Upload folder using huggingface_hub

.devops/cloud-v-pipeline

@@ -0,0 +1,22 @@
+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    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 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}

.devops/full-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip
+    apt-get install -y build-essential python3 python3-pip git
 
 COPY requirements.txt requirements.txt
 

.devops/full-rocm.Dockerfile

@@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

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

@@ -0,0 +1,84 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-clblast
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        OpenCL Inference of LLaMA model in C/C++
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git mesa-libOpenCL-devel clblast-devel
+Requires:       clblast
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CLBLAST=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+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
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamaclblast
+%{_bindir}/llamaclblastserver
+%{_bindir}/llamaclblastsimple
+/usr/lib/systemd/system/llamaclblast.service
+%config /etc/sysconfig/llama
+
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

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

@@ -0,0 +1,83 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-cublas
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git cuda-toolkit
+Requires:       cuda-toolkit
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CUBLAS=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacppcublas
+cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacublas.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppcublas
+%{_bindir}/llamacppcublasserver
+%{_bindir}/llamacppcublassimple
+/usr/lib/systemd/system/llamacublas.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.devops/llama-cpp.srpm.spec

@@ -0,0 +1,85 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+#    In the meantime, YYYYMMDD format will be used.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git libstdc++-devel
+Requires:       libstdc++
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+Models are not included in this package and must be downloaded separately.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+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
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llama
+%{_bindir}/llamaserver
+%{_bindir}/llamasimple
+/usr/lib/systemd/system/llama.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.devops/main-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential
+    apt-get install -y build-essential git
 
 WORKDIR /app
 

.devops/main-rocm.Dockerfile

@@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/main" ]

.editorconfig

@@ -17,3 +17,6 @@ indent_style = tab
 
 [prompts/*.txt]
 insert_final_newline = unset
+
+[examples/server/public/*]
+indent_size = 2

.github/workflows/code-coverage.yml

@@ -0,0 +1,36 @@
+name: Code Coverage
+on: [push, pull_request]
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+
+jobs:
+  run:
+    runs-on: ubuntu-20.04
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - 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/gguf-publish.yml

@@ -0,0 +1,43 @@
+# This workflow will upload a Python Package using Twine when a GGUF release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# See `gguf-py/README.md` for how to make a release.
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  workflow_dispatch:
+  push:
+    # Pattern matched against refs/tags
+    tags:
+      - 'gguf-v*'           # Push events to every version tag
+
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v3
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.9.x'
+    - name: Install dependencies
+      run: |
+        cd gguf-py
+        python -m pip install poetry
+        poetry install
+
+    - name: Build package
+      run: poetry build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        password: ${{ secrets.PYPI_API_TOKEN }}

.gitignore

@@ -1,6 +1,6 @@
 *.o
 *.a
-*.so
+*.bin
 .DS_Store
 .build/
 .cache/
@@ -12,20 +12,7 @@
 .vs/
 .vscode/
 
-build/
-build-em/
-build-debug/
-build-release/
-build-ci-debug/
-build-ci-release/
-build-static/
-build-cublas/
-build-opencl/
-build-metal/
-build-mpi/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build*/
 out/
 tmp/
 
@@ -39,19 +26,24 @@ models-mnt
 /perplexity
 /embedding
 /train-text-from-scratch
+/convert-llama2c-to-ggml
 /simple
 /benchmark-matmult
 /vdot
 /server
 /Pipfile
 /embd-input-test
+/gguf
+/gguf-llama-simple
 /libllama.so
-
+/llama-bench
+build-info.h
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
 
 __pycache__
+dist
 
 dist/
 *.spec
@@ -65,11 +57,11 @@ perf-*.txt
 
 examples/jeopardy/results.txt
 
-pyproject.toml
 poetry.lock
 poetry.toml
 
 # Test binaries
+tests/test-grammar-parser
 tests/test-double-float
 tests/test-grad0
 tests/test-opt
@@ -78,16 +70,21 @@ tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0
 
-koboldcpp.so
-koboldcpp_failsafe.so
-koboldcpp_openblas.so
-koboldcpp_noavx2.so
-koboldcpp_clblast.so
-koboldcpp.dll
-koboldcpp_failsafe.dll
-koboldcpp_openblas.dll
-koboldcpp_noavx2.dll
-koboldcpp_clblast.dll
-koboldcpp_cublas.dll
-cublas64_11.dll
-cublasLt64_11.dll
+/koboldcpp_default.so
+/koboldcpp_failsafe.so
+/koboldcpp_openblas.so
+/koboldcpp_noavx2.so
+/koboldcpp_clblast.so
+/koboldcpp_cublas.so
+/koboldcpp_default.dll
+/koboldcpp_failsafe.dll
+/koboldcpp_openblas.dll
+/koboldcpp_noavx2.dll
+/koboldcpp_clblast.dll
+/koboldcpp_cublas.dll
+/cublas64_11.dll
+/cublasLt64_11.dll
+/rocblas/
+rocblas.dll
+hipblas.dll
+koboldcpp_hipblas.so

CMakeLists.txt

@@ -50,6 +50,9 @@ set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA
 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 dmmv CUDA kernels"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                             "llama: max. batch size for using peer access")
+option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 
 
@@ -65,6 +68,11 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 
 add_compile_definitions(GGML_USE_K_QUANTS)
+add_compile_definitions(LOG_DISABLE_LOGS)
+
+set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+set(GGML_V2_CUDA_SOURCES otherarch/ggml_v2-cuda.cu otherarch/ggml_v2-cuda.h)
+set(GGML_V2_LEGACY_CUDA_SOURCES otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h)
 
 if (LLAMA_CUBLAS)
     cmake_minimum_required(VERSION 3.17)
@@ -75,10 +83,6 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
-        set(GGML_V2_CUDA_SOURCES otherarch/ggml_v2-cuda.cu otherarch/ggml_v2-cuda.h)
-        set(GGML_V2_LEGACY_CUDA_SOURCES otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h)
-
         add_compile_definitions(GGML_USE_CUBLAS)
         #add_compile_definitions(GGML_CUDA_CUBLAS) #remove to not use cublas
         add_compile_definitions(GGML_CUDA_MMQ_Y=${LLAMA_CUDA_MMQ_Y})
@@ -91,6 +95,7 @@ if (LLAMA_CUBLAS)
             add_compile_definitions(GGML_CUDA_F16)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -121,6 +126,75 @@ if (LLAMA_CUBLAS)
     endif()
 endif()
 
+if (LLAMA_HIPBLAS)
+    if (MSVC)
+        list(APPEND CMAKE_PREFIX_PATH "C:/Program Files/AMD/ROCm/5.5")
+    else()
+        list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    endif()
+
+
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ${GGML_SOURCES_CUDA})
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PUBLIC GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-rocm PUBLIC GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PUBLIC GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PUBLIC K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        target_compile_definitions(ggml-rocm PUBLIC CC_TURING=1000000000)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PUBLIC hip::device hip::host roc::rocblas roc::hipblas)
+
+
+        add_library(ggml-v2-rocm OBJECT ${GGML_V2_CUDA_SOURCES})
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-v2-rocm PUBLIC GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-v2-rocm PUBLIC GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-v2-rocm PUBLIC GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-v2-rocm PUBLIC K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        target_compile_definitions(ggml-v2-rocm PUBLIC CC_TURING=1000000000)
+        set_source_files_properties(otherarch/ggml_v2-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-v2-rocm PUBLIC hip::device hip::host roc::rocblas roc::hipblas)
+
+
+        add_library(ggml-v2-legacy-rocm OBJECT ${GGML_V2_LEGACY_CUDA_SOURCES})
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-v2-legacy-rocm PUBLIC GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-v2-legacy-rocm PUBLIC GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-v2-legacy-rocm PUBLIC GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-v2-legacy-rocm PUBLIC K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        target_compile_definitions(ggml-v2-legacy-rocm PUBLIC CC_TURING=1000000000)
+        set_source_files_properties(otherarch/ggml_v2-cuda-legacy.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-v2-legacy-rocm PUBLIC hip::device hip::host roc::rocblas roc::hipblas)
+
+
+
+
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm ggml-v2-rocm ggml-v2-legacy-rocm)
+    else()
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    endif()
+endif()
+
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
         set(c_flags
@@ -133,15 +207,22 @@ if (LLAMA_ALL_WARNINGS)
             -Wstrict-prototypes
             -Wpointer-arith
             -Wmissing-prototypes
+            -Werror=implicit-int
+            -Wno-unused-function
         )
         set(cxx_flags
             -Wall
             -Wextra
             -Wpedantic
             -Wcast-qual
+            -Wmissing-declarations
             -Wno-unused-function
             -Wno-multichar
         )
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+            # g++ only
+            set(cxx_flags ${cxx_flags} -Wno-format-truncation -Wno-array-bounds)
+        endif()
     else()
         # todo : msvc
     endif()
@@ -153,7 +234,7 @@ if (LLAMA_ALL_WARNINGS)
 
 endif()
 
-if (MSVC)
+if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
     if (BUILD_SHARED_LIBS)
@@ -190,7 +271,7 @@ if (NOT MSVC)
     endif()
 endif()
 
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64"))
     message(STATUS "ARM detected")
     if (MSVC)
         # TODO: arm msvc?
@@ -301,37 +382,52 @@ target_link_libraries(ggml_v2 PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 set_target_properties(ggml_v2 PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
 add_library(common2
-            examples/common.cpp
-            examples/common.h)
-target_include_directories(common2 PUBLIC . ./otherarch ./otherarch/tools ./examples)
+            common/common.cpp
+            common/common.h
+            common/grammar-parser.h
+            common/grammar-parser.cpp)
+target_include_directories(common2 PUBLIC . ./otherarch ./otherarch/tools ./examples ./common)
 target_compile_features(common2 PUBLIC cxx_std_11) # don't bump
 target_link_libraries(common2 PRIVATE ggml ${LLAMA_EXTRA_LIBS})
 set_target_properties(common2 PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
 add_library(gpttype_adapter
             gpttype_adapter.cpp)
-target_include_directories(gpttype_adapter PUBLIC . ./otherarch ./otherarch/tools ./examples)
+target_include_directories(gpttype_adapter PUBLIC . ./otherarch ./otherarch/tools ./examples ./common)
 target_compile_features(gpttype_adapter PUBLIC cxx_std_11) # don't bump
 target_link_libraries(gpttype_adapter PRIVATE common2 ggml ${LLAMA_EXTRA_LIBS})
 set_target_properties(gpttype_adapter PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
+if (LLAMA_CUBLAS)
+    set(TARGET koboldcpp_cublas)
+    add_library(${TARGET} SHARED expose.cpp expose.h)
+    target_include_directories(${TARGET} PUBLIC . ./otherarch ./otherarch/tools ./examples ./common)
+    target_compile_features(${TARGET} PUBLIC cxx_std_11) # don't bump
+    set_target_properties(${TARGET} PROPERTIES PREFIX "")
+    set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME "koboldcpp_cublas")
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${TARGET} PUBLIC Threads::Threads ggml ggml_v1 ggml_v2 common2 gpttype_adapter ${LLAMA_EXTRA_LIBS})
+    target_compile_features(${TARGET} PRIVATE cxx_std_11)
+endif()
 
-set(TARGET koboldcpp_cublas)
-add_library(${TARGET} SHARED expose.cpp expose.h)
-target_include_directories(${TARGET} PUBLIC . ./otherarch ./otherarch/tools ./examples)
-target_compile_features(${TARGET} PUBLIC cxx_std_11) # don't bump
-set_target_properties(${TARGET} PROPERTIES PREFIX "")
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME "koboldcpp_cublas")
-set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-target_link_libraries(${TARGET} PUBLIC ggml ggml_v1 ggml_v2 common2 gpttype_adapter ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if (LLAMA_HIPBLAS)
+    set(TARGET koboldcpp_hipblas)
+    add_library(${TARGET} SHARED expose.cpp expose.h)
+    target_include_directories(${TARGET} PUBLIC . ./otherarch ./otherarch/tools ./examples ./common)
+    target_compile_features(${TARGET} PUBLIC cxx_std_11) # don't bump
+    set_target_properties(${TARGET} PROPERTIES PREFIX "")
+    set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME "koboldcpp_hipblas")
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${TARGET} PUBLIC Threads::Threads ggml ggml_v1 ggml_v2 common2 gpttype_adapter ${LLAMA_EXTRA_LIBS})
+    target_compile_features(${TARGET} PRIVATE cxx_std_11)
+endif()
 
 
 if (MAKE_MISC_FILES)
+add_subdirectory(common)
 add_library(llama
         llama.cpp
         llama.h
-        llama-util.h
         )
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump

Dockerfile

@@ -5,6 +5,7 @@ RUN apt update \
  && apt install build-essential wget libopenblas-dev make -y \
  && make LLAMA_OPENBLAS=1 \
  && wget https://huggingface.co/notstoic/pygmalion-13b-ggml/resolve/main/pygmalion-13b-ggml-q4_0.bin \
- && apt remove build-essential wget make -y
+ && apt remove build-essential wget make -y \
+ && rm -fr *.bat convert-* ci docs examples otherarchs tests 
 
  ENTRYPOINT ["python", "koboldcpp.py", "pygmalion-13b-ggml-q4_0.bin", "--port", "7860"]

MIT_LICENSE_GGML_LLAMACPP_ONLY

@@ -23,4 +23,4 @@ SOFTWARE.
 ===================================
 
 Note that the above license applies ONLY to the GGML library and llama.cpp by ggerganov which are licensed under the MIT License
-Kobold Lite by Concedo and the provided python ctypes bindings in koboldcpp.dll are licensed under the AGPL v3.0 License
+Kobold Lite by Concedo and the provided python ctypes bindings in koboldcpp dlls are licensed under the AGPL v3.0 License

Makefile

@@ -1,4 +1,4 @@
-default: koboldcpp koboldcpp_failsafe koboldcpp_openblas koboldcpp_noavx2 koboldcpp_clblast koboldcpp_cublas
+default: koboldcpp_default koboldcpp_failsafe koboldcpp_openblas koboldcpp_noavx2 koboldcpp_clblast koboldcpp_cublas koboldcpp_hipblas
 tools: quantize_gpt2 quantize_gptj quantize_llama quantize_neox quantize_mpt
 dev: koboldcpp_openblas
 dev2: koboldcpp_clblast
@@ -20,8 +20,6 @@ ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/nul
 ARCH_ADD = -lcblas
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
-CXXV := $(shell $(CXX) --version | head -n 1)
 
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
@@ -41,8 +39,8 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11   -fPIC -DGGML_USE_K_QUANTS
-CXXFLAGS = -I. -I./examples -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS
+CFLAGS   = -I.              -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11   -fPIC -DGGML_USE_K_QUANTS -DLOG_DISABLE_LOGS -D_GNU_SOURCE
+CXXFLAGS = -I. -I./common -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS -DLOG_DISABLE_LOGS -D_GNU_SOURCE
 LDFLAGS  =
 
 # these are used on windows, to build some libraries with extra old device compatibility
@@ -110,7 +108,8 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 # old library NEEDS mf16c to work. so we must build with it. new one doesnt
 	ifeq ($(OS),Windows_NT)
 		CFLAGS +=
-		NONECFLAGS += -mno-sse3
+		NONECFLAGS +=
+# -mno-sse3
 		SIMPLECFLAGS += -mavx -msse3
 		FULLCFLAGS += -mavx2 -msse3 -mfma -mf16c -mavx
 	else
@@ -195,6 +194,42 @@ ggml_v2-cuda-legacy.o: otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-l
 	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
 
+ifdef LLAMA_HIPBLAS
+	ROCM_PATH	?= /opt/rocm
+	CC         := $(ROCM_PATH)/llvm/bin/clang
+	CXX        := $(ROCM_PATH)/llvm/bin/clang++
+	GPU_TARGETS ?= gfx803 gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	LLAMA_CUDA_DMMV_X ?= 32
+	LLAMA_CUDA_MMV_Y ?= 2
+	LLAMA_CUDA_KQUANTS_ITER ?= 2
+	HIPFLAGS   += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS $(shell $(ROCM_PATH)/bin/hipconfig -C)
+ifdef LLAMA_CUDA_FORCE_DMMV
+	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
+endif # LLAMA_CUDA_FORCE_DMMV
+	HIPLDFLAGS    += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib -lhipblas -lamdhip64 -lrocblas
+	HIP_OBJS       += ggml-cuda.o ggml_v2-cuda.o ggml_v2-cuda-legacy.o
+ggml-cuda.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
+						-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
+                        -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
+                        -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+ggml_v2-cuda.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
+						-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
+                        -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
+                        -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+ggml_v2-cuda-legacy.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
+						-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
+                        -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
+                        -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(CXX) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+ggml_v2-cuda.o: otherarch/ggml_v2-cuda.cu otherarch/ggml_v2-cuda.h
+	$(CXX) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+ggml_v2-cuda-legacy.o: otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h
+	$(CXX) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # LLAMA_HIPBLAS
+
+
+
 ifdef LLAMA_METAL
 	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
 	CXXFLAGS += -DGGML_USE_METAL
@@ -224,12 +259,16 @@ ifneq ($(filter armv8%,$(UNAME_M)),)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
+CCV := $(shell $(CC) --version | head -n 1)
+CXXV := $(shell $(CXX) --version | head -n 1)
+
 DEFAULT_BUILD =
 FAILSAFE_BUILD =
 OPENBLAS_BUILD =
 NOAVX2_BUILD =
 CLBLAST_BUILD =
 CUBLAS_BUILD =
+HIPBLAS_BUILD =
 
 ifeq ($(OS),Windows_NT)
 	DEFAULT_BUILD = $(CXX) $(CXXFLAGS)  $^ -shared -o $@.dll $(LDFLAGS)
@@ -238,10 +277,12 @@ ifeq ($(OS),Windows_NT)
 	NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o $@.dll $(LDFLAGS)
 	CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ lib/OpenCL.lib lib/clblast.lib -shared -o $@.dll $(LDFLAGS)
 
-ifdef LLAMA_CUBLAS
-	CUBLAS_BUILD = $(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $^ -shared -o $@.dll $(CUBLASLD_FLAGS) $(LDFLAGS)
-endif
-
+	ifdef LLAMA_CUBLAS
+		CUBLAS_BUILD = $(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $^ -shared -o $@.dll $(CUBLASLD_FLAGS) $(LDFLAGS)
+	endif
+	ifdef LLAMA_HIPBLAS
+		HIPBLAS_BUILD = $(CXX) $(CXXFLAGS) $(HIPFLAGS) $^ -shared -o $@.dll $(HIPLDFLAGS) $(LDFLAGS)
+	endif
 else
 	DEFAULT_BUILD = $(CXX) $(CXXFLAGS)  $^ -shared -o $@.so $(LDFLAGS)
 	FAILSAFE_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o $@.so $(LDFLAGS)
@@ -250,24 +291,29 @@ else
 	NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
 	endif
 	ifdef LLAMA_CLBLAST
-        ifeq ($(UNAME_S),Darwin)
-                CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -framework OpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
-        else
-                CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -lOpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
-        endif
+		ifeq ($(UNAME_S),Darwin)
+				CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -framework OpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
+		else
+				CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -lOpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
+		endif
 	endif
 
-ifdef LLAMA_CUBLAS
-	CUBLAS_BUILD = $(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $^ -shared -o $@.so $(CUBLASLD_FLAGS) $(LDFLAGS)
-endif
+	ifdef LLAMA_CUBLAS
+		CUBLAS_BUILD = $(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $^ -shared -o $@.so $(CUBLASLD_FLAGS) $(LDFLAGS)
+	endif
+	ifdef LLAMA_HIPBLAS
+		HIPBLAS_BUILD = $(CXX) $(CXXFLAGS) $(HIPFLAGS) $^ -shared -o $@.so $(HIPLDFLAGS) $(LDFLAGS)
+	endif
 
 	ifndef LLAMA_OPENBLAS
 	ifndef LLAMA_CLBLAST
 	ifndef LLAMA_CUBLAS
+	ifndef LLAMA_HIPBLAS
 	OPENBLAS_BUILD = @echo 'Your OS $(OS) does not appear to be Windows. For faster speeds, install and link a BLAS library. Set LLAMA_OPENBLAS=1 to compile with OpenBLAS support or LLAMA_CLBLAST=1 to compile with ClBlast support. This is just a reminder, not an error.'
 	endif
 	endif
 	endif
+	endif
 endif
 
 
@@ -293,16 +339,16 @@ $(info )
 
 ggml.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) -c $< -o $@
-ggml_openblas.o: ggml.c ggml.h
+ggml_openblas.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(OPENBLAS_FLAGS) -c $< -o $@
-ggml_failsafe.o: ggml.c ggml.h
+ggml_failsafe.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(NONECFLAGS) -c $< -o $@
-ggml_noavx2.o: ggml.c ggml.h
+ggml_noavx2.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) -c $< -o $@
-ggml_clblast.o: ggml.c ggml.h
+ggml_clblast.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
-ggml_cublas.o: ggml.c ggml.h
-	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CUBLAS_FLAGS) -c $< -o $@
+ggml_cublas.o: ggml.c ggml.h ggml-cuda.h k_quants.h
+	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CUBLAS_FLAGS) $(HIPFLAGS) -c $< -o $@
 
 #quants K
 k_quants.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
@@ -328,7 +374,7 @@ ggml_v2_noavx2.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 ggml_v2_clblast.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
 ggml_v2_cublas.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
-	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CUBLAS_FLAGS) -c $< -o $@
+	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CUBLAS_FLAGS) $(HIPFLAGS) -c $< -o $@
 
 #extreme old version compat
 ggml_v1.o: otherarch/ggml_v1.c otherarch/ggml_v1.h
@@ -345,19 +391,19 @@ ggml_v2-opencl-legacy.o: otherarch/ggml_v2-opencl-legacy.c otherarch/ggml_v2-ope
 	$(CC) $(CFLAGS) -c $< -o $@
 
 # intermediate objects
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h otherarch/llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
-common.o: examples/common.cpp examples/common.h
+common.o: common/common.cpp common/common.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
-console.o: examples/console.cpp examples/console.h
+console.o: common/console.cpp common/console.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
-grammar-parser.o: examples/grammar-parser.cpp examples/grammar-parser.h
+grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 expose.o: expose.cpp expose.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 # idiotic "for easier compilation"
-GPTTYPE_ADAPTER = gpttype_adapter.cpp otherarch/llama_v2.cpp llama.cpp otherarch/utils.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp ggml.h ggml-cuda.h llama.h llama-util.h
+GPTTYPE_ADAPTER = gpttype_adapter.cpp otherarch/llama_v2.cpp otherarch/llama_v3.cpp llama.cpp otherarch/utils.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp ggml.h ggml-cuda.h llama.h otherarch/llama-util.h
 gpttype_adapter_failsafe.o: $(GPTTYPE_ADAPTER)
 	$(CXX) $(CXXFLAGS) $(FAILSAFE_FLAGS) -c $< -o $@
 gpttype_adapter.o: $(GPTTYPE_ADAPTER)
@@ -365,10 +411,10 @@ gpttype_adapter.o: $(GPTTYPE_ADAPTER)
 gpttype_adapter_clblast.o: $(GPTTYPE_ADAPTER)
 	$(CXX) $(CXXFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
 gpttype_adapter_cublas.o: $(GPTTYPE_ADAPTER)
-	$(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) -c $< -o $@
+	$(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $(HIPFLAGS) -c $< -o $@
 
 clean:
-	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so
+	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state gguf gguf.exe main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp_default.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp_hipblas.dll koboldcpp_default.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so koboldcpp_hipblas.so
 
 main: examples/main/main.cpp build-info.h ggml.o k_quants.o ggml-alloc.o llama.o common.o console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -376,19 +422,24 @@ main: examples/main/main.cpp build-info.h ggml.o k_quants.o ggml-alloc.o llama.o
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+gguf: examples/gguf/gguf.cpp build-info.h ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 #generated libraries
-koboldcpp: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o $(OBJS)
+koboldcpp_default: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
 	$(DEFAULT_BUILD)
-koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o $(OBJS)
+koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
 	$(OPENBLAS_BUILD)
-koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_failsafe.o ggml-alloc.o $(OBJS)
+koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_failsafe.o ggml-alloc.o grammar-parser.o $(OBJS)
 	$(FAILSAFE_BUILD)
-koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_noavx2.o ggml-alloc.o $(OBJS)
+koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_noavx2.o ggml-alloc.o grammar-parser.o $(OBJS)
 	$(NOAVX2_BUILD)
-koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o k_quants.o ggml-alloc.o $(OBJS)
+koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
 	$(CLBLAST_BUILD)
-koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o $(CUBLAS_OBJS) $(OBJS)
+koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o grammar-parser.o $(CUBLAS_OBJS) $(OBJS)
 	$(CUBLAS_BUILD)
+koboldcpp_hipblas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o grammar-parser.o $(HIP_OBJS) $(OBJS)
+	$(HIPBLAS_BUILD)
 
 quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o k_quants.o ggml-alloc.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)

