civ/hot-step-cpp-ROCm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial release

Browse Source
This commit is contained in:
civ
2026-08-16 18:24:52 +07:00
commit 876886a39a
13244 changed files with 2353959 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
engine/.clang-format
+170
View File
@@ -0,0 +1,170 @@
---
Language: Cpp
AlignAfterOpenBracket: Align
AlignArrayOfStructures: Left
AlignConsecutiveAssignments: AcrossComments
AlignConsecutiveBitFields: AcrossComments
AlignConsecutiveDeclarations: AcrossComments
AlignConsecutiveMacros: AcrossComments
# AlignConsecutiveShortCaseStatements: AcrossComments
AlignEscapedNewlines: Left # LeftWithLastLine
AlignOperands: Align
AlignTrailingComments:
Kind: Always
OverEmptyLines: 1
AllowAllArgumentsOnNextLine: true
AllowAllParametersOfDeclarationOnNextLine: false
# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen
AllowShortBlocksOnASingleLine: Never
AllowShortCaseLabelsOnASingleLine: false
AllowShortFunctionsOnASingleLine: Inline
AllowShortIfStatementsOnASingleLine: Never
AllowShortLambdasOnASingleLine: Inline
AllowShortLoopsOnASingleLine: false
AlwaysBreakBeforeMultilineStrings: true
# Treat CUDA keywords/attributes as "attribute macros" and avoid breaking lines inside them
AttributeMacros:
- __host__
- __device__
- __global__
- __forceinline__
- __launch_bounds__
BinPackArguments: true
BinPackParameters: false # OnePerLine
BitFieldColonSpacing: Both
BreakBeforeBraces: Custom # Attach
BraceWrapping:
AfterCaseLabel: true
AfterClass: false
AfterControlStatement: false
AfterEnum: false
AfterFunction: false
AfterNamespace: false
AfterObjCDeclaration: false
AfterStruct: false
AfterUnion: false
AfterExternBlock: false
BeforeCatch: false
BeforeElse: false
BeforeLambdaBody: false
BeforeWhile: false
IndentBraces: false
SplitEmptyFunction: false
SplitEmptyRecord: false
SplitEmptyNamespace: false
# BreakAdjacentStringLiterals: true
BreakAfterAttributes: Never
BreakBeforeBinaryOperators: None
BreakBeforeInlineASMColon: OnlyMultiline
BreakBeforeTernaryOperators: false
# BreakBinaryOperations: Never
BreakConstructorInitializers: AfterColon
# BreakFunctionDefinitionParameters: false
BreakInheritanceList: AfterComma
BreakStringLiterals: true
# BreakTemplateDeclarations: Yes
ColumnLimit: 120
CommentPragmas: '^ IWYU pragma:'
CompactNamespaces: false
ConstructorInitializerIndentWidth: 4
ContinuationIndentWidth: 4
Cpp11BracedListStyle: false
DerivePointerAlignment: false
DisableFormat: false
EmptyLineBeforeAccessModifier: Leave
EmptyLineAfterAccessModifier: Never
ExperimentalAutoDetectBinPacking: false
FixNamespaceComments: true
IncludeBlocks: Regroup
IncludeCategories:
- Regex: '".*"'
Priority: 1
SortPriority: 0
- Regex: '^<.*\.h>'
Priority: 2
SortPriority: 0
- Regex: '^<.*'
Priority: 3
SortPriority: 0
- Regex: '.*'
Priority: 4
SortPriority: 0
IncludeIsMainRegex: '([-_](test|unittest))?$'
IncludeIsMainSourceRegex: ''
IndentAccessModifiers: false
IndentCaseBlocks: true
IndentCaseLabels: true
IndentExternBlock: NoIndent
IndentGotoLabels: false
IndentPPDirectives: AfterHash
IndentWidth: 4
IndentWrappedFunctionNames: false
InsertBraces: true # NOTE: may lead to incorrect formatting
InsertNewlineAtEOF: true
JavaScriptQuotes: Leave
JavaScriptWrapImports: true
KeepEmptyLinesAtTheStartOfBlocks: false
LambdaBodyIndentation: Signature
LineEnding: LF
MacroBlockBegin: ''
MacroBlockEnd: ''
MaxEmptyLinesToKeep: 1
NamespaceIndentation: None
ObjCBinPackProtocolList: Auto
ObjCBlockIndentWidth: 4
ObjCSpaceAfterProperty: true
ObjCSpaceBeforeProtocolList: true
PPIndentWidth: -1
PackConstructorInitializers: CurrentLine
PenaltyBreakAssignment: 2
PenaltyBreakBeforeFirstCallParameter: 1
PenaltyBreakComment: 300
PenaltyBreakFirstLessLess: 120
PenaltyBreakString: 1000
PenaltyBreakTemplateDeclaration: 10
PenaltyExcessCharacter: 1000000
PenaltyReturnTypeOnItsOwnLine: 200
PointerAlignment: Middle
QualifierAlignment: Left
#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict']
RawStringFormats:
- Language: Cpp
Delimiters:
- cc
- CC
- cpp
- Cpp
- CPP
- 'c++'
- 'C++'
CanonicalDelimiter: ''
ReferenceAlignment: Middle
ReflowComments: false # IndentOnly
SeparateDefinitionBlocks: Always
SortIncludes: CaseInsensitive
SortUsingDeclarations: LexicographicNumeric
SpaceAfterCStyleCast: true
SpaceAfterLogicalNot: false
SpaceAfterTemplateKeyword: true
SpaceBeforeAssignmentOperators: true
SpaceBeforeCpp11BracedList: false
SpaceBeforeCtorInitializerColon: true
SpaceBeforeInheritanceColon: true
SpaceBeforeParens: ControlStatements
SpaceBeforeRangeBasedForLoopColon: true
SpaceInEmptyBlock: false
SpaceInEmptyParentheses: false
SpacesBeforeTrailingComments: 2
SpacesInAngles: Never
SpacesInContainerLiterals: true
SpacesInLineCommentPrefix:
Minimum: 1
Maximum: -1
SpacesInParentheses: false
SpacesInSquareBrackets: false
SpaceBeforeSquareBrackets: false
Standard: c++17
TabWidth: 4
UseTab: Never
WhitespaceSensitiveMacros: ['STRINGIZE']
...
engine/.gitattributes
+1
View File
@@ -0,0 +1 @@
* text=auto eol=lf
engine/.github/workflows/ci-build.yml
Vendored
+77
View File
@@ -0,0 +1,77 @@
# Validate that the project builds on Ubuntu and macOS (no model download).
name: CI Build
on:
pull_request:
workflow_dispatch:
jobs:
build:
runs-on: ${{ matrix.os }}
strategy:
fail-fast: false
matrix:
os: [ubuntu-latest, macos-latest]
steps:
- name: Checkout
uses: actions/checkout@v4
with:
submodules: recursive
- name: Build (Ubuntu)
if: matrix.os == 'ubuntu-latest'
run: |
sudo apt-get update -qq
sudo apt-get install -y -qq cmake build-essential pkg-config libopenblas-dev
mkdir build && cd build
cmake .. -DGGML_BLAS=ON
cmake --build . --config Release -j$(nproc)
- name: Build (macOS)
if: matrix.os == 'macos-latest'
run: |
mkdir build && cd build
cmake ..
cmake --build . --config Release -j$(sysctl -n hw.ncpu)
- name: Smoke test
run: |
./build/ace-lm --help 2>&1 | head -5
./build/ace-synth --help 2>&1 | head -5
./build/quantize --help 2>&1 | head -3
lint:
name: Lint & Static Analysis
runs-on: ubuntu-latest
if: github.event_name == 'pull_request'
steps:
- uses: actions/checkout@v4
- name: Install lint tools
run: |
sudo apt-get update -qq
sudo apt-get install -y -qq clang-format clang-tidy cppcheck
- name: Run clang-format (check mode)
run: |
find . \
\( -path './.git' -o -path './ggml' -o -path './build' -o -path './vendor' -o -path './mp3' \) -prune -o \
-type f \( -name '*.c' -o -name '*.h' -o -name '*.cc' -o -name '*.cpp' -o -name '*.hpp' \) \
-print0 | xargs -0 clang-format --dry-run --Werror
- name: Run cppcheck
run: |
cppcheck --enable=all --error-exitcode=1 --inline-suppr \
--suppress=missingIncludeSystem \
--suppress=missingInclude \
--suppress=cstyleCast \
--suppress=constVariable \
--suppress=constVariablePointer \
--suppress=constParameterPointer \
--suppress=variableScope \
--suppress=uselessCallsSubstr \
--suppress=useStlAlgorithm \
--suppress=shiftNegativeLHS \
-i ggml -i build -i .git -i mp3 \
.
engine/.github/workflows/release.yml
Vendored
+263
View File
@@ -0,0 +1,263 @@
name: Build & Release Binaries
on:
release:
types: [published]
workflow_dispatch:
inputs:
release_tag:
description: 'Release tag to attach binaries to (e.g. v0.1.0)'
required: true
type: string
skip_linux:
description: 'Skip Linux build'
required: false
type: boolean
default: false
skip_mac:
description: 'Skip macOS build'
required: false
type: boolean
default: false
skip_windows:
description: 'Skip Windows build'
required: false
type: boolean
default: false
permissions:
contents: write
jobs:
build-linux:
name: Build · linux-x64
runs-on: ubuntu-22.04
if: github.event_name != 'workflow_dispatch' || !inputs.skip_linux
steps:
- name: Checkout
uses: actions/checkout@v4
with:
submodules: recursive
- name: Install build tools
run: |
sudo apt-get update -qq
sudo apt-get install -y -qq cmake build-essential pkg-config libopenblas-dev
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2
with:
create-symlink: true
key: build-linux-ubuntu-22.04
- name: Install CUDA toolkit
uses: Jimver/cuda-toolkit@v0.2.30
with:
log-file-suffix: 'ubuntu-22.04.txt'
- name: Install Vulkan SDK
uses: humbletim/install-vulkan-sdk@v1.2
with:
version: 1.4.309.0
cache: true
- name: Configure & Build
run: |
./buildall.sh
- name: Smoke test
continue-on-error: true
shell: bash
run: |
BIN="build"
"$BIN/ace-lm" 2>&1 | head -5
"$BIN/ace-synth" 2>&1 | head -5
"$BIN/ace-understand" 2>&1 | head -5
"$BIN/neural-codec" 2>&1 | head -5
"$BIN/quantize" 2>&1 | head -3
"$BIN/mp3-codec" 2>&1 | head -3
- name: Resolve release tag
id: tag
shell: bash
run: |
if [ "${{ github.event_name }}" = "release" ]; then
echo "value=${{ github.event.release.tag_name }}" >> $GITHUB_OUTPUT
else
echo "value=${{ inputs.release_tag }}" >> $GITHUB_OUTPUT
fi
- name: Package binaries
run: |
mkdir -p dist
cp build/ace-* \
build/quantize build/neural-codec build/mp3-codec build/*.so dist/
tar -C dist -czf "acestep-linux-x64.tar.gz" .
- name: Upload to release
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: |
gh release upload "${{ steps.tag.outputs.value }}" \
"acestep-linux-x64.tar.gz" \
--clobber
build-mac:
name: Build · macos-arm64-metal
runs-on: macos-latest
if: github.event_name != 'workflow_dispatch' || !inputs.skip_mac
steps:
- name: Checkout
uses: actions/checkout@v4
with:
submodules: recursive
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2
with:
create-symlink: true
key: build-mac-macos-latest
- name: Configure & Build
run: |
mkdir build
cd build
cmake ..
cmake --build . --config Release -j "$(nproc)"
- name: Smoke test
continue-on-error: true
shell: bash
run: |
BIN="build"
"$BIN/ace-lm" 2>&1 | head -5
"$BIN/ace-synth" 2>&1 | head -5
"$BIN/ace-understand" 2>&1 | head -5
"$BIN/neural-codec" 2>&1 | head -5
"$BIN/quantize" 2>&1 | head -3
"$BIN/mp3-codec" 2>&1 | head -3
- name: Resolve release tag
id: tag
shell: bash
run: |
if [ "${{ github.event_name }}" = "release" ]; then
echo "value=${{ github.event.release.tag_name }}" >> $GITHUB_OUTPUT
else
echo "value=${{ inputs.release_tag }}" >> $GITHUB_OUTPUT
fi
- name: Package binaries
run: |
mkdir -p dist
cd build
for bin in ace-* quantize neural-codec mp3-codec; do
install_name_tool -add_rpath @executable_path "$bin"
done
cp -P ace-* quantize neural-codec mp3-codec libacestep*.a libggml*.dylib ../dist/
cd ..
tar -C dist -czf "acestep-macos-arm64-metal.tar.gz" .
