llama.cpp : print gguf version

The use of special characters within source files can break compiling on some computers with different region and language settings. Using Unicode escape sequences should allow for the code to be compiled on all setups without needing to change your computers settings or switch regions.

.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/lamma-cpp-clblast.srpm.spec

@@ -0,0 +1,58 @@
+# 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:        master
+Release:        1%{?dist}
+Summary:        OpenCL Inference of LLaMA model in pure 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
+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}/llamacppclblast
+cp -p server %{buildroot}%{_bindir}/llamacppclblastserver
+cp -p simple %{buildroot}%{_bindir}/llamacppclblastsimple
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppclblast
+%{_bindir}/llamacppclblastserver
+%{_bindir}/llamacppclblastsimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

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

@@ -0,0 +1,59 @@
+# 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:        master
+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
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppcublas
+%{_bindir}/llamacppcublasserver
+%{_bindir}/llamacppcublassimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.devops/llama-cpp.srpm.spec

@@ -0,0 +1,58 @@
+# 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
+Version:        master
+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
+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
+%autosetup
+
+%build
+make -j
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacpp
+cp -p server %{buildroot}%{_bindir}/llamacppserver
+cp -p simple %{buildroot}%{_bindir}/llamacppsimple
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacpp
+%{_bindir}/llamacppserver
+%{_bindir}/llamacppsimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.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" ]

.dockerignore

@@ -5,14 +5,7 @@
 .vscode/
 .DS_Store
 
-build/
-build-em/
-build-debug/
-build-release/
-build-static/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build*/
 
 models/*
 

.github/workflows/build.yml

@@ -291,24 +291,32 @@ jobs:
           cd build
           ctest -C Release --verbose --timeout 900
 
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
 
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v3
         with:
           path: |
-            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
+            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
 
   windows-latest-cmake-cublas:
     runs-on: windows-latest
@@ -338,23 +346,31 @@ jobs:
           cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
           cmake --build . --config Release
 
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
 
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v3
         with:
           path: |
-            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
         if: ${{ matrix.cuda == '12.1.0' }}
@@ -400,21 +416,34 @@ jobs:
       - windows-latest-cmake-cublas
 
     steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
       - name: Download artifacts
         id: download-artifact
         uses: actions/download-artifact@v3
 
-      - name: Get commit hash
-        id: commit
-        uses: pr-mpt/actions-commit-hash@v2
-
       - name: Create release
         id: create_release
         uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ steps.tag.outputs.name }}
 
       - name: Upload release
         id: upload_release

.gitignore

@@ -3,6 +3,8 @@
 *.so
 *.gguf
 *.bin
+*.exe
+*.dll
 .DS_Store
 .build/
 .cache/
@@ -14,20 +16,7 @@
 .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/
 
@@ -58,6 +47,7 @@ compile_commands.json
 CMakeSettings.json
 
 __pycache__
+dist
 
 zig-out/
 zig-cache/
@@ -68,7 +58,6 @@ perf-*.txt
 
 examples/jeopardy/results.txt
 
-pyproject.toml
 poetry.lock
 poetry.toml
 
@@ -81,4 +70,3 @@ tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0
-

CMakeLists.txt

@@ -74,6 +74,7 @@ set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
@@ -352,6 +353,43 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+
+    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 ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-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

Makefile

@@ -280,6 +280,30 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
+ifdef LLAMA_HIPBLAS
+	ROCM_PATH	?= /opt/rocm
+	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	LLAMA_CUDA_DMMV_X       ?= 32
+	LLAMA_CUDA_MMV_Y        ?= 1
+	LLAMA_CUDA_KQUANTS_ITER ?= 2
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	LDFLAGS		+= -lhipblas -lamdhip64 -lrocblas
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	HIPFLAGS    += -DCC_TURING=1000000000
+ifdef LLAMA_CUDA_FORCE_DMMV
+	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
+endif # LLAMA_CUDA_FORCE_DMMV
+	OBJS        += ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # LLAMA_HIPBLAS
+
 ifdef LLAMA_METAL
 	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
 	CXXFLAGS += -DGGML_USE_METAL

README.md

@@ -11,15 +11,17 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398)
+- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717
 
-Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa)
+- A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398)
 
-### Current `master` should be considered in Beta - expect some issues for a few days!
+  Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa)
 
-### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up!
+  ### Current `master` should be considered in Beta - expect some issues for a few days!
 
-### Issues with non-GGUF models will be considered with low priority!
+  ### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up!
+
+  ### Issues with non-GGUF models will be considered with low priority!
 
 ----
 
@@ -39,6 +41,7 @@ Last revision compatible with the old format: [dadbed9](https://github.com/ggerg
         <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
+        <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
         <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
         <li><a href="#using-openllama">Using OpenLLaMA</a></li>
         <li><a href="#using-gpt4all">Using GPT4All</a></li>
@@ -65,12 +68,11 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 - Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
-- 4-bit, 5-bit and 8-bit integer quantization support
-- Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS
-- cuBLAS and CLBlast support
+- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
+- CUDA, Metal and OpenCL GPU backend support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
-Since then, the project has improved significantly thanks to many contributions. This project is for educational purposes and serves
+Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves
 as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
 
 **Supported platforms:**
@@ -84,6 +86,7 @@ as the main playground for developing new features for the [ggml](https://github
 
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [X] Falcon
 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
 - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
@@ -114,90 +117,84 @@ as the main playground for developing new features for the [ggml](https://github
 
 ---
 
-Here is a typical run using LLaMA-7B:
+Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
 ```java
-make -j && ./main -m ./models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
+$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
 I UNAME_P:  arm
 I UNAME_M:  arm64
-I CFLAGS:   -I.              -O3 -DNDEBUG -std=c11   -fPIC -pthread -DGGML_USE_ACCELERATE
-I CXXFLAGS: -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -pthread
+I CFLAGS:   -I.            -O3 -std=c11   -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -pthread -DGGML_USE_K_QUANTS -DGGML_USE_ACCELERATE
+I CXXFLAGS: -I. -I./common -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -DGGML_USE_K_QUANTS
 I LDFLAGS:   -framework Accelerate
-I CC:       Apple clang version 14.0.0 (clang-1400.0.29.202)
-I CXX:      Apple clang version 14.0.0 (clang-1400.0.29.202)
+I CC:       Apple clang version 14.0.3 (clang-1403.0.22.14.1)
+I CXX:      Apple clang version 14.0.3 (clang-1403.0.22.14.1)
 
 make: Nothing to be done for `default'.
-main: seed = 1678486056
-llama_model_load: loading model from './models/7B/ggml-model-q4_0.bin' - please wait ...
-llama_model_load: n_vocab = 32000
-llama_model_load: n_ctx   = 512
-llama_model_load: n_embd  = 4096
-llama_model_load: n_mult  = 256
-llama_model_load: n_head  = 32
-llama_model_load: n_layer = 32
-llama_model_load: n_rot   = 128
-llama_model_load: f16     = 2
-llama_model_load: n_ff    = 11008
-llama_model_load: ggml ctx size = 4529.34 MB
-llama_model_load: memory_size =   512.00 MB, n_mem = 16384
-llama_model_load: .................................... done
-llama_model_load: model size =  4017.27 MB / num tensors = 291
+main: build = 1041 (cf658ad)
+main: seed  = 1692823051
+llama_model_loader: loaded meta data with 16 key-value pairs and 363 tensors from models/llama-13b-v2/ggml-model-q4_0.gguf (version GGUF V1 (latest))
+llama_model_loader: - type  f32:   81 tensors
+llama_model_loader: - type q4_0:  281 tensors
+llama_model_loader: - type q6_K:    1 tensors
+llm_load_print_meta: format         = GGUF V1 (latest)
+llm_load_print_meta: arch           = llama
+llm_load_print_meta: vocab type     = SPM
+llm_load_print_meta: n_vocab        = 32000
+llm_load_print_meta: n_merges       = 0
+llm_load_print_meta: n_ctx_train    = 4096
+llm_load_print_meta: n_ctx          = 512
+llm_load_print_meta: n_embd         = 5120
+llm_load_print_meta: n_head         = 40
+llm_load_print_meta: n_head_kv      = 40
+llm_load_print_meta: n_layer        = 40
+llm_load_print_meta: n_rot          = 128
+llm_load_print_meta: n_gqa          = 1
+llm_load_print_meta: f_norm_eps     = 1.0e-05
+llm_load_print_meta: f_norm_rms_eps = 1.0e-05
+llm_load_print_meta: n_ff           = 13824
+llm_load_print_meta: freq_base      = 10000.0
+llm_load_print_meta: freq_scale     = 1
+llm_load_print_meta: model type     = 13B
+llm_load_print_meta: model ftype    = mostly Q4_0
+llm_load_print_meta: model size     = 13.02 B
+llm_load_print_meta: general.name   = LLaMA v2
+llm_load_print_meta: BOS token = 1 '<s>'
+llm_load_print_meta: EOS token = 2 '</s>'
+llm_load_print_meta: UNK token = 0 '<unk>'
+llm_load_print_meta: LF token  = 13 '<0x0A>'
+llm_load_tensors: ggml ctx size =    0.11 MB
+llm_load_tensors: mem required  = 7024.01 MB (+  400.00 MB per state)
+...................................................................................................
+llama_new_context_with_model: kv self size  =  400.00 MB
+llama_new_context_with_model: compute buffer total size =   75.41 MB
 
-main: prompt: 'Building a website can be done in 10 simple steps:'
-main: number of tokens in prompt = 15
-     1 -> ''
-  8893 -> 'Build'
-   292 -> 'ing'
-   263 -> ' a'
-  4700 -> ' website'
-   508 -> ' can'
-   367 -> ' be'
-  2309 -> ' done'
-   297 -> ' in'
- 29871 -> ' '
- 29896 -> '1'
- 29900 -> '0'
-  2560 -> ' simple'
-  6576 -> ' steps'
- 29901 -> ':'
-
-sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000
+system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
+sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
+generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0
 
 
-Building a website can be done in 10 simple steps:
-1) Select a domain name and web hosting plan
-2) Complete a sitemap
-3) List your products
-4) Write product descriptions
-5) Create a user account
-6) Build the template
-7) Start building the website
-8) Advertise the website
-9) Provide email support
-10) Submit the website to search engines
-A website is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves.
-The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user's browser.
-The web pages are stored in a web server. The web server is also called a host. When the website is accessed, it is retrieved from the server and displayed on the user's computer.
-A website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server.
-A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user's screen.
-A website can also be viewed on different devices such as desktops, tablets and smartphones.
-Hence, to have a website displayed on a browser, the website must be hosted.
-A domain name is an address of a website. It is the name of the website.
-The website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server.
-A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user’s screen.
-A website can also be viewed on different devices such as desktops, tablets and smartphones. Hence, to have a website displayed on a browser, the website must be hosted.
-A domain name is an address of a website. It is the name of the website.
-A website is an address of a website. It is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves.
-The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user’s browser.
-A website is known as a website when it is hosted
-
-main: mem per token = 14434244 bytes
-main:     load time =  1332.48 ms
-main:   sample time =  1081.40 ms
-main:  predict time = 31378.77 ms / 61.41 ms per token
-main:    total time = 34036.74 ms
+ Building a website can be done in 10 simple steps:
+Step 1: Find the right website platform.
+Step 2: Choose your domain name and hosting plan.
+Step 3: Design your website layout.
+Step 4: Write your website content and add images.
+Step 5: Install security features to protect your site from hackers or spammers
+Step 6: Test your website on multiple browsers, mobile devices, operating systems etc…
+Step 7: Test it again with people who are not related to you personally – friends or family members will work just fine!
+Step 8: Start marketing and promoting the website via social media channels or paid ads
+Step 9: Analyze how many visitors have come to your site so far, what type of people visit more often than others (e.g., men vs women) etc…
+Step 10: Continue to improve upon all aspects mentioned above by following trends in web design and staying up-to-date on new technologies that can enhance user experience even further!
+How does a Website Work?
+A website works by having pages, which are made of HTML code. This code tells your computer how to display the content on each page you visit – whether it’s an image or text file (like PDFs). In order for someone else’s browser not only be able but also want those same results when accessing any given URL; some additional steps need taken by way of programming scripts that will add functionality such as making links clickable!
+The most common type is called static HTML pages because they remain unchanged over time unless modified manually (either through editing files directly or using an interface such as WordPress). They are usually served up via HTTP protocols – this means anyone can access them without having any special privileges like being part of a group who is allowed into restricted areas online; however, there may still exist some limitations depending upon where one lives geographically speaking.
+How to
+llama_print_timings:        load time =   576.45 ms
+llama_print_timings:      sample time =   283.10 ms /   400 runs   (    0.71 ms per token,  1412.91 tokens per second)
+llama_print_timings: prompt eval time =   599.83 ms /    19 tokens (   31.57 ms per token,    31.68 tokens per second)
+llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms per token,    16.28 tokens per second)
+llama_print_timings:       total time = 25431.49 ms
 ```
 
 And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
@@ -425,6 +422,35 @@ Building the program with BLAS support may lead to some performance improvements
   | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 
+- #### hipBLAS
+
+  This provide BLAS acceleation on HIP supported GPU like AMD GPU.
+  Make sure to have ROCm installed.
+  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
+  Windows support is coming soon...
+
+  - Using `make`:
+    ```bash
+    make LLAMA_HIPBLAS=1
+    ```
+  - Using `CMake`:
+    ```bash
+    mkdir build
+    cd build
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
+    cmake --build .
+    ```
+
+  The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
+  If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
+  The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
+  | Option                  | Legal values           | Default | Description |
+  |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+
 - #### CLBlast
 
   OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
@@ -542,6 +568,8 @@ As the models are currently fully loaded into memory, you will need adequate dis
 
 Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
 
+*(outdated)*
+
 | Model | Measure      | F16    | Q4_0   | Q4_1   | Q5_0   | Q5_1   | Q8_0   |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
 |    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
@@ -604,6 +632,16 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
     CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
 ```
 
+### Constrained output with grammars
+
+`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
+
+```bash
+./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+```
+
+The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder
@@ -885,3 +923,4 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 - [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [GBNF grammars](./grammars/README.md)

ci/run.sh

@@ -391,6 +391,7 @@ 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

common/common.cpp

@@ -289,7 +289,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_batch = std::stoi(argv[i]);
-            params.n_batch = std::min(512, params.n_batch);
         } else if (arg == "--keep") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -388,11 +387,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #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 == "--mul-mat-q" || arg == "-mmq") {
+        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
+            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
@@ -418,6 +417,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } 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") {
@@ -600,11 +611,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+#ifdef GGML_USE_CUBLAS
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif // GGML_USE_CUBLAS
 #endif
     fprintf(stdout, "  --mtest               compute maximum memory usage\n");
     fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -733,35 +746,3 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok
 
     return std::string(result.data(), result.size());
 }
-
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_bos) {
-    int n_tokens = text.length() + add_bos;
-    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-    return result;
-}
-
-std::string llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token) {
-    std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        const int check = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-
-    return std::string(result.data(), result.size());
-}
-

common/common.h

@@ -28,6 +28,7 @@ struct gpt_params {
     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
 
@@ -64,11 +65,15 @@ struct gpt_params {
     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         = false; // if true, use experimental mul_mat_q kernels
+    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
@@ -116,15 +121,6 @@ std::vector<llama_token> llama_tokenize(
            const std::string & text,
                         bool   add_bos);
 
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_bos);
-
 std::string llama_token_to_str(
         const struct llama_context * ctx,
                        llama_token   token);
-
-std::string llama_token_to_str_bpe(
-    const struct llama_context * ctx,
-                   llama_token   token);

convert-falcon-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF falcon--> gguf conversion
 
 import gguf
@@ -94,21 +95,27 @@ print("gguf: get model metadata")
 
 block_count = hparams["n_layer"]
 
-gguf_writer.add_name(last_dir)
+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"])
+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[str] = []
+scores: List[float] = []
+toktypes: List[int] = []
 merges: List[str] = []
 
 
@@ -152,41 +159,30 @@ if Path(dir_model + "/tokenizer.json").is_file():
             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)
 
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
+print("gguf: get special token ids")
+# Look for special tokens in config.json
 
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
+if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
+    gguf_writer.add_bos_token_id(hparams["bos_token_id"])
 
-        # find special token ids
+if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
+    gguf_writer.add_eos_token_id(hparams["eos_token_id"])
 
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
+if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
+    gguf_writer.add_unk_token_id(hparams["unk_token_id"])
 
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
+if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
+    gguf_writer.add_sep_token_id(hparams["sep_token_id"])
 
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
+    gguf_writer.add_pad_token_id(hparams["pad_token_id"])
 
 
 # TENSORS
@@ -194,8 +190,9 @@ if Path(dir_model + "/tokenizer.json").is_file():
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
 # params for qkv transform
-n_head = hparams["n_head"]
+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

convert-gptneox-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF gptneox--> gguf conversion
 
 import gguf

convert-llama-7b-pth-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # 7b pth llama --> gguf conversion
 # Only models with a single datafile are supported, like 7B
 # HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model

convert-llama-ggmlv3-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys, struct, math, argparse
 from pathlib import Path
 
@@ -93,7 +94,7 @@ class Tensor:
         pad = ((offset + 31) & ~31) - offset
         offset += pad
         n_elems = np.prod(self.dims)
-        n_bytes = (n_elems * tysize) // blksize
+        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
@@ -215,15 +216,10 @@ class GGMLToGGUF:
         if self.vocab_override is not None:
             vo = self.vocab_override
             print('* Adding vocab item(s)')
-            for (idx, vitem) in enumerate(vo.all_tokens()):
-                if len(vitem) == 3:
-                    tokens.append(vitem[0])
-                    scores.append(vitem[1])
-                    toktypes.append(vitem[2])
-                else:
-                    # Maybe try to guess the token type here?
-                    tokens.append(vitem[0])
-                    scores.append(vitem[1])
+            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)
@@ -231,13 +227,24 @@ class GGMLToGGUF:
                 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
-            if len(vbytes) == 0:
+            # 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:
-                hv = hex(vbytes[0])[2:].upper()
-                vbytes = bytes(f'<0x{hv}>', encoding = 'UTF-8')
+                vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8')
                 tt = 6 # Byte
             else:
                 vbytes = vbytes.replace(b' ', b'\xe2\x96\x81')
@@ -247,6 +254,9 @@ class GGMLToGGUF:
         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):
         nm = self.name_map
@@ -331,4 +341,5 @@ def main():
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
-main()
+if __name__ == '__main__':
+    main()

convert-llama-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF llama --> gguf conversion
 
 import gguf

convert-lora-to-ggml.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 import json
 import os
 import re
@@ -6,23 +6,22 @@ import struct
 import sys
 from typing import Any, Dict, Sequence, TextIO
 
+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 +38,7 @@ 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:
@@ -54,12 +53,14 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
     # 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
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(
@@ -67,7 +68,7 @@ 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-pth-to-ggml.py

@@ -1,13 +0,0 @@
-# Compatibility stub
-
-import argparse
-
-import convert
-
-parser = argparse.ArgumentParser(
-    description="""[DEPRECATED - use `convert.py` instead]
-    Convert a LLaMA model checkpoint to a ggml compatible file""")
-parser.add_argument('dir_model',  help='directory containing the model checkpoint')
-parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
-args = parser.parse_args()
-convert.main(['--outtype', 'f16' if args.ftype == 1 else 'f32', '--', args.dir_model])

convert.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 
 import gguf
 import argparse
@@ -69,7 +69,10 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
     'I32': DT_I32,
 }
 
-class GGMLFileType(enum.Enum):
+# 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
 
@@ -101,6 +104,14 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
+    f_rope_freq_base: Optional[float] = None
+    f_rope_scale: Optional[float] = None
+
+    ftype: Optional[GGMLFileType] = None
+
+    # path to the directory containing the model files
+    path_model: Optional['Path'] = None
+
     @staticmethod
     def find_n_mult(n_ff: int, n_embd: int) -> int:
         # hardcoded magic range
@@ -150,13 +161,20 @@ class 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_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"]
+        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 = Params.find_n_mult(n_ff, n_embd)
 
@@ -169,15 +187,17 @@ class Params:
                             "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_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,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_mult           = n_mult,
+            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
@@ -186,15 +206,26 @@ class 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_layer    = config["n_layers"]
-        n_mult     = config["multiple_of"]
-        n_ctx      = 2048 if config["norm_eps"] == 1e-06 else 4096 # hack to determine LLaMA v1 vs v2
-        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"]
+        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
+        n_embd           = config["dim"]
+        n_layer          = config["n_layers"]
+        n_mult           = config["multiple_of"]
+        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 and 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]
@@ -203,15 +234,16 @@ class Params:
             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_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,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_mult           = n_mult,
+            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
@@ -226,6 +258,8 @@ class Params:
         else:
             params = Params.guessed(model_plus.model)
 
+        params.path_model = model_plus.paths[0].parent
+
         return params
 
 
@@ -728,7 +762,13 @@ class OutputFile:
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
     def add_meta_arch(self, params: Params) -> None:
-        self.gguf.add_name                ("LLaMA")
+        name = "LLaMA"
+        if (params.n_ctx == 4096):
+            name = "LLaMA v2"
+            if params.path_model:
+                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)
@@ -738,6 +778,15 @@ class OutputFile:
         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:
+            self.gguf.add_rope_freq_base(params.f_rope_freq_base)
+
+        if params.f_rope_scale:
+            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+
+        if params.ftype:
+            self.gguf.add_file_type(params.ftype)
+
     def add_meta_vocab(self, vocab: Vocab) -> None:
         tokens = []
         scores = []
@@ -956,7 +1005,7 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence
             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}'")
@@ -1020,6 +1069,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
                             " - LLaMA v2: --ctx 4096\n")
         params.n_ctx = args.ctx
 
+    if args.outtype:
+        params.ftype = {
+            "f32": GGMLFileType.AllF32,
+            "f16": GGMLFileType.MostlyF16,
+        }[args.outtype]
+
     print(f"params = {params}")
 
     vocab: Vocab
@@ -1040,11 +1095,14 @@ def main(args_in: Optional[List[str]] = None) -> None:
             vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
             vocab = load_vocab(vocab_dir, args.vocabtype)
 
-        model       = model_plus.model
-        model       = convert_model_names(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)
+        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, params, model, vocab)
         print(f"Wrote {outfile}")

examples/CMakeLists.txt

@@ -25,6 +25,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(embd-input)
     add_subdirectory(llama-bench)
+    add_subdirectory(beam_search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/beam_search/CMakeLists.txt

@@ -0,0 +1,8 @@
+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)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/beam_search/beam_search.cpp

@@ -0,0 +1,188 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
+#include "common.h"
+#include "llama.h"
+#include "build-info.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
+#define NOMINMAX
+#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;
+};
+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_str(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.
+bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const 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.
+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("%lu", 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 (%lu tokens, max %lu)\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_str(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_str(ctx,token_id);
+    }
+    std::cout << std::endl;
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    llama_backend_free();
+
+    return 0;
+}

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

@@ -12,15 +12,19 @@ usage: ./convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
-  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'models/ggml-vocab.bin')
+  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'tokenizer.bin')
   --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 is as follows:
+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 <ggml-vocab.bin> --llama2c-model <llama2.c model path> --llama2c-output-model <ggml output model path>`
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin`
 
-Now you can use the model with command like:
+For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually:
 
-`$ ./main -m <ggml output model path> -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5`
+`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output 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

@@ -17,6 +17,9 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_VERSION_GGJT_V3   3
+
 //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
 typedef struct {
     int dim; // transformer dimension
@@ -49,10 +52,10 @@ typedef struct {
     // 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;
+    float* wcls;
 } TransformerWeights;
 
-void malloc_weights(TransformerWeights* w, Config* p) {
+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);
@@ -86,9 +89,16 @@ void malloc_weights(TransformerWeights* w, Config* p) {
 
     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);
+    }
 }
 
-int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
+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;
@@ -100,6 +110,22 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
     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;
 }
 
@@ -115,6 +141,7 @@ void free_weights(TransformerWeights* w) {
     delete w->w2;
     delete w->w3;
     delete w->rms_final_weight;
+    if (w->wcls) delete w->wcls;
 }
 
 void print_sample_weights(TransformerWeights *w){
@@ -131,6 +158,7 @@ void print_sample_weights(TransformerWeights *w){
     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]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -509,26 +537,28 @@ bool is_ggml_file(const char *filename) {
 }
 
 void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
-    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
-    if (is_ggml_file(filename)) {
-
-        struct llama_context_params llama_params = llama_context_default_params();
-        llama_params.vocab_only = true;
-
-        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
-        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
-
-        const int n_vocab = llama_n_vocab(lctx);
-        vocab->id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
-            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
-            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
-            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
-            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
-        }
-        llama_free(lctx);
-        llama_free_model(lmodel);
-    } else { // assume llama2.c vocabulary
+#pragma message("TODO: implement reading vocabulary using gguf")
+//    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
+//    if (is_ggml_file(filename)) {
+//
+//        struct llama_context_params llama_params = llama_context_default_params();
+//        llama_params.vocab_only = true;
+//
+//        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
+//        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+//
+//        const int n_vocab = llama_n_vocab(lctx);
+//        vocab->id_to_token.resize(n_vocab);
+//        for (int i=0; i<n_vocab; ++i) {
+//            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
+//            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
+//            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
+//            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
+//        }
+//        llama_free(lctx);
+//        llama_free_model(lmodel);
+//    } else
+    { // assume llama2.c vocabulary
         printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
         llama_file file(filename, "rb");
         const int  n_vocab = config->vocab_size;
@@ -538,6 +568,12 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
+            // Special-case handling of <0xXX> single byte tokens.
+            char byte_val;
+            if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+                char cstr[2] = { byte_val, 0 };
+                text = cstr;
+            }
             vocab->id_to_token[i].text = text;
             vocab->id_to_token[i].score = score;
             vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
@@ -589,83 +625,80 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
     }
 
 #pragma message("TODO: implement file saving using gguf")
-    (void) vocab;
-    (void) model;
-    (void) w;
-//    // write_magic
-//    file.write_u32(LLAMA_FILE_MAGIC);   // magic
-//    file.write_u32(LLAMA_FILE_VERSION); // version
-//    // write_hparams
-//    file.write_u32(model->hparams.n_vocab);
-//    file.write_u32(model->hparams.n_embd);
-//    file.write_u32(model->hparams.n_mult);
-//    file.write_u32(model->hparams.n_head);
-//    file.write_u32(model->hparams.n_layer);
-//    file.write_u32(model->hparams.n_rot);
-//    file.write_u32(LLAMA_FTYPE_ALL_F32);
-//
-//    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
-//    uint32_t n_vocab = model->hparams.n_vocab;
-//    for (uint32_t i = 0; i < n_vocab; i++) {
-//        const auto & token_data = vocab->id_to_token.at(i);
-//        file.write_u32((uint32_t) token_data.tok.size());
-//        file.write_raw(token_data.tok.data(), token_data.tok.size());
-//        file.write_raw(&token_data.score, sizeof(token_data.score));
-//    }
-//
-//    // stuff AK weights into GG weights one by one.
-//    // w->token_embedding_table -> model->tok_embeddings
-//    // float*                   -> struct ggml_tensor
-//    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
-//    stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
-//
-//    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
-//    //print_row(model->norm, 0);
-//
-//    // for rms-att-weight
-//    int row_length = model->hparams.n_embd;
-//    const auto & hparams = model->hparams;
-//    //int n_ff = model->hparams.n_embd;
-//    int n_ff = get_n_ff(&hparams);
-//
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
-//        auto & layer = model->layers[i];
-//        // 1d
-//        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
-//        stuff_karpathy_weights_into_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
-//        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
-//
-//        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
-//        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
-//    }
-//    // write tensors
-//    write_tensor(&file, model->tok_embeddings);
-//    write_tensor(&file, model->norm);
-//    write_tensor(&file, model->output); // ?
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
-//        auto & layer = model->layers[i];
-//
-//        write_tensor(&file, layer.attention_norm);
-//        write_tensor(&file, layer.wq);
-//        write_tensor(&file, layer.wk);
-//        write_tensor(&file, layer.wv);
-//        write_tensor(&file, layer.wo);
-//        write_tensor(&file, layer.ffn_norm);
-//        write_tensor(&file, layer.w1);
-//        write_tensor(&file, layer.w2);
-//        write_tensor(&file, layer.w3);
-//    }
+    // write_magic
+    file.write_u32(LLAMA_FILE_MAGIC_GGJT);   // magic
+    file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // version
+    // write_hparams
+    file.write_u32(model->hparams.n_vocab);
+    file.write_u32(model->hparams.n_embd);
+    file.write_u32(model->hparams.n_mult);
+    file.write_u32(model->hparams.n_head);
+    file.write_u32(model->hparams.n_layer);
+    file.write_u32(model->hparams.n_rot);
+    file.write_u32(LLAMA_FTYPE_ALL_F32);
+
+    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
+    uint32_t n_vocab = model->hparams.n_vocab;
+    for (uint32_t i = 0; i < n_vocab; i++) {
+        const auto & token_data = vocab->id_to_token.at(i);
+        file.write_u32((uint32_t) token_data.text.size());
+        file.write_raw(token_data.text.data(), token_data.text.size());
+        file.write_raw(&token_data.score, sizeof(token_data.score));
+    }
+
+    // stuff AK weights into GG weights one by one.
+    // w->token_embedding_table -> model->tok_embeddings
+    // float*                   -> struct ggml_tensor
+    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
+    stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+
+    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
+    //print_row(model->norm, 0);
+
+    // for rms-att-weight
+    int row_length = model->hparams.n_embd;
+    const auto & hparams = model->hparams;
+    //int n_ff = model->hparams.n_embd;
+    int n_ff = get_n_ff(&hparams);
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
+        auto & layer = model->layers[i];
+        // 1d
+        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        stuff_karpathy_weights_into_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
+        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+
+        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
+        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
+        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
+    }
+    // write tensors
+    write_tensor(&file, model->tok_embeddings);
+    write_tensor(&file, model->norm);
+    write_tensor(&file, model->output); // ?
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        write_tensor(&file, layer.attention_norm);
+        write_tensor(&file, layer.wq);
+        write_tensor(&file, layer.wk);
+        write_tensor(&file, layer.wv);
+        write_tensor(&file, layer.wo);
