removed unused semicolon

Temporary fix for stablecode and starcoder on mac

.dockerignore

@@ -1 +1,4 @@
 ggml/build
+build
+models
+assets

.github/workflows/build-commit.yml

@@ -25,18 +25,14 @@ jobs:
     - name: Dependencies
       id: depends
       run: |
-        brew update
-        brew install cmake boost
+        brew install cmake boost asio
     - name: Build
       id: make_build
       run: |
-        cd ${{github.workspace}}/ggml
-        cmake -B ${{github.workspace}}/ggml/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
-        cd ${{github.workspace}}/ggml/build 
-        make codegen codegen-serve codegen-quantize
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen 
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen-serve 
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen-quantize
+        cmake -B ${{github.workspace}}/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+        cd ${{github.workspace}}/build 
+        make 
+        chmod +x ${{github.workspace}}/build/bin/turbopilot
 
     - uses: benjlevesque/short-sha@v2.2
       id: short-sha
@@ -50,22 +46,22 @@ jobs:
         # Artifact name
         name: turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ steps.short-sha.outputs.sha }} # optional, default is artifact
         # A file, directory or wildcard pattern that describes what to upload
-        path: ${{github.workspace}}/ggml/build/bin/codegen*
+        path: ${{github.workspace}}/build/bin/turbopilot
         # The desired behavior if no files are found using the provided path.
         
     - name: package artifacts for release
       if: startsWith(github.ref, 'refs/tags/')
 
       run: |
-        cd ${{github.workspace}}/ggml/build/bin 
-        zip turbopilot-${{ runner.os }}-${{ runner.arch }}.zip ./codegen*
+        cd ${{github.workspace}}/build/bin
+        zip turbopilot-${{ runner.os }}-${{ runner.arch }}.zip ./turbopilot
 
     - name: Upload binaries to release
       uses: softprops/action-gh-release@v1
       if: startsWith(github.ref, 'refs/tags/')
       with:
         token: ${{ secrets.PUBLISH_TOKEN }}
-        files: ${{github.workspace}}/ggml/build/bin/turbopilot-${{ runner.os }}-${{ runner.arch }}.zip
+        files: ${{github.workspace}}/build/bin/turbopilot-${{ runner.os }}-${{ runner.arch }}.zip
 
 
 
@@ -95,7 +91,7 @@ jobs:
         submodules: true
 
     - name: Install Dependencies
-      run: sudo apt-get update && sudo apt-get install -yq libboost-dev
+      run: sudo apt-get update && sudo apt-get install -yq libboost-dev libasio-dev libboost-thread-dev
 
     - name: Install OpenBlas
       if: ${{ matrix.build == 'avx2-openblas' }}
@@ -112,13 +108,10 @@ jobs:
       # Configure CMake in a 'build' subdirectory. `CMAKE_BUILD_TYPE` is only required if you are using a single-configuration generator such as make.
       # See https://cmake.org/cmake/help/latest/variable/CMAKE_BUILD_TYPE.html?highlight=cmake_build_type
       run: |
-        cd ${{github.workspace}}/ggml
-        cmake -B ${{github.workspace}}/ggml/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} ${{ matrix.defines }}
-        cd ${{github.workspace}}/ggml/build 
-        make codegen codegen-serve codegen-quantize
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen 
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen-serve 
-        chmod +x ${{github.workspace}}/ggml/build/bin/codegen-quantize
+        cmake -B ${{github.workspace}}/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} ${{ matrix.defines }}
+        cd ${{github.workspace}}/build 
+        make
+        chmod +x ${{github.workspace}}/build/bin/turbopilot
         
         
     - uses: benjlevesque/short-sha@v2.2
@@ -133,15 +126,15 @@ jobs:
         # Artifact name
         name: turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}-${{ steps.short-sha.outputs.sha }} # optional, default is artifact
         # A file, directory or wildcard pattern that describes what to upload
-        path: ${{github.workspace}}/ggml/build/bin/codegen*
+        path: ${{github.workspace}}/build/bin/turbopilot
         # The desired behavior if no files are found using the provided path.
         
     - name: package artifacts for release
       if: startsWith(github.ref, 'refs/tags/')
 
       run: |
-        cd ${{github.workspace}}/ggml/build/bin 
-        zip turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip ./codegen*
+        cd ${{github.workspace}}/build/bin
+        zip turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip ./turbopilot
 
 
     - name: Upload binaries to release
@@ -149,7 +142,7 @@ jobs:
       if: startsWith(github.ref, 'refs/tags/')
       with:
         token: ${{ secrets.PUBLISH_TOKEN }}
-        files: ${{github.workspace}}/ggml/build/bin/turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip
+        files: ${{github.workspace}}/build/bin/turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip
 
   windows-latest-cmake:
     runs-on: windows-latest
@@ -213,16 +206,16 @@ jobs:
           $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
           $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
           & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+
       - name: Build
         id: cmake_build
         env:
           BOOST_ROOT: ${{ steps.install-boost.outputs.BOOST_ROOT }}
         run: |
-          cd ${{github.workspace}}/ggml
           mkdir build
           cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release --target codegen  codegen-serve codegen-quantize
+          cmake .. ${{ matrix.defines }} -DBoost_LIBRARYDIRS=${{ steps.install-boost.outputs.BOOST_ROOT }}/lib
+          cmake --build . --config Release --target turbopilot
 
       # - name: Add libopenblas.dll
       #   id: add_libopenblas_dll
@@ -244,11 +237,6 @@ jobs:
           & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
           .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
 
-          - uses: benjlevesque/short-sha@v2.2
-          id: short-sha
-          with:
-            length: 6
-
   
       - name: Upload Build Artifacts
         uses: actions/upload-artifact@v3.1.2
@@ -256,15 +244,14 @@ jobs:
           # Artifact name
           name: turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ steps.short-sha.outputs.sha }}-${{ matrix.build }} # optional, default is artifact
           # A file, directory or wildcard pattern that describes what to upload
-          path: ${{github.workspace}}\\ggml\\build\\bin\\Release\\codegen*
+          path: ${{github.workspace}}\\build\\bin\\Release\\turbopilot.exe
           # The desired behavior if no files are found using the provided path.
           
       - name: package artifacts for release
         if: startsWith(github.ref, 'refs/tags/')
   
         run: |
-          cd ${{github.workspace}}\ggml\build\bin\\Release\
-          7z a ${{github.workspace}}\ggml\build\bin\Release\turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip ./codegen*
+          7z a ${{github.workspace}}\build\bin\Release\turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip ${{github.workspace}}\\build\\bin\\Release\\turbopilot.exe
   
     
       - name: Convert backslashes
@@ -280,4 +267,4 @@ jobs:
         if: startsWith(github.ref, 'refs/tags/')
         with:
           token: ${{ secrets.PUBLISH_TOKEN }}
-          files: ${{ steps.convert_backslashes.outputs.converted_path }}/ggml/build/bin/Release/turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip
+          files: ${{ steps.convert_backslashes.outputs.converted_path }}/build/bin/Release/turbopilot-${{ runner.os }}-${{ runner.arch }}-${{ matrix.build }}.zip

.github/workflows/docker-image.yml

@@ -16,8 +16,52 @@ jobs:
     strategy:
       matrix:
         config:
-          - {tag: "", dockerfile: "./Dockerfile.default"}
-          - {tag: "-cuda", dockerfile: "./Dockerfile.cuda"}
+          - tag: 
+            dockerfile: ./Dockerfile.default
+            platforms: linux/amd64,linux/arm64
+            build_base: ubuntu:22.04
+            runtime_base: ubuntu:22.04
+
+
+          - tag: -openblas
+            dockerfile: ./Dockerfile.default
+            platforms: linux/amd64,linux/arm64 
+            build_base: ubuntu:22.04
+            runtime_base: ubuntu:22.04
+            extra_deps: libopenblas-dev
+            cmake_args: -DGGML_OPENBLAS=On
+
+            
+          - tag: -cuda11-7
+            dockerfile: ./Dockerfile.default
+            platforms: linux/amd64
+            build_base: nvidia/cuda:11.7.1-devel-ubuntu22.04
+            runtime_base: nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu22.04
+            cmake_args: -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
+            
+          - tag: -cuda12-0
+            dockerfile: ./Dockerfile.default
+            platforms: linux/amd64
+            build_base: nvidia/cuda:12.0.0-devel-ubuntu22.04
+            runtime_base: nvidia/cuda:12.0.0-runtime-ubuntu22.04
+            cmake_args: -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
+            
+          - tag: -cuda12-2
+            dockerfile: ./Dockerfile.default
+            platforms: linux/amd64
+            build_base: nvidia/cuda:12.2.0-devel-ubuntu22.04
+            runtime_base: nvidia/cuda:12.2.0-runtime-ubuntu22.04
+            cmake_args: -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
+          
+          # - tag: -clblast
+          #   dockerfile: ./Dockerfile.default
+          #   platforms: linux/amd64
+          #   build_base: ubuntu:22.04
+          #   runtime_base: ubuntu:22.04
+          #   runtime_deps: libclblast1
+          #   extra_deps: libclblast-dev
+          #   cmake_args: -DGGML_CLBLAST=On
+            
 
     steps:
     
@@ -44,17 +88,23 @@ jobs:
         password: ${{ secrets.GH_TOKEN }}
 
     - name: Build and push incremental
-      uses: docker/build-push-action@v4
+      uses: docker/build-push-action@v4.1.1
       if: (!startsWith(github.ref, 'refs/tags/'))
       with:
         file: ${{matrix.config.dockerfile}}
         push: true
         tags: ghcr.io/ravenscroftj/turbopilot:nightly${{matrix.config.tag}}-${{ github.sha }}
         context: ${{github.workspace}}
-        platforms: linux/amd64,linux/arm64
+        platforms: ${{matrix.config.platforms}}
+        build-args: |
+          EXTRA_DEPS=${{matrix.config.extra_deps}}
+          CMAKE_ARGS=${{matrix.config.cmake_args}}
+          BUILD_BASE=${{matrix.config.build_base}}
+          RUNTIME_BASE=${{matrix.config.runtime_base}}
+          RUNTIME_DEPS=${{matrix.config.runtime_deps}}
 
 
-    - name: Build and push release
+    - name: Build and push release (Main Latest Build)
       uses: docker/build-push-action@v4
       if: startsWith(github.ref, 'refs/tags/') && matrix.config.tag == ''
       with:
@@ -62,15 +112,27 @@ jobs:
         push: true
         tags: ghcr.io/ravenscroftj/turbopilot:${{ github.ref_name }}, ghcr.io/ravenscroftj/turbopilot:latest
         context: ${{github.workspace}}
-        platforms: linux/amd64,linux/arm64  
+        platforms: ${{matrix.config.platforms}} 
+        build-args: |
+          EXTRA_DEPS=${{matrix.config.extra_deps}}
+          CMAKE_ARGS=${{matrix.config.cmake_args}}
+          BUILD_BASE=${{matrix.config.build_base}}
+          RUNTIME_BASE=${{matrix.config.runtime_base}}
+          RUNTIME_DEPS=${{matrix.config.runtime_deps}}
 