- name: Upload to release
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: |
gh release upload "${{ steps.tag.outputs.value }}" \
"acestep-macos-arm64-metal.tar.gz" \
--clobber
build-windows:
name: Build · windows-x64
runs-on: windows-latest
if: github.event_name != 'workflow_dispatch' || !inputs.skip_windows
steps:
- name: Checkout
uses: actions/checkout@v4
with:
submodules: recursive
- name: Cache CUDA toolkit
id: cache-cuda
uses: actions/cache@v4
with:
path: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
key: cuda-12-windows-latest
- name: Install CUDA toolkit
if: steps.cache-cuda.outputs.cache-hit != 'true'
uses: Jimver/cuda-toolkit@v0.2.30
with:
log-file-suffix: 'windows-latest.txt'
- name: Install Vulkan SDK
uses: humbletim/install-vulkan-sdk@v1.2
with:
version: 1.4.309.0
cache: true
- name: Cache build directory
uses: actions/cache@v4
with:
path: build-msvc
key: build-msvc-${{ github.sha }}
restore-keys: |
build-msvc-
- name: Configure & Build
shell: pwsh
run: |
# Configure — only print errors
cmake -S . -B build-msvc `
-DGGML_CPU_ALL_VARIANTS=ON `
-DGGML_CUDA=ON `
-DGGML_VULKAN=ON `
-DGGML_BACKEND_DL=ON `
--log-level=ERROR 2>&1 | Where-Object { $_ -notmatch '^--' }
# Build — suppress per-file progress, only show warnings/errors
cmake --build build-msvc --config Release -j $env:NUMBER_OF_PROCESSORS `
-- /v:minimal /consoleloggerparameters:ErrorsOnly 2>&1 `
| Where-Object { $_ -match '(error|warning|FAILED|fatal)' -or $_ -eq '' } `
| Select-Object -Last 50
- name: Smoke test
continue-on-error: true
shell: bash
run: |
BIN="build-msvc/Release"
"$BIN/ace-lm.exe" 2>&1 | head -5
"$BIN/ace-synth.exe" 2>&1 | head -5
"$BIN/ace-understand.exe" 2>&1 | head -5
"$BIN/neural-codec.exe" 2>&1 | head -5
"$BIN/quantize.exe" 2>&1 | head -3
"$BIN/mp3-codec.exe" 2>&1 | head -3
- name: Resolve release tag
id: tag
shell: bash
run: |
if [ "${{ github.event_name }}" = "release" ]; then
echo "value=${{ github.event.release.tag_name }}" >> $GITHUB_OUTPUT
else
echo "value=${{ inputs.release_tag }}" >> $GITHUB_OUTPUT
fi
- name: Package binaries
shell: pwsh
run: |
New-Item -ItemType Directory -Path dist | Out-Null
Copy-Item "build-msvc\Release\*.exe" dist\ -ErrorAction SilentlyContinue
Copy-Item "build-msvc\Release\*.dll" dist\ -ErrorAction SilentlyContinue
Compress-Archive -Path dist\* -DestinationPath "acestep-windows-x64.zip"
- name: Upload to release
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
shell: pwsh
run: |
gh release upload "${{ steps.tag.outputs.value }}" `
"acestep-windows-x64.zip" `
--clobber
engine/.gitignore
+13
View File
@@ -0,0 +1,13 @@
build/
*.wav
*.bf16
tests/*/
checkpoints/
models/*.gguf
adapters/*/
adapters/*.safetensors
__pycache__/
node_modules/
tools/webui/dist/
engine/.gitmodules
+3
View File
@@ -0,0 +1,3 @@
[submodule "ggml"]
path = ggml
url = https://github.com/ServeurpersoCom/ggml.git
engine/CMakeLists.txt
+742
View File
@@ -0,0 +1,742 @@
cmake_minimum_required(VERSION 3.21)
project(acestep-ggml LANGUAGES C CXX)
# CI cache generation: 2 (2026-07-16). This file is hashed into the GitHub
# Actions build-cache key — bump this comment to force cold builds when the
# cached objects themselves are suspect (v1.1.3 stale-cache mixed-ABI crash,
# issues #82/#83). Routine drift is handled by the .built-commit stamp guard
# in release.yml/cache-warm.yml; this is the manual override.
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
# version.h: embed git commit hash into all binaries.
# runs on every build, only rewrites if the hash changed.
set(VERSION_OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/version.h")
add_custom_target(version ALL
COMMAND "${CMAKE_COMMAND}" "-DSRC_DIR=${CMAKE_CURRENT_SOURCE_DIR}" "-DOUTPUT=${VERSION_OUTPUT}"
-P "${CMAKE_CURRENT_SOURCE_DIR}/tools/version.cmake"
BYPRODUCTS "${VERSION_OUTPUT}"
COMMENT "Checking git version"
)
# pthread: required explicitly on older glibc (< 2.34) where libpthread
# is not merged into libc. Modern distros link it implicitly but aarch64
# and older x86_64 toolchains need the explicit dependency.
find_package(Threads REQUIRED)
# Suppress MSVC fopen/sprintf deprecation warnings and Windows.h macro pollution.
# NOMINMAX: prevents Windows.h from defining min/max macros that collide with
# std::min/std::max (causes C2589 errors in solvers/schedulers).
# WIN32_LEAN_AND_MEAN: reduces Windows.h header bloat.
if(MSVC)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS NOMINMAX WIN32_LEAN_AND_MEAN)
endif()
# Static MSVC runtime (/MT) for portable release builds.
# Eliminates the VC++ Redistributable dependency for end users.
# Only enable during release builds: -DHOT_STEP_STATIC_RUNTIME=ON
option(HOT_STEP_STATIC_RUNTIME "Use static MSVC runtime (/MT) for portable builds" OFF)
if(HOT_STEP_STATIC_RUNTIME AND MSVC)
set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
message(STATUS "MSVC runtime: static (/MT)")
endif()
# Put executables and backend .so in the same directory (build root).
# Without this, ggml defaults to bin/ for .so but executables stay in root,
# and ggml_backend_load_all() can't find the backends at runtime.
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
# macOS rpath: make binaries relocatable (portable release support).
# Without this, CMake bakes the absolute build directory into LC_RPATH,
# which breaks on any machine other than the one that built it.
# @executable_path tells dyld to look for dylibs next to the binary.
if(APPLE)
set(CMAKE_INSTALL_RPATH "@executable_path")
set(CMAKE_BUILD_WITH_INSTALL_RPATH TRUE)
set(CMAKE_MACOSX_RPATH TRUE)
endif()
# Linux rpath: make binaries relocatable (portable release support).
# $ORIGIN tells the dynamic linker to search for .so files next to the binary.
if(UNIX AND NOT APPLE)
set(CMAKE_INSTALL_RPATH "$ORIGIN")
set(CMAKE_BUILD_WITH_INSTALL_RPATH TRUE)
endif()
# DiT tensor names can exceed default GGML_MAX_NAME of 64
add_compile_definitions(GGML_MAX_NAME=128)
# Harden: mark fread/fwrite/etc with warn_unused_result on all platforms
if(NOT MSVC)
add_compile_definitions(_FORTIFY_SOURCE=2)
endif()
# CUDA architectures: cover Turing to Blackwell for distributed binaries (CI: CUDA 13.1 / 12.8).
# The CUDA 12.8 build additionally targets legacy Pascal/Volta (see below).
# Users can override with -DCMAKE_CUDA_ARCHITECTURES=native for local builds.
if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
find_package(CUDAToolkit QUIET)
# Base arch list: Turing through Lovelace/Ada
set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-real;89-real;90-real")
if(CUDAToolkit_FOUND)
# Legacy GPUs — Pascal only (sm_60 P100, sm_61 GTX 10xx / P40 / P4).
# CUDA 12.x can still compile real SASS for these; CUDA 13.0 removed
# offline compilation support, so they go only into the cuda12.8 variant.
# Real SASS is required — the 75-virtual PTX above can't JIT backwards
# onto pre-Turing cards.
#
# NOTE: Volta (sm_70) is deliberately EXCLUDED. ggml's mma-based MMQ and
# flash-attention kernels have no device code for sm_70 (they need
# Turing+), so a sm_70 build crashes on Volta. Worse, sm_70 SASS is
# binary-compatible upward to Turing (sm_75), so the driver loads it on
# 75 cards in preference to JIT-ing the compute_75 PTX — which regressed
# all Turing users on the cuda12.8 bundle in v1.1.1 (#63). Pascal (major
# 6) is unaffected since Turing can't load major-6 SASS.
if(CUDAToolkit_VERSION VERSION_LESS "13.0")
list(APPEND CMAKE_CUDA_ARCHITECTURES "60-real;61-real")
endif()
# Blackwell sm_120a: CUDA 12.8+
if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
list(APPEND CMAKE_CUDA_ARCHITECTURES "120a-real")
endif()
# Blackwell sm_121a: CUDA 12.9+ (same as upstream GGML)
if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.9")
list(APPEND CMAKE_CUDA_ARCHITECTURES "121a-real")
endif()
endif()
endif()
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
find_package(hip CONFIG QUIET)
if(hip_FOUND)
message(STATUS "building with AMD ROCm support")
add_compile_definitions(__HIP_PLATFORM_AMD__)
endif()
# Disable flash attention (cuda12-volta variant). ggml's mma flash-attention has
# no device code for Volta (sm_70); with this defined the engine takes the manual
# attention path instead. Pair with -DGGML_CUDA_FORCE_CUBLAS=ON (ggml) which
# likewise replaces the mma MMQ kernels. See engine/src/hot-step-build-flags.h.
option(HOT_STEP_DISABLE_FA "Disable flash attention (Volta / pre-Turing GPUs)" OFF)
if(HOT_STEP_DISABLE_FA)
add_compile_definitions(HOT_STEP_DISABLE_FA)
message(STATUS "[HOT-Step] Flash attention DISABLED (HOT_STEP_DISABLE_FA)")
endif()
# ggml as subdirectory, inherits GGML_CUDA, GGML_METAL, etc. from cmake flags
# CUDA graphs default on: standalone ggml ships them off. Overridable with
# -DGGML_CUDA_GRAPHS=OFF or at runtime with GGML_CUDA_DISABLE_GRAPHS=1.
if(NOT DEFINED GGML_CUDA_GRAPHS)
set(GGML_CUDA_GRAPHS_DEFAULT ON)
endif()
add_subdirectory(ggml)
# cpp-httplib (HTTP server library, used by ace-server)
add_subdirectory(vendor/cpp-httplib)
# Shared compile options and ggml linkage
macro(link_ggml_backends target)
target_include_directories(${target} PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_CURRENT_BINARY_DIR}
)
target_include_directories(${target} SYSTEM PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/ggml/include
)
if(MSVC)
target_compile_options(${target} PRIVATE /W4 /wd4100 /wd4505)
else()
target_compile_options(${target} PRIVATE -Wall -Wextra -Wshadow -Wconversion
-Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
endif()
target_link_libraries(${target} PRIVATE ggml Threads::Threads)
if(TARGET ggml-base)
target_link_libraries(${target} PRIVATE ggml-base)
endif()
foreach(backend cpu blas cuda metal vulkan)
if(TARGET ggml-${backend})
get_target_property(CURRENT_BACKEND_TYPE ggml-${backend} TYPE)
if (CURRENT_BACKEND_TYPE STREQUAL "MODULE_LIBRARY")
# DL mode: backend is loaded at runtime via dlopen,
# skip all link-time deps.
continue()
endif()
target_link_libraries(${target} PRIVATE ggml-${backend})
endif()
endforeach()
add_dependencies(${target} version)
endmacro()
# yyjson (MIT, fast JSON parser/writer)
add_library(yyjson STATIC vendor/yyjson/yyjson.c)
target_include_directories(yyjson PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/yyjson)
if(MSVC)
target_compile_options(yyjson PRIVATE /W0)
else()
target_compile_options(yyjson PRIVATE -w)
endif()
# Lua 5.4 (MIT, embedded scripting for plugin system)
# All .c files except lua.c (standalone interpreter) and luac.c (compiler)
file(GLOB LUA_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/vendor/lua/*.c")
list(FILTER LUA_SOURCES EXCLUDE REGEX "(lua|luac)\\.c$")
add_library(lua54 STATIC ${LUA_SOURCES})
target_include_directories(lua54 PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/lua)
if(MSVC)
target_compile_options(lua54 PRIVATE /W0)
else()
target_compile_options(lua54 PRIVATE -w)
endif()
# ─────────────────────────────────────────────────────────────────────────────
# ONNX Runtime (SuperSep stem separation + VAE-ORT TensorRT acceleration)
# ─────────────────────────────────────────────────────────────────────────────
# Pre-built ORT GPU package. Resolution order:
# 1. ORT_ROOT cmake variable
# 2. ONNXRUNTIME_ROOT environment variable
# 3. Auto-detect from engine/deps/onnxruntime/ (populated by buildall.cmd)
#
# CUDAToolkit detection: the find_package at L73 is conditional on
# CMAKE_CUDA_ARCHITECTURES, so on cached re-configures CUDAToolkit_FOUND
# may be unset. Ensure it's always available for CUDA EP support.
find_package(CUDAToolkit QUIET)
# SuperSep itself is pure GGML and always built. This option now only controls
# whether the ONNX Runtime-dependent paths (StableStep's ONNX backend and the
# ONNX VAE/text-encoder) are compiled — they are the last ORT consumers.
option(HOT_STEP_SUPERSEP "Build the ONNX Runtime paths (StableStep ONNX backend, ONNX VAE)" ON)
set(ORT_ROOT "" CACHE PATH "Path to ONNX Runtime pre-built package")
if(NOT ORT_ROOT AND DEFINED ENV{ONNXRUNTIME_ROOT})
set(ORT_ROOT "$ENV{ONNXRUNTIME_ROOT}")
endif()
# Auto-detect from deps directory (buildall.cmd downloads here)
if(NOT ORT_ROOT)
set(_ORT_DEPS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/deps/onnxruntime")
if(EXISTS "${_ORT_DEPS_DIR}/include/onnxruntime_cxx_api.h")
set(ORT_ROOT "${_ORT_DEPS_DIR}")
message(STATUS "ORT auto-detected at ${ORT_ROOT}")
endif()
endif()
set(SUPERSEP_ENABLED FALSE)
if(HOT_STEP_SUPERSEP AND ORT_ROOT)
if(EXISTS "${ORT_ROOT}/include/onnxruntime_cxx_api.h")
set(SUPERSEP_ENABLED TRUE)
else()
message(WARNING "ORT_ROOT set but onnxruntime_cxx_api.h not found at ${ORT_ROOT}/include")
endif()
endif()
# SuperSep library (STFT + ONNX Runtime for stages 1-4, GGML for the
# BS-Roformer-Leap Xe pair used by SUPERSEP_STABLESTEP)
add_library(supersep STATIC src/supersep.cpp)
target_include_directories(supersep PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}/vendor/pocketfft
${CMAKE_CURRENT_BINARY_DIR}
)
# bs-roformer-ggml.h needs ggml headers/symbols.
target_link_libraries(supersep PUBLIC ggml)
if(MSVC)
target_compile_options(supersep PRIVATE /W4 /wd4100 /wd4505 /wd4244 /wd4267)
else()
target_compile_options(supersep PRIVATE -Wall -Wextra -Wno-unused-parameter -Wno-sign-conversion)
endif()
# SuperSep is now pure GGML (bs-roformer-ggml.h / mdx23c-ggml.h) — no ONNX
# Runtime, so it builds and runs on every backend including Vulkan, Metal and
# plain CPU. The definition is kept because supersep.cpp/.h still guard on it.
target_compile_definitions(supersep PUBLIC HOT_STEP_SUPERSEP)
message(STATUS "SuperSep: ENABLED (native GGML)")
if(CUDAToolkit_FOUND)
target_compile_definitions(supersep PUBLIC GGML_USE_CUDA)
endif()