Package.swift

@@ -2,8 +2,30 @@
 
 import PackageDescription
 
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v11),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_SWIFT"),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
     name: "llama",
+    platforms: platforms,
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -11,14 +33,23 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: ["ggml-metal.metal"],
-            sources: ["ggml.c", "llama.cpp"],
+            exclude: exclude,
+            sources: [
+                "ggml.c",
+                "llama.cpp",
+                "ggml-alloc.c",
+                "k_quants.c",
+            ] + additionalSources,
             publicHeadersPath: "spm-headers",
-            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32"]),
+                .define("GGML_USE_K_QUANTS"),
+                .define("GGML_USE_ACCELERATE")
+            ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]
-        ),
+        )
     ],
     cxxLanguageStandard: .cxx11
 )

README.md

@@ -3,4 +3,4 @@ sdk: docker
 emoji: 🚀
 colorFrom: yellow
 colorTo: blue
----
+---

build-info.h

@@ -3,5 +3,7 @@
 
 #define BUILD_NUMBER 999
 #define BUILD_COMMIT "KOBOLDCPP"
+#define BUILD_COMPILER "KCPP"
+#define BUILD_TARGET "KCPP"
 
 #endif // BUILD_INFO_H

ci/run.sh

@@ -159,17 +159,17 @@ function gg_run_open_llama_3b_v2 {
 
     python3 ../convert.py ${path_models}
 
-    model_f16="${path_models}/ggml-model-f16.bin"
-    model_q8_0="${path_models}/ggml-model-q8_0.bin"
-    model_q4_0="${path_models}/ggml-model-q4_0.bin"
-    model_q4_1="${path_models}/ggml-model-q4_1.bin"
-    model_q5_0="${path_models}/ggml-model-q5_0.bin"
-    model_q5_1="${path_models}/ggml-model-q5_1.bin"
-    model_q2_k="${path_models}/ggml-model-q2_k.bin"
-    model_q3_k="${path_models}/ggml-model-q3_k.bin"
-    model_q4_k="${path_models}/ggml-model-q4_k.bin"
-    model_q5_k="${path_models}/ggml-model-q5_k.bin"
-    model_q6_k="${path_models}/ggml-model-q6_k.bin"
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
@@ -196,17 +196,17 @@ function gg_run_open_llama_3b_v2 {
     (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/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 3 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     function check_ppl {
         qnt="$1"
@@ -233,6 +233,48 @@ function gg_run_open_llama_3b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/3B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+
     set +e
 }
 
@@ -242,6 +284,7 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf 'OpenLLaMA 3B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -253,6 +296,11 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # open_llama_7b_v2
@@ -285,17 +333,17 @@ function gg_run_open_llama_7b_v2 {
 
     python3 ../convert.py ${path_models}
 
-    model_f16="${path_models}/ggml-model-f16.bin"
-    model_q8_0="${path_models}/ggml-model-q8_0.bin"
-    model_q4_0="${path_models}/ggml-model-q4_0.bin"
-    model_q4_1="${path_models}/ggml-model-q4_1.bin"
-    model_q5_0="${path_models}/ggml-model-q5_0.bin"
-    model_q5_1="${path_models}/ggml-model-q5_1.bin"
-    model_q2_k="${path_models}/ggml-model-q2_k.bin"
-    model_q3_k="${path_models}/ggml-model-q3_k.bin"
-    model_q4_k="${path_models}/ggml-model-q4_k.bin"
-    model_q5_k="${path_models}/ggml-model-q5_k.bin"
-    model_q6_k="${path_models}/ggml-model-q6_k.bin"
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
 
     wiki_test="${path_wiki}/wiki.test.raw"
 
@@ -310,17 +358,17 @@ function gg_run_open_llama_7b_v2 {
     ./bin/quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/main --model ${model_f16}  -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} -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} -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} -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} -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} -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} -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} -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} -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} -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} -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/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
@@ -359,6 +407,48 @@ function gg_run_open_llama_7b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/7B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # currently not supported by the CUDA backend
+    # q8_0
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
     set +e
 }
 
@@ -368,6 +458,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -379,6 +470,11 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 ## main
@@ -391,6 +487,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
     python3 -m pip install -r ${SRC}/requirements.txt
+    python3 -m pip install --editable gguf-py
 fi
 
 ret=0

class.py

@@ -0,0 +1,313 @@
+## KoboldCpp based GGML Backend by Concedo
+## For use as a custom backend in KoboldAI United
+## Not intended for general use.
+
+from __future__ import annotations
+
+import time, json
+import torch
+import requests
+import numpy as np
+from typing import List, Optional, Union
+import os
+from . import koboldcpp
+
+import utils
+from logger import logger
+from modeling.inference_model import (
+    GenerationResult,
+    GenerationSettings,
+    InferenceModel,
+)
+
+model_backend_name = "koboldcpp" #specific instead of ggml
+model_backend_type = "ggml" #This should be a generic name in case multiple model backends are compatible (think Hugging Face Custom and Basic Hugging Face)
+
+kcpp_backend_loaded = False
+
+class KoboldCppException(Exception):
+    """To be used for errors on cpp side of KoboldCpp."""
+
+class KcppArgsObject:
+    def __init__(self, **kwargs):
+        self.__dict__.update(kwargs)
+
+class model_backend(InferenceModel):
+    def __init__(self) -> None:
+        super().__init__()
+
+    def is_valid(self, model_name, model_path, menu_path):
+
+        foundfile = False
+        try:
+            files = os.listdir(model_path)
+            foundfile = len([filename for filename in files if (("ggml" in filename.lower() and ".bin" in filename.lower()) or ".gguf" in filename.lower())])>0
+        except:
+            pass
+        return foundfile
+
+    def get_requested_parameters(self, model_name, model_path, menu_path, parameters = {}):
+
+        self.kcpp_threads = 5
+        self.model_name = "GGML_Model"
+        self.kcpp_ctxsize = 2048
+        self.kcpp_blasbatchsize = 512
+        self.kcpp_gpulayers = 0
+        self.kcpp_smartcontext = False
+        self.kcpp_ropescale = 0.0
+        self.kcpp_ropebase = 10000.0
+        self.kcpp_useclblast = None
+        self.kcpp_usecublas = None
+        self.kcpp_noblas = False
+        self.kcpp_noavx2 = False
+        self.kcpp_nommap = False
+        self.kcpp_debugmode = 0
+        self.kcpp_tensor_split_str = ""
+        self.kcpp_tensor_split = None
+
+        files = os.listdir(model_path)
+        foundfiles = [filename for filename in files if (("ggml" in filename.lower() and ".bin" in filename.lower()) or ".gguf" in filename.lower())]
+
+        requested_parameters = []
+        foldermdls = []
+        for ff in foundfiles:
+            foldermdls.append({'text': ff, 'value': os.path.join(model_path, ff)})
+        requested_parameters.append({
+                                    "uitype": "dropdown",
+                                    "unit": "string",
+                                    "label": "GGML DataFile Name",
+                                    "id": "kcpp_filename",
+                                    "default": os.path.join(model_path, foundfiles[0]) if len(foundfiles)>0 else model_name,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Actual GGML DataFile Name",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": "",
+                                    'children': foldermdls
+                                    })
+        requested_parameters.append({
+                                    "uitype": "dropdown",
+                                    "unit": "int",
+                                    "label": "KoboldCpp Accelerator",
+                                    "id": "kcpp_accelerator",
+                                    "default": 0,
+                                    "check": {"value": "", 'check': "!="},
+                                    'multiple': False,
+                                    "tooltip": "KoboldCpp Accelerator",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": "",
+                                    'children': [{'text': 'Use No BLAS', 'value': 0}, {'text': 'Use OpenBLAS', 'value': 1}, {'text': 'Use CuBLAS', 'value': 2},
+                                    {'text': 'Use CLBLast GPU #1', 'value': 3},{'text': 'Use CLBLast GPU #2', 'value': 4},{'text': 'Use CLBLast GPU #3', 'value': 5}
+                                    ,{'text': 'NoAVX2 Mode (Old CPU)', 'value': 6},{'text': 'Failsafe Mode (Old CPU)', 'value': 7}],
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "Threads",
+                                    "id": "kcpp_threads",
+                                    "default": self.kcpp_threads,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Thread Count",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "Max Context Size",
+                                    "id": "kcpp_ctxsize",
+                                    "default": self.kcpp_ctxsize,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Max Context Size",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "BLAS Batch Size",
+                                    "id": "kcpp_blasbatchsize",
+                                    "default": self.kcpp_blasbatchsize,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "BLAS Batch Size",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "GPU Layers",
+                                    "id": "kcpp_gpulayers",
+                                    "default": self.kcpp_gpulayers,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "GPU Layers",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "Rope Scale",
+                                    "id": "kcpp_ropescale",
+                                    "default": self.kcpp_ropescale,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Rope Scale",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "int",
+                                    "label": "Rope Base",
+                                    "id": "kcpp_ropebase",
+                                    "default": self.kcpp_ropebase,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Rope Base",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        requested_parameters.append({
+                                    "uitype": "dropdown",
+                                    "unit": "int",
+                                    "label": "Smart Context",
+                                    "id": "kcpp_smartcontext",
+                                    "default": self.kcpp_smartcontext,
+                                    "check": {"value": "", 'check': "!="},
+                                    'multiple': False,
+                                    "tooltip": "Smart Context",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": "",
+                                    'children': [{'text': 'False', 'value': False}, {'text': 'True', 'value': True}],
+                                    })
+        requested_parameters.append({
+                                    "uitype": "dropdown",
+                                    "unit": "int",
+                                    "label": "Debug Mode",
+                                    "id": "kcpp_debugmode",
+                                    "default": self.kcpp_debugmode,
+                                    "check": {"value": "", 'check': "!="},
+                                    'multiple': False,
+                                    "tooltip": "Debug Mode",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": "",
+                                    'children': [{'text': 'False', 'value': 0}, {'text': 'True', 'value': 1}],
+                                    })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "text",
+                                    "label": "Tensor Split",
+                                    "id": "kcpp_tensor_split_str",
+                                    "default": self.kcpp_tensor_split_str,
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Tensor Split, values are space separated",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
+        return requested_parameters
+
+    def set_input_parameters(self, parameters):
+        self.kcpp_threads = parameters["kcpp_threads"]
+        self.kcpp_filename = parameters["kcpp_filename"]
+        self.kcpp_ctxsize = parameters["kcpp_ctxsize"]
+        self.kcpp_blasbatchsize = parameters["kcpp_blasbatchsize"]
+        self.kcpp_gpulayers = parameters["kcpp_gpulayers"]
+        self.kcpp_smartcontext = parameters["kcpp_smartcontext"]
+        self.kcpp_ropescale = parameters["kcpp_ropescale"]
+        self.kcpp_ropebase = parameters["kcpp_ropebase"]
+        self.kcpp_debugmode = parameters["kcpp_debugmode"]
+        self.kcpp_tensor_split_str = parameters["kcpp_tensor_split_str"]
+        if self.kcpp_tensor_split_str and self.kcpp_tensor_split_str!="":
+            splits = self.kcpp_tensor_split_str.split()
+            self.kcpp_tensor_split = []
+            for s in splits:
+                self.kcpp_tensor_split.append(int(s))
+
+        accel = parameters["kcpp_accelerator"]
+        if accel==0:
+            self.kcpp_noblas = True
+        elif accel==1:
+           pass
+        elif accel==2:
+            self.kcpp_usecublas = ["normal"]
+        elif accel==3:
+            self.kcpp_useclblast = [0,0]
+        elif accel==4:
+            self.kcpp_useclblast = [1,0]
+        elif accel==5:
+            self.kcpp_useclblast = [0,1]
+        elif accel==6:
+            self.kcpp_noavx2 = True
+        elif accel==7:
+            self.kcpp_noavx2 = True
+            self.kcpp_noblas = True
+            self.kcpp_nommap = True
+        pass
+
+    def unload(self):
+        print("Attemping to unload library")
+        koboldcpp.unload_libs()
+        global kcpp_backend_loaded
+        kcpp_backend_loaded = False
+        pass
+
+    def _load(self, save_model: bool, initial_load: bool) -> None:
+        global kcpp_backend_loaded
+        self.tokenizer = self._get_tokenizer("gpt2")
+        if not kcpp_backend_loaded:
+            kcppargs = KcppArgsObject(model=self.kcpp_filename, model_param=self.kcpp_filename,
+            port=5001, port_param=5001, host='', launch=False, lora=None, threads=self.kcpp_threads, blasthreads=self.kcpp_threads,
+            psutil_set_threads=False, highpriority=False, contextsize=self.kcpp_ctxsize,
+            blasbatchsize=self.kcpp_blasbatchsize, ropeconfig=[self.kcpp_ropescale, self.kcpp_ropebase], stream=False, smartcontext=self.kcpp_smartcontext,
+            unbantokens=False, bantokens=None, usemirostat=None, forceversion=0, nommap=self.kcpp_nommap,
+            usemlock=False, noavx2=self.kcpp_noavx2, debugmode=self.kcpp_debugmode, skiplauncher=True, hordeconfig=None, noblas=self.kcpp_noblas,
+            useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, gpulayers=self.kcpp_gpulayers, tensor_split=self.kcpp_tensor_split, config=None, onready='', multiuser=False)
+
+            koboldcpp.main(kcppargs,False) #initialize library without enabling Lite http server
+            kcpp_backend_loaded = True
+        pass
+
+    def _save_settings(self):
+        pass
+
+    def _raw_generate(
+        self,
+        prompt_tokens: Union[List[int], torch.Tensor],
+        max_new: int,
+        gen_settings: GenerationSettings,
+        single_line: bool = False,
+        batch_count: int = 1,
+        seed: Optional[int] = None,
+        **kwargs,
+    ) -> GenerationResult:
+
+        decoded_prompt = utils.decodenewlines(self.tokenizer.decode(prompt_tokens))
+
+        # Store context in memory to use it for comparison with generated content
+        utils.koboldai_vars.lastctx = decoded_prompt
+
+        genresult = koboldcpp.generate(decoded_prompt,max_new,utils.koboldai_vars.max_length,
+        gen_settings.temp,int(gen_settings.top_k),gen_settings.top_a,gen_settings.top_p,
+        gen_settings.typical,gen_settings.tfs,gen_settings.rep_pen,gen_settings.rep_pen_range,
+        sampler_order=gen_settings.sampler_order,use_default_badwordsids=utils.koboldai_vars.use_default_badwordsids)
+
+        outputs = [genresult]
+        return GenerationResult(
+            model=self,
+            out_batches=np.array(
+                [self.tokenizer.encode(x) for x in outputs]
+            ),
+            prompt=prompt_tokens,
+            is_whole_generation=True,
+            single_line=single_line,
+        )

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

colab.ipynb

@@ -0,0 +1,61 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 0,
+  "metadata": {
+    "colab": {
+      "private_outputs": true,
+      "provenance": [],
+      "gpuType": "T4",
+      "authorship_tag": "ABX9TyOv14c2MWENhO6RJ3uy6vD7",
+      "include_colab_link": true
+    },
+    "kernelspec": {
+      "name": "python3",
+      "display_name": "Python 3"
+    },
+    "language_info": {
+      "name": "python"
+    },
+    "accelerator": "GPU"
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "view-in-github",
+        "colab_type": "text"
+      },
+      "source": [
+        "<a href=\"https://colab.research.google.com/github/henk717/koboldcpp/blob/concedo/colab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "cellView": "form",
+        "id": "uJS9i_Dltv8Y"
+      },
+      "outputs": [],
+      "source": [
+        "#@title <b>v-- Enter your model below and then click this to start Koboldcpp</b>\n",
+        "\n",
+        "Model = \"https://huggingface.co/TheBloke/airoboros-l2-13B-gpt4-1.4.1-GGML/resolve/main/airoboros-l2-13b-gpt4-1.4.1.ggmlv3.q4_0.bin\" #@param [\"\"]{allow-input: true}\n",
+        "Layers = 43 #@param [43]{allow-input: true}\n",
+        "\n",
+        "%cd /content\n",
+        "!git clone https://github.com/LostRuins/koboldcpp\n",
+        "%cd /content/koboldcpp\n",
+        "!make LLAMA_CUBLAS=1\n",
+        "\n",
+        "!wget $Model -O model.ggml\n",
+        "!wget -c https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64\n",
+        "!chmod +x cloudflared-linux-amd64\n",
+        "!nohup ./cloudflared-linux-amd64 tunnel --url http://localhost:5001 &\n",
+        "!sleep 10\n",
+        "!cat nohup.out\n",
+        "!python koboldcpp.py model.ggml --stream --usecublas 0 --gpulayers $Layers --hordeconfig concedo\n"
+      ]
+    }
+  ]
+}

common/CMakeLists.txt

@@ -0,0 +1,20 @@
+# common
+
+set(TARGET common)
+
+add_library(${TARGET} OBJECT
+    common.h
+    common.cpp
+    console.h
+    console.cpp
+    grammar-parser.h
+    grammar-parser.cpp
+    )
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+target_include_directories(${TARGET} PUBLIC .)
+target_compile_features(${TARGET} PUBLIC cxx_std_11)
+target_link_libraries(${TARGET} PRIVATE llama)

