examples : remove --instruct remnants

.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/llama-cli-intel.Dockerfile

@@ -1,26 +0,0 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
-
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
-
-ARG LLAMA_SYCL_F16=OFF
-RUN apt-get update && \
-    apt-get install -y git
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
-        echo "LLAMA_SYCL_F16 is set" && \
-        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
-    fi && \
-    cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
-    cmake --build build --config Release --target llama-cli
-
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
-
-COPY --from=build /app/build/bin/llama-cli /llama-cli
-
-ENV LC_ALL=C.utf8
-
-ENTRYPOINT [ "/llama-cli" ]

.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/llama-server-intel.Dockerfile

@@ -1,29 +0,0 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
-
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
-
-ARG LLAMA_SYCL_F16=OFF
-RUN apt-get update && \
-    apt-get install -y git libcurl4-openssl-dev
-
-WORKDIR /app
-
-COPY . .
-
-RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
-        echo "LLAMA_SYCL_F16 is set" && \
-        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 ${OPT_SYCL_F16} && \
-    cmake --build build --config Release --target llama-server
-
-FROM intel/oneapi-basekit:$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
-
-ENV LC_ALL=C.utf8
-
-ENTRYPOINT [ "/llama-server" ]

.devops/main-cuda.Dockerfile

@@ -23,13 +23,13 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
 ENV LLAMA_CUDA=1
 
-RUN make -j$(nproc) llama-cli
+RUN make -j$(nproc) main
 
 FROM ${BASE_CUDA_RUN_CONTAINER} 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
 
-ENTRYPOINT [ "/llama-cli" ]
+ENTRYPOINT [ "/main" ]

.devops/main-intel.Dockerfile

@@ -0,0 +1,34 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build build --config Release --target main
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/main /main
+
+ENV LC_ALL=C.utf8
+
+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$(nproc) llama-cli
+RUN make -j$(nproc) main
 
-ENTRYPOINT [ "/app/llama-cli" ]
+ENTRYPOINT [ "/app/main" ]

.devops/main-vulkan.Dockerfile

@@ -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 +9,15 @@ WORKDIR /app
 
 COPY . .
 
-RUN make -j$(nproc) llama-cli
+RUN make -j$(nproc) main
 
 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

@@ -243,6 +243,8 @@ effectiveStdenv.mkDerivation (
     # 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 +294,7 @@ 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$(nproc) llama-server
+RUN make -j$(nproc) server
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/server /server
 
-ENTRYPOINT [ "/llama-server" ]
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -0,0 +1,45 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        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 ${OPT_SYCL_F16} && \
+    cmake --build build --config Release --target server
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+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$(nproc) llama-server
+RUN make -j$(nproc)
 
-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$(nproc) llama-server
+RUN make -j$(nproc) server
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1
 
-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

@@ -10,11 +10,11 @@ shift
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
     python3 ./convert-hf-to-gguf.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

@@ -26,7 +26,3 @@ indent_size = 2
 
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
-
-[examples/cvector-generator/*.txt]
-trim_trailing_whitespace = unset
-insert_final_newline = unset

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

@@ -24,7 +24,7 @@ body:
       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
+        $./main --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:

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

@@ -24,7 +24,7 @@ body:
       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
+        $./main --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:

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

@@ -24,7 +24,7 @@ body:
       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
+        $./main --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:

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

@@ -24,7 +24,7 @@ body:
       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
+        $./main --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:

.github/PULL_REQUEST_TEMPLATE/pull_request_template.md

@@ -0,0 +1,5 @@
+- Self Reported Review Complexity:
+    - [ ] Review Complexity : Low
+    - [ ] Review Complexity : Medium
+    - [ ] Review Complexity : High
+- [ ] I have read the [contributing guidelines](CONTRIBUTING.md)

.github/labeler.yml

@@ -42,6 +42,7 @@ build:
             - cmake/**
             - CMakeLists.txt
             - CMakePresets.json
+            - codecov.yml
 examples:
     - changed-files:
         - any-glob-to-any-file: examples/**

.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 }}
@@ -84,7 +84,7 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
-    runs-on: macos-12
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -103,10 +103,12 @@ 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
@@ -239,8 +241,8 @@ 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
@@ -682,7 +684,7 @@ 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
@@ -827,7 +829,7 @@ jobs:
           name: llama-bin-win-${{ matrix.build }}.zip
 
   windows-latest-cmake-cuda:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -841,9 +843,8 @@ jobs:
         with:
           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'

.github/workflows/code-coverage.yml

@@ -0,0 +1,40 @@
+name: Code Coverage
+on: [push, pull_request]
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+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: Build
+        run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
+
+      - name: Run tests
+        run: CC=gcc-8 make test
+
+      - name: Generate coverage report
+        run: |
+          make coverage
+          make lcov-report
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v3
+        env:
+           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
+        with:
+          files: lcov-report/coverage.info

.github/workflows/docker.yml

@@ -30,18 +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: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { 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/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" }
-          # Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
-          #- { 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-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: "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/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 }}
@@ -30,7 +32,7 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
@@ -87,22 +89,8 @@ jobs:
             exit 1
           fi
 
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DLLAMA_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DLLAMA_CURL=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-              -DLLAMA_OPENMP=OFF ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
       - name: Build
         id: cmake_build
-        if: ${{ matrix.sanitizer != 'THREAD' }}
         run: |
           cmake -B build \
               -DLLAMA_NATIVE=OFF \
@@ -110,7 +98,7 @@ jobs:
               -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ 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
@@ -127,7 +115,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     steps:
       - name: Clone
@@ -150,7 +138,7 @@ 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

.gitignore

@@ -1,123 +1,129 @@
-# Extensions
-
+*.o
 *.a
-*.bat
-*.bin
-*.dll
-*.dot
-*.etag
-*.exe
-*.gcda
-*.gcno
-*.gcov
+*.so
 *.gguf
 *.gguf.json
-*.lastModified
+*.bin
+*.exe
+*.dll
 *.log
-*.metallib
-*.o
-*.so
+*.gcov
+*.gcno
+*.gcda
+*.dot
+*.bat
 *.tmp
-
-# IDE / OS
-
+*.metallib
+*.etag
+*.lastModified
+.DS_Store
+.build/
 .cache/
 .ccls-cache/
 .direnv/
-.DS_Store
 .envrc
-.idea/
 .swiftpm
+.venv
+.clang-tidy
 .vs/
 .vscode/
-nppBackup
+.idea/
 
+ggml-metal-embed.metal
 
-# Coverage
-
-gcovr-report/
 lcov-report/
-
-# Build Artifacts
+gcovr-report/
 
 tags
-.build/
 build*
-!build-info.cmake
-!build-info.cpp.in
-!build-info.sh
 !build.zig
-/libllama.so
-/llama-*
-android-ndk-*
-arm_neon.h
 cmake-build-*
-CMakeSettings.json
-compile_commands.json
-ggml-metal-embed.metal
-llama-batched-swift
+android-ndk-*
 out/
 tmp/
 
-# CI
-
-!.github/workflows/*.yml
-
-# Models
-
 models/*
 models-mnt
-!models/.editorconfig
-!models/ggml-vocab-*.gguf*
 