# NOTE: SUPERSEP_ENABLED below is now a misnomer kept for the ORT plumbing that
# OTHER features still need. SuperSep itself no longer touches ONNX Runtime —
# but sa3-refine.h (StableStep's ONNX backend) and model-store.h's
# vae-ort / cond-enc-ort / text-enc-ort / vae-enc-ort do, so ace-server still
# links ORT and the runtime DLLs are still required by those paths.
# TensorRT Native SDK (DiT TRT acceleration + LoRA refitting)
# ─────────────────────────────────────────────────────────────────────────────
# Platform-specific TRT detection:
# Windows: vendored SDK in engine/deps/tensorrt/ (versioned import libs)
# Linux: system-installed TRT packages via find_path/find_library
# The runtime DLLs/SOs are expected on the library search path at runtime.
set(TRT_ENABLED FALSE)
# --- Windows: vendored SDK in engine/deps/tensorrt/ ---
set(_TRT_DEPS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/deps/tensorrt")
if(WIN32 AND EXISTS "${_TRT_DEPS_DIR}/include/NvInfer.h" AND EXISTS "${_TRT_DEPS_DIR}/lib/nvinfer_10.lib")
set(TRT_ENABLED TRUE)
set(_TRT_INCLUDE_DIR "${_TRT_DEPS_DIR}/include")
set(_TRT_LIB_DIR "${_TRT_DEPS_DIR}/lib")
# Windows vendored SDK ships versioned import libs
set(_TRT_NVINFER_LIB nvinfer_10)
set(_TRT_NVONNXPARSER_LIB nvonnxparser_10)
message(STATUS "[TRT] Found vendored SDK at ${_TRT_DEPS_DIR}")
# --- Linux: system-installed TRT packages ---
elseif(NOT WIN32)
find_path(_TRT_INCLUDE_DIR
NAMES NvInfer.h
PATHS /usr/include/x86_64-linux-gnu /usr/local/include /usr/include
)
find_library(_TRT_NVINFER_LIB
NAMES nvinfer
PATHS /usr/lib/x86_64-linux-gnu /usr/local/lib /usr/lib
)
find_library(_TRT_NVONNXPARSER_LIB
NAMES nvonnxparser
PATHS /usr/lib/x86_64-linux-gnu /usr/local/lib /usr/lib
)
if(_TRT_INCLUDE_DIR AND _TRT_NVINFER_LIB AND _TRT_NVONNXPARSER_LIB)
set(TRT_ENABLED TRUE)
message(STATUS "[TRT] Found system TRT: ${_TRT_NVINFER_LIB}")
else()
message(STATUS "[TRT] Not found (install libnvinfer-dev + libnvonnxparsers-dev to enable)")
endif()
endif()
if(NOT TRT_ENABLED AND WIN32)
message(STATUS "[TRT] Not found (set engine/deps/tensorrt/ to enable DiT TRT)")
endif()
# Core library (shared between binaries)
add_library(acestep-core STATIC
src/request.cpp
src/model-store.cpp
src/pipeline-lm.cpp
src/pipeline-synth.cpp
src/pipeline-synth-ops.cpp
src/pipeline-understand.cpp
)
if(hip_FOUND)
target_link_libraries(acestep-core PUBLIC yyjson lua54 supersep hip::host)
else()
target_link_libraries(acestep-core PUBLIC yyjson lua54 supersep)
endif()
link_ggml_backends(acestep-core)
# ONNX Runtime for the paths that still use it: sa3-refine.h (StableStep's ONNX
# backend) and model-store.h's vae-ort / cond-enc-ort / text-enc-ort /
# vae-enc-ort. This used to arrive transitively from supersep, which is now
# pure GGML — so acestep-core declares it directly.
if(SUPERSEP_ENABLED)
target_compile_definitions(acestep-core PUBLIC HOT_STEP_ORT)
target_include_directories(acestep-core PUBLIC "${ORT_ROOT}/include")
target_link_directories(acestep-core PUBLIC "${ORT_ROOT}/lib")
target_link_libraries(acestep-core PUBLIC onnxruntime)
if(APPLE)
target_link_libraries(acestep-core PUBLIC "-framework CoreML" "-framework Foundation")
endif()
message(STATUS "ONNX Runtime paths: ENABLED (ORT at ${ORT_ROOT})")
else()
message(STATUS "ONNX Runtime paths: DISABLED (StableStep uses its GGML backend)")
endif()
# TRT native linkage for DiT acceleration (shared across platforms)
if(TRT_ENABLED)
target_compile_definitions(acestep-core PUBLIC HOT_STEP_TRT)
target_include_directories(acestep-core PUBLIC "${_TRT_INCLUDE_DIR}")
if(_TRT_LIB_DIR)
target_link_directories(acestep-core PUBLIC "${_TRT_LIB_DIR}")
endif()
target_link_libraries(acestep-core PUBLIC ${_TRT_NVINFER_LIB} ${_TRT_NVONNXPARSER_LIB})
# TRT headers include cuda_runtime_api.h — on Linux g++ needs the CUDA
# include path explicitly (nvcc gets it automatically, but acestep-core
# compiles as plain C++). Windows CUDA Toolkit puts headers on PATH.
if(NOT WIN32)
find_package(CUDAToolkit QUIET)
if(CUDAToolkit_FOUND)
target_include_directories(acestep-core PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
target_link_libraries(acestep-core PUBLIC CUDA::cudart)
message(STATUS "[TRT] CUDA include: ${CUDAToolkit_INCLUDE_DIRS}")
endif()
endif()
message(STATUS "[TRT] DiT TRT acceleration: ENABLED")
endif()
# acestep-core compiles as plain C++ (not nvcc), but pipeline-synth.cpp includes
# <cuda_runtime.h> for VRAM instrumentation whenever GGML_USE_CUDA is defined
# (propagated from ggml). The host compiler needs the CUDA Toolkit include path
# explicitly on EVERY CUDA build — Linux g++ and Windows MSVC (under Ninja) both
# fail to find it otherwise. (The TRT block above also adds this, but only when
# TRT is enabled, so plain CUDA release builds were missing it.)
if(GGML_CUDA)
find_package(CUDAToolkit QUIET)
if(CUDAToolkit_FOUND)
target_include_directories(acestep-core PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
target_link_libraries(acestep-core PUBLIC CUDA::cudart)
message(STATUS "[CUDA] acestep-core include: ${CUDAToolkit_INCLUDE_DIRS}")
endif()
endif()
# ─────────────────────────────────────────────────────────────────────────────
# TRT-LLM Executor (C++ Executor API for LM inference)
# ─────────────────────────────────────────────────────────────────────────────
# Pre-built TRT-LLM SDK (tensorrt_llm.dll + plugin DLL).
# Auto-detect from engine/trtllm-libs/.
# This is SEPARATE from HOT_STEP_TRT (raw NvInfer for DiT). Both can coexist.
#
# STATUS: DISABLED (2026-06-02). Native Windows TRT-LLM is not viable:
# - FMHA/XQA cubin embedding requires GCC inline asm (INCBIN), impossible on MSVC
# - Docker-built engines have Linux platform tags, can't deserialize on Windows
# - TRT version mismatch between Docker (10.14) and Windows SDK (10.16)
# - ONNX-rebuilt engines lack TRT-LLM tensor bindings (kv_cache_block_offsets etc.)
# The code remains intact behind #ifdef HOT_STEP_TRTLLM for future WSL2 or
# cross-platform engine support. To re-enable, set HOT_STEP_TRTLLM_ENABLE=ON.
option(HOT_STEP_TRTLLM_ENABLE "Enable TRT-LLM Executor (currently broken on native Windows)" OFF)
set(TRTLLM_ENABLED FALSE)
set(_TRTLLM_LIBS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/trtllm-libs")
set(_TRTLLM_INC_DIR "${CMAKE_CURRENT_SOURCE_DIR}/trtllm-include")
if(HOT_STEP_TRTLLM_ENABLE AND WIN32
AND EXISTS "${_TRTLLM_LIBS_DIR}/tensorrt_llm.lib"
AND EXISTS "${_TRTLLM_INC_DIR}/tensorrt_llm/executor/executor.h")
set(TRTLLM_ENABLED TRUE)
message(STATUS "[TRT-LLM] Found at ${_TRTLLM_LIBS_DIR}")
target_compile_definitions(acestep-core PUBLIC HOT_STEP_TRTLLM)
target_include_directories(acestep-core PUBLIC "${_TRTLLM_INC_DIR}")
target_link_directories(acestep-core PUBLIC "${_TRTLLM_LIBS_DIR}")
target_link_libraries(acestep-core PUBLIC tensorrt_llm)
# NOTE: nvinfer_plugin_tensorrt_llm is loaded dynamically via LoadLibrary
# in lm-trtllm.h to avoid pulling its dependency chain at process startup.
message(STATUS "[TRT-LLM] LM Executor: ENABLED")
else()
if(HOT_STEP_TRTLLM_ENABLE)
message(STATUS "[TRT-LLM] Not found (set engine/trtllm-libs/ + trtllm-include/ to enable)")
else()
message(STATUS "[TRT-LLM] DISABLED (set -DHOT_STEP_TRTLLM_ENABLE=ON to re-enable)")
endif()
endif()
# ace-synth: full pipeline (text-enc + cond + dit + vae + wav)
add_executable(ace-synth tools/ace-synth.cpp)
target_link_libraries(ace-synth PRIVATE acestep-core)
link_ggml_backends(ace-synth)
# CUDA runtime for TRT DiT path (cudaMalloc, cudaMemcpy, etc.)
if(CUDAToolkit_FOUND)
if(GGML_STATIC)
target_link_libraries(ace-synth PRIVATE CUDA::cudart_static)
else()
target_link_libraries(ace-synth PRIVATE CUDA::cudart)
endif()
endif()
# ace-lm: LLM inference (CoT + audio codes)
add_executable(ace-lm tools/ace-lm.cpp)
target_link_libraries(ace-lm PRIVATE acestep-core)
link_ggml_backends(ace-lm)
# CUDA runtime for TRT LM path (cudaMalloc, cudaMemcpy, etc.)
if(CUDAToolkit_FOUND)
if(GGML_STATIC)
target_link_libraries(ace-lm PRIVATE CUDA::cudart_static)
else()
target_link_libraries(ace-lm PRIVATE CUDA::cudart)
endif()
endif()
# webui: convert tools/webui/public/index.html.gz to a C header for embedding.
# the .gz is committed to git so the C++ build works without npm.
# to update: cd tools/webui && npm install && npm run build, then rebuild ace-server.
set(WEBUI_INPUT "${CMAKE_CURRENT_SOURCE_DIR}/tools/public/index.html.gz")
set(WEBUI_OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/index.html.gz.hpp")
add_custom_command(
OUTPUT "${WEBUI_OUTPUT}"
COMMAND "${CMAKE_COMMAND}" "-DINPUT=${WEBUI_INPUT}" "-DOUTPUT=${WEBUI_OUTPUT}" -P "${CMAKE_CURRENT_SOURCE_DIR}/tools/xxd.cmake"
DEPENDS "${WEBUI_INPUT}"
COMMENT "Embedding webui into index.html.gz.hpp"
)
set_source_files_properties(${WEBUI_OUTPUT} PROPERTIES GENERATED TRUE)
# hot-step-server: HOT-Step HTTP server (LM + synth endpoints + embedded webui)
# NOTE: upstream ace-server.cpp is kept as reference but NOT compiled.
# Our server binary is hot-step-server.cpp with extension layer support.
add_executable(ace-server tools/hot-step-server.cpp ${WEBUI_OUTPUT})
target_include_directories(ace-server PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
if(hip_FOUND)
target_link_libraries(ace-server PRIVATE acestep-core httplib supersep hip::host)
else()
target_link_libraries(ace-server PRIVATE acestep-core httplib supersep)
endif()
link_ggml_backends(ace-server)
# CUDA runtime for GET /vram (cudaMemGetInfo)
find_package(CUDAToolkit QUIET)
if(CUDAToolkit_FOUND)
target_compile_definitions(ace-server PRIVATE GGML_USE_CUDA)
if(GGML_STATIC)
target_link_libraries(ace-server PRIVATE CUDA::cudart_static)
else()
target_link_libraries(ace-server PRIVATE CUDA::cudart)
endif()
# Copy CUDA runtime DLL to build dir (needed for portable release builds —
# end users don't have the CUDA Toolkit, so cudart64_*.dll must ship with the binary)
if(WIN32)
get_target_property(_CUDART_LOC CUDA::cudart IMPORTED_LOCATION)
if(_CUDART_LOC)
add_custom_command(TARGET ace-server POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
"${_CUDART_LOC}"
"$<TARGET_FILE_DIR:ace-server>"
COMMENT "Copying cudart DLL for portable release"
)
endif()
endif()
# Write CUDA version marker next to ace-server (server reads this to
# select the correct runtime DLLs — cuBLAS/cudart/cuDNN for 12 vs 13)
file(GENERATE OUTPUT "$<TARGET_FILE_DIR:ace-server>/.cuda-version"
CONTENT "${CUDAToolkit_VERSION_MAJOR}")
endif()
# Copy ONNX Runtime DLLs next to ace-server at build time
if(SUPERSEP_ENABLED AND WIN32)
set(_ORT_DLL_DIR "${ORT_ROOT}/lib")
foreach(_dll onnxruntime.dll onnxruntime_providers_shared.dll onnxruntime_providers_cuda.dll onnxruntime_providers_tensorrt.dll)
if(EXISTS "${_ORT_DLL_DIR}/${_dll}")
add_custom_command(TARGET ace-server POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
"${_ORT_DLL_DIR}/${_dll}"
"$<TARGET_FILE_DIR:ace-server>/${_dll}"
COMMENT "Copying ${_dll}"
)
endif()
endforeach()
endif()
# Copy ONNX Runtime dylibs next to ace-server at build time (macOS)
if(SUPERSEP_ENABLED AND APPLE)
set(_ORT_LIB_DIR "${ORT_ROOT}/lib")
file(GLOB _ORT_DYLIBS "${_ORT_LIB_DIR}/libonnxruntime*.dylib")
foreach(_dylib ${_ORT_DYLIBS})
add_custom_command(TARGET ace-server POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
"${_dylib}" "$<TARGET_FILE_DIR:ace-server>"
COMMENT "Copying ${_dylib}"
)
endforeach()
endif()
# Copy ONNX Runtime shared libs next to ace-server at build time (Linux)
if(SUPERSEP_ENABLED AND UNIX AND NOT APPLE)
set(_ORT_LIB_DIR "${ORT_ROOT}/lib")
file(GLOB _ORT_SOLIBS "${_ORT_LIB_DIR}/libonnxruntime*.so*")
foreach(_solib ${_ORT_SOLIBS})
add_custom_command(TARGET ace-server POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
"${_solib}" "$<TARGET_FILE_DIR:ace-server>"
COMMENT "Copying ${_solib}"
)
endforeach()
endif()
add_executable(ace-understand tools/ace-understand.cpp)
target_link_libraries(ace-understand PRIVATE acestep-core)
link_ggml_backends(ace-understand)
# quantize: GGUF requantizer (BF16 -> K-quants)
add_executable(quantize tools/quantize.cpp)
link_ggml_backends(quantize)
# neural-codec: Oobleck VAE neural audio codec (encode/decode WAV <-> latent)
add_executable(neural-codec tools/neural-codec.cpp)
link_ggml_backends(neural-codec)
# sa3-ggml-test: parity tests for the StableStep GGML SA3 modules vs goldens
add_executable(sa3-ggml-test tools/sa3-ggml-test.cpp)
target_link_libraries(sa3-ggml-test PRIVATE yyjson)
link_ggml_backends(sa3-ggml-test)
