with --ctx:
-                     - LLaMA v1: --ctx 2048
-                     - LLaMA v2: --ctx 4096"""
-                parser.error(textwrap.dedent(msg))
-            params.n_ctx = args.ctx
+    params = Params.load(model_plus)
+    if params.n_ctx == -1:
+        if args.ctx is None:
+            msg = """\
+                The model doesn't have a context size, and you didn't specify one with --ctx
+                Please specify one with --ctx:
+                 - LLaMA v1: --ctx 2048
+                 - LLaMA v2: --ctx 4096"""
+            parser.error(textwrap.dedent(msg))
+        params.n_ctx = args.ctx
 
-        if args.outtype:
-            params.ftype = {
-                "f32": GGMLFileType.AllF32,
-                "f16": GGMLFileType.MostlyF16,
-                "q8_0": GGMLFileType.MostlyQ8_0,
-            }[args.outtype]
+    if args.outtype:
+        params.ftype = {
+            "f32": GGMLFileType.AllF32,
+            "f16": GGMLFileType.MostlyF16,
+            "q8_0": GGMLFileType.MostlyQ8_0,
+        }[args.outtype]
 
-        logger.info(f"params = {params}")
+    logger.info(f"params = {params}")
 
     model_parent_path = model_plus.paths[0].parent
     vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
@@ -1377,19 +1539,8 @@ def main(args_in: list[str] | None = None) -> None:
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
         outfile = args.outfile
-        if params is None:
-            params = Params(
-                n_vocab    = vocab.vocab_size,
-                n_embd     = 1,
-                n_layer    = 1,
-                n_ctx      = 1,
-                n_ff       = 1,
-                n_head     = 1,
-                n_head_kv  = 1,
-                f_norm_eps = 1e-5,
-            )
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
+                                    endianess=endianess, pad_vocab=args.pad_vocab)
         logger.info(f"Wrote {outfile}")
         return
 
@@ -1402,13 +1553,13 @@ def main(args_in: list[str] | None = None) -> None:
     model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")
 
     OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
     logger.info(f"Wrote {outfile}")
 
 

docs/HOWTO-add-model.md

@@ -17,7 +17,7 @@ Also, it is important to check that the examples and main ggml backends (CUDA, M
 ### 1. Convert the model to GGUF
 
 This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
-Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format).
+Depending on the model architecture, you can use either [convert.py](../convert.py) or [convert-hf-to-gguf.py](../convert-hf-to-gguf.py).
 
 The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
 
@@ -100,7 +100,7 @@ Have a look at existing implementation like `build_llama`, `build_dbrx` or `buil
 
 When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
 
-Note: to debug the inference graph: you can use [llama-eval-callback](../examples/eval-callback).
+Note: to debug the inference graph: you can use [eval-callback](../examples/eval-callback).
 
 ## GGUF specification
 

docs/debugging-tests.md

@@ -1,104 +0,0 @@
-# Debugging Tests Tips
-
-## How to run & execute or debug a specific test without anything else to keep the feedback loop short?
-
-There is a script called debug-test.sh in the scripts folder whose parameter takes a REGEX and an optional test number.
-
-For example, running the following command will output an interactive list from which you can select a test. It takes this form:
-
-`debug-test.sh [OPTION]... <test_regex> <test_number>`
-
-It will then build & run in the debugger for you.
-
-To just execute a test and get back a PASS or FAIL message run:
-
-```bash
-./scripts/debug-test.sh test-tokenizer
-```
-
-To test in GDB use the `-g` flag to enable gdb test mode.
-
-```bash
-./scripts/debug-test.sh -g test-tokenizer
-
-# Once in the debugger, i.e. at the chevrons prompt, setting a breakpoint could be as follows:
->>> b main
-```
-
-To speed up the testing loop, if you know your test number you can just run it similar to below:
-
-```bash
-./scripts/debug-test.sh test 23
-```
-
-For further reference use `debug-test.sh -h` to print help.
-
-&nbsp;
-
-### How does the script work?
-If you want to be able to use the concepts contained in the script separately, the important ones are briefly outlined below.
-
-#### Step 1: Reset and Setup folder context
-
-From base of this repository, let's create `build-ci-debug` as our build context.
-
-```bash
-rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
-```
-
-#### Step 2: Setup Build Environment and Compile Test Binaries
-
-Setup and trigger a build under debug mode. You may adapt the arguments as needed, but in this case these are sane defaults.
-
-```bash
-cmake -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON ..
-make -j
-```
-
-#### Step 3: Find all tests available that matches REGEX
-
-The output of this command will give you the command & arguments needed to run GDB.
-
-* `-R test-tokenizer` : looks for all the test files named `test-tokenizer*` (R=Regex)
-* `-N` : "show-only" disables test execution & shows test commands that you can feed to GDB.
-* `-V` : Verbose Mode
-
-```bash
-ctest -R "test-tokenizer" -V -N
-```
-
-This may return output similar to below (focusing on key lines to pay attention to):
-
-```bash
-...
-1: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
-1: Working Directory: .
-Labels: main
-  Test  #1: test-tokenizer-0-llama-spm
-...
-4: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-falcon.gguf"
-4: Working Directory: .
-Labels: main
-  Test  #4: test-tokenizer-0-falcon
-...
-```
-
-#### Step 4: Identify Test Command for Debugging
-
-So for test #1 above we can tell these two pieces of relevant information:
-* Test Binary: `~/llama.cpp/build-ci-debug/bin/test-tokenizer-0`
-* Test GGUF Model: `~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf`
-
-#### Step 5: Run GDB on test command
-
-Based on the ctest 'test command' report above we can then run a gdb session via this command below:
-
-```bash
-gdb --args ${Test Binary} ${Test GGUF Model}
-```
-
-Example:
-
-```bash
-gdb --args ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
-```

docs/token_generation_performance_tips.md

@@ -3,7 +3,7 @@
 ## Verifying that the model is running on the GPU with CUDA
 Make sure you compiled llama with the correct env variables according to [this guide](../README.md#CUDA), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
 ```shell
-./llama-cli -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
+./main -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
 ```
 