+        write_tensor(&file, layer.ffn_norm);
+        write_tensor(&file, layer.w1);
+        write_tensor(&file, layer.w2);
+        write_tensor(&file, layer.w3);
+    }
 }
 
 struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/ggml-vocab.bin";
+    params.fn_vocab_model    = "tokenizer.bin";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
     params.fn_train_data     = "shakespeare.txt";
     params.fn_checkpoint_in  = "checkpoint.bin";
@@ -718,7 +751,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
     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    llama2.c vocabulary or ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggmlv3 model path 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");
@@ -791,9 +824,12 @@ int main(int argc, char ** argv) {
         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);
-        if(checkpoint_init_weights(&weights, &config, file)) { return 1; }
+        malloc_weights(&weights, &config, shared_weights);
+        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
         fclose(file);
     }
 

examples/embd-input/embd_input.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import ctypes
 from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
 import numpy as np

examples/embd-input/llava.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/embd-input/minigpt4.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/embd-input/panda_gpt.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/embedding/embedding.cpp

@@ -72,23 +72,30 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "\n");
     }
 
-    if (params.embedding){
-        if (embd_inp.size() > 0) {
-            if (llama_eval(ctx, embd_inp.data(), embd_inp.size(), n_past, params.n_threads)) {
-                fprintf(stderr, "%s : failed to eval\n", __func__);
-                return 1;
-            }
-        }
-
-        const int n_embd = llama_n_embd(ctx);
-        const auto embeddings = llama_get_embeddings(ctx);
-
-        for (int i = 0; i < n_embd; i++) {
-            printf("%f ", embeddings[i]);
-        }
-        printf("\n");
+    if (embd_inp.size() > (size_t)params.n_ctx) {
+        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
+                __func__, embd_inp.size(), params.n_ctx);
+        return 1;
     }
 
+    while (!embd_inp.empty()) {
+        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
+        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return 1;
+        }
+        n_past += n_tokens;
+        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
+    }
+
+    const int n_embd = llama_n_embd(ctx);
+    const auto embeddings = llama_get_embeddings(ctx);
+
+    for (int i = 0; i < n_embd; i++) {
+        printf("%f ", embeddings[i]);
+    }
+    printf("\n");
+
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);

examples/gguf/gguf.cpp

@@ -30,6 +30,9 @@ bool gguf_ex_write(const std::string & fname) {
     gguf_set_val_u32 (ctx, "some.parameter.uint32",   0x12345678);
     gguf_set_val_i32 (ctx, "some.parameter.int32",   -0x12345679);
     gguf_set_val_f32 (ctx, "some.parameter.float32",  0.123456789f);
+    gguf_set_val_u64 (ctx, "some.parameter.uint64",   0x123456789abcdef0ull);
+    gguf_set_val_i64 (ctx, "some.parameter.int64",   -0x123456789abcdef1ll);
+    gguf_set_val_f64 (ctx, "some.parameter.float64",  0.1234567890123456789);
     gguf_set_val_bool(ctx, "some.parameter.bool",     true);
     gguf_set_val_str (ctx, "some.parameter.string",   "hello world");
 

examples/jeopardy/graph.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import matplotlib.pyplot as plt
 import os
 import csv

examples/json-schema-to-grammar.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import argparse
 import json
 import re

examples/llama-bench/llama-bench.cpp

@@ -18,9 +18,7 @@
 #include "llama.h"
 #include "common.h"
 #include "build-info.h"
-#ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#endif
 
 // utils
 static uint64_t get_time_ns() {
@@ -148,7 +146,7 @@ struct cmd_params {
 };
 
 static const cmd_params cmd_params_defaults = {
-    /* model         */ {"models/7B/ggml-model-q4_0.bin"},
+    /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
     /* n_batch       */ {512},
@@ -179,12 +177,12 @@ static void print_usage(int /* argc */, char ** argv) {
     fprintf(stdout, "  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     fprintf(stdout, "  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
     fprintf(stdout, "  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    fprintf(stdout, "  -ts, --tensor_split <ts>                       \n");
+    fprintf(stdout, "  -ts, --tensor_split <ts0/ts1/..>               \n");
     fprintf(stdout, "  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : "md");
+    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
     fprintf(stdout, "  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     fprintf(stdout, "\n");
-    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by repeating the parameter.\n");
+    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 
 }
 
@@ -443,6 +441,8 @@ struct test {
     static const std::string gpu_info;
     std::string model_filename;
     std::string model_type;
+    uint64_t model_size;
+    uint64_t model_n_params;
     int n_batch;
     int n_threads;
     bool f32_kv;
@@ -459,8 +459,10 @@ struct test {
     test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
         model_filename = inst.model;
         char buf[128];
-        llama_model_type(lmodel, buf, sizeof(buf));
+        llama_model_desc(lmodel, buf, sizeof(buf));
         model_type = buf;
+        model_size = llama_model_size(lmodel);
+        model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
         f32_kv = inst.f32_kv;
@@ -504,7 +506,7 @@ struct test {
 
     static std::string get_backend() {
         if (cuda) {
-            return "CUDA";
+            return GGML_CUDA_NAME;
         }
         if (opencl) {
             return "OpenCL";
@@ -526,7 +528,7 @@ struct test {
             "build_commit", "build_number",
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
-            "model_filename", "model_type",
+            "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "f16_kv",
             "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
             "n_prompt", "n_gen", "test_time",
@@ -540,6 +542,7 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+            field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
             field == "avg_ns" || field == "stddev_ns") {
@@ -575,7 +578,7 @@ struct test {
             build_commit, std::to_string(build_number),
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
-            model_filename, model_type,
+            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
             std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
@@ -711,8 +714,15 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 15;
+            return 16;
         }
+        if (field == "size" || field == "params") {
+            return 10;
+        }
+        if (field == "n_gpu_layers") {
+            return 3;
+        }
+
         int width = std::max((int)field.length(), 10);
 
         if (test::get_field_type(field) == test::STRING) {
@@ -721,14 +731,33 @@ struct markdown_printer : public printer {
         return width;
     }
 
+    static std::string get_field_display_name(const std::string & field) {
+        if (field == "n_gpu_layers") {
+            return "ngl";
+        }
+        if (field == "n_threads") {
+            return "threads";
+        }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
+        if (field == "tensor_split") {
+            return "ts";
+        }
+        return field;
+    }
+
     void print_header(const cmd_params & params) override {
         // select fields to print
-        fields = { "model", "backend" };
+        fields.push_back("model");
+        fields.push_back("size");
+        fields.push_back("params");
+        fields.push_back("backend");
         bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
         if (!is_cpu_backend) {
             fields.push_back("n_gpu_layers");
         }
-        if (params.n_batch.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
+        if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
             fields.push_back("n_threads");
         }
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
@@ -754,7 +783,7 @@ struct markdown_printer : public printer {
 
         fprintf(fout, "|");
         for (const auto & field : fields) {
-            fprintf(fout, " %*s |", get_field_width(field), field.c_str());
+            fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str());
         }
         fprintf(fout, "\n");
         fprintf(fout, "|");
@@ -771,12 +800,26 @@ struct markdown_printer : public printer {
         fprintf(fout, "|");
         for (const auto & field : fields) {
             std::string value;
+            char buf[128];
             if (field == "model") {
                 value = t.model_type;
+            } else if (field == "size") {
+                if (t.model_size < 1024*1024*1024) {
+                    snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
+                }
+                value = buf;
+            } else if (field == "params") {
+                if (t.model_n_params < 1000*1000*1000) {
+                    snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
+                }
+                value = buf;
             } else if (field == "backend") {
                 value = test::get_backend();
             } else if (field == "test") {
-                char buf[128];
                 if (t.n_prompt > 0 && t.n_gen == 0) {
                     snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
                 } else if (t.n_gen > 0 && t.n_prompt == 0) {
@@ -787,7 +830,6 @@ struct markdown_printer : public printer {
                 }
                 value = buf;
             } else if (field == "t/s") {
-                char buf[128];
                 snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
                 value = buf;
             } else if (vmap.find(field) != vmap.end()) {

examples/main/README.md

@@ -288,6 +288,10 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 -   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
+### Grammars
+
+-   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
+
 ### Quantization
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run).

examples/main/main.cpp

@@ -43,7 +43,7 @@ static bool is_interacting = false;
 void sigint_handler(int signo) {
     if (signo == SIGINT) {
         if (!is_interacting) {
-            is_interacting=true;
+            is_interacting = true;
         } else {
             console::cleanup();
             printf("\n");
@@ -189,23 +189,37 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // Add BOS if SPM tokenizer
+    const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+
     // tokenize the prompt
     std::vector<llama_token> embd_inp;
+
+    if (llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM) {
+        // Add a space in front of the first character to match OG llama tokenizer behavior
+        params.prompt.insert(0, 1, ' ');
+    }
+
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
-        embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
         embd_inp = session_tokens;
     }
 
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+    }
+
     // Tokenize negative prompt
     std::vector<llama_token> guidance_inp;
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
         params.cfg_negative_prompt.insert(0, 1, ' ');
-        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, true);
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
     }
@@ -252,8 +266,8 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
@@ -590,7 +604,12 @@ int main(int argc, char ** argv) {
                     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(ctx)] = nl_logit;
+                    for (size_t idx = 0; idx < candidates_p.size; idx++) {
+                        if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
+                            candidates_p.data[idx].logit = nl_logit;
+                            break;
+                        }
+                    }
                 }
 
                 if (grammar != NULL) {
@@ -791,7 +810,8 @@ int main(int argc, char ** argv) {
         }
 
         // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
-        if (params.interactive && n_remain <= 0 && params.n_predict != -1) {
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
             n_remain = params.n_predict;
             is_interacting = true;
         }

examples/make-ggml.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 This script converts Hugging Face llama models to GGML and quantizes them.
 

examples/perplexity/perplexity.cpp

@@ -6,6 +6,8 @@
 #include <ctime>
 #include <sstream>
 #include <cstring>
+#include <thread>
+#include <mutex>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -27,14 +29,69 @@ std::vector<float> softmax(const std::vector<float>& logits) {
     return probs;
 }
 
-void perplexity(llama_context * ctx, const gpt_params & params) {
+float log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
+    return logits[tok] - max_logit - log(sum_exp);
+}
+
+void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread>& workers,
+        double& nll, double& nll2) {
+
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, n_vocab, logits, tokens, n_token] () {
+        double local_nll = 0, local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            double v = -log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            local_nll += v;
+            local_nll2 += v*v;
+        }
+    };
+    for (auto& w : workers) w = std::thread(compute);
+    compute();
+    for (auto& w : workers) w.join();
+
+}
+
+void perplexity_v2(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
-    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int n_chunk_max = tokens.size() / params.n_ctx;
+    if (params.ppl_stride <= 0) {
+        fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
+        return;
+    }
+
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    const int calc_chunk = params.n_ctx;
+
+    fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
+
+    if (int(tokens.size()) <= calc_chunk) {
+        fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
+                tokens.size(), params.n_ctx, params.ppl_stride);
+        return;
+    }
+
+    const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_vocab = llama_n_vocab(ctx);
@@ -45,6 +102,112 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
 
     fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * params.ppl_stride;
+        const int end   = start + calc_chunk;
+
+        const int num_batches = (calc_chunk + n_batch - 1) / n_batch;
+        //fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
+            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+                //fprintf(stderr, "%s : failed to eval\n", __func__);
+                return;
+            }
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(ctx);
+            }
+
+            const auto batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+
+            if (j == 0) {
+                tokens[batch_start] = token_org;
+            }
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
+        for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) {
+
+            // Calculate probability of next token, given the previous ones.
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+
+            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
+            nll += -std::log(prob);
+            ++count;
+        }
+        // perplexity is e^(average negative log-likelihood)
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            printf("%8d  %.4lf\n", i*params.ppl_stride, std::exp(nll / count));
+        }
+        fflush(stdout);
+    }
+    printf("\n");
+}
+
+void perplexity(llama_context * ctx, const gpt_params & params) {
+    if (params.ppl_stride > 0) {
+        perplexity_v2(ctx, params);
+        return;
+    }
+
+    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+    // Output: `perplexity: 13.5106 [114/114]`
+    // BOS tokens will be added for each chunk before eval
+
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    const int n_chunk_max = tokens.size() / params.n_ctx;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(ctx);
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+    double nll2 = 0.0;
+
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
     for (int i = 0; i < n_chunk; ++i) {
         const int start =     i * params.n_ctx;
         const int end   = start + params.n_ctx;
@@ -63,7 +226,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             const auto token_org = tokens[batch_start];
 
             // add BOS token for the first batch of each chunk
-            if (j == 0) {
+            if (add_bos && j == 0) {
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
@@ -104,22 +267,32 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // Example, we have a context window of 512, we will compute perplexity for each of the
         // last 256 tokens.  Then, we split the input up into context window size chunks to
         // process the entire prompt.
-        for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
-            // Calculate probability of next token, given the previous ones.
-            const std::vector<float> tok_logits(
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
+        const int first = std::min(512, params.n_ctx/2);
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, workers, nll, nll2);
+        count += params.n_ctx - first - 1;
 
-            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
-
-            nll += -std::log(prob);
-            ++count;
-        }
         // perplexity is e^(average negative log-likelihood)
-        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            double av = nll/count;
+            double av2 = nll2/count - av*av;
+            if (av2 > 0) av2 = sqrt(av2/(count-1));
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*params.n_ctx, std::exp(nll / count), av, av2);
+        }
         fflush(stdout);
     }
     printf("\n");
+    nll2 /= count;
+    nll /= count;
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        double ppl = exp(nll);
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
 }
 
 std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
@@ -177,8 +350,11 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    fprintf(stderr, "================================= is_spm = %d\n", is_spm);
+
     // This is needed as usual for LLaMA models
-    bool prepend_bos = true;
+    const bool add_bos = is_spm;
 
     // Number of tasks to use when computing the score
     if ( params.hellaswag_tasks < hs_task_count  ) {
@@ -231,19 +407,30 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     double acc = 0.0f;
     const int n_vocab = llama_n_vocab(ctx);
 
+    std::vector<std::vector<int>> ending_tokens(4);
+
     std::vector<float> tok_logits(n_vocab);
 
     for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
-
         // Tokenize the context to count tokens
-        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
+        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
         size_t context_size = context_embd.size();
 
+        for (int i = 0; i < 4; ++i) {
+            ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[i], add_bos);
+            for (int k = 0; k < int(context_size); ++k) {
+                if (ending_tokens[i][k] != context_embd[k]) {
+                    fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k);
+                    break;
+                }
+            }
+        }
+
         // Do the 1st ending
         // In this case we include the context when evaluating
-        auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], prepend_bos);
+        //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
+        auto query_embd = ending_tokens[0];
         auto query_size = query_embd.size();
-        //printf("First query: %d\n",(int)query_size);
 
         // Stop if query wont fit the ctx window
         if (query_size > (size_t)params.n_ctx) {
@@ -288,7 +475,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) {
 
             // Tokenize the query
-            query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false);
+            query_embd.resize(ending_tokens[ending_idx].size() - context_size);
+            std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int));
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window
@@ -369,6 +557,12 @@ int main(int argc, char ** argv) {
     params.perplexity = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
+    if (params.ppl_stride > 0) {
+        fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
+                params.n_ctx, params.n_ctx + params.ppl_stride/2);
+        params.n_ctx += params.ppl_stride/2;
+    }
+
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);

examples/quantize/quantize.cpp

@@ -14,25 +14,25 @@ struct quant_option {
 };
 
 static const std::vector<struct quant_option> QUANT_OPTIONS = {
-    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.50G, +0.2499 ppl @ 7B", },
-    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1846 ppl @ 7B", },
-    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.30G, +0.0796 ppl @ 7B", },
-    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0415 ppl @ 7B", },
+    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.56G, +0.2166 ppl @ LLaMA-v1-7B", },
+    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
+    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
+    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
 #ifdef GGML_USE_K_QUANTS
-    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.67G, +0.8698 ppl @ 7B", },
+    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
-    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5505 ppl @ 7B", },
-    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.06G, +0.2437 ppl @ 7B", },
-    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1803 ppl @ 7B", },
+    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
+    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
+    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
-    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.56G, +0.1149 ppl @ 7B", },
-    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0535 ppl @ 7B", },
+    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
+    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
     { "Q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
-    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0353 ppl @ 7B", },
-    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0142 ppl @ 7B", },
-    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0044 ppl @ 7B", },
+    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
+    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
+    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
 #endif
-    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ 7B", },
+    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
 };

examples/reason-act.sh

@@ -1,4 +1,3 @@
-
 #!/bin/bash
 
 cd `dirname $0`

examples/server/README.md

@@ -77,34 +77,31 @@ You need to have [Node.js](https://nodejs.org/en) installed.
 ```bash
 mkdir llama-client
 cd llama-client
-npm init
-npm install axios
 ```
 
 Create a index.js file and put inside this:
 