# bs-roformer-test: parity test for the GGML BS-RoFormer (SuperSep StableStep
# separation) vs PyTorch goldens. See scripts/dump_bs_roformer_goldens.py.
add_executable(bs-roformer-test tools/bs-roformer-test.cpp)
link_ggml_backends(bs-roformer-test)
# mdx23c-test: parity test for the GGML MDX23C drum separator (SuperSep stage 3).
add_executable(mdx23c-test tools/mdx23c-test.cpp)
link_ggml_backends(mdx23c-test)
# ace-midi: MuScriptor audio->MIDI transcription (GGML port, in development —
# docs/plans/muscriptor-cpp-port.md). Standalone tool, no acestep-core needed.
add_executable(ace-midi tools/ace-midi.cpp)
link_ggml_backends(ace-midi)
# ace-train: training toolchain (phase 2: dataset tensor preprocessing —
# docs/plans/2026-07-27-preprocess-implementation.md). Standalone tool: every
# engine module it uses is header-only, so no acestep-core link is needed.
add_executable(ace-train tools/ace-train.cpp)
target_link_libraries(ace-train PRIVATE yyjson)
link_ggml_backends(ace-train)
# mp3-codec: MP3 encoder/decoder (standalone, no ggml needed)
# The mp3/ headers are header-only and usable by ace-synth too via #include "mp3/mp3enc.h"
add_executable(mp3-codec tools/mp3-codec.cpp)
target_include_directories(mp3-codec PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_CURRENT_BINARY_DIR}
)
add_dependencies(mp3-codec version)
if(MSVC)
target_compile_options(mp3-codec PRIVATE /W4 /wd4100 /wd4505)
else()
target_compile_options(mp3-codec PRIVATE -Wall -Wextra -Wconversion
-Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
target_link_libraries(mp3-codec PRIVATE m)
endif()
target_link_libraries(mp3-codec PRIVATE Threads::Threads)
# mastering: reference-based audio mastering (standalone, no ggml needed)
# Implements the matchering algorithm using pocketfft for FFT.
add_executable(mastering tools/mastering.cpp)
target_include_directories(mastering PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}/vendor/pocketfft
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_CURRENT_BINARY_DIR}
)
add_dependencies(mastering version)
if(MSVC)
target_compile_options(mastering PRIVATE /W4 /wd4100 /wd4505 /wd4244 /wd4267)
else()
target_compile_options(mastering PRIVATE -Wall -Wextra -Wconversion
-Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
target_link_libraries(mastering PRIVATE m)
endif()
target_link_libraries(mastering PRIVATE Threads::Threads)
# ─────────────────────────────────────────────────────────────────────────────
# VST3 Hosting Library + vst-host tool
# ─────────────────────────────────────────────────────────────────────────────
# We compile only the necessary VST3 SDK source files for hosting (loading
# plugins, processing audio, state save/restore). We do NOT use the SDK's
# own CMakeLists.txt because it requires cmake 3.25+ and pulls in the full
# build system including plugin examples and VSTGUI.
set(VST3SDK_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/vendor/vst3sdk")
# Static library: vst3-hosting
# Platform-specific hosting sources:
# Windows: module_win32.cpp + threadchecker_win32.cpp
# macOS: module_mac.mm (ARC) + threadchecker_mac.mm
# Linux: module_linux.cpp + threadchecker_linux.cpp
if(APPLE)
set(VST3_PLATFORM_SOURCES
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/module_mac.mm
${VST3SDK_ROOT}/public.sdk/source/common/threadchecker_mac.mm
)
# module_mac.mm requires Objective-C ARC
set_source_files_properties(
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/module_mac.mm
${VST3SDK_ROOT}/public.sdk/source/common/threadchecker_mac.mm
PROPERTIES COMPILE_FLAGS "-fobjc-arc"
)
elseif(WIN32)
set(VST3_PLATFORM_SOURCES
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/module_win32.cpp
${VST3SDK_ROOT}/public.sdk/source/common/threadchecker_win32.cpp
)
else()
set(VST3_PLATFORM_SOURCES
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/module_linux.cpp
${VST3SDK_ROOT}/public.sdk/source/common/threadchecker_linux.cpp
)
endif()
add_library(vst3-hosting STATIC
# Base library
${VST3SDK_ROOT}/base/source/baseiids.cpp
${VST3SDK_ROOT}/base/source/fobject.cpp
${VST3SDK_ROOT}/base/source/fdebug.cpp
${VST3SDK_ROOT}/base/source/fstreamer.cpp
${VST3SDK_ROOT}/base/source/fbuffer.cpp
${VST3SDK_ROOT}/base/source/updatehandler.cpp
${VST3SDK_ROOT}/base/source/timer.cpp
${VST3SDK_ROOT}/base/source/fstring.cpp
# Pluginterfaces
${VST3SDK_ROOT}/pluginterfaces/base/funknown.cpp
${VST3SDK_ROOT}/pluginterfaces/base/ustring.cpp
${VST3SDK_ROOT}/pluginterfaces/base/coreiids.cpp
${VST3SDK_ROOT}/pluginterfaces/base/conststringtable.cpp
# Public SDK hosting (cross-platform + platform-specific)
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/module.cpp
${VST3_PLATFORM_SOURCES}
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/hostclasses.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/plugprovider.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/processdata.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/parameterchanges.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/pluginterfacesupport.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/eventlist.cpp
${VST3SDK_ROOT}/public.sdk/source/vst/hosting/connectionproxy.cpp
# Public SDK common (memory streams for state I/O)
${VST3SDK_ROOT}/public.sdk/source/common/memorystream.cpp
${VST3SDK_ROOT}/public.sdk/source/common/commoniids.cpp
${VST3SDK_ROOT}/public.sdk/source/common/pluginview.cpp
${VST3SDK_ROOT}/public.sdk/source/common/commonstringconvert.cpp
# VST interface ID definitions (all DEF_CLASS_IID symbols)
${VST3SDK_ROOT}/public.sdk/source/vst/vstinitiids.cpp
# VST utilities
${VST3SDK_ROOT}/public.sdk/source/vst/utility/stringconvert.cpp
)
target_include_directories(vst3-hosting PUBLIC
${VST3SDK_ROOT}
${VST3SDK_ROOT}/pluginterfaces
${VST3SDK_ROOT}/public.sdk
)
# The SDK defines DEVELOPMENT=1 for debug builds
target_compile_definitions(vst3-hosting PRIVATE
$<$<CONFIG:Debug>:DEVELOPMENT=1>
$<$<CONFIG:Release>:RELEASE=1>
)
if(MSVC)
# Suppress noisy SDK warnings
target_compile_options(vst3-hosting PRIVATE /W0)
else()
target_compile_options(vst3-hosting PRIVATE -w)
endif()
# macOS: VST3 hosting needs Cocoa + CoreFoundation frameworks
if(APPLE)
find_library(COCOA_FRAMEWORK Cocoa)
find_library(COREFOUNDATION_FRAMEWORK CoreFoundation)
target_link_libraries(vst3-hosting PRIVATE
${COCOA_FRAMEWORK}
${COREFOUNDATION_FRAMEWORK}
)
endif()
# vst-host: standalone VST3 host tool (GUI + offline processing + chain)
add_executable(vst-host tools/vst-host.cpp)
target_include_directories(vst-host PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/src
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_CURRENT_BINARY_DIR}
${VST3SDK_ROOT}
)
target_link_libraries(vst-host PRIVATE vst3-hosting yyjson Threads::Threads)
if(MSVC)
target_compile_options(vst-host PRIVATE /W4 /wd4100 /wd4505)
target_compile_definitions(vst-host PRIVATE _CRT_SECURE_NO_WARNINGS NOMINMAX WIN32_LEAN_AND_MEAN)
else()
target_compile_options(vst-host PRIVATE -Wall -Wextra -Wconversion
-Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
endif()
add_dependencies(vst-host version)
# macOS: vst-host also needs Apple frameworks
if(APPLE)
target_link_libraries(vst-host PRIVATE ${COCOA_FRAMEWORK} ${COREFOUNDATION_FRAMEWORK})
endif()
engine/LICENSE
+21
View File
@@ -0,0 +1,21 @@
MIT License
Copyright (c) 2023-2026 The acestep.cpp authors
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
engine/README.md
+185
View File
@@ -0,0 +1,185 @@
# acestep.cpp
Local AI music generation server with browser UI, powered by GGML.
Describe a song, get stereo 48kHz audio. Runs on CPU, CUDA, Metal, Vulkan.
<img width="1704" height="773" alt="Light" src="https://github.com/user-attachments/assets/aeda150a-46a2-4542-a2d6-57d238a7bbb4" />
<img width="1705" height="771" alt="Dark" src="https://github.com/user-attachments/assets/4941cec9-b6ff-4e09-8905-bdc3ee06d222" />
## Download models
Grab one GGUF of each type from Hugging Face and drop them in the `models/` folder:
https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/tree/main
| Type | Pick one | Size |
|------|----------|------|
| LM | acestep-5Hz-lm-4B-Q8_0.gguf | 4.2 GB |
| Text encoder | Qwen3-Embedding-0.6B-Q8_0.gguf | 748 MB |
| DiT | acestep-v15-turbo-Q8_0.gguf | 2.4 GB |
| VAE | vae-BF16.gguf (always this one) | 322 MB |
Three LM sizes available: 0.6B (fast), 1.7B, 4B (best quality).
Multiple DiT variants: turbo (8 steps), sft (50 steps, higher quality), base, shift1, shift3, continuous.
Alternative: `./models.sh` downloads the default set automatically (needs `pip install hf`).
## Build
```
git clone --recurse-submodules https://github.com/ServeurpersoCom/acestep.cpp.git
cd acestep.cpp
```
### Windows
Pre-built binaries (until CI is set up): https://www.serveurperso.com/temp/acestep.cpp-win64/
To build from source, install
[Visual C++ Build Tools](https://visualstudio.microsoft.com/visual-cpp-build-tools/)
(select "Desktop development with C++" workload) and optionally the
[CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) and/or the
[Vulkan SDK](https://vulkan.lunarg.com/sdk/home).
```cmd
buildcuda.cmd # NVIDIA GPU
buildvulkan.cmd # AMD/Intel GPU (Vulkan)
buildall.cmd # all backends (CUDA + Vulkan + CPU, runtime loading)
```
### Linux / macOS
```bash
./buildcuda.sh # NVIDIA GPU
./buildvulkan.sh # AMD/Intel GPU (Vulkan)
./buildcpu.sh # CPU only (with BLAS)
./buildall.sh # all backends (CUDA + Vulkan + CPU, runtime loading)
```
macOS auto-enables Metal and Accelerate BLAS with any of the above.
## Run
```bash
./server.sh # Linux / macOS
server.cmd # Windows
```
Open http://localhost:8085 in your browser. The WebUI handles everything:
write a caption, set lyrics and metadata, generate, play, and download tracks.
Models are loaded on first request (zero GPU at startup) and swapped
automatically when you pick a different one in the UI.
## Adapters
Drop adapters in the `adapters/` folder and restart the server.
Supports LoRA today in two flavours: PEFT directories (with
`adapter_model.safetensors` + `adapter_config.json`) and ComfyUI single
`.safetensors` files. Select the active adapter from the WebUI.
## Server options
```
--models <dir> Model directory (required)
--adapters <dir> Adapter directory (LoRA today, LoKr soon)
--host <addr> Listen address (default: 127.0.0.1)
--port <N> Listen port (default: 8080)
--max-batch <N> LM batch limit 1-9 (default: 1)
--vae-chunk <N> VAE tile size (default: 256, lower = less VRAM)
--mp3-bitrate <N> MP3 kbps (default: 128)
```
<details>
<summary>API endpoints</summary>
The server exposes three POST endpoints and two GET endpoints:
**POST /lm** - Generate lyrics and audio codes from a caption. Returns JSON.
**POST /synth** - Render audio codes into MP3 or WAV (`?wav=1`).
Accepts JSON or multipart (with source audio for cover/repaint modes).
**POST /understand** - Reverse pipeline: audio in, metadata + lyrics + codes out.
Accepts multipart (audio file) or JSON (codes-only).