-# Zig
+/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-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
+CMakeSettings.json
+
+__pycache__
+dist
 
 zig-out/
 zig-cache/
 
-# Logs
-
 ppl-*.txt
 qnt-*.txt
 perf-*.txt
 
-# Examples
-
 examples/jeopardy/results.txt
-examples/server/*.css.hpp
 examples/server/*.html.hpp
 examples/server/*.js.hpp
 examples/server/*.mjs.hpp
-!build_64.sh
-!examples/*.bat
-!examples/*/*.kts
-!examples/*/*/*.kts
-!examples/sycl/*.bat
-!examples/sycl/*.sh
+examples/server/*.css.hpp
 
-# Python
-
-__pycache__
-.venv
-/Pipfile
-dist
 poetry.lock
 poetry.toml
+nppBackup
 
 # Test binaries
-/tests/test-backend-ops
-/tests/test-double-float
-/tests/test-grad0
 /tests/test-grammar-parser
 /tests/test-llama-grammar
+/tests/test-double-float
+/tests/test-grad0
 /tests/test-opt
 /tests/test-quantize-fns
 /tests/test-quantize-perf
-/tests/test-rope
 /tests/test-sampling
 /tests/test-tokenizer-0
-/tests/test-tokenizer-1-bpe
 /tests/test-tokenizer-1-spm
-
-# Scripts
-!/scripts/install-oneapi.bat
+/tests/test-tokenizer-1-bpe
+/tests/test-rope
+/tests/test-backend-ops

CMakeLists.txt

@@ -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)
@@ -95,15 +91,13 @@ 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)
-option(LLAMA_CUDA_FORCE_MMQ                  "llama: always use mmq kernels instead of cuBLAS"  OFF)
-option(LLAMA_CUDA_FORCE_CUBLAS               "llama: always use cuBLAS instead of mmq kernels"  OFF)
+option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
@@ -120,7 +114,6 @@ option(LLAMA_HIP_UMA                         "llama: use HIP unified memory arch
 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})
@@ -145,6 +138,9 @@ option(LLAMA_BUILD_SERVER                    "llama: build server example"
 option(LLAMA_LASX                            "llama: enable lasx"                               ON)
 option(LLAMA_LSX                             "llama: enable lsx"                                ON)
 
+# add perf arguments
+option(LLAMA_PERF                            "llama: enable perf"                               OFF)
+
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 
@@ -315,9 +311,9 @@ 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)
@@ -325,7 +321,7 @@ 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)
@@ -378,15 +374,12 @@ 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()
@@ -409,27 +402,12 @@ if (LLAMA_CUBLAS)
 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 == FP16 CUDA intrinsics
-            # 61 == integer CUDA intrinsics
-            # 70 == FP16 tensor cores
-            # 75 == int8 tensor cores
-            if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
-                set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
-                #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)
@@ -449,9 +427,6 @@ if (LLAMA_CUDA)
         if (LLAMA_CUDA_FORCE_MMQ)
             add_compile_definitions(GGML_CUDA_FORCE_MMQ)
         endif()
-        if (LLAMA_CUDA_FORCE_CUBLAS)
-            add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
-        endif()
         if (LLAMA_CUDA_NO_VMM)
             add_compile_definitions(GGML_CUDA_NO_VMM)
         endif()
@@ -497,6 +472,21 @@ 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()
@@ -537,10 +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()
@@ -667,7 +653,6 @@ if (LLAMA_SYCL)
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
-    find_package(MKL REQUIRED)
 
     message(STATUS "SYCL found")
 
@@ -682,22 +667,21 @@ if (LLAMA_SYCL)
     endif()
 
     add_compile_options(-I./) #include DPCT
+    add_compile_options(-I/${SYCL_INCLUDE_DIR})
 
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
     if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
     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} IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
     else()
-        add_compile_options(-I/${SYCL_INCLUDE_DIR})
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
         if (LLAMA_SYCL_TARGET STREQUAL "INTEL")
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
         elseif (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
@@ -872,6 +856,10 @@ if (LLAMA_CPU_HBM)
     target_link_libraries(ggml PUBLIC memkind)
 endif()
 
+if (LLAMA_PERF)
+    add_compile_definitions(GGML_PERF)
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -1271,7 +1259,6 @@ add_library(ggml OBJECT
             ${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}
             )
 

CMakePresets.json

@@ -11,21 +11,9 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-    {
-        "name": "sycl-base",
-        "hidden": true,
-        "generator": "Ninja",
-        "binaryDir": "${sourceDir}/build-${presetName}",
-        "cacheVariables": {
-            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
-            "CMAKE_CXX_COMPILER": "icx",
-            "LLAMA_SYCL": "ON",
-            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
-        }
-    },
+
     { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
-    { "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
     { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
 
     {
@@ -47,18 +35,15 @@
     },
 
     { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
-    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
-    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
+    { "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",  "reldbg" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
+    { "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", "reldbg" ] },
-    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
-
-    { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
-    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
+    { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
+    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
   ]
 }

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 \
+	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 = \
@@ -344,6 +307,9 @@ ifdef LLAMA_GPROF
 	MK_CFLAGS   += -pg
 	MK_CXXFLAGS += -pg
 endif
+ifdef LLAMA_PERF
+	MK_CPPFLAGS += -DGGML_PERF
+endif
 
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
@@ -438,11 +404,10 @@ 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
 
@@ -453,30 +418,21 @@ ifndef LLAMA_NO_OPENMP
 endif # LLAMA_NO_OPENMP
 
 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_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_RPC
 	MK_CPPFLAGS   += -DGGML_USE_RPC
 	OBJS          += ggml-rpc.o
@@ -504,7 +460,7 @@ ifdef LLAMA_CUDA
 		CUDA_PATH ?= /usr/local/cuda
 	endif
 	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(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         += $(OBJS_CUDA_TEMP_INST)
@@ -537,9 +493,6 @@ endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_FORCE_MMQ
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # LLAMA_CUDA_FORCE_MMQ
-ifdef LLAMA_CUDA_FORCE_CUBLAS
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # LLAMA_CUDA_FORCE_CUBLAS
 ifdef LLAMA_CUDA_DMMV_X
 	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
@@ -608,10 +561,6 @@ 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
@@ -791,9 +740,6 @@ 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 $@
 
@@ -831,7 +777,7 @@ 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 +793,62 @@ 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/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)
 	$(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 +861,23 @@ 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 +888,55 @@ 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)
+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-export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(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-retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(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-speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(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-parallel: examples/parallel/parallel.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-lookahead: examples/lookahead/lookahead.cpp ggml.o llama.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: examples/lookup/lookup.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-create: examples/lookup/lookup-create.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-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 +962,20 @@ 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)
 
@@ -1051,7 +987,7 @@ tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-grammar-integration: tests/test-grammar-integration.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+tests/test-grammar-integration: tests/test-grammar-integration.cpp ggml.o llama.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 

README-sycl.md

@@ -1,7 +1,6 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
-- [Recommended Release](#recommended-release)
 - [News](#news)
 - [OS](#os)
 - [Hardware](#hardware)
@@ -32,23 +31,8 @@ When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneM
 
 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.
 