 
-    - name: Build and push release (CUDA)
+    - name: Build and push release (Accelerated Builds)
       uses: docker/build-push-action@v4
       if: startsWith(github.ref, 'refs/tags/') && matrix.config.tag != ''
       with:
         file: ${{matrix.config.dockerfile}}
         push: true
-        tags: ghcr.io/ravenscroftj/turbopilot:${{ github.ref_name }}
+        tags: ghcr.io/ravenscroftj/turbopilot:${{ github.ref_name }}${{matrix.config.tag}}
         context: ${{github.workspace}}
-        platforms: linux/amd64,linux/arm64  
+        platforms: ${{matrix.config.platforms}}
+        build-args: |
+          EXTRA_DEPS=${{matrix.config.extra_deps}}
+          CMAKE_ARGS=${{matrix.config.cmake_args}}
+          BUILD_BASE=${{matrix.config.build_base}}
+          RUNTIME_BASE=${{matrix.config.runtime_base}}
+          RUNTIME_DEPS=${{matrix.config.runtime_deps}}

.gitignore

@@ -0,0 +1,2 @@
+build/
+models/

.gitmodules

@@ -1,3 +1,12 @@
 [submodule "ggml"]
-	path = ggml
-	url = git@github.com:ravenscroftj/ggml.git
+	path = extern/ggml
+	url = git@github.com:ggerganov/ggml.git
+[submodule "extern/argparse"]
+	path = extern/argparse
+	url = https://github.com/p-ranav/argparse.git
+[submodule "extern/sbdlog"]
+	path = extern/spdlog
+	url = https://github.com/gabime/spdlog.git
+[submodule "extern/ggml"]
+	path = extern/ggml
+	url = https://github.com/ggerganov/ggml

.vscode/c_cpp_properties.json

@@ -0,0 +1,20 @@
+{
+    "configurations": [
+        {
+            "name": "Linux",
+            "includePath": [
+                "${workspaceFolder}/**",
+                "${workspaceFolder}/extern/crow/include",
+                "${workspaceFolder}/include",
+                "${workspaceFolder}/include"
+            ],
+            "defines": [],
+            "compilerPath": "/usr/bin/gcc",
+            "cStandard": "c17",
+            "cppStandard": "gnu++17",
+            "intelliSenseMode": "linux-gcc-x64",
+            "configurationProvider": "ms-vscode.cmake-tools"
+        }
+    ],
+    "version": 4
+}

.vscode/launch.json

@@ -0,0 +1,70 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "(gdb) Launch TBP",
+            "type": "cppdbg",
+            "request": "launch",
+            "program": "/home/james/workspace/rafael-llm/turbopilot/build/bin/turbopilot",
+            "args": [
+                //TBP ARGS
+                "-v",
+                "-f",
+                "/home/james/Downloads/replit-code-v1-3b-q4_0.bin",
+                "-m",
+                "replit",
+            ],
+            "stopAtEntry": false,
+            "cwd": "${workspaceFolder}",
+            "environment": [],
+            "externalConsole": false,
+            "MIMode": "gdb",
+            "setupCommands": [
+                {
+                    "description": "Enable pretty-printing for gdb",
+                    "text": "-enable-pretty-printing",
+                    "ignoreFailures": true
+                },
+                {
+                    "description": "Set Disassembly Flavor to Intel",
+                    "text": "-gdb-set disassembly-flavor intel",
+                    "ignoreFailures": true
+                }
+            ]
+        },
+        {
+            "name": "(gdb) Launch Replut",
+            "type": "cppdbg",
+            "request": "launch",
+            "program": "/home/james/workspace/rafael-llm/turbopilot/extern/ggml/build/bin/replit",
+            "args": [
+                // REPLIT ARGS
+                "-m",
+                "/home/james/Downloads/replit-code-v1-3b-q4_0.bin",
+                "-f",
+                "/home/james/workspace/rafael-llm/turbopilot/test.txt"
+            ],
+            "stopAtEntry": false,
+            "cwd": "${workspaceFolder}",
+            "environment": [],
+            "externalConsole": false,
+            "MIMode": "gdb",
+            "setupCommands": [
+                {
+                    "description": "Enable pretty-printing for gdb",
+                    "text": "-enable-pretty-printing",
+                    "ignoreFailures": true
+                },
+                {
+                    "description": "Set Disassembly Flavor to Intel",
+                    "text": "-gdb-set disassembly-flavor intel",
+                    "ignoreFailures": true
+                }
+            ]
+        },
+
+    ]
+}

.vscode/tasks.json

@@ -0,0 +1,28 @@
+{
+    "tasks": [
+        {
+            "type": "cppbuild",
+            "label": "C/C++: g++ build active file",
+            "command": "/usr/bin/g++",
+            "args": [
+                "-fdiagnostics-color=always",
+                "-g",
+                "${file}",
+                "-o",
+                "${fileDirname}/${fileBasenameNoExtension}"
+            ],
+            "options": {
+                "cwd": "${fileDirname}"
+            },
+            "problemMatcher": [
+                "$gcc"
+            ],
+            "group": {
+                "kind": "build",
+                "isDefault": true
+            },
+            "detail": "Task generated by Debugger."
+        }
+    ],
+    "version": "2.0.0"
+}

BUILD.md

@@ -32,25 +32,19 @@ Make sure the ggml subproject is checked out with `git submodule init` and `git
 Configure cmake to build the project with the following:
 
 ```bash
-mkdir ggml/build
-cd ggml/build
+mkdir build
+cd build
 cmake ..
 ```
 
 If you are running on linux you can optionally compile a static build with `cmake -D CMAKE_EXE_LINKER_FLAGS="-static" ..` which should make your binary portable across different flavours of the OS.
 
-From here you can now build the components that make up turbopilot:
+From here you can now build the components that make up turbopilot by running:
 
 ```bash
-make codegen codegen-quantize codegen-serve
+make
 ```
 
-Where:
-
-- *codegen* is a command line tool for testing out prompts in a lightweight way (a lot like llama.cpp)
-- *codegen-serve* is the actual REST server that can be used to connect to VSCode
-- *codegen-quantize* is the tool for quantizing models exported by the conversion script. For more details see [Converting and Quantizing The Models](https://github.com/ravenscroftj/turbopilot/wiki/Converting-and-Quantizing-The-Models).
-
 ### Building with OpenBLAS
 
 

CMakeLists.txt

@@ -0,0 +1,73 @@
+cmake_minimum_required (VERSION 3.0)
+project(turbopilot VERSION 0.1.0)
+set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+find_package(Threads REQUIRED)
+
+
+# option(BUILD_SHARED_LIBS "Build using shared libraries" OFF)
+
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")
+
+option(GGML_CLBLAST                 "ggml: use clBLAST"                  OFF)
+option(GGML_CUBLAS                  "ggml: use cuBLAS"                   OFF)
+
+
+
+if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+    message(STATUS "ARM detected")
+    if (MSVC)
+        # TODO: arm msvc?
+        message(STATUS "ARM+MSVC= :( ")
+    else()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            # Raspberry Pi 2
+            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            add_compile_options(-mfp16-format=ieee -mno-unaligned-access)
+        endif()
+    endif()
+endif()
+
+
+
+
+
+add_subdirectory(extern/ggml)
+add_subdirectory(extern/argparse)
+add_subdirectory(extern/spdlog)
+
+if (GGML_STATIC)
+    SET(CMAKE_FIND_LIBRARY_SUFFIXES ".a")
+    SET(BUILD_SHARED_LIBS OFF)
+    SET(CMAKE_EXE_LINKER_FLAGS "-static")
+endif()
+
+if (GGML_CUBLAS)
+    cmake_minimum_required(VERSION 3.17)
+
+    find_package(CUDAToolkit)
+    if (CUDAToolkit_FOUND)
+        add_compile_definitions(GGML_USE_CUBLAS)
+    else()
+        message(WARNING "cuBLAS not found")
+    endif()
+endif()
+
+
+
+add_subdirectory(src)
+
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)

Dockerfile.cuda

@@ -1,31 +0,0 @@
-FROM nvidia/cuda:11.7.1-devel-ubuntu22.04 AS build
-
-RUN apt-get update && apt-get install -y build-essential cmake libboost-dev
-
-ADD ./ggml /build/ggml
-
-RUN mkdir /build/ggml/build
-
-WORKDIR  /build/ggml/build
-
-RUN cmake -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc ..
-RUN make codegen-serve
-
-FROM nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu22.04 AS runtime
-
-
-WORKDIR /app
-
-COPY --from=build /build/ggml/build/bin/codegen-serve /app/codegen-serve
-
-ENV THREADS=4
-
-ENV MODEL="/models/codegen-2B-multi-ggml-4bit-quant.bin"
-
-ENV BATCHSIZE=64
-
-COPY ./run.sh /app/
-
-EXPOSE 18080
-
-CMD /app/run.sh

Dockerfile.default

@@ -1,22 +1,42 @@
-FROM alpine AS build
+ARG BUILD_BASE="ubuntu:22.04"
+ARG RUNTIME_BASE="ubuntu:22.04"
 
-RUN apk add --update alpine-sdk boost-dev cmake
+FROM ${BUILD_BASE} AS build
 
-ADD ./ggml /build/ggml
+ARG EXTRA_DEPS=""
+ARG CMAKE_ARGS=""
 
-RUN mkdir /build/ggml/build
+RUN echo "CMAKE_ARGS: ${CMAKE_ARGS}"
+RUN echo "EXTRA_DEPS: ${EXTRA_DEPS}"
 
-WORKDIR  /build/ggml/build
+ENV DEBIAN_FRONTEND=noninteractive
 
-RUN cmake -D GGML_STATIC=ON ..
-RUN make codegen-serve
+# inlude kitware apt repo to allow us to grab latest cmake
+RUN apt-get update && apt-get install -y ca-certificates gpg wget
+RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+RUN echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' | tee /etc/apt/sources.list.d/kitware.list >/dev/null
 
-FROM alpine AS runtime
+RUN apt-get update && apt-get install -y build-essential cmake libboost-dev libboost-thread-dev ${EXTRA_DEPS}
+
+ADD ./ /turbopilot
+
+RUN mkdir /turbopilot/build
+
+WORKDIR  /turbopilot/build
+
+RUN cmake .. ${CMAKE_ARGS}
+RUN make turbopilot
+
+FROM ${RUNTIME_BASE} AS runtime
+
+ARG RUNTIME_DEPS=""
+
+RUN if [[ -z "${RUNTIME_DEPS}" ]] ; then echo "No runtime libs required" ; else  apt-get update && apt-get install -y ${RUNTIME_DEPS} ; fi
 
 
 WORKDIR /app
 
-COPY --from=build /build/ggml/build/bin/codegen-serve /app/codegen-serve
+COPY --from=build /turbopilot/build/bin/turbopilot /app/turbopilot
 