**GET /health** - Returns `{"status":"ok"}`.
**GET /props** - Available models, server config, default parameters.
See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full API reference
and AceRequest JSON specification.
</details>
<details>
<summary>CLI tools (advanced)</summary>
For scripting without the server, `ace-lm` and `ace-synth` work as a pipe:
```bash
# LM generates lyrics + codes
./build/ace-lm \
--request /tmp/request.json \
--lm models/acestep-5Hz-lm-4B-Q8_0.gguf
# DiT + VAE render to audio
./build/ace-synth \
--request /tmp/request0.json \
--embedding models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit models/acestep-v15-turbo-Q8_0.gguf \
--vae models/vae-BF16.gguf
```
See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full JSON reference,
task types, batching, and understand pipeline.
</details>
## Technical documentation
[docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) covers the complete AceRequest JSON
reference, all task types (text2music, cover, repaint, lego, extract, complete),
FSM constrained decoding, custom GGML operators, quantization, and architecture
internals.
## Community
### ACE-Step official documentation
- [A Musician's Guide](https://github.com/ace-step/ACE-Step-1.5/discussions/235) - non-technical guide for music makers
- [Tutorial](https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/Tutorial.md) - design philosophy, model architecture, input control, inference hyperparameters
### Third-party UIs for acestep.cpp
- [acestep-cpp-ui](https://github.com/audiohacking/acestep-cpp-ui)
- [acestep.cpp-simple-GUI](https://github.com/Nurb4000/acestep.cpp-simple-GUI)
- [aceradio](https://github.com/IMbackK/aceradio)
## Samples
https://github.com/user-attachments/assets/9a50c1f4-9ec0-474a-bd14-e8c6b00622a1
https://github.com/user-attachments/assets/fb606249-0269-4153-b651-bf78e05baf22
https://github.com/user-attachments/assets/e0580468-5e33-4a1f-a0f4-b914e4b9a8c2
https://github.com/user-attachments/assets/292a31f1-f97e-4060-9207-ed8364d9a794
https://github.com/user-attachments/assets/34b1b781-a5bc-46c4-90a6-615a10bc2c6a
## Acknowledgements
Independent C++ implementation based on
[ACE-Step 1.5](https://github.com/ace-step/ACE-Step-1.5) by ACE Studio and StepFun.
All model weights are theirs, this is just a native backend.
```bibtex
@misc{gong2026acestep,
title={ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation},
author={Junmin Gong, Yulin Song, Wenxiao Zhao, Sen Wang, Shengyuan Xu, Jing Guo},
howpublished={\url{https://github.com/ace-step/ACE-Step-1.5}},
year={2026},
note={GitHub repository}
}
```
engine/UPSTREAM_SYNC
+4
View File
@@ -0,0 +1,4 @@
UPSTREAM_REPO=acestep.cpp
UPSTREAM_COMMIT=9d38f00267d71f8a1d7a6d3e325476472ce09179
SYNC_DATE=2026-07-14
SYNC_NOTES=LM perf overhaul (graph arenas, static decode graph replay, set_rows KV, padded attn windows), snake autofuse + fused CUDA/Vulkan kernels via ggml bump to ServeurpersoCom b677b63c, K/V F16 cast before flash_attn_ext, dead full-attention padding mask removed (incl. hot-step-sampler.h), GGML_CUDA_GRAPHS default ON, cpp-httplib 0.44, WAV PCM24 fixes, locale-immune float parser. Upstream src/solvers/ NOT copied (fork has own).
engine/build-mac.sh
Executable
+79
View File
@@ -0,0 +1,79 @@
#!/bin/bash
# HOT-Step CPP — macOS build script (Apple Silicon + Metal)
#
# Prerequisites:
# - Xcode (full install, not just command line tools — needed for Metal)
# - CMake 3.20+ (brew install cmake)
# - Ninja (brew install ninja) — optional but faster
#
# Usage:
# ./build-mac.sh # Release build with Metal
# ./build-mac.sh Debug # Debug build
#
# ONNX Runtime (SuperSep):
# To enable stem separation, download the macOS ONNX Runtime package:
# brew install onnxruntime
# Or download from: https://github.com/microsoft/onnxruntime/releases
# Place in engine/deps/onnxruntime-osx-arm64/ (or set ORT_ROOT)
set -e
cd "$(dirname "$0")"
BUILD_TYPE="${1:-Release}"
BUILD_DIR="build"
echo "╔══════════════════════════════════════════╗"
echo "║ HOT-Step CPP — macOS Build ║"
echo "║ Apple Silicon + Metal ║"
echo "╚══════════════════════════════════════════╝"
echo ""
echo "Build type: ${BUILD_TYPE}"
mkdir -p "${BUILD_DIR}"
cd "${BUILD_DIR}"
# Detect Ninja
GENERATOR=""
if command -v ninja &>/dev/null; then
GENERATOR="-G Ninja"
echo "Generator: Ninja"
else
echo "Generator: Unix Makefiles (install ninja for faster builds: brew install ninja)"
fi
# Auto-detect ONNX Runtime
ORT_FLAG=""
if [ -n "${ORT_ROOT}" ]; then
echo "ONNX Runtime: ${ORT_ROOT}"
elif [ -d "../deps/onnxruntime-osx-arm64" ]; then
export ORT_ROOT="../deps/onnxruntime-osx-arm64"
echo "ONNX Runtime: auto-detected at ${ORT_ROOT}"
else
echo "ONNX Runtime: not found (SuperSep will be disabled)"
echo " → To enable: brew install onnxruntime, or download manually"
fi
cmake .. \
${GENERATOR} \
-DCMAKE_BUILD_TYPE="${BUILD_TYPE}" \
-DGGML_METAL=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_BACKEND_DL=OFF
# Build using all available cores
NCPU=$(sysctl -n hw.ncpu 2>/dev/null || echo 4)
cmake --build . --config "${BUILD_TYPE}" -j "${NCPU}"
echo ""
echo "═══════════════════════════════════════════"
echo " Build complete! Binaries in: $(pwd)/"
echo ""
echo " Targets built:"
[ -f ace-server ] && echo " ✓ ace-server"
[ -f mastering ] && echo " ✓ mastering"
[ -f mp3-codec ] && echo " ✓ mp3-codec"
[ -f vst-host ] && echo " ✓ vst-host"
echo ""
echo " Next: cd .. && ./launch.sh"
echo "═══════════════════════════════════════════"
engine/build.cmd
+151
View File
@@ -0,0 +1,151 @@
@echo off
REM HOT-Step engine build (CUDA, native arch only)
REM Compiles ONLY for the local GPU — fast dev builds.
REM
REM Automatically finds Visual Studio / Build Tools via vswhere.
REM Automatically downloads ONNX Runtime GPU SDK for SuperSep support.
REM --- Find vcvars64.bat dynamically ---
REM vswhere ships with VS 2017+ and VS BuildTools.
REM
REM IMPORTANT: %ProgramFiles(x86)% contains parentheses which break
REM batch for-loop parsing. We write the vswhere output to a temp file
REM and read from that instead.
set "VSWHERE=%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"
if not exist "%VSWHERE%" (
echo ERROR: vswhere.exe not found. Is Visual Studio or Build Tools installed?
exit /b 1
)
set "VCVARS_TMP=%TEMP%\vcvars_path.txt"
"%VSWHERE%" -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -find "VC\Auxiliary\Build\vcvars64.bat" > "%VCVARS_TMP%" 2>nul
set "VCVARS="
for /f "usebackq tokens=*" %%i in ("%VCVARS_TMP%") do set "VCVARS=%%i"
del "%VCVARS_TMP%" 2>nul
if not defined VCVARS (
echo ERROR: Could not find vcvars64.bat via vswhere.
echo Install the "Desktop development with C++" workload.
exit /b 1
)
REM Skip vcvars if already sourced (prevents PATH overflow on repeated runs)
if defined VSCMD_VER (
echo Using cached VS environment ^(VSCMD_VER=%VSCMD_VER%^)
) else (
echo Using: %VCVARS%
call "%VCVARS%"
)
REM ── ONNX Runtime GPU SDK (for SuperSep stem separation) ────────────
REM Auto-downloads from Microsoft's GitHub Releases if not present.
REM Users can skip this by setting ONNXRUNTIME_ROOT env var.
set "ORT_VERSION=1.25.1"
set "ORT_DIR=%~dp0deps\onnxruntime"
set "ORT_MARKER=%ORT_DIR%\include\onnxruntime_cxx_api.h"
if defined ONNXRUNTIME_ROOT (
echo [ORT] Using ONNXRUNTIME_ROOT=%ONNXRUNTIME_ROOT%
goto :cudnn
)
if exist "%ORT_MARKER%" (
echo [ORT] Found at %ORT_DIR%
goto :cudnn
)
echo.
echo [ORT] ONNX Runtime GPU SDK not found. Downloading v%ORT_VERSION%...
echo [ORT] (one-time download for SuperSep stem separation)
echo.
set "ORT_ZIP=%TEMP%\onnxruntime-win-x64-gpu-%ORT_VERSION%.zip"
set "ORT_URL=https://github.com/microsoft/onnxruntime/releases/download/v%ORT_VERSION%/onnxruntime-win-x64-gpu-%ORT_VERSION%.zip"
echo [ORT] Downloading from %ORT_URL%
curl -L -o "%ORT_ZIP%" "%ORT_URL%"
if errorlevel 1 (
echo [ORT] WARNING: Download failed. Building without SuperSep support.
goto :build
)
echo [ORT] Extracting...
mkdir "%~dp0deps" 2>nul
powershell -NoProfile -Command "Expand-Archive -Path '%ORT_ZIP%' -DestinationPath '%~dp0deps' -Force"
if errorlevel 1 (
echo [ORT] WARNING: Extraction failed. Building without SuperSep support.
goto :build
)
REM Rename extracted folder (it has version in the name)
if exist "%~dp0deps\onnxruntime-win-x64-gpu-%ORT_VERSION%" (
ren "%~dp0deps\onnxruntime-win-x64-gpu-%ORT_VERSION%" onnxruntime
)
del "%ORT_ZIP%" 2>nul
if exist "%ORT_MARKER%" (
echo [ORT] Successfully installed to %ORT_DIR%
) else (
echo [ORT] WARNING: Installation may have failed. Check %ORT_DIR%
)
REM ── cuDNN 9 (required for ONNX Runtime CUDA EP) ────────────────────
:cudnn
REM ORT GPU needs cudnn64_9.dll which isn't bundled. We get it from
REM the nvidia-cudnn-cu12 pip package (no NVIDIA login required).
REM Only the runtime DLLs are needed — copied next to the exe.
set "CUDNN_MARKER=%~dp0build\Release\cudnn64_9.dll"
if exist "%CUDNN_MARKER%" (
echo [cuDNN] Found cudnn64_9.dll
goto :build
)
echo.
echo [cuDNN] cudnn64_9.dll not found. Installing via pip...
echo [cuDNN] (one-time download for CUDA-accelerated SuperSep)
echo.
python -m pip install --quiet nvidia-cudnn-cu12 2>nul
if errorlevel 1 (
echo [cuDNN] WARNING: pip install failed. CUDA EP will be disabled.
echo [cuDNN] To fix: pip install nvidia-cudnn-cu12
goto :build
)
REM Find the installed DLLs and copy them to build/Release
for /f "tokens=*" %%d in ('python -c "import nvidia.cudnn; import os; print(os.path.join(nvidia.cudnn.__path__[0], 'bin'))" 2^>nul') do (
if exist "%%d\cudnn64_9.dll" (
echo [cuDNN] Copying DLLs from %%d
mkdir "%~dp0build\Release" 2>nul
copy /y "%%d\cudnn*.dll" "%~dp0build\Release\" >nul 2>nul
echo [cuDNN] Done
) else (
echo [cuDNN] WARNING: Could not find cudnn64_9.dll in pip package
echo [cuDNN] path checked: %%d
)
)
:build
cd /d "%~dp0"
mkdir build 2>nul
cd build
REM Only run cmake configure if not yet configured (avoids invalidating incremental builds)
if not exist "CMakeCache.txt" (
REM HOT_STEP_CMAKE_FLAGS can be set by update.bat for auto-detected backends.
REM When unset, defaults to CUDA-only (native dev build).
if defined HOT_STEP_CMAKE_FLAGS (
cmake .. %HOT_STEP_CMAKE_FLAGS% -DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON
) else (
cmake .. -DGGML_CUDA=ON -DGGML_CUDA_GRAPHS=ON -DCMAKE_CUDA_ARCHITECTURES="75;80;86;89;90;120a" -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON
)
)
cmake --build . --config Release -j %NUMBER_OF_PROCESSORS%
cd ..
engine/buildall.cmd
+100
View File
@@ -0,0 +1,100 @@
@echo off
REM HOT-Step engine build (all backends: CPU variants + CUDA + Vulkan)
REM Uses vswhere to find any Visual Studio edition automatically.
REM Automatically downloads ONNX Runtime GPU SDK for SuperSep support.
REM Skip vcvars if already sourced (prevents PATH overflow on repeated runs)
if defined VSCMD_VER goto :deps
set "VSWHERE=%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"
if not exist "%VSWHERE%" (
echo ERROR: vswhere.exe not found. Is Visual Studio or Build Tools installed?
exit /b 1
)
set "VCVARS_TMP=%TEMP%\vcvars_path.txt"
"%VSWHERE%" -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -find "VC\Auxiliary\Build\vcvars64.bat" > "%VCVARS_TMP%" 2>nul
set "VCVARS="
for /f "usebackq tokens=*" %%i in ("%VCVARS_TMP%") do set "VCVARS=%%i"
del "%VCVARS_TMP%" 2>nul
if not defined VCVARS (
echo ERROR: Could not find vcvars64.bat via vswhere.
echo Install the "Desktop development with C++" workload.
exit /b 1
)
echo Using: %VCVARS%
call "%VCVARS%"
:deps
REM ── ONNX Runtime GPU SDK (for SuperSep stem separation) ────────────
REM Auto-downloads from Microsoft's GitHub Releases if not present.