common/common.cpp

@@ -0,0 +1,1270 @@
+#include "common.h"
+#include "build-info.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iterator>
+#include <iostream>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_set>
+#include <vector>
+#include <cinttypes>
+
+#if defined(__APPLE__) && defined(__MACH__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <codecvt>
+#include <locale>
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#else
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int32_t get_num_physical_cores() {
+#ifdef __linux__
+    // enumerate the set of thread siblings, num entries is num cores
+    std::unordered_set<std::string> siblings;
+    for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu"
+            + std::to_string(cpu) + "/topology/thread_siblings");
+        if (!thread_siblings.is_open()) {
+            break; // no more cpus
+        }
+        std::string line;
+        if (std::getline(thread_siblings, line)) {
+            siblings.insert(line);
+        }
+    }
+    if (!siblings.empty()) {
+        return static_cast<int32_t>(siblings.size());
+    }
+#elif defined(__APPLE__) && defined(__MACH__)
+    int32_t num_physical_cores;
+    size_t len = sizeof(num_physical_cores);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+#elif defined(_WIN32)
+    //TODO: Implement
+#endif
+    unsigned int n_threads = std::thread::hardware_concurrency();
+    return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
+}
+
+static void process_escapes(std::string& input) {
+    std::size_t input_len = input.length();
+    std::size_t output_idx = 0;
+
+    for (std::size_t input_idx = 0; input_idx < input_len; ++input_idx) {
+        if (input[input_idx] == '\\' && input_idx + 1 < input_len) {
+            switch (input[++input_idx]) {
+                case 'n':  input[output_idx++] = '\n'; break;
+                case 'r':  input[output_idx++] = '\r'; break;
+                case 't':  input[output_idx++] = '\t'; break;
+                case '\'': input[output_idx++] = '\''; break;
+                case '\"': input[output_idx++] = '\"'; break;
+                case '\\': input[output_idx++] = '\\'; break;
+                default:   input[output_idx++] = '\\';
+                           input[output_idx++] = input[input_idx]; break;
+            }
+        } else {
+            input[output_idx++] = input[input_idx];
+        }
+    }
+
+    input.resize(output_idx);
+}
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    bool invalid_param = false;
+    std::string arg;
+    gpt_params default_params;
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-s" || arg == "--seed") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.seed = std::stoul(argv[i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                params.n_threads = std::thread::hardware_concurrency();
+            }
+        } else if (arg == "-p" || arg == "--prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.prompt = argv[i];
+        } else if (arg == "-e" || arg == "--escape") {
+            params.escape = true;
+        } else if (arg == "--prompt-cache") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.path_prompt_cache = argv[i];
+        } else if (arg == "--prompt-cache-all") {
+            params.prompt_cache_all = true;
+        } else if (arg == "--prompt-cache-ro") {
+            params.prompt_cache_ro = true;
+        } else if (arg == "-f" || arg == "--file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        } else if (arg == "-n" || arg == "--n-predict") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_predict = std::stoi(argv[i]);
+        } else if (arg == "--top-k") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.top_k = std::stoi(argv[i]);
+        } else if (arg == "-c" || arg == "--ctx-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--rope-freq-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_base = std::stof(argv[i]);
+        } else if (arg == "--rope-freq-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = 1.0f/std::stof(argv[i]);
+        } else if (arg == "--memory-f32") {
+            params.memory_f16 = false;
+        } else if (arg == "--top-p") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.top_p = std::stof(argv[i]);
+        } else if (arg == "--temp") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.temp = std::stof(argv[i]);
+        } else if (arg == "--tfs") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.tfs_z = std::stof(argv[i]);
+        } else if (arg == "--typical") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.typical_p = std::stof(argv[i]);
+        } else if (arg == "--repeat-last-n") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.repeat_last_n = std::stoi(argv[i]);
+        } else if (arg == "--repeat-penalty") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.repeat_penalty = std::stof(argv[i]);
+        } else if (arg == "--frequency-penalty") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.frequency_penalty = std::stof(argv[i]);
+        } else if (arg == "--presence-penalty") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.presence_penalty = std::stof(argv[i]);
+        } else if (arg == "--mirostat") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.mirostat = std::stoi(argv[i]);
+        } else if (arg == "--mirostat-lr") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.mirostat_eta = std::stof(argv[i]);
+        } else if (arg == "--mirostat-ent") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.mirostat_tau = std::stof(argv[i]);
+        } else if (arg == "--cfg-negative-prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.cfg_negative_prompt = argv[i];
+        } else if (arg == "--cfg-negative-prompt-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
+            if (params.cfg_negative_prompt.back() == '\n') {
+                params.cfg_negative_prompt.pop_back();
+            }
+        } else if (arg == "--cfg-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.cfg_scale = std::stof(argv[i]);
+        } else if (arg == "-b" || arg == "--batch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_batch = std::stoi(argv[i]);
+        } else if (arg == "--keep") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_keep = std::stoi(argv[i]);
+        } else if (arg == "--draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_draft = std::stoi(argv[i]);
+        } else if (arg == "--chunks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_chunks = std::stoi(argv[i]);
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model = argv[i];
+        } else if (arg == "-md" || arg == "--model-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_draft = argv[i];
+        } else if (arg == "-a" || arg == "--alias") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_alias = argv[i];
+        } else if (arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter = argv[i];
+            params.use_mmap = false;
+        } else if (arg == "--lora-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
+        } else if (arg == "-i" || arg == "--interactive") {
+            params.interactive = true;
+        } else if (arg == "--embedding") {
+            params.embedding = true;
+        } else if (arg == "--interactive-first") {
+            params.interactive_first = true;
+        } else if (arg == "-ins" || arg == "--instruct") {
+            params.instruct = true;
+        } else if (arg == "--multiline-input") {
+            params.multiline_input = true;
+        } else if (arg == "--simple-io") {
+            params.simple_io = true;
+        } else if (arg == "--color") {
+            params.use_color = true;
+        } else if (arg == "--mlock") {
+            params.use_mlock = true;
+        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers_draft = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--main-gpu" || arg == "-mg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            params.main_gpu = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+#endif
+        } else if (arg == "--tensor-split" || arg == "-ts") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            std::string arg_next = argv[i];
+
+            // split string by , and /
+            const std::regex regex{R"([,/]+)"};
+            std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
+            std::vector<std::string> split_arg{it, {}};
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+            for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                if (i < split_arg.size()) {
+                    params.tensor_split[i] = std::stof(split_arg[i]);
+                } else {
+                    params.tensor_split[i] = 0.0f;
+                }
+            }
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = false;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--no-mmap") {
+            params.use_mmap = false;
+        } else if (arg == "--numa") {
+            params.numa = true;
+        } else if (arg == "--export") {
+            params.export_cgraph = true;
+        } else if (arg == "--verbose-prompt") {
+            params.verbose_prompt = true;
+        } else if (arg == "-r" || arg == "--reverse-prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.antiprompt.push_back(argv[i]);
+        } else if (arg == "-ld" || arg == "--logdir") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.logdir = argv[i];
+
+            if (params.logdir.back() != DIRECTORY_SEPARATOR) {
+                params.logdir += DIRECTORY_SEPARATOR;
+            }
+        } else if (arg == "--perplexity") {
+            params.perplexity = true;
+        } else if (arg == "--ppl-stride") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.ppl_stride = std::stoi(argv[i]);
+        } else if (arg == "--ppl-output-type") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.ppl_output_type = std::stoi(argv[i]);
+        } else if (arg == "--hellaswag") {
+            params.hellaswag = true;
+        } else if (arg == "--hellaswag-tasks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.hellaswag_tasks = std::stoi(argv[i]);
+        } else if (arg == "--ignore-eos") {
+            params.ignore_eos = true;
+        } else if (arg == "--no-penalize-nl") {
+            params.penalize_nl = false;
+        } else if (arg == "-l" || arg == "--logit-bias") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::stringstream ss(argv[i]);
+            llama_token key;
+            char sign;
+            std::string value_str;
+            try {
+                if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
+                    params.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
+                } else {
+                    throw std::exception();
+                }
+            } catch (const std::exception&) {
+                invalid_param = true;
+                break;
+            }
+        } else if (arg == "-h" || arg == "--help") {
+            gpt_print_usage(argc, argv, default_params);
+#ifndef LOG_DISABLE_LOGS
+            log_print_usage();
+#endif // LOG_DISABLE_LOGS
+            exit(0);
+        } else if (arg == "--random-prompt") {
+            params.random_prompt = true;
+        } else if (arg == "--in-prefix-bos") {
+            params.input_prefix_bos = true;
+        } else if (arg == "--in-prefix") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.input_prefix = argv[i];
+        } else if (arg == "--in-suffix") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.input_suffix = argv[i];
+        } else if (arg == "--grammar") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.grammar = argv[i];
+        } else if (arg == "--grammar-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(params.grammar)
+            );
+#ifndef LOG_DISABLE_LOGS
+        // Parse args for logging parameters
+        } else if ( log_param_single_parse( argv[i] ) ) {
+            // Do nothing, log_param_single_parse automatically does it's thing
+            //  and returns if a match was found and parsed.
+        } else if ( log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i] ) ) {
+            // We have a matching known parameter requiring an argument,
+            //  now we need to check if there is anything after this argv
+            //  and flag invalid_param or parse it.
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            if( !log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i-1], argv[i]) ) {
+                invalid_param = true;
+                break;
+            }
+        // End of Parse args for logging parameters
+#endif // LOG_DISABLE_LOGS
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            gpt_print_usage(argc, argv, default_params);
+            exit(1);
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        gpt_print_usage(argc, argv, default_params);
+        exit(1);
+    }
+    if (params.prompt_cache_all &&
+            (params.interactive || params.interactive_first ||
+             params.instruct)) {
+        fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
+        gpt_print_usage(argc, argv, default_params);
+        exit(1);
+    }
+
+    if (params.escape) {
+        process_escapes(params.prompt);
+        process_escapes(params.input_prefix);
+        process_escapes(params.input_suffix);
+    }
+
+    return true;
+}
+
+void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --interactive-first   run in interactive mode and wait for input right away\n");
+    printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    printf("  -r PROMPT, --reverse-prompt PROMPT\n");
+    printf("                        halt generation at PROMPT, return control in interactive mode\n");
+    printf("                        (can be specified more than once for multiple prompts).\n");
+    printf("  --color               colorise output to distinguish prompt and user input from generations\n");
+    printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -p PROMPT, --prompt PROMPT\n");
+    printf("                        prompt to start generation with (default: empty)\n");
+    printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    printf("  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    printf("  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    printf("                        not supported with --interactive or other interactive options\n");
+    printf("  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    printf("  --random-prompt       start with a randomized prompt.\n");
+    printf("  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
+    printf("  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    printf("  -f FNAME, --file FNAME\n");
+    printf("                        prompt file to start generation.\n");
+    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    printf("  --mirostat N          use Mirostat sampling.\n");
+    printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    printf("                        modifies the likelihood of token appearing in the completion,\n");
+    printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
+    printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  --cfg-negative-prompt PROMPT\n");
+    printf("                        negative prompt to use for guidance. (default: empty)\n");
+    printf("  --cfg-negative-prompt-file FNAME\n");
+    printf("                        negative prompt file to use for guidance. (default: empty)\n");
+    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    printf("  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
+    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    if (llama_mlock_supported()) {
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+    }
+    if (llama_mmap_supported()) {
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+    }
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ngld N, --n-gpu-layers-draft N\n");
+    printf("                        number of layers to store in VRAM for the draft model\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+#ifdef GGML_USE_CUBLAS
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif // GGML_USE_CUBLAS
+#endif
+    printf("  --export              export the computation graph to 'llama.ggml'\n");
+    printf("  --verbose-prompt      print prompt before generation\n");
+    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -md FNAME, --model-draft FNAME\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("  -ld LOGDIR, --logdir LOGDIR\n");
+    printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("\n");
+}
+
+std::string gpt_random_prompt(std::mt19937 & rng) {
+    const int r = rng() % 10;
+    switch (r) {
+        case 0: return "So";
+        case 1: return "Once upon a time";
+        case 2: return "When";
+        case 3: return "The";
+        case 4: return "After";
+        case 5: return "If";
+        case 6: return "import";
+        case 7: return "He";
+        case 8: return "She";
+        case 9: return "They";
+        default: return "To";
+    }
+
+    return "The";
+}
+
+//
+// Model utils
+//
+
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
+    auto lparams = llama_context_default_params();
+
+    lparams.n_ctx           = params.n_ctx;
+    lparams.n_batch         = params.n_batch;
+    if (params.n_gpu_layers != -1) {
+        lparams.n_gpu_layers = params.n_gpu_layers;
+    }
+    lparams.main_gpu        = params.main_gpu;
+    lparams.tensor_split    = params.tensor_split;
+    lparams.low_vram        = params.low_vram;
+    lparams.mul_mat_q       = params.mul_mat_q;
+    lparams.seed            = params.seed;
+    lparams.f16_kv          = params.memory_f16;
+    lparams.use_mmap        = params.use_mmap;
+    lparams.use_mlock       = params.use_mlock;
+    lparams.logits_all      = params.perplexity;
+    lparams.embedding       = params.embedding;
+    lparams.rope_freq_base  = params.rope_freq_base;
+    lparams.rope_freq_scale = params.rope_freq_scale;
+
+    return lparams;
+}
+
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
+    auto lparams = llama_context_params_from_gpt_params(params);
+
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+        return std::make_tuple(nullptr, nullptr);
+    }
+
+    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    if (lctx == NULL) {
+        fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        llama_free_model(model);
+        return std::make_tuple(nullptr, nullptr);
+    }
+
+    if (!params.lora_adapter.empty()) {
+        int err = llama_model_apply_lora_from_file(model,
+                                             params.lora_adapter.c_str(),
+                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             params.n_threads);
+        if (err != 0) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+            llama_free(lctx);
+            llama_free_model(model);
+            return std::make_tuple(nullptr, nullptr);
+        }
+    }
+
+    if (params.ignore_eos) {
+        params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+    }
+
+    {
+        LOG("warming up the model with an empty run\n");
+
+        const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
+        llama_reset_timings(lctx);
+    }
+
+    return std::make_tuple(model, lctx);
+}
+
+//
+// Vocab utils
+//
+
+std::vector<llama_token> llama_tokenize(
+        struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos) {
+    // upper limit for the number of tokens
+    int n_tokens = text.length() + add_bos;
+    std::vector<llama_token> result(n_tokens);
+    n_tokens = llama_tokenize(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_tokenize(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+    return result;
+}
+
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+    std::vector<char> result(8, 0);
+    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+
+    return std::string(result.data(), result.size());
+}
+
+std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    const llama_token bos_id = llama_token_bos(ctx);
+
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        // remove the leading space of the first non-BOS token
+        if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') {
+            piece = piece.substr(1);
+        }
+
+        result += piece;
+    }
+
+    return result;
+}
+
+std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        result += piece;
+    }
+
+    return result;
+}
+
+//
+// Sampling utils
+//
+
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx) {
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(ctx);
+
+    const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const float   repeat_penalty  = params.repeat_penalty;
+    const float   alpha_presence  = params.presence_penalty;
+    const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+
+    float * logits = llama_get_logits(ctx) + idx * n_vocab;
+
+    // Apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    candidates.clear();
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+    if (ctx_guidance) {
+        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    }
+
+    // apply penalties
+    if (!last_tokens.empty()) {
+        const float nl_logit = logits[llama_token_nl(ctx)];
+        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+
+        llama_sample_repetition_penalty(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    if (grammar != NULL) {
+        llama_sample_grammar(ctx, &cur_p, grammar);
+    }
+
+    if (temp <= 0) {
+        // Greedy sampling
+        id = llama_sample_token_greedy(ctx, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            const int mirostat_m = 100;
+            llama_sample_temperature(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+        } else if (mirostat == 2) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            llama_sample_temperature(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+        } else {
+            // Temperature sampling
+            llama_sample_top_k      (ctx, &cur_p, top_k, 1);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
+            llama_sample_typical    (ctx, &cur_p, typical_p, 1);
+            llama_sample_top_p      (ctx, &cur_p, top_p, 1);
+            llama_sample_temperature(ctx, &cur_p, temp);
+
+            {
+                const int n_top = 10;
+                LOG("top %d candidates:\n", n_top);
+
+                for (int i = 0; i < n_top; i++) {
+                    const llama_token id = cur_p.data[i].id;
+                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                }
+            }
+
+            id = llama_sample_token(ctx, &cur_p);
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+        }
+    }
+    // printf("`%d`", candidates_p.size);
+
+    if (grammar != NULL) {
+        llama_grammar_accept_token(ctx, grammar, id);
+    }
+
+    return id;
+}
+
+//
+// YAML utils
+//
+
+// returns true if successful, false otherwise
+bool create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%e, ", data[i]);
+    }
+    fprintf(stream, "%e]\n", data.back());
+}
+
+void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%d, ", data[i]);
+    }
+    fprintf(stream, "%d]\n", data.back());
+}
+
+void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data) {
+    std::string data_str(data == NULL ? "" : data);
+
+    if (data_str.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    size_t pos_start = 0;
+    size_t pos_found = 0;
+
+    if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) {
+        data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
+        data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
+        data_str = "\"" + data_str + "\"";
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    if (data_str.find('\n') == std::string::npos) {
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    fprintf(stream, "%s: |\n", prop_name);
+    while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
+        fprintf(stream, "  %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
+        pos_start = pos_found + 1;
+    }
+}
+
+std::string get_sortable_timestamp() {
+    using clock = std::chrono::system_clock;
+
+    const clock::time_point current_time = clock::now();
+    const time_t as_time_t = clock::to_time_t(current_time);
+    char timestamp_no_ns[100];
+    std::strftime(timestamp_no_ns, 100, "%Y_%m_%d-%H_%M_%S", std::localtime(&as_time_t));
+
+    const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+        current_time.time_since_epoch() % 1000000000).count();
+    char timestamp_ns[11];
+    snprintf(timestamp_ns, 11, "%09" PRId64, ns);
+
+    return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
+}
+
+void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
+                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
+    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
+    fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
+
+#ifdef NDEBUG
+    fprintf(stream, "debug: false\n");
+#else
+    fprintf(stream, "debug: true\n");
+#endif // NDEBUG
+
+    fprintf(stream, "model_desc: %s\n", model_desc);
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(lctx));
+
+#ifdef __OPTIMIZE__
+    fprintf(stream, "optimize: true\n");
+#else
+    fprintf(stream, "optimize: false\n");
+#endif // __OPTIMIZE__
+
+    fprintf(stream, "time: %s\n", timestamp.c_str());
+
+    fprintf(stream, "\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "# User Inputs #\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
+    fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
+    dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str());
+    fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale);
+    fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
+    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
+    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
+    fprintf(stream, "export: %s # default: false\n", params.export_cgraph ? "true" : "false");
+    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
+    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
+    fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
+    fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
+    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
+
+    const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
+    const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
+    fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
+
+    dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str());
+    fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
+    dump_string_yaml_multiline(stream, "in_suffix", params.input_prefix.c_str());
+    fprintf(stream, "instruct: %s # default: false\n", params.instruct ? "true" : "false");
+    fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
+    fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
+    fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
+    fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
+
+    fprintf(stream, "logit_bias:\n");
+    for (std::pair<llama_token, float> lb : params.logit_bias) {
+        if (ignore_eos && lb.first == logit_bias_eos->first) {
+            continue;
+        }
+        fprintf(stream, "  %d: %f", lb.first, lb.second);
+    }
+
+    fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
+    fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
+    fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
+    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
+    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
+    fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
+    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
+    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
+    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
+    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
+    fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false");
+    fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
+    fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
+    fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty);
+    dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
+    fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
+    fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
+    fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
+    dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
+    fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty);
+
+    fprintf(stream, "reverse_prompt:\n");
+    for (std::string ap : params.antiprompt) {
+        size_t pos = 0;
+        while ((pos = ap.find('\n', pos)) != std::string::npos) {
+            ap.replace(pos, 1, "\\n");
+            pos += 1;
+        }
+
+        fprintf(stream, "  - %s\n", ap.c_str());
+    }
+
+    fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
+    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
+    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
+
+    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
+    dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
+
+    fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z);
+    fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "top_k: %d # default: 40\n", params.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p);
+    fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p);
+    fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+}

common/common.h

@@ -0,0 +1,211 @@
+// Various helper functions and utilities
+
+#pragma once
+
+#include "llama.h"
+
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
+#include <string>
+#include <vector>
+#include <random>
+#include <thread>
+#include <unordered_map>
+#include <tuple>
+
+#ifdef _WIN32
+#define DIRECTORY_SEPARATOR '\\'
+#else
+#define DIRECTORY_SEPARATOR '/'
+#endif // _WIN32
+
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define print_build_info() do {                                                             \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+} while(0)
+
+//
+// CLI argument parsing
+//
+int32_t get_num_physical_cores();
+
+struct gpt_params {
+    uint32_t seed                           = -1;   // RNG seed
+    int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_predict                       = -1;   // new tokens to predict
+    int32_t n_ctx                           = 512;  // context size
+    int32_t n_batch                         = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
+    int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
+    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
+    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
+    int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
+    int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
+    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+
+    // sampling parameters
+    int32_t top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typical_p         = 1.00f; // 1.0 = disabled
+    float   temp              = 0.80f; // 1.0 = disabled
+    float   repeat_penalty    = 1.10f; // 1.0 = disabled
+    int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   frequency_penalty = 0.00f; // 0.0 = disabled
+    float   presence_penalty  = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    float   mirostat_eta      = 0.10f; // learning rate
+
+    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
+
+    // Classifier-Free Guidance
+    // https://arxiv.org/abs/2306.17806
+    std::string cfg_negative_prompt;       // string to help guidance
+    float       cfg_scale         = 1.f;   // How strong is guidance
+
+    std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft       = "";                              // draft model for speculative decoding
+    std::string model_alias       = "unknown"; // model alias
+    std::string prompt            = "";
+    std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix      = "";  // string to prefix user inputs with
+    std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
+    std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
+    std::string logdir            = "";  // directory in which to save YAML log files
+
+    std::string lora_adapter = "";  // lora adapter path
+    std::string lora_base    = "";  // base model path for the lora adapter
+
+    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
+    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
+                                    //                                       (which is more convenient to use for plotting)
+                                    //
+    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
+
+    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
+    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
+    bool random_prompt     = false; // do not randomize prompt if none provided
+    bool use_color         = false; // use color to distinguish generations and inputs
+    bool interactive       = false; // interactive mode
+    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            = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
+    bool interactive_first = false; // wait for user input immediately
+    bool multiline_input   = false; // reverse the usage of `\`
+    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
+
+    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool ignore_eos        = false; // ignore generated EOS tokens
+    bool instruct          = false; // instruction mode (used for Alpaca models)
+    bool penalize_nl       = true;  // consider newlines as a repeatable token
+    bool perplexity        = false; // compute perplexity over the prompt
+    bool use_mmap          = true;  // use mmap for faster loads
+    bool use_mlock         = false; // use mlock to keep model in memory
+    bool numa              = false; // attempt optimizations that help on some NUMA systems
+    bool export_cgraph     = false; // export the computation graph
+    bool verbose_prompt    = false; // print prompt tokens before generation
+};
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+//
+// Model utils
+//
+
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
+
+//
+// Vocab utils
+//
+
+// tokenizes a string into a vector of tokens
+// should work similar to Python's `tokenizer.encode`
+std::vector<llama_token> llama_tokenize(
+        struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos);
+
+// tokenizes a token into a piece
+// should work similar to Python's `tokenizer.id_to_piece`
+std::string llama_token_to_piece(
+        const struct llama_context * ctx,
+                       llama_token   token);
+
+// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
+//       that takes into account the tokenizer type and decides how to handle the leading space
+//
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+// removes the leading space from the first non-BOS token
+std::string llama_detokenize_spm(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+std::string llama_detokenize_bpe(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+//
+// Sampling utils
+//
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+//
+// required:
+//  - ctx:    context to use for sampling
+//  - params: sampling parameters
+//
+// optional:
+//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - grammar:       grammar to use for sampling, ignore if NULL
+//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:           sample from llama_get_logits(ctx) + idx * n_vocab
+//
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
+//
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx = 0);
+
+//
+// YAML utils
+//
+
+bool create_directory_with_parents(const std::string & path);
+void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
+void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
+void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
+std::string get_sortable_timestamp();
+
+void dump_non_result_info_yaml(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

common/console.cpp

@@ -0,0 +1,501 @@
+#include "console.h"
+#include <vector>
+#include <iostream>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#ifndef ENABLE_VIRTUAL_TERMINAL_PROCESSING
+#define ENABLE_VIRTUAL_TERMINAL_PROCESSING 0x0004
+#endif
+#else
+#include <climits>
+#include <sys/ioctl.h>
+#include <unistd.h>
+#include <wchar.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <termios.h>
+#endif
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_YELLOW  "\x1b[33m"
+#define ANSI_COLOR_BLUE    "\x1b[34m"
+#define ANSI_COLOR_MAGENTA "\x1b[35m"
+#define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_BOLD          "\x1b[1m"
+
+namespace console {
+
+    //
+    // Console state
+    //
+
+    static bool      advanced_display = false;
+    static bool      simple_io        = true;
+    static display_t current_display  = reset;
+
+    static FILE*     out              = stdout;
+
+#if defined (_WIN32)
+    static void*     hConsole;
+#else
+    static FILE*     tty              = nullptr;
+    static termios   initial_state;
+#endif
+
+    //
+    // Init and cleanup
+    //
+
+    void init(bool use_simple_io, bool use_advanced_display) {
+        advanced_display = use_advanced_display;
+        simple_io = use_simple_io;
+#if defined(_WIN32)
+        // Windows-specific console initialization
+        DWORD dwMode = 0;
+        hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            hConsole = GetStdHandle(STD_ERROR_HANDLE);
+            if (hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(hConsole, &dwMode))) {
+                hConsole = nullptr;
+                simple_io = true;
+            }
+        }
+        if (hConsole) {
+            // Check conditions combined to reduce nesting
+            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING) &&
+                !SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
+                advanced_display = false;
+            }
+            // Set console output codepage to UTF8
+            SetConsoleOutputCP(CP_UTF8);
+        }
+        HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
+        if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
+            // Set console input codepage to UTF16
+            _setmode(_fileno(stdin), _O_WTEXT);
+
+            // Set ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
+            if (simple_io) {
+                dwMode |= ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT;
+            } else {
+                dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
+            }
+            if (!SetConsoleMode(hConIn, dwMode)) {
+                simple_io = true;
+            }
+        }
+#else
+        // POSIX-specific console initialization
+        if (!simple_io) {
+            struct termios new_termios;
+            tcgetattr(STDIN_FILENO, &initial_state);
+            new_termios = initial_state;
+            new_termios.c_lflag &= ~(ICANON | ECHO);
+            new_termios.c_cc[VMIN] = 1;
+            new_termios.c_cc[VTIME] = 0;
+            tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
+
+            tty = fopen("/dev/tty", "w+");
+            if (tty != nullptr) {
+                out = tty;
+            }
+        }
+
+        setlocale(LC_ALL, "");
+#endif
+    }
+
+    void cleanup() {
+        // Reset console display
+        set_display(reset);
+
+#if !defined(_WIN32)
+        // Restore settings on POSIX systems
+        if (!simple_io) {
+            if (tty != nullptr) {
+                out = stdout;
+                fclose(tty);
+                tty = nullptr;
+            }
+            tcsetattr(STDIN_FILENO, TCSANOW, &initial_state);
+        }
+#endif
+    }
+
+    //
+    // Display and IO
+    //
+
+    // Keep track of current display and only emit ANSI code if it changes
+    void set_display(display_t display) {
+        if (advanced_display && current_display != display) {
+            fflush(stdout);
+            switch(display) {
+                case reset:
+                    fprintf(out, ANSI_COLOR_RESET);
+                    break;
+                case prompt:
+                    fprintf(out, ANSI_COLOR_YELLOW);
+                    break;
+                case user_input:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
+                    break;
+                case error:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
+            }
+            current_display = display;
+            fflush(out);
+        }
+    }
+
+    static char32_t getchar32() {
+#if defined(_WIN32)
+        HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+        wchar_t high_surrogate = 0;
+
+        while (true) {
+            INPUT_RECORD record;
+            DWORD count;
+            if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+                return WEOF;
+            }
+
+            if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+                wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+                if (wc == 0) {
+                    continue;
+                }
+
+                if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                    high_surrogate = wc;
+                    continue;
+                }
+                if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                    if (high_surrogate != 0) { // Check if we have a high surrogate
+                        return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                    }
+                }
+
+                high_surrogate = 0; // Reset the high surrogate
+                return static_cast<char32_t>(wc);
+            }
+        }
+#else
+        wchar_t wc = getwchar();
+        if (static_cast<wint_t>(wc) == WEOF) {
+            return WEOF;
+        }
+
+#if WCHAR_MAX == 0xFFFF
+        if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+            wchar_t low_surrogate = getwchar();
+            if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
+                return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
+            }
+        }
+        if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
+            return 0xFFFD; // Return the replacement character U+FFFD
+        }
+#endif
+
+        return static_cast<char32_t>(wc);
+#endif
+    }
+
+    static void pop_cursor() {
+#if defined(_WIN32)
+        if (hConsole != NULL) {
+            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
+
+            COORD newCursorPosition = bufferInfo.dwCursorPosition;
+            if (newCursorPosition.X == 0) {
+                newCursorPosition.X = bufferInfo.dwSize.X - 1;
+                newCursorPosition.Y -= 1;
+            } else {
+                newCursorPosition.X -= 1;
+            }
+
+            SetConsoleCursorPosition(hConsole, newCursorPosition);
+            return;
+        }
+#endif
+        putc('\b', out);
+    }
+
+    static int estimateWidth(char32_t codepoint) {
+#if defined(_WIN32)
+        (void)codepoint;
+        return 1;
+#else
+        return wcwidth(codepoint);
+#endif
+    }
+
+    static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+#if defined(_WIN32)
+        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+        if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
+            // go with the default
+            return expectedWidth;
+        }
+        COORD initialPosition = bufferInfo.dwCursorPosition;
+        DWORD nNumberOfChars = length;
+        WriteConsole(hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
+
+        CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
+        GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+
+        // Figure out our real position if we're in the last column
+        if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
+            DWORD nNumberOfChars;
+            WriteConsole(hConsole, &" \b", 2, &nNumberOfChars, NULL);
+            GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+        }
+
+        int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
+        if (width < 0) {
+            width += newBufferInfo.dwSize.X;
+        }
+        return width;
+#else
+        // We can trust expectedWidth if we've got one
+        if (expectedWidth >= 0 || tty == nullptr) {
+            fwrite(utf8_codepoint, length, 1, out);
+            return expectedWidth;
+        }
+
+        fputs("\033[6n", tty); // Query cursor position
+        int x1;
+        int y1;
+        int x2;
+        int y2;
+        int results = 0;
+        results = fscanf(tty, "\033[%d;%dR", &y1, &x1);
+
+        fwrite(utf8_codepoint, length, 1, tty);
+
+        fputs("\033[6n", tty); // Query cursor position
+        results += fscanf(tty, "\033[%d;%dR", &y2, &x2);
+
+        if (results != 4) {
+            return expectedWidth;
+        }
+
+        int width = x2 - x1;
+        if (width < 0) {
+            // Calculate the width considering text wrapping
+            struct winsize w;
+            ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
+            width += w.ws_col;
+        }
+        return width;
+#endif
+    }
+
+    static void replace_last(char ch) {
+#if defined(_WIN32)
+        pop_cursor();
+        put_codepoint(&ch, 1, 1);
+#else
+        fprintf(out, "\b%c", ch);
+#endif
+    }
+
+    static void append_utf8(char32_t ch, std::string & out) {
+        if (ch <= 0x7F) {
+            out.push_back(static_cast<unsigned char>(ch));
+        } else if (ch <= 0x7FF) {
+            out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0xFFFF) {
+            out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0x10FFFF) {
+            out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else {
+            // Invalid Unicode code point
+        }
+    }
+
+    // Helper function to remove the last UTF-8 character from a string
+    static void pop_back_utf8_char(std::string & line) {
+        if (line.empty()) {
+            return;
+        }
+
+        size_t pos = line.length() - 1;
+
+        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
+        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
+            if ((line[pos] & 0xC0) != 0x80) {
+                break; // Found the start of the character
+            }
+        }
+        line.erase(pos);
+    }
+
+    static bool readline_advanced(std::string & line, bool multiline_input) {
+        if (out != stdout) {
+            fflush(stdout);
+        }
+
+        line.clear();
+        std::vector<int> widths;
+        bool is_special_char = false;
+        bool end_of_stream = false;
+
+        char32_t input_char;
+        while (true) {
+            fflush(out); // Ensure all output is displayed before waiting for input
+            input_char = getchar32();
+
+            if (input_char == '\r' || input_char == '\n') {
+                break;
+            }
+
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
+                end_of_stream = true;
+                break;
+            }
+
+            if (is_special_char) {
+                set_display(user_input);
+                replace_last(line.back());
+                is_special_char = false;
+            }
+
+            if (input_char == '\033') { // Escape sequence
+                char32_t code = getchar32();
+                if (code == '[' || code == 0x1B) {
+                    // Discard the rest of the escape sequence
+                    while ((code = getchar32()) != (char32_t) WEOF) {
+                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
+                            break;
+                        }
+                    }
+                }
+            } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
+                if (!widths.empty()) {
+                    int count;
+                    do {
+                        count = widths.back();
+                        widths.pop_back();
+                        // Move cursor back, print space, and move cursor back again
+                        for (int i = 0; i < count; i++) {
+                            replace_last(' ');
+                            pop_cursor();
+                        }
+                        pop_back_utf8_char(line);
+                    } while (count == 0 && !widths.empty());
+                }
+            } else {
+                int offset = line.length();
+                append_utf8(input_char, line);
+                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
+                if (width < 0) {
+                    width = 0;
+                }
+                widths.push_back(width);
+            }
+
+            if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
+                set_display(prompt);
+                replace_last(line.back());
+                is_special_char = true;
+            }
+        }
+
+        bool has_more = multiline_input;
+        if (is_special_char) {
+            replace_last(' ');
+            pop_cursor();
+
+            char last = line.back();
+            line.pop_back();
+            if (last == '\\') {
+                line += '\n';
+                fputc('\n', out);
+                has_more = !has_more;
+            } else {
+                // llama will just eat the single space, it won't act as a space
+                if (line.length() == 1 && line.back() == ' ') {
+                    line.clear();
+                    pop_cursor();
+                }
+                has_more = false;
+            }
+        } else {
+            if (end_of_stream) {
+                has_more = false;
+            } else {
+                line += '\n';
+                fputc('\n', out);
+            }
+        }
+
+        fflush(out);
+        return has_more;
+    }
+
+    static bool readline_simple(std::string & line, bool multiline_input) {
+#if defined(_WIN32)
+        std::wstring wline;
+        if (!std::getline(std::wcin, wline)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            GenerateConsoleCtrlEvent(CTRL_C_EVENT, 0);
+            return false;
+        }
+
+        int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), NULL, 0, NULL, NULL);
+        line.resize(size_needed);
+        WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), &line[0], size_needed, NULL, NULL);
+#else
+        if (!std::getline(std::cin, line)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            return false;
+        }
+#endif
+        if (!line.empty()) {
+            char last = line.back();
+            if (last == '/') { // Always return control on '/' symbol
+                line.pop_back();
+                return false;
+            }
+            if (last == '\\') { // '\\' changes the default action
+                line.pop_back();
+                multiline_input = !multiline_input;
+            }
+        }
+        line += '\n';
+
+        // By default, continue input if multiline_input is set
+        return multiline_input;
+    }
+
+    bool readline(std::string & line, bool multiline_input) {
+        set_display(user_input);
+
+        if (simple_io) {
+            return readline_simple(line, multiline_input);
+        }
+        return readline_advanced(line, multiline_input);
+    }
+
+}