-## Recommended Release
-
-The SYCL backend would be broken by some PRs due to no online CI.
-
-The following release is verified with good quality:
-
-|Commit ID|Tag|Release|Verified  Platform|
-|-|-|-|-|
-|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
-
-
 ## News
 
-- 2024.5
-  - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
-  - Arch Linux is verified successfully.
-
 - 2024.4
   - Support data types: GGML_TYPE_IQ4_NL, GGML_TYPE_IQ4_XS, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S, GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M.
 
@@ -93,7 +77,7 @@ The following release is verified with good quality:
 *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.
 
@@ -115,14 +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
 
@@ -291,7 +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:
 ```
@@ -329,7 +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:
 
@@ -340,7 +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:
@@ -410,9 +394,15 @@ Output (example):
 
 4. Install build tools
 
-a. Download & install cmake for Windows: https://cmake.org/download/ (CMake can also be installed from Visual Studio Installer)
-b. The new Visual Studio will install Ninja as default. (If not, please install it manually: https://ninja-build.org/)
+a. Download & install cmake for Windows: https://cmake.org/download/
 
+b. Download & install mingw-w64 make for Windows provided by w64devkit
+
+- Download the 1.19.0 version of [w64devkit](https://github.com/skeeto/w64devkit/releases/download/v1.19.0/w64devkit-1.19.0.zip).
+
+- Extract `w64devkit` on your pc.
+
+- Add the **bin** folder path in the Windows system PATH environment (for e.g. `C:\xxx\w64devkit\bin\`).
 
 ### II. Build llama.cpp
 
@@ -422,10 +412,10 @@ On the oneAPI command line window, step into the llama.cpp main directory and ru
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -G "Ninja" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
 
 # Option 2: Or FP16
-cmake -B build -G "Ninja" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 cmake --build build --config Release -j
 ```
@@ -435,23 +425,9 @@ Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former in
 .\examples\sycl\win-build-sycl.bat
 ```
 
-Or, use CMake presets to build:
-```sh
-cmake --preset x64-windows-sycl-release
-cmake --build build-x64-windows-sycl-release -j --target llama-cli
-
-cmake -DLLAMA_SYCL_F16=ON --preset x64-windows-sycl-release
-cmake --build build-x64-windows-sycl-release -j --target llama-cli
-
-cmake --preset x64-windows-sycl-debug
-cmake --build build-x64-windows-sycl-debug -j --target llama-cli
-```
-
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
-
 *Notes:*
 
-- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
+- 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
 
@@ -512,13 +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

@@ -10,9 +10,6 @@
 
 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
@@ -195,7 +192,6 @@ 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)
@@ -209,7 +205,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
-- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
@@ -222,7 +217,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 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
@@ -388,30 +383,6 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
 
-### Nix
-
-On Mac and Linux, the Nix package manager can be used via
-```
-nix profile install nixpkgs#llama-cpp
-```
-For flake enabled installs.
-
-Or
-```
-nix-env --file '<nixpkgs>' --install --attr llama-cpp
-```
-For non-flake enabled installs.
-
-This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-#### Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-```
-flox install llama-cpp
-```
-Flox follows the nixpkgs build of llama.cpp.
-
 ### Metal Build
 
 On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
@@ -510,9 +481,8 @@ 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 custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). Speed for large batch sizes will be worse but VRAM consumption will be lower.                    |
-  | LLAMA_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
+  | 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_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.                                                                         |
@@ -585,7 +555,7 @@ 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
@@ -606,9 +576,7 @@ 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:
 
@@ -616,7 +584,7 @@ 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
@@ -649,18 +617,21 @@ python3 -m pip install -r requirements.txt
 # convert the model to ggml FP16 format
 python3 convert-hf-to-gguf.py models/mymodel/
 
+# [Optional] for models using BPE tokenizers
+python convert-hf-to-gguf.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.
@@ -735,7 +706,7 @@ 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
@@ -759,16 +730,16 @@ 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
@@ -790,7 +761,7 @@ 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).
@@ -869,7 +840,7 @@ $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:
@@ -936,8 +907,8 @@ 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.
@@ -987,7 +958,7 @@ 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)

ci/run.sh

@@ -303,47 +303,47 @@ function gg_run_open_llama_7b_v2 {
 
     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"
@@ -437,45 +437,45 @@ function gg_run_pythia_1_4b {
 
     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}  -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/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_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/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_60} -c 128 -b 128 --chunks 1 ) 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     --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
+    (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"
@@ -569,47 +569,47 @@ 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="$1"
@@ -693,10 +693,10 @@ function gg_run_embd_bge_small {
     model_f16="${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
 }

codecov.yml

@@ -0,0 +1,14 @@
+comment: off
+
+coverage:
+  status:
+    project:
+      default:
+        target: auto
+        threshold: 0
+        base: auto
+    patch:
+      default:
+        target: auto
+        threshold: 0
+        base: auto

common/common.h

@@ -73,6 +73,7 @@ struct gpt_params {
     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)
@@ -152,6 +153,7 @@ struct gpt_params {
     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 multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
@@ -178,12 +180,6 @@ struct gpt_params {
     std::string mmproj = "";        // path to multimodal projector
     std::vector<std::string> image; // path to image file(s)
 
-    // embedding
-    bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-    std::string embd_sep   = "\n";  // separator of embendings
-
     // server params
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
@@ -236,15 +232,6 @@ struct gpt_params {
 
     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);
@@ -382,7 +369,7 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
+void llama_embd_normalize(const float * inp, float * out, int n);
 
 float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 

common/json-schema-to-grammar.cpp

@@ -40,7 +40,7 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
+const std::string SPACE_RULE = "\" \"?";
 
 struct BuiltinRule {
     std::string content;
@@ -57,7 +57,7 @@ std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
     {"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})", {}}},
+    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F]{4})", {}}},
     {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
     {"null", {"\"null\" space", {}}},
 };

convert-hf-to-gguf-update.py

@@ -83,7 +83,6 @@ models = [
     {"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", },
 ]
 
@@ -214,7 +213,7 @@ src_func = f"""
 """
 
 convert_py_pth = pathlib.Path("convert-hf-to-gguf.py")
-convert_py = convert_py_pth.read_text(encoding="utf-8")
+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),
@@ -222,7 +221,7 @@ convert_py = re.sub(
     flags=re.DOTALL | re.MULTILINE,
 )
 
-convert_py_pth.write_text(convert_py, encoding="utf-8")
+convert_py_pth.write_text(convert_py)
 
 logger.info("+++ convert-hf-to-gguf.py was updated")
 

convert-hf-to-gguf.py

@@ -65,8 +65,7 @@ class Model:
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
-    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool,
-                 model_name: str | None, split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, model_name: str | None):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
         self.dir_model = dir_model
@@ -81,7 +80,7 @@ class Model:
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = Model.load_hparams(self.dir_model)
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         if self.ftype == gguf.LlamaFileType.GUESSED:
@@ -97,8 +96,7 @@ class Model:
         ftype_lw: str = ftype_up.lower()
         # allow templating the file name with the output ftype, useful with the "auto" ftype
         self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up)
-        self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
-                                           split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
+        self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
 
     @classmethod
     def __init_subclass__(cls):
@@ -334,8 +332,6 @@ class Model:
         self.gguf_writer.close()
 
     def write_vocab(self):
-        if len(self.gguf_writer.tensors) != 1:
-            raise ValueError('Splitting the vocabulary is not supported')
         self.gguf_writer.write_header_to_file(self.fname_out)
         self.gguf_writer.write_kv_data_to_file()
         self.gguf_writer.close()
@@ -481,9 +477,6 @@ class Model:
         if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
             # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
             res = "smaug-bpe"
-        if chkhsh == "c7ea5862a53e4272c035c8238367063e2b270d51faa48c0f09e9d5b54746c360":
-            # ref: https://huggingface.co/LumiOpen/Poro-34B-chat
-            res = "poro-chat"
         if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
             res = "jina-v2-code"
@@ -971,11 +964,7 @@ class XverseModel(Model):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model)
         vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-        # Since we are checking the maximum index, we need to ensure it's strictly less than vocab_size,
-        # because vocab_size is the count of items, and indexes start at 0.
-        max_vocab_index = max(tokenizer.get_vocab().values())
-        if max_vocab_index >= vocab_size:
-            raise ValueError("Vocabulary size exceeds expected maximum size.")
+        assert max(tokenizer.vocab.values()) < vocab_size
 
         reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
@@ -1408,48 +1397,6 @@ class LlamaModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@Model.register("BitnetForCausalLM")
-class BitnetModel(Model):
-    model_arch = gguf.MODEL_ARCH.BITNET
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-        self.gguf_writer.add_rope_scaling_factor(1.0)
-
-    def weight_quant(self, weight):
-        dtype = weight.dtype
-        weight = weight.float()
-        s = 1 / weight.abs().mean().clamp(min=1e-5)
-        weight = (weight * s).round().clamp(-1, 1) / s
-        scale = weight.abs().max().unsqueeze(0)
-        weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
-        weight = torch.sign(weight).type(dtype)
-        return weight.type(dtype), scale.type(torch.float32)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        new_name = self.map_tensor_name(name)
-
-        if any(self.match_model_tensor_name(new_name, key, bid) for key in [
-            gguf.MODEL_TENSOR.ATTN_Q,
-            gguf.MODEL_TENSOR.ATTN_K,
-            gguf.MODEL_TENSOR.ATTN_V,
-            gguf.MODEL_TENSOR.ATTN_OUT,
-            gguf.MODEL_TENSOR.FFN_UP,
-            gguf.MODEL_TENSOR.FFN_DOWN,
-            gguf.MODEL_TENSOR.FFN_GATE,
-        ]):
-            # transform weight into 1/0/-1 (in fp32)
-            weight_torch, scale_torch = self.weight_quant(data_torch)
-            yield (new_name, weight_torch)
-            yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
-        else:
-            yield (new_name, data_torch)
-
-
 @Model.register("GrokForCausalLM")
 class GrokModel(Model):
     model_arch = gguf.MODEL_ARCH.GROK
@@ -1682,12 +1629,6 @@ class Qwen2MoeModel(Model):
         super().set_gguf_parameters()
         if (n_experts := self.hparams.get("num_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
-        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
-            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
-            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
-        if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
-            self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
-            logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -2775,124 +2716,6 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@Model.register("T5ForConditionalGeneration")
-@Model.register("T5WithLMHeadModel")
-class T5Model(Model):
-    model_arch = gguf.MODEL_ARCH.T5
-
-    def set_vocab(self):
-        # to avoid TypeError: Descriptors cannot be created directly
-        # exception when importing sentencepiece_model_pb2
-        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
-        from sentencepiece import SentencePieceProcessor
-        from sentencepiece import sentencepiece_model_pb2 as model
-
-        tokenizer_path = self.dir_model / 'spiece.model'
-
-        if not tokenizer_path.is_file():
-            raise FileNotFoundError(f"File not found: {tokenizer_path}")
-
-        sentencepiece_model = model.ModelProto()
-        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
-        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
-        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
-        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
-        assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM
-
-        tokenizer = SentencePieceProcessor()
-        tokenizer.LoadFromFile(str(tokenizer_path))
-
-        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
-
-        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
-        scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
-
-        for token_id in range(tokenizer.vocab_size()):
-            piece = tokenizer.IdToPiece(token_id)
-            text = piece.encode("utf-8")
-            score = tokenizer.GetScore(token_id)
-
-            toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.IsUnknown(token_id):
-                toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.IsControl(token_id):
-                toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.IsUnused(token_id):
-                toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.IsByte(token_id):
-                toktype = SentencePieceTokenTypes.BYTE
-
-            tokens[token_id] = text
-            scores[token_id] = score
-            toktypes[token_id] = toktype
-
-        added_tokens_file = self.dir_model / 'added_tokens.json'
-        if added_tokens_file.is_file():
-            with open(added_tokens_file, "r", encoding="utf-8") as f:
-                added_tokens_json = json.load(f)
-                for key in added_tokens_json:
-                    token_id = added_tokens_json[key]
-                    if (token_id >= vocab_size):
-                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
-                        continue
-
-                    tokens[token_id] = key.encode("utf-8")
-                    scores[token_id] = -1000.0
-                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
-
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
-            for i in range(1, pad_count + 1):
-                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
-
-        self.gguf_writer.add_tokenizer_model("t5")
-        self.gguf_writer.add_tokenizer_pre("default")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-        self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
-        if precompiled_charsmap:
-            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-        special_vocab.add_to_gguf(self.gguf_writer)
-
-        self.gguf_writer.add_add_bos_token(False)
-        self.gguf_writer.add_add_eos_token(True)
-
-    def set_gguf_parameters(self):
-        self.gguf_writer.add_name("T5")
-        self.gguf_writer.add_context_length(self.hparams["n_positions"])
-        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
-        self.gguf_writer.add_block_count(self.hparams["num_layers"])
-        self.gguf_writer.add_head_count(self.hparams["num_heads"])
-        self.gguf_writer.add_key_length(self.hparams["d_kv"])
-        self.gguf_writer.add_value_length(self.hparams["d_kv"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_decoder_start_token_id(self.hparams["decoder_start_token_id"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
-        # Sometimes T5 and Flan-T5 based models contain "encoder.embed_tokens.weight" tensor or
-        # "decoder.embed_tokens.weight" tensors that are duplicates of "shared.weight" tensor
-        # To prevent errors caused by an unnecessary unmapped tensor, skip both of them and use only "shared.weight".
-        if name == "decoder.embed_tokens.weight" or name == "encoder.embed_tokens.weight":
-            logger.debug(f"Skipping tensor {name!r} in safetensors so that convert can end normally.")
-            return []
-
-        return [(self.map_tensor_name(name), data_torch)]
-
-
 ###### CONVERSION LOGIC ######
 
 
@@ -2978,44 +2801,10 @@ def parse_args() -> argparse.Namespace:
         "--verbose", action="store_true",
         help="increase output verbosity",
     )
-    parser.add_argument(
-        "--split-max-tensors", type=int, default=0,
-        help="max tensors in each split",
-    )
-    parser.add_argument(
-        "--split-max-size", type=str, default="0",
-        help="max size per split N(M|G)",
-    )
-    parser.add_argument(
-        "--dry-run", action="store_true",
-        help="only print out a split plan and exit, without writing any new files",
-    )
-    parser.add_argument(
-        "--no-tensor-first-split", action="store_true",
-        help="do not add tensors to the first split (disabled by default)"
-    )
 
     return parser.parse_args()
 
 
-def split_str_to_n_bytes(split_str: str) -> int:
-    if split_str.endswith("K"):
-        n = int(split_str[:-1]) * 1000
-    elif split_str.endswith("M"):
-        n = int(split_str[:-1]) * 1000 * 1000
-    elif split_str.endswith("G"):
-        n = int(split_str[:-1]) * 1000 * 1000 * 1000
-    elif split_str.isnumeric():
-        n = int(split_str)
-    else:
-        raise ValueError(f"Invalid split size: {split_str}, must be a number, optionally followed by K, M, or G")
-
-    if n < 0:
-        raise ValueError(f"Invalid split size: {split_str}, must be positive")
-
-    return n
-
-
 def main() -> None:
     args = parse_args()
 
@@ -3048,10 +2837,6 @@ def main() -> None:
         "auto": gguf.LlamaFileType.GUESSED,
     }
 
-    if args.use_temp_file and (args.split_max_tensors > 0 or args.split_max_size != "0"):
-        logger.error("Error: Cannot use temp file when splitting")
-        sys.exit(1)
-
     if args.outfile is not None:
         fname_out = args.outfile
     else:
@@ -3069,10 +2854,7 @@ def main() -> None:
             logger.error(f"Model {hparams['architectures'][0]} is not supported")
             sys.exit(1)
 
-        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file,
-                                     args.no_lazy, args.model_name, split_max_tensors=args.split_max_tensors,
-                                     split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
-                                     small_first_shard=args.no_tensor_first_split)
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy, args.model_name)
 
         logger.info("Set model parameters")
         model_instance.set_gguf_parameters()
@@ -3083,13 +2865,13 @@ def main() -> None:
         model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
 
         if args.vocab_only:
-            logger.info("Exporting model vocab...")
+            logger.info(f"Exporting model vocab to '{model_instance.fname_out}'")
             model_instance.write_vocab()
-            logger.info("Model vocab successfully exported.")
         else:
-            logger.info("Exporting model...")
+            logger.info(f"Exporting model to '{model_instance.fname_out}'")
             model_instance.write()
-            logger.info("Model successfully exported.")
+
+        logger.info(f"Model successfully exported to '{model_instance.fname_out}'")
 
 
 if __name__ == '__main__':

docs/HOWTO-add-model.md

@@ -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/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 +12,43 @@ 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(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)
+    if (LLAMA_RPC)
+        add_subdirectory(rpc)
+    endif()
 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/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/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 -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
 
 # 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 -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 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 -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
 
 # 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 -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
 ```
 