REM Users can skip this by setting ORT_ROOT or ONNXRUNTIME_ROOT env var.
set "ORT_VERSION=1.25.1"
set "ORT_DIR=%~dp0deps\onnxruntime"
set "ORT_MARKER=%ORT_DIR%\include\onnxruntime_cxx_api.h"
if defined ONNXRUNTIME_ROOT (
echo [ORT] Using ONNXRUNTIME_ROOT=%ONNXRUNTIME_ROOT%
goto :build
)
if exist "%ORT_MARKER%" (
echo [ORT] Found at %ORT_DIR%
goto :build
)
echo.
echo ══════════════════════════════════════════════════════════════════
echo ONNX Runtime GPU SDK not found. Downloading v%ORT_VERSION%...
echo This is needed for SuperSep stem separation (one-time download).
echo ══════════════════════════════════════════════════════════════════
echo.
set "ORT_ZIP=%TEMP%\onnxruntime-win-x64-gpu-%ORT_VERSION%.zip"
set "ORT_URL=https://github.com/microsoft/onnxruntime/releases/download/v%ORT_VERSION%/onnxruntime-win-x64-gpu-%ORT_VERSION%.zip"
echo [ORT] Downloading from %ORT_URL%
curl -L -o "%ORT_ZIP%" "%ORT_URL%"
if errorlevel 1 (
echo [ORT] WARNING: Download failed. Building without SuperSep support.
echo [ORT] You can manually download and extract to: %ORT_DIR%
goto :build
)
echo [ORT] Extracting...
mkdir "%~dp0deps" 2>nul
powershell -NoProfile -Command "Expand-Archive -Path '%ORT_ZIP%' -DestinationPath '%~dp0deps' -Force"
if errorlevel 1 (
echo [ORT] WARNING: Extraction failed. Building without SuperSep support.
goto :build
)
REM Rename extracted folder (it has version in the name)
if exist "%~dp0deps\onnxruntime-win-x64-gpu-%ORT_VERSION%" (
ren "%~dp0deps\onnxruntime-win-x64-gpu-%ORT_VERSION%" onnxruntime
)
REM Clean up zip
del "%ORT_ZIP%" 2>nul
if exist "%ORT_MARKER%" (
echo [ORT] Successfully installed to %ORT_DIR%
) else (
echo [ORT] WARNING: Installation may have failed. Check %ORT_DIR%
)
:build
cd /d "%~dp0"
rem rd /s /q build 2>nul
mkdir build 2>nul
cd build
REM Build with ORT auto-detection from deps directory
cmake .. -DGGML_CPU_ALL_VARIANTS=ON -DGGML_CUDA=ON -DGGML_VULKAN=ON -DGGML_BACKEND_DL=ON %RELEASE_CMAKE_EXTRA%
cmake --build . --config Release -j %NUMBER_OF_PROCESSORS%
cd ..
engine/buildall.sh
+10
View File
@@ -0,0 +1,10 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
export PATH=/usr/local/cuda/bin:$PATH
cmake .. -DGGML_CPU_ALL_VARIANTS=ON -DGGML_CUDA=ON -DGGML_VULKAN=ON -DGGML_BACKEND_DL=ON
cmake --build . --config Release -j "$(nproc)"
engine/buildcpu.sh
+8
View File
@@ -0,0 +1,8 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
cmake .. -DGGML_BLAS=ON
cmake --build . --config Release -j "$(nproc)"
engine/buildcuda.cmd
+39
View File
@@ -0,0 +1,39 @@
@echo off
REM HOT-Step engine build (CUDA only)
REM Uses vswhere to find any Visual Studio edition automatically.
REM Skip vcvars if already sourced (prevents PATH overflow on repeated runs)
if defined VSCMD_VER goto :build
set "VSWHERE=%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"
if not exist "%VSWHERE%" (
echo ERROR: vswhere.exe not found. Is Visual Studio or Build Tools installed?
exit /b 1
)
set "VCVARS_TMP=%TEMP%\vcvars_path.txt"
"%VSWHERE%" -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -find "VC\Auxiliary\Build\vcvars64.bat" > "%VCVARS_TMP%" 2>nul
set "VCVARS="
for /f "usebackq tokens=*" %%i in ("%VCVARS_TMP%") do set "VCVARS=%%i"
del "%VCVARS_TMP%" 2>nul
if not defined VCVARS (
echo ERROR: Could not find vcvars64.bat via vswhere.
echo Install the "Desktop development with C++" workload.
exit /b 1
)
echo Using: %VCVARS%
call "%VCVARS%"
:build
cd /d "%~dp0"
rem rd /s /q build 2>nul
mkdir build 2>nul
cd build
cmake .. -DGGML_CUDA=ON
cmake --build . --config Release -j %NUMBER_OF_PROCESSORS%
cd ..
engine/buildcuda.sh
+8
View File
@@ -0,0 +1,8 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
cmake --build . --config Release -j "$(nproc)"
engine/buildhip.sh
Executable
+16
View File
@@ -0,0 +1,16 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
if [ -z "${GFX_NAME}" ]; then
echo "Warn: Couldn't detect AMD GPU for HIP! Using fallback value (gfx1030).";
GFX_NAME="gfx1030";
else
echo "Building for GPU arch: ${GFX_NAME}";
fi
cmake .. -DGGML_HIP=ON -DGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_TYPE=Release
cmake --build . --config Release -j `nproc --ignore=1`
engine/buildtermux.sh
+8
View File
@@ -0,0 +1,8 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
cmake .. -DGGML_BLAS=ON -DBLAS_INCLUDE_DIRS=$PREFIX/include/openblas
cmake --build . --config Release -j "$(nproc)"
engine/buildvulkan.cmd
+39
View File
@@ -0,0 +1,39 @@
@echo off
REM HOT-Step engine build (Vulkan only)
REM Uses vswhere to find any Visual Studio edition automatically.
REM Skip vcvars if already sourced (prevents PATH overflow on repeated runs)
if defined VSCMD_VER goto :build
set "VSWHERE=%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"
if not exist "%VSWHERE%" (
echo ERROR: vswhere.exe not found. Is Visual Studio or Build Tools installed?
exit /b 1
)
set "VCVARS_TMP=%TEMP%\vcvars_path.txt"
"%VSWHERE%" -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -find "VC\Auxiliary\Build\vcvars64.bat" > "%VCVARS_TMP%" 2>nul
set "VCVARS="
for /f "usebackq tokens=*" %%i in ("%VCVARS_TMP%") do set "VCVARS=%%i"
del "%VCVARS_TMP%" 2>nul
if not defined VCVARS (
echo ERROR: Could not find vcvars64.bat via vswhere.
echo Install the "Desktop development with C++" workload.
exit /b 1
)
echo Using: %VCVARS%
call "%VCVARS%"
:build
cd /d "%~dp0"
rem rd /s /q build 2>nul
mkdir build 2>nul
cd build
cmake .. -DGGML_VULKAN=ON
cmake --build . --config Release -j %NUMBER_OF_PROCESSORS%
cd ..
engine/buildvulkan.sh
+8
View File
@@ -0,0 +1,8 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
cmake .. -DGGML_VULKAN=ON
cmake --build . --config Release -j "$(nproc)"
engine/buildwebui.sh
+9
View File
@@ -0,0 +1,9 @@
#!/bin/bash
cd tools/webui
rm -rf node_modules
npm install
npm run format
npm run lint
npm run check
npm run build
engine/checkpoints.sh
+63
View File
@@ -0,0 +1,63 @@
#!/bin/bash
# Download ACE-Step checkpoints from HuggingFace
# Usage: ./checkpoints.sh [--all]
# default: Qwen3-Embedding-0.6B + acestep-5Hz-lm-4B + acestep-v15-turbo + vae
# --all: + all LM variants + all DiT variants (incl. XL 4B) from ACE-Step registry
set -eu
DIR="checkpoints"
mkdir -p "$DIR"
HF="hf download --quiet"
MAIN="ACE-Step/Ace-Step1.5"
dl_main() {
local name="$1"
local target="$DIR/$name"
if [ -d "$target" ] && [ "$(ls "$target"/*.safetensors "$target"/*.bin 2>/dev/null | wc -l)" -gt 0 ]; then
echo "[OK] $name"
return
fi
echo "[Download] $name <- $MAIN"
$HF "$MAIN" --include "$name/*" --local-dir "$DIR"
}
dl_repo() {
local name="$1" repo="$2"
local target="$DIR/$name"
if [ -d "$target" ] && [ "$(ls "$target"/*.safetensors "$target"/*.bin 2>/dev/null | wc -l)" -gt 0 ]; then
echo "[OK] $name"
return
fi
echo "[Download] $name <- $repo"
$HF "$repo" --local-dir "$target"
}
# Core (required)
dl_main "Qwen3-Embedding-0.6B"
dl_repo "acestep-5Hz-lm-4B" "ACE-Step/acestep-5Hz-lm-4B"
dl_main "acestep-v15-turbo"
dl_main "vae"
# Every model from ACE-Step registry
if [ "${1:-}" = "--all" ]; then
# LM variants (from main repo)
dl_main "acestep-5Hz-lm-1.7B"
# LM variants (separate repos)
dl_repo "acestep-5Hz-lm-0.6B" "ACE-Step/acestep-5Hz-lm-0.6B"
# DiT variants (separate repos)
dl_repo "acestep-v15-turbo-shift3" "ACE-Step/acestep-v15-turbo-shift3"
dl_repo "acestep-v15-turbo-shift1" "ACE-Step/acestep-v15-turbo-shift1"
dl_repo "acestep-v15-turbo-continuous" "ACE-Step/acestep-v15-turbo-continuous"
dl_repo "acestep-v15-sft" "ACE-Step/acestep-v15-sft"
dl_repo "acestep-v15-base" "ACE-Step/acestep-v15-base"
# XL (4B DiT) variants
dl_repo "acestep-v15-xl-turbo" "ACE-Step/acestep-v15-xl-turbo"
dl_repo "acestep-v15-xl-sft" "ACE-Step/acestep-v15-xl-sft"
dl_repo "acestep-v15-xl-base" "ACE-Step/acestep-v15-xl-base"
fi
find "$DIR" -name '.cache' -type d -exec rm -rf {} + 2>/dev/null
echo "[Done] Checkpoints ready in $DIR"
echo "[Done] Run: python3 convert.py"
engine/convert-comfy-int8.py
+258
View File
@@ -0,0 +1,258 @@
#!/usr/bin/env python3
# convert-comfy-int8.py: ComfyUI int8 (comfy_quant) DiT safetensors -> Q8_0 GGUF
#
# ComfyUI int8 checkpoints (convert_to_quant output) store each quantized
# linear as:
# <base>.weight I8 (out, in) -- quantized weight
# <base>.weight_scale F32 scalar or (out,) -- per-tensor / per-row scale
# <base>.comfy_quant U8 json blob -- {"format": "int8_tensorwise",
# "convrot": true?,
# "convrot_groupsize": G?, ...}
# Dequant is w = int8 * scale. Both per-tensor and per-row int8 grids are
# exactly representable in GGUF Q8_0 (fp16 scale + 32x int8 per block; every
# block scale in row r = that row's scale), so we repack bit-faithfully
# instead of dequantizing + requantizing.
#
# ConvRot (--convrot): weights are stored PRE-ROTATED (W' = W @ H_block^T per
# input-dim group, H = regular Hadamard, power-of-4 group size). Handling:
# - decoder.* weights keep the rotation; the layer is recorded in the GGUF
# KV "acestep.convrot_map" ("name:group;...") and the engine applies the
# matching activation rotation at inference (see engine/src/dit-graph.h).
# - all other components (encoder/tokenizer/detokenizer) are dequantized and
# UNROTATED offline to BF16 — they run once per generation, so keeping
# them int8 isn't worth wiring rotation through their graph builders.
#
# Non-quantized tensors follow convert.py conventions (F32 -> BF16 truncate).
# The DiT GGUF must also carry silence_latent + the acestep.* config KVs,
# which ComfyUI files lack -- both are copied from a donor GGUF of the same
# architecture (any convert.py-produced acestep-v15-*.gguf that matches).
#
# Usage:
# python convert-comfy-int8.py <comfy.safetensors> <donor.gguf> <out.gguf> [--name NAME]
import argparse
import json
import os
import sys
import numpy as np
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, SCRIPT_DIR)
import gguf # noqa: E402
import convert # noqa: E402 (engine/convert.py: read_sf_header, add_metadata)
Q8_0 = gguf.GGMLQuantizationType.Q8_0
BF16 = gguf.GGMLQuantizationType.BF16
QK8_0 = 32 # ggml Q8_0 block size
def log(msg):
print("[COMFY-INT8] %s" % msg, file=sys.stderr, flush=True)
def f32_to_bf16(arr_f32):
"""Truncate F32 to BF16 (as uint16), matching convert.py."""
a = np.ascontiguousarray(arr_f32, dtype=np.float32)
return (a.view(np.uint32) >> 16).astype(np.uint16)
def pack_q8_0(qs, row_scales):
"""Pack int8 weights (rows, cols) + per-row f32 scales into raw Q8_0 blocks.
Q8_0 block layout: 2-byte fp16 scale followed by 32 int8 values.
Returns a uint8 array of shape (rows, cols//32 * 34).
"""
rows, cols = qs.shape
nb = cols // QK8_0
d = np.repeat(row_scales.astype(np.float16).reshape(rows, 1, 1), nb, axis=1)
blocks = np.concatenate([d.view(np.uint8), qs.reshape(rows, nb, QK8_0).view(np.uint8)], axis=2)
return np.ascontiguousarray(blocks.reshape(rows, nb * (QK8_0 + 2)))
def build_hadamard(size):
"""Regular Hadamard (H4 Kronecker powers, normalized). Power-of-4 sizes.
Mirrors convert_to_quant utils/convrot.py build_hadamard(); symmetric.