common/console.h

@@ -0,0 +1,19 @@
+// Console functions
+
+#pragma once
+
+#include <string>
+
+namespace console {
+    enum display_t {
+        reset = 0,
+        prompt,
+        user_input,
+        error
+    };
+
+    void init(bool use_simple_io, bool use_advanced_display);
+    void cleanup();
+    void set_display(display_t display);
+    bool readline(std::string & line, bool multiline_input);
+}

common/grammar-parser.cpp

@@ -0,0 +1,424 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from llama.cpp
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    static void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<llama_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    static bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    static const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    static const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<llama_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = LLAMA_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum llama_gretype type = last_sym_start < out_elements.size()
+                        ? LLAMA_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<llama_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<llama_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({LLAMA_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    static const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    static void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    static bool is_char_element(llama_grammar_element elem) {
+        switch (elem.type) {
+            case LLAMA_GRETYPE_CHAR:           return true;
+            case LLAMA_GRETYPE_CHAR_NOT:       return true;
+            case LLAMA_GRETYPE_CHAR_ALT:       return true;
+            case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
+                case LLAMA_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case LLAMA_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case LLAMA_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                case LLAMA_GRETYPE_ALT:
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                case LLAMA_GRETYPE_CHAR_NOT:
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    static void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<llama_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            llama_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case LLAMA_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case LLAMA_GRETYPE_CHAR_ALT:
+                    case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const llama_grammar_element *> parse_state::c_rules() {
+        std::vector<const llama_grammar_element *> ret;
+        ret.reserve(rules.size());
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

common/grammar-parser.h

@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by llama.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "llama.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                 symbol_ids;
+        std::vector<std::vector<llama_grammar_element>> rules;
+
+        std::vector<const llama_grammar_element *> c_rules();
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}

common/log.h

@@ -0,0 +1,643 @@
+#pragma once
+
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
+
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a diffrent function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_diffrent()
+//  #include "log.h"
+//
+//  FILE* log_handler_diffrent()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
+#endif
+
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
+
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+    static std::string pid;
+    if (pid.empty())
+    {
+        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+        //  it's not the same as "pid" but is unique enough to solve multiple instances
+        //  trying to write to the same log.
+        std::stringstream ss;
+        ss << std::this_thread::get_id();
+        pid = ss.str();
+    }
+
+    return pid;
+}
+
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
+
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    std::stringstream buf;
+
+    buf << log_file_basename;
+    buf << ".";
+    buf << log_get_pid();
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_IMPL(str, ...)                                                                                          \
+    {                                                                                                               \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
+    }
+#else
+    #define LOG_IMPL(str, ...)                                                                                               \
+    {                                                                                                                    \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    }
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_TEE_IMPL(str, ...)                                                                                                          \
+    {                                                                                                                                   \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    }
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                               \
+    {                                                                                                                                        \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    }
+#endif
+
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
+
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#ifndef _MSC_VER
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#ifndef _MSC_VER
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#ifndef _MSC_VER
+    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
+    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
+#else
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized{false};
+    static bool _disabled{(filename.empty() && target == nullptr)};
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
+    if (_initialized)
+    {
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  untill enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n")
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET))
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n")
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n")
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n")
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n")
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n")
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n")
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n")
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n")
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n")
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n")
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n")
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n")
+#ifdef _MSC_VER
+    LOG_TEE("15 Hello msvc TEE without arguments\n")
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test")
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n")
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test")
+    LOG("19 Hello msvc LOG without arguments\n")
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test")
+    LOGLN("21 Hello msvc LOGLN without arguments\n")
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test")
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    printf("log options:\n");
+    /* format
+    printf("  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str())
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+    [&tokens, &ctx]()                                                        \
+    {                                                                        \
+        std::stringstream buf;                                               \
+        buf << "[ ";                                                         \
+                                                                             \
+        bool first = true;                                                   \
+        for (const auto &token : tokens)                                     \
+        {                                                                    \
+            if (!first)                                                      \
+                buf << ", ";                                                 \
+            else                                                             \
+                first = false;                                               \
+                                                                             \
+            auto detokenized = llama_token_to_piece(ctx, token);             \
+                                                                             \
+            detokenized.erase(                                               \
+                std::remove_if(                                              \
+                    detokenized.begin(),                                     \
+                    detokenized.end(),                                       \
+                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end());                                          \
+                                                                             \
+            buf                                                              \
+                << "'" << detokenized << "'"                                 \
+                << ":" << std::to_string(token);                             \
+        }                                                                    \
+        buf << " ]";                                                         \
+                                                                             \
+        return buf.str();                                                    \
+    }()                                                                      \
+        .c_str()
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS

convert-baichuan-hf-to-gguf.py

@@ -0,0 +1,304 @@
+#!/usr/bin/env python3
+# HF baichuan --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+import itertools
+import gguf
+import numpy as np
+import torch
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
+
+# reverse HF permute back to original pth layout
+
+
+def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
+
+    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape))
+
+def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray:
+        r = weights.shape[0] // 3
+        return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
+
+def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray:
+        r = weights.shape[0] // 3
+        return weights[r * n_part : r * n_part + r, ...]
+
+def count_model_parts(dir_model: str) -> int:
+    num_parts = 0
+
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+
+    return num_parts
+
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+print("hello print: ",hparams["architectures"][0])
+if hparams["architectures"][0] != "BaichuanForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+print(f"num_parts:{num_parts}\n")
+ARCH=gguf.MODEL_ARCH.BAICHUAN
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+head_count = hparams["num_attention_heads"]
+
+if "num_key_value_heads" in hparams:
+    head_count_kv = hparams["num_key_value_heads"]
+else:
+    head_count_kv = head_count
+
+if "_name_or_path" in hparams:
+    hf_repo = hparams["_name_or_path"]
+else:
+    hf_repo = ""
+
+if "max_sequence_length" in hparams:
+    ctx_length = hparams["max_sequence_length"]
+elif "max_position_embeddings" in hparams:
+    ctx_length = hparams["max_position_embeddings"]
+elif "model_max_length" in hparams:
+    ctx_length = hparams["model_max_length"]
+else:
+    print("gguf: can not find ctx length parameter.")
+
+    sys.exit()
+
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_source_hf_repo(hf_repo)
+gguf_writer.add_tensor_data_layout("Meta AI original pth")
+gguf_writer.add_context_length(ctx_length)
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+gguf_writer.add_head_count(head_count)
+gguf_writer.add_head_count_kv(head_count_kv)
+gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+
+if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
+    if "type" in hparams["rope_scaling"]:
+        if hparams["rope_scaling"]["type"] == "linear":
+            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytes] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
+
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab, scores and token types")
+
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
+
+for i in range(tokenizer.vocab_size()):
+    text: bytes
+    score: float
+
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
+
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
+
+    # toktype = 4 is user-defined = tokens from added_tokens.json
+
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
+
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
+
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
+
+        print("gguf: get added tokens")
+
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
+
+
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    tmp=model_part
+    for i in range(block_count):
+        if f"model.layers.{i}.self_attn.W_pack.weight" in model_part:
+            print(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+        # we don't need these
+        if name.endswith(".rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-falcon-hf-to-gguf.py

@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+# HF falcon--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def bytes_to_unicode():
+    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "RWForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.FALCON
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Falcon")
+gguf_writer.add_context_length(2048) # not in config.json
+gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+if "n_head_kv" in hparams:
+    gguf_writer.add_head_count_kv(hparams["n_head_kv"])
+else:
+    gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+    scores.append(0.0)                      # dymmy
+    toktypes.append(gguf.TokenType.NORMAL)  # dummy
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# params for qkv transform
+n_head    = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
+head_dim = hparams["hidden_size"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        # QKV tensor transform
+        # The original query_key_value tensor contains n_head_kv "kv groups",
+        # each consisting of n_head/n_head_kv query weights followed by one key
+        # and one value weight (shared by all query heads in the kv group).
+        # This layout makes it a big pain to work with in GGML.
+        # So we rearrange them here,, so that we have n_head query weights
+        # followed by n_head_kv key weights followed by n_head_kv value weights,
+        # in contiguous fashion.
+        # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+
+        if "query_key_value" in name:
+            qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+            q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
+            k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+            v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+            data = torch.cat((q,k,v)).reshape_as(data)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-gptneox-hf-to-gguf.py

@@ -0,0 +1,251 @@
+#!/usr/bin/env python3
+# HF gptneox--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTNeoXForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.GPTNEOX
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_context_length(hparams["max_position_embeddings"])
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"])))
+gguf_writer.add_head_count(hparams["num_attention_heads"])
+gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+vocab_size = len(tokenizer_json["model"]["vocab"])
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+
+gguf_writer.add_token_list(tokens)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        # we don't need these
+        if name.endswith(".attention.masked_bias") or name.endswith(".attention.bias") or name.endswith(".attention.rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-llama-ggml-to-gguf.py

@@ -0,0 +1,451 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import math
+import struct
+import sys
+from enum import IntEnum
+from pathlib import Path
+
+import numpy as np
+
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+# Note: Does not support GGML_QKK_64
+QK_K = 256
+# Items here are (block size, type size)
+GGML_QUANT_SIZES = {
+    gguf.GGMLQuantizationType.F32  : (1, 4),
+    gguf.GGMLQuantizationType.F16  : (1, 2),
+    gguf.GGMLQuantizationType.Q4_0 : (32, 2 + 16),
+    gguf.GGMLQuantizationType.Q4_1 : (32, 2 + 2 + 16),
+    gguf.GGMLQuantizationType.Q5_0 : (32, 2 + 4 + 16),
+    gguf.GGMLQuantizationType.Q5_1 : (32, 2 + 2 + 4 + 16),
+    gguf.GGMLQuantizationType.Q8_0 : (32, 2 + 32),
+    gguf.GGMLQuantizationType.Q8_1 : (32, 4 + 4 + 32),
+    gguf.GGMLQuantizationType.Q2_K : (256, 2 + 2 + QK_K // 16 + QK_K // 4),
+    gguf.GGMLQuantizationType.Q3_K : (256, 2 + QK_K // 4 + QK_K // 8 + 12),
+    gguf.GGMLQuantizationType.Q4_K : (256, 2 + 2 + QK_K // 2 + 12),
+    gguf.GGMLQuantizationType.Q5_K : (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12),
+    gguf.GGMLQuantizationType.Q6_K : (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16),
+    gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
+}
+
+class GGMLFormat(IntEnum):
+    GGML = 0
+    GGMF = 1
+    GGJT = 2
+
+class GGMLFType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1
+    MOSTLY_Q4_0          = 2
+    MOSTLY_Q4_1          = 3
+    MOSTLY_Q4_1_SOME_F16 = 4
+    MOSTLY_Q8_0          = 7
+    MOSTLY_Q5_0          = 8
+    MOSTLY_Q5_1          = 9
+    MOSTLY_Q2_K          = 10
+    MOSTLY_Q3_K_S        = 11
+    MOSTLY_Q3_K_M        = 12
+    MOSTLY_Q3_K_L        = 13
+    MOSTLY_Q4_K_S        = 14
+    MOSTLY_Q4_K_M        = 15
+    MOSTLY_Q5_K_S        = 16
+    MOSTLY_Q5_K_M        = 17
+    MOSTLY_Q6_K          = 18
+
+class Hyperparameters:
+    def __init__(self):
+        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
+        self.n_layer = self.n_rot = self.n_ff = 0
+        self.ftype = GGMLFType.ALL_F32
+
+    def set_n_ff(self, model):
+        ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
+        assert ff_tensor_idx is not None, 'Missing layer 0 FF tensor'
+        ff_tensor = model.tensors[ff_tensor_idx]
+        self.n_ff = ff_tensor.dims[1]
+
+    def load(self, data, offset):
+        (
+            self.n_vocab,
+            self.n_embd,
+            self.n_mult,
+            self.n_head,
+            self.n_layer,
+            self.n_rot,
+            ftype,
+        ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
+        try:
+            self.ftype = GGMLFType(ftype)
+        except ValueError:
+            raise ValueError(f'Invalid ftype {ftype}')
+        return 4 * 7
+
+    def __str__(self):
+        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
+
+class Vocab:
+    def __init__(self, load_scores = True):
+        self.items = []
+        self.load_scores = load_scores
+
+    def load(self, data, offset, n_vocab):
+        orig_offset = offset
+        for _ in range(n_vocab):
+            itemlen = struct.unpack('<I', data[offset:offset + 4])[0]
+            assert itemlen < 4096, 'Absurd vocab item length'
+            offset += 4
+            item_text = bytes(data[offset:offset + itemlen])
+            offset += itemlen
+            if self.load_scores:
+                item_score = struct.unpack('<f', data[offset:offset + 4])[0]
+                offset += 4
+            else:
+                item_score = 0.0
+            self.items.append((item_text, item_score))
+        return offset - orig_offset
+
+class Tensor:
+    def __init__(self, use_padding = True):
+        self.name = None
+        self.dims: tuple[int, ...] = ()
+        self.dtype = None
+        self.start_offset = 0
+        self.len_bytes = np.int64(0)
+        self.use_padding = use_padding
+
+    def load(self, data, offset):
+        orig_offset = offset
+        (n_dims, name_len, dtype) = struct.unpack('<3I', data[offset:offset + 12])
+        assert n_dims >= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}'
+        assert name_len < 4096, 'Absurd tensor name length'
+        quant = GGML_QUANT_SIZES.get(dtype)
+        assert quant is not None, 'Unknown tensor type'
+        (blksize, tysize) = quant
+        offset += 12
+        self.dtype= dtype
+        self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
+        offset += 4 * n_dims
+        self.name = bytes(data[offset:offset + name_len])
+        offset += name_len
+        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
+        offset += pad
+        n_elems = np.prod(self.dims)
+        n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
+        self.start_offset = offset
+        self.len_bytes = n_bytes
+        offset += n_bytes
+        # print(n_dims, name_len, dtype, self.dims, self.name, pad)
+        return offset - orig_offset
+
+class GGMLModel:
+    def __init__(self):
+        self.hyperparameters = None
+        self.vocab = None
+        self.tensor_map = {}
+        self.tensors = []
+
+    def validate_header(self, data, offset):
+        magic = bytes(data[offset:offset + 4])
+        if magic == b'GGUF':
+            raise ValueError('File is already in GGUF format.')
+        if magic == b'lmgg':
+            self.file_format = GGMLFormat.GGML
+            self.format_version = 1
+            return 4
+        version = struct.unpack('<I', data[offset + 4:offset + 8])[0]
+        if magic == b'fmgg':
+            if version != 1:
+                raise ValueError(f'Cannot handle unexpected GGMF file version {version}')
+            self.file_format = GGMLFormat.GGMF
+            self.format_version = version
+            return 8
+        if magic == b'tjgg':
+            if version < 1 or version > 3:
+                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
+            self.file_format = GGMLFormat.GGJT
+            self.format_version = version
+            return 8
+        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
+
+    def validate_conversion(self, ftype):
+        err = ''
+        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
+            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
+                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
+        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
+            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
+        if len(err) > 0:
+            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
+
+    def load(self, data, offset):
+        offset += self.validate_header(data, offset)
+        hp = Hyperparameters()
+        offset += hp.load(data, offset)
+        print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
+        self.validate_conversion(hp.ftype)
+        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
+        offset += vocab.load(data, offset, hp.n_vocab)
+        tensors: list[Tensor] = []
+        tensor_map = {}
+        while offset < len(data):
+            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
+            offset += tensor.load(data, offset)
+            tensor_map[tensor.name] = len(tensors)
+            tensors.append(tensor)
+        self.hyperparameters = hp
+        self.vocab = vocab
+        self.tensors = tensors
+        self.tensor_map = tensor_map
+        hp.set_n_ff(self)
+        return offset
+
+class GGMLToGGUF:
+    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
+        hp = ggml_model.hyperparameters
+        self.model = ggml_model
+        self.data = data
+        self.cfg = cfg
+        self.params_override = params_override
+        self.vocab_override = vocab_override
+        self.special_vocab = special_vocab
+        if params_override is not None:
+            n_kv_head = params_override.n_head_kv
+        else:
+            if cfg.gqa == 1:
+                n_kv_head = hp.n_head
+            else:
+                gqa = float(cfg.gqa)
+                n_kv_head = None
+                for x in range(1, 256):
+                    if float(hp.n_head) / float(x) == gqa:
+                        n_kv_head = x
+                assert n_kv_head is not None, "Couldn't determine n_kv_head from GQA param"
+                print(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}')
+        self.n_kv_head = n_kv_head
+        self.name_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.LLAMA, ggml_model.hyperparameters.n_layer)
+
+    def save(self):
+        print('* Preparing to save GGUF file')
+        gguf_writer = gguf.GGUFWriter(
+            self.cfg.output,
+            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
+            use_temp_file = False )
+        self.add_params(gguf_writer)
+        self.add_vocab(gguf_writer)
+        if self.special_vocab is not None:
+            self.special_vocab.add_to_gguf(gguf_writer)
+        self.add_tensors(gguf_writer)
+        print("    gguf: write header")
+        gguf_writer.write_header_to_file()
+        print("    gguf: write metadata")
+        gguf_writer.write_kv_data_to_file()
+        print("    gguf: write tensors")
+        gguf_writer.write_tensors_to_file()
+        gguf_writer.close()
+
+    def add_params(self, gguf_writer):
+        hp = self.model.hyperparameters
+        cfg = self.cfg
+        if cfg.desc is not None:
+            desc = cfg.desc
+        else:
+            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
+        try:
+            # Filenames aren't necessarily valid UTF8.
+            name = cfg.name if cfg.name is not None else cfg.input.name
+        except UnicodeDecodeError:
+            name = None
+        print('* Adding model parameters and KV items')
+        if name is not None:
+            gguf_writer.add_name(name)
+        gguf_writer.add_description(desc)
+        gguf_writer.add_file_type(int(hp.ftype))
+        if self.params_override is not None:
+            po = self.params_override
+            assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
+            assert po.n_layer == hp.n_layer, 'Model hyperparams mismatch'
+            assert po.n_head == hp.n_head, 'Model hyperparams mismatch'
+            gguf_writer.add_context_length      (po.n_ctx)
+            gguf_writer.add_embedding_length    (po.n_embd)
+            gguf_writer.add_block_count         (po.n_layer)
+            gguf_writer.add_feed_forward_length (po.n_ff)
+            gguf_writer.add_rope_dimension_count(po.n_embd // po.n_head)
+            gguf_writer.add_head_count          (po.n_head)
+            gguf_writer.add_head_count_kv       (po.n_head_kv)
+            gguf_writer.add_layer_norm_rms_eps  (po.f_norm_eps)
+            return
+        gguf_writer.add_context_length(cfg.context_length)
+        gguf_writer.add_embedding_length(hp.n_embd)
+        gguf_writer.add_block_count(hp.n_layer)
+        gguf_writer.add_feed_forward_length(hp.n_ff)
+        gguf_writer.add_rope_dimension_count(hp.n_embd // hp.n_head)
+        gguf_writer.add_head_count(hp.n_head)
+        gguf_writer.add_head_count_kv(self.n_kv_head)
+        gguf_writer.add_layer_norm_rms_eps(float(cfg.eps))
+
+    def add_vocab(self, gguf_writer):
+        hp = self.model.hyperparameters
+        gguf_writer.add_tokenizer_model('llama')
+        tokens = []
+        scores = []
+        toktypes = []
+        if self.vocab_override is not None:
+            vo = self.vocab_override
+            print('* Adding vocab item(s)')
+            for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
+                tokens.append(vbytes)
+                scores.append(score)
+                toktypes.append(ttype)
+            assert len(tokens) == hp.n_vocab, \
+                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
+            gguf_writer.add_token_list(tokens)
+            gguf_writer.add_token_scores(scores)
+            if len(toktypes) > 0:
+                gguf_writer.add_token_types(toktypes)
+            return
+        print(f'* Adding {hp.n_vocab} vocab item(s)')
+        assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab'
+        for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items):
+            tt = 1 # Normal
+            # Special handling for UNK, BOS, EOS tokens.
+            if tokid <= 2:
+                if tokid == 0:
+                    vbytes = b'<unk>'
+                    tt = 2
+                elif tokid == 1:
+                    vbytes = b'<s>'
+                    tt = 3
+                else:
+                    vbytes = b'</s>'
+                    tt = 3
+            elif len(vbytes) == 0:
+                tt = 3 # Control
+            elif tokid >= 3 and tokid <= 258 and len(vbytes) == 1:
+                vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8')
+                tt = 6 # Byte
+            else:
+                vbytes = vbytes.replace(b' ', b'\xe2\x96\x81')
+            toktypes.append(tt)
+            tokens.append(vbytes)
+            scores.append(vscore)
+        gguf_writer.add_token_list(tokens)
+        gguf_writer.add_token_scores(scores)
+        gguf_writer.add_token_types(toktypes)
+        gguf_writer.add_unk_token_id(0)
+        gguf_writer.add_bos_token_id(1)
+        gguf_writer.add_eos_token_id(2)
+
+    def add_tensors(self, gguf_writer):
+        tensor_map = self.name_map
+        data = self.data
+        print(f'* Adding {len(self.model.tensors)} tensor(s)')
+        for tensor in self.model.tensors:
+            name = str(tensor.name, 'UTF-8')
+            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+            assert mapped_name is not None, f'Bad name {name}'
+            tempdims = list(tensor.dims[:])
+            if len(tempdims) > 1:
+                temp = tempdims[1]
+                tempdims[1] = tempdims[0]
+                tempdims[0] = temp
+            # print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}')
+            gguf_writer.add_tensor(
+                mapped_name,
+                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
+                raw_shape = tempdims,
+                raw_dtype = tensor.dtype )
+
+def handle_metadata(cfg, hp):
+    import convert
+    assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
+    hf_config_path   = cfg.model_metadata_dir / "config.json"
+    orig_config_path = cfg.model_metadata_dir / "params.json"
+    # We pass a fake model here. "original" mode will check the shapes of some
+    # tensors if information is missing in the .json file: other than that, the
+    # model data isn't used so this should be safe (at least for now).
+    fakemodel = {
+        'tok_embeddings.weight': convert.LazyTensor.__new__(convert.LazyTensor),
+        'layers.0.feed_forward.w1.weight': convert.LazyTensor.__new__(convert.LazyTensor),
+    }
+    fakemodel['tok_embeddings.weight'].shape = [hp.n_vocab]
+    fakemodel['layers.0.feed_forward.w1.weight'].shape = [hp.n_ff]
+    if hf_config_path.exists():
+        params = convert.Params.loadHFTransformerJson(fakemodel, hf_config_path)
+    elif orig_config_path.exists():
+        params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
+    else:
+        raise ValueError('Unable to load metadata')
+    vocab = convert.load_vocab(
+        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+        cfg.vocabtype )
+    # FIXME: Respect cfg.vocab_dir?
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
+    convert.check_vocab_size(params, vocab)
+    return (params, vocab, svocab)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
+    parser.add_argument('--input', '-i', type = Path, required = True,
+        help = 'Input GGMLv3 filename')
+    parser.add_argument('--output', '-o', type = Path, required = True,
+        help ='Output GGUF filename')
+    parser.add_argument('--name',
+        help = 'Set model name')
+    parser.add_argument('--desc',
+        help = 'Set model description')
+    parser.add_argument('--gqa', type = int, default = 1,
+        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+    parser.add_argument('--eps', default = '5.0e-06',
+        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+    parser.add_argument('--context-length', '-c', type=int, default = 2048,
+        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+    parser.add_argument('--model-metadata-dir', '-m', type = Path,
+        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+    parser.add_argument("--vocab-dir", type=Path,
+        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(f'* Using config: {cfg}')
+    print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n')
+    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
+        print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
+    data = np.memmap(cfg.input, mode = 'r')
+    model = GGMLModel()
+    print('* Scanning GGML input file')
+    offset = model.load(data, 0)
+    print(f'* GGML model hyperparameters: {model.hyperparameters}')
+    vocab_override = None
+    params_override = None
+    special_vocab = None
+    if cfg.model_metadata_dir is not None:
+        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
+        print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
+        print(f'* Overriding params: {params_override}')
+        print(f'* Overriding vocab: {vocab_override}')
+        print(f'* Special vocab: {special_vocab}')
+    else:
+        print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
+        if model.file_format == GGMLFormat.GGML:
+            print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
+    converter = GGMLToGGUF(model, data, cfg,
+        params_override = params_override,
+        vocab_override = vocab_override,
+        special_vocab = special_vocab )
+    converter.save()
+    print(f'* Successful completion. Output saved to: {cfg.output}')
+
+if __name__ == '__main__':
+    main()

convert-lora-to-ggml.py

@@ -1,28 +1,29 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
+from __future__ import annotations
+
 import json
 import os
 import re
 import struct
 import sys
-from typing import Any, Dict, Sequence, TextIO
+from typing import Any, BinaryIO, Sequence
 