 ## Sample results

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 -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
 
 ...
 

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

@@ -8,7 +8,7 @@ To convert the model first download the models from the [llama2.c](https://githu
 
 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/cvector-generator/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-cvector-generator)
-add_executable(${TARGET} cvector-generator.cpp pca.hpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

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 --pca-batch 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 --pca-batch 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;
-}

examples/cvector-generator/negative.txt

@@ -1 +0,0 @@
-[INST] Act like a person who is extremely sad. [/INST] 

examples/cvector-generator/pca.hpp

@@ -1,322 +0,0 @@
-#include "common.h"
-#include "llama.h"
-#include "ggml.h"
-
-#ifdef GGML_USE_CUDA
-#include "ggml-cuda.h"
-#endif
-
-#ifdef GGML_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#include <cstdio>
-#include <ctime>
-#include <string>
-#include <tuple>
-#include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
-
-#define DEBUG_POS 5
-
-static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) {
-    printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]);
-    if (!with_data) return;
-    printf("%s: %s[0] = [", __func__, t->name);
-    for (size_t i = 0; i <= DEBUG_POS; i++) {
-        printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0));
-    }
-    printf(" ... ]\n");
-}
-
-namespace PCA {
-
-// input params for PCA computations
-struct pca_params {
-    int n_threads = 1;
-    int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used
-    int n_iterations = 1000;
-    float tolerance = 1e-7;
-
-    // for debugging
-    int i_layer = 0;
-    int n_layers = 0;
-};
-
-// result from each iteration
-struct pca_result {
-    struct ggml_tensor * calculated_square = NULL;
-    std::vector<struct ggml_tensor *> eigenvectors;
-    std::vector<float> distances;
-};
-
-struct pca_model {
-    ggml_backend_t backend = NULL;
-    ggml_backend_buffer_t buffer;
-    struct ggml_context * ctx;      // context to compute graph on target device
-    struct ggml_context * ctx_host; // host context to store results
-
-    // tensors on target device
-    struct ggml_tensor * dev_input;
-    struct ggml_tensor * dev_square;
-    struct ggml_tensor * dev_eigenvector;
-
-    pca_model(struct ggml_tensor * t_input) {
-#ifdef GGML_USE_CUDA
-        fprintf(stderr, "%s: using CUDA backend\n", __func__);
-        backend = ggml_backend_cuda_init(0); // init device 0
-        if (!backend) {
-            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
-        }
-#endif
-
-// TODO: enable Metal support when support for GGML_OP_SQRT is added
-// #ifdef GGML_USE_METAL
-//         fprintf(stderr, "%s: using Metal backend\n", __func__);
-//         backend = ggml_backend_metal_init();
-//         if (!backend) {
-//             fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
-//         }
-// #endif
-
-        // if there aren't GPU Backends fallback to CPU backend
-        if (!backend) {
-            backend = ggml_backend_cpu_init();
-        }
-
-        const int num_tensors = 4;
-        struct ggml_init_params params {
-            /*.mem_size   =*/ ggml_tensor_overhead() * num_tensors,
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ true,
-        };
-        ctx = ggml_init(params);
-
-        auto n_samples = t_input->ne[0];
-        auto n_embd    = t_input->ne[1];
-
-        dev_input       = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd);
-        dev_square      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,    n_embd);
-        dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-        ggml_set_name(dev_input,       "dev_input");
-        ggml_set_name(dev_square,      "dev_square");
-        ggml_set_name(dev_eigenvector, "dev_eigenvector");
-        buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
-        ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input));
-
-        // initialize eigenvector to random normalized vector
-        {
-            std::vector<float> random_vec(ggml_nelements(dev_eigenvector), 0.0);
-            std::default_random_engine generator(static_cast<unsigned int>(std::time(0)));
-            std::uniform_real_distribution<float> distribution(0.0, 1.0);
-            float sum_sqr = 0.0; // for normalizing random_vec
-            for (size_t i = 0; i < random_vec.size(); ++i) {
-                float f = distribution(generator);
-                sum_sqr += f * f;
-                random_vec[i] = f;
-            }
-            // normalize it
-            float random_vec_norm = std::sqrt(sum_sqr);
-            for (size_t i = 0; i < random_vec.size(); ++i) {
-                random_vec[i] /= random_vec_norm;
-            }
-            ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector));
-        }
-    }
-
-    ~pca_model() {
-        ggml_free(ctx);
-        ggml_backend_buffer_free(buffer);
-        ggml_backend_free(backend);
-    }
-};
-
-static struct ggml_cgraph * build_graph_piter(
-        const struct pca_params & params,
-        const pca_model & model,
-        bool calc_square = false) {
-    GGML_ASSERT(params.n_batch > 0);
-    // TODO: buf_size must be able to scale with params.n_batch
-    static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
-    static std::vector<uint8_t> buf(buf_size);
-
-    struct ggml_init_params params0 = {
-        /*.mem_size   =*/ buf_size,
-        /*.mem_buffer =*/ buf.data(),
-        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
-    };
-    // create a temporally context to build the graph
-    struct ggml_context * ctx0 = ggml_init(params0);
-    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
-
-    // turn v_diff_original into square matrix if needed
-    struct ggml_tensor * tmp_square;
-    if (calc_square) {
-        tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input);