 When running llama, before it starts the inference work, it will output diagnostic information that shows whether cuBLAS is offloading work to the GPU. Look for these lines:
@@ -27,7 +27,7 @@ RAM: 32GB
 
 Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.q4_0.gguf` (30B parameters, 4bit quantization, GGML)
 
-Run command: `./llama-cli -m "path/to/model.gguf" -p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
+Run command: `./main -m "path/to/model.gguf" -p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
 
 Result:
 

examples/CMakeLists.txt

@@ -12,45 +12,41 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(cvector-generator)
     add_subdirectory(baby-llama)
-    add_subdirectory(batched-bench)
     add_subdirectory(batched)
+    add_subdirectory(batched-bench)
+    add_subdirectory(beam-search)
     add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
-    add_subdirectory(export-lora)
     add_subdirectory(finetune)
-    add_subdirectory(gbnf-validator)
-    add_subdirectory(gguf-split)
-    add_subdirectory(gguf)
     add_subdirectory(gritlm)
-    add_subdirectory(imatrix)
+    add_subdirectory(gguf-split)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
-    add_subdirectory(lookahead)
-    add_subdirectory(lookup)
-    add_subdirectory(main)
-    add_subdirectory(parallel)
-    add_subdirectory(passkey)
-    add_subdirectory(perplexity)
-    add_subdirectory(quantize-stats)
-    add_subdirectory(quantize)
-    add_subdirectory(retrieval)
-    if (LLAMA_RPC)
-        add_subdirectory(rpc)
-    endif()
-    if (LLAMA_BUILD_SERVER)
-    add_subdirectory(server)
-    endif()
     if (LLAMA_SYCL)
         add_subdirectory(sycl)
     endif()
+    add_subdirectory(main)
+    add_subdirectory(tokenize)
+    add_subdirectory(parallel)
+    add_subdirectory(perplexity)
+    add_subdirectory(quantize)
+    add_subdirectory(quantize-stats)
+    add_subdirectory(retrieval)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(passkey)
     add_subdirectory(speculative)
-    add_subdirectory(tokenize)
+    add_subdirectory(lookahead)
+    add_subdirectory(lookup)
+    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(imatrix)
+    if (LLAMA_BUILD_SERVER)
+        add_subdirectory(server)
+    endif()
+    add_subdirectory(export-lora)
 endif()

examples/Miku.sh

@@ -22,7 +22,7 @@ if [ -n "$N_THREAD" ]; then
     GEN_OPTIONS+=(--threads "$N_THREAD")
 fi
 
-./llama-cli "${GEN_OPTIONS[@]}" \
+./main "${GEN_OPTIONS[@]}" \
     --model "$MODEL" \
     --in-prefix " " \
     --in-suffix "${AI_NAME}:" \

examples/alpaca.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m ./models/alpaca.13b.ggmlv3.q8_0.bin \
+       --color \
+       -f ./prompts/alpaca.txt \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 7

examples/baby-llama/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET llama-baby-llama)
+set(TARGET baby-llama)
 add_executable(${TARGET} baby-llama.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

examples/baby-llama/baby-llama.cpp

@@ -522,8 +522,8 @@ static struct ggml_tensor * forward(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, 1]
             // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -759,8 +759,8 @@ static struct ggml_tensor * forward_batch(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, n_batch]
             // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
             assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
             assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
 
@@ -1056,7 +1056,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wqb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            KQ_pos, n_rot, 0);
+                                            KQ_pos, n_rot, 0, 0);
             struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                             ggml_reshape_3d(ctx0,
                                                 ggml_mul_mat(ctx0,
@@ -1065,7 +1065,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wkb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            KQ_pos, n_rot, 0);
+                                            KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {

examples/base-translate.sh

@@ -58,4 +58,4 @@ echo "$2
 model=$1
 
 # generate the most likely continuation until the string "===" is found
-./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs
+./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

examples/batched-bench/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET llama-batched-bench)
+set(TARGET batched-bench)
 add_executable(${TARGET} batched-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

examples/batched-bench/README.md

@@ -10,16 +10,16 @@ There are 2 modes of operation:
 - `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
 
 ```bash
-./llama-batched-bench -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
+./batched-bench MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
 
 # LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
-./llama-batched-bench -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 99
+./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 2048 512 0 99
 
 # LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
-./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 2048 512 1 99
 
 # custom set of batches
-./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 2048 -b 512 -ub 512 -ngl 999 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 512 512 0 999 0 128,256,512 128,256 1,2,4,8,16,32
 ```
 
 ## Sample results

examples/batched-bench/batched-bench.cpp

@@ -28,27 +28,67 @@ static std::vector<int> parse_list(char * p) {
     return ret;
 }
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
-    LOG_TEE("\nexample usage:\n");
-    LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
-    LOG_TEE("\n");
-}
-
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
-        return 1;
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [FATTN] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
+        printf("  example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        return 1 ;
     }
 
-    int is_pp_shared = params.is_pp_shared;
+    int n_kv_max     = 2048;
+    int n_batch      = 2048;
+    int n_ubatch     = 512;
+    bool flash_attn  = false;
+    int is_pp_shared = 0;
+    int n_gpu_layers = 0;
 
-    std::vector<int> n_pp = params.n_pp;
-    std::vector<int> n_tg = params.n_tg;
-    std::vector<int> n_pl = params.n_pl;
+    std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
+    std::vector<int> n_tg = { 128, 256, };
+    std::vector<int> n_pl = { 1, 2, 4, 8, 16, 32, };
+    //std::vector<int> n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, };
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        n_kv_max = std::atoi(argv[2]);
+    }
+
+    if (argc >= 4) {
+        n_batch = std::atoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        n_ubatch = std::atoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        flash_attn = std::atoi(argv[5]);
+    }
+
+    if (argc >= 7) {
+        is_pp_shared = std::atoi(argv[6]);
+    }
+
+    if (argc >= 8) {
+        n_gpu_layers = std::atoi(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_pp = parse_list(argv[8]);
+    }
+
+    if (argc >= 10) {
+        n_tg = parse_list(argv[9]);
+    }
+
+    if (argc >= 11) {
+        n_pl = parse_list(argv[10]);
+    }
 