 ENV THREADS=4
 

MODELS.md

@@ -0,0 +1,80 @@
+# Models Directory
+
+## StableCode Instruct State-of-the-art for low Spec machines(Released 8th August 2023)
+
+[StableCode](https://stability.ai/blog/stablecode-llm-generative-ai-coding) Instruct is a new model from [Stability.ai](https://stability.ai/) which provides reasonable autocomplete suggestions in approx 3GiB of RAM.
+
+| Model Name          | RAM Requirement | Direct Download  | HF Project Link |
+|---------------------|-----------------|-----------------|-----------------|
+| StableCode   | ~3GiB        |   [:arrow_down:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/resolve/main/stablecode-instruct-alpha-3b.ggmlv1.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/)           |
+
+To run in Turbopilot set model type `-m stablecode`
+
+## "Coder" family models
+
+WizardCoder, StarCoder and SantaCoder are current "state-of-the-art" autocomplete models 
+
+### SantaCoder (Small Model, Reasonable on lower spec machines - Released 13/4/2023)
+
+[SantaCoder](https://huggingface.co/bigcode/santacoder) is a smaller version of the StarCoder and WizardCoder family with only 1.1 Billion parameters. The model is trained with fill-in-the-middle objective allowing it to be used to auto-complete function parameters.
+
+This model is primarily trained on Python, Java and Javscript.
+
+
+| Model Name          | RAM Requirement | Direct Download  | HF Project Link |
+|---------------------|-----------------|-----------------|-----------------|
+| SantaCoder   | ~2GiB        |   [:arrow_down:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/resolve/main/santacoder-q4_0.bin)           |   [:hugs:](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/)           |
+
+To run in Turbopilot set model type `-m starcoder`
+
+
+### WizardCoder 15B Best Autocomplete Performance, Compute-Hungry (Released 15/6/2023)
+
+[WizardCoder](https://github.com/nlpxucan/WizardLM/tree/main/WizardCoder) is the current SOTA auto complete model, it is an updated version of StarCoder that achieves 57.1 pass@1 on HumanEval benchmarks (essentially in 57% of cases it correctly solves a given challenge. Read more about how this metric works in the scientific paper [here](https://arxiv.org/pdf/2107.03374.pdf) ).
+
+Even when quantized, WizardCoder is a large model that takes up a significant amount of RAM.
+
+
+| Model Name          | RAM Requirement | Direct Download  | HF Project Link |
+|---------------------|-----------------|-----------------|-----------------|
+| WizardCoder   | ~12GiB        |   [:arrow_down:](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/resolve/main/WizardCoder-15B-1.0.ggmlv3.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/)           |
+
+To run in Turbopilot set model type `-m wizardcoder`
+
+
+### StarCoder (Released 4/5/2023)
+
+[StarCoder](https://huggingface.co/blog/starcoder) held the previous title of state-of-the-art coding model back in May 2023. It is still a reasonably good model by comparison but it is a similar size and has similar RAM and compute requirements to WizardCoder so you may be better off just running that. Links below provided for posterity.
+
+
+| Model Name          | RAM Requirement | Direct Download  | HF Project Link |
+|---------------------|-----------------|------------------|-----------------|
+| StarCoder   | ~12GiB        | [:arrow_down:](https://huggingface.co/NeoDim/starcoder-GGML/resolve/main/starcoder-ggml-q4_0.bin)           |   [:hugs:](https://huggingface.co/NeoDim/starcoder-GGML/)           |
+| StarCoder Plus   | ~12GiB        | [:arrow_down:](https://huggingface.co/TheBloke/starcoderplus-GGML/resolve/main/starcoderplus.ggmlv3.q4_0.bin)           |   [:hugs:](https://huggingface.co/TheBloke/starcoderplus-GGML/)           |
+
+To run in Turbopilot set model type `-m starcoder`
+
+
+
+
+
+
+## CodeGen 1.0
+
+The CodeGen models were the first models supported by Turbopilot. They perform less well than the newer Wizardcoder/Starcoder/Santacoder variant models.
+
+
+The `multi` flavour models can provide auto-complete suggestions for `C`, `C++`, `Go`, `Java`, `JavaScript`, and `Python`.
+
+The `mono` flavour models can provide auto-complete suggestions for `Python` only (but the quality of Python-specific suggestions may be higher).
+
+Pre-converted and pre-quantized models are available for download from here:
+
+| Model Name          | RAM Requirement | Supported Languages       | Direct Download  | HF Project Link |
+|---------------------|-----------------|---------------------------|-----------------|-----------------|
+| CodeGen 350M multi   | ~800MiB        | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`  |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-350M-multi-ggml-quant/resolve/main/codegen-350M-multi-ggml-4bit-quant.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-350M-multi-ggml-quant)           |
+| CodeGen 350M mono   | ~800MiB   | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-350M-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-350M-mono-ggml-quant)           |
+| CodeGen 2B multi   | ~4GiB  | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`          |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-2B-multi-ggml-quant/resolve/main/codegen-2B-multi-ggml-4bit-quant_q4_0.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-2B-multi-ggml-quant)          |
+| CodeGen 2B mono   | ~4GiB  | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-2B-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-2B-mono-ggml-quant/)          |
+| CodeGen 6B multi   | ~8GiB  | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`          |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-6B-multi-ggml-quant/resolve/main/codegen-6B-multi-ggml-4bit-quant.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-6B-multi-ggml-quant)          |
+| CodeGen 6B mono   | ~8GiB  | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-6B-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-6B-mono-ggml-quant/)          |

README.md

@@ -9,8 +9,11 @@ TurboPilot is a self-hosted [copilot](https://github.com/features/copilot) clone
 
 ![a screen recording of turbopilot running through fauxpilot plugin](assets/vscode-status.gif)
 
+**✨ Now Supports [StableCode 3B Instruct](https://huggingface.co/stabilityai/stablecode-instruct-alpha-3b)** simply use [TheBloke's Quantized GGML models](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML) and set `-m stablecode`.
 
-**NEW:** As of v0.0.5 turbopilot supports cuda inference which greatly accelerates suggestions when working with longer prompts (i.e. longer existing code files).
+**✨ New: Refactored + Simplified**: The source code has been improved to make it easier to extend and add new models to Turbopilot. The system now supports multiple flavours of model
+
+**✨ New: Wizardcoder, Starcoder, Santacoder support** - Turbopilot now supports state of the art local code completion models which provide more programming languages and "fill in the middle" support.
 
 ## 🤝 Contributing
 
@@ -23,6 +26,7 @@ Make a fork, make your changes and then open a [PR](https://github.com/ravenscro
 
 The easiest way to try the project out is to grab the pre-processed models and then run the server in docker.
 
+
 ### Getting The Models
 
 You have 2 options for getting the model
@@ -31,20 +35,11 @@ You have 2 options for getting the model
 
 You can download the pre-converted, pre-quantized models from Huggingface.
 
-The `multi` flavour models can provide auto-complete suggestions for `C`, `C++`, `Go`, `Java`, `JavaScript`, and `Python`.
+For low RAM users (4-8 GiB), I recommend [StableCode](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML) and for high power users (16+ GiB RAM, discrete GPU or apple silicon) I recomnmend [WizardCoder](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/resolve/main/WizardCoder-15B-1.0.ggmlv3.q4_0.bin).
 
-The `mono` flavour models can provide auto-complete suggestions for `Python` only (but the quality of Python-specific suggestions may be higher).
+Turbopilot still supports the first generation codegen models from `v0.0.5` and earlier builds. Although old models do need to be requantized.
 
-Pre-converted and pre-quantized models are available for download from here:
-
-| Model Name          | RAM Requirement | Supported Languages       | Direct Download  | HF Project Link |
-|---------------------|-----------------|---------------------------|-----------------|-----------------|
-| CodeGen 350M multi   | ~800MiB        | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`  |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-350M-multi-ggml-quant/resolve/main/codegen-350M-multi-ggml-4bit-quant.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-350M-multi-ggml-quant)           |
-| CodeGen 350M mono   | ~800MiB   | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-350M-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-350M-mono-ggml-quant)           |
-| CodeGen 2B multi   | ~4GiB  | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`          |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-2B-multi-ggml-quant/resolve/main/codegen-2B-multi-ggml-4bit-quant.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-2B-multi-ggml-quant)          |
-| CodeGen 2B mono   | ~4GiB  | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-2B-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-2B-mono-ggml-quant/)          |
-| CodeGen 6B multi   | ~8GiB  | `C`, `C++`, `Go`, `Java`, `JavaScript`, `Python`          |   [:arrow_down:](https://huggingface.co/ravenscroftj/CodeGen-6B-multi-ggml-quant/resolve/main/codegen-6B-multi-ggml-4bit-quant.bin)           |   [:hugs:](https://huggingface.co/ravenscroftj/CodeGen-6B-multi-ggml-quant)          |
-| CodeGen 6B mono   | ~8GiB  | `Python`          |   [:arrow_down:](https://huggingface.co/Guglielmo/CodeGen-6B-mono-ggml-quant/resolve/main/ggml-model-quant.bin)           |   [:hugs:](https://huggingface.co/Guglielmo/CodeGen-6B-mono-ggml-quant/)          |
+You can find a full catalogue of models in [MODELS.md](MODELS.md).
 
 
 #### Option B: Convert The Models Yourself - Hard, More Flexible
@@ -58,17 +53,21 @@ Download the [latest binary](https://github.com/ravenscroftj/turbopilot/releases
 Run:
 
 ```bash
-./codegen-serve -m ./models/codegen-6B-multi-ggml-4bit-quant.bin
+./turbopilot -m starcoder -f ./models/santacoder-q4_0.bin
 ```
 
-The application should start a server on port `18080`
+The application should start a server on port `18080`, you can change this with the `-p` option but this is the default port that vscode-fauxpilot tries to connect to so you probably want to leave this alone unless you are sure you know what you're doing.
 