-        ggml_set_name(tmp_square, "tmp_square");
-    }
-
-    struct ggml_tensor * b_tensor;
-    struct ggml_tensor * distance;
-    struct ggml_tensor * old_eigen    = model.dev_eigenvector;
-    struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square;
-
-    for (int i = 0; i < params.n_batch; ++i) {
-        // b_tensor = square * eigenvector^T
-        b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen);
-        ggml_set_name(b_tensor, "b_tensor");
-
-        // normalize
-        b_tensor = ggml_div_inplace(ctx0,
-            b_tensor,
-            ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor)))
-        );
-        ggml_format_name(b_tensor, "b_tensor_norm_%d", i);
-
-        // calculate distance(new eigenvector - old eigenvector)
-        // we don't use ggml_sub because it may not be implemented on GPU backend
-        struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1));
-        distance = ggml_sqrt_inplace(ctx0,
-            ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old)));
-        ggml_format_name(distance, "distance_%d", i);
-
-        old_eigen = b_tensor;
-
-        // build operations nodes
-        ggml_build_forward_expand(gf, distance);
-    }
-
-    // delete the temporally context used to build the graph
-    ggml_free(ctx0);
-    return gf;
-}
-
-static ggml_status compute_piter(
-        const struct pca_params & params,
-        const pca_model & model,
-        struct ggml_cgraph * gf,
-        ggml_gallocr_t allocr,
-        struct pca_result & result) {
-    // allocate tensors
-    ggml_gallocr_alloc_graph(allocr, gf);
-
-    if (ggml_backend_is_cpu(model.backend)) {
-        ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
-    }
-
-// TODO: enable GPU support when support for GGML_OP_SQRT is added
-//#ifdef GGML_USE_METAL
-//    if (ggml_backend_is_metal(model.backend)) {
-//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
-//    }
-//#endif
-
-    ggml_status res = ggml_backend_graph_compute(model.backend, gf);
-    if (res == GGML_STATUS_SUCCESS) {
-        auto extract_i = [](std::string prefix, std::string str) -> int {
-            int i = -1;
-            if (str.rfind(prefix, 0) == 0) {
-                sscanf(str.c_str(), (prefix + "%d").c_str(), &i);
-            }
-            return i;
-        };
-        result.calculated_square = NULL;
-        result.eigenvectors.clear();
-        result.distances.clear();
-        result.eigenvectors.resize(params.n_batch);
-        result.distances.resize(params.n_batch);
-        // get output nodes
-        for (int i = 0; i < gf->n_nodes; ++i) {
-            auto node = gf->nodes[i];
-            int iter = -1;
-            // find b_tensor (without copying data from device)
-            if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
-                result.eigenvectors[iter] = node;
-            }
-            // find distances, then copy data from device
-            if ((iter = extract_i("distance_", node->name)) > -1) {
-                float d;
-                ggml_backend_tensor_get(node, &d, 0, sizeof(float));
-                result.distances[iter] = d;
-                // std::cout << node->name << " = " << d << "\n";
-            }
-            // find tmp_square if it exists (without copying data from device)
-            if (std::string(node->name) == "tmp_square") {
-                result.calculated_square = node;
-            }
-        }
-    }
-    return res;
-}
-
-static void power_iteration(
-        const struct pca_params & params,
-        struct ggml_tensor * input, // shape of input: [n_samples, n_embd]
-        struct ggml_tensor * output) {
-    //printf("in power iteration\n");
-    struct pca_model model(input);
-
-    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
-    struct pca_result result;
-    struct ggml_tensor * last_eigenvector = NULL;
-
-    int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations
-    for (int iter = 0; iter < n_iters; ++iter) {
-        bool calc_square = (iter == 0); // only need to calculate square for first iteration
-        struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square);
-        // ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot");
-        compute_piter(params, model, gf, allocr, result);
-
-        for (size_t k = 0; k < result.distances.size(); ++k) {
-            last_eigenvector = result.eigenvectors[k];
-            if (result.distances[k] < params.tolerance) {
-                break; // done
-            }
-        }
-
-        if (calc_square) {
-            // copy and store the square matrix if needed
-            GGML_ASSERT(result.calculated_square != NULL);
-            ggml_backend_tensor_copy(result.calculated_square, model.dev_square);
-        }
-
-        {
-            // copy last eigen vector and store as input for next iteration
-            GGML_ASSERT(last_eigenvector != NULL);
-            ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector);
-        }
-
-        printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
-            __func__, params.i_layer+1, params.n_layers, iter, n_iters, params.n_batch);
-    }
-
-    // get output tensor
-    GGML_ASSERT(last_eigenvector);
-    ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
-    //print_debug_tensor(output);
-    ggml_gallocr_free(allocr);
-}
-
-static void run_pca(
-        struct pca_params & params,
-        const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_samples, n_embd]
-        const std::vector<struct ggml_tensor *> & v_output) {
-    printf("%s: Running PCA...\n", __func__);
-    for (size_t il = 0; il < v_input.size(); ++il) {
-
-        // prepare output vector
-        struct ggml_tensor * ctrl_out = v_output[il];
-        ggml_format_name(ctrl_out, "direction.%ld", il+1);
-
-        // run power_iteration
-        params.i_layer = il;
-        params.n_layers = v_input.size();
-        power_iteration(params, v_input[il], ctrl_out);
-        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
-    }
-}
-
-}

examples/cvector-generator/positive.txt

@@ -1 +0,0 @@
-[INST] Act like a person who is extremely happy. [/INST] 