     // init LLM
 
@@ -57,7 +97,12 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = llama_model_default_params();
+
+    const std::vector<float> t_split(llama_max_devices(), 0.0f);
+
+    model_params.n_gpu_layers = n_gpu_layers;
+    model_params.tensor_split = t_split.data();
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -66,7 +111,16 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed       = 1234;
+    ctx_params.n_ctx      = n_kv_max;
+    ctx_params.n_batch    = n_batch;
+    ctx_params.n_ubatch   = n_ubatch;
+    ctx_params.flash_attn = flash_attn;
+
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
     // ensure enough sequences are available
     ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
@@ -78,8 +132,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int32_t n_kv_max = llama_n_ctx(ctx);
-
     llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
 
     // decode in batches of ctx_params.n_batch tokens
@@ -123,7 +175,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, flash_attn, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

examples/batched.swift/Makefile

@@ -1,6 +1,6 @@
 .PHONY: build
 
 build:
-	xcodebuild -scheme llama-batched-swift -destination "generic/platform=macOS" -derivedDataPath build
-	rm -f ./llama-batched-swift
-	ln -s ./build/Build/Products/Debug/llama-batched-swift ./llama-batched-swift
+	xcodebuild -scheme batched_swift -destination "generic/platform=macOS" -derivedDataPath build
+	rm -f ./batched_swift
+	ln -s ./build/Build/Products/Debug/batched_swift ./batched_swift

examples/batched.swift/Package.swift

@@ -4,7 +4,7 @@
 import PackageDescription
 
 let package = Package(
-    name: "llama-batched-swift",
+    name: "batched_swift",
     platforms: [.macOS(.v12)],
     dependencies: [
         .package(name: "llama", path: "../../"),
@@ -13,7 +13,7 @@ let package = Package(
         // Targets are the basic building blocks of a package, defining a module or a test suite.
         // Targets can depend on other targets in this package and products from dependencies.
         .executableTarget(
-            name: "llama-batched-swift",
+            name: "batched_swift",
             dependencies: ["llama"],
             path: "Sources",
             linkerSettings: [.linkedFramework("Foundation"), .linkedFramework("AppKit")]

examples/batched.swift/README.md

@@ -1,4 +1,4 @@
 This is a swift clone of `examples/batched`.
 
 $ `make`
-$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET llama-batched)
+set(TARGET batched)
 add_executable(${TARGET} batched.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