 ```javascript
-const axios = require("axios");
-
 const prompt = `Building a website can be done in 10 simple steps:`;
 
 async function Test() {
-    let result = await axios.post("http://127.0.0.1:8080/completion", {
-        prompt,
-        n_predict: 512,
-    });
-
-    // the response is received until completion finish
-    console.log(result.data.content);
+    let response = await fetch("http://127.0.0.1:8080/completion", {
+        method: 'POST',
+        body: JSON.stringify({
+            prompt,
+            n_predict: 512,
+        })
+    })
+    console.log((await response.json()).content)
 }
 
-Test();
+Test()
 ```
 
 And run it:
 
 ```bash
-node .
+node index.js
 ```
 
 ## API Endpoints
@@ -126,7 +123,7 @@ node .
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
-    `prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. A space is inserted in the front like main.cpp does.
+    `prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does.
 
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).

examples/server/api_like_OAI.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import argparse
 from flask import Flask, jsonify, request, Response
 import urllib.parse

examples/server/public/index.html

@@ -102,6 +102,17 @@
       padding: 0.5em;
     }
 
+    .prob-set {
+      padding: 0.3em;
+      border-bottom: 1px solid #ccc;
+    }
+
+    .popover-content {
+      position: absolute;
+      background-color: white;
+      padding: 0.2em;
+      box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);
+    }
 
     textarea {
       padding: 5px;
@@ -133,11 +144,17 @@
       font-size: 80%;
       color: #888;
     }
+
+    @media (prefers-color-scheme: dark) {
+      .popover-content {
+        background-color: black;
+      }
+    }
   </style>
 
   <script type="module">
     import {
-      html, h, signal, effect, computed, render, useSignal, useEffect, useRef
+      html, h, signal, effect, computed, render, useSignal, useEffect, useRef, Component
     } from '/index.js';
 
     import { llama } from '/completion.js';
@@ -168,6 +185,7 @@
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
       grammar: '',
+      n_probs: 0, // no completion_probabilities
     })
 
     /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -334,10 +352,21 @@
 
       const prompt = template(session.value.template, {
         message: msg,
-        history: session.value.transcript.flatMap(([name, message]) => template(session.value.historyTemplate, {name, message})).join("\n"),
+        history: session.value.transcript.flatMap(
+          ([name, data]) =>
+            template(
+              session.value.historyTemplate,
+              {
+                name,
+                message: Array.isArray(data) ?
+                  data.map(msg => msg.content).join('').replace(/^\s/, '') :
+                  data,
+              }
+            )
+        ).join("\n"),
       });
 
-      let currentMessage = '';
+      const currentMessages = [];
       const history = session.value.transcript
 
       const llamaParams = {
@@ -347,15 +376,19 @@
 
       for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
         const data = chunk.data;
-        currentMessage += data.content;
-
-        // remove leading whitespace
-        currentMessage = currentMessage.replace(/^\s+/, "")
-
-        transcriptUpdate([...history, ["{{char}}", currentMessage]])
 
         if (data.stop) {
-          console.log("Completion finished: '", currentMessage, "', summary: ", data);
+          while (
+            currentMessages.length > 0 &&
+            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
+          ) {
+            currentMessages.pop();
+          }
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
+          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
+        } else {
+          currentMessages.push(data);
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
         }
 
         if (data.timings) {
@@ -420,8 +453,18 @@
         }
       }, [messages])
 
-      const chatLine = ([user, msg]) => {
-        return html`<p key=${msg}><strong>${template(user)}:</strong> <${Markdownish} text=${template(msg)} /></p>`
+      const chatLine = ([user, data], index) => {
+        let message
+        const isArrayMessage = Array.isArray(data)
+        if (params.value.n_probs > 0 && isArrayMessage) {
+          message = html`<${Probabilities} data=${data} />`
+        } else {
+          const text = isArrayMessage ?
+            data.map(msg => msg.content).join('').replace(/^\s+/, '') :
+            data;
+          message = html`<${Markdownish} text=${template(text)} />`
+        }
+        return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
       };
 
       return html`
@@ -568,10 +611,71 @@
               ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
               ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
             </fieldset>
+            <fieldset>
+              ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
+            </fieldset>
           </details>
         </form>
       `
     }
+
+    const probColor = (p) => {
+      const r = Math.floor(192 * (1 - p));
+      const g = Math.floor(192 * p);
+      return `rgba(${r},${g},0,0.3)`;
+    }
+
+    const Probabilities = (params) => {
+      return params.data.map(msg => {
+        const { completion_probabilities } = msg;
+        if (
+          !completion_probabilities ||
+          completion_probabilities.length === 0
+        ) return msg.content
+
+        if (completion_probabilities.length > 1) {
+          // Not for byte pair
+          if (completion_probabilities[0].content.startsWith('byte: \\')) return msg.content
+
+          const splitData = completion_probabilities.map(prob => ({
+            content: prob.content,
+            completion_probabilities: [prob]
+          }))
+          return html`<${Probabilities} data=${splitData} />`
+        }
+
+        const { probs, content } = completion_probabilities[0]
+        const found = probs.find(p => p.tok_str === msg.content)
+        const pColor = found ? probColor(found.prob) : 'transparent'
+
+        const popoverChildren = html`
+          <div class="prob-set">
+            ${probs.map((p, index) => {
+              return html`
+                <div
+                  key=${index}
+                  title=${`prob: ${p.prob}`}
+                  style=${{
+                    padding: '0.3em',
+                    backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
+                  }}
+                >
+                  <span>${p.tok_str}: </span>
+                  <span>${Math.floor(p.prob * 100)}%</span>
+                </div>
+              `
+            })}
+          </div>
+        `
+
+        return html`
+          <${Popover} style=${{ backgroundColor: pColor }} popoverChildren=${popoverChildren}>
+            ${msg.content.match(/\n/gim) ? html`<br />` : msg.content}
+          </>
+        `
+      });
+    }
+
     // poor mans markdown replacement
     const Markdownish = (params) => {
       const md = params.text
@@ -600,10 +704,121 @@
       `
     }
 
+    // simple popover impl
+    const Popover = (props) => {
+      const isOpen = useSignal(false);
+      const position = useSignal({ top: '0px', left: '0px' });
+      const buttonRef = useRef(null);
+      const popoverRef = useRef(null);
+
+      const togglePopover = () => {
+        if (buttonRef.current) {
+          const rect = buttonRef.current.getBoundingClientRect();
+          position.value = {
+            top: `${rect.bottom + window.scrollY}px`,
+            left: `${rect.left + window.scrollX}px`,
+          };
+        }
+        isOpen.value = !isOpen.value;
+      };
+
+      const handleClickOutside = (event) => {
+        if (popoverRef.current && !popoverRef.current.contains(event.target) && !buttonRef.current.contains(event.target)) {
+          isOpen.value = false;
+        }
+      };
+
+      useEffect(() => {
+        document.addEventListener('mousedown', handleClickOutside);
+        return () => {
+          document.removeEventListener('mousedown', handleClickOutside);
+        };
+      }, []);
+
+      return html`
+        <span style=${props.style} ref=${buttonRef} onClick=${togglePopover}>${props.children}</span>
+        ${isOpen.value && html`
+          <${Portal} into="#portal">
+            <div
+              ref=${popoverRef}
+              class="popover-content"
+              style=${{
+                top: position.value.top,
+                left: position.value.left,
+              }}
+            >
+              ${props.popoverChildren}
+            </div>
+          </${Portal}>
+        `}
+      `;
+    };
+
+    // Source: preact-portal (https://github.com/developit/preact-portal/blob/master/src/preact-portal.js)
+    /** Redirect rendering of descendants into the given CSS selector */
+    class Portal extends Component {
+      componentDidUpdate(props) {
+        for (let i in props) {
+          if (props[i] !== this.props[i]) {
+            return setTimeout(this.renderLayer);
+          }
+        }
+      }
+
+      componentDidMount() {
+        this.isMounted = true;
+        this.renderLayer = this.renderLayer.bind(this);
+        this.renderLayer();
+      }
+
+      componentWillUnmount() {
+        this.renderLayer(false);
+        this.isMounted = false;
+        if (this.remote && this.remote.parentNode) this.remote.parentNode.removeChild(this.remote);
+      }
+
+      findNode(node) {
+        return typeof node === 'string' ? document.querySelector(node) : node;
+      }
+
+      renderLayer(show = true) {
+        if (!this.isMounted) return;
+
+        // clean up old node if moving bases:
+        if (this.props.into !== this.intoPointer) {
+          this.intoPointer = this.props.into;
+          if (this.into && this.remote) {
+            this.remote = render(html`<${PortalProxy} />`, this.into, this.remote);
+          }
+          this.into = this.findNode(this.props.into);
+        }
+
+        this.remote = render(html`
+          <${PortalProxy} context=${this.context}>
+            ${show && this.props.children || null}
+          </${PortalProxy}>
+        `, this.into, this.remote);
+      }
+
+      render() {
+        return null;
+      }
+    }
+    // high-order component that renders its first child if it exists.
+    // used as a conditional rendering proxy.
+    class PortalProxy extends Component {
+      getChildContext() {
+        return this.props.context;
+      }
+      render({ children }) {
+        return children || null;
+      }
+    }
+
     function App(props) {
 
       return html`
-        <div id="container">
+        <div>
           <header>
             <h1>llama.cpp</h1>
           </header>
@@ -624,11 +839,13 @@
       `;
     }
 
-    render(h(App), document.body);
+    render(h(App), document.querySelector('#container'));
   </script>
 </head>
 
 <body>
+  <div id="container"></div>
+  <div id="portal"></div>
 </body>
 
 </html>

examples/server/server.cpp

@@ -124,8 +124,9 @@ static void server_log(const char *level, const char *function, int line,
 static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
 {
     std::string out = token == -1 ? "" : llama_token_to_str(ctx, token);
-    // if first bit is 1, meaning it's a partial character
-    if (out.size() > 0 && (out[0] & 0x80) == 0x80)
+    // if the size is 1 and first bit is 1, meaning it's a partial character
+    //   (size > 1 meaning it's already a known token)
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
     {
         std::stringstream ss;
         ss << std::hex << (out[0] & 0xff);
@@ -190,6 +191,7 @@ struct llama_server_context
     size_t n_past = 0;
     size_t n_remain = 0;
 
+    json prompt;
     std::vector<llama_token> embd;
     std::vector<llama_token> last_n_tokens;
 
@@ -267,6 +269,53 @@ struct llama_server_context
         return true;
     }
 
+    std::vector<llama_token> tokenize(json json_prompt, bool add_bos)
+    {
+        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+        // or the first element of the json_prompt array is a string.
+        std::vector<llama_token> prompt_tokens;
+
+        if (json_prompt.is_array())
+        {
+            bool first = true;
+            for (const auto& p : json_prompt)
+            {
+                if (p.is_string())
+                {
+                    auto s = p.template get<std::string>();
+                    std::vector<llama_token> p;
+                    if (first)
+                    {
+                        s.insert(0, 1, ' '); // add a space if it's the first
+                        p = ::llama_tokenize(ctx, s, add_bos);
+                        first = false;
+                    }
+                    else
+                    {
+                        p = ::llama_tokenize(ctx, s, false);
+                    }
+                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+                }
+                else
+                {
+                    if (first)
+                    {
+                        first = false;
+                    }
+                    prompt_tokens.push_back(p.template get<llama_token>());
+                }
+            }
+        }
+        else
+        {
+            auto s = json_prompt.template get<std::string>();
+            s.insert(0, 1, ' '); // always add a first space
+            prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
+        }
+
+        return prompt_tokens;
+    }
+
     bool loadGrammar()
     {
         if (!params.grammar.empty()) {
@@ -294,8 +343,8 @@ struct llama_server_context
 
     void loadPrompt()
     {
-        params.prompt.insert(0, 1, ' '); // always add a first space
-        std::vector<llama_token> prompt_tokens = ::llama_tokenize(ctx, params.prompt, true);
+        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+
         num_prompt_tokens = prompt_tokens.size();
 
         if (params.n_keep < 0)
@@ -671,12 +720,11 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
     fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
     fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
@@ -867,12 +915,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
 #endif // GGML_USE_CUBLAS
         }
-        else if (arg == "--mul-mat-q" || arg == "-mmq")
+        else if (arg == "--no-mul-mat-q" || arg == "-nommq")
         {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")
@@ -1017,7 +1065,7 @@ static json format_final_response(llama_server_context &llama, const std::string
         {"tokens_predicted", llama.num_tokens_predicted},
         {"tokens_evaluated", llama.num_prompt_tokens},
         {"generation_settings", format_generation_settings(llama)},
-        {"prompt", llama.params.prompt},
+        {"prompt", llama.prompt},
         {"truncated", llama.truncated},
         {"stopped_eos", llama.stopped_eos},
         {"stopped_word", llama.stopped_word},
@@ -1056,33 +1104,50 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
         {"tokens", tokens}};
 }
 
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
 static void parse_options_completion(const json &body, llama_server_context &llama)
 {
     gpt_params default_params;
 
-    llama.stream = body.value("stream", false);
-    llama.params.n_predict = body.value("n_predict", default_params.n_predict);
-    llama.params.top_k = body.value("top_k", default_params.top_k);
-    llama.params.top_p = body.value("top_p", default_params.top_p);
-    llama.params.tfs_z = body.value("tfs_z", default_params.tfs_z);
-    llama.params.typical_p = body.value("typical_p", default_params.typical_p);
-    llama.params.repeat_last_n = body.value("repeat_last_n", default_params.repeat_last_n);
-    llama.params.temp = body.value("temperature", default_params.temp);
-    llama.params.repeat_penalty = body.value("repeat_penalty", default_params.repeat_penalty);
-    llama.params.presence_penalty = body.value("presence_penalty", default_params.presence_penalty);
-    llama.params.frequency_penalty = body.value("frequency_penalty", default_params.frequency_penalty);
-    llama.params.mirostat = body.value("mirostat", default_params.mirostat);
-    llama.params.mirostat_tau = body.value("mirostat_tau", default_params.mirostat_tau);
-    llama.params.mirostat_eta = body.value("mirostat_eta", default_params.mirostat_eta);
-    llama.params.penalize_nl = body.value("penalize_nl", default_params.penalize_nl);
-    llama.params.n_keep = body.value("n_keep", default_params.n_keep);
-    llama.params.seed = body.value("seed", default_params.seed);
-    llama.params.prompt = body.value("prompt", default_params.prompt);
-    llama.params.grammar = body.value("grammar", default_params.grammar);
-    llama.params.n_probs = body.value("n_probs", default_params.n_probs);
+    llama.stream = json_value(body, "stream", false);
+    llama.params.n_predict = json_value(body, "n_predict", default_params.n_predict);
+    llama.params.top_k = json_value(body, "top_k", default_params.top_k);
+    llama.params.top_p = json_value(body, "top_p", default_params.top_p);
+    llama.params.tfs_z = json_value(body, "tfs_z", default_params.tfs_z);
+    llama.params.typical_p = json_value(body, "typical_p", default_params.typical_p);
+    llama.params.repeat_last_n = json_value(body, "repeat_last_n", default_params.repeat_last_n);
+    llama.params.temp = json_value(body, "temperature", default_params.temp);
+    llama.params.repeat_penalty = json_value(body, "repeat_penalty", default_params.repeat_penalty);
+    llama.params.presence_penalty = json_value(body, "presence_penalty", default_params.presence_penalty);
+    llama.params.frequency_penalty = json_value(body, "frequency_penalty", default_params.frequency_penalty);
+    llama.params.mirostat = json_value(body, "mirostat", default_params.mirostat);
+    llama.params.mirostat_tau = json_value(body, "mirostat_tau", default_params.mirostat_tau);
+    llama.params.mirostat_eta = json_value(body, "mirostat_eta", default_params.mirostat_eta);
+    llama.params.penalize_nl = json_value(body, "penalize_nl", default_params.penalize_nl);
+    llama.params.n_keep = json_value(body, "n_keep", default_params.n_keep);
+    llama.params.seed = json_value(body, "seed", default_params.seed);
+    llama.params.grammar = json_value(body, "grammar", default_params.grammar);
+    llama.params.n_probs = json_value(body, "n_probs", default_params.n_probs);
+
+    if (body.count("prompt") != 0)
+    {
+        llama.prompt = body["prompt"];
+    }
+    else
+    {
+        llama.prompt = "";
+    }
 
     llama.params.logit_bias.clear();
-    if (body.value("ignore_eos", false))
+    if (json_value(body, "ignore_eos", false))
     {
         llama.params.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
     }
@@ -1144,6 +1209,62 @@ static void log_server_request(const Request &req, const Response &res)
                            });
 }
 
+bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.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.
+void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
+    auto & llama = *static_cast<llama_server_context*>(callback_data);
+    // 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(llama, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        llama.generated_token_probs.resize(llama.generated_token_probs.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
+        std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
+        printf("%lu", n);
+    }
+    fflush(stdout);
+#if 0 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams:\n";
+    for (size_t i=0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{state.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+struct token_translator {
+    llama_context * ctx;
+    std::string operator()(llama_token tok) const { return llama_token_to_str(ctx, tok); }
+    std::string operator()(completion_token_output cto) const { return (*this)(cto.tok); }
+};
+
+void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
+    auto & gtps = llama.generated_token_probs;
+    auto translator = token_translator{llama.ctx};
+    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
+    const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
+    if (llama.generated_text.capacity() < llama.generated_text.size() + len) {
+        llama.generated_text.reserve(llama.generated_text.size() + len);
+    }
+    for (const completion_token_output & cto : gtps) {
+        llama.generated_text += translator(cto);
+    }
+}
+
 int main(int argc, char **argv)
 {
     // own arguments required by this example
@@ -1226,22 +1347,30 @@ int main(int argc, char **argv)
         llama.beginCompletion();
 
         if (!llama.stream) {
-            size_t stop_pos = std::string::npos;
+            if (llama.params.n_beams) {
+                // Fill llama.generated_token_probs vector with final beam.
+                llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
+                                  llama.n_past, llama.n_remain, llama.params.n_threads);
+                // Translate llama.generated_token_probs to llama.generated_text.
+                append_to_generated_text_from_generated_token_probs(llama);
+            } else {
+                size_t stop_pos = std::string::npos;
 
-            while (llama.has_next_token) {
-                const completion_token_output token_with_probs = llama.doCompletion();
-                const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
+                while (llama.has_next_token) {
+                    const completion_token_output token_with_probs = llama.doCompletion();
+                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
 
-                stop_pos = llama.findStoppingStrings(llama.generated_text,
-                    token_text.size(), STOP_FULL);
-            }
+                    stop_pos = llama.findStoppingStrings(llama.generated_text,
+                        token_text.size(), STOP_FULL);
+                }
 
-            if (stop_pos == std::string::npos) {
-                stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
-            }
-            if (stop_pos != std::string::npos) {
-                llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
-                    llama.generated_text.end());
+                if (stop_pos == std::string::npos) {
+                    stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
+                }
+                if (stop_pos != std::string::npos) {
+                    llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
+                        llama.generated_text.end());
+                }
             }
 
             const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
@@ -1257,59 +1386,86 @@ int main(int argc, char **argv)
 
                 while (llama.has_next_token) {
                     const completion_token_output token_with_probs = llama.doCompletion();