"""
import math
if size < 4 or (size & (size - 1)) != 0 or (math.log(size, 4) % 1 != 0):
raise ValueError("unsupported Hadamard size %d (power of 4 only)" % size)
H4 = np.array([[1, 1, 1, -1], [1, 1, -1, 1], [1, -1, 1, 1], [-1, 1, 1, 1]], dtype=np.float64)
H = H4
while H.shape[0] < size:
H = np.kron(H, H4)
return H / math.sqrt(size)
def unrotate_weight(w, group_size):
"""Undo the offline ConvRot rotation: W = W_rot @ H per input-dim group
(H symmetric orthogonal, so H^T == H and W_rot @ H recovers W)."""
out_f, in_f = w.shape
H = build_hadamard(group_size)
wg = w.reshape(out_f, in_f // group_size, group_size)
return (wg @ H).reshape(out_f, in_f).astype(np.float32)
def read_donor(donor_path):
"""Return (config dict, donor name, silence_latent f32 array, {name: shape})."""
r = gguf.GGUFReader(donor_path)
fields = {f.name: f for f in r.fields.values()}
def get_str(key):
f = fields.get(key)
return bytes(f.parts[f.data[0]]).decode() if f else None
cfg_json = get_str("acestep.config_json")
if not cfg_json:
log("FATAL: donor %s has no acestep.config_json (not a convert.py DiT GGUF?)" % donor_path)
sys.exit(1)
silence = None
shapes = {}
for t in r.tensors:
shapes[t.name] = tuple(reversed([int(d) for d in t.shape])) # ne order -> torch order
if t.name == "silence_latent":
silence = np.asarray(t.data).flatten().reshape(15000, 64).astype(np.float32)
if silence is None:
log("FATAL: donor has no silence_latent tensor")
sys.exit(1)
return json.loads(cfg_json), get_str("general.name"), silence, shapes
def main():
ap = argparse.ArgumentParser()
ap.add_argument("comfy_st")
ap.add_argument("donor_gguf")
ap.add_argument("out_gguf")
ap.add_argument("--name", default=None, help="general.name for the output (default: derived from output filename)")
ap.add_argument("--no-runtime-rotation", action="store_true",
help="Dequantize + unrotate EVERYTHING to BF16 (no acestep.convrot_map, no engine "
"rotation needed). Numerically equivalent reference build — 2x the size; used "
"to A/B-validate the engine's runtime rotation path.")
args = ap.parse_args()
name = args.name or os.path.basename(args.out_gguf).rsplit(".", 1)[0]
cfg, donor_name, silence, donor_shapes = read_donor(args.donor_gguf)
log("donor: %s (%d tensors)" % (donor_name, len(donor_shapes)))
meta, hdr_size = convert.read_sf_header(args.comfy_st)
w = gguf.GGUFWriter(args.out_gguf, "acestep-dit", use_temp_file=True)
w.add_name(name)
convert.add_metadata(w, cfg, "dit")
n_q8, n_bf16, n_dequant, total = 0, 0, 0, 0
shape_mismatches = []
convrot_map = [] # (tensor name, group size) kept rotated in the GGUF
with open(args.comfy_st, "rb") as f:
def read_tensor(tname):
info = meta[tname]
f.seek(hdr_size + info["data_offsets"][0])
return f.read(info["data_offsets"][1] - info["data_offsets"][0]), info
def layer_quant_config(base):
key = base + ".comfy_quant"
if key not in meta:
return {}
raw, _ = read_tensor(key)
try:
return json.loads(raw.decode())
except Exception:
return {}
for tname in sorted(meta.keys()):
if tname.endswith(".weight_scale") or tname.endswith(".comfy_quant"):
continue
info = meta[tname]
dtype, shape = info["dtype"], info["shape"]
donor_shape = donor_shapes.get(tname)
if donor_shape is None:
log(" WARNING: %s not in donor GGUF -- writing anyway" % tname)
elif donor_shape != tuple(shape):
shape_mismatches.append((tname, tuple(shape), donor_shape))
raw, _ = read_tensor(tname)
if dtype == "I8":
base = tname[: -len(".weight")]
sraw, sinfo = read_tensor(base + ".weight_scale")
scales = np.frombuffer(sraw, dtype=np.float32)
qcfg = layer_quant_config(base)
rot_group = int(qcfg.get("convrot_groupsize", 0)) if qcfg.get("convrot") else 0
qs = np.frombuffer(raw, dtype=np.int8).reshape(-1, shape[-1])
rows, cols = qs.shape
# scalar scale -> broadcast per-row; per-row scale as-is
if scales.size == 1:
row_scales = np.full(rows, scales[0], dtype=np.float32)
elif scales.size == rows:
row_scales = scales.reshape(rows)
else:
log(" FATAL: %s weight_scale has %d entries for %d rows" % (tname, scales.size, rows))
sys.exit(1)
keep_rotation = rot_group > 0 and tname.startswith("decoder.") and not args.no_runtime_rotation
packable = len(shape) == 2 and cols % QK8_0 == 0 and not (rot_group > 0 and args.no_runtime_rotation)
if packable and (rot_group == 0 or keep_rotation):
packed = pack_q8_0(qs, row_scales)
w.add_tensor(tname, packed, raw_dtype=Q8_0)
if keep_rotation:
convrot_map.append((tname, rot_group))
n_q8 += 1
total += packed.nbytes
else:
# dequant fallback; undo rotation so no runtime support needed
deq = qs.astype(np.float32) * row_scales[:, None]
if rot_group > 0:
deq = unrotate_weight(deq, rot_group)
w.add_tensor(tname, f32_to_bf16(deq.reshape(shape)), raw_dtype=BF16)
n_dequant += 1
total += deq.size * 2
elif dtype == "BF16":
arr = np.frombuffer(raw, dtype=np.uint16).reshape(shape)
w.add_tensor(tname, arr, raw_dtype=BF16)
n_bf16 += 1
total += arr.nbytes
elif dtype == "F32":
arr = f32_to_bf16(np.frombuffer(raw, dtype=np.float32)).reshape(shape)
w.add_tensor(tname, arr, raw_dtype=BF16)
n_bf16 += 1
total += arr.nbytes
else:
log(" FATAL: %s has unsupported dtype %s" % (tname, dtype))
sys.exit(1)
w.add_tensor("silence_latent", silence)
total += silence.nbytes
if shape_mismatches:
for tname, got, want in shape_mismatches:
log(" FATAL: shape mismatch %s: comfy %s vs donor %s" % (tname, got, want))
sys.exit(1)
if convrot_map:
w.add_string("acestep.convrot_map", ";".join("%s:%d" % (n, g) for n, g in convrot_map))
groups = sorted(set(g for _, g in convrot_map))
log("convrot: %d rotated decoder weights kept (group sizes %s) -> acestep.convrot_map"
% (len(convrot_map), groups))
log("tensors: %d Q8_0, %d BF16, %d dequant-fallback, + silence_latent (%.2f GB)"
% (n_q8, n_bf16, n_dequant, total / (1 << 30)))
w.write_header_to_file()
w.write_kv_data_to_file()
w.write_tensors_to_file(progress=True)
w.close()
log("wrote %.0f MB -> %s" % (os.path.getsize(args.out_gguf) / (1 << 20), args.out_gguf))
if __name__ == "__main__":
main()
engine/convert-sa3.py
+175
View File
@@ -0,0 +1,175 @@
#!/usr/bin/env python3
# convert-sa3.py: Stable Audio 3 checkpoints -> GGUF for the StableStep GGML backend.
#
# Produces four GGUFs into models/:
# sa3-dit-BF16.gguf arch "sa3-dit" from stable-audio-3-medium (model.* keys)
# sa3-same-enc-F16.gguf arch "sa3-same-enc" from pretransform.model.* (encoder side)
# sa3-same-dec-F16.gguf arch "sa3-same-dec" from pretransform.model.* (decoder side)
# sa3-text-enc-BF16.gguf arch "sa3-t5gemma" from the t5gemma-b-b-ul2 subfolder
#
# The SAME pair is stored F16 (not BF16): the decoder's sinusoidal FF blocks
# amplify per-weight rounding noise across its 12 layers (see write_sa3_gguf).
#
# Tensor policy: >=2D weights -> BF16; 1D tensors (norms, biases, scales) -> F32
# (precision finding from the ONNX leg: this model's norm/timestep paths are
# fp32-sensitive — measured cosine 0.966 with blanket fp16 vs 0.9995 scoped).
# Tensor names are the source names minus the strip prefix; the C++ graph
# builders consume them as-is. The full model_config.json is embedded verbatim
# under metadata key "sa3.config_json" (the C++ side parses what it needs).
#
# Runs in the StableAudio3 uv venv:
# cd d:/Ace-Step-Latest/StableAudio3
# uv run --with gguf python d:/Ace-Step-Latest/hot-step-cpp/engine/convert-sa3.py
import json
import os
import sys
import numpy as np
import gguf
sys.path.insert(0, r"d:/Ace-Step-Latest/StableAudio3")
from huggingface_hub import hf_hub_download
from safetensors import safe_open
OUTPUT_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "models")
REPO = "stabilityai/stable-audio-3-medium"
def log(msg):
print(f"[convert-sa3] {msg}", file=sys.stderr, flush=True)
def to_np(t):
import torch
if t.dtype == torch.bfloat16:
return t.float().numpy()
return t.numpy()
def write_sa3_gguf(out_path, arch, tensors, config_json, extra_meta=None, half="bf16"):
"""tensors: list of (name, np.float32 array). >=2D stored in `half`
("bf16" or "f16"), 1D tensors (norms, biases, scales) stored F32.
half="f16" is used for the SAME autoencoder halves: the decoder's
sinusoidal FF layers amplify weight rounding noise layer over layer
(bf16 ~0.4% rel error -> parity cosine 0.9987 < 0.999; f16 ~0.05%
passes). Same file size either way; weight magnitudes are far inside
f16 range."""
w = gguf.GGUFWriter(out_path, arch)
w.add_string("sa3.config_json", config_json)
for k, v in (extra_meta or {}).items():
w.add_string(k, v)
import torch
n_half = n_f32 = 0
for name, arr in tensors:
arr = np.ascontiguousarray(arr, dtype=np.float32)
if arr.size == 0:
# e.g. bottleneck.noise_scaling_factor (1, 0, 1) when
# noise_augment_dim == 0. ggml's gguf reader hits an integer
# divide-by-zero on ne==0 tensors, and the C++ side never reads
# them — drop.
log(f" skipping zero-element tensor {name} {arr.shape}")
continue
if arr.ndim >= 2:
# raw_dtype does NOT convert — it labels. Convert to 16-bit bytes
# explicitly (uint16 view keeps the logical shape).
if half == "f16":
h = torch.from_numpy(arr).to(torch.float16).view(torch.uint16).numpy()
w.add_tensor(name, h, raw_dtype=gguf.GGMLQuantizationType.F16)
else:
h = torch.from_numpy(arr).to(torch.bfloat16).view(torch.uint16).numpy()
w.add_tensor(name, h, raw_dtype=gguf.GGMLQuantizationType.BF16)
n_half += 1
else:
w.add_tensor(name, arr) # F32
n_f32 += 1
w.write_header_to_file()
w.write_kv_data_to_file()
w.write_tensors_to_file()
w.close()
size = os.path.getsize(out_path) / 1e9
log(f"{os.path.basename(out_path)}: {n_half} {half.upper()} + {n_f32} F32 tensors, {size:.2f} GB")
def main():
# Local-checkpoint support (e.g. LoRA-merged models): --ckpt/--config
# override the HF download; --out-dir redirects output; --dit-only skips
# the SAME + T5Gemma GGUFs (unchanged when only the DiT was fine-tuned).
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt", default=None, help="Local model.safetensors (default: HF download)")
ap.add_argument("--config", default=None, help="Local model_config.json (default: HF download)")
ap.add_argument("--out-dir", default=OUTPUT_DIR)
ap.add_argument("--dit-only", action="store_true")
args = ap.parse_args()
out_dir = args.out_dir
os.makedirs(out_dir, exist_ok=True)
cfg_path = args.config or hf_hub_download(REPO, "model_config.json")
ckpt_path = args.ckpt or hf_hub_download(REPO, "model.safetensors")
with open(cfg_path) as f:
config_json = f.read()
# ── Split the combined checkpoint by prefix ─────────────────────────
dit_tensors, enc_tensors, dec_tensors = [], [], []
with safe_open(ckpt_path, framework="pt", device="cpu") as f:
for key in f.keys():
if key.startswith("model."):
dit_tensors.append((key[len("model."):], to_np(f.get_tensor(key))))
elif key.startswith("pretransform.model."):
sub = key[len("pretransform.model."):]
# AudioAutoencoder members: encoder.*, decoder.*, bottleneck.*,
# pretransform.* (patched — no weights). Bottleneck params go to BOTH
# (encoder applies scale/bias+running_std, decoder inverts).
if sub.startswith("encoder."):
enc_tensors.append((sub, to_np(f.get_tensor(key))))
elif sub.startswith("decoder."):
dec_tensors.append((sub, to_np(f.get_tensor(key))))
elif sub.startswith("bottleneck."):
t = to_np(f.get_tensor(key))
enc_tensors.append((sub, t))
dec_tensors.append((sub, t))
# conditioner.* (learned padding, seconds embedder) rides with the DiT
# gguf — small and needed by the same backend module.
elif key.startswith("conditioner."):
dit_tensors.append((key, to_np(f.get_tensor(key))))
write_sa3_gguf(os.path.join(out_dir, "sa3-dit-BF16.gguf"),
"sa3-dit", dit_tensors, config_json)
if args.dit_only:
log("Done (dit-only).")
return
write_sa3_gguf(os.path.join(out_dir, "sa3-same-enc-F16.gguf"),
"sa3-same-enc", enc_tensors, config_json, half="f16")
write_sa3_gguf(os.path.join(out_dir, "sa3-same-dec-F16.gguf"),
"sa3-same-dec", dec_tensors, config_json, half="f16")
# ── T5Gemma encoder (separate HF model in the repo subfolder) ───────
t5_cfg = hf_hub_download(REPO, "config.json", subfolder="t5gemma-b-b-ul2")
t5_ckpt = hf_hub_download(REPO, "model.safetensors", subfolder="t5gemma-b-b-ul2")
with open(t5_cfg) as f:
t5_config_json = f.read()
t5_tensors = []
with safe_open(t5_ckpt, framework="pt", device="cpu") as f:
for key in f.keys():
# Encoder-only: drop the decoder half (never used by SA3)
if key.startswith("decoder."):
continue
t5_tensors.append((key, to_np(f.get_tensor(key))))
# The SA3 conditioner's learned padding embedding is applied to the text
# encoder's output (padded positions replaced) — it belongs to this module,
# so duplicate it here (it also rides in the DiT gguf with the rest of
# conditioner.*).
with safe_open(ckpt_path, framework="pt", device="cpu") as f:
key = "conditioner.conditioners.prompt.padding_embedding"
t5_tensors.append((key, to_np(f.get_tensor(key))))
write_sa3_gguf(os.path.join(out_dir, "sa3-text-enc-BF16.gguf"),
"sa3-t5gemma", t5_tensors, t5_config_json,
extra_meta={"sa3.parent_config_json": config_json})
log("Done.")