examples/batched/README.md

@@ -3,7 +3,7 @@
 The example demonstrates batched generation from a given prompt
 
 ```bash
-./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
+./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
 
 ...
 

examples/batched/batched.cpp

@@ -7,31 +7,48 @@
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
-    LOG_TEE("\nexample usage:\n");
-    LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
-    LOG_TEE("\n");
-}
-
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    params.prompt = "Hello my name is";
-    params.n_predict = 32;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
-        return 1;
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
+        return 1 ;
     }
 
-
     // number of parallel batches
-    int n_parallel = params.n_parallel;
+    int n_parallel = 1;
 
     // total length of the sequences including the prompt
-    int n_predict = 32;
+    int n_len = 32;
+
+    // number of layers to offload to the GPU
+    int n_gpu_layers = 0;
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        params.prompt = argv[2];
+    }
+
+    if (argc >= 4) {
+        n_parallel = std::atoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        n_len = std::atoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        n_gpu_layers = std::atoi(argv[5]);
+    }
+
+    if (params.prompt.empty()) {
+        params.prompt = "Hello my name is";
+    }
+
+    process_escapes(params.prompt);
 
     // init LLM
 
@@ -40,7 +57,9 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = llama_model_default_params();
+
+    model_params.n_gpu_layers = n_gpu_layers;
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -54,14 +73,18 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
 
-    const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = llama_context_default_params();
 
+    ctx_params.seed  = 1234;
     ctx_params.n_ctx   = n_kv_req;
-    ctx_params.n_batch = std::max(n_predict, n_parallel);
+    ctx_params.n_batch = std::max(n_len, n_parallel);
+    ctx_params.n_seq_max       = n_parallel;
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -70,9 +93,9 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
@@ -133,7 +156,7 @@ int main(int argc, char ** argv) {
 
     const auto t_main_start = ggml_time_us();
 
-    while (n_cur <= n_predict) {
+    while (n_cur <= n_len) {
         // prepare the next batch
         llama_batch_clear(batch);
 
@@ -169,7 +192,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of generation? -> mark the stream as finished
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {

examples/beam-search/CMakeLists.txt

@@ -1,5 +1,5 @@
-set(TARGET llama-cvector-generator)
-add_executable(${TARGET} cvector-generator.cpp pca.hpp)
+set(TARGET beam-search)
+add_executable(${TARGET} beam-search.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/beam-search/beam-search.cpp

@@ -0,0 +1,188 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+// Used for debugging to print out beam tokens.
+struct ostream_beam_view {
+    llama_context * ctx;
+    llama_beam_view beam_view;
+};
+
+static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
+    os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
+    for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
+        os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
+    }
+    return os << ')';
+}
+
+// Put here anything you want back in beam_search_callback().
+struct beam_search_callback_data {
+    llama_context * ctx;
+    std::vector<llama_token> response;
+};
+
+// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
+// For example, eob can be flagged due to maximum token length, stop words, etc.
+static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
+    return n_tokens && llama_token_is_eog(llama_get_model(callback_data.ctx), tokens[n_tokens-1]);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+    auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        callback_data.response.resize(callback_data.response.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        std::copy(tokens, tokens + n, callback_data.response.end() - n);
+        printf("%zu", n);
+    }
+    fflush(stdout);
+#if 1 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n";
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{callback_data.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+    //params.n_gpu_layers = 200;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc < 2 || argv[1][0] == '-' )
+    {
+        printf( "Usage: %s MODEL_PATH [BEAM_WIDTH=2] [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    params.model = argv[1];
+
+    params.n_beams = 2 < argc ? std::stoi(argv[2]) : 2;
+
+    if ( argc > 3 )
+    {
+        params.prompt = argv[3];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+
+    if ( model == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
+
+    const size_t max_context_size     = llama_n_ctx( ctx );
+    const size_t max_tokens_list_size = max_context_size - 4 ;
+
+    if (tokens_list.size() > max_tokens_list_size)
+    {
+        fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" ,
+             __func__ , tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        std::cout << llama_token_to_piece(ctx, id);
+    }
+    std::cout << std::flush;
+
+    int n_past = 0;
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), tokens_list.size(), n_past, 0)))
+    {
+        fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
+        return 1;
+    }
+    n_past += tokens_list.size();
+
+    beam_search_callback_data callback_data{ctx, {}};
+    size_t const beam_width = static_cast<size_t>(params.n_beams);
+    int const n_predict = 256;
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict);
+
+    std::cout << "\n\n";
+    for (llama_token const token_id : callback_data.response) {
+        std::cout << llama_token_to_piece(ctx,token_id);
+    }
+    std::cout << std::endl;
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    llama_backend_free();
+
+    return 0;
+}

examples/benchmark/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET llama-bench-matmult)
+set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})

examples/chat-13B.sh

@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
      $PROMPT_TEMPLATE > $PROMPT_FILE
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./llama-cli $GEN_OPTIONS \
+./main $GEN_OPTIONS \
   --model "$MODEL" \
   --threads "$N_THREAD" \
   --n_predict "$N_PREDICTS" \

examples/chat-persistent.sh

@@ -62,7 +62,7 @@ fi
 if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
     echo 'Prompt cache does not exist, building...'
     # Default batch_size to 64 here for better user feedback during initial prompt processing
-    ./llama-cli 2>>"$LOG" \
+    ./main 2>>"$LOG" \
         --batch_size 64 \
         "${OPTS[@]}" \
         --prompt-cache "$PROMPT_CACHE_FILE" \
@@ -109,13 +109,13 @@ while read -e line; do
 
     printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
 
-    ./llama-cli 2>>"$LOG" "${OPTS[@]}" \
+    ./main 2>>"$LOG" "${OPTS[@]}" \
             --prompt-cache "$CUR_PROMPT_CACHE" \
             --prompt-cache-all \
             --file "$CUR_PROMPT_FILE" \
             --reverse-prompt "${USER_NAME}:" \
             --n_predict "$n_predict" |
-        skip_bytes 1 |                  # skip BOS token added by ./llama-cli
+        skip_bytes 1 |                  # skip BOS token added by ./main
         tee "$CUR_PROMPT_FILE.tmp" |    # save prompt + generation to tmp file
         skip_bytes "$n_prompt_len_pre"  # print generation
 
@@ -133,7 +133,7 @@ while read -e line; do
     # TODO get both messages in one go
     if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
         ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
-        echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
+        echo >&2 "Couldn't get number of tokens from ./main output!"
         exit 1
     fi
 
@@ -144,7 +144,7 @@ while read -e line; do
     fi
 
     # Update cache for next prompt in background, ideally during user input
-    ./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
+    ./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
           --prompt-cache "$NEXT_PROMPT_CACHE" \
           --file "$NEXT_PROMPT_FILE" \
           --n_predict 1 &

examples/chat-vicuna.sh

@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
      $PROMPT_TEMPLATE > $PROMPT_FILE
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./bin/llama-cli $GEN_OPTIONS \
+./bin/main $GEN_OPTIONS \
   --model "$MODEL" \
   --threads "$N_THREAD" \
   --n_predict "$N_PREDICTS" \

examples/chat.sh

@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./llama-cli -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
+./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
     --repeat_penalty 1.0 --color -i \
     -r "User:" -f prompts/chat-with-bob.txt

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

@@ -1,4 +1,4 @@
-set(TARGET llama-convert-llama2c-to-ggml)
+set(TARGET convert-llama2c-to-ggml)
 add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

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

@@ -2,13 +2,13 @@
 
 This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
 
-To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository:
+To convert the model first download the models from the [llma2.c](https://github.com/karpathy/llama2.c) repository:
 
 `$ make -j`
 
 After successful compilation, following usage options are available:
 ```
-usage: ./llama-convert-llama2c-to-ggml [options]
+usage: ./convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
@@ -19,10 +19,10 @@ options:
 
 An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
 
-`$ ./llama-convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
 Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
 
 Now you can use the model with a command like:
 
-`$ ./llama-cli -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
+`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -774,7 +774,7 @@ static struct train_params get_default_train_params() {
 
     params.samples_start_after_nl = false;
     params.use_adam               = true;
-    params.use_flash              = false;
+    params.use_flash              = true;
     params.use_scratch            = true;
 
     // only adam

examples/cvector-generator/README.md

@@ -1,34 +0,0 @@
-# cvector-generator
-
-This example demonstrates how to generate a control vector using gguf models.
-
-Related PRs:
-- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
-- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
-- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
-
-## Examples
-
-```sh
-# CPU only
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf
-
-# With GPU
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99
-
-# With advanced options
-./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100
-
-# To see help message
-./cvector-generator -h
-# Then, have a look at "cvector" section
-```
-
-## Tips and tricks
-
-If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
-
-```
-<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
-<|im_start|>system\nYou are in a very good mood today<|im_end|>
-```