-                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
-                    if (llama.multibyte_pending > 0) {
+                    if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
                         continue;
                     }
+                    const std::string token_text = llama_token_to_str(llama.ctx, token_with_probs.tok);
 
                     size_t pos = std::min(sent_count, llama.generated_text.size());
 
                     const std::string str_test = llama.generated_text.substr(pos);
+                    bool is_stop_full = false;
                     size_t stop_pos =
                         llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
                     if (stop_pos != std::string::npos) {
+                        is_stop_full = true;
                         llama.generated_text.erase(
                             llama.generated_text.begin() + pos + stop_pos,
                             llama.generated_text.end());
                         pos = std::min(sent_count, llama.generated_text.size());
                     } else {
+                        is_stop_full = false;
                         stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
                             STOP_PARTIAL);
                     }
 
-                    const std::string to_send = llama.generated_text.substr(pos, stop_pos);
-                    sent_count += to_send.size();
+                    if (
+                        stop_pos == std::string::npos ||
+                        // Send rest of the text if we are at the end of the generation
+                        (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                    ) {
+                        const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
 
-                    std::vector<completion_token_output> probs_output = {};
+                        sent_count += to_send.size();
 
-                    if (llama.params.n_probs > 0) {
-                        const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
-                        size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
-                        size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
-                        if (probs_pos < probs_stop_pos) {
-                            probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                        std::vector<completion_token_output> probs_output = {};
+
+                        if (llama.params.n_probs > 0) {
+                            const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                            size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                            size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                            if (probs_pos < probs_stop_pos) {
+                                probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                            }
+                            sent_token_probs_index = probs_stop_pos;
+                        }
+
+                        const json data = format_partial_response(llama, to_send, probs_output);
+
+                        const std::string str =
+                            "data: " +
+                            data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                            "\n\n";
+
+                        LOG_VERBOSE("data stream", {
+                            { "to_send", str }
+                        });
+
+                        if (!sink.write(str.data(), str.size())) {
+                            LOG_VERBOSE("stream closed", {});
+                            llama_print_timings(llama.ctx);
+                            return false;
                         }
-                        sent_token_probs_index = probs_stop_pos;
                     }
 
-                    const json data = llama.has_next_token
-                                          ? format_partial_response(llama, to_send, probs_output)
-                                          // Generation is done, send extra information.
-                                          : format_final_response(llama, to_send, llama.generated_token_probs);
+                    if (!llama.has_next_token) {
+                        // Generation is done, send extra information.
+                        const json data = format_final_response(llama, "", llama.generated_token_probs);
 
-                    const std::string str =
-                        "data: " +
-                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
-                        "\n\n";
+                        const std::string str =
+                            "data: " +
+                            data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                            "\n\n";
 
-                    LOG_VERBOSE("data stream", {
-                        { "to_send", str }
-                    });
+                        LOG_VERBOSE("data stream", {
+                            { "to_send", str }
+                        });
 
-                    if (!sink.write(str.data(), str.size())) {
-                        LOG_VERBOSE("stream closed", {});
-                        llama_print_timings(llama.ctx);
-                        return false;
+                        if (!sink.write(str.data(), str.size())) {
+                            LOG_VERBOSE("stream closed", {});
+                            llama_print_timings(llama.ctx);
+                            return false;
+                        }
                     }
                 }
 
@@ -1337,8 +1493,11 @@ int main(int argc, char **argv)
         auto lock = llama.lock();
 
         const json body = json::parse(req.body);
-        const std::string content = body.value("content", "");
-        const std::vector<llama_token> tokens = llama_tokenize(llama.ctx, content, false);
+        std::vector<llama_token> tokens;
+        if (body.count("content") != 0)
+        {
+            tokens = llama.tokenize(body["content"], false);
+        }
         const json data = format_tokenizer_response(tokens);
         return res.set_content(data.dump(), "application/json"); });
 
@@ -1350,7 +1509,14 @@ int main(int argc, char **argv)
 
         llama.rewind();
         llama_reset_timings(llama.ctx);
-        llama.params.prompt = body.value("content", "");
+        if (body.count("content") != 0)
+        {
+            llama.prompt = body["content"];
+        }
+        else
+        {
+            llama.prompt = "";
+        }
         llama.params.n_predict = 0;
         llama.loadPrompt();
         llama.beginCompletion();
@@ -1379,7 +1545,7 @@ int main(int argc, char **argv)
                           {
         if (res.status == 400) {
             res.set_content("Invalid request", "text/plain");
-        } else {
+        } else if (res.status != 500) {
             res.set_content("File Not Found", "text/plain");
             res.status = 404;
         } });

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1868,10 +1868,10 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1));                                            assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
         t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd));                 assert_shape_2d(t11->grad, N*n_batch, n_embd);
         t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3));                                            assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
-        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
+        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false));        assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
         t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch));                                  assert_shape_2d(t08->grad, n_embd, N*n_batch);
         t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3));                                            assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
-        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
+        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false));        assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
         t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch));                                  assert_shape_2d(t05->grad, n_embd, N*n_batch);
         t04->grad = expand(gb, ggml_add_inplace(ctx0,
                         ggml_add_inplace(ctx0,

flake.lock

@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1685518550,
-        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
+        "lastModified": 1692799911,
+        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
+        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1685931219,
-        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
+        "lastModified": 1692913444,
+        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
+        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
         "type": "github"
       },
       "original": {

flake.nix

@@ -6,6 +6,9 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
+        name = "llama.cpp";
+        src = ./.;
+        meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
         osSpecific = with pkgs; buildInputs ++
@@ -21,11 +24,17 @@
               CoreGraphics
               CoreVideo
             ]
+          else if isDarwin then
+            with pkgs.darwin.apple_sdk.frameworks; [
+              Accelerate
+              CoreGraphics
+              CoreVideo
+            ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ];
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
         postPatch = ''
@@ -38,35 +47,35 @@
           mv $out/bin/server $out/bin/llama-server
         '';
         cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in {
+      in
+      {
         packages.default = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
-          buildInputs = osSpecific;
+          inherit name src meta postPatch nativeBuildInputs buildInputs postInstall;
           cmakeFlags = cmakeFlags
             ++ (if isAarch64 && isDarwin then [
-              "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-              "-DLLAMA_METAL=ON"
-            ] else [
-              "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=OpenBLAS"
+            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
+            "-DLLAMA_METAL=ON"
+          ] else [
+            "-DLLAMA_BLAS=ON"
+            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
         };
         packages.opencl = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
+          inherit name src meta postPatch nativeBuildInputs postInstall;
           buildInputs = with pkgs; buildInputs ++ [ clblast ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_CLBLAST=ON"
           ];
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
+        };
+        packages.rocm = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_HIPBLAS=1"
+            "-DCMAKE_C_COMPILER=hipcc"
+            "-DCMAKE_CXX_COMPILER=hipcc"
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+          ];
         };
         apps.llama-server = {
           type = "app";
@@ -80,8 +89,13 @@
           type = "app";
           program = "${self.packages.${system}.default}/bin/llama";
         };
+        apps.quantize = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/quantize";
+        };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
+          buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
       });

ggml-alloc.c

@@ -8,6 +8,7 @@
 
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
 
 //#define GGML_ALLOCATOR_DEBUG
 
@@ -67,8 +68,8 @@ struct ggml_allocr {
     struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
     size_t max_size;
     bool measure;
-    int parse_seq[GGML_MAX_NODES];
-    bool has_parse_seq;
+    int parse_seq[GGML_MAX_CONCUR];
+    int parse_seq_len;
 
 #ifdef GGML_ALLOCATOR_DEBUG
     struct ggml_tensor * allocated_tensors[1024];
@@ -76,7 +77,7 @@ struct ggml_allocr {
 };
 
 #ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
         if (alloc->allocated_tensors[i] == NULL) {
             alloc->allocated_tensors[i] = tensor;
@@ -85,7 +86,7 @@ static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tens
     }
     GGML_ASSERT(!"out of allocated_tensors");
 }
-static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
         if (alloc->allocated_tensors[i] == tensor ||
             (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
@@ -238,15 +239,11 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t
     alloc->n_free_blocks++;
 }
 
-void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n) {
-    int pos = 0;
+void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) {
     for (int i = 0; i < n; i++) {
-        if (list[i] != -1) {
-            alloc->parse_seq[pos] = list[i];
-            pos++;
-        }
+        alloc->parse_seq[i] = list[i];
     }
-    alloc->has_parse_seq = true;
+    alloc->parse_seq_len = n;
 }
 
 void ggml_allocr_reset(struct ggml_allocr * alloc) {
@@ -269,7 +266,7 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
         /*.max_size      = */ 0,
         /*.measure       = */ false,
         /*.parse_seq     = */ {0},
-        /*.has_parse_seq = */ false,
+        /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ = {0},
 #endif
@@ -298,7 +295,7 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
         /*.max_size      = */ 0,
         /*.measure       = */ true,
         /*.parse_seq     = */ {0},
-        /*.has_parse_seq = */ false,
+        /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
         /*.allocated_tensors = */ = {0},
 #endif
@@ -445,8 +442,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                         else {
                             AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
                             node->data = parent->data;
+                            return;
                         }
-                        return;
                     }
                 }
             }
@@ -497,69 +494,86 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 allocate_node(alloc, input);
             }
         }
-        for (int ind = 0; ind < gf->n_nodes; ind++) {
-            int i;
-            if (alloc->has_parse_seq) {
-                i = alloc->parse_seq[ind];
-            } else {
-                i = ind;
-            }
-            struct ggml_tensor * node = gf->nodes[i];
+        // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
+        int last_barrier_pos = 0;
+        int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes;
 
-            // allocate parents (leafs)
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
+        for (int ind = 0; ind < n_nodes; ind++) {
+            // allocate a node if there is no parse_seq or this is not a barrier
+            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) {
+                int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind;
+                struct ggml_tensor * node = gf->nodes[i];
+
+                // allocate parents (leafs)
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * parent = node->src[j];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    allocate_node(alloc, parent);
                 }
-                allocate_node(alloc, parent);
+
+                // allocate node
+                allocate_node(alloc, node);
+
+                AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * parent = node->src[j];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    AT_PRINTF("%s", parent->name);
+                    if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                        AT_PRINTF(", ");
+                    }
+                }
+                AT_PRINTF("\n");
             }
 
-            // allocate node
-            allocate_node(alloc, node);
-
-            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                AT_PRINTF("%s", parent->name);
-                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                    AT_PRINTF(", ");
-                }
-            }
-            AT_PRINTF("\n");
 
             // update parents
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
+            // update immediately if there is no parse_seq
+            // update only at barriers if there is parse_seq
+            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
+                int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
+                int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
+                for (int i = update_start; i < update_end; i++) {
+                    int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i;
+                    struct ggml_tensor * node = gf->nodes[node_i];
+
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * parent = node->src[j];
+                        if (parent == NULL) {
+                            break;
+                        }
+                        struct hash_node * p_hn = hash_get(ht, parent);
+                        p_hn->n_children -= 1;
+
+                        //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                        if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                            if (ggml_is_view(parent)) {
+                                struct ggml_tensor * view_src = get_view_source(parent);
+                                struct hash_node * view_src_hn = hash_get(ht, view_src);
+                                view_src_hn->n_views -= 1;
+                                AT_PRINTF("view_src %s\n", view_src->name);
+                                if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                                    ggml_allocator_free_tensor(alloc, view_src);
+                                }
+                            }
+                            else {
+                                if (parent->data != node->data) {
+                                    ggml_allocator_free_tensor(alloc, parent);
+                                }
+                            }
+                        }
+                    }
                 }
-                struct hash_node * p_hn = hash_get(ht, parent);
-                p_hn->n_children -= 1;
-
-                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
-                        struct ggml_tensor * view_src = get_view_source(parent);
-                        struct hash_node * view_src_hn = hash_get(ht, view_src);
-                        view_src_hn->n_views -= 1;
-                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
-                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, view_src);
-                        }
-                    }
-                    else {
-                        if (parent->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, parent);
-                        }
-                    }
+                AT_PRINTF("\n");
+                if (alloc->parse_seq_len) {
+                    last_barrier_pos = ind + 1;
                 }
             }
-            AT_PRINTF("\n");
         }
         // free graph outputs here that wouldn't be freed otherwise because they have no children
         if (outputs != NULL && outputs[g] != NULL) {

ggml-alloc.h

@@ -12,7 +12,7 @@ GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n);
+GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
 
 GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
 GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);

ggml-cuda.cu

@@ -6,15 +6,116 @@
 #include <atomic>
 #include <assert.h>
 
+#if defined(GGML_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasCreate hipblasCreate
+#define cublasGemmEx hipblasGemmEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event) hipStreamWaitEvent(stream, event, 0)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#else
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
 #include <cuda_fp16.h>
+#endif
 
 #include "ggml-cuda.h"
 #include "ggml.h"
 
 #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#ifndef CC_TURING
 #define CC_TURING   700
+#endif
+
+#if defined(GGML_USE_HIPBLAS)
+#define __CUDA_ARCH__ 1300
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int&>(c);
+}
+
+static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
+#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
+    c = __builtin_amdgcn_sdot4(a, b, c, false);
+#elif defined(__gfx1100__)
+    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+#elif defined(__gfx1010__) || defined(__gfx900__)
+    int tmp1;
+    int tmp2;
+    asm("\n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        "
+        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
+        : "v"(a), "v"(b)
+    );
+#else
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
+#endif
+    return c;
+}
+#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -259,6 +360,7 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_CPY_BLOCK_SIZE 32
 #define CUDA_SCALE_BLOCK_SIZE 256
 #define CUDA_ROPE_BLOCK_SIZE 256
+#define CUDA_ALIBI_BLOCK_SIZE 32
 #define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
 #define CUDA_QUANTIZE_BLOCK_SIZE 256
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
@@ -286,7 +388,7 @@ static int g_device_count = -1;
 static int g_main_device = 0;
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
-static bool g_mul_mat_q = false;
+static bool g_mul_mat_q = true;
 
 static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default
@@ -423,8 +525,8 @@ static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
 
     const int vui = x[ib].qs[iqs];
 
@@ -466,8 +568,8 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -519,8 +621,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
     const uint8_t q = x[i].qs[32*n + l];
     float * y = yy + i*QK_K + 128*n;
 
-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
     y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
@@ -530,8 +632,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
     const int il = tid%16;  // 0...15
     const uint8_t q = x[i].qs[il] >> (2*is);
     float * y = yy + i*QK_K + 16*is + il;
-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
     y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
     y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
 #endif
@@ -617,8 +719,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float
 
     float * y = yy + i*QK_K + 64*il + n*ir;
 
-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
 
     const uint8_t * q = x[i].qs + 32*il + n*ir;
 
@@ -656,8 +758,8 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float
 
     float * y = yy + i*QK_K + 64*il + 2*ir;
 
-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
 
     const uint8_t * ql = x[i].qs + 32*il + 2*ir;
     const uint8_t * qh = x[i].qh + 2*ir;
@@ -769,8 +871,8 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