+import numpy as np
 import torch
 
-from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType
+NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
+
 
 HF_SUBLAYER_TO_GGML = {
-    "self_attn.q_proj": "attention.wq",
-    "self_attn.k_proj": "attention.wk",
-    "self_attn.v_proj": "attention.wv",
-    "self_attn.o_proj": "attention.wo",
-    "mlp.gate_proj": "feed_forward.w1",
-    "mlp.down_proj": "feed_forward.w2",
-    "mlp.up_proj": "feed_forward.w3",
-    "input_layernorm": "attention_norm",
+    "self_attn.q_proj": "attn_q",
+    "self_attn.k_proj": "attn_k",
+    "self_attn.v_proj": "attn_v",
+    "self_attn.o_proj": "attn_output",
+    "mlp.gate_proj": "ffn_gate",
+    "mlp.down_proj": "ffn_down",
+    "mlp.up_proj": "ffn_up",
+    "input_layernorm": "attn_norm",
     "post_attention_layernorm": "ffn_norm",
-    # "norm": "norm",
-    # "embed_tokens": "tok_embeddings",
-    # "lm_head": "output",
 }
 
 
@@ -39,7 +40,7 @@ def translate_tensor_name(t: str) -> str:
             sys.exit(1)
 
         output_string = (
-            f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+            f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
         )
         return output_string
     else:
@@ -47,19 +48,21 @@ def translate_tensor_name(t: str) -> str:
         sys.exit(1)
 
 
-def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
     fout.write(struct.pack("i", params["r"]))
     # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
     # but some models ship a float value instead
     # let's convert to int, but fail if lossless conversion is not possible
-    assert int(params["lora_alpha"]) == params["lora_alpha"], "cannot convert float to int losslessly"
+    assert (
+        int(params["lora_alpha"]) == params["lora_alpha"]
+    ), "cannot convert float to int losslessly"
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
 def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: DataType
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(
@@ -67,7 +70,7 @@ def write_tensor_header(
             "iii",
             len(shape),
             len(sname),
-            DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]],
+            NUMPY_TYPE_TO_FTYPE[data_type.name],
         )
     )
     fout.write(struct.pack("i" * len(shape), *shape[::-1]))

convert-starcoder-hf-to-gguf.py

@@ -0,0 +1,248 @@
+#!/usr/bin/env python3
+# HF starcoder --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def bytes_to_unicode():
+    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
+    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.STARCODER
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("StarCoder")
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+    scores.append(0.0)                      # dymmy
+    toktypes.append(gguf.TokenType.NORMAL)  # dummy
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# params for qkv transform
+n_head    = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert.py

@@ -1,4 +1,6 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
+from __future__ import annotations
+
 import argparse
 import concurrent.futures
 import copy
@@ -15,141 +17,151 @@ import re
 import signal
 import struct
 import sys
+import time
 import zipfile
 from abc import ABCMeta, abstractmethod
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List,
-                    Literal, Optional, Sequence, Tuple, TypeVar, Union)
+from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
 
 import numpy as np
-from sentencepiece import SentencePieceProcessor  # type: ignore
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 if TYPE_CHECKING:
-    from typing_extensions import TypeAlias
+    from typing import TypeAlias
 
 if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
     faulthandler.register(signal.SIGUSR1)
 
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 
+ARCH=gguf.MODEL_ARCH.LLAMA
+NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
+
+DEFAULT_CONCURRENCY = 8
+#
+# data types
+#
 
 @dataclass(frozen=True)
-class UnquantizedDataType:
+class DataType:
     name: str
+    dtype: np.dtype[Any]
+    valid_conversions: list[str]
 
-
-DT_F16 = UnquantizedDataType('F16')
-DT_F32 = UnquantizedDataType('F32')
-DT_I32 = UnquantizedDataType('I32')
-DT_BF16 = UnquantizedDataType('BF16')
-
+    def elements_to_bytes(self, n_elements: int) -> int:
+        return n_elements * self.dtype.itemsize
 
 @dataclass(frozen=True)
-class QuantizedDataType:
-    groupsize: int
-    have_addends: bool
-    have_g_idx: bool
-
+class UnquantizedDataType(DataType):
+    pass
 
-DT_Q4_0 = QuantizedDataType(groupsize=32, have_addends=False, have_g_idx=False)
-DT_Q4_1 = QuantizedDataType(groupsize=32, have_addends=True, have_g_idx=False)
+DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
 
-DataType = Union[UnquantizedDataType, QuantizedDataType]
+@dataclass(frozen=True)
+class QuantizedDataType(DataType):
+    block_size: int
+    quantized_dtype: np.dtype[Any]
+    ggml_type: gguf.GGMLQuantizationType
 
-DATA_TYPE_TO_FTYPE: Dict[DataType, int] = {
-    DT_F32: 0,
-    DT_F16: 1,
-    DT_Q4_0: 2,
-    DT_Q4_1: 3,
-}
+    def quantize(self, arr: NDArray) -> NDArray:
+        raise NotImplementedError(f'Quantization for {self.name} not implemented')
 
-FTYPE_TO_DATA_TYPE: Dict[int, DataType] = \
-    {ftype: dtype for (dtype, ftype) in DATA_TYPE_TO_FTYPE.items()}
+    def elements_to_bytes(self, n_elements: int) -> int:
+        assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
+        return self.quantized_dtype.itemsize * (n_elements // self.block_size)
 
-DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = {
-    DT_BF16: np.dtype(np.uint16),
-    DT_F16: np.dtype(np.float16),
-    DT_F32: np.dtype(np.float32),
-    DT_I32: np.dtype(np.int32),
+@dataclass(frozen=True)
+class Q8_0QuantizedDataType(QuantizedDataType):
+    # Mini Q8_0 quantization in Python!
+    def quantize(self, arr: NDArray) -> NDArray:
+        assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
+        assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
+        n_blocks = arr.size // self.block_size
+        blocks = arr.reshape((n_blocks, self.block_size))
+        # Much faster implementation of block quantization contributed by @Cebtenzzre
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
+            d = abs(blocks).max(axis = 1) / np.float32(127)
+            with np.errstate(divide = 'ignore'):
+                qs = (blocks / d[:, None]).round()
+            qs[d == 0] = 0
+            yield from zip(d, qs)
+        return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
+
+DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
+    dtype = np.dtype(np.float32), valid_conversions = [],
+    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+
+# Quantized types skipped here because they may also map to np.float32
+NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
+for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
+    if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
+        raise ValueError(f'Invalid duplicate data type {dt}')
+    NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
+
+SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
+    'BF16': DT_BF16,
+    'F16': DT_F16,
+    'F32': DT_F32,
+    'I32': DT_I32,
 }
 
-NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \
-    {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()}
-
-
-class GGMLFileType(enum.Enum):
-    AllF32 = 0
-    MostlyF16 = 1  # except 1d tensors
-    MostlyQ4_0 = 2  # except 1d tensors
-    MostlyQ4_1 = 3  # except 1d tensors
-    PerLayerIsQ4_1 = 4  # but tok_embeddings.weight and output.weight are F16
-
-    def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType:
-        if len(tensor.shape) == 1:
-            # 1D tensors are always F32.
-            return DT_F32
-        elif self == GGMLFileType.AllF32:
-            return DT_F32
-        elif self == GGMLFileType.MostlyF16:
-            return DT_F16
-        elif self == GGMLFileType.MostlyQ4_0:
-            return DT_Q4_0
-        elif self == GGMLFileType.MostlyQ4_1:
-            return DT_Q4_1
-        elif self == GGMLFileType.PerLayerIsQ4_1:
-            if name in ('output.weight', 'tok_embeddings.weight'):
-                return DT_F16
-            else:
-                return DT_Q4_1
-        else:
+# TODO: match this with `llama_ftype`
+# TODO: rename to LLAMAFileType
+# TODO: move to `gguf.py`
+class GGMLFileType(enum.IntEnum):
+    AllF32     = 0
+    MostlyF16  = 1  # except 1d tensors
+    MostlyQ8_0 = 7  # except 1d tensors
+
+    def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType:
+        dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
+        if dt is None:
             raise ValueError(self)
+        # 1D tensors are always F32.
+        return dt if len(tensor.shape) > 1 else DT_F32
 
+GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
+    GGMLFileType.AllF32    : DT_F32,
+    GGMLFileType.MostlyF16 : DT_F16,
+    GGMLFileType.MostlyQ8_0: DT_Q8_0,
+}
 
-def make_tensors_list() -> List[str]:
-    ret = [
-        'tok_embeddings.weight',
-        'norm.weight',
-        'output.weight',
-    ]
-    for i in range(80):  # maximum number of layer
-        ret += [
-            f'layers.{i}.attention.wq.weight',
-            f'layers.{i}.attention.wk.weight',
-            f'layers.{i}.attention.wv.weight',
-            f'layers.{i}.attention.wo.weight',
-            f'layers.{i}.attention_norm.weight',
-            f'layers.{i}.feed_forward.w1.weight',
-            f'layers.{i}.feed_forward.w2.weight',
-            f'layers.{i}.feed_forward.w3.weight',
-            f'layers.{i}.ffn_norm.weight',
-        ]
-    return ret
+#
+# hparams loading
+#
 
+@dataclass
+class Params:
+    n_vocab:    int
+    n_embd:     int
+    n_layer:    int
+    n_ctx:      int
+    n_ff:       int
+    n_head:     int
+    n_head_kv:  int
+    f_norm_eps: float
 
-TENSORS_LIST = make_tensors_list()
-TENSORS_SET = set(TENSORS_LIST)
+    f_rope_freq_base: float | None = None
+    f_rope_scale: float | None = None
 
+    ftype: GGMLFileType | None = None
 
-def find_n_mult(n_ff: int, n_embd: int) -> int:
-    # hardcoded magic range
-    for n_mult in range(8192, 1, -1):
-        calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
-        if calc_ff == n_ff:
-            return n_mult
-    raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
-
-@dataclass
-class Params:
-    n_vocab:   int
-    n_embd:    int
-    n_mult:    int
-    n_head:    int
-    n_layer:   int
-    n_kv_head: Optional[int]  # This parameter is only used for Llama 2
+    # path to the directory containing the model files
+    path_model: Path | None = None
 
     @staticmethod
-    def guessed(model: 'LazyModel') -> 'Params':
+    def guessed(model: LazyModel) -> Params:
         # try transformer naming first
         n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
 
@@ -165,65 +177,110 @@ class Params:
             raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
-        n_head=n_embd // 128 # guessed
+        n_head = n_embd // 128 # guessed
+        n_mult = 256           # guessed
+
+        # TODO: verify this
+        n_ff = int(2 * (4 * n_embd) / 3)
+        n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult)
 
         return Params(
-            n_vocab   = n_vocab,
-            n_embd    = n_embd,
-            n_mult    = 256,
-            n_head    = n_head,
-            n_layer   = n_layer,
-            n_kv_head = None,
+            n_vocab    = n_vocab,
+            n_embd     = n_embd,
+            n_layer    = n_layer,
+            n_ctx      = -1,
+            n_ff       = n_ff,
+            n_head     = n_head,
+            n_head_kv  = n_head,
+            f_norm_eps = 1e-5,
         )
 
     @staticmethod
-    def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab = config["vocab_size"];
-        n_embd  = config["hidden_size"];
-        n_head  = config["num_attention_heads"];
-        n_layer = config["num_hidden_layers"];
-        n_ff    = config["intermediate_size"];
-        n_kv_head = config.get("num_key_value_heads")
+        n_vocab          = config["vocab_size"]
+        n_embd           = config["hidden_size"]
+        n_layer          = config["num_hidden_layers"]
+        n_ff             = config["intermediate_size"]
+        n_head           = config["num_attention_heads"]
+        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
+        f_norm_eps       = config["rms_norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
+
+        rope_scaling = config.get("rope_scaling")
+        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
+            f_rope_scale = config["rope_scaling"].get("factor")
+        else:
+            f_rope_scale = None
 
-        n_mult = find_n_mult(n_ff, n_embd);
+        if "max_sequence_length" in config:
+            n_ctx = config["max_sequence_length"]
+        elif "max_position_embeddings" in config:
+            n_ctx = config["max_position_embeddings"]
+        else:
+            raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
+                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
         return Params(
-            n_vocab   = n_vocab,
-            n_embd    = n_embd,
-            n_mult    = n_mult,
-            n_head    = n_head,
-            n_layer   = n_layer,
-            n_kv_head = n_kv_head,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
+            f_rope_scale     = f_rope_scale,
         )
 
     # LLaMA v2 70B params.json
-    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
     @staticmethod
-    def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab   = config["vocab_size"];
-        n_embd    = config["dim"];
-        n_head    = config["n_heads"];
-        n_layer   = config["n_layers"];
-        n_mult    = config["multiple_of"];
+        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
+        n_embd           = config["dim"]
+        n_layer          = config["n_layers"]
+        n_ff             = -1
+        n_head           = config["n_heads"]
+        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
+        f_norm_eps       = config["norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
+
+        # hack to determine LLaMA v1 vs v2 vs CodeLlama
+        if f_rope_freq_base == 1000000:
+            # CodeLlama
+            n_ctx = 16384
+        elif config["norm_eps"] == 1e-05:
+            # LLaMA v2
+            n_ctx = 4096
+        else:
+            # LLaMA v1
+            n_ctx = 2048
 
         if n_vocab == -1:
             n_vocab = model["tok_embeddings.weight"].shape[0]
 
+        if n_ff == -1:
+            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
+
         return Params(
-            n_vocab   = n_vocab,
-            n_embd    = n_embd,
-            n_mult    = n_mult,
-            n_head    = n_head,
-            n_layer   = n_layer,
-            n_kv_head = None,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
         )
 
     @staticmethod
-    def load(model_plus: 'ModelPlus') -> 'Params':
+    def load(model_plus: ModelPlus) -> Params:
         hf_config_path   = model_plus.paths[0].parent / "config.json"
         orig_config_path = model_plus.paths[0].parent / "params.json"
 
@@ -231,33 +288,104 @@ class Params:
             params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
         elif orig_config_path.exists():
             params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
-        else:
+        elif model_plus.format != 'none':
             params = Params.guessed(model_plus.model)
+        else:
+            raise ValueError('Cannot guess params when model format is none')
+
+        params.path_model = model_plus.paths[0].parent
 
-        print(f'params: n_vocab:{params.n_vocab} n_embd:{params.n_embd} n_mult:{params.n_mult} n_head:{params.n_head} n_layer:{params.n_layer}')
         return params
 
 
-class SentencePieceVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path], vocabtype: Optional[str]) -> None:
-        self.vocabtype = vocabtype
-        if self.vocabtype == "bpe":
-          self.sentencepiece_tokenizer = json.loads(open(str(fname_tokenizer)).read())
+#
+# vocab
+#
+
+class BpeVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
+        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
+        added_tokens: dict[str, int]
+        if fname_added_tokens is not None:
+            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
-          self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: Dict[str, int]
+            # Fall back to trying to find the added tokens in tokenizer.json
+            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
+            if not tokenizer_json_file.is_file():
+                added_tokens = {}
+            else:
+                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
+                added_tokens = dict(
+                    (item['content'], item['id'])
+                    for item in tokenizer_json.get('added_tokens', [])
+                    # Added tokens here can be duplicates of the main vocabulary.
+                    if item['content'] not in self.bpe_tokenizer )
+
+        vocab_size: int = len(self.bpe_tokenizer)
+        expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids      = sorted(added_tokens.values())
+        if expected_ids != actual_ids:
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
+
+        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_list    = [text for (text, idx) in items]
+        self.vocab_size_base: int = vocab_size
+        self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer      = fname_tokenizer
+        self.fname_added_tokens   = fname_added_tokens
+
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        tokenizer = self.bpe_tokenizer
+        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
+        byte_encoder = tokenization_gpt2.bytes_to_unicode()
+        byte_decoder = {v: k for k, v in byte_encoder.items()}
+        score = 0.0
+        for i, item in enumerate(tokenizer):
+            text: bytes = item.encode("utf-8")
+            # FIXME: These shouldn't be hardcoded, but it's probably better than the current behavior?
+            if i <= 258 and text.startswith(b'<') and text.endswith(b'>'):
+                if i == 0 and text == b'<unk>':
+                    toktype = gguf.TokenType.UNKNOWN
+                elif i == 1 or i == 2:
+                    toktype = gguf.TokenType.CONTROL
+                elif i >= 3 and text.startswith(b'<0x'):
+                    toktype = gguf.TokenType.BYTE
+                else:
+                    toktype = gguf.TokenType.NORMAL
+            else:
+                toktype = gguf.TokenType.NORMAL
+            yield text, score, toktype
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.bpe_tokens()
+        yield from self.added_tokens()
+
+    def __repr__(self) -> str:
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+class SentencePieceVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
+        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        added_tokens: dict[str, int]
         if fname_added_tokens is not None:
-            added_tokens = json.load(open(fname_added_tokens))
+            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
             added_tokens = {}
-        if self.vocabtype == "bpe":
-          vocab_size: int = len(self.sentencepiece_tokenizer)
-        else:
-          vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+
+        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
         expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids = sorted(added_tokens.values())
+        actual_ids   = sorted(added_tokens.values())
         if expected_ids != actual_ids:
             raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
+
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_list = [text for (text, idx) in items]
         self.vocab_size_base: int = vocab_size
@@ -265,126 +393,74 @@ class SentencePieceVocab:
         self.fname_tokenizer = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
-    def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float]]:
+    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.sentencepiece_tokenizer
-        if self.vocabtype == "bpe":
-          from transformers.models.gpt2 import tokenization_gpt2
-          byte_encoder = tokenization_gpt2.bytes_to_unicode()
-          byte_decoder = {v: k for k, v in byte_encoder.items()}
-          for i, item in enumerate(tokenizer):
-            text: bytes
-            text = b''.join([x.to_bytes(1, byteorder='big') for x in [byte_decoder[y] for y in item]])
-            score: float = -i
-            yield text, score
-        else:
-          for i in range(tokenizer.vocab_size()):
-              text: bytes
-              if tokenizer.is_unknown(i):
-                  text = " \u2047 ".encode("utf-8")
-              elif tokenizer.is_control(i):
-                  text = b""
-              elif tokenizer.is_byte(i):
-                  piece = tokenizer.id_to_piece(i)
-                  if len(piece) != 6:
-                      raise Exception(f"Invalid token: {piece}")
-                  byte_value = int(piece[3:-1], 16)
-                  text = struct.pack("B", byte_value)
-              else:
-                  text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
-              score: float = tokenizer.get_score(i)
-              yield text, score
-
-    def added_tokens(self) -> Iterable[Tuple[bytes, float]]:
+        for i in range(tokenizer.vocab_size()):
+            piece = tokenizer.id_to_piece(i)
+            text: bytes = piece.encode("utf-8")
+            score: float = tokenizer.get_score(i)
+
+            toktype = gguf.TokenType.NORMAL
+            if tokenizer.is_unknown(i):
+                toktype = gguf.TokenType.UNKNOWN
+            if tokenizer.is_control(i):
+                toktype = gguf.TokenType.CONTROL
+
+            # NOTE: I think added_tokens are user defined.
+            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
+            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
+
+            if tokenizer.is_unused(i):
+                toktype = gguf.TokenType.UNUSED
+            if tokenizer.is_byte(i):
+                toktype = gguf.TokenType.BYTE
+
+            yield text, score, toktype
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             score = -1000.0
-            yield text.encode("utf-8"), score
+            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float]]:
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.sentencepiece_tokens()
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
+Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
-class GGMLVocab:
-    def __init__(self, tokens: List[Tuple[bytes, float]]):
-        self.tokens = tokens
-        self.vocab_size = len(tokens)
+#
+# data loading
+# TODO: reuse (probably move to gguf.py?)
+#
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float]]:
-        return self.tokens
-
-    def __repr__(self) -> str:
-        return f"<GGMLVocab with {self.vocab_size} tokens>"
-
-
-Vocab = Union[SentencePieceVocab, GGMLVocab]
-
-
-def permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
-    if n_kv_head is not None and n_head != n_kv_head:
-        n_head //= n_kv_head
+def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
+    #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
+    if n_head_kv is not None and n_head != n_head_kv:
+        n_head //= n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                 .swapaxes(1, 2)
                 .reshape(weights.shape))
 
 
-def dequantize_q4(qvalues_pack32: NDArray, scales: NDArray, addends: Optional[NDArray], g_idx: Optional[NDArray]) -> NDArray:
-    # First reinterpret each row from a list of int32s containing 8 values each
-    # to a list of uint8s containing 2 values each.
-    qvalues_pack8 = qvalues_pack32.view(np.uint8)
-
-    # Then split out the two values per int8 (which requires an actual
-    # conversion because numpy doesn't natively support int4s).
-    qvalues = np.zeros([qvalues_pack8.shape[0], qvalues_pack8.shape[1] * 2], dtype=np.uint8)
-    qvalues[:, 0::2] = qvalues_pack8 & 0xf
-    qvalues[:, 1::2] = qvalues_pack8 >> 4
-
-    assert addends is None or addends.shape == scales.shape
-    assert qvalues.shape[0] == scales.shape[0]
-    assert qvalues.shape[1] % scales.shape[1] == 0
-    if g_idx is None:
-        repeat_count = qvalues.shape[1] // scales.shape[1]
-        scales = scales[:, :, np.newaxis]
-        if addends is not None:
-            addends = addends[:, :, np.newaxis]
-        # Reshape so that the below computation broadcasts over scales and addends:
-        qvalues.shape = (qvalues.shape[0], scales.shape[1], int(repeat_count))
-    else:
-        # In this case the scale and addend is selected for each column by g_idx:
-        assert addends is not None
-        scales = scales[:, g_idx]
-        addends = addends[:, g_idx]
-    if addends is None:
-        # Q4_0
-        qvalues = qvalues.view(np.int8)
-        qvalues -= 8
-    # And do the actual 'value = scale * qvalue + addend' computation.
-    values = scales * qvalues
-    if addends is not None:
-        values += addends
-    if g_idx is None:
-        values.shape = (values.shape[0], values.shape[1] * values.shape[2])
-    return values
-
-
 class Tensor(metaclass=ABCMeta):
     data_type: DataType
 
     @abstractmethod
-    def astype(self, data_type: DataType) -> 'Tensor': ...
+    def astype(self, data_type: DataType) -> Tensor: ...
     @abstractmethod
-    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'Tensor': ...
+    def permute(self, n_head: int, n_head_kv: int) -> Tensor: ...
     @abstractmethod
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def part(self, n_part: int) -> 'UnquantizedTensor': ...
+    def part(self, n_part: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def to_ggml(self) -> 'GGMLCompatibleTensor': ...
+    def to_ggml(self) -> GGMLCompatibleTensor: ...
 