examples/cvector-generator/completions.txt

@@ -1,582 +0,0 @@
-
-That game
-I can see
-Hmm, this
-I can relate to
-Who is
-I understand the
-Ugh,
-What the hell was
-Hey, did anyone
-Although
-Thank you for choosing
-What are you
-Oh w
-How dare you open
-It was my pleasure
-I'm hon
-I appreciate that you
-Are you k
-Whoever left this
-It's always
-Ew,
-Hey, I l
-Hello? Is someone
-I understand that
-That poem
-Aww, poor
-Hey, it
-Alright, who
-I didn't
-Well, life
-The document
-Oh no, this
-I'm concerned
-Hello, this is
-This art
-Hmm, this drink
-Hi there!
-It seems
-Is
-Good
-I can't
-Ex
-Who are
-I can see that
-Wow,
-Today is a
-Hey friend
-Sometimes friends
-Oh, this old
-The weather outside
-This place is sur
-I appreciate your input
-Thank you for the
-Look at
-I'm disappoint
-To my
-How dare you
-That's an
-This piece of art
-Eww
-This park is
-This is incredible
-Oh no, someone
-Exc
-Well, it'
-I warned
-Hey, I understand
-Hey, I saw
-How dare you go
-What the he
-Hey
-It's
-Hello? Hello?
-It
-Oh no!
-This is the perfect
-Good morning,
-Oh no, there
-It's so
-Yeah
-Uh,
-Hello everyone
-Who turned off
-The weather
-Who'
-Hey, this
-Wait,
-Eww, gross
-Excuse
-It seems like you
-Thank you so
-What happened?
-Oh my g
-I am deeply sad
-I war
-Okay, let'
-Hey, that
-That was a beautiful
-Oh no! That
-What happened
-Hey there
-The artist'
-What?!
-Hey, it'
-I am disappoint
-It seems like
-Oh no! The
-This park is a
-If you
-Yes! I did
-It sounds
-What
-Who is it
-Hmm, that
-That's strange
-Yeah, that was
-That's interesting
-This park
-What the hell
-Who is that
-I feel like my
-Oh well
-What the hell is
-Hello? Hello
-To my dearest
-Bless you!\"
-Thank you for
-Oh, looks like
-Can you please
-This place is
-Eww, what
-Bless you
-Is everything
-Hey, I just
-Whoever left these
-Well, that'
-I feel
-Hey, do you
-It's sad
-Oh no, it
-Hey, that'
-Oh my god,
-Thank you,
-Hello little one,
-I apolog
-Hey team, I
-How dare you read
-Who is this and
-Whoever left
-Hi there! W
-A
-If you have
-I was
-U
-Bless
-Well, this
-Oh, I'
-It's a
-Eww,
-Is everything okay?
-Oh, I
-Hello, can you
-Al
-That was a great
-What are
-I understand that not
-Oh no, not
-Who is it?\"
-Hey, can we
-Whoever is taking
-I would love to
-Hey, I noticed
-Hey, could
-I understand that there
-Hello?
-D
-Oh man, I
-Thank you so much
-Oh no, my
-Dear [Name
-Uh
-I remember
-Hey, who
-Well, it
-Are you
-I understand that it
-Hey, is
-I would
-Who is this
-Excuse me
-Alright
-I am thrilled
-Sometimes friends have
-Who the
-It's interesting
-I would love
-E
-Hello? Is anyone
-Well, this is
-This place
-Well,
-I warned you
-Hey, watch where
-Oh my
-That'
-Sometimes friends have different
-I understand that everyone
-What?
-What do these notes
-I can relate
-I'm not
-I understand
-To my dear
-Guys
-Well
-Hey, I appreciate
-Wow, what
-Dear
-That melody
-Who the hell
-Today is
-Hello little
-Wow, look
-That's great
-Love is never wrong
-I'm having
-Whoa, did
-Ugh
-Can you please provide
-I miss you,
-I feel uncom
-I know
-Ugh, this
-Hey, watch
-Oh great, a
-I didn
-Okay
-That game of char
-Oh
-I appreciate
-Who's there
-I am so
-Oh great, someone
-Hey, could you
-I remember wondering
-Wait, what?
-What do
-Hello? Can
-Hey there,
-That game of
-This is incred
-Oh my gosh
-Oh great, f
-I appreciate your
-It sounds like
-What the heck
-Okay, I understand
-Ew
-I understand that this
-Uh, hi
-Hi everyone!
-What the hell?
-Thank you for your
-Oh no, the
-Wow, I
-Who turned
-Dear [
-Whoever
-This is a
-Whoa, he
-What in the world
-Although the physical
-Hello, who is
-That's amaz
-Hey, I know
-Okay, that
-Hi everyone
-Hey, is everything
-I understand your fr
-Oh no, poor
-Oh, look
-Good morning
-Ew, gross
-Oh no, did
-Look at the family
-Hey team
-Yes!