         const float   * y = yy + i * QK_K + y_offset;
         const uint8_t * q = x[i].qs + q_offset;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
         aux[0] = a[0] & 0x0f0f0f0f;
@@ -990,8 +1092,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         const float   * y1 = yy + i*QK_K + y_offset;
         const float   * y2 = y1 + 128;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux[0] = a[im+0] & kmask1;
@@ -1123,8 +1225,8 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
         const float   * y1  = yy + i*QK_K + y_offset;
         const float   * y2  = y1 + 128;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux[0] = a[im+0] & kmask1;
@@ -1347,8 +1449,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
         return;
     }
 
-    y[ib].ds.x = d;
-    y[ib].ds.y = sum;
+    reinterpret_cast<half&>(y[ib].ds.x) = d;
+    reinterpret_cast<half&>(y[ib].ds.y) = sum;
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
@@ -2345,7 +2447,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
         u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
     }
 
-    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, bq8_1->ds.x);
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
 }
 
 template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
@@ -2431,7 +2533,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR2_K; ++ i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@@ -2550,7 +2652,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR3_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@@ -2719,7 +2821,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
 
     for (int i = 0; i < QR4_K; ++i) {
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2half(bq8i->ds);
 
         const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
         u[2*i+0] = q8[0];
@@ -2746,8 +2848,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     const float dall = bq4_K->d[0];
     const float dmin = bq4_K->d[1];
 
-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2float(bq8_1[0].ds);
+    const float d8_2 = __low2float(bq8_1[1].ds);
 
     const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
     const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@@ -2900,7 +3002,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR5_K; ++i) {
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2float(bq8i->ds);
 
         const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
         u[2*i+0] = q8[0];
@@ -2918,8 +3020,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 
     const float d = bq5_K->d;
 
-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2half(bq8_1[0].ds);
+    const float d8_2 = __low2half(bq8_1[1].ds);
 
     const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
     const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@@ -3074,7 +3176,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR6_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + 2*i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
     }
 
     return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@@ -3242,7 +3344,7 @@ static __device__ __forceinline__ void mul_mat_q(
                     *dsi_dst = *dsi_src;
                 } else {
                     float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = (*dsi_src).x;
+                    *dfi_dst = __low2half(*dsi_src);
                 }
             }
 
@@ -3886,13 +3988,13 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
 // rope == RoPE == rotary positional embedding
 static __global__ void rope_f32(const float * x, float * dst, const int ncols, const float p0,
                                 const float p_delta, const int p_delta_rows, const float theta_scale) {
-    const int col = 2*(blockDim.x*blockIdx.x + threadIdx.x);
+    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (col >= ncols) {
         return;
     }
 
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
     const int i = row*ncols + col;
 
     const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2);
@@ -3906,6 +4008,28 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c
     dst[i + 1] = x0*sin_theta + x1*cos_theta;
 }
 
+static __global__ void rope_neox_f32(const float * x, float * dst, const int ncols, const float p0,
+                                const float p_delta, const int p_delta_rows, const float theta_scale) {
+    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i = row*ncols + col/2;
+
+    const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2);
+    const float sin_theta = sinf(theta);
+    const float cos_theta = cosf(theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + ncols/2];
+
+    dst[i + 0]       = x0*cos_theta - x1*sin_theta;
+    dst[i + ncols/2] = x0*sin_theta + x1*cos_theta;
+}
+
 static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) {
     const int col = blockDim.x*blockIdx.x + threadIdx.x;
     const int half_n_dims = ncols/4;
@@ -3940,9 +4064,32 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol
     dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
 }
 
-static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
+static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
+                                 const int n_heads_log2_floor, const float m0, const float m1) {
     const int col = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
     const int row = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i = row*ncols + col;
+
+    const int k = row/k_rows;
+
+    float m_k;
+    if (k < n_heads_log2_floor) {
+        m_k = powf(m0, k + 1);
+    } else {
+        m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
+    }
+
+    dst[i] = col * m_k + x[i];
+}
+
+static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
+    const int col = blockDim.y*blockIdx.y + threadIdx.y;
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (col >= ncols) {
         return;
@@ -3955,24 +4102,29 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
 
 // the CUDA soft max implementation differs from the CPU implementation
 // instead of doubles floats are used
-// values are also not normalized to the maximum value by subtracting it in the exponential function
-// theoretically these changes could cause problems with rounding error and arithmetic overflow but for LLaMa it seems to be fine
 static __global__ void soft_max_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int block_size = blockDim.x;
-    const int tid = threadIdx.x;
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+    const int block_size = blockDim.y;
+    const int tid = threadIdx.y;
 
-    float tmp = 0.0;
-
-    for (int block_start = 0; block_start < ncols; block_start += block_size) {
-        const int col = block_start + tid;
-
-        if (col >= ncols) {
-            break;
-        }
+    float max_val = -INFINITY;
 
+    for (int col = tid; col < ncols; col += block_size) {
         const int i = row*ncols + col;
-        const float val = expf(x[i]);
+        max_val = max(max_val, x[i]);
+    }
+
+    // find the max value in the block
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        max_val = max(max_val, __shfl_xor_sync(0xffffffff, max_val, mask, 32));
+    }
+
+    float tmp = 0.f;
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const int i = row*ncols + col;
+        const float val = expf(x[i] - max_val);
         tmp += val;
         dst[i] = val;
     }
@@ -3983,15 +4135,11 @@ static __global__ void soft_max_f32(const float * x, float * dst, const int ncol
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
     }
 
-    for (int block_start = 0; block_start < ncols; block_start += block_size) {
-        const int col = block_start + tid;
-
-        if (col >= ncols) {
-            break;
-        }
+    const float inv_tmp = 1.f / tmp;
 
+    for (int col = tid; col < ncols; col += block_size) {
         const int i = row*ncols + col;
-        dst[i] /= tmp;
+        dst[i] *= inv_tmp;
     }
 }
 
@@ -4751,13 +4899,21 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons
 
 static void rope_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0,
                           const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
-    GGML_ASSERT(nrows % 2 == 0);
-    const dim3 block_dims(2*CUDA_ROPE_BLOCK_SIZE, 1, 1);
+    GGML_ASSERT(nrows % 2 == 0); // GG: is this assert really needed? I don't see why
+    const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1);
     const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
-    const dim3 block_nums(num_blocks_x, nrows, 1);
+    const dim3 block_nums(nrows, num_blocks_x, 1);
     rope_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale);
 }
 
+static void rope_neox_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0,
+                          const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
+    const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nrows, num_blocks_x, 1);
+    rope_neox_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale);
+}
+
 static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float block_p, const float theta_scale, cudaStream_t stream) {
     GGML_ASSERT(nrows % 4 == 0);
     const dim3 block_dims(4*CUDA_ROPE_BLOCK_SIZE, 1, 1);
@@ -4766,16 +4922,25 @@ static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, con
     rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p, block_p, theta_scale);
 }
 
+static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
+                           const int k_rows, const int n_heads_log2_floor, const float m0,
+                           const float m1, cudaStream_t stream) {
+    const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
+    const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
+    const dim3 block_nums(num_blocks_x, nrows, 1);
+    alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
+}
+
 static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
-    const dim3 block_dims(CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1, 1);
+    const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
     const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
-    const dim3 block_nums(block_num_x, nrows_x, 1);
+    const dim3 block_nums(nrows_x, block_num_x, 1);
     diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
 }
 
 static void soft_max_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, cudaStream_t stream) {
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    const dim3 block_nums(1, nrows_x, 1);
+    const dim3 block_dims(1, WARP_SIZE, 1);
+    const dim3 block_nums(nrows_x, 1, 1);
     soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x);
 }
 
@@ -4880,10 +5045,18 @@ void ggml_init_cublas() {
     static bool initialized = false;
 
     if (!initialized) {
+
+#ifdef __HIP_PLATFORM_AMD__
+        // Workaround for a rocBLAS bug when using multiple graphics cards:
+        // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
+        rocblas_initialize();
+        CUDA_CHECK(cudaDeviceSynchronize());
+#endif
+
         CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
         GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
         int64_t total_vram = 0;
-        fprintf(stderr, "%s: found %d CUDA devices:\n", __func__, g_device_count);
+        fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
         for (int id = 0; id < g_device_count; ++id) {
             cudaDeviceProp prop;
             CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
@@ -5481,7 +5654,8 @@ inline void ggml_cuda_op_rope(
 
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
-    const bool is_glm = mode & 4;
+    const bool is_neox = mode & 2;
+    const bool is_glm  = mode & 4;
 
     // compute
     if (is_glm) {
@@ -5489,6 +5663,10 @@ inline void ggml_cuda_op_rope(
         const float id_p = min(p, n_ctx - 2.f);
         const float block_p = max(p - (n_ctx - 2.f), 0.f);
         rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main);
+    } else if (is_neox) {
+        GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet");
+        const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale;
+        rope_neox_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main);
     } else {
         const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale;
         rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main);
@@ -5501,6 +5679,41 @@ inline void ggml_cuda_op_rope(
     (void) i1;
 }
 
+inline void ggml_cuda_op_alibi(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
+    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
+    cudaStream_t & cudaStream_main){
+
+    GGML_ASSERT(src0_ddf_i != nullptr);
+    GGML_ASSERT(dst_ddf_i != nullptr);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t i01_diff = i01_high - i01_low;
+
+    const int n_past = ((int32_t *) dst->op_params)[0];
+    const int n_head = ((int32_t *) dst->op_params)[1];
+    float max_bias;
+    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
+
+    GGML_ASSERT(ne01 + n_past == ne00);
+    GGML_ASSERT(n_head == ne02);
+
+    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
+
+    // compute
+    alibi_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_heads_log2_floor, m0, m1, cudaStream_main);
+
+    (void) src1;
+    (void) src0_ddq_i;
+    (void) src1_ddf_i;
+    (void) i1;
+}
+
 inline void ggml_cuda_op_diag_mask_inf(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
     float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
@@ -6121,6 +6334,11 @@ void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_ten
     ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm
 }
 
+void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_alibi, true, true);
+}
+
 void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     (void) src0;
     (void) src1;
@@ -6240,7 +6458,7 @@ static struct ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
     return extra;
 }
 
-void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace) {
+void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace, bool no_alloc) {
     if (scratch && g_scratch_size == 0) {
         return;
     }
@@ -6249,14 +6467,19 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
     if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
         const ggml_op src0_op = tensor->src[0]->op;
         if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) {
-            ggml_cuda_assign_buffers_impl(tensor->src[0], scratch, force_inplace);
+            ggml_cuda_assign_buffers_impl(tensor->src[0], scratch, force_inplace, no_alloc);
         }
     }
     if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) {
-        ggml_cuda_assign_buffers_impl(tensor->src[1], scratch, force_inplace);
+        ggml_cuda_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc);
     }
 
     tensor->backend = GGML_BACKEND_GPU;
+
+    if (scratch && no_alloc) {
+        return;
+    }
+
     struct ggml_tensor_extra_gpu * extra;
 
     const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
@@ -6308,16 +6531,48 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
     tensor->extra = extra;
 }
 
+void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset) {
+    if (g_scratch_size == 0) {
+        return;
+    }
+    if (g_scratch_buffer == nullptr) {
+        CUDA_CHECK(cudaMalloc(&g_scratch_buffer, g_scratch_size));
+    }
+
+    struct ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
+
+    const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
+        tensor->op == GGML_OP_VIEW;
+
+    if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
+        struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
+        char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
+        size_t view_offset = 0;
+        if (tensor->op == GGML_OP_VIEW) {
+            memcpy(&view_offset, tensor->op_params, sizeof(size_t));
+        }
+        extra->data_device[g_main_device] = src0_ddc + view_offset;
+    } else {
+        extra->data_device[g_main_device] = (char *) g_scratch_buffer + offset;
+    }
+
+    tensor->extra = extra;
+}
+
 void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, true, false);
+    ggml_cuda_assign_buffers_impl(tensor, true, false, false);
+}
+
+void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor) {
+    ggml_cuda_assign_buffers_impl(tensor, true, false, true);
 }
 
 void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, false, false);
+    ggml_cuda_assign_buffers_impl(tensor, false, false, false);
 }
 
 void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, false, true);
+    ggml_cuda_assign_buffers_impl(tensor, false, true, false);
 }
 
 void ggml_cuda_set_main_device(int main_device) {
@@ -6456,6 +6711,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             }
             func = ggml_cuda_rope;
             break;
+        case GGML_OP_ALIBI:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cuda_alibi;
+            break;
         default:
             return false;
     }

ggml-cuda.h

@@ -2,6 +2,14 @@
 
 #include "ggml.h"
 
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -16,9 +24,14 @@ GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const str
 GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
 GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+
 GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+
+GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+
 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);

ggml-metal.m

@@ -63,6 +63,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q8_0);
     GGML_METAL_DECL_KERNEL(get_rows_q2_K);
     GGML_METAL_DECL_KERNEL(get_rows_q3_K);
     GGML_METAL_DECL_KERNEL(get_rows_q4_K);
@@ -73,6 +74,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
@@ -81,6 +83,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
@@ -167,7 +170,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
         ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
+        fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
+                (int) ctx->pipeline_##name.threadExecutionWidth); \
         if (error) { \
             fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
@@ -186,6 +191,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q8_0);
         GGML_METAL_ADD_KERNEL(get_rows_q2_K);
         GGML_METAL_ADD_KERNEL(get_rows_q3_K);
         GGML_METAL_ADD_KERNEL(get_rows_q4_K);
@@ -196,6 +202,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
@@ -203,6 +210,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
@@ -218,12 +226,12 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    fprintf(stderr, "%s: hasUnifiedMemory             = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    fprintf(stderr, "%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    fprintf(stderr, "%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
     if (ctx->device.maxTransferRate != 0) {
-        fprintf(stderr, "%s: maxTransferRate              = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        fprintf(stderr, "%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
-        fprintf(stderr, "%s: maxTransferRate              = built-in GPU\n", __func__);
+        fprintf(stderr, "%s: maxTransferRate               = built-in GPU\n", __func__);
     }
 
     return ctx;
@@ -537,8 +545,8 @@ void ggml_metal_graph_compute(
 
             id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
 
-            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
 
             for (int ind = node_start; ind < node_end; ++ind) {
                 const int i = has_concur ? ctx->concur_list[ind] : ind;
@@ -744,32 +752,32 @@ void ggml_metal_graph_compute(
                                 [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                                 ne00%32 == 0 &&
                                 ne11 > 1) {
-                                    switch (src0->type) {
-                                        case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
-                                        case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
-                                        case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
-                                        case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
-                                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
-                                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
-                                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
-                                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
-                                        default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
-                                    }
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
-                                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
-                                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
-                                    [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
-                                    [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
-                                    [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
-                                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                switch (src0->type) {
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
+                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
+                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
+                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
+                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
+                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
+                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
+                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                 }
-                            else {
+                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
 
@@ -799,6 +807,15 @@ void ggml_metal_graph_compute(
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
                                         } break;
+                                    case GGML_TYPE_Q8_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32];
+                                        } break;
                                     case GGML_TYPE_Q2_K:
                                         {
                                             GGML_ASSERT(ne02 == 1);
@@ -868,24 +885,24 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
-                                [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
+                                [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -895,9 +912,10 @@ void ggml_metal_graph_compute(
                     case GGML_OP_GET_ROWS:
                         {
                             switch (src0->type) {
-                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