 
-def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray:
+def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
     assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
     fp32_arr = bf16_arr.astype(np.uint32) << 16
     return fp32_arr.view(np.float32)
@@ -397,27 +473,27 @@ class UnquantizedTensor(Tensor):
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
 
     def astype(self, data_type: DataType) -> Tensor:
-        dtype = DATA_TYPE_TO_NUMPY[data_type]
+        dtype = data_type.dtype
         if self.data_type == DT_BF16:
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
 
-    def to_ggml(self) -> 'UnquantizedTensor':
+    def to_ggml(self) -> UnquantizedTensor:
         return self
 
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
 
-    def part(self, n_part: int) -> 'UnquantizedTensor':
+    def part(self, n_part: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'UnquantizedTensor':
-        return UnquantizedTensor(permute(self.ndarray, n_head, n_kv_head))
+    def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor:
+        return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv))
 
 
-def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
+def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray:
     tensor = lazy_tensor.load()
     assert isinstance(tensor, UnquantizedTensor)
 
@@ -433,196 +509,24 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv
     return tensor.ndarray
 
 
-class GGMLQuantizedTensor(Tensor):
-    data_type: QuantizedDataType
-
-    def __init__(self, ndarray: NDArray, shape: List[int], data_type: DataType) -> None:
-        rows, columns = shape
-        assert data_type in (DT_Q4_1, DT_Q4_0)  # for now
-        assert isinstance(data_type, QuantizedDataType)  # redundant, but mypy complains without this
-        assert columns % data_type.groupsize == 0
-        words_in_block = 6 if data_type == DT_Q4_1 else 5
-        self.ndarray = ndarray.view(dtype=np.uint32).reshape((rows, columns // data_type.groupsize, words_in_block))
-        self.shape = shape[:]
-        self.data_type = data_type
-
-    def astype(self, data_type: DataType) -> Tensor:
-        if data_type == self.data_type:
-            return self
-        scales = self.ndarray[:, :, 0].view(np.float32)
-        if self.data_type.have_addends:
-            addends = self.ndarray[:, :, 1].view(np.float32)
-        else:
-            addends = None
-        qweights = self.ndarray[:, :, -4:].reshape([self.shape[0], self.shape[1] // 8])
-
-        dq = dequantize_q4(qweights, scales, addends, g_idx=None)
-        return UnquantizedTensor(dq).astype(data_type)
-
-    def to_ggml(self) -> 'GGMLQuantizedTensor':
-        return self
-
-    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'GGMLQuantizedTensor':
-        return GGMLQuantizedTensor(permute(self.ndarray, n_head, n_kv_head), self.shape, self.data_type)
-
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
-        r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
-
-    def part(self, n_part: int) -> 'UnquantizedTensor':
-        r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
-
-GGMLCompatibleTensor = Union[UnquantizedTensor, GGMLQuantizedTensor]
-
-
-class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int, n_kv_head: Optional[int] = None) -> None:
-        self.base = base
-        self.n_head = n_head
-        self.n_kv_head = n_kv_head
-        self.data_type = self.base.data_type
-
-    def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head, self.n_kv_head)
-
-    def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head, self.n_kv_head)
-
-    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
-        raise Exception("shouldn't permute twice")
-
-
-class GPTQForLLaMaQuantizedTensor(Tensor):
-    def __init__(self, model: 'LazyModel', namebase: str) -> None:
-        qweight = load_unquantized(model[f"{namebase}.qweight"], np.int32)
-        scales = load_unquantized(model[f"{namebase}.scales"], np.float32, convert=True)
-
-        bias = model.get(f"{namebase}.bias")
-        if bias is not None:
-            # Q4_1 does not support bias; good thing the bias is always all zeros.
-            assert not np.any(load_unquantized(bias))
-
-        if f"{namebase}.zeros" in model:
-            zeros = load_unquantized(model[f"{namebase}.zeros"], np.float32)
-        else:
-            qzeros = load_unquantized(model[f"{namebase}.qzeros"], np.int32)
-            assert qzeros.dtype == np.int32
-            zeros = dequantize_q4(qzeros, scales, scales, g_idx=None)
-            assert zeros.dtype == np.float32
-
-        assert zeros.shape == scales.shape
-
-        # Output is transposed compared to the input, and addends have their sign flipped.
-        # Scales and zeros similarly must be transposed but only for newer
-        # versions of GPTQ-for-LLaMa; the older versions can be identified by
-        # having shape (n_embd, 1).
-        qweight = qweight.T
-        if scales.shape[1] != 1:
-            scales = scales.T
-            zeros = zeros.T
-
-        # Output also has signs flipped for the addends.
-        self.qweight = qweight
-        self.scales = scales
-        self.addends = -zeros
-
-        self.g_idx: Optional[NDArray]
-        if f"{namebase}.g_idx" in model:
-            self.g_idx = load_unquantized(model[f"{namebase}.g_idx"], np.int32)
-            assert self.g_idx.shape == (qweight.shape[1] * 8,)
-        else:
-            self.g_idx = None
-
-        self.shape = [self.qweight.shape[0], self.qweight.shape[1] * 8]
-        self.data_type = QuantizedDataType(groupsize=self.groupsize(), have_addends=True,
-                                           have_g_idx=(self.g_idx is not None))
-
-    def inspect(self, row: int, col: int) -> None:
-        '''For debugging.'''
-        qweight = (self.qweight[row, col // 8] >> (4 * (col & 7))) & 0xf
-        if self.g_idx is not None:
-            group = self.g_idx[col]
-        else:
-            group = int(col // self.groupsize())
-        scale = self.scales[row, group]
-        addend = self.addends[row, group]
-        with np.printoptions(precision=None, suppress=True):
-            print(f'scale:{scale} addend:{addend} qweight:{qweight}')
-            print('possible values:', np.arange(16) * scale + addend)
-            print('actual value:', qweight * scale + addend)
-
-    def astype(self, data_type: DataType) -> Tensor:
-        if isinstance(data_type, QuantizedDataType):
-            assert self.g_idx is None and data_type.have_addends is True and data_type.have_g_idx is False
-            return self.regroup(data_type.groupsize)
-
-        dequantized = dequantize_q4(np.ascontiguousarray(self.qweight), self.scales, self.addends, self.g_idx)
-        return UnquantizedTensor(dequantized).astype(data_type)
-
-    def groupsize(self) -> int:
-        assert self.addends.shape == self.scales.shape
-        assert self.shape[1] % self.scales.shape[1] == 0
-        return self.shape[1] // self.scales.shape[1]
-
-    def regroup(self, new_groupsize: int = 32) -> 'GPTQForLLaMaQuantizedTensor':
-        # Old versions of GPTQ-for-LLaMa shared scales and addends between all the
-        # columns in a row.  Newer versions share them between every set of N
-        # columns in a row, where N is the `groupsize` parameter, usually 128.  The
-        # output format shares them between every set of 32 columns.  To handle
-        # this, duplicate scales and addends for every smaller group.
-        # (In the above, 'row' and 'column' are in the sense of the output.)
-        assert self.g_idx is None
-        old_groupsize = self.groupsize()
-        assert old_groupsize >= new_groupsize and old_groupsize % new_groupsize == 0, old_groupsize
-        ret = copy.copy(self)
-        ret.addends = self.addends.repeat(old_groupsize // new_groupsize, axis=1)
-        ret.scales = self.scales.repeat(old_groupsize // new_groupsize, axis=1)
-        ret.data_type = QuantizedDataType(groupsize=new_groupsize, have_addends=True, have_g_idx=False)
-        return ret
-
-    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
-        return DeferredPermutedTensor(self, n_head, n_kv_head)
-
-    def to_ggml(self) -> GGMLQuantizedTensor:
-        # The output format looks like this:
-        # For each row:
-        #   For each group of 32 columns:
-        #     - addend (float32, 4 bytes)
-        #     - scale (float32, 4 bytes)
-        #     - weights (int4 * 32, 16 bytes)
-
-        if self.groupsize() != 32:
-            raise Exception("should have been regrouped before converting to ggml")
-
-        # Since the output format is mixed between integers and floats, we have
-        # to hackily view the floats as int32s just so numpy will let us
-        # concatenate them.
-        addends_view = self.addends.view(dtype=np.int32)[:, :, np.newaxis]
-        scales_view = self.scales.view(dtype=np.int32)[:, :, np.newaxis]
-
-        # Split into groups of 4 columns (i.e. 32 columns of quantized data):
-        grouped = self.qweight.reshape([self.qweight.shape[0], self.qweight.shape[1] // 4, 4])
-
-        # And concatenate:
-        grouped = np.concatenate([scales_view, addends_view, grouped], axis=2, casting='no')
-
-        return GGMLQuantizedTensor(grouped, self.shape, DT_Q4_1)
+GGMLCompatibleTensor = UnquantizedTensor
 
 
 @dataclass
 class LazyTensor:
     _load: Callable[[], Tensor]
-    shape: List[int]
+    shape: list[int]
     data_type: DataType
     description: str
 
     def load(self) -> Tensor:
         ret = self._load()
-        assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description)
+        # Should be okay if it maps to the same numpy type?
+        assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
+                (self.data_type, ret.data_type, self.description)
         return ret
 
-    def astype(self, data_type: DataType) -> 'LazyTensor':
+    def astype(self, data_type: DataType) -> LazyTensor:
         self.validate_conversion_to(data_type)
 
         def load() -> Tensor:
@@ -630,39 +534,28 @@ class LazyTensor:
         return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
 
     def validate_conversion_to(self, data_type: DataType) -> None:
-        if data_type == self.data_type:
-            return
-        if isinstance(data_type, QuantizedDataType):
-            if not isinstance(self.data_type, QuantizedDataType):
-                raise Exception(f"Can't turn an unquantized tensor into a quantized type ({data_type})")
-            if self.data_type.have_g_idx:
-                sys.stderr.write(
-                    "Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), "
-                    "which is not yet natively supported by GGML. "
-                    "For now you can still convert this model by passing `--outtype f16` to dequantize, "
-                    "but that will result in a much larger output file for no quality benefit.\n")
-                sys.exit(1)
-            assert not data_type.have_g_idx and self.data_type.have_addends and data_type.have_addends
+        if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
+            raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
 
 
-LazyModel = Dict[str, LazyTensor]
+LazyModel: TypeAlias = 'dict[str, LazyTensor]'
 
 
 @dataclass
 class ModelPlus:
     model: LazyModel
-    paths: List[Path]  # Where this was read from.
-    format: Literal['ggml', 'torch', 'safetensors']
-    vocab: Optional[Vocab]  # For GGML models (which have vocab built in), the vocab.
+    paths: list[Path]  # Where this was read from.
+    format: Literal['ggml', 'torch', 'safetensors', 'none']
+    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
-def merge_sharded(models: List[LazyModel]) -> LazyModel:
+def merge_sharded(models: list[LazyModel]) -> LazyModel:
     # Original LLaMA models have each file contain one part of each tensor.
     # Use a dict instead of a set to preserve order.
     names = {name: None for model in models for name in model}
 
     def convert(name: str) -> LazyTensor:
-        lazy_tensors: List[LazyTensor] = [model[name] for model in models]
+        lazy_tensors: list[LazyTensor] = [model[name] for model in models]
         if len(lazy_tensors) == 1:
             # only one file; don't go through this procedure since there might
             # be quantized tensors
@@ -690,7 +583,7 @@ def merge_sharded(models: List[LazyModel]) -> LazyModel:
     return {name: convert(name) for name in names}
 
 
-def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
+def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     formats = set(mp.format for mp in models_plus)
     assert len(formats) == 1, "different formats?"
     format = formats.pop()
@@ -713,17 +606,17 @@ def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
     return ModelPlus(model, paths, format, vocab)
 
 
-def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_kv_head: Optional[int] = None) -> LazyTensor:
+def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute(n_head, n_kv_head)
-    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_kv_head}) ' + lazy_tensor.description)
+        return lazy_tensor.load().permute(n_head, n_head_kv)
+    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute_part(n_part, n_head)
+        return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
     s = lazy_tensor.shape.copy()
     s[0] = s[0] // 3
-    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
@@ -732,66 +625,6 @@ def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     s[0] = s[0] // 3
     return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
 
-def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
-    out: LazyModel = {}
-    out["tok_embeddings.weight"] = model["model.embed_tokens.weight"]
-    out["norm.weight"] = model["model.norm.weight"]
-    out["output.weight"] = model["lm_head.weight"]
-
-    for i in itertools.count():
-        if f"model.layers.{i}.self_attn.q_proj.weight" in model:
-            out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
-            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_kv_head)
-            out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
-        elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
-            out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)
-            out[f"layers.{i}.attention.wk.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head)
-            out[f"layers.{i}.attention.wv.weight"] = part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
-        else:
-            break
-
-        out[f"layers.{i}.attention.wo.weight"] = model[f"model.layers.{i}.self_attn.o_proj.weight"]
-
-        out[f"layers.{i}.feed_forward.w1.weight"] = model[f"model.layers.{i}.mlp.gate_proj.weight"]
-        out[f"layers.{i}.feed_forward.w2.weight"] = model[f"model.layers.{i}.mlp.down_proj.weight"]
-        out[f"layers.{i}.feed_forward.w3.weight"] = model[f"model.layers.{i}.mlp.up_proj.weight"]
-
-        out[f"layers.{i}.attention_norm.weight"] = model[f"model.layers.{i}.input_layernorm.weight"]
-        out[f"layers.{i}.ffn_norm.weight"] = model[f"model.layers.{i}.post_attention_layernorm.weight"]
-    return out
-
-
-def handle_quantization(model: LazyModel) -> LazyModel:
-    '''Convert a model with entries for 'foo.qweight', 'foo.scales', etc.
-    (which resolve to UnquantizedTensors with the raw data) to one with entries
-    for 'foo.weight' (which resolve to QuantizedTensors).
-    '''
-    def convert(name: str) -> Tuple[str, LazyTensor]:
-        if name.endswith(".qweight"):
-            namebase = name.rsplit('.', 1)[0]
-            orig_name = namebase + ".weight"
-
-            lazy_tensor = model[name]
-            assert len(lazy_tensor.shape) == 2
-            real_shape = [lazy_tensor.shape[1], lazy_tensor.shape[0] * 8]
-
-            # Calculate type.  This replicates the logic in
-            # GPTQForLLaMaQuantizedTensor (which is executed when the modelis
-            # actually loaded).
-            lazy_scales = model[f"{namebase}.scales"]
-            scales_width = 1 if lazy_scales.shape[1] == 1 else lazy_scales.shape[0]
-            assert real_shape[1] % scales_width == 0
-            groupsize = real_shape[1] // scales_width
-            have_g_idx = f"{namebase}.g_idx" in model
-            data_type = QuantizedDataType(groupsize=groupsize, have_addends=True, have_g_idx=have_g_idx)
-
-            def load() -> Tensor:
-                return GPTQForLLaMaQuantizedTensor(model, namebase)
-
-            return (orig_name, LazyTensor(load, real_shape, data_type, '[quantized]'))
-        else:
-            return (name, model[name])
-    return dict(convert(name) for name in model)
 
 # Functionality that simulates `torch.load` but where individual tensors are
 # only loaded into memory on demand, not all at once.
@@ -824,13 +657,11 @@ class LazyUnpickler(pickle.Unpickler):
         assert isinstance(pid[1], LazyStorageKind)
         data_type = pid[1].data_type
         filename_stem = pid[2]
-        filename = self.data_base_path + '/' + filename_stem
+        filename = f'{self.data_base_path}/{filename_stem}'
         info = self.zip_file.getinfo(filename)
 
         def load(offset: int, elm_count: int) -> NDArray:
-            dtype = DATA_TYPE_TO_NUMPY.get(data_type)
-            if dtype is None:
-                raise Exception("tensor stored in unsupported format")
+            dtype = data_type.dtype
             fp = self.zip_file.open(info)
             fp.seek(offset * dtype.itemsize)
             size = elm_count * dtype.itemsize
@@ -840,9 +671,8 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
-    # @staticmethod
+    @staticmethod
     def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
-                               # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
 
@@ -852,13 +682,15 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'pickled storage_offset={storage_offset} in {storage.description}'
         return LazyTensor(load, list(size), storage.kind.data_type, description)
 
-    # @staticmethod
+    @staticmethod
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: Dict[Any, Any] = {
-        ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2,
-        ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2,
+    CLASSES: dict[tuple[str, str], Any] = {
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # the staticmethods have a __func__ attribute.
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
         ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
         ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
         ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
@@ -885,25 +717,17 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
     return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None)
 
 
-SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
-    'BF16': DT_BF16,
-    'F16': DT_F16,
-    'F32': DT_F32,
-    'I32': DT_I32,
-}
-
-
 def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
     header_size, = struct.unpack('<Q', fp.read(8))
-    header: Dict[str, Dict[str, Any]] = json.loads(fp.read(header_size))
+    header: dict[str, dict[str, Any]] = json.loads(fp.read(header_size))
     # Use mmap for the actual data to avoid race conditions with the file offset.
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
     byte_buf = mapped[8 + header_size:]
 
-    def convert(info: Dict[str, Any]) -> LazyTensor:
+    def convert(info: dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
-        numpy_dtype = DATA_TYPE_TO_NUMPY[data_type]
-        shape: List[int] = info['shape']
+        numpy_dtype = data_type.dtype
+        shape: list[int] = info['shape']
         begin, end = info['data_offsets']
         assert 0 <= begin <= end <= len(byte_buf)
         assert end - begin == math.prod(shape) * numpy_dtype.itemsize
@@ -924,84 +748,6 @@ def must_read(fp: IO[bytes], length: int) -> bytes:
     return ret
 
 
-def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
-    magic = must_read(fp, 4)[::-1]
-    if magic in (b'ggmf', b'ggjt'):
-        version, = struct.unpack("i", must_read(fp, 4))
-        assert version == 1
-    else:
-        assert magic == b'ggml'
-        version = None
-    n_vocab, n_embd, n_mult, n_head, n_layer, rot, file_type = struct.unpack('<7i', must_read(fp, 28))
-
-    tokens: List[Tuple[bytes, float]] = []
-    for i in range(n_vocab):
-        if i == 32000:
-            # HACK: GPT4All messed with the format without changing the magic
-            # number.  Specifically, they changed the vocab section to contain
-            # `n_vocab - 1` tokens instead of `n_vocab` (i.e. omitting the
-            # extra pad token).  Try to detect if we're reading a file like
-            # this.
-            orig_pos = fp.tell()
-            fp.seek(20, io.SEEK_CUR)
-            is_gpt4all = fp.read(21) == b'tok_embeddings.weight'
-            fp.seek(orig_pos)
-            if is_gpt4all:
-                break
-
-        length, = struct.unpack("i", must_read(fp, 4))
-        text = must_read(fp, length)
-        if magic != b'ggml':
-            score, = struct.unpack("f", must_read(fp, 4))
-            tokens.append((text, score))
-    vocab = GGMLVocab(tokens) if magic != b'ggml' else None
-
-    model: LazyModel = {}
-    # Use mmap for the actual data to avoid race conditions with the file offset.
-    off = fp.raw.tell()
-    mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
-    fp.raw.seek(off)  # needed on Windows
-
-    def read_tensor() -> None:  # this is a function so that variables captured in `load` don't change
-        shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
-        assert 0 <= shape_len <= 3
-        shape: List[int] = list(struct.unpack(f"{shape_len}i", must_read(fp, 4 * shape_len)))
-        shape = shape[::-1]
-        name = must_read(fp, name_len).decode('utf-8')
-        data_type = FTYPE_TO_DATA_TYPE[ftype]
-
-        if magic == b'ggjt':
-            fp.seek((fp.tell() + 31) & -32)
-
-        if data_type == DT_Q4_1:
-            # See GPTQForLLaMaQuantizedTensor.ggml_ndarray()
-            size = 24 * (shape[1] // 32) * shape[0]
-        elif data_type == DT_Q4_0:
-            size = 20 * (shape[1] // 32) * shape[0]
-        else:
-            numpy_dtype = DATA_TYPE_TO_NUMPY[data_type]
-            elm_count = math.prod(shape)
-            size = elm_count * numpy_dtype.itemsize
-        offset = fp.tell()
-        buf = mapped[offset:offset+size]
-        fp.seek(size, io.SEEK_CUR)
-
-        def load() -> Tensor:
-            if isinstance(data_type, QuantizedDataType):
-                ndarray = np.frombuffer(buf, dtype=np.uint32)
-                return GGMLQuantizedTensor(ndarray, shape, data_type)
-            else:
-                return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape))
-        description = f'ggml offset={offset} type={data_type} path={path}'
-        model[name] = LazyTensor(load, shape, data_type, description)
-
-    while fp.read(1) != b'':
-        fp.seek(-1, io.SEEK_CUR)
-        read_tensor()
-
-    return ModelPlus(model=model, paths=[path], format='ggml', vocab=vocab)
-
-
 @functools.lru_cache(maxsize=None)
 def lazy_load_file(path: Path) -> ModelPlus:
     fp = open(path, 'rb')
@@ -1010,9 +756,6 @@ def lazy_load_file(path: Path) -> ModelPlus:
     if first8[:2] == b'PK':
         # A zip file, i.e. PyTorch format
         return lazy_load_torch_file(fp, path)
-    elif first8[2:4] == b'gg':
-        # GGML format
-        return lazy_load_ggml_file(fp, path)
     elif struct.unpack('<Q', first8)[0] < 16 * 1024 * 1024:
         # Probably safetensors
         return lazy_load_safetensors_file(fp, path)
@@ -1023,28 +766,43 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
     output value buffered per thread.'''
-    with concurrent.futures.ThreadPoolExecutor() as executor:
-        futures: List[concurrent.futures.Future[Out]] = []
-        items_rev = list(iterable)[::-1]
-        for i in range(min(concurrency, len(items_rev))):
-            futures.append(executor.submit(func, items_rev.pop()))
+    if concurrency < 2:
+        yield from map(func, iterable)
+        # Not reached.
+    iterable = iter(iterable)
+    executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor]
+    if use_processpool_executor:
+        executor_class = ProcessPoolExecutor
+    else:
+        executor_class = ThreadPoolExecutor
+    with executor_class(max_workers = max_workers) as executor:
+        futures: list[concurrent.futures.Future[Out]] = []
+        done = False
+        for _ in range(concurrency):
+            try:
+                futures.append(executor.submit(func, next(iterable)))
+            except StopIteration:
+                done = True
+                break
+
         while futures:
             result = futures.pop(0).result()
-            if items_rev:
-                futures.append(executor.submit(func, items_rev.pop()))
+            while not done and len(futures) < concurrency:
+                try:
+                    futures.append(executor.submit(func, next(iterable)))
+                except StopIteration:
+                    done = True
+                    break
             yield result
 