 If you have a multi-core system you can control how many CPUs are used with the `-t` option - for example, on my AMD Ryzen 5000 which has 6 cores/12 threads I use:
 
 ```bash
-./codegen-serve -t 6 -m ./models/codegen-6B-multi-ggml-4bit-quant.bin
+./codegen-serve -t 6 -m starcoder -f ./models/santacoder-q4_0.bin
 ```
 
+To run the legacy codegen models. Just change the model type flag `-m` to `codegen` instead.
+
+**NOTE: Turbopilot 0.1.0 and newer re-quantize your codegen models old models from v0.0.5 and older. I am working on providing updated quantized codegen models**
+
 ### 📦 Running From Docker
 
 You can also run Turbopilot from the pre-built docker image supplied [here](https://github.com/users/ravenscroftj/packages/container/package/turbopilot)
@@ -79,7 +78,8 @@ You will still need to download the models separately, then you can run:
 docker run --rm -it \
   -v ./models:/models \
   -e THREADS=6 \
-  -e MODEL="/models/codegen-2B-multi-ggml-4bit-quant.bin" \
+  -e MODEL_TYPE=starcoder \
+  -e MODEL="/models/santacoder-q4_0.bin" \
   -p 18080:18080 \
   ghcr.io/ravenscroftj/turbopilot:latest
 ```
@@ -92,18 +92,26 @@ As of release v0.0.5 turbocode now supports CUDA inference. In order to run the
 docker run --gpus=all --rm -it \
   -v ./models:/models \
   -e THREADS=6 \
-  -e MODEL="/models/codegen-2B-multi-ggml-4bit-quant.bin" \
+  -e MODEL_TYPE=starcoder \
+  -e MODEL="/models/santacoder-q4_0.bin" \
+  -e GPU_LAYERS=32 \
   -p 18080:18080 \
-  ghcr.io/ravenscroftj/turbopilot:v0.0.5-cuda
+  ghcr.io/ravenscroftj/turbopilot:v0.2.0-cuda11-7
 ```
 
-You will need CUDA 11 or later to run this container. You should be able to see `/app/codegen-serve` listed when you run `nvidia-smi`.
+If you have a big enough GPU then setting `GPU_LAYERS` will allow turbopilot to fully offload computation onto your GPU rather than copying data backwards and forwards, dramatically speeding up inference. 
+
+Swap `ghcr.io/ravenscroftj/turbopilot:v0.1.0-cuda11` for `ghcr.io/ravenscroftj/turbopilot:v0.2.0-cuda12-0` or `ghcr.io/ravenscroftj/turbopilot:v0.2.0-cuda12-2` if you are using CUDA 12.0 or 12.2 respectively.
+
+You will need CUDA 11 or CUDA 12 later to run this container. You should be able to see `/app/turbopilot` listed when you run `nvidia-smi`.
 
 
 #### Executable and CUDA
 
 As of v0.0.5 a CUDA version of the linux executable is available - it requires that libcublas 11 be installed on the machine - I might build ubuntu debs at some point but for now running in docker may be more convenient if you want to use a CUDA GPU.
 
+You can use GPU offloading via the `--ngl` option.
+
 ### 🌐 Using the API
 
 #### Support for the official Copilot Plugin
@@ -174,12 +182,7 @@ Should get you something like this:
 
 ## 👉 Known Limitations
 
-Again I want to set expectations around this being a proof-of-concept project. With that in mind. Here are some current known limitations.
-
-As of **v0.0.2**:
-- The models can be quite slow - especially the 6B ones. It can take ~30-40s to make suggestions across 4 CPU cores.
-- I've only tested the system on Ubuntu 22.04 but I am now supplying ARM docker images and soon I'll be providing ARM binary releases.
-- Sometimes suggestions get truncated in nonsensical places - e.g. part way through a variable name or string name. This is due to a hard limit of 2048 on the context length (prompt + suggestion).
+- Currently Turbopilot only supports one GPU device at a time (it will not try to make use of multiple devices).
 
 ## 👏 Acknowledgements
 

convert-codegen-to-ggml.py

@@ -171,10 +171,10 @@ for name in list_vars.keys():
         print("  Skipping variable: " + name)
         continue
 
-    n_dims = len(data.shape);
+    n_dims = len(data.shape)
 
     # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype_cur = 0;
+    ftype_cur = 0
     if ftype != 0:
         if name[-7:] == ".weight" and n_dims == 2:
             print("  Converting to float16")
@@ -211,7 +211,7 @@ for name in list_vars.keys():
     fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
     for i in range(n_dims):
         fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str);
+    fout.write(str)
 
     # data
     data.tofile(fout)

extern/argparse

@@ -0,0 +1 @@
+Subproject commit b0930ab0288185815d6dc67af59de7014a6272f7

extern/ggml

@@ -0,0 +1 @@
+Subproject commit 1a5d5f331de1d3c7ace40d86fe2373021a42f9ce

extern/spdlog

@@ -0,0 +1 @@
+Subproject commit 2ee8bac78e6525a8ad9a9196e65d502ce390d83a

ggml

@@ -1 +0,0 @@
-Subproject commit 6c4fe0ef5e50b76dd2539130c109e12179da0bd2

include/turbopilot/gptj.hpp

@@ -0,0 +1,83 @@
+#ifndef __TURBOPILOT_GPTJ_H
+#define __TURBOPILOT_GPTJ_H
+
+#include <turbopilot/model.hpp>
+
+#include <vector>
+#include <map>
+
+// default hparams (GPT-J 6B)
+struct gptj_hparams {
+    int32_t n_vocab = 50400;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 16;
+    int32_t n_layer = 28;
+    int32_t n_rot   = 64;
+    int32_t ftype   = 1;
+};
+
+struct gptj_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_q_proj_w;
+    struct ggml_tensor * c_attn_k_proj_w;
+    struct ggml_tensor * c_attn_v_proj_w;
+
+    struct ggml_tensor * c_attn_proj_w;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gptj_model {
+    gptj_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<gptj_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+
+
+class GPTJModel : public TurbopilotModel {
+
+public:
+    GPTJModel(ModelConfig config, std::mt19937 &rng) : TurbopilotModel(config, rng){
+        this->model = new gptj_model{};
+        this->vocab = new gpt_vocab{};
+    }
+    virtual ~GPTJModel();
+    bool load_model(std::string path);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
+
+private:
+    gptj_model *model = NULL;
+    gpt_vocab *vocab = NULL;
+
+
+};
+
+#endif

include/turbopilot/gptneox.hpp

@@ -0,0 +1,87 @@
+#ifndef __TURBOPILOT_GPTNEOX_H
+#define __TURBOPILOT_GPTNEOX_H
+
+#include <turbopilot/model.hpp>
+
+#include <vector>
+#include <map>
+
+// default hparams (StableLM 3B)
+struct gpt_neox_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 4096;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 32;
+    int32_t n_layer = 16;
+    int32_t n_rot   = 32; // rotary_pct * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = 1;
+};
+
+struct gpt_neox_layer {
+    // pre normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // post normalization
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gpt_neox_model {
+    gpt_neox_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    //struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<gpt_neox_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+
+class GPTNEOXModel : public TurbopilotModel {
+
+public:
+    GPTNEOXModel(ModelConfig config, std::mt19937 &rng) : TurbopilotModel(config, rng){
+        this->model = new gpt_neox_model{};
+        this->vocab = new gpt_vocab{};
+    }
+    virtual ~GPTNEOXModel();
+    bool load_model(std::string path);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
+
+private:
+    gpt_neox_model *model = NULL;
+    gpt_vocab *vocab = NULL;
+
+
+};
+
+#endif // __TURBOPILOT_GPTNEOX_H

include/turbopilot/model.hpp

@@ -0,0 +1,72 @@
+#ifndef __TURBOPILOT_MODEL_H
+#define __TURBOPILOT_MODEL_H
+
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <map>
+#include <vector>
+#include <random>
+#include <mutex>
+
+typedef void (*offload_func_t)(struct ggml_tensor * tensor);
+void ggml_nop(struct ggml_tensor * tensor);
+
+struct gpt_vocab
+{
+    using id = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+    std::vector<std::string> special_tokens;
+
+    void add_special_token(const std::string &token);
+};
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab &vocab, const std::string &text);
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+    const gpt_vocab &vocab,
+    const float *logits,
+    int top_k,
+    double top_p,
+    double temp,
+    std::mt19937 &rng);
+
+struct ModelConfig
+{
+    int n_threads = 4;
+    int32_t top_k = 40;
+    float top_p = 0.95f;
+    float temp = 0.80f;
+    float repeat_penalty = 1.10f;
+    int32_t seed = -1;     // RNG seed
+    int32_t n_ctx = 512;   // context size
+    int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_gpu_layers = 0;
+};
+
+class TurbopilotModel
+{
+
+public:
+    TurbopilotModel(ModelConfig config, std::mt19937 &rng) : 
+        config(config), 
+        rng(rng) 
+    {}
+    virtual bool load_model(std::string model_path) = 0;
+    std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
+    void lock();
+    void unlock();
+
+
+protected:
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt) = 0;
+    ModelConfig config;
+    std::mt19937 &rng;
+    std::mutex model_lock;
+};
+
+
+#endif //__TURBOPILOT_MODEL_H

include/turbopilot/server.hpp

@@ -0,0 +1,57 @@
+#ifndef __TURBOPILOT_SERVER_H
+#define __TURBOPILOT_SERVER_H
+
+
+#include <spdlog/spdlog.h>
+
+#include "turbopilot/model.hpp"
+
+#include "crow_all.h"
+
+crow::response handle_openai_request(TurbopilotModel *model, const crow::request& req);
+
+crow::response handle_hf_request(TurbopilotModel *model, const crow::request& req);
+
+class TBPLogger : public crow::ILogHandler {
+ public:
+  TBPLogger() {}
+  void log(std::string message, crow::LogLevel crow_level) {
+    // "message" doesn't contain the timestamp and loglevel
+    // prefix the default logger does and it doesn't end
+    // in a newline.
+
+    spdlog::level::level_enum level = spdlog::level::info;
+
+    switch(crow_level){
+        case crow::LogLevel::Critical:
+            level = spdlog::level::critical;
+            break;
+
+        case crow::LogLevel::Error:
+            level = spdlog::level::err;
+            break;
+
+        case crow::LogLevel::Warning:
+            level = spdlog::level::warn;
+            break;
+
+        case crow::LogLevel::Info:
+            level = spdlog::level::info;
+            break;
+
+        case crow::LogLevel::Debug:
+            level = spdlog::level::debug;
+            break;
+
+        default:
+            // if case is not a known value, assume the worst
+            level = spdlog::level::critical;
+    }
+
+    spdlog::log(level, message);
+  }
+};
+
+
+#endif // __TURBOPILOT_SERVER_H
+

include/turbopilot/starcoder.hpp

@@ -0,0 +1,79 @@
+#ifndef __TURBOPILOT_STARCODER_H
+#define __TURBOPILOT_STARCODER_H
+
+#include <turbopilot/model.hpp>
+
+// default hparams (GPT-2 117M)
+// https://huggingface.co/bigcode/gpt_bigcode-santacoder/blob/main/config.json
+struct starcoder_hparams {
+    int32_t n_vocab = 49280;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 2048;
+    int32_t n_head  = 16;
+    int32_t n_layer = 24;
+    int32_t ftype   = 1;
+};
+
+struct starcoder_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // mlp
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct starcoder_model {
+    starcoder_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
+
+    std::vector<starcoder_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+
+class StarcoderModel : public TurbopilotModel {
+public:
+    StarcoderModel(ModelConfig config, std::mt19937 &rng) : TurbopilotModel(config, rng){
+        this->model = new starcoder_model{};
+        this->vocab = new gpt_vocab{};
+    }
+    virtual ~StarcoderModel();
+    bool load_model(std::string path);
+    virtual std::stringstream predict_impl(std::string prompt, int max_length, bool include_prompt);
+
+private:
+    starcoder_model *model = NULL;
+    gpt_vocab *vocab = NULL;
+};
+
+
+#endif //__TURBOPILOT_STARCODER_H