+                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
                                 case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
                                 case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
                                 case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
@@ -938,16 +956,17 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_NORM:
                         {
-                            const float eps = 1e-5f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
                             const int nth = 256;
 
                             [encoder setComputePipelineState:ctx->pipeline_norm];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
-                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
                             [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
@@ -990,7 +1009,9 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
                             [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
                             [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
+
                             const int nth = 32;
+
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ROPE:
@@ -1005,8 +1026,8 @@ void ggml_metal_graph_compute(
                             memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
                             [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
                             [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
@@ -1057,24 +1078,24 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;

ggml-metal.metal

@@ -18,6 +18,12 @@ typedef struct {
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
 
+#define QK8_0 32
+typedef struct {
+    half    d;         // delta
+    int8_t  qs[QK8_0]; // quants
+} block_q8_0;
+
 kernel void kernel_add(
         device const float * src0,
         device const float * src1,
@@ -87,7 +93,12 @@ kernel void kernel_gelu(
     device       float * dst,
     uint tpig[[thread_position_in_grid]]) {
     float x = src0[tpig];
-    dst[tpig] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
 kernel void kernel_soft_max(
@@ -352,7 +363,7 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
     const int first_row = (r0 * nsg + sgitg) * nr;
     const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
     device const block_q_type * x = (device const block_q_type *) src0 + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
     float yl[16];       // src1 vector cache
     float sumf[nr]={0.f};
 
@@ -424,6 +435,68 @@ kernel void kernel_mul_mat_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
+kernel void kernel_mul_mat_q8_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = ne00/QK8_0;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * nsg + sgitg) * nr;
+    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+    device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];
+    float sumf[nr]={0.f};
+
+    const int ix = tiisg/2;
+    const int il = tiisg%2;
+
+    device const float * yb = y + ix * QK8_0 + 16*il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += nw/2) {
+        for (int i = 0; i < 16; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            device const int8_t * qs = x[ib+row*nb].qs + 16*il;
+            float sumq = 0.f;
+            for (int iq = 0; iq < 16; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*x[ib+row*nb].d;
+        }
+
+        yb += QK8_0 * 16;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+        }
+    }
+}
+
 kernel void kernel_mul_mat_f16_f32(
         device const  char * src0,
         device const  char * src1,
@@ -475,7 +548,6 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
-
 kernel void kernel_alibi_f32(
         device const float * src0,
         device       float * dst,
@@ -571,7 +643,25 @@ kernel void kernel_rope(
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         }
     } else {
-        // TODO: implement
+        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
+            for (int64_t ic = 0; ic < n_dims; ic += 2) {
+                const float cos_theta = cos(theta);
+                const float sin_theta = sin(theta);
+
+                theta *= theta_scale;
+
+                const int64_t i0 = ib*n_dims + ic/2;
+
+                device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                const float x0 = src[0];
+                const float x1 = src[n_dims/2];
+
+                dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            }
+        }
     }
 }
 
@@ -1598,12 +1688,12 @@ template <typename type4x4>
 void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 1);
     const half d = il ? (xb->d / 16.h) : xb->d;
-    const half m = il ? (-8.h * 16.h) : -8.h;
+    const half m = il ? ( -8.h * 16.h) : -8.h;
     const ushort mask0 = il ? 0x00F0 : 0x000F;
     const ushort mask1 = il ? 0xF000 : 0x0F00;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)] = (((qs[i] & mask0)) + m) * d;
+        reg[i/2][2*(i%2)]   = (((qs[i] & mask0)     ) + m) * d;
         reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) + m) * d;
     }
 }
@@ -1617,11 +1707,21 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
     const ushort mask1 = il ? 0xF000 : 0x0F00;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)] = (((qs[i] & mask0)) * d) + m;
+        reg[i/2][2*(i%2)]   = (((qs[i] & mask0)     ) * d) + m;
         reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) * d) + m;
     }
 }
 
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const half d = xb->d;
+
+    for (int i=0;i<16;i++) {
+        reg[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+}
+
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
     const half d = xb->d;
@@ -1850,6 +1950,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         //load data and store to threadgroup memory
         half4x4 temp_a;
         dequantize_func(x, il, temp_a);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
         #pragma unroll(16)
         for (int i = 0; i < 16; i++) {
             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
@@ -1895,14 +1996,14 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
         for (int i = 0; i < 8; i++) {
-            threadgroup_barrier(mem_flags::mem_device);
             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
-        threadgroup_barrier(mem_flags::mem_device);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
         device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
         if (sgitg==0) {
             for (int i = 0; i < n_rows; i++) {
@@ -1923,9 +2024,10 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
                           constant uint64_t &, constant uint64_t &, uint, uint, uint);
 
-template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows<half4x4, 1, dequantize_f16>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
@@ -1936,9 +2038,10 @@ typedef void (mat_mm_t)(device const uchar *, device const float *, device float
                              constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \
                              constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
+template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;

ggml.h

@@ -211,11 +211,12 @@
 #define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
+
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
 #define GGUF_MAGIC   0x46554747 // "GGUF"
-#define GGUF_VERSION 1
+#define GGUF_VERSION 2
 
 #define GGUF_DEFAULT_ALIGNMENT 32
 
@@ -259,8 +260,9 @@
 extern "C" {
 #endif
 
-#ifdef __ARM_NEON
-    // we use the built-in 16-bit float type
+#if defined(__ARM_NEON) && defined(__CUDACC__)
+    typedef half ggml_fp16_t;
+#elif defined(__ARM_NEON)
     typedef __fp16 ggml_fp16_t;
 #else
     typedef uint16_t ggml_fp16_t;
@@ -344,10 +346,12 @@ extern "C" {
         GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
+        GGML_OP_CONCAT,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
         GGML_OP_RMS_NORM_BACK,
+        GGML_OP_GROUP_NORM,
 
         GGML_OP_MUL_MAT,
         GGML_OP_OUT_PROD,
@@ -373,14 +377,19 @@ extern "C" {
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
 
+        GGML_OP_UPSCALE, // nearest interpolate
+
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
+        GGML_OP_GET_REL_POS,
+        GGML_OP_ADD_REL_POS,
 
         GGML_OP_UNARY,
 
@@ -804,6 +813,13 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // concat a and b on dim 2
+    // used in stable-diffusion
+    GGML_API struct ggml_tensor * ggml_concat(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -893,14 +909,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // normalize along rows
-    // TODO: eps is hardcoded to 1e-5 for now
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
@@ -912,6 +929,19 @@ extern "C" {
             struct ggml_tensor  * a,
             float                 eps);
 
+    // group normalize along ne0*ne1*n_groups
+    // used in stable-diffusion
+    // TODO: eps is hardcoded to 1e-6 for now
+    GGML_API struct ggml_tensor * ggml_group_norm(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_groups);
+
+    GGML_API struct ggml_tensor * ggml_group_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_groups);
+
     // a - x
     // b - dy
     // TODO: update with configurable eps
@@ -1212,6 +1242,15 @@ extern "C" {
             float                 freq_base,
             float                 freq_scale);
 
+    // xPos RoPE, in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            float                 base,
+            bool                  down);
+
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
     GGML_API struct ggml_tensor * ggml_rope_back(
@@ -1220,7 +1259,11 @@ extern "C" {
             int                   n_past,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx);
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 xpos_base,
+            bool                  xpos_down);
 
     // alibi position embedding
     // in-place, returns view(a)
@@ -1247,6 +1290,15 @@ extern "C" {
             int                   p0,  // padding
             int                   d0); // dilation
 
+    // conv_1d with padding = half
+    // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
+    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s,
+            int                   d);
+
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1258,14 +1310,38 @@ extern "C" {
             int                   d0,
             int                   d1);
 
-    // conv_1d with padding = half
-    // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
-    GGML_API struct ggml_tensor * ggml_conv_1d_ph(
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is equal to kernel size
+    // padding is zero
+    // example:
+    // a:     16   16    3  768
+    // b:   1024 1024    3    1
+    // res:   64   64  768    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is 1
+    // padding is half
+    // example:
+    // a:      3    3    256  256
+    // b:     64   64    256    1
+    // res:   64   64    256    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            int                   s,
-            int                   d);
+            int                   stride);
 
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
@@ -1292,6 +1368,13 @@ extern "C" {
             int                   p0,
             int                   p1);
 
+    // nearest interpolate
+    // used in stable-diffusion
+    GGML_API struct ggml_tensor * ggml_upscale(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   scale_factor);
+
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1345,6 +1428,27 @@ extern "C" {
         struct ggml_tensor  * a,
         enum ggml_unary_op op);
 
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_get_rel_pos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   qh,
+            int                   kh);
+
+    // used in sam
+
+    GGML_API struct ggml_tensor * ggml_add_rel_pos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * pw,
+            struct ggml_tensor  * ph);
+
+    GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * pw,
+            struct ggml_tensor  * ph);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@@ -1499,6 +1603,7 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * tensor);
 
+
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
@@ -1722,6 +1827,9 @@ extern "C" {
         GGUF_TYPE_BOOL    = 7,
         GGUF_TYPE_STRING  = 8,
         GGUF_TYPE_ARRAY   = 9,
+        GGUF_TYPE_UINT64  = 10,
+        GGUF_TYPE_INT64   = 11,
+        GGUF_TYPE_FLOAT64 = 12,
         GGUF_TYPE_COUNT,       // marks the end of the enum
     };
 
@@ -1762,6 +1870,9 @@ extern "C" {
     GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);
     GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);
     GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);
+    GGML_API uint64_t     gguf_get_val_u64 (struct gguf_context * ctx, int i);
+    GGML_API int64_t      gguf_get_val_i64 (struct gguf_context * ctx, int i);
+    GGML_API double       gguf_get_val_f64 (struct gguf_context * ctx, int i);
     GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);
     GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
     GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);
@@ -1781,6 +1892,9 @@ extern "C" {
     GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
     GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t  val);
     GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float    val);
+    GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
+    GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t  val);
+    GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double   val);
     GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool     val);
     GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
     GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);

gguf-py/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

gguf-py/README.md

@@ -0,0 +1,55 @@
+## gguf
+
+This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302)
+(GGML Universal File) format.
+
+See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py)
+as an example for its usage.
+
+## Installation
+```sh
+pip install gguf
+```
+
+## Development
+Maintainers who participate in development of this package are advised to install it in editable mode:
+
+```sh
+cd /path/to/llama.cpp/gguf-py
+
+pip install --editable .
+```
+
+**Note**: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires `setup.py`.
+In this case, upgrade Pip to the latest:
+
+```sh
+pip install --upgrade pip
+```
+
+## Publishing
+To publish the package, you need to have `twine` and `build` installed:
+
+```sh
+pip install build twine
+```
+
+Then, folow these steps to release a new version:
+
+1. Update the version in `pyproject.toml`.
+2. Build the package:
+
+```sh
+python -m build
+```
+
+3. Upload the generated distribution archives:
+
+```sh
+python -m twine upload dist/*
+```
+
+## TODO
+- [ ] Add tests
+- [ ] Include conversion scripts as command line entry points in this package.
+- Add CI workflow for releasing the package.

gguf-py/gguf/__init__.py

@@ -0,0 +1 @@
+from .gguf import *

gguf-py/gguf/gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import shutil
 import sys
 import struct
@@ -12,7 +13,7 @@ from typing import Any, IO, List, Optional
 #
 
 GGUF_MAGIC             = 0x46554747
-GGUF_VERSION           = 1
+GGUF_VERSION           = 2
 GGUF_DEFAULT_ALIGNMENT = 32
 
 # general
@@ -26,14 +27,15 @@ KEY_GENERAL_DESCRIPTION          = "general.description"
 KEY_GENERAL_LICENSE              = "general.license"
 KEY_GENERAL_SOURCE_URL           = "general.source.url"
 KEY_GENERAL_SOURCE_HF_REPO       = "general.source.hugginface.repository"
+KEY_GENERAL_FILE_TYPE            = "general.file_type"
 
 # LLM
-KEY_LLM_CONTEXT_LENGTH        = "{arch}.context_length"
-KEY_LLM_EMBEDDING_LENGTH      = "{arch}.embedding_length"
-KEY_LLM_BLOCK_COUNT           = "{arch}.block_count"
-KEY_LLM_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
-KEY_LLM_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
-KEY_LLM_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
+KEY_CONTEXT_LENGTH        = "{arch}.context_length"
+KEY_EMBEDDING_LENGTH      = "{arch}.embedding_length"
+KEY_BLOCK_COUNT           = "{arch}.block_count"
+KEY_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
+KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
+KEY_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
 
 # attention
 KEY_ATTENTION_HEAD_COUNT        = "{arch}.attention.head_count"
@@ -45,6 +47,7 @@ KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
 
 # RoPE
 KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count"
+KEY_ROPE_FREQ_BASE       = "{arch}.rope.freq_base"
 KEY_ROPE_SCALE_LINEAR    = "{arch}.rope.scale_linear"
 
 # tokenization
@@ -362,6 +365,9 @@ class GGUFValueType(IntEnum):
     BOOL    = 7
     STRING  = 8
     ARRAY   = 9
+    UINT64  = 10
+    INT64   = 11
+    FLOAT64 = 12
 
     @staticmethod
     def get_type(val):
@@ -375,6 +381,7 @@ class GGUFValueType(IntEnum):
             return GGUFValueType.BOOL
         elif isinstance(val, int):
             return GGUFValueType.INT32
+        # TODO: need help with 64-bit types in Python
         else:
             print("Unknown type: "+str(type(val)))
             sys.exit()
@@ -397,8 +404,8 @@ class GGUFWriter:
     def write_header_to_file(self):
         self.fout.write(struct.pack("<I", GGUF_MAGIC))
         self.fout.write(struct.pack("<I", GGUF_VERSION))
-        self.fout.write(struct.pack("<I", self.ti_data_count))
-        self.fout.write(struct.pack("<I", self.kv_data_count))
+        self.fout.write(struct.pack("<Q", self.ti_data_count))
+        self.fout.write(struct.pack("<Q", self.kv_data_count))
         self.flush()
 #        print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
 
@@ -441,6 +448,18 @@ class GGUFWriter:
         self.add_key(key)
         self.add_val(val, GGUFValueType.FLOAT32)
 
+    def add_uint64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT64)
+
+    def add_int64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT64)
+
+    def add_float64(self, key: str, val: float):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.FLOAT64)
+
     def add_bool(self, key: str, val: bool):
         self.add_key(key)
         self.add_val(val, GGUFValueType.BOOL)
@@ -480,17 +499,23 @@ class GGUFWriter:
             self.kv_data += struct.pack("<i", val)
         elif vtype == GGUFValueType.FLOAT32:
             self.kv_data += struct.pack("<f", val)
+        elif vtype == GGUFValueType.UINT64:
+            self.kv_data += struct.pack("<Q", val)
+        elif vtype == GGUFValueType.INT64:
+            self.kv_data += struct.pack("<q", val)
+        elif vtype == GGUFValueType.FLOAT64:
+            self.kv_data += struct.pack("<d", val)
         elif vtype == GGUFValueType.BOOL:
             self.kv_data += struct.pack("?", val)
         elif vtype == GGUFValueType.STRING:
             encoded_val = val.encode("utf8") if isinstance(val, str) else val
-            self.kv_data += struct.pack("<I", len(encoded_val))
+            self.kv_data += struct.pack("<Q", len(encoded_val))
             self.kv_data += encoded_val
         elif vtype == GGUFValueType.ARRAY:
             ltype = set([GGUFValueType.get_type(item) for item in val])
             assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
             self.kv_data += struct.pack("<I", list(ltype)[0])
-            self.kv_data += struct.pack("<I", len(val))
+            self.kv_data += struct.pack("<Q", len(val))
             for item in val:
                 self.add_val(item, add_vtype=False)
         else:
@@ -504,12 +529,12 @@ class GGUFWriter:
         assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
 
         encoded_name = name.encode("utf8")
-        self.ti_data += struct.pack("<I", len(encoded_name))
+        self.ti_data += struct.pack("<Q", len(encoded_name))
         self.ti_data += encoded_name
         n_dims = len(tensor_shape)
         self.ti_data += struct.pack("<I", n_dims)
         for i in range(n_dims):
-            self.ti_data += struct.pack("<I", tensor_shape[n_dims - 1 - i])
+            self.ti_data += struct.pack("<Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
             dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
         else:
@@ -581,7 +606,7 @@ class GGUFWriter:
         self.add_string(KEY_GENERAL_AUTHOR, author)
 
     def add_tensor_data_layout(self, layout: str):
-        self.add_string(KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+        self.add_string(KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
 
     def add_url(self, url: str):
         self.add_string(KEY_GENERAL_URL, url)
@@ -595,6 +620,9 @@ class GGUFWriter:
     def add_source_hf_repo(self, repo: str):
         self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
 
+    def add_file_type(self, ftype: int):
+        self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
+
     def add_name(self, name: str):
         self.add_string(KEY_GENERAL_NAME, name)
 
@@ -608,27 +636,27 @@ class GGUFWriter:
 
     def add_context_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_CONTEXT_LENGTH.format(arch=self.arch), length)
+            KEY_CONTEXT_LENGTH.format(arch=self.arch), length)
 
     def add_embedding_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_EMBEDDING_LENGTH.format(arch=self.arch), length)
+            KEY_EMBEDDING_LENGTH.format(arch=self.arch), length)
 
     def add_block_count(self, length: int):
         self.add_uint32(
-            KEY_LLM_BLOCK_COUNT.format(arch=self.arch), length)
+            KEY_BLOCK_COUNT.format(arch=self.arch), length)
 
     def add_feed_forward_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+            KEY_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
     def add_parallel_residual(self, use: bool):
         self.add_bool(
-            KEY_LLM_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
+            KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
 
     def add_tensor_data_layout(self, layout: str):
         self.add_string(
-            KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+            KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
 
     def add_head_count(self, count: int):
         self.add_uint32(
@@ -658,7 +686,10 @@ class GGUFWriter:
         self.add_uint32(
             KEY_ROPE_DIMENSION_COUNT.format(arch=self.arch), count)
 
-    def add_rope_scale_linear(self, value:  float):
+    def add_rope_freq_base(self, value: float):
+        self.add_float32(KEY_ROPE_FREQ_BASE.format(arch=self.arch), value)
+
+    def add_rope_scale_linear(self, value: float):
         self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value)
 