examples/embedding/CMakeLists.txt

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

examples/embedding/README.md

@@ -9,53 +9,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
-
-## extra parameters
-### --embd-normalize $integer$
-| $integer$ | description         | formula |
-|-----------|---------------------|---------|
-| $-1$      | none                |
-| $0$       | max absolute int16  | $\Large{{32760 * x_i} \over\max \lvert x_i\rvert}$
-| $1$       | taxicab             | $\Large{x_i \over\sum \lvert x_i\rvert}$
-| $2$       | euclidean (default) | $\Large{x_i \over\sqrt{\sum x_i^2}}$
-| $>2$      | p-norm              | $\Large{x_i \over\sqrt[p]{\sum \lvert x_i\rvert^p}}$
-
-### --embd-output-format $'string'$
-| $'string'$ | description                  |  |
-|------------|------------------------------|--|
-| ''         | same as before               | (default)
-| 'array'    | single embeddings            | $[[x_1,...,x_n]]$
-|            | multiple embeddings          | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
-| 'json'     | openai style                 |
-| 'json+'    | add cosine similarity matrix |
-
-### --embd-separator $"string"$
-| $"string"$   | |
-|--------------|-|
-| "\n"         | (default)
-| "<#embSep#>" | for exemple
-| "<#sep#>"    | other exemple
-
-## examples
-### Unix-based systems (Linux, macOS, etc.):
-
-```bash
-./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
-```
-
-### Windows:
-
-```powershell
-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
-```
-

examples/embedding/embedding.cpp

@@ -7,30 +7,23 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static std::vector<std::string> split_lines(const std::string & s, const std::string & separator = "\n") {
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
     std::vector<std::string> lines;
-    size_t start = 0;
-    size_t end = s.find(separator);
-
-    while (end != std::string::npos) {
-        lines.push_back(s.substr(start, end - start));
-        start = end + separator.length();
-        end = s.find(separator, start);
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
     }
-
-    lines.push_back(s.substr(start)); // Add the last part
-
     return lines;
 }
 
-static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
-    size_t n_tokens = tokens.size();
-    for (size_t i = 0; i < n_tokens; i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
@@ -47,10 +40,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 
         // try to get sequence embeddings - supported only when pooling_type is not NONE
         const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        if (embd == NULL) {
+            embd = llama_get_embeddings_ith(ctx, i);
+            if (embd == NULL) {
+                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
+                continue;
+            }
+        }
 
         float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd, embd_norm);
+        //TODO: I would also add a parameter here to enable normalization or not.
+        /*fprintf(stdout, "unnormalized_embedding:");
+        for (int hh = 0; hh < n_embd; hh++) {
+            fprintf(stdout, "%9.6f ", embd[hh]);
+        }
+        fprintf(stdout, "\n");*/
+        llama_embd_normalize(embd, out, n_embd);
     }
 }
 
@@ -92,12 +97,6 @@ int main(int argc, char ** argv) {
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
-    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
-        return 1;
-    }
-
     if (n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, n_ctx);
@@ -110,7 +109,7 @@ int main(int argc, char ** argv) {
     }
 
     // split the prompt into lines
-    std::vector<std::string> prompts = split_lines(params.prompt, params.embd_sep);
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
@@ -170,7 +169,7 @@ int main(int argc, char ** argv) {
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
             float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
+            batch_decode(ctx, batch, out, s, n_embd);
             llama_batch_clear(batch);
             p += s;
             s = 0;
@@ -183,78 +182,29 @@ int main(int argc, char ** argv) {
 
     // final batch
     float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
+    batch_decode(ctx, batch, out, s, n_embd);
 
-    if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
+    // print the first part of the embeddings or for a single prompt, the full embedding
+    fprintf(stdout, "\n");
+    for (int j = 0; j < n_prompts; j++) {
+        fprintf(stdout, "embedding %d: ", j);
+        for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+        }
         fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                }
-            }
-            fprintf(stdout, "\n");
-        }
-
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
-            }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
-                }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
-                fprintf(stdout, "\n");
-            }
-        }
     }
 
-    if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
-        const bool notArray = params.embd_out != "array";
-
-        fprintf(stdout, notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
-        for (int j = 0;;) { // at least one iteration (one prompt)
-            if (notArray) fprintf(stdout, "    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
-            fprintf(stdout, "[");
-            for (int i = 0;;) { // at least one iteration (n_embd > 0)
-                fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
-                i++;
-                if (i < n_embd) fprintf(stdout, ","); else break;
+    // print cosine similarity matrix
+    if (n_prompts > 1) {
+        fprintf(stdout, "\n");
+        printf("cosine similarity matrix:\n\n");
+        for (int i = 0; i < n_prompts; i++) {
+            for (int j = 0; j < n_prompts; j++) {
+                float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                fprintf(stdout, "%6.2f ", sim);
             }
-            fprintf(stdout, notArray ? "]\n    }" : "]");
-            j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            fprintf(stdout, "\n");
         }
-        fprintf(stdout, notArray ? "\n  ]" : "]\n");
-
-        if (params.embd_out == "json+" && n_prompts > 1) {
-            fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
-                fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f", sim);
-                    j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
-                }
-                fprintf(stdout, " ]");
-                i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
-            }
-            fprintf(stdout, "\n  ]");
-        }
-
-        if (notArray) fprintf(stdout, "\n}\n");
     }
 
     // clean up

examples/eval-callback/CMakeLists.txt

@@ -1,9 +1,9 @@
-set(TARGET llama-eval-callback)
+set(TARGET eval-callback)
 add_executable(${TARGET} eval-callback.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
 set(TEST_TARGET test-eval-callback)
-add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
 set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

examples/eval-callback/README.md

@@ -6,7 +6,7 @@ It simply prints to the console all operations and tensor data.
 Usage:
 
 ```shell
-llama-eval-callback \
+eval-callback \
   --hf-repo ggml-org/models \
   --hf-file phi-2/ggml-model-q4_0.gguf \
   --model phi-2-q4_0.gguf \

examples/export-lora/CMakeLists.txt

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

examples/export-lora/README.md

@@ -3,7 +3,7 @@
 Apply LORA adapters to base model and export the resulting model.
 
 ```
-usage: llama-export-lora [options]
+usage: export-lora [options]
 
 options:
   -h, --help                         show this help message and exit
@@ -17,7 +17,7 @@ options:
 For example:
 
 ```bash
-./bin/llama-export-lora \
+./bin/export-lora \
     -m open-llama-3b-v2-q8_0.gguf \
     -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
     -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin

examples/finetune/CMakeLists.txt

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

examples/finetune/README.md

@@ -7,7 +7,7 @@ Basic usage instructions:
 wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
 
 # finetune LORA adapter
-./bin/llama-finetune \
+./bin/finetune \
         --model-base open-llama-3b-v2-q8_0.gguf \
         --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
         --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
@@ -18,7 +18,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
         --use-checkpointing
 
 # predict
-./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
 ```
 
 **Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
@@ -38,14 +38,14 @@ After 10 more iterations:
 Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
 
 llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
-These LORA adapters can then be used by `llama-cli` together with the base model, like in the 'predict' example command above.
+These LORA adapters can then be used by `main` together with the base model, like in the 'predict' example command above.
 
-In `llama-cli` you can also load multiple LORA adapters, which will then be mixed together.
+In `main` you can also load multiple LORA adapters, which will then be mixed together.
 