-
 def check_vocab_size(params: Params, vocab: Vocab) -> None:
     if params.n_vocab != vocab.vocab_size:
-        # GGMLVocab comes from the same file as the model so shouldn't mismatch:
-        assert isinstance(vocab, SentencePieceVocab)
+        assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
         if params.n_vocab == vocab.vocab_size_base:
             print("Ignoring added_tokens.json since model matches vocab size without it.")
             vocab.added_tokens_list = []
@@ -1061,105 +819,200 @@ def check_vocab_size(params: Params, vocab: Vocab) -> None:
 
 class OutputFile:
     def __init__(self, fname_out: Path) -> None:
-        self.fout = open(fname_out, "wb")
-
-    def write_file_header(self, params: Params, file_type: GGMLFileType) -> None:
-        self.fout.write(b"ggjt"[::-1])  # magic
-        values = [
-            1,  # file version
-            params.n_vocab,
-            params.n_embd,
-            params.n_mult,
-            params.n_head,
-            params.n_layer,
-            params.n_embd // params.n_head,  # rot (obsolete)
-            file_type.value,
-        ]
-        self.fout.write(struct.pack("i" * len(values), *values))
-
-    def write_tensor_header(self, name: str, shape: Sequence[int], data_type: DataType) -> None:
-        sname = name.encode('utf-8')
-        self.fout.write(struct.pack("iii", len(shape), len(sname), DATA_TYPE_TO_FTYPE[data_type]))
-        self.fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-        self.fout.write(sname)
-        self.fout.seek((self.fout.tell() + 31) & -32)
-
-    def write_vocab(self, vocab: Vocab) -> None:
-        for text, score in vocab.all_tokens():
-            self.fout.write(struct.pack("i", len(text)))
-            self.fout.write(text)
-            self.fout.write(struct.pack("f", score))
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+    def add_meta_arch(self, params: Params) -> None:
+        name = "LLaMA"
+
+        # TODO: better logic to determine model name
+        if params.n_ctx == 4096:
+            name = "LLaMA v2"
+        elif params.path_model is not None:
+            name = str(params.path_model.parent).split('/')[-1]
+
+        self.gguf.add_name                (name)
+        self.gguf.add_context_length      (params.n_ctx)
+        self.gguf.add_embedding_length    (params.n_embd)
+        self.gguf.add_block_count         (params.n_layer)
+        self.gguf.add_feed_forward_length (params.n_ff)
+        self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
+        self.gguf.add_head_count          (params.n_head)
+        self.gguf.add_head_count_kv       (params.n_head_kv)
+        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
+
+        if params.f_rope_freq_base is not None:
+            self.gguf.add_rope_freq_base(params.f_rope_freq_base)
+
+        if params.f_rope_scale is not None:
+            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+
+        if params.ftype is not None:
+            self.gguf.add_file_type(params.ftype)
+
+    def add_meta_vocab(self, vocab: Vocab) -> None:
+        tokens = []
+        scores = []
+        toktypes = []
+        # NOTE: `all_tokens` returns the base vocabulary and added tokens
+        for text, score, toktype in vocab.all_tokens():
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
+        self.gguf.add_token_list(tokens)
+        self.gguf.add_token_scores(scores)
+        self.gguf.add_token_types(toktypes)
+
+    def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
+        svocab.add_to_gguf(self.gguf)
+
+    def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
+        n_elements = int(np.prod(tensor.shape))
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
+        data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype)
+
+    def write_meta(self) -> None:
+        self.gguf.write_header_to_file()
+        self.gguf.write_kv_data_to_file()
+
+    def write_tensor_info(self) -> None:
+        self.gguf.write_ti_data_to_file()
+
+    def close(self) -> None:
+        self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, vocab: Vocab) -> None:
-        of = OutputFile(fname_out)
-        params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0, n_head=1, n_layer=0)
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
+        check_vocab_size(params, vocab)
+
         of = OutputFile(fname_out)
-        of.write_file_header(params, file_type=GGMLFileType.AllF32)
-        of.write_vocab(vocab)
-        of.fout.close()
+
+        # meta data
+        of.add_meta_arch(params)
+        of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
+        of.write_meta()
+
+        of.close()
+
+    @staticmethod
+    def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]:
+        name, lazy_tensor = item
+        tensor = lazy_tensor.load().to_ggml()
+        return (lazy_tensor.data_type, tensor.ndarray)
 
     @staticmethod
-    def write_all(fname_out: Path, params: Params, file_type: GGMLFileType, model: LazyModel, vocab: Vocab) -> None:
+    def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray:
+        dt, arr = item
+        if not isinstance(dt, QuantizedDataType):
+            return arr
+        return dt.quantize(arr)
+
+    @staticmethod
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
+
         of = OutputFile(fname_out)
-        of.write_file_header(params, file_type)
-        print("Writing vocab...")
-        of.write_vocab(vocab)
 
-        def do_item(item: Tuple[str, LazyTensor]) -> NDArray:
-            name, lazy_tensor = item
-            return lazy_tensor.load().to_ggml().ndarray
+        # meta data
+        of.add_meta_arch(params)
+        of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
+        # tensor info
+        for name, lazy_tensor in model.items():
+            of.add_tensor_info(name, lazy_tensor)
+
+        of.write_meta()
+        of.write_tensor_info()
+
+        # tensor data
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True)
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
 
-        ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
+        start = time.time()
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
-            of.write_tensor_header(name, lazy_tensor.shape, lazy_tensor.data_type)
-            ndarray.tofile(of.fout)
-        of.fout.close()
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
+            of.gguf.write_tensor_data(ndarray)
 
+        of.close()
 
-def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType:
-    wq_type = model["layers.0.attention.wq.weight"].data_type
-    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
+def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
+    wq_type = model[NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
+
+    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
-    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
+    if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
         return GGMLFileType.MostlyF16
-    if output_type_str == "q4_1" or (output_type_str is None and isinstance(wq_type, QuantizedDataType) and
-                                     wq_type.have_addends):
-        if isinstance(model["output.weight"].data_type, QuantizedDataType):
-            return GGMLFileType.MostlyQ4_1
-        else:
-            return GGMLFileType.PerLayerIsQ4_1
-    if output_type_str == "q4_0" or (output_type_str is None and isinstance(wq_type, QuantizedDataType)):
-        return GGMLFileType.MostlyQ4_0
+    if output_type_str == "q8_0":
+        return GGMLFileType.MostlyQ8_0
+
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
+
     raise Exception(f"Unexpected combination of types: {name_to_type}")
 
+def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
+    return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
+            for (name, tensor) in model.items()}
 
-def do_necessary_conversions(model: LazyModel, params: Params) -> LazyModel:
-    model = handle_quantization(model)
+def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
+    tmap = gguf.TensorNameMap(ARCH, params.n_layer)
+    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
-    if "lm_head.weight" in model:
-        model = convert_transformers_to_orig(model, params)
-    model = filter_and_sort_tensors(model)
+    tmp = model
 
-    return model
+    # HF models permut or pack some of the tensors, so we need to undo that
+    for i in itertools.count():
+        if f"model.layers.{i}.self_attn.q_proj.weight" in model:
+            print(f"Permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
+           #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
+        elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
+            print(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+        else:
+            break
 
+    out: LazyModel = {}
+    for name, lazy_tensor in model.items():
+        tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
+        if name_new is None:
+            raise Exception(f"Unexpected tensor name: {name}")
 
-def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
-    return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
-            for (name, tensor) in model.items()}
+        if tensor_type in should_skip:
+            print(f"skipping tensor {name_new}")
+            continue
 
+        print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
+        out[name_new] = lazy_tensor
 
-def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
+    return out
+
+def nth_multifile_path(path: Path, n: int) -> Path | None:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns: List[Tuple[str, str]] = [
+    patterns: list[tuple[str, str]] = [
         # - x.00.pth, x.01.pth, etc.
         (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
         # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
@@ -1175,11 +1028,11 @@ def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
     return None
 
 
-def find_multifile_paths(path: Path) -> List[Path]:
+def find_multifile_paths(path: Path) -> list[Path]:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the whole list of paths in the model.
     '''
-    ret: List[Path] = []
+    ret: list[Path] = []
     for i in itertools.count():
         nth_path = nth_multifile_path(path, i)
         if nth_path is None:
@@ -1203,11 +1056,6 @@ def load_some_model(path: Path) -> ModelPlus:
             # Try the PyTorch patterns too, with lower priority
             globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
             files = [file for glob in globs for file in path.glob(glob)]
-        if not files:
-            # Try GGML too, but with lower priority, since if both a non-GGML
-            # model and a GGML model exist in the same directory, we assume the
-            # latter was converted from the former.
-            files = list(path.glob("ggml-model*.bin*"))
         if not files:
             raise Exception(f"Can't find model in directory {path}")
         if len(files) > 1:
@@ -1215,7 +1063,7 @@ def load_some_model(path: Path) -> ModelPlus:
         path = files[0]
 
     paths = find_multifile_paths(path)
-    models_plus: List[ModelPlus] = []
+    models_plus: list[ModelPlus] = []
     for path in paths:
         print(f"Loading model file {path}")
         models_plus.append(lazy_load_file(path))
@@ -1224,19 +1072,14 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-def filter_and_sort_tensors(model: LazyModel) -> LazyModel:
-    return {name: model[name] for name in TENSORS_LIST if name in model}
-
-
-def load_vocab(path: Path, vocabtype: Optional[str]) -> SentencePieceVocab:
-    print(f"vocabtype: {vocabtype}")
+def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
     # Be extra-friendly and accept either a file or a directory.  Also, if it's
     # a directory, it might be the model directory, and tokenizer.model might
     # be in the parent of that.
     if path.is_dir():
         vocab_file = "tokenizer.model"
         if vocabtype == 'bpe':
-          vocab_file = "vocab.json"
+            vocab_file = "vocab.json"
         path2 = path / vocab_file
         # Use `.parent` instead of /.. to handle the symlink case better.
         path3 = path.parent / vocab_file
@@ -1246,23 +1089,27 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> SentencePieceVocab:
             path = path3
         else:
             raise FileNotFoundError(
-                f"Could not find tokenizer.model in {path} or its parent; "
+                f"Could not find {vocab_file} in {path} or its parent; "
                 "if it's in another directory, pass the directory as --vocab-dir")
+
+    print(f"Loading vocab file '{path}', type '{vocabtype}'")
+
     added_tokens_path = path.parent / "added_tokens.json"
-    print(f"Loading vocab file {path}")
-    return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None,
-                              vocabtype)
+    if vocabtype == "bpe":
+        return BpeVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    elif vocabtype == "spm":
+        return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    else:
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
 
-def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
     namestr = {
-        GGMLFileType.AllF32: "f32",
+        GGMLFileType.AllF32:    "f32",
         GGMLFileType.MostlyF16: "f16",
-        GGMLFileType.MostlyQ4_0: "q4_0",
-        GGMLFileType.MostlyQ4_1: "q4_1",
-        GGMLFileType.PerLayerIsQ4_1: "q4_1",
+        GGMLFileType.MostlyQ8_0:"q8_0",
     }[file_type]
-    ret = model_paths[0].parent / f"ggml-model-{namestr}.bin"
+    ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
     if ret in model_paths:
         sys.stderr.write(
             f"Error: Default output path ({ret}) would overwrite the input. "
@@ -1279,47 +1126,82 @@ def do_dump_model(model_plus: ModelPlus) -> None:
         print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}")
 
 
-def main(args_in: Optional[List[str]] = None) -> None:
+def main(args_in: list[str] | None = None) -> None:
     parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
-    parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
-    parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)")
-    parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
-    parser.add_argument("model", type=Path,
-                        help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
-    parser.add_argument("--vocabtype", default='spm', choices=["spm", "bpe"], help="vocab format (default: spm)")
+    parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outtype",     choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
+    parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
     args = parser.parse_args(args_in)
 
-    vocab: Vocab
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
-    elif args.vocab_only:
-        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        return
+
+    if not args.vocab_only:
+        model_plus = load_some_model(args.model)
+    else:
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
+
+    if args.dump:
+        do_dump_model(model_plus)
+        return
+
+    params = Params.load(model_plus)
+    if params.n_ctx == -1:
+        if args.ctx is None:
+            raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n"
+                            "Please specify one with --ctx:\n"
+                            " - LLaMA v1: --ctx 2048\n"
+                            " - LLaMA v2: --ctx 4096\n")
+        params.n_ctx = args.ctx
+
+    if args.outtype:
+        params.ftype = {
+            "f32": GGMLFileType.AllF32,
+            "f16": GGMLFileType.MostlyF16,
+            "q8_0": GGMLFileType.MostlyQ8_0,
+        }[args.outtype]
+
+    print(f"params = {params}")
+
+    vocab: Vocab
+    if args.vocab_only:
         assert args.outfile, "need --outfile if using --vocab-only"
+        # FIXME: Try to respect vocab_dir somehow?
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
+        return
+
+    if model_plus.vocab is not None and args.vocab_dir is None:
+        vocab = model_plus.vocab
     else:
-        model_plus = load_some_model(args.model)
-        if args.dump:
-            do_dump_model(model_plus)
-            return
-        if model_plus.vocab is not None and args.vocab_dir is None:
-            vocab = model_plus.vocab
-        else:
-            vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-            vocab = load_vocab(vocab_dir, args.vocabtype)
-        params = Params.load(model_plus)
-        model = model_plus.model
-        model = do_necessary_conversions(model, params)
-        output_type = pick_output_type(model, args.outtype)
-        model = convert_to_output_type(model, output_type)
-        outfile = args.outfile or default_outfile(model_plus.paths, output_type)
-        OutputFile.write_all(outfile, params, output_type, model, vocab)
-        print(f"Wrote {outfile}")
+        vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
+        vocab = load_vocab(vocab_dir, args.vocabtype)
+    # FIXME: Try to respect vocab_dir somehow?
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+
+    model   = model_plus.model
+    model   = convert_model_names(model, params)
+    ftype   = pick_output_type(model, args.outtype)
+    model   = convert_to_output_type(model, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+
+    params.ftype = ftype
+    print(f"Writing {outfile}, format {ftype}")
+
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    print(f"Wrote {outfile}")
 
 
 if __name__ == '__main__':

docs/token_generation_performance_tips.md

@@ -3,7 +3,7 @@
 ## Verifying that the model is running on the GPU with cuBLAS
 Make sure you compiled llama with the correct env variables according to [this guide](../README.md#cublas), 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
-./main -m "path/to/model.bin" -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:
@@ -25,9 +25,9 @@ GPU: A6000 (48GB VRAM)
 CPU: 7 physical cores
 RAM: 32GB
 
-Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.ggmlv3.q4_0.bin` (30B parameters, 4bit quantization, GGML)
+Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.q4_0.gguf` (30B parameters, 4bit quantization, GGML)
 
-Run command: `./main -m "path/to/model.bin" -p "-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

@@ -6,27 +6,6 @@ find_package(Threads REQUIRED)
 
 # ...
 
-# common
-
-set(TARGET common)
-
-add_library(${TARGET} OBJECT
-    common.h
-    common.cpp
-    console.h
-    console.cpp
-    grammar-parser.h
-    grammar-parser.cpp
-    )
-
-if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-endif()
-
-target_include_directories(${TARGET} PUBLIC .)
-target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
-
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -42,8 +21,12 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
+    add_subdirectory(speculative)
     add_subdirectory(embd-input)
+    add_subdirectory(llama-bench)
+    add_subdirectory(beam-search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -9,12 +9,12 @@
 #endif
 
 #ifdef LLAMA_DEFAULT_RMS_EPS
-static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
+constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
 #else
-static const float rms_norm_eps = 5e-6f;
+constexpr float rms_norm_eps = 5e-6f;
 #endif
 
-float frand() {
+static float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
 
@@ -25,19 +25,21 @@ struct random_normal_distribution {
     float max;
 };
 
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
+static void init_random_normal_distribution(
+    struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max
+) {
     rnd->gen = std::mt19937(seed);
     rnd->nd = std::normal_distribution<float>{mean, std};
     rnd->min = min;
     rnd->max = max;
 }
 
-float frand_normal(struct random_normal_distribution * rnd) {
+static float frand_normal(struct random_normal_distribution * rnd) {
     const float r = rnd->nd(rnd->gen);
     return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
 }
 
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -48,13 +50,9 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-struct ggml_tensor * randomize_tensor(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        float fmin,
-        float fmax) {
-
+static struct ggml_tensor * randomize_tensor(
+    struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax
+) {
     switch (ndims) {
         case 1:
             for (int i0 = 0; i0 < ne[0]; i0++) {
@@ -95,11 +93,9 @@ struct ggml_tensor * randomize_tensor(
     return tensor;
 }
 
-struct ggml_tensor * randomize_tensor_normal(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        struct random_normal_distribution * rnd) {
+static struct ggml_tensor * randomize_tensor_normal(
+    struct ggml_tensor * tensor, int ndims, const int64_t ne[], struct random_normal_distribution * rnd
+) {
     float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
@@ -159,7 +155,7 @@ struct llama_hparams {
     }
 };
 
-uint32_t get_n_ff(const struct llama_hparams* hparams) {
+static uint32_t get_n_ff(const struct llama_hparams* hparams) {
     const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
     return n_ff;
 }
@@ -260,7 +256,7 @@ struct llama_model_lora {
     std::vector<llama_layer_lora> layers;
 };
 
-void init_model(struct llama_model * model) {
+static void init_model(struct llama_model * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -297,7 +293,7 @@ void init_model(struct llama_model * model) {
 }
 
 
-void init_model_lora(struct llama_model_lora * model) {
+static void init_model_lora(struct llama_model_lora * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -340,7 +336,7 @@ void init_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void set_param_model(struct llama_model * model) {
+static void set_param_model(struct llama_model * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -366,7 +362,7 @@ void set_param_model(struct llama_model * model) {
     }
 }
 
-void set_param_model_lora(struct llama_model_lora * model) {
+static void set_param_model_lora(struct llama_model_lora * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -397,7 +393,7 @@ void set_param_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -426,7 +422,9 @@ void randomize_model(struct llama_model * model, int seed, float mean, float std
 }
 
 
-void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model_lora(
+    struct llama_model_lora * model, int seed, float mean, float std, float min, float max
+) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -459,7 +457,7 @@ void randomize_model_lora(struct llama_model_lora * model, int seed, float mean,
     }
 }
 
-bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_ctx   = hparams.n_ctx;
@@ -495,7 +493,7 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int
     return true;
 }
 
-bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
+static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_ctx   = hparams.n_ctx;
@@ -531,15 +529,15 @@ bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora *
     return true;
 }
 
-struct ggml_tensor * forward(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past) {
-
+static struct ggml_tensor * forward(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -756,25 +754,25 @@ struct ggml_tensor * forward(
     return inpL;
 }
 
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+static void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
     GGML_ASSERT(tensor->n_dims == 1);
     GGML_ASSERT(tensor->ne[0] == ne0);
 }
 
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
+static void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
     GGML_ASSERT(tensor->n_dims == 2);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
 }
 
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
+static void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
     GGML_ASSERT(tensor->n_dims == 3);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
 }
 
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+static void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
     GGML_ASSERT(tensor->n_dims == 4);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
@@ -782,16 +780,16 @@ void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int6
     GGML_ASSERT(tensor->ne[3] == ne3);
 }
 
-struct ggml_tensor * forward_batch(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past,
-        const  int              n_batch) {
-
+static struct ggml_tensor * forward_batch(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past,
+    const  int              n_batch
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -1073,16 +1071,15 @@ struct ggml_tensor * forward_batch(
     return inpL;
 }
 
-
-struct ggml_tensor * forward_lora(
-        struct llama_model_lora * model,
-        struct llama_kv_cache   * cache,
-        struct ggml_context     * ctx0,
-        struct ggml_cgraph      * gf,
-        struct ggml_tensor      * tokens_input,
-        const  int                n_tokens,
-        const  int                n_past) {
-
+static struct ggml_tensor * forward_lora(
+    struct llama_model_lora * model,
+    struct llama_kv_cache   * cache,
+    struct ggml_context     * ctx0,
+    struct ggml_cgraph      * gf,
+    struct ggml_tensor      * tokens_input,
+    const  int                n_tokens,
+    const  int                n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -1328,7 +1325,7 @@ struct ggml_tensor * forward_lora(
     return inpL;
 }
 
-void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
     assert(logits->n_dims == 2);
     assert(probs->n_dims == 2);
     assert(best_samples->n_dims == 1);
@@ -1359,7 +1356,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str
     }
 }
 
-void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax_batch(
+    struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
+    struct ggml_tensor * best_samples
+) {
     GGML_ASSERT(best_samples->n_dims == 2);
     GGML_ASSERT(logits->n_dims == 3);
     GGML_ASSERT(probs->n_dims == 3);
@@ -1393,7 +1393,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits
     }
 }
 
-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
         printf(" %.2f", p);
@@ -1401,7 +1401,7 @@ void print_row(struct ggml_tensor * probs, int i) {
     printf("\n");
 }
 
-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
     assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
@@ -1412,7 +1412,7 @@ void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-void print_token(int token, int n_vocab) {
+static void print_token(int token, int n_vocab) {
     for (int k = 0; k < token; ++k) {
         printf(" ");
     }
@@ -1423,14 +1423,14 @@ void print_token(int token, int n_vocab) {
     printf("\n");
 }
 
-void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
+static void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
     for (int i=0; i<tokens->ne[0]; ++i) {
         int token = ggml_get_i32_1d(tokens, i);
         print_token(token, n_vocab);
     }
 }
 
-void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     float randomness = 0.0f;
@@ -1451,7 +1451,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
     }
 }
 
-void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets_batch(
+    struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
+) {
     GGML_ASSERT(tokens_input->n_dims == 2);
     GGML_ASSERT(     targets->n_dims == 3);
     int n_tokens = tokens_input->ne[0];
@@ -1474,7 +1476,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct
     }
 }
 
-void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
+static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     for (int i=0; i<n_tokens-n_shift; ++i) {
@@ -1485,12 +1487,16 @@ void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * tar
     }
 }
 
-struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * square_error_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     // todo: instead of a-b: a[1:]-b[:-1]
     return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, a, b)));
 }
 
-struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * cross_entropy_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     const float eps = 1e-3f;
     return
         ggml_sum(ctx,
@@ -1617,15 +1623,10 @@ int main(int argc, char ** argv) {
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
         struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
-        opt_params_adam.print_forward_graph = false;
-        opt_params_adam.print_backward_graph = false;
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
-        opt_params_adam.adam.n_iter = 16;
         opt_params_lbfgs.lbfgs.n_iter = 16;
-        // ggml_opt(ctx0, opt_params_adam, e);
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
         ggml_build_forward_expand(&gf, e);

examples/beam-search/CMakeLists.txt

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

examples/beam-search/beam-search.cpp

@@ -0,0 +1,186 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+// Used for debugging to print out beam tokens.
+struct ostream_beam_view {
+    llama_context * ctx;
+    llama_beam_view beam_view;
+};
+
+static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
+    os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
+    for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
+        os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
+    }
+    return os << ')';
+}
+
+// Put here anything you want back in beam_search_callback().
+struct beam_search_callback_data {
+    llama_context * ctx;
+    std::vector<llama_token> response;
+};
+
+// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
+// For example, eob can be flagged due to maximum token length, stop words, etc.
+static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+    auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        callback_data.response.resize(callback_data.response.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        std::copy(tokens, tokens + n, callback_data.response.end() - n);
+        printf("%zu", n);
+    }
+    fflush(stdout);
+#if 1 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n";
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{callback_data.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+    //params.n_gpu_layers = 200;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc < 2 || argv[1][0] == '-' )
+    {
+        printf( "Usage: %s MODEL_PATH [BEAM_WIDTH=2] [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    params.model = argv[1];
+
+    params.n_beams = 2 < argc ? std::stoi(argv[2]) : 2;
+
+    if ( argc > 3 )
+    {
+        params.prompt = argv[3];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+
+    if ( model == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
+
+    const size_t max_context_size     = llama_n_ctx( ctx );
+    const size_t max_tokens_list_size = max_context_size - 4 ;
+
+    if (tokens_list.size() > max_tokens_list_size)
+    {
+        fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" ,
+             __func__ , tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        std::cout << llama_token_to_piece(ctx, id);
+    }
+    std::cout << std::flush;
+
+    int n_past = llama_get_kv_cache_token_count(ctx);
+    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    {
+        fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
+        return 1;
+    }
+    n_past += tokens_list.size();
+
+    beam_search_callback_data callback_data{ctx, {}};
+    size_t const beam_width = static_cast<size_t>(params.n_beams);
+    int const n_predict = 256;
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
+
+    std::cout << "\n\n";
+    for (llama_token const token_id : callback_data.response) {
+        std::cout << llama_token_to_piece(ctx,token_id);
+    }
+    std::cout << std::endl;
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    llama_backend_free();
+
+    return 0;
+}

examples/benchmark/CMakeLists.txt

@@ -1,7 +1,8 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/benchmark/benchmark-matmult.cpp

@@ -1,5 +1,6 @@
-#include "ggml.h"
 #include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 
 #include <locale.h>
 #include <assert.h>
@@ -20,7 +21,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -31,19 +32,19 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-float tensor_sum_elements(const ggml_tensor * tensor) {
-    float sum = 0;
-    if (tensor->type==GGML_TYPE_F32) {
+static float tensor_sum_elements(const ggml_tensor * tensor) {
+    double sum = 0;
+    if (tensor->type == GGML_TYPE_F32) {
         for (int j = 0; j < tensor->ne[1]; j++) {
             for (int k = 0; k < tensor->ne[0]; k++) {
-                sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];
+                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
             }
         }
     }
     return sum;
 }
 
-void tensor_dump(const ggml_tensor * tensor, const char * name) {
+static void tensor_dump(const ggml_tensor * tensor, const char * name) {
     printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
         tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
@@ -58,7 +59,7 @@ struct benchmark_params_struct {
     int32_t n_iterations  = 10;
 };
 
-void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
+static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
@@ -99,7 +100,7 @@ int main(int argc, char ** argv)  {
         exit(1);
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
     printf("Starting Test\n");
 
     // create the ggml context
@@ -125,12 +126,15 @@ int main(int argc, char ** argv)  {
 
     //printf("Memsize required = %i\n", sizex*sizex);
 
+    // TODO: perform the bench for all types or for a user specified type
+    const ggml_type qtype = GGML_TYPE_Q4_1;
+
     size_t ctx_size = 0;
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
@@ -163,7 +167,7 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
     ggml_set_f32(m2, 2.0f);
 
-    printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
     // printf("Creating new tensor m11xm2\n");
     struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
 
@@ -181,17 +185,16 @@ int main(int argc, char ** argv)  {
 
     TENSOR_DUMP(gf.nodes[0]);
 
-    printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
     int32_t nelements = sizex*sizey;
-    int32_t ne[2] = { sizex, sizey };
 
     std::vector<int64_t> hist_cur(1 << 4, 0);
 
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -202,8 +205,8 @@ int main(int argc, char ** argv)  {
 