run.sh

@@ -1,3 +1,6 @@
 #!/bin/sh
-
-/app/codegen-serve -t $THREADS -m $MODEL -b $BATCHSIZE
+if [ -z "$GPU_LAYERS" ]; then 
+    /app/turbopilot -t $THREADS -m $MODEL_TYPE -f $MODEL 
+else
+    /app/turbopilot -t $THREADS -m $MODEL_TYPE -f $MODEL --ngl $GPU_LAYERS
+fi

src/CMakeLists.txt

@@ -0,0 +1,32 @@
+set(TURBOPILOT_TARGET turbopilot)
+
+find_package(Boost COMPONENTS thread system REQUIRED)
+
+include_directories(${Boost_INCLUDE_DIRS})
+
+add_executable(${TURBOPILOT_TARGET} 
+  main.cpp 
+  gptj.cpp
+  gptneox.cpp
+  common.cpp
+  server.cpp
+  starcoder.cpp
+  ../include/turbopilot/model.hpp
+  ../include/turbopilot/gptj.hpp
+  ../include/turbopilot/gptneox.hpp
+  ../include/turbopilot/starcoder.hpp
+  )
+
+#set(THREADS_PREFER_PTHREAD_FLAG TRUE)
+#find_package(Threads REQUIRED)
+
+
+target_include_directories(${TURBOPILOT_TARGET} PRIVATE
+  ../include
+  ../extern/spdlog/include
+  ../extern/crow/include
+)
+
+#target_compile_features(${TURBOPILOT_TARGET} PRIVATE cxx_std_11)
+
+target_link_libraries(${TURBOPILOT_TARGET} PRIVATE ggml argparse)

src/common.cpp

@@ -0,0 +1,184 @@
+#include "turbopilot/model.hpp"
+
+#include <regex>
+#include <cmath>
+#include <random>
+
+
+void TurbopilotModel::lock(){
+    this->model_lock.lock();
+}
+
+void TurbopilotModel::unlock(){
+    this->model_lock.unlock();
+}
+
+std::stringstream TurbopilotModel::predict(std::string prompt, int max_length, bool include_prompt){
+    lock();
+    auto result = predict_impl(prompt, max_length, include_prompt);
+    unlock();
+    return result;
+}
+
+void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
+    (void) tensor;
+}
+
+void gpt_vocab::add_special_token(const std::string & token) {
+    special_tokens.push_back(token);
+}
+
+
+void gpt_split_words(std::string str, std::vector<std::string>& words) {
+    const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+    const std::regex re(pattern);
+    std::smatch m;
+
+    while (std::regex_search(str, m, re)) {
+        for (auto x : m) {
+            words.push_back(x);
+        }
+        str = m.suffix();
+    }
+}
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+
+        // Generate the subpattern from the special_tokens vector if it's not empty
+        if (!vocab.special_tokens.empty()) {
+            const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])");
+            std::string special_tokens_subpattern;
+            for (const auto & token : vocab.special_tokens) {
+                if (!special_tokens_subpattern.empty()) {
+                    special_tokens_subpattern += "|";
+                }
+                special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)");
+            }
+
+            std::regex re(special_tokens_subpattern);
+            std::smatch m;
+            // Split the text by special tokens.
+            while (std::regex_search(str, m, re)) {
+                // Split the substrings in-between special tokens into words.
+                gpt_split_words(m.prefix(), words);
+                // Add matched special tokens as words.
+                for (auto x : m) {
+                    words.push_back(x);
+                }
+                str = m.suffix();
+            }
+            // Remaining text without special tokens will be handled below.
+        }
+
+        gpt_split_words(str, words);
+    }
+
+    // find the longest token that forms each word in words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        for (int i = 0; i < (int) word.size(); ){
+            for (int j = word.size() - 1; j >= i; j--){
+                auto cand = word.substr(i, j-i+1);
+                auto it = vocab.token_to_id.find(cand);
+                if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab
+                    tokens.push_back(it->second);
+                    i = j + 1;
+                    break;
+                }
+                else if (j == i){ // word.substr(i, 1) has no matching
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
+                    i++;
+                }
+            }
+        }
+    }
+
+    return tokens;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[i]*scale, i));
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) probs.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+

src/gptj.cpp

@@ -0,0 +1,702 @@
+#include <turbopilot/gptj.hpp>
+#include <spdlog/spdlog.h>
+
+#include <ggml/ggml.h>
+
+#include <iostream>
+#include <fstream>
+
+
+#ifdef GGML_USE_CLBLAST
+#include "ggml-opencl.h"
+#endif
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+// The GPT-J model requires about 16MB of memory per input token.
+//
+bool gptj_eval(
+        const gptj_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n{}: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            spdlog::error("{}: failed to allocate {} bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            cur = ggml_norm(ctx0, inpL);
+
+            // cur = ln_1_g*cur + ln_1_b
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        struct ggml_tensor * inpSA = cur;
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+
+            // store key and value to memory
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+        }
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+        {
+            // note here we pass inpSA instead of cur
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    inpSA);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // cur = proj_w*cur + proj_b
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // self-attention + FF
+        cur  = ggml_add(ctx0, cur, inpFF);
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpL);
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        inpL = ggml_add(ctx0,
+                ggml_repeat(ctx0, model.lmh_b, inpL),
+                inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-j.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+
+GPTJModel::~GPTJModel(){
+    ggml_free(model->ctx);
+    free(model);
+    free(vocab);
+}
+
+bool GPTJModel::load_model(std::string fname) {
+    spdlog::info("{}: loading model from '{}' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        spdlog::error("{}: failed to open '{}'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            spdlog::error("{}: invalid model file '{}' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model->hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        spdlog::info("{}: n_vocab = {}\n", __func__, hparams.n_vocab);
+        spdlog::info("{}: n_ctx   = {}\n", __func__, hparams.n_ctx);
+        spdlog::info("{}: n_embd  = {}\n", __func__, hparams.n_embd);
+        spdlog::info("{}: n_head  = {}\n", __func__, hparams.n_head);
+        spdlog::info("{}: n_layer = {}\n", __func__, hparams.n_layer);
+        spdlog::info("{}: n_rot   = {}\n", __func__, hparams.n_rot);
+        spdlog::info("{}: ftype   = {}\n", __func__, hparams.ftype);
+        spdlog::info("{}: qntvr   = {}\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model->hparams.n_vocab) {
+            spdlog::error("{}: invalid model file '{}' (bad vocab size {} != {})\n",
+                    __func__, fname.c_str(), n_vocab, model->hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab->token_to_id[word] = i;
+            vocab->id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model->hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        spdlog::error("{}: invalid model file '{}' (bad ftype value {})\n",
+                __func__, fname.c_str(), model->hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model->ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
+        ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_v
+
+        ctx_size += (5 + 10*n_layer)*512; // object overhead
+
+        spdlog::info("{}: ggml ctx size = {} MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model->ctx = ggml_init(params);
+        if (!model->ctx) {
+            spdlog::error("{}: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model->layers.resize(n_layer);
+
+        model->wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model->ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model->ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model->lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model->lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model->tensors["transformer.wte.weight"] = model->wte;
+
+        model->tensors["transformer.ln_f.weight"] = model->ln_f_g;
+        model->tensors["transformer.ln_f.bias"]   = model->ln_f_b;
+
+        model->tensors["lm_head.weight"] = model->lmh_g;
+        model->tensors["lm_head.bias"]   = model->lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model->layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model->tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"]          = layer.ln_1_g;
+            model->tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"]            = layer.ln_1_b;
+
+            model->tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"]   = layer.c_attn_q_proj_w;
+            model->tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"]   = layer.c_attn_k_proj_w;
+            model->tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"]   = layer.c_attn_v_proj_w;
+
+            model->tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
+
+            model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"]     = layer.c_mlp_fc_w;
+            model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"]       = layer.c_mlp_fc_b;
+
+            model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"]    = layer.c_mlp_proj_w;
+            model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"]      = layer.c_mlp_proj_b;
+        }
+    }
+
+
+
+
+    // key + value memory
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model->memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model->memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model->memory_k) + ggml_nbytes(model->memory_v);
+
+        spdlog::info("{}: memory_size = {} MB, n_mem = {}\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        spdlog::info("{}: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model->tensors.find(name.data()) == model->tensors.end()) {
+                spdlog::error("{}: unknown tensor '{}' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model->tensors[name.data()];
+            if (ggml_nelements(tensor) != nelements) {
+                spdlog::error("{}: tensor '{}' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                spdlog::error("{}: tensor '{}' has wrong shape in model file: got [{}, {}], expected [{}, {}]\n",
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                spdlog::error("{}: tensor '{}' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+            
+        }
+
+        printf("\n");
+        spdlog::info(" done\n");
+
+        spdlog::info("{}: model size = {:06.2f} MB / num tensors = {}\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+
+
+    #if defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+    if(config.n_gpu_layers > 0){
+        size_t vram_total = 0;
+        int gpu_layers = std::min(config.n_gpu_layers, model->hparams.n_layer);
+        spdlog::info("Attempting to offload {} layers to GPU", gpu_layers);
+
+        for(int i=0; i < gpu_layers; i++) {
+            const auto & layer = model->layers[i];
+            layer.c_attn_q_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_attn_k_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_attn_v_proj_w->backend = GGML_BACKEND_GPU;
+            
+            layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
+
+            #if defined(GGML_USE_CLBLAST)
+            ggml_cl_transform_tensor(layer.c_attn_q_proj_w->data,layer.c_attn_q_proj_w); vram_total += ggml_nbytes(layer.c_attn_q_proj_w);
+            ggml_cl_transform_tensor(layer.c_attn_k_proj_w->data,layer.c_attn_k_proj_w); vram_total += ggml_nbytes(layer.c_attn_k_proj_w);
+            ggml_cl_transform_tensor(layer.c_attn_v_proj_w->data,layer.c_attn_v_proj_w); vram_total += ggml_nbytes(layer.c_attn_v_proj_w);
+            ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #else
+            ggml_cuda_transform_tensor(layer.c_attn_q_proj_w->data,layer.c_attn_q_proj_w); vram_total += ggml_nbytes(layer.c_attn_q_proj_w);
+            ggml_cuda_transform_tensor(layer.c_attn_k_proj_w->data,layer.c_attn_k_proj_w); vram_total += ggml_nbytes(layer.c_attn_k_proj_w);
+            ggml_cuda_transform_tensor(layer.c_attn_v_proj_w->data,layer.c_attn_v_proj_w); vram_total += ggml_nbytes(layer.c_attn_v_proj_w);
+            ggml_cuda_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #endif
+        }
+
+        spdlog::info("{}: [GPU] total VRAM used: {} MB\n", __func__, vram_total / 1024 / 1024);
+    }
+    
+    #endif // defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+    return true;
+}
+
+std::stringstream GPTJModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
+
+    std::stringstream result;
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize((*vocab), prompt);
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    int n_predict = std::min(max_length, model->hparams.n_ctx - (int) embd_inp.size());
+
+    spdlog::debug("{}: number of tokens in prompt = {}", __func__, embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+
+    std::vector<float> logits;
+
+    gptj_eval((*model), config.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!gptj_eval((*model), config.n_threads, n_past, embd, logits, mem_per_token)) {
+                throw std::runtime_error("Failed to predict");
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = config.top_k;
+            const float top_p = config.top_p;
+            const float temp  = config.temp;
+
+            const int n_vocab = model->hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p((*vocab), logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+
+            if(id != 50256){
+                result << vocab->id_to_token[id].c_str();
+            }
+            
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (int k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+
+                if(include_prompt){
+                    result << vocab->id_to_token[embd_inp[k]].c_str();
+                }
+
+                if (embd.size() > config.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+
+    spdlog::debug("{}:   sample time = {:8.2f} ms\n", __func__, t_sample_us/1000.0f);
+    spdlog::debug("{}:  predict time = {:8.2f} ms / {:.2f} ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+    
+
+    return result;
+}