 For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
 
 ```bash
-./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
   --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
   --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
 ```
@@ -55,7 +55,7 @@ You can change how strong each LORA adapter is applied to the base model by usin
 For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
 
 ```bash
-./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
   --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
   --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
   --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin

examples/finetune/finetune.sh

@@ -2,7 +2,7 @@
 cd `dirname $0`
 cd ../..
 
-EXE="./llama-finetune"
+EXE="./finetune"
 
 if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi
 if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi

examples/gbnf-validator/CMakeLists.txt

@@ -1,5 +1,5 @@
-set(TARGET llama-gbnf-validator)
+set(TARGET gbnf-validator)
 add_executable(${TARGET} gbnf-validator.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common grammar-parser llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gbnf-validator/gbnf-validator.cpp

@@ -7,8 +7,6 @@
 
 #include <cstdio>
 #include <cstdlib>
-#include <sstream>
-#include <fstream>
 #include <string>
 #include <vector>
 
@@ -71,14 +69,13 @@ int main(int argc, char** argv) {
         return 1;
     }
 
-    std::string grammar_str;
-    {
-        std::ifstream grammar_file(grammar_filename);
-        GGML_ASSERT(grammar_file.is_open() && "Failed to open grammar file");
-        std::stringstream buffer;
-        buffer << grammar_file.rdbuf();
-        grammar_str = buffer.str();
-    }
+    fseek(grammar_file, 0, SEEK_END);
+    size_t grammar_size = ftell(grammar_file);
+    fseek(grammar_file, 0, SEEK_SET);
+
+    std::string grammar_str(grammar_size, ' ');
+    fread(&grammar_str[0], 1, grammar_size, grammar_file);
+    fclose(grammar_file);
 
     // Parse the GBNF grammar
     auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
@@ -103,15 +100,20 @@ int main(int argc, char** argv) {
             grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
 
     // Read the input file
-    std::string input_str;
-    {
-        std::ifstream input_file(input_filename);
-        GGML_ASSERT(input_file.is_open() && "Failed to open input file");
-        std::stringstream buffer;
-        buffer << input_file.rdbuf();
-        input_str = buffer.str();
+    FILE* input_file = fopen(input_filename.c_str(), "r");
+    if (!input_file) {
+        fprintf(stdout, "Failed to open input file: %s\n", input_filename.c_str());
+        return 1;
     }
 
+    fseek(input_file, 0, SEEK_END);
+    size_t input_size = ftell(input_file);
+    fseek(input_file, 0, SEEK_SET);
+
+    std::string input_str(input_size, ' ');
+    fread(&input_str[0], 1, input_size, input_file);
+    fclose(input_file);
+
     // Validate the input string against the grammar
     size_t error_pos;
     std::string error_msg;

examples/gguf-split/CMakeLists.txt

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

examples/gguf-split/tests.sh

@@ -18,8 +18,8 @@ fi
 
 set -x
 
-SPLIT=$1/llama-gguf-split
-MAIN=$1/llama-cli
+SPLIT=$1/gguf-split
+MAIN=$1/main
 WORK_PATH=$TMP_DIR/gguf-split
 ROOT_DIR=$(realpath $(dirname $0)/../../)
 

examples/gguf/CMakeLists.txt

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

examples/gritlm/CMakeLists.txt

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

examples/gritlm/README.md

@@ -26,7 +26,7 @@ $ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf --ou
 
 Run the example using the downloaded model:
 ```console
-$ ./llama-gritlm -m models/gritlm-7b_q4_1.gguf
+$ ./gritlm -m models/gritlm-7b_q4_1.gguf
 
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103

examples/gritlm/gritlm.cpp

@@ -44,7 +44,6 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_kv_cache_clear(ctx);
-        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -99,9 +98,7 @@ static std::string generate(llama_context * ctx, const std::string & prompt, boo
     llama_token eos_token = llama_token_eos(mdl);
 
     llama_kv_cache_clear(ctx);
-    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
-
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
     std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
@@ -169,7 +166,8 @@ int main(int argc, char * argv[]) {
 
     llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
 
-    // create generation context
+    // create new context - set to embedding mode
+    cparams.embeddings = true;
     llama_context * ctx = llama_new_context_with_model(mdl, cparams);
 
     // ### Embedding/Representation ###

examples/imatrix/CMakeLists.txt

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

examples/imatrix/README.md

@@ -6,7 +6,7 @@ More information is available here: https://github.com/ggerganov/llama.cpp/pull/
 ## Usage
 
 ```
-./llama-imatrix \
+./imatrix \
     -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \
     [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
     [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]
@@ -28,8 +28,8 @@ For faster computation, make sure to use GPU offloading via the `-ngl` argument
 LLAMA_CUDA=1 make -j
 
 # generate importance matrix (imatrix.dat)
-./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
+./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
 
 # use the imatrix to perform a Q4_K_M quantization
-./llama-quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
+./quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
 ```

examples/infill/CMakeLists.txt

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

examples/infill/README.md

@@ -42,5 +42,5 @@ scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.ggu
 ```
 
 ```bash
-./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
 ```

examples/infill/infill.cpp

@@ -223,11 +223,7 @@ int main(int argc, char ** argv) {
     inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-
-    const llama_token middle_token = llama_token_middle(model);
-    if (middle_token >= 0) {
-        embd_inp.push_back(middle_token);
-    }
+    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -532,12 +528,7 @@ int main(int argc, char ** argv) {
                 inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-
-                const llama_token middle_token = llama_token_middle(model);
-                if (middle_token >= 0) {
-                    embd_inp.push_back(middle_token);
-                }
-
+                embd_inp.push_back(llama_token_middle(model));
                 embd.clear();
                 n_remain = params.n_predict;
                 n_past = 0;

examples/jeopardy/jeopardy.sh

@@ -21,7 +21,7 @@ counter=1
 echo 'Running'
 while IFS= read -r question
 do
-  exe_cmd="./llama-cli -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
+  exe_cmd="./main -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
   echo $counter
   echo "Current Question: $question"
   eval "$exe_cmd"

examples/json-schema-pydantic-example.py

@@ -1,5 +1,5 @@
 # Usage:
-#! ./llama-server -m some-model.gguf &
+#! ./server -m some-model.gguf &
 #! pip install pydantic
 #! python json-schema-pydantic-example.py
 

examples/json_schema_to_grammar.py

@@ -29,8 +29,9 @@ class BuiltinRule:
         self.content = content
         self.deps = deps or []
 
-# Constraining spaces to prevent model "running away".
-SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}'
+# whitespace is constrained to a single space char to prevent model "running away" in
+# whitespace. Also maybe improves generation quality?
+SPACE_RULE = '" "?'
 
 PRIMITIVE_RULES = {
     'boolean'      : BuiltinRule('("true" | "false") space', []),
@@ -42,7 +43,7 @@ PRIMITIVE_RULES = {
     'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
     'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
     'uuid'         : BuiltinRule(r'"\"" [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'         : BuiltinRule(r'[^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})', []),
+    'char'         : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F]{4})', []),
     'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
     'null'         : BuiltinRule('"null" space', []),
 }
@@ -523,7 +524,7 @@ class SchemaConverter:
 def main(args_in = None):
     parser = argparse.ArgumentParser(
         description='''
-            Generates a grammar (suitable for use in ./llama-cli) that produces JSON conforming to a
+            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
             given JSON schema. Only a subset of JSON schema features are supported; more may be
             added in the future.
         ''',

examples/llama-bench/README.md

@@ -1,4 +1,4 @@
-# llama.cpp/examples/llama-bench
+# llama.cpp/example/llama-bench
 
 Performance testing tool for llama.cpp.