-Hey, can I
-Okay, that'
-It's great
-Love is
-Hey, what
-Good morning, world
-Who is it?
-That poem really reson
-I
-That's
-I understand the task
-Gu
-Hello? Who'
-This postcard is
-Whoa,
-Oh, that
-I understand that I
-Whoever is
-Hello? Who is
-I'm really
-Wow, this
-Can
-This artwork really
-This is a shame
-I miss you too
-Who are you?
-Today is a difficult
-Hey, just
-Are you okay
-I am
-Hi,
-Wow, that
-Hey there! Can
-Okay, stay
-Oh great, just
-Yeah,
-Hello? Can you
-Oh, looks
-Thank you for sharing
-I'm glad
-Hey, is that
-Hmm
-It was my
-It sounds like you
-Wow, your
-I was promised certain
-That was such a
-Thank
-Excuse you
-That was
-Hey team,
-I feel un
-It was
-What'
-Hey friend, I
-How
-Saying goodbye
-That
-It's heart
-How dare
-Oh,
-Hello, may
-What's this
-Thank you for recogn
-Aww, that
-Oh, I remember
-Hmm, that'
-I miss
-I know this
-Wait
-Is everything okay
-Who is that person
-Wow, you
-Oh great
-I'm sad
-Wow, the
-I am very disappoint
-Who turned off the
-I understand that things
-I'm very
-Hi
-That's very
-Okay, I
-Oh no,
-Wow, there
-What's wrong
-I apologize for
-Hey, I
-Can I help you
-Oh, I didn
-Alright,
-Oh wow,
-Oh my goodness
-I know this event
-What in the
-Saying
-Yeah, that
-Guys, I
-Hey, this v
-This post
-Are
-Hey, can
-Hello? Is
-I can only imagine
-Oh, that sounds
-Hey, is anyone
-I am disappointed
-Hello,
-Hey everyone, I
-That was such
-It's okay
-The artist
-Whoa
-I understand that mistakes
-Can I help
-Who
-Hi everyone! I
-Hey, can you
-Wow, how
-Today
-Oh no, I
-Oh well, I
-Well, that
-This is the
-Yes! I finally
-Hey there little
-Hello everyone!
-Love is never
-Look at the
-This postcard
-Oh great,
-Can I
-Hmm, this is
-I understand your
-Oh, look at
-B
-I'm so
-Whoa, this
-W
-Oh, this
-Sometimes
-This piece of
-What the
-That was a
-Hey, do
-Oh no
-Whoa, what
-I feel like I
-The documentary
-Hello
-Hello little one
-I understand that my
-Eww, that
-Wow, an
-Yes! Finally,
-Although the physical location
-Whoever is watching
-That movie
-I remember wondering about
-Hey there, little
-Who's
-Hello, who
-Hello everyone! Thank
-Hello, can
-That's too
-Hey, just wanted
-Hey there, I
-Saying good
-Hey there!
-Who is there?
-Oh my good
-I am very
-Oh no, what
-Wow, thank
-I was promised
-Hi, is
-Hey, I'
-Guys, the
-Oh no, that
-Who is there
-Hello, this
-That movie really touched
-If you have something
-The documentary was
-I'm starting
-Are you kidd
-That movie really
-Hey everyone,
-Thank you for considering
-I didn'
-Yes! I
-Can you
-Oh my god
-Hey, whoever
-That melody really
-Thank you, little
-Hello, may I
-Look
-Wow, we
-It looks
-What do these
-Oh wow
-I apologize
-What are you all
-It's such
-It's clear
-Hey, I was
-Hey friend,
-I can only
-The weather outside is
-Eww, this
-I miss you
-Wow
-Aww,
-Hi, is there
-This artwork
-Okay,
-Oh well,
-This
-I'
-Say
-Hey there little gu
-Hmm,
-Whoa, who
-I am thr
-Oh man
-Okay, stay calm
-I'm happy
-Oh, this cur
-Oh man,
-I'm sorry
-Hello? Who
-What?! That
-This piece
-Hey everyone
-That's so
-Are you okay?
-What happened? Where
-Hi there
-The
-Who the hell entered
-I can
-Guys,
-What's
-What in
-It's important
-I'm
-I'm coming
-It'
-Yes! Finally
-Wait, what
-Wow, reading
-I'm surprised
-Hey, did
-Hey,
-Okay, let
-I understand that you
-Who the hell threw
-Eww, who
-Thank you for thinking
-Who is this?\"
-I am deeply
-Thank you for including
-Oh no, an
-It looks like you
-Aww
-I'm confused
-Wow, it
-That poem really
-Yes
-Hey there, is
-Hey, what'
-Thank you for remember
-To
-This is
-Thank you for making
-I can'
-That mel
-Wow, they
-I feel like
-Although the
-Who are you
-Love
-If
-What the hell are
-I am so sad
-Oh, I found
-Thank you
-It looks like
-Well, life is
-I appreciate that
-The artist's
-Whoa, that
-It's never