src/gptneox.cpp

@@ -0,0 +1,776 @@
+#include <turbopilot/gptneox.hpp>
+#include <spdlog/spdlog.h>
+
+#include <ggml/ggml.h>
+
+#ifdef GGML_USE_CLBLAST
+#include "ggml-opencl.h"
+#endif
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+#include <cinttypes>
+
+#include <iostream>
+#include <fstream>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+
+// feed-forward network
+ggml_tensor * gpt_neox_ff(
+        const gpt_neox_layer &layer,
+        ggml_context * ctx0,
+        ggml_tensor * inp) {
+    ggml_tensor * cur = ggml_norm(ctx0, inp);
+
+    cur = ggml_add(ctx0,
+        ggml_mul(ctx0,
+            ggml_repeat(ctx0, layer.ln_2_g, cur),
+            cur),
+        ggml_repeat(ctx0, layer.ln_2_b, cur));
+
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_fc_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_fc_b, cur),
+            cur);
+
+    // GELU activation
+    cur = ggml_gelu(ctx0, cur);
+
+    // projection
+    // cur = proj_w*cur + proj_b
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_proj_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_proj_b, cur),
+            cur);
+    return cur;
+}
+
+
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool gpt_neox_eval(
+        const gpt_neox_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 512u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scr0_size = 512*1024*1024;
+    static void * scr0 = malloc(scr0_size);
+
+    static size_t scr1_size = 512*1024*1024;
+    static void * scr1 = malloc(scr1_size);
+    
+
+    if (mem_per_token > 0 && (mem_per_token*N) > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        spdlog::debug("{}: reallocating buffer from {} to {} bytes\n", __func__, buf_size, buf_size_new);
+
+        
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+        // self-attention
+        {
+            {
+                cur = ggml_norm(ctx0, inpL);
+
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            }
+
+            // compute QKV
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_attn_w,
+                        cur);
+
+                cur = ggml_add(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
+
+            // using mode = 2 for GPT-NeoX mode
+            Qcur = ggml_rope_inplace(ctx0, Qcur, n_past, n_rot, 2, 0);
+            Kcur = ggml_rope_inplace(ctx0, Kcur, n_past, n_rot, 2, 0);
+
+            // store key and value to memory
+            {
+                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_proj_w,
+                        cur);
+
+                cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
+            }
+        }
+
+        ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
+
+        if (hparams.par_res == 0) {
+            struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
+
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpFF);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpFF);
+        } else {
+            struct ggml_tensor * inpFF = cur;
+
+            // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+            // note here we pass inpL instead of cur
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpL);
+
+            // layer input + FF
+            cur  = ggml_add(ctx0, cur, inpFF);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpL);
+        }
+    }
+
+    ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        //inpL = ggml_add(ctx0,
+        //        ggml_repeat(ctx0, model.lmh_b, inpL),
+        //        inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+    
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+
+    spdlog::debug("used_mem = {}\n", ggml_used_mem(ctx0));
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0) /  N; //* 4;
+        spdlog::debug("Set mem_per_token={} / {} * {} = {}", ggml_used_mem(ctx0), N, 4, mem_per_token);
+        
+    }
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+
+GPTNEOXModel::~GPTNEOXModel(){
+    ggml_free(model->ctx);
+    free(model);
+    free(vocab);
+}
+
+bool GPTNEOXModel::load_model(std::string fname) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model->hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: par_res = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model->hparams.n_vocab;
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab->token_to_id[word] = i;
+            vocab->id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model->hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model->hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model->ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model->hparams;
+
+        const size_t n_embd  = hparams.n_embd;
+        const size_t n_layer = hparams.n_layer;
+        const size_t n_ctx   = hparams.n_ctx;
+        const size_t n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);           // lmh_g
+        //ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 16*n_layer)*1024; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model->ctx = ggml_init(params);
+        if (!model->ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model->layers.resize(n_layer);
+
+        model->wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model->ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model->ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model->lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        //model->lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model->tensors["gpt_neox.embed_in.weight"] = model->wte;
+
+        model->tensors["gpt_neox.final_layer_norm.weight"] = model->ln_f_g;
+        model->tensors["gpt_neox.final_layer_norm.bias"]   = model->ln_f_b;
+
+        model->tensors["embed_out.weight"] = model->lmh_g;
+        //model->tensors["lm_head.bias"]   = model->lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model->layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"]   = layer.ln_1_b;
+
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"]   = layer.c_attn_attn_b;
+
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"]   = layer.c_attn_proj_b;
+
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"]   = layer.ln_2_b;
+
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"]   = layer.c_mlp_fc_b;
+
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
+            model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"]   = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int64_t n_mem      = n_layer*n_ctx;
+        const int64_t n_elements = n_embd*n_mem;
+
+        model->memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model->memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model->memory_k) + ggml_nbytes(model->memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model->tensors.find(name.data()) == model->tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model->tensors[name.data()];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+    fin.close();
+
+    #if defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+
+    if(config.n_gpu_layers > 0){
+        size_t vram_total = 0;
+        int gpu_layers = std::min(config.n_gpu_layers, model->hparams.n_layer);
+        spdlog::info("Attempting to offload {} layers to GPU", gpu_layers);
+
+
+        for(int i=0; i < gpu_layers; i++) {
+            const auto & layer = model->layers[i];
+            layer.c_attn_attn_w->backend = GGML_BACKEND_GPU;
+            layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
+
+            #if defined(GGML_USE_CLBLAST)
+            ggml_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+            ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #else
+            ggml_cuda_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+            ggml_cuda_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #endif
+        }
+
+        spdlog::info("{}: [GPU] total VRAM used: {} MB\n", __func__, vram_total / 1024 / 1024);
+    }
+    
+    #endif // defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+    return true;
+}
+
+std::stringstream GPTNEOXModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
+
+    std::stringstream result;
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize((*vocab), prompt);
+
+    auto END_TOKEN_ID = vocab->token_to_id["<|endoftext|>"];
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    int n_predict = std::min(max_length, model->hparams.n_ctx - (int) embd_inp.size());
+
+    spdlog::debug("{}: number of tokens in prompt = {}", __func__, embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd;
+
+
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+
+    std::vector<float> logits;
+
+    std::vector<gpt_vocab::id> test = {};
+
+    for(int i=0;i<64;i++){
+        test.push_back(i);
+    }
+
+    gpt_neox_eval((*model), config.n_threads, 0, test, logits, mem_per_token);
+
+    const int64_t t_start_us = ggml_time_us();
+
+    int64_t t_prompt_us = 0;
+    
+    int64_t t_response_us = 0;
+
+    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!gpt_neox_eval((*model), config.n_threads, n_past, embd, logits, mem_per_token)) {
+                throw std::runtime_error("Failed to predict");
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = config.top_k;
+            const float top_p = config.top_p;
+            const float temp  = config.temp;
+
+            const int n_vocab = model->hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p((*vocab), logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+
+            // do not actually add endoftext char to the end string
+            if(id != END_TOKEN_ID){            
+                result << vocab->id_to_token[id].c_str();
+            }
+            
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (int k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+
+                if(include_prompt){
+                    result << vocab->id_to_token[embd_inp[k]].c_str();
+                }
+
+                if (embd.size() > config.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+
+
+        // end of text token
+        //if (embd.back() == 50256) {
+        if(embd.back() == END_TOKEN_ID){     
+            break;
+        }
+    }
+
+    spdlog::debug("{}:   sample time = {:8.2f} ms\n", __func__, t_sample_us/1000.0f);
+    spdlog::debug("{}:  predict time = {:8.2f} ms / {:.2f} ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+    
+    
+
+    return result;
+}
+