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
@@ -220,7 +223,7 @@ int main(int argc, char ** argv)  {
     printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
 
 
-    // Let's use the F32 result from above as a reference for the q4_0 multiplication
+    // Let's use the F32 result from above as a reference for the quantized multiplication
     float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
@@ -250,7 +253,7 @@ int main(int argc, char ** argv)  {
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
         float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
-        float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
+        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 
         if (delta > allowed_delta)  {

examples/chat.sh

@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./main -m ./models/7B/ggml-model-q4_0.bin -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

@@ -0,0 +1,5 @@
+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 ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

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

@@ -0,0 +1,26 @@
+## Convert llama2.c model to ggml
+
+This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
+
+To convert the model first download the models from the [llma2.c](https://github.com/karpathy/llama2.c) repository:
+
+`$ make -j`
+
+After successful compilation, following usage options are available:
+```
+usage: ./convert-llama2c-to-ggml [options]
+
+options:
+  -h, --help                       show this help message and exit
+  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default 'models/7B/ggml-model-f16.gguf')
+  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model
+  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default ak_llama_model.bin')
+```
+
+An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
+
+`$ ./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`
+
+Now you can use the model with a command like:
+
+`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

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

@@ -0,0 +1,963 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+
+#include <unordered_map>
+#include <vector>
+#include <cassert>
+#include <climits>
+#include <cstring>
+#include <cstdarg>
+#include <ctime>
+#include <random>
+#include <stdexcept>
+#include <sstream>
+#include <algorithm>
+#include <string>
+
+// GGUF keys & tensor names.
+
+#define KV_GENERAL_ARCHITECTURE          "general.architecture"
+#define KV_GENERAL_NAME                  "general.name"
+
+#define KV_TOKENIZER_MODEL               "tokenizer.ggml.model"
+#define KV_TOKENIZER_LIST                "tokenizer.ggml.tokens"
+#define KV_TOKENIZER_TOKEN_TYPE          "tokenizer.ggml.token_type"
+#define KV_TOKENIZER_SCORES              "tokenizer.ggml.scores"
+#define KV_TOKENIZER_BOS_ID              "tokenizer.ggml.bos_token_id"
+#define KV_TOKENIZER_EOS_ID              "tokenizer.ggml.eos_token_id"
+#define KV_TOKENIZER_UNK_ID              "tokenizer.ggml.unknown_token_id"
+#define KV_TOKENIZER_SEP_ID              "tokenizer.ggml.seperator_token_id"
+#define KV_TOKENIZER_PAD_ID              "tokenizer.ggml.padding_token_id"
+#define KV_TOKENIZER_HF_JSON             "tokenizer.huggingface.json"
+
+#define KV_CONTEXT_LENGTH                "llama.context_length"
+#define KV_EMBEDDING_LENGTH              "llama.embedding_length"
+#define KV_BLOCK_COUNT                   "llama.block_count"
+#define KV_FEED_FORWARD_LENGTH           "llama.feed_forward_length"
+#define KV_ATTENTION_HEAD_COUNT          "llama.attention.head_count"
+#define KV_ATTENTION_HEAD_COUNT_KV       "llama.attention.head_count_kv"
+#define KV_ATTENTION_LAYERNORM_RMS_EPS   "llama.attention.layer_norm_rms_epsilon"
+#define KV_ROPE_DIMENSION_COUNT          "llama.rope.dimension_count"
+
+#define TN_TOKEN_EMBD  "token_embd.weight"
+#define TN_OUTPUT_NORM "output_norm.weight"
+#define TN_OUTPUT      "output.weight"
+#define TN_ATTN_NORM   "blk.%d.attn_norm.weight"
+#define TN_ATTN_Q      "blk.%d.attn_q.weight"
+#define TN_ATTN_K      "blk.%d.attn_k.weight"
+#define TN_ATTN_V      "blk.%d.attn_v.weight"
+#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
+#define TN_FFN_NORM    "blk.%d.ffn_norm.weight"
+#define TN_FFN_GATE    "blk.%d.ffn_gate.weight"
+#define TN_FFN_DOWN    "blk.%d.ffn_down.weight"
+#define TN_FFN_UP      "blk.%d.ffn_up.weight"
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_VERSION_GGJT_V3   3
+
+#define TOKENIZER_NAME "llama"
+#define UNKNOWN_TOKEN_ID 0
+#define BOS_TOKEN_ID 1
+#define EOS_TOKEN_ID 2
+
+//////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
+typedef struct {
+    int dim; // transformer dimension
+    int hidden_dim; // for ffn layers
+    int n_layers; // number of layers
+    int n_heads; // number of query heads
+    int n_kv_heads; // number of key/value heads (can be < query heads because of multiquery)
+    int vocab_size; // vocabulary size, usually 256 (byte-level)
+    int seq_len; // max sequence length
+} Config;
+
+struct TransformerWeights {
+    // token embedding table
+    float* token_embedding_table;    // (vocab_size, dim)
+    // weights for rmsnorms
+    float* rms_att_weight; // (layer, dim) rmsnorm weights
+    float* rms_ffn_weight; // (layer, dim)
+    // weights for matmuls
+    float* wq; // (layer, dim, dim)
+    float* wk; // (layer, dim, dim)
+    float* wv; // (layer, dim, dim)
+    float* wo; // (layer, dim, dim)
+    // weights for ffn
+    float* w1; // (layer, hidden_dim, dim)
+    float* w2; // (layer, dim, hidden_dim)
+    float* w3; // (layer, hidden_dim, dim)
+    // final rmsnorm
+    float* rms_final_weight; // (dim,)
+    // freq_cis for RoPE relatively positional embeddings
+    // float* freq_cis_real; // (seq_len, dim/2)
+    // float* freq_cis_imag; // (seq_len, dim/2)
+    // (optional) classifier weights for the logits, on the last layer
+    float* wcls;
+
+    ~TransformerWeights() {
+        delete[] token_embedding_table;
+        delete[] rms_att_weight;
+        delete[] rms_ffn_weight;
+        delete[] wq;
+        delete[] wk;
+        delete[] wv;
+        delete[] wo;
+        delete[] w1;
+        delete[] w2;
+        delete[] w3;
+        delete[] rms_final_weight;
+        delete[] wcls;
+    }
+};
+
+static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
+    // we calloc instead of malloc to keep valgrind happy
+    w->token_embedding_table = new float[p->vocab_size * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+
+    w->rms_att_weight = new float[p->n_layers * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+
+    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+
+    w->wq = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wk = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wv = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->wo = new float[p->n_layers * p->dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+
+    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
+    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+
+    w->rms_final_weight = new float[p->dim]();
+    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+
+    if (shared_weights) {
+        w->wcls = NULL;
+    } else {
+        w->wcls = new float[p->vocab_size * p->dim]();
+        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+    }
+}
+
+static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
+    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
+    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
+    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
+    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
+    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
+    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
+    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+
+    // Skip freq_cis_real & freq_cis_imag
+    int head_size = p->dim / p->n_heads;
+    fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
+
+    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+
+    // Check we didn't forget to read anything
+    auto curr = ftell(f);
+    fseek(f, 0, SEEK_END);
+    auto end = ftell(f);
+    if (curr != end) {
+        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        return 1;
+    }
+
+    return 0;
+}
+
+static void print_sample_weights(TransformerWeights *w){
+    printf("----- Quick print of first of the weight vales of all the variables\n");
+    printf("%f\n", w->token_embedding_table[0]);
+    printf("%f\n", w->rms_att_weight[0]);
+    printf("%f\n", w->rms_ffn_weight[0]);
+
+    printf("%f\n", w->wq[0]);
+    printf("%f\n", w->wk[0]);
+    printf("%f\n", w->wv[0]);
+    printf("%f\n", w->wo[0]);
+    printf("%f\n", w->w1[0]);
+    printf("%f\n", w->w2[0]);
+    printf("%f\n", w->w3[0]);
+    printf("%f\n", w->rms_att_weight[0]);
+    if (w->wcls) printf("%f\n", w->wcls[0]);
+}
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+//////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
+
+struct llama_vocab {
+    using id    = int32_t;
+    using token = std::string;
+    using ttype = llama_token_type;
+
+    struct token_data {
+        token text;
+        float score;
+        ttype type;
+    };
+
+    std::unordered_map<token, id> token_to_id;
+    std::vector<token_data> id_to_token;
+};
+
+struct my_llama_hparams {
+    uint32_t n_vocab = 32000;
+    uint32_t n_ctx   = 512;   // this is provided as user input?
+    uint32_t n_embd  = 4096;
+    uint32_t n_ff    = 11008;
+    uint32_t n_mult  = 4;
+    uint32_t n_head  = 32;
+    uint32_t n_layer = 32;
+    uint32_t n_rot   = 64;
+    bool operator!=(const my_llama_hparams& other) const {
+        return memcmp(this, &other, sizeof(my_llama_hparams));
+    }
+};
+
+struct my_llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct my_llama_model {
+    struct ggml_context * ctx = NULL;
+
+    std::string name;
+
+    my_llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<my_llama_layer> layers;
+
+    uint32_t train_its = 0;
+    uint32_t train_samples = 0;
+    uint32_t train_tokens = 0;
+};
+
+struct train_params {
+    const char * fn_vocab_model;
+    const char * fn_llama2c_model;
+    const char * fn_llama2c_output_model;
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * fn_model_out;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_embd;
+    int n_mult;
+    int n_head;
+    int n_layer;
+    int n_rotmax;
+
+    int n_threads;
+    int n_batch;
+    int n_examples;
+    int n_predict;
+
+    int print_info_interval;
+    int print_details_interval;
+
+    bool samples_start_after_nl;
+    bool use_adam;
+    bool use_flash;
+    bool use_scratch;
+
+    // only adam
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_alpha;
+
+    int   lbfgs_n_iter;
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_decay;
+
+    int mem_model_gb;
+    int mem_compute_gb;
+    int mem_compute0_gb;
+    int mem_compute1_gb;
+};
+
+static void print_params(struct my_llama_hparams * params) {
+    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
+    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
+    printf("%s: n_head:  %d\n", __func__, params->n_head);
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
+    printf("%s: n_layer: %d\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+}
+
+static void init_model(struct my_llama_model * model) {
+    const auto & hparams = model->hparams;
+
+    const uint32_t n_embd  = hparams.n_embd;
+    const uint32_t n_layer = hparams.n_layer;
+    const uint32_t n_vocab = hparams.n_vocab;
+
+    const uint32_t n_ff = hparams.n_ff;
+    struct ggml_context * ctx = model->ctx;
+
+    model->train_its = 0;
+    model->train_samples = 0;
+    model->train_tokens = 0;
+
+    model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
+
+    model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
+
+    model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
+
+    // printing the per-layer allocations here so we dont print in the for loop.
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+
+    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
+
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+
+    ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
+    ggml_set_name(model->norm,           "norm.weight");
+    ggml_set_name(model->output,         "output.weight");
+
+    model->layers.resize(n_layer);
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        std::string layers_i = "layers." + std::to_string(i);
+
+        layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+
+        layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
+        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+
+        ggml_set_name(layer.attention_norm, (layers_i + ".attention_norm.weight").c_str());
+
+        ggml_set_name(layer.wq, (layers_i + ".attention.wq.weight").c_str());
+        ggml_set_name(layer.wk, (layers_i + ".attention.wk.weight").c_str());
+        ggml_set_name(layer.wv, (layers_i + ".attention.wv.weight").c_str());
+        ggml_set_name(layer.wo, (layers_i + ".attention.wo.weight").c_str());
+
+        ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str());
+
+        ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str());
+        ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
+        ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
+    }
+}
+
+static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+    return *ptr;
+}
+
+static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+    int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+    return *ptr;
+}
+
+static void print_row(struct ggml_tensor * probs, int i) {
+    for (int k = 0; k < probs->ne[0]; ++k) {
+        float p = get_f32_2d(probs, k, i);
+        printf(" %f", p);
+    }
+    printf("\n");
+}
+
+static void print_matrix(struct ggml_tensor * probs) {
+    assert(probs->n_dims == 2);
+    for (int i = 0; i < probs->ne[1]; ++i) {
+        for (int k = 0; k < probs->ne[0]; ++k) {
+            float p = get_f32_2d(probs, k, i);
+            printf(" %.2f", p);
+        }
+        printf("\n");
+    }
+}
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("fread failed: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+    std::float_t read_f32() {
+        std::float_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static bool is_ggml_file(const char * filename) {
+    llama_file file(filename, "rb");
+    if (file.size < 4) {
+        return false;
+    }
+    uint32_t magic = file.read_u32();
+    return magic == GGUF_MAGIC;
+}
+
+static std::string llama_escape_whitespaces(const std::string & text) {
+    std::ostringstream out;
+    for (char c : text) {
+        if (c == ' ') out << "\xe2\x96\x81";
+        else out << c;
+    }
+    return out.str();
+}
+
+static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+    if (is_ggml_file(filename)) {
+        struct ggml_context * ctx_data = NULL;
+
+        struct gguf_init_params params = {
+            /*.no_alloc = */ false,
+            /*.ctx      = */ &ctx_data,
+        };
+
+        struct gguf_context * ctx = gguf_init_from_file(filename, params);
+        GGML_ASSERT(ctx != NULL);
+
+        const int model_idx = gguf_find_key(ctx, KV_TOKENIZER_MODEL);
+        GGML_ASSERT(model_idx >= 0);
+        std::string tokenizer_name = gguf_get_val_str(ctx, model_idx);
+        GGML_ASSERT(tokenizer_name == TOKENIZER_NAME);
+
+        const int token_idx = gguf_find_key(ctx, KV_TOKENIZER_LIST);
+        GGML_ASSERT(token_idx >= 0);
+
+        const int score_idx = gguf_find_key(ctx, KV_TOKENIZER_SCORES);
+        GGML_ASSERT(score_idx >= 0);
+        const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+
+        const int toktype_idx = gguf_find_key(ctx, KV_TOKENIZER_TOKEN_TYPE);
+        GGML_ASSERT(toktype_idx >= 0);
+        const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
+
+        const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+
+        vocab->id_to_token.resize(n_vocab);
+
+        for (uint32_t i = 0; i < n_vocab; i++) {
+            std::string word = gguf_get_arr_str(ctx, token_idx, i);
+
+            vocab->token_to_id[word] = i;
+
+            auto & token_data = vocab->id_to_token[i];
+            token_data.text  = std::move(word);
+            token_data.score = scores[i];
+            token_data.type  = (llama_token_type) toktypes[i];
+        }
+        ggml_free(ctx_data);
+        gguf_free(ctx);
+    } else {
+        // assume llama2.c vocabulary
+        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        llama_file file(filename, "rb");
+        if (!file.fp) {
+            die_fmt("%s: %s", strerror(errno), filename);
+        }
+        const int  n_vocab = config->vocab_size;
+        /* uint32_t max_token_length =  */ file.read_u32(); // unused
+        vocab->id_to_token.resize(n_vocab);
+        for (llama_vocab::id id=0; id<n_vocab; ++id) {
+            float_t score = file.read_f32();
+            uint32_t len = file.read_u32();
+            std::string text = file.read_string(len);
+
+            unsigned char byte_val;
+            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            if (id == UNKNOWN_TOKEN_ID) {
+                text = "<unk>";
+                type = LLAMA_TOKEN_TYPE_UNKNOWN;
+            } else if (id == BOS_TOKEN_ID) {
+                text = "<s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (id == EOS_TOKEN_ID) {
+                text = "</s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (text.empty()) {
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+                // Text of byte tokens is already in the expected format.
+                type = LLAMA_TOKEN_TYPE_BYTE;
+            } else {
+                type = LLAMA_TOKEN_TYPE_NORMAL;
+            }
+            text = llama_escape_whitespaces(text);
+
+            vocab->id_to_token[id].text = text;
+            vocab->id_to_token[id].score = score;
+            vocab->id_to_token[id].type = type;
+            vocab->token_to_id.emplace(text, id);
+        }
+    }
+}
+
+static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
+    int ct;
+    switch (gg_weights->n_dims){
+        case 1:
+            ct = 0;
+            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
+                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
+                *ptr = karpathy_weights[ct];
+                ct++;
+            }
+            break;
+        case 2:
+            ct = 0;
+            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
+                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
+                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
+                    *ptr = karpathy_weights[ct];
+                    ct++;
+                }
+            }
+            break;
+        case 3:
+            ct = 0;
+            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
+                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
+                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
+                        *ptr = karpathy_weights[ct];
+                        ct++;
+                    }
+                }
+            }
+            break;
+    }
+}
+
+static void save_as_llama_model(
+    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+) {
+    // convert AK weights into GG weights one by one.
+    // w->token_embedding_table -> model->tok_embeddings
+    // float*                   -> struct ggml_tensor
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
+    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    //print_row(model->norm, 0);
+
+    // for rms-att-weight
+    int row_length = model->hparams.n_embd;
+    int n_ff = model->hparams.n_ff;
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
+        auto & layer = model->layers[i];
+        // 1d
+        convert_weights_ak_to_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        convert_weights_ak_to_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
+
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
+        convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+
+        convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
+        convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
+        convert_weights_ak_to_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
+    }
+
+    struct gguf_context * ctx = gguf_init_empty();
+
+    std::vector<const char*> tokens;
+    std::vector<float> scores;
+    std::vector<llama_token_type> token_types;
+    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+        tokens.push_back(token_data.text.c_str());
+        scores.push_back(token_data.score);
+        token_types.push_back(token_data.type);
+    }
+    gguf_set_arr_str(ctx, KV_TOKENIZER_LIST, tokens.data(), tokens.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_SCORES, GGUF_TYPE_FLOAT32, scores.data(), scores.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_TOKEN_TYPE, GGUF_TYPE_INT32, token_types.data(), token_types.size());
+
+    gguf_set_val_str(ctx, KV_TOKENIZER_MODEL, TOKENIZER_NAME);
+
+    gguf_set_val_str(ctx, KV_GENERAL_ARCHITECTURE, "llama");
+    gguf_set_val_str(ctx, KV_GENERAL_NAME, "llama");
+
+    // special tokens
+    gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
+
+    gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
+    gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
+    gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    // n_head_kv is optional, default to n_head
+    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
+    gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
+    gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
+
+    // write tensors
+    ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD);
+    gguf_add_tensor(ctx, model->tok_embeddings);
+
+    ggml_set_name(model->norm, TN_OUTPUT_NORM);
+    gguf_add_tensor(ctx, model->norm);
+
+    ggml_set_name(model->output, TN_OUTPUT);
+    gguf_add_tensor(ctx, model->output);
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        ggml_format_name(layer.wq, TN_ATTN_Q, i);
+        gguf_add_tensor(ctx, layer.wq);
+
+        ggml_format_name(layer.wk, TN_ATTN_K, i);
+        gguf_add_tensor(ctx, layer.wk);
+
+        ggml_format_name(layer.wv, TN_ATTN_V, i);
+        gguf_add_tensor(ctx, layer.wv);
+
+        ggml_format_name(layer.wo, TN_ATTN_OUTPUT, i);
+        gguf_add_tensor(ctx, layer.wo);
+
+        ggml_format_name(layer.attention_norm, TN_ATTN_NORM, i);
+        gguf_add_tensor(ctx, layer.attention_norm);
+
+        ggml_format_name(layer.w1, TN_FFN_GATE, i);
+        gguf_add_tensor(ctx, layer.w1);
+
+        ggml_format_name(layer.w2, TN_FFN_DOWN, i);
+        gguf_add_tensor(ctx, layer.w2);
+
+        ggml_format_name(layer.w3, TN_FFN_UP, i);
+        gguf_add_tensor(ctx, layer.w3);
+
+        ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i);
+        gguf_add_tensor(ctx, layer.ffn_norm);
+    }
+
+    gguf_write_to_file(ctx, filename, false);
+    gguf_free(ctx);
+}
+
+static struct train_params get_default_train_params() {
+    struct train_params params;
+    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_llama2c_output_model = "ak_llama_model.bin";
+    params.fn_train_data     = "shakespeare.txt";
+    params.fn_checkpoint_in  = "checkpoint.bin";
+    params.fn_checkpoint_out = "checkpoint.bin";
+    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+
+    params.seed       =   -1;
+
+    params.n_ctx      =  128;
+    params.n_embd     =  256;
+    params.n_mult     =  256;
+    params.n_head     =    8;
+    params.n_layer    =   16;
+    params.n_rotmax   =   64;
+
+    params.n_threads  =    6;
+    params.n_batch    =    8;
+    params.n_examples =    8;
+    params.n_predict  = 1024;
+
+    params.print_info_interval    = 1;
+    params.print_details_interval = 2;
+
+    params.samples_start_after_nl = false;
+    params.use_adam               = true;
+    params.use_flash              = true;
+    params.use_scratch            = true;
+
+    // only adam
+    params.warmup            =  100;
+    params.cos_decay_steps   = 1000;
+    params.cos_decay_restart = 1.1f;
+    params.cos_decay_alpha   = 0.0f;
+
+    params.lbfgs_n_iter      = 16;
+    params.adam_n_iter       = 16;
+    params.adam_alpha        = 1e-3f;
+    params.adam_decay        = 1e-3f;
+
+    params.mem_model_gb   = 2;
+    params.mem_compute_gb = 24;
+    params.mem_compute0_gb = 8;
+    params.mem_compute1_gb = 2;
+
+    return params;
+}
+
+static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                       show this help message and exit\n");
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
+    fprintf(stderr, "  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
+    fprintf(stderr, "\n");
+}
+
+static bool params_parse(int argc, char ** argv, struct train_params * params) {
+    bool invalid_param = false;
+    bool reqd_param_found = false;
+    std::string arg;
+    struct train_params default_params = get_default_train_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "--copy-vocab-from-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_vocab_model = argv[i];
+        } else if (arg == "--llama2c-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            reqd_param_found = true;
+            params->fn_llama2c_model = argv[i];
+        } else if (arg == "--llama2c-output-model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_llama2c_output_model = argv[i];
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, &default_params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (!reqd_param_found){
+        fprintf(stderr, "error: please specify a llama2.c .bin file to be converted with argument --llama2c-model\n");
+        print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+
+    return true;
+}
+
+static std::string basename(const std::string &path) {
+    size_t pos = path.find_last_of("/\\");
+    if (pos == std::string::npos) {
+        return path;
+    }
+    return path.substr(pos + 1);
+}
+
+int main(int argc, char ** argv) {
+    struct train_params params = get_default_train_params();
+    if (!params_parse(argc, argv, &params)) {
+        return 1;
+    }
+    Config config;
+    TransformerWeights weights = {};
+    {
+        FILE *file = fopen(params.fn_llama2c_model, "rb");
+        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        // read in the config header
+        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        auto shared_weights = config.vocab_size > 0;
+        config.vocab_size = abs(config.vocab_size);
+
+        // read in the Transformer weights
+        malloc_weights(&weights, &config, shared_weights);
+        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        fclose(file);
+    }
+
+    struct llama_vocab vocab;
+    load_vocab(params.fn_vocab_model, &config, &vocab);
+
+    struct my_llama_model model;
+    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx   = params.n_ctx;
+    model.hparams.n_embd  = config.dim; //params.n_embd;
+    model.hparams.n_ff    = config.hidden_dim;
+    model.hparams.n_mult  = 32;//params.n_mult;
+    model.hparams.n_head  = config.n_heads; //params.n_head;
+    model.hparams.n_layer = config.n_layers; //params.n_layer;
+    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    print_params(&model.hparams);
+    struct ggml_init_params lcparams;
+    lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
+    lcparams.mem_buffer = NULL;
+    lcparams.no_alloc   = false;
+
+    model.ctx = ggml_init(lcparams);
+
+    init_model(&model);
+    model.name = basename(params.fn_llama2c_model);
+    save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
+
+    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+
+    ggml_free(model.ctx);
+    return 0;
+}

examples/embd-input/embd-input-lib.cpp

@@ -1,8 +1,5 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
+#include "build-info.h"
+#include "common.h"
 #include "embd-input.h"
 
 #include <cassert>
@@ -23,11 +20,11 @@ extern "C" {
 struct MyModel* create_mymodel(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return nullptr;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = uint32_t(time(NULL));
@@ -167,7 +164,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
         // TODO: Apply penalties
-        // float nl_logit = logits[llama_token_nl()];
+        // float nl_logit = logits[llama_token_nl(ctx)];
         // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
         // llama_sample_repetition_penalty(ctx, &candidates_p,
         //      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
@@ -176,7 +173,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
         // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
         // last_n_repeat, alpha_frequency, alpha_presence);
         // if (!penalize_nl) {
-        //     logits[llama_token_nl()] = nl_logit;
+        //     logits[llama_token_nl(ctx)] = nl_logit;
         // }
 
         if (temp <= 0) {
@@ -211,10 +208,10 @@ const char * sampling(struct MyModel * mymodel) {
     llama_context * ctx = mymodel->ctx;
     int id = sampling_id(mymodel);
     static std::string ret;
-    if (id == llama_token_eos()) {
+    if (id == llama_token_eos(ctx)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_str(ctx, id);
+        ret = llama_token_to_piece(ctx, id);
     }
     eval_id(mymodel, id);
     return ret.c_str();