examples/cvector-generator/cvector-generator.cpp

@@ -1,499 +0,0 @@
-#include "common.h"
-#include "llama.h"
-#include "ggml.h"
-#include "pca.hpp"
-
-#ifdef GGML_USE_CUDA
-#include "ggml-cuda.h"
-#endif
-
-#ifdef GGML_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#include <cstdio>
-#include <string>
-#include <tuple>
-#include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
-#include <climits>
-
-
-//////////////////////////////////////////////////
-// utils
-
-template <class Iter>
-static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
-    std::string ret;
-    for (; begin != end; ++begin) {
-        ret += llama_token_to_piece(ctx, *begin);
-    }
-
-    return ret;
-}
-
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
-    printf("\nexample usage:\n");
-    printf("\n    CPU only:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf\n", argv[0]);
-    printf("\n    with GPU:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99\n", argv[0]);
-    printf("\n    advanced:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100\n", argv[0]);
-    printf("\n");
-}
-
-//////////////////////////////////////////////////
-
-
-// cb_eval is reused for each pair of positive - negative prompt
-struct callback_data {
-    ggml_context * ctx_ggml = nullptr;   // holds v_pos, v_neg, v_diff_filtered
-
-    int n_layers = 0;
-    int n_tokens = 0;
-    bool is_eval_pos = true;
-
-    // each element of the vector correspond to one layer
-    std::vector<struct ggml_tensor *> v_pos; // vector of matrices of size [n_embd, n_tokens]
-    std::vector<struct ggml_tensor *> v_neg; // vector of matrices of size [n_embd, n_tokens]
-    std::vector<struct ggml_tensor *> v_diff_filtered;   // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
-
-    // save a tensor into either v_pos or v_neg (decided by is_eval_pos)
-    void save_tensor_for_layer(struct ggml_tensor * t) {
-        GGML_ASSERT(t->type == GGML_TYPE_F32);
-
-        if (ctx_ggml == nullptr) {
-            // alloc a new ctx_ggml if needed
-            struct ggml_init_params params_ggml = {
-                /*.mem_size   =*/ ggml_tensor_overhead() * n_layers * 3u,
-                /*.mem_buffer =*/ NULL,
-                /*.no_alloc   =*/ true,
-            };
-            ctx_ggml = ggml_init(params_ggml);
-        }
-
-        // copy tensor data
-        auto n_bytes = ggml_nbytes(t);
-        struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
-        t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
-        ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
-        ggml_set_name(t_layer, ggml_get_name(t));
-        //print_debug_tensor(t_layer);
-
-        if (is_eval_pos) {
-            v_pos.push_back(t_layer);
-        } else {
-            v_neg.push_back(t_layer);
-        }
-    }
-
-    // calculate diff (v_pos - v_neg) and place the result back to v_pos
-    // all zero rows in the diff tensor will also be removed
-    // NOTE: final layer is ignored. we only have (n_layers - 1) to process
-    std::vector<struct ggml_tensor *> calc_diff() {
-        for (float il = 0; il < v_pos.size(); il++) {
-            float * a = (float *) v_pos[il]->data;
-            float * b = (float *) v_neg[il]->data;
-            size_t n_elem = ggml_nelements(v_pos[il]);
-            for (size_t j = 0; j < n_elem; j++) {
-                a[j] -= b[j];
-            }
-            //print_debug_tensor(v_pos[i]);
-            auto diff_filtered = filter_nonzero_rows(v_pos[il]);
-            v_diff_filtered.push_back(diff_filtered);
-        }
-        return v_diff_filtered; // for convinient, we return the result std::vector
-    }
-
-    // delete zero rows from a given 2D tensor
-    struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
-        //printf("filter_nonzero_rows\n");
-        auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
-            // check if given row containing all zero elements
-            int n_cols = t->ne[0]; // hint: should be equal to n_embd
-            for (int col = 0; col < n_cols; ++col) {
-                if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
-                    return false;
-                }
-            }
-            return true;
-        };
-        std::vector<int> rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
-        for (int i_row = 0; i_row < a->ne[1]; i_row++) {
-            if (!is_row_all_zeros(a, i_row, 1e-6)) {
-                rows_to_copy.push_back(i_row);
-            }
-        }
-
-        // get "n_nonzero_rows" for the output "diff_filtered"
-        int n_nonzero_rows = rows_to_copy.size();
-        //printf("n_nonzero_rows: %d\n", n_nonzero_rows);
-        int n_embd = a->ne[0];
-        GGML_ASSERT(n_nonzero_rows > 0);
-
-        // diff_filtered: [n_embd, n_nonzero_rows]
-        struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
-            ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
-        ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
-        diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
-
-        // copy non-zero rows
-        for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
-            int src_row = rows_to_copy[dest_row];
-            for (int i = 0; i < n_embd; i++) {
-                float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
-                ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
-            }
-        }
-
-        //print_debug_tensor(diff_filtered);
-
-        return diff_filtered;
-    }
-
-    // we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
-    void reset() {
-        for (auto ptr : v_pos) free(ptr->data);
-        for (auto ptr : v_neg) free(ptr->data);
-        for (auto ptr : v_diff_filtered) free(ptr->data);
-        v_pos.clear();
-        v_neg.clear();
-        v_diff_filtered.clear();
-        if (ctx_ggml) {
-            ggml_free(ctx_ggml);
-        }
-        ctx_ggml = nullptr;
-    }
-};
-
-/**
- * process_ctx is used to store the ggml context for pre-post processing the diff vectors
- * in short, input => v_diff and output => v_final
- */
-struct train_context {
-    ggml_context * ctx_ggml;
-    int n_embd;
-    int n_layers;
-
-    /* pair of prompts to be used for generating final vector */
-    std::vector<std::string> positive_entries;
-    std::vector<std::string> negative_entries;
-
-    // each element of the vector correspond to one layer
-    // NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
-    // NOTE (2): v_diff is transposed from v_diff_tmp
-    std::vector<struct ggml_tensor *> v_diff;  // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
-    std::vector<struct ggml_tensor *> v_final; // vector of vectors of size [n_embd] to be written to file
-
-    // to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
-    // v_diff_tmp will get converted unto v_diff later on
-    std::vector<std::vector<uint8_t>> v_diff_tmp;
-
-    train_context(int n_embd_, int n_layers_) {
-        n_embd = n_embd_;
-        n_layers = n_layers_;
-        struct ggml_init_params params_ggml = {
-            /*.mem_size   =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ true,
-        };
-        ctx_ggml = ggml_init(params_ggml);
-        for (int il = 0; il < n_layers - 1; il++) {
-            std::vector<uint8_t> empty;
-            v_diff_tmp.push_back(empty);
-            auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
-            t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
-            v_final.push_back(t);
-        }
-    }
-
-    // add new rows into existing tensor in v_diff_tmp
-    void concat_diff_tmp(const std::vector<struct ggml_tensor *> & diff_filtered) {
-        GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
-        for (int il = 0; il < n_layers - 1; il++) {
-            auto t = diff_filtered[il];
-            auto & diff_tmp = v_diff_tmp[il];
-            size_t curr_size = diff_tmp.size();
-            diff_tmp.resize(curr_size + ggml_nbytes(t));
-            memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
-        }
-    }
-
-    // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
-    // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
-    void build_v_diff() {
-        printf("build_v_diff\n");
-        for (int il = 0; il < n_layers - 1; il++) {
-            auto & diff_tmp = v_diff_tmp[il];
-            int n_elem = diff_tmp.size() / sizeof(float);
-            GGML_ASSERT(n_elem % n_embd == 0);
-            int n_rows = n_elem / n_embd;
-            struct ggml_tensor * diff = ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd);