src/main.cpp

@@ -0,0 +1,196 @@
+#include <iostream>
+#include <cstdio>
+#include <ggml/ggml.h>
+
+#include <spdlog/spdlog.h>
+
+#include <turbopilot/crow_all.h>
+
+#include <argparse/argparse.hpp>
+
+#include "turbopilot/model.hpp"
+#include "turbopilot/starcoder.hpp"
+#include "turbopilot/gptj.hpp"
+#include "turbopilot/gptneox.hpp"
+#include "turbopilot/server.hpp"
+
+
+#define TURBOPILOT_VERSION "1.1.0"
+
+int main(int argc, char **argv)
+{
+
+    argparse::ArgumentParser program("turbopilot", TURBOPILOT_VERSION);
+
+    program.add_argument("--debug")
+        .default_value(false)
+        .help("Output verbose logs and timings")
+        .implicit_value(true);
+
+    program.add_argument("-f", "--model-file")
+        .help("Path to the model that turbopilot should serve")
+        .required();
+
+    program.add_argument("-m", "--model-type")
+        .help("The type of model to load. Can be codegen,starcoder,wizardcoder,stablecode")
+        .default_value("codegen");
+
+    program.add_argument("-t", "--threads")
+        .help("The number of CPU threads turbopilot is allowed to use. Defaults to 4")
+        .default_value(4)
+        .scan<'i', int>();
+
+
+    program.add_argument("--ngl", "--n-gpu-layers")
+        .help("The number of layers to offload to GPU")
+        .default_value(0)
+        .scan<'i', int>();
+
+
+    program.add_argument("-p", "--port")
+        .help("The tcp port that turbopilot should listen on")
+        .default_value(18080)
+        .scan<'i', int>();
+
+    program.add_argument("-r", "--random-seed")
+        .help("Set the random seed for RNG functions")
+        .default_value(-1)
+        .scan<'i', int>();
+
+    program.add_argument("--temperature")
+        .help("Set the generation temperature")
+        .default_value(0.2f)
+        .scan<'g', float>();
+
+    program.add_argument("--top-p")
+        .help("Set the generation top_p")
+        .default_value(0.1f)
+        .scan<'g', float>();
+
+    program.add_argument("-b", "--batch-size")
+        .help("set batch size for model completion")
+        .default_value(512)
+        .scan<'i',int>();
+
+
+    program.add_argument("prompt").remaining();
+
+
+
+    try
+    {
+        program.parse_args(argc, argv);
+    }
+    catch (const std::runtime_error &err)
+    {
+        std::cerr << err.what() << std::endl;
+        std::cerr << program;
+        return 1;
+    }
+
+    if(program.get<bool>("--debug")){
+        spdlog::set_level(spdlog::level::level_enum::debug);
+        spdlog::debug("debug logging enabled");
+    }
+
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+
+    TurbopilotModel *model = NULL;
+
+    auto model_type = program.get<std::string>("--model-type");
+
+    ModelConfig config{};
+    std::mt19937 rng(program.get<int>("--random-seed"));
+
+    config.n_threads = program.get<int>("--threads");
+    config.temp = program.get<float>("--temperature");
+    config.top_p = program.get<float>("--top-p");
+    config.n_batch = program.get<int>("--batch-size");
+    config.n_gpu_layers = program.get<int>("--ngl");
+
+    if(model_type.compare("codegen") == 0) {
+        spdlog::info("Initializing GPT-J type model for '{}' model", model_type);
+        model = new GPTJModel(config, rng);
+    }else if(model_type.compare("starcoder") == 0 || model_type.compare("wizardcoder") == 0){
+        spdlog::info("Initializing Starcoder/Wizardcoder type model for '{}' model type", model_type);
+        model = new StarcoderModel(config, rng);
+    }else if(model_type.compare("stablecode") == 0){
+        spdlog::info("Initializing StableLM type model for '{}' model type", model_type);
+        model = new GPTNEOXModel(config, rng);
+    }else{
+        spdlog::error("Invalid model type: {}", model_type);
+    }
+
+    spdlog::info("Attempt to load model from {}", program.get<std::string>("--model-type"));
+    int64_t t_load_us = 0;
+    const int64_t t_start_us = ggml_time_us();
+    auto loaded = model->load_model(program.get<std::string>("--model-file"));
+
+    if(!loaded){
+        spdlog::error("Failed to load model");
+        return -1;
+    }
+
+    t_load_us = ggml_time_us() - t_start_us;
+    
+
+    spdlog::info("Loaded model in {:0.2f}ms", t_load_us/1000.0f);
+
+
+    crow::SimpleApp app;
+
+    TBPLogger logger;
+
+    crow::logger::setHandler(&logger);
+
+    CROW_ROUTE(app, "/")([](){
+        return "Hello world";
+    });
+
+
+    CROW_ROUTE(app, "/copilot_internal/v2/token")([](){
+        //return "Hello world";
+
+        crow::json::wvalue response = {{"token","1"}, {"expires_at", static_cast<std::uint64_t>(2600000000)}, {"refresh_in",900}};
+
+        crow::response res;
+        res.code = 200;
+        res.set_header("Content-Type", "application/json");
+        res.body = response.dump();
+        return res;
+    });
+
+    //huggingface code compatible endpoint
+    CROW_ROUTE(app, "/api/generate").methods(crow::HTTPMethod::Post)
+    ([&model](const crow::request& req) {
+        return handle_hf_request(model, req);
+    });
+
+    CROW_ROUTE(app, "/v1/completions").methods(crow::HTTPMethod::Post)
+    ([&model](const crow::request& req) {
+        return handle_openai_request(model, req);
+    });
+
+    CROW_ROUTE(app, "/v1/engines/codegen/completions").methods(crow::HTTPMethod::Post)
+    ([&model](const crow::request& req) {
+        return handle_openai_request(model, req);
+    });
+
+
+    CROW_ROUTE(app, "/v1/engines/copilot-codex/completions").methods(crow::HTTPMethod::Post)
+    ([&model](const crow::request& req) {
+        return handle_openai_request(model, req);
+    });
+    
+
+    app.port(program.get<int>("--port")).multithreaded().run();
+
+    
+
+    free(model);
+}
+
+

src/server.cpp

@@ -0,0 +1,110 @@
+
+#include "turbopilot/server.hpp"
+#include "turbopilot/model.hpp"
+
+#include <boost/lexical_cast.hpp>
+#include <boost/uuid/uuid.hpp>            // uuid class
+#include <boost/uuid/uuid_generators.hpp> // generators
+#include <boost/uuid/uuid_io.hpp>         // streaming operators etc.
+
+
+
+/**
+ * This function serves requests for autocompletion from crow
+ * 
+*/
+crow::response handle_hf_request(TurbopilotModel *model, const crow::request& req){
+
+        crow::json::rvalue data = crow::json::load(req.body);
+
+        if(!data.has("inputs")){
+            crow::response res;
+            res.code = 400;
+            res.set_header("Content-Type", "application/json");
+            res.body = "{\"message\":\"you must specify inputs field or\"}";
+            return res;
+        }
+
+        // std::string suffix = data["suffix"].s();
+        int maxTokens = 200;
+        if(data.has("max_tokens")){
+            maxTokens = data["max_tokens"].i();
+        }
+
+
+        auto result = model->predict(data["inputs"].s(), maxTokens, false);
+
+        crow::json::wvalue response = {
+            {"generated_text", result.str()},
+        };
+
+        
+        crow::response res;
+        res.code = 200;
+        res.set_header("Content-Type", "application/json");
+        res.body = response.dump(); //ss.str();
+        return res;
+}
+
+/**
+ * This function serves requests for autocompletion from crow
+ * 
+*/
+crow::response handle_openai_request(TurbopilotModel *model, const crow::request& req){
+
+        crow::json::rvalue data = crow::json::load(req.body);
+
+        if(!data.has("prompt") && !data.has("input_ids")){
+            crow::response res;
+            res.code = 400;
+            res.set_header("Content-Type", "application/json");
+            res.body = "{\"message\":\"you must specify a prompt or input_ids\"}";
+            return res;
+        }
+
+        // std::string suffix = data["suffix"].s();
+        int maxTokens = 200;
+        if(data.has("max_tokens")){
+            maxTokens = data["max_tokens"].i();
+        }
+
+
+        auto result = model->predict(data["prompt"].s(), maxTokens, false);
+
+
+        boost::uuids::uuid uuid = boost::uuids::random_generator()();
+
+
+        // Generate a mock response based on the input parameters
+        crow::json::wvalue choice = {
+            {"text", result.str()},
+            {"index",0},
+            {"finish_reason", "length"},
+            {"logprobs", nullptr}
+        };
+        crow::json::wvalue::list choices = {choice};
+
+
+        crow::json::wvalue usage = {
+            {"completion_tokens", 0},
+            // {"prompt_tokens", static_cast<std::uint64_t>(embd_inp.size())},
+            {"prompt_tokens", 0},
+            {"total_tokens", 0}
+        };
+
+        crow::json::wvalue response = {
+            {"id", boost::lexical_cast<std::string>(uuid)},
+            {"model", "codegen"},
+            {"object","text_completion"},
+            {"created", static_cast<std::int64_t>(std::time(nullptr))},
+            {"choices", choices },
+            {"usage", usage}
+        };
+
+        crow::response res;
+        res.code = 200;
+        res.set_header("Content-Type", "application/json");
+
+        res.body = response.dump(); //ss.str();
+        return res;
+}