     def add_tokenizer_model(self, model: str):

gguf-py/pyproject.toml

@@ -0,0 +1,28 @@
+[tool.poetry]
+name = "gguf"
+version = "0.2.1"
+description = "Write ML models in GGUF for GGML"
+authors = ["GGML <ggml@ggml.ai>"]
+packages = [
+    {include = "gguf"},
+]
+readme = "README.md"
+homepage = "https://ggml.ai"
+repository = "https://github.com/ggerganov/llama.cpp"
+keywords = ["ggml", "gguf", "llama.cpp"]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+]
+
+[tool.poetry.dependencies]
+python = ">=3.8"
+numpy = ">=1.17"
+
+[tool.poetry.dev-dependencies]
+pytest = "^5.2"
+
+[build-system]
+requires = ["poetry-core>=1.0.0"]
+build-backend = "poetry.core.masonry.api"

gguf-py/tests/test_gguf.py

@@ -0,0 +1,7 @@
+import gguf
+
+# TODO: add tests
+
+
+def test_write_gguf():
+    pass

grammars/README.md

@@ -0,0 +1,91 @@
+# GBNF Guide
+
+GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `examples/main` and `examples/server`.
+
+## Background
+
+[Bakus-Naur Form (BNF)](https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form) is a notation for describing the syntax of formal languages like programming languages, file formats, and protocols. GBNF is an extension of BNF that primarily adds a few modern regex-like features.
+
+## Basics
+
+In GBNF, we define *production rules* that specify how a *non-terminal* (rule name) can be replaced with sequences of *terminals* (characters, specifically Unicode [code points](https://en.wikipedia.org/wiki/Code_point)) and other non-terminals. The basic format of a production rule is `nonterminal ::= sequence...`.
+
+## Example
+
+Before going deeper, let's look at some of the features demonstrated in `grammars/chess.gbnf`, a small chess notation grammar:
+```
+# `root` specifies the pattern for the overall output
+root ::= (
+    # it must start with the characters "1. " followed by a sequence
+    # of characters that match the `move` rule, followed by a space, followed
+    # by another move, and then a newline
+    "1. " move " " move "\n"
+
+    # it's followed by one or more subsequent moves, numbered with one or two digits
+    ([1-9] [0-9]? ". " move " " move "\n")+
+)
+
+# `move` is an abstract representation, which can be a pawn, nonpawn, or castle.
+# The `[+#]?` denotes the possibility of checking or mate signs after moves
+move ::= (pawn | nonpawn | castle) [+#]?
+
+pawn ::= ...
+nonpawn ::= ...
+castle ::= ...
+```
+
+## Non-Terminals and Terminals
+
+Non-terminal symbols (rule names) stand for a pattern of terminals and other non-terminals. They are required to be a dashed lowercase word, like `move`, `castle`, or `check-mate`.
+
+Terminals are actual characters ([code points](https://en.wikipedia.org/wiki/Code_point)). They can be specified as a sequence like `"1"` or `"O-O"` or as ranges like `[1-9]` or `[NBKQR]`.
+
+## Characters and character ranges
+
+Terminals support the full range of Unicode. Unicode characters can be specified directly in the grammar, for example `hiragana ::= [ぁ-ゟ]`, or with escapes: 8-bit (`\xXX`), 16-bit (`\uXXXX`) or 32-bit (`\UXXXXXXXX`).
+
+Character ranges can be negated with `^`:
+```
+single-line ::= [^\n]+ "\n"`
+```
+
+## Sequences and Alternatives
+
+The order of symbols in a sequence matter. For example, in `"1. " move " " move "\n"`, the `"1. "` must come before the first `move`, etc.
+
+Alternatives, denoted by `|`, give different sequences that are acceptable. For example, in `move ::= pawn | nonpawn | castle`, `move` can be a `pawn` move, a `nonpawn` move, or a `castle`.
+
+Parentheses `()` can be used to group sequences, which allows for embedding alternatives in a larger rule or applying repetition and optptional symbols (below) to a sequence.
+
+## Repetition and Optional Symbols
+
+- `*` after a symbol or sequence means that it can be repeated zero or more times.
+- `+` denotes that the symbol or sequence should appear one or more times.
+- `?` makes the preceding symbol or sequence optional.
+
+## Comments and newlines
+
+Comments can be specified with `#`:
+```
+# defines optional whitspace
+ws ::= [ \t\n]+
+```
+
+Newlines are allowed between rules and between symbols or sequences nested inside parentheses. Additionally, a newline after an alternate marker `|` will continue the current rule, even outside of parentheses.
+
+## The root rule
+
+In a full grammar, the `root` rule always defines the starting point of the grammar. In other words, it specifies what the entire output must match.
+
+```
+# a grammar for lists
+root ::= ("- " item)+
+item ::= [^\n]+ "\n"
+```
+
+## Next steps
+
+This guide provides a brief overview. Check out the GBNF files in this directory (`grammars/`) for examples of full grammars. You can try them out with:
+```
+./main -m <model> --grammar-file grammars/some-grammar.gbnf -p 'Some prompt'
+```

k_quants.c

@@ -77,6 +77,11 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         }
         return 1/iscale;
     }
+    bool return_early = false;
+    if (rmse_type < 0) {
+        rmse_type = -rmse_type;
+        return_early = true;
+    }
     int weight_type = rmse_type%2;
     float sumlx = 0;
     float suml2 = 0;
@@ -89,56 +94,9 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         suml2 += w*l*l;
     }
     float scale = sumlx/suml2;
+    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
     float best = scale * sumlx;
-    for (int itry = 0; itry < 3; ++itry) {
-        iscale = 1/scale;
-        float slx = 0;
-        float sl2 = 0;
-        bool changed = false;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            if (l + nmax != L[i]) { changed = true; }
-            float w = weight_type == 1 ? x[i] * x[i] : 1.f;
-            slx += w*x[i]*l;
-            sl2 += w*l*l;
-        }
-        if (!changed || sl2 == 0 || slx*slx <= best*sl2) { break; }
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-        }
-        sumlx = slx; suml2 = sl2;
-        scale = sumlx/suml2;
-        best = scale * sumlx;
-    }
-    for (int itry = 0; itry < 5; ++itry) {
-        int n_changed = 0;
-        for (int i = 0; i < n; ++i) {
-            float w = weight_type == 1 ? x[i]*x[i] : 1;
-            int l = L[i] - nmax;
-            float slx = sumlx - w*x[i]*l;
-            if (slx > 0) {
-                float sl2 = suml2 - w*l*l;
-                int new_l = nearest_int(x[i] * sl2 / slx);
-                new_l = MAX(-nmax, MIN(nmax-1, new_l));
-                if (new_l != l) {
-                    slx += w*x[i]*new_l;
-                    sl2 += w*new_l*new_l;
-                    if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
-                        L[i] = nmax + new_l; sumlx = slx; suml2 = sl2;
-                        scale = sumlx / suml2; best = scale * sumlx;
-                        ++n_changed;
-                    }
-                }
-            }
-        }
-        if (!n_changed) { break; }
-    }
-    if (rmse_type < 3) {
-        return scale;
-    }
-    for (int is = -4; is <= 4; ++is) {
+    for (int is = -9; is <= 9; ++is) {
         if (is == 0) {
             continue;
         }
@@ -221,12 +179,17 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
     return 1/iscale;
 }
 
-static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, int ntry) {
+static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+        int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
+    float sum_x = 0;
+    float sum_x2 = 0;
     for (int i = 1; i < n; ++i) {
         if (x[i] < min) min = x[i];
         if (x[i] > max) max = x[i];
+        sum_x += x[i];
+        sum_x2 += x[i]*x[i];
     }
     if (max == min) {
         for (int i = 0; i < n; ++i) L[i] = 0;
@@ -254,7 +217,7 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
         for (int i = 0; i < n; ++i) {
             sum += x[i] - scale*L[i];
         }
-        min = sum/n;
+        min = alpha*min + (1 - alpha)*sum/n;
         if (min > 0) min = 0;
         iscale = 1/scale;
         if (!did_change) break;
@@ -263,6 +226,82 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
     return scale;
 }
 
+static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights[0];
+    float sum_x = sum_w * x[0];
+    for (int i = 1; i < n; ++i) {
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) min = 0;
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff * diff;
+        float w = weights[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights[i];
+            sum_l += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff * diff;
+                float w = weights[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
 #if QK_K == 256
 static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
     if (j < 4) {
@@ -281,6 +320,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
     const int nb = k / QK_K;
 
     uint8_t L[QK_K];
+    uint8_t Laux[16];
+    float   weights[16];
     float mins[QK_K/16];
     float scales[QK_K/16];
 
@@ -291,7 +332,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/16; ++j) {
-            scales[j] = make_qkx1_quants(16, 3, x + 16*j, L + 16*j, &mins[j], 5);
+            for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
+            scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;
@@ -637,6 +679,8 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
     const int nb = k / QK_K;
 
     uint8_t L[QK_K];
+    uint8_t Laux[32];
+    float   weights[32];
     float mins[QK_K/32];
     float scales[QK_K/32];
 
@@ -645,7 +689,12 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
-            scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 5);
+            //scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;
@@ -798,6 +847,8 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
     uint8_t L[QK_K];
     float mins[QK_K/32];
     float scales[QK_K/32];
+    float weights[32];
+    uint8_t Laux[32];
 #else
     int8_t L[QK_K];
     float scales[QK_K/16];
@@ -810,7 +861,12 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
-            scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 5);
+            //scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;

llama.h

@@ -103,6 +103,8 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+
+        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
     typedef struct llama_token_data {
@@ -245,12 +247,18 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx);
+
     LLAMA_API int llama_model_n_vocab(const struct llama_model * model);
     LLAMA_API int llama_model_n_ctx  (const struct llama_model * model);
     LLAMA_API int llama_model_n_embd (const struct llama_model * model);
 
     // Get a string describing the model type
-    LLAMA_API int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size);
+    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+    // Returns the total size of all the tensors in the model in bytes
+    LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+    // Returns the total number of parameters in the model
+    LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
 
     // Returns 0 on success
     LLAMA_API int llama_model_quantize(
@@ -344,7 +352,7 @@ extern "C" {
 
     LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
 
-    LLAMA_API llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
@@ -366,13 +374,6 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
-    LLAMA_API int llama_tokenize_bpe(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
     LLAMA_API int llama_tokenize_with_model(
         const struct llama_model * model,
                       const char * text,
@@ -388,12 +389,6 @@ extern "C" {
                                   char * buf,
                                   int    length);
 
-    LLAMA_API int llama_token_to_str_bpe(
-            const struct llama_context * ctx,
-                           llama_token   token,
-                                  char * buf,
-                                  int    length);
-
     LLAMA_API int llama_token_to_str_with_model(
               const struct llama_model * model,
                            llama_token   token,
@@ -474,6 +469,43 @@ extern "C" {
     /// @details Accepts the sampled token into the grammar
     LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
 
+    //
+    // Beam search
+    //
+
+    struct llama_beam_view {
+        const llama_token * tokens;
+        size_t n_tokens;
+        float p;   // Cumulative beam probability (renormalized relative to all beams)
+        bool eob;  // Callback should set this to true when a beam is at end-of-beam.
+    };
+
+    // Passed to beam_search_callback function.
+    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
+    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
+    // These pointers are valid only during the synchronous callback, so should not be saved.
+    struct llama_beams_state {
+        struct llama_beam_view * beam_views;
+        size_t n_beams;               // Number of elements in beam_views[].
+        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
+        bool last_call;               // True iff this is the last callback invocation.
+    };
+
+    // Type of pointer to the beam_search_callback function.
+    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
+    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
+    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, llama_beams_state);
+
+    /// @details Deterministically returns entire sentence constructed by a beam search.
+    /// @param ctx Pointer to the llama_context.
+    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
+    /// @param callback_data A pointer that is simply passed back to callback.
+    /// @param n_beams Number of beams to use.
+    /// @param n_past Number of tokens already evaluated.
+    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
+    /// @param n_threads Number of threads as passed to llama_eval().
+    LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);

requirements.txt

@@ -1,2 +1,3 @@
 numpy==1.24
 sentencepiece==0.1.98
+gguf>=0.1.0

scripts/perf-run-all.sh

@@ -1,93 +0,0 @@
-#!/bin/bash
-#
-# Measure the performance (time per token) of the various quantization techniques
-#
-
-QUANTIZE=0
-if [ "$1" != "" ]; then
-    echo "Quantizing"
-    QUANTIZE=1
-fi
-
-if [ "$QUANTIZE" != "0" ]; then
-    #
-    # quantize
-    #
-
-    # 7B
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
-
-    # 13B
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
-fi
-
-#
-# perf
-# run each command twice
-#
-
-set -x
-
-# 7B - 4 threads
-     ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings
-
-# 7B - 8 threads
-     ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings
-
-# 13B - 4 threads
-     ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings
-
-# 13B - 8 threads
-     ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings

scripts/ppl-run-all.sh

@@ -1,39 +0,0 @@
-#!/bin/bash
-
-#
-# quantize
-#
-
-# 7B
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
-
-# 13B
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
-
-#
-# perplexity
-#
-
-# 7B
-time ./bin/perplexity -m ../models/7B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-f16.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_0.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_1.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_0.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_1.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q8_0.txt
-
-# 13B
-time ./bin/perplexity -m ../models/13B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-f16.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_0.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_1.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_0.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_1.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q8_0.txt

scripts/qnt-all.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+qnt=(q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args=""
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+model="$1"
+out="../tmp/results-${model}"
+
+mkdir -p ${out}
+
+for q in ${qnt[@]}; do
+    time ./bin/quantize ../models/${model}/ggml-model-f16.gguf ../models/${model}/ggml-model-${q}.gguf ${q} 2>&1 ${args} | tee ${out}/qnt-${q}.txt
+done

scripts/run-all-perf.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args="-ngl 999 -n 64 -p 512"
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+model="$1"
+out="../tmp/results-${model}"
+
+mkdir -p ${out}
+
+mstr=""
+
+for q in ${qnt[@]}; do
+    mstr="${mstr} -m ../models/${model}/ggml-model-${q}.gguf"
+done
+
+./bin/llama-bench ${mstr} ${args} 2> /dev/null

scripts/run-all-ppl.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args="-ngl 999 -t 8"
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+model="$1"
+out="../tmp/results-${model}"
+
+mkdir -p ${out}
+
+for q in ${qnt[@]}; do
+    time ./bin/perplexity -m ../models/${model}/ggml-model-f16.gguf -f ./wiki.test.raw ${args} 2>&1 | tee ${out}/ppl-${q}.txt
+done

scripts/sync-ggml.sh

@@ -1,14 +1,16 @@
 #!/bin/bash
 
-cp -rpv ../ggml/src/ggml.c           ./ggml.c
-cp -rpv ../ggml/src/ggml-cuda.h      ./ggml-cuda.h
-cp -rpv ../ggml/src/ggml-cuda.cu     ./ggml-cuda.cu
-cp -rpv ../ggml/src/ggml-opencl.h    ./ggml-opencl.h
-cp -rpv ../ggml/src/ggml-opencl.cpp  ./ggml-opencl.cpp
-cp -rpv ../ggml/src/ggml-metal.h     ./ggml-metal.h
-cp -rpv ../ggml/src/ggml-metal.m     ./ggml-metal.m
-cp -rpv ../ggml/src/ggml-metal.metal ./ggml-metal.metal
-cp -rpv ../ggml/include/ggml/ggml.h  ./ggml.h
+cp -rpv ../ggml/src/ggml.c                ./ggml.c
+cp -rpv ../ggml/src/ggml-alloc.c          ./ggml-alloc.c
+cp -rpv ../ggml/src/ggml-cuda.h           ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-cuda.cu          ./ggml-cuda.cu
+cp -rpv ../ggml/src/ggml-opencl.h         ./ggml-opencl.h
+cp -rpv ../ggml/src/ggml-opencl.cpp       ./ggml-opencl.cpp
+cp -rpv ../ggml/src/ggml-metal.h          ./ggml-metal.h
+cp -rpv ../ggml/src/ggml-metal.m          ./ggml-metal.m
+cp -rpv ../ggml/src/ggml-metal.metal      ./ggml-metal.metal
+cp -rpv ../ggml/include/ggml/ggml.h       ./ggml.h
+cp -rpv ../ggml/include/ggml/ggml-alloc.h ./ggml-alloc.h
 
 cp -rpv ../ggml/tests/test-opt.cpp    ./tests/test-opt.cpp
 cp -rpv ../ggml/tests/test-grad0.cpp  ./tests/test-grad0.cpp

tests/CMakeLists.txt

@@ -28,7 +28,8 @@ llama_build_and_test_executable(test-sampling.cpp)
 llama_build_executable(test-tokenizer-0.cpp)
 llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 llama_build_executable(test-tokenizer-1.cpp)
-llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+# test-tokenizer-1 requires a BPE vocab. re-enable when we have one.
+#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)

tests/test-tokenizer-0.cpp

@@ -17,6 +17,8 @@ static std::string unescape_whitespace(llama_context* ctx, const std::vector<lla
 static const std::map<std::string, std::vector<llama_token>> & k_tests() {
     static std::map<std::string, std::vector<llama_token>> _k_tests = {
         { " ",                      {1,    259, }, },
+        { "  ",                     { 1,    1678, }, },
+        { "   ",                    { 1,     268, }, },
         { "\t",                     { 1,    29871,   12, }, },
         { "\n",                     { 1,    29871,   13, }, },
         { "\t\n",                   { 1,    29871,   12,     13, }, },
@@ -38,6 +40,12 @@ static const std::map<std::string, std::vector<llama_token>> & k_tests() {
                 243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,
                 313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,
                 313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
+        { "Hello",                  { 1,    15043 }, },
+        { " Hello",                 { 1,    29871,  15043 }, },
+        { "  Hello",                { 1,    259,    15043 }, },
+        { "   Hello",               { 1,    1678,   15043 }, },
+        { "    Hello",              { 1,    268,    15043 }, },
+        { "    Hello\n    Hello",   { 1,    268,    15043,  13,     1678,   15043 }, },
     };
 
     return _k_tests;
@@ -92,7 +100,8 @@ int main(int argc, char **argv) {
     bool success = true;
 
     for (const auto & test_kv : k_tests()) {
-        std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, true);
+        // Add a space in front of the first character to match OG llama tokenizer behavior
+        std::vector<llama_token> res = llama_tokenize(ctx, " " + test_kv.first, true);
         fprintf(stderr, "%s : '%s' tokenized to '%s'\n",
             __func__, test_kv.first.c_str(), unescape_whitespace(ctx, res).c_str());
 
@@ -106,7 +115,8 @@ int main(int argc, char **argv) {
 
         if (!correct) {
             fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s'\n", __func__, unescape_whitespace(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str());
             fprintf(stderr, "%s : expected tokens: ", __func__);
             for (const auto & t : test_kv.second) {
                 fprintf(stderr, "%6d, ", t);

tests/test-tokenizer-1.cpp

@@ -11,18 +11,11 @@
 #include <locale>
 
 static std::string escape_whitespace(const std::string& text) {
-    std::string result;
-    bool escaping = false;
-    result += "\xe2\x96\x81";
+    std::string result = "\xe2\x96\x81";
     for (size_t offs = 0; offs < text.length(); ++offs) {
         if (text[offs] == ' ') {
-            if (!escaping) {
-                result += "\xe2\x96\x81";
-                escaping = true;
-            }
-        }
-        else {
-            escaping = false;
+            result += "\xe2\x96\x81";
+        } else {
             result += text[offs];
         }
     }
@@ -74,11 +67,13 @@ int main(int argc, char **argv) {
         }
     }
 
+    GGML_ASSERT(llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_BPE);
+
     const int n_vocab = llama_n_vocab(ctx);
 
     for (int i = 0; i < n_vocab; ++i) {
-        std::string forward = llama_token_to_str_bpe(ctx, i);
-        std::vector<llama_token> tokens = llama_tokenize_bpe(ctx, forward, false);
+        std::string forward = llama_token_to_str(ctx, i);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, forward, false);
         if (tokens.size() == 1) {
             if (i != tokens[0]) {
                 std::string backward = llama_token_to_str(ctx, tokens[0]);
@@ -86,16 +81,6 @@ int main(int argc, char **argv) {
                     __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str());
                 return 2;
             }
-        } else {
-            llama_token_type type = llama_token_get_type(ctx, i);
-            if (type == LLAMA_TOKEN_TYPE_UNKNOWN || type == LLAMA_TOKEN_TYPE_CONTROL || type == LLAMA_TOKEN_TYPE_BYTE) {
-                fprintf(stderr, "%s : info: token %d is string %s and bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-            } else {
-                fprintf(stderr, "%s : error: token %d is string %s but bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-                return 2;
-            }
         }
     }