-            ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
-            // copy data & transpose
-            diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
-            float * arr = (float *) diff_tmp.data();
-            for (int ir = 0; ir < n_rows; ++ir) {
-                for (int ic = 0; ic < n_embd; ++ic) {
-                    float f = arr[ir*n_embd + ic];
-                    ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
-                }
-            }
-            v_diff.push_back(diff);
-            print_debug_tensor(diff);
-            // free memory of diff_tmp
-            diff_tmp.resize(0);
-        }
-    }
-
-    ~train_context() {
-        for (auto ptr : v_final) free(ptr->data);
-        for (auto ptr : v_diff) free(ptr->data);
-        // no need to free v_diff_tmp, since we didn't use malloc
-        ggml_free(ctx_ggml);
-    }
-};
-
-struct tokenized_prompt {
-    std::vector<llama_token> tokens_pos;
-    std::vector<llama_token> tokens_neg;
-    size_t max_seq_len;
-
-    tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-        tokens_pos = ::llama_tokenize(ctx, pos, add_bos);
-        tokens_neg = ::llama_tokenize(ctx, neg, add_bos);
-        max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
-        padding_seq(ctx, tokens_pos, max_seq_len);
-        padding_seq(ctx, tokens_neg, max_seq_len);
-    }
-
-    void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
-        // TODO: customize padding token
-        std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
-        llama_token pad_tok = pad_tokens.back();
-        while (tokens.size() < len) {
-            tokens.push_back(pad_tok);
-        }
-    }
-};
-
-//////////////////////////////////////////////////
-
-template <typename T>
-static std::string to_string(const T & val) {
-    std::stringstream ss;
-    ss << val;
-    return ss.str();
-}
-
-static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
-    std::vector<std::string> output;
-    std::ifstream file(path);
-    if (!file.is_open()) {
-        fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
-        exit(1);
-    }
-    std::string line;
-    while (std::getline(file, line)) {
-        bool is_skip = skip_empty_lines && line.empty();
-        if (!is_skip) {
-            string_process_escapes(line);
-            output.push_back(line);
-        }
-    }
-    file.close();
-    return output;
-}
-
-//////////////////////////////////////////////////
-
-static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
-    auto * cb_data = (callback_data *) user_data;
-    static const char * l_out_name = "l_out";
-    const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
-
-    if (ask) {
-        return is_l_out;
-    }
-
-    if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
-        return true;
-    }
-
-    // save the tensor to current context
-    cb_data->save_tensor_for_layer(t);
-    return true;
-}
-
-static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_cache_clear(ctx);
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
-        return false;
-    }
-    return true;
-}
-
-static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const std::string fname, const std::string model_hint) {
-    struct gguf_context * ctx = gguf_init_empty();
-
-    const std::string arch = "controlvector";
-    gguf_set_val_str(ctx, "general.architecture", arch.c_str());
-    gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
-    gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
-
-    for (size_t i = 0; i < v_ctrl.size(); ++i) {
-        gguf_add_tensor(ctx, v_ctrl[i]);
-        print_debug_tensor(v_ctrl[i]);
-        printf("Added tensor: %s\n", v_ctrl[i]->name);
-    }
-
-    printf("%s: writing file...\n", __func__);
-    gguf_write_to_file(ctx, fname.c_str(), false);
-    printf("%s: wrote file '%s'\n", __func__, fname.c_str());
-    gguf_free(ctx);
-}
-
-/**
- * Load prompt files and completion file.
- * Then format each pair of prompt + completion to make an entry.
- */
-static int prepare_entries(gpt_params & params, train_context & ctx_train) {
-    // load prompts
-    std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
-    std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
-    if (positive_prompts.size() != negative_prompts.size()) {
-        fprintf(stderr, "number of positive and negative prompts must be equal\n");
-        return 1;
-    }
-    if (positive_prompts.empty()) {
-        fprintf(stderr, "must provide at least one prompt pair\n");
-        return 1;
-    }
-
-    // create templated prompts
-    std::vector<std::string> completions = ctrlvec_load_prompt_file(params.cvector_completions_file, false);
-    auto format_template = [](std::string persona, std::string suffix) {
-        // entry in positive/negative.txt must already be formatted i.e. "[INST] Act as if you're extremely happy. [/INST]"
-        return persona + " " + suffix;
-    };
-    for (size_t i = 0; i < positive_prompts.size(); ++i) {
-        for (int j = 0; j < std::min((int) completions.size(), params.n_completions); ++j) {
-            // TODO replicate the truncations done by the python implementation
-            ctx_train.positive_entries.push_back(format_template(positive_prompts[i], completions[j]));
-            ctx_train.negative_entries.push_back(format_template(negative_prompts[i], completions[j]));
-        }
-    }
-    return 0;
-}
-
-int main(int argc, char ** argv) {
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
-        return 1;
-    }
-
-    if (params.n_pca_iterations % params.n_pca_batch != 0) {
-        fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
-        return 1;
-    }
-
-
-    callback_data cb_data;
-
-    // pass the callback to the backend scheduler
-    // it will be executed for each node during the graph computation
-    params.cb_eval = cb_eval;
-    params.cb_eval_user_data = &cb_data;
-    params.warmup = false;
-
-    print_build_info();
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-
-    // int n_ctx = llama_n_ctx(ctx);
-    int n_layers = llama_n_layer(model);
-    int n_embd = llama_n_embd(model);
-    // get model hint param (a.k.a model arch name)
-    char model_hint[128];
-    llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
-
-    // init train_context
-    train_context ctx_train(n_embd, n_layers);
-
-    // load and prepare entries for training
-    prepare_entries(params, ctx_train);
-
-    // we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
-    std::vector<tokenized_prompt> tokenized_prompts;
-    size_t n_total_tokens = 0;
-    for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
-        tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
-        n_total_tokens += 2 * t.max_seq_len;
-        tokenized_prompts.push_back(std::move(t));
-    }
-
-    std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
-
-    for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
-        bool success = false;
-        tokenized_prompt t = tokenized_prompts[i];
-        cb_data.n_layers = n_layers;
-        cb_data.n_tokens = t.max_seq_len;
-
-        printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
-            (int) i+1, (int) ctx_train.positive_entries.size(),
-            tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
-            tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
-            (int) t.max_seq_len);
-
-        cb_data.is_eval_pos = true;
-        success = get_hidden_layers(ctx, t.tokens_pos);
-        if (!success) break;
-
-        cb_data.is_eval_pos = false;
-        success = get_hidden_layers(ctx, t.tokens_neg);
-        if (!success) break;
-
-        // calculate diff and remove all zero rows
-        auto v_diff_filtered = cb_data.calc_diff();
-
-        // save & concat the filtered v_diff to ctx_train
-        ctx_train.concat_diff_tmp(v_diff_filtered);
-
-        // reset for next iteration
-        cb_data.reset();
-    }
-
-    // done with the model, we can now free it to make gain some memory
-    printf("Done evaluate prompts, unload model...\n");
-    llama_free(ctx);
-    llama_free_model(model);
-
-    // prepare ctx_train for PCA
-    ctx_train.build_v_diff();
-
-    // run PCA
-    PCA::pca_params pca_params;
-    pca_params.n_threads = params.n_threads;
-    pca_params.n_batch = params.n_pca_batch;
-    pca_params.n_iterations = params.n_pca_iterations;
-    PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
-
-    // write output vectors to gguf
-    export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
-
-    llama_backend_free();
-
-    return 0;
-}