src/starcoder.cpp

@@ -0,0 +1,834 @@
+#include <iostream>
+#include <fstream>
+
+#include <turbopilot/starcoder.hpp>
+#include <ggml/ggml.h>
+#include <spdlog/spdlog.h>
+
+#ifdef GGML_USE_CLBLAST
+#include "ggml-opencl.h"
+#endif
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool starcoder_eval(
+        const starcoder_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+
+    static size_t buf_size = 512u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scr0_size = 512u*1024*1024;
+    static void * scr0 = malloc(scr0_size);
+
+    static size_t scr1_size = 512u*1024*1024;
+    static void * scr1 = malloc(scr1_size);
+
+    if (mem_per_token > 0 && 2*mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 2*(mem_per_token*N); // add 10% to account for ggml object overhead
+
+        if(buf_size_new > buf_size){
+            spdlog::debug("{}: reallocating buffer from {} to {} bytes\n", __func__, buf_size, buf_size_new);
+
+            // reallocate
+            buf_size = buf_size_new;
+            buf = realloc(buf, buf_size);
+            if (buf == nullptr) {
+                spdlog::error("{}: failed to allocate {} bytes\n", __func__, buf_size);
+                return false;
+            } 
+        }
+
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    for (int i = 0; i < N; ++i) {
+        ((int32_t *) position->data)[i] = n_past + i;
+    }
+
+    // wte + wpe
+    struct ggml_tensor * inpL =
+        ggml_add(ctx0,
+                ggml_get_rows(ctx0, model.wte, embd),
+                ggml_get_rows(ctx0, model.wpe, position));
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+        // norm
+        {
+            // [ 768, N]
+            cur = ggml_norm(ctx0, inpL);
+
+            // cur = ln_1_g*cur + ln_1_b
+            // [ 768, N]
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        // attn
+        // [2304, 768] - model.layers[il].c_attn_attn_w
+        // [2304,   1] - model.layers[il].c_attn_attn_b
+        // [ 768,   N] - cur (in)
+        // [2304,   N] - cur (out)
+        //
+        // cur = attn_w*cur + attn_b
+        // [2304, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_attn_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                    cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [64, N, 12]
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            // [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3); //TODO: need to be tiled
+
+            // GG: flash attention
+            //struct ggml_tensor * V =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0,
+            //                ggml_reshape_3d(ctx0,
+            //                    ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+            //                    n_embd/n_head, n_head, n_past + N),
+            //                1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
+
+            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
+
+            // K * Q
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //TODO: check if it broadcasts
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [64, N, 12]
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            // [64, 12, N]
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            // [768, N]
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+        }
+
+        // projection
+        // [ 768, 768] - model.layers[il].c_attn_proj_w
+        // [ 768,   1] - model.layers[il].c_attn_proj_b
+        // [ 768,   N] - cur (in)
+        // [ 768,   N] - cur (out)
+        //
+        // cur = proj_w*cur + proj_b
+        // [768, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
+                    cur);
+        }
+
+        // add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF);
+
+                // cur = ln_2_g*cur + ln_2_b
+                // [ 768, N]
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            }
+
+            // fully connected
+            // [3072, 768] - model.layers[il].c_mlp_fc_w
+            // [3072,   1] - model.layers[il].c_mlp_fc_b
+            // [ 768,   N] - cur (in)
+            // [3072,   N] - cur (out)
+            //
+            // cur = fc_w*cur + fc_b
+            // [3072, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            // [3072, N]
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
+            // [ 768,    1] - model.layers[il].c_mlp_proj_b
+            // [3072,    N] - cur (in)
+            // [ 768,    N] - cur (out)
+            //
+            // cur = proj_w*cur + proj_b
+            // [768, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+    // norm
+    {
+        // [ 768, N]
+        inpL = ggml_norm(ctx0, inpL);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        // [ 768, N]
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.lm_head
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result just for the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+
+    spdlog::debug("{}: used mem buf={} bytes", __func__, ggml_used_mem(ctx0));
+    
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+
+StarcoderModel::~StarcoderModel(){
+    ggml_free(model->ctx);
+    free(model);
+    free(vocab);
+}
+
+bool StarcoderModel::load_model(std::string fname) {
+  printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model->hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model->hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model->hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab->token_to_id[word] = i;
+            vocab->id_to_token[i] = word;
+
+            // if (i < 10) fprintf(stderr, "%.s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
+        }
+
+        // Add StarChat special tokens.
+        for (const std::string & token : {
+                "<|system|>",
+                "<|user|>",
+                "<|assistant|>",
+                "<|end|>",
+                "<fim-prefix>",
+                "<fim-middle>",
+                "<fim-suffix>",
+                "<fim-pad>",
+                "<|end_of_turn|>"
+            }) {
+            if (vocab->token_to_id.find(token) != vocab->token_to_id.end()) {
+                vocab->add_special_token(token);
+            }
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model->hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model->hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model->ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w // TODO:
+        ctx_size += n_layer*(       (n_embd + 2*kv_dim)*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 12*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model->ctx = ggml_init(params);
+        if (!model->ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+        model->layers.resize(n_layer);
+
+        model->ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model->ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model->wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model->wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model->lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model->tensors["model/ln_f/g"] = model->ln_f_g;
+        model->tensors["model/ln_f/b"] = model->ln_f_b;
+
+        model->tensors["model/wte"]     = model->wte;
+        model->tensors["model/wpe"]     = model->wpe;
+        model->tensors["model/lm_head"] = model->lm_head;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model->layers[i];
+
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd + 2*kv_dim);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd + 2*kv_dim);
+
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd); //TODO: 4*n_embd = config.n_inner
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model->tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
+            model->tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
+
+            model->tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
+            model->tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
+
+            model->tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
+            model->tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
+
+            model->tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
+            model->tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
+
+            model->tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
+            model->tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
+
+            model->tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
+            model->tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model->hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model->memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model->memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model->memory_k) + ggml_nbytes(model->memory_v);
+
+        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        bool has_lm_head = false;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model->tensors.find(name.data()) == model->tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model->tensors[name.data()];
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file. got %d, expected %d\n",
+                        __func__, name.data(), (int) ggml_nelements(tensor), nelements);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model->lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
+            total_size += ggml_nbytes(tensor);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+
+    #if defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+    if(config.n_gpu_layers > 0){
+        size_t vram_total = 0;
+        int gpu_layers = std::min(config.n_gpu_layers, model->hparams.n_layer);
+        spdlog::info("Attempting to offload {} layers to GPU", gpu_layers);
+
+
+        for(int i=0; i < gpu_layers; i++) {
+            const auto & layer = model->layers[i];
+            layer.c_attn_attn_w->backend = GGML_BACKEND_GPU;
+            layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
+
+            #if defined(GGML_USE_CLBLAST)
+            ggml_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+            ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #else
+            ggml_cuda_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+            ggml_cuda_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cuda_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+            #endif
+        }
+
+        spdlog::info("{}: [GPU] total VRAM used: {} MB\n", __func__, vram_total / 1024 / 1024);
+    }
+    
+    #endif // defined(GGML_USE_CLBLAST) || defined(GGML_USE_CUBLAS)
+
+
+    return true;
+}
+
+
+std::stringstream StarcoderModel::predict_impl(std::string prompt, int max_length, bool include_prompt) {
+
+    std::stringstream result;
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize((*vocab), prompt);
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    int n_predict = std::min(max_length, model->hparams.n_ctx - (int) embd_inp.size());
+
+    spdlog::debug("{}: number of tokens in prompt = {}", __func__, embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+
+    std::vector<float> logits;
+        std::vector<gpt_vocab::id> test = {};
+
+    for(int i=0;i<64;i++){
+        test.push_back(i);
+    }
+
+    spdlog::debug("{}: calculate required memory per token", __func__);
+    starcoder_eval((*model), config.n_threads, 0, test, logits, mem_per_token);
+    spdlog::debug("{}: mem_per_token={}", __func__, mem_per_token);
+    spdlog::debug("{}: total mem needed for prompt = {}*{}={}", __func__, embd_inp.size(), mem_per_token, embd_inp.size()*mem_per_token);
+
+
+    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+        // predict
+        spdlog::debug("{}: process token #{}: ", __func__, i);
+
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!starcoder_eval((*model), config.n_threads, n_past, embd, logits, mem_per_token)) {
+                throw std::runtime_error("Failed to predict");
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = config.top_k;
+            const float top_p = config.top_p;
+            const float temp  = config.temp;
+
+            const int n_vocab = model->hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p((*vocab), logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+
+            if(id != 50256){
+                result << vocab->id_to_token[id].c_str();
+            }
+            
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (int k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+
+                if(include_prompt){
+                    result << vocab->id_to_token[embd_inp[k]].c_str();
+                }
+
+                if (embd.size() > config.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+    spdlog::debug("{}:   sample time = {:8.2f} ms\n", __func__, t_sample_us/1000.0f);
+    spdlog::debug("{}:  predict time = {:8.2f} ms / {:.2f} ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+    
+    
+
+    return result;
+}

test.txt

@@ -0,0 +1,10 @@
+#%%
+import os
+import cats
+
+
+
+
+
+
+

test_codegen2.py

@@ -0,0 +1,65 @@
+#%%
+from transformers import AutoTokenizer, AutoModelForCausalLM
+tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen2-1B")
+model = AutoModelForCausalLM.from_pretrained("Salesforce/codegen2-1B", trust_remote_code=True, revision="main")
+
+
+#%%
+model = model.to(device="cuda")
+
+#%%
+text = """
+import os
+
+def post_to_pastebin"""
+input_ids = tokenizer(text, return_tensors="pt").to("cuda").input_ids
+generated_ids = model.generate(input_ids, max_length=512)
+print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
+
+# %%
+
+def format_model_input(prefix, suffix):
+  return prefix + "<mask_1>" + suffix + "<|endoftext|>" + "<sep>" + "<mask_1>"
+
+
+prefix = """
+import os
+
+def post_to_pastebin"""
+suffix = "result = post_to_pastebin(content)"
+text = format_model_input(prefix, suffix)
+input_ids = tokenizer(text, return_tensors="pt").to("cuda").input_ids
+generated_ids = model.generate(input_ids, max_length=128)
+print(tokenizer.decode(generated_ids[0], skip_special_tokens=False))
+# %%
+def main():
+  text = """
+
+  print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
+
+if __name__ == '__main__':
+    main()
+
+  print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
+  # %%
+
+  import os
+
+  def post_to_pastebin"""
+  input_ids = tokenizer(text, return_tensors="pt").to("cuda").input_ids
+  generated_ids = model.generate(input_ids, max_length=512)
+
+
+ print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
+
+# %%
+
+def post_to_pastebin(content):
+  input_ids = tokenizer(content, return_tensors="pt").to("cuda").input_ids
+  generated_ids = model.generate(input_ids, max_length=512)
+  return tokenizer.decode(generated_ids[0], skip_special_tokens=True)
+
+
+
+
+  

test_santa.py

@@ -0,0 +1,45 @@
+#%%
+import torch
+from transformers import CodeGenTokenizer, GPTJForCausalLM
+
+
+checkpoint = "/home/james/workspace/rafael-llm/codegen-2B-multi-gptj"
+device = "cuda" # for GPU usage or "cpu" for CPU usage
+
+tokenizer = CodeGenTokenizer.from_pretrained("Salesforce/codegen-350M-multi")
+model = GPTJForCausalLM.from_pretrained(checkpoint).to(device)
+
+
+#model = AutoModel.from_pretrained(checkpoint, trust_remote_code=True).to(device)
+#%%
+
+# define the user model
+class User:
+
+
+# %%
+code = """import os
+import requests
+
+#send the json data to pastebin
+def send_data"""
+inputs = tokenizer.encode(code, return_tensors="pt").to(device)
+outputs = model.generate(inputs, max_length=200)
+response = tokenizer.decode(outputs[0])
+
+print(response)
+
+import requests
+
+#send the json data to pastebin
+def send_data(data):
+    url = "http://pastebin.com/api_post.php"
+    data = {"api_dev_key": "<api_key>", "api_user_key": "<user_key>", "api_content": data}
+    response = requests.post(url, data=data).text
+    return response
+
+
+
+# %%
+code
+# %%

turbopilot.code-workspace

@@ -0,0 +1,94 @@
+{
+	"folders": [
+		{
+			"path": "."
+		},
+		{
+			"path": "extern/ggml"
+		},
+		{
+			"path": "../../pymicrocosm"
+		}
+	],
+	"settings": {
+		"files.associations": {
+			"array": "cpp",
+			"atomic": "cpp",
+			"bit": "cpp",
+			"*.tcc": "cpp",
+			"bitset": "cpp",
+			"cctype": "cpp",
+			"chrono": "cpp",
+			"clocale": "cpp",
+			"cmath": "cpp",
+			"compare": "cpp",
+			"concepts": "cpp",
+			"cstdint": "cpp",
+			"cstdio": "cpp",
+			"cstdlib": "cpp",
+			"cstring": "cpp",
+			"ctime": "cpp",
+			"cwchar": "cpp",
+			"cwctype": "cpp",
+			"deque": "cpp",
+			"map": "cpp",
+			"unordered_map": "cpp",
+			"vector": "cpp",
+			"exception": "cpp",
+			"fstream": "cpp",
+			"functional": "cpp",
+			"initializer_list": "cpp",
+			"iosfwd": "cpp",
+			"istream": "cpp",
+			"limits": "cpp",
+			"memory": "cpp",
+			"new": "cpp",
+			"numbers": "cpp",
+			"numeric": "cpp",
+			"ostream": "cpp",
+			"ratio": "cpp",
+			"regex": "cpp",
+			"semaphore": "cpp",
+			"sstream": "cpp",
+			"stdexcept": "cpp",
+			"stop_token": "cpp",
+			"streambuf": "cpp",
+			"string": "cpp",
+			"string_view": "cpp",
+			"system_error": "cpp",
+			"thread": "cpp",
+			"type_traits": "cpp",
+			"tuple": "cpp",
+			"typeinfo": "cpp",
+			"utility": "cpp",
+			"csignal": "cpp",
+			"cstdarg": "cpp",
+			"cstddef": "cpp",
+			"any": "cpp",
+			"strstream": "cpp",
+			"charconv": "cpp",
+			"cinttypes": "cpp",
+			"codecvt": "cpp",
+			"complex": "cpp",
+			"condition_variable": "cpp",
+			"coroutine": "cpp",
+			"list": "cpp",
+			"set": "cpp",
+			"algorithm": "cpp",
+			"iterator": "cpp",
+			"memory_resource": "cpp",
+			"optional": "cpp",
+			"random": "cpp",
+			"source_location": "cpp",
+			"future": "cpp",
+			"iomanip": "cpp",
+			"iostream": "cpp",
+			"mutex": "cpp",
+			"span": "cpp",
+			"cfenv": "cpp",
+			"typeindex": "cpp",
+			"variant": "cpp",
+			"unordered_set": "cpp"
+		}
+	}
+}