if __name__ == "__main__":
main()
engine/convert.py
+304
View File
@@ -0,0 +1,304 @@
#!/usr/bin/env python3
# convert.py: safetensors to GGUF for ACE-Step (LM, DiT, TextEncoder, VAE)
# Reads from checkpoints/, writes GGUF to models/
# Each GGUF is self-contained: weights + config + tokenizer + silence_latent
import os
import sys
import json
import struct
import zipfile
import numpy as np
import gguf
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
CHECKPOINT_DIR = os.path.join(SCRIPT_DIR, "checkpoints")
OUTPUT_DIR = os.path.join(SCRIPT_DIR, "models")
BF16 = gguf.GGMLQuantizationType.BF16
def log(tag, msg):
print("[%s] %s" % (tag, msg), file=sys.stderr, flush=True)
# Safetensors reader
def read_sf_header(path):
with open(path, "rb") as f:
n = struct.unpack("<Q", f.read(8))[0]
meta = json.loads(f.read(n))
meta.pop("__metadata__", None)
return meta, 8 + n
def find_sf_files(model_dir):
"""Return list of safetensors paths (single, sharded, or diffusers VAE)."""
single = os.path.join(model_dir, "model.safetensors")
if os.path.exists(single):
return [single]
index = os.path.join(model_dir, "model.safetensors.index.json")
if os.path.exists(index):
with open(index, "r", encoding="utf-8") as f:
idx = json.load(f)
shards = sorted(set(idx["weight_map"].values()))
return [os.path.join(model_dir, s) for s in shards]
diffusers = os.path.join(model_dir, "diffusion_pytorch_model.safetensors")
if os.path.exists(diffusers):
return [diffusers]
return []
# Model classification
ARCHS = {
"lm": "acestep-lm",
"dit": "acestep-dit",
"text-enc": "acestep-text-enc",
"vae": "acestep-vae",
}
def classify(name):
if name.startswith("acestep-5Hz-lm"):
return "lm"
if name.startswith("acestep-v15"):
return "dit"
if name.startswith("Qwen3-Embedding"):
return "text-enc"
if name == "vae":
return "vae"
return None
# GGUF metadata from config.json
def add_metadata(w, cfg, model_type):
if "num_hidden_layers" in cfg:
w.add_block_count(cfg["num_hidden_layers"])
if "hidden_size" in cfg:
w.add_embedding_length(cfg["hidden_size"])
if "intermediate_size" in cfg:
w.add_feed_forward_length(cfg["intermediate_size"])
if "num_attention_heads" in cfg:
w.add_head_count(cfg["num_attention_heads"])
if "num_key_value_heads" in cfg:
w.add_head_count_kv(cfg["num_key_value_heads"])
if "head_dim" in cfg:
w.add_key_length(cfg["head_dim"])
if "vocab_size" in cfg:
w.add_vocab_size(cfg["vocab_size"])
if "max_position_embeddings" in cfg:
w.add_context_length(cfg["max_position_embeddings"])
if "rms_norm_eps" in cfg:
w.add_layer_norm_rms_eps(cfg["rms_norm_eps"])
rope = cfg.get("rope_theta")
if rope:
w.add_rope_freq_base(float(rope))
if model_type == "lm":
if cfg.get("tie_word_embeddings"):
w.add_bool("acestep.tie_word_embeddings", True)
if model_type == "dit":
for key in [
"in_channels", "audio_acoustic_hidden_dim", "patch_size",
"sliding_window", "fsq_dim", "text_hidden_dim", "timbre_hidden_dim",
"num_lyric_encoder_hidden_layers", "num_timbre_encoder_hidden_layers",
"num_audio_decoder_hidden_layers", "num_attention_pooler_hidden_layers",
]:
if key in cfg:
w.add_uint32("acestep.%s" % key, cfg[key])
# XL models have separate encoder dimensions (2B models omit these)
for key in [
"encoder_hidden_size", "encoder_intermediate_size",
"encoder_num_attention_heads", "encoder_num_key_value_heads",
]:
if key in cfg:
w.add_uint32("acestep.%s" % key, cfg[key])
if cfg.get("is_turbo"):
w.add_bool("acestep.is_turbo", True)
levels = cfg.get("fsq_input_levels")
if levels:
w.add_array("acestep.fsq_input_levels", levels)
w.add_string("acestep.config_json", json.dumps(cfg, separators=(",", ":")))
# Tensor packing from safetensors
def add_tensors_from_sf(w, sf_path, tag, model_type):
meta, hdr_size = read_sf_header(sf_path)
names = sorted(meta.keys())
with open(sf_path, "rb") as f:
count = 0
total = 0
for name in names:
info = meta[name]
# normalize: some upstream checkpoints omit the "model." prefix
if model_type == "lm" and not name.startswith("model."):
name = "model." + name
dtype_str = info["dtype"]
shape = info["shape"]
off0, off1 = info["data_offsets"]
nbytes = off1 - off0
f.seek(hdr_size + off0)
raw = f.read(nbytes)
if dtype_str == "BF16":
arr = np.frombuffer(raw, dtype=np.uint16).reshape(shape)
w.add_tensor(name, arr, raw_dtype=BF16)
elif dtype_str == "F16":
arr = np.frombuffer(raw, dtype=np.float16).reshape(shape)
w.add_tensor(name, arr)
elif dtype_str == "F32":
# convert F32 to BF16: truncate lower 16 mantissa bits
arr = np.frombuffer(raw, dtype=np.uint32).reshape(shape)
arr = (arr >> 16).astype(np.uint16)
w.add_tensor(name, arr, raw_dtype=BF16)
nbytes = nbytes // 2 # actual stored size
else:
log(tag, " skip %s: dtype %s" % (name, dtype_str))
continue
count += 1
total += nbytes
return count, total
# silence_latent.pt reader (replaces pt2bin C++ tool)
# PyTorch .pt is a ZIP with entry "*/data/0" containing f32 [64, 15000]
# We transpose to [15000, 64] (ggml layout: 64 contiguous per frame)
def read_silence_latent(model_dir):
pt_path = os.path.join(model_dir, "silence_latent.pt")
if not os.path.exists(pt_path):
return None
with zipfile.ZipFile(pt_path) as z:
for entry in z.namelist():
if entry.endswith("/data/0"):
raw = z.read(entry)
src = np.frombuffer(raw, dtype=np.float32).reshape(64, 15000)
return src.T.copy()
return None
# BPE tokenizer embedding (vocab.json + merges.txt -> GGUF KV)
def add_bpe_tokenizer(w, model_dir, tag):
vocab_path = os.path.join(model_dir, "vocab.json")
merges_path = os.path.join(model_dir, "merges.txt")
if not os.path.exists(vocab_path) or not os.path.exists(merges_path):
return False
with open(vocab_path, "r", encoding="utf-8") as f:
vocab = json.load(f)
tokens = [""] * len(vocab)
for tok_str, tok_id in vocab.items():
if 0 <= tok_id < len(tokens):
tokens[tok_id] = tok_str
with open(merges_path, "r", encoding="utf-8") as f:
merges = []
for line in f:
line = line.rstrip("\n\r")
if not line:
continue
if line.startswith("#version:"):
continue
merges.append(line)
w.add_tokenizer_model("gpt2")
w.add_token_list(tokens)
w.add_token_merges(merges)
log(tag, " tokenizer: %d vocab, %d merges" % (len(tokens), len(merges)))
return True
# Main conversion
def convert_model(name, model_dir, output_path, model_type):
tag = "GGUF"
cfg_path = os.path.join(model_dir, "config.json")
if not os.path.exists(cfg_path):
log(tag, "skip %s: no config.json" % name)
return False
with open(cfg_path, "r", encoding="utf-8") as f:
cfg = json.load(f)
sf_files = find_sf_files(model_dir)
if not sf_files:
log(tag, "skip %s: no safetensors" % name)
return False
arch = ARCHS[model_type]
log(tag, "%s (%s, %d shard%s) -> %s" % (
name, arch, len(sf_files), "" if len(sf_files) == 1 else "s",
os.path.basename(output_path)))
w = gguf.GGUFWriter(output_path, arch, use_temp_file=True)
w.add_name(name)
add_metadata(w, cfg, model_type)
# BPE tokenizer for LM and text encoder
if model_type in ("lm", "text-enc"):
add_bpe_tokenizer(w, model_dir, tag)
# Model weights
n_tensors = 0
n_bytes = 0
for sf in sf_files:
c, b = add_tensors_from_sf(w, sf, tag, model_type)
n_tensors += c
n_bytes += b
if len(sf_files) > 1:
log(tag, " %s: %d tensors" % (os.path.basename(sf), c))
# silence_latent for DiT (read .pt, transpose, embed as f32 tensor)
if model_type == "dit":
sl = read_silence_latent(model_dir)
if sl is not None:
w.add_tensor("silence_latent", sl)
n_tensors += 1
n_bytes += sl.nbytes
log(tag, " silence_latent: [%d, %d] f32 (%.1f MB)" % (
sl.shape[0], sl.shape[1], sl.nbytes / (1 << 20)))
else:
log(tag, " WARNING: no silence_latent.pt found")
log(tag, " total: %d tensors, %.1f GB" % (n_tensors, n_bytes / (1 << 30)))
w.write_header_to_file()
w.write_kv_data_to_file()
w.write_tensors_to_file(progress=True)
w.close()
out_mb = os.path.getsize(output_path) / (1 << 20)
log(tag, " wrote %.0f MB -> %s" % (out_mb, output_path))
return True
def main():
if not os.path.isdir(CHECKPOINT_DIR):
log("GGUF", "checkpoints/ not found, run checkpoints.sh first")
sys.exit(1)
os.makedirs(OUTPUT_DIR, exist_ok=True)
entries = sorted(os.listdir(CHECKPOINT_DIR))
converted = 0
skipped = []
for name in entries:
model_dir = os.path.join(CHECKPOINT_DIR, name)
if not os.path.isdir(model_dir):
continue
model_type = classify(name)
if model_type is None:
skipped.append(name)
continue
output_path = os.path.join(OUTPUT_DIR, "%s-BF16.gguf" % name)
if os.path.exists(output_path):
log("GGUF", "skip %s: %s exists" % (name, os.path.basename(output_path)))
converted += 1
continue
if convert_model(name, model_dir, output_path, model_type):
converted += 1
if skipped:
log("GGUF", "skipped (unknown): %s" % ", ".join(skipped))
log("GGUF", "done: %d model(s) in %s" % (converted, OUTPUT_DIR))
if __name__ == "__main__":
main()
engine/docs/ARCHITECTURE.md
+1176
View File
File diff suppressed because it is too large Load Diff
engine/examples/ace-understand.sh
+37
View File
@@ -0,0 +1,37 @@
#!/bin/bash
# Roundtrip: audio -> understand -> SFT DiT -> MP3
#
# Usage: ./ace-understand.sh input.wav (or input.mp3)
#
# understand:
# input -> ace-understand.json (audio codes + metadata)
#
# ace-synth:
# ace-understand.json -> ace-understand0.mp3
set -eu
if [ $# -lt 1 ]; then
echo "Usage: $0 <input.wav|input.mp3>"
exit 1
fi
input="$1"
../build/ace-understand \
--src-audio "$input" \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf \
-o ace-understand.json
sed -i \
's/"audio_cover_strength": *[0-9.]*/"audio_cover_strength": 0.04/' \
ace-understand.json
../build/ace-synth \
--src-audio "$input" \
--request ace-understand.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/client-batch.py
+104
View File
@@ -0,0 +1,104 @@
#!/usr/bin/env python3
# client-batch.py: test batching via ace-server
#
# POST /lm (lm_batch_size=2 in JSON) -> 2 enriched requests
# POST /synth (JSON array of 2 requests) -> 2 MP3s in one GPU batch
#
# Start the server first: ./server.sh
import json
import sys
import urllib.error
import urllib.request
URL = "http://127.0.0.1:8085"
def post_json(endpoint, data):
body = json.dumps(data).encode()
req = urllib.request.Request(
URL + endpoint,
data=body,
headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(req) as resp:
return resp.read(), resp.headers
def parse_multipart_mixed(data, content_type):
"""Parse multipart/mixed response into list of body bytes."""
boundary = None
for part in content_type.split(";"):
part = part.strip()
if part.startswith("boundary="):
boundary = part[len("boundary="):].strip().encode()
break
if not boundary:
raise ValueError("no boundary in content-type: " + content_type)
delimiter = b"--" + boundary
parts = []
for chunk in data.split(delimiter):
if not chunk or chunk.startswith(b"--"):
continue
chunk = chunk.strip(b"\r\n")
if not chunk:
continue
sep = chunk.find(b"\r\n\r\n")
if sep < 0:
continue
body = chunk[sep + 4:]
if body.endswith(b"\r\n"):
body = body[:-2]
parts.append(body)
return parts
# Phase 1: LM generates N variations
try:
with open("simple-batch.json") as f:
request_json = json.load(f)
except FileNotFoundError:
print("ERROR: simple-batch.json not found (run from the examples/ directory)")
sys.exit(1)
try:
lm_batch_size = request_json.get("lm_batch_size", 1)
print("POST /lm (lm_batch_size=%d)..." % lm_batch_size)
lm_data, _ = post_json("/lm", request_json)
except urllib.error.URLError as e:
print("ERROR: cannot connect to %s (%s)" % (URL, e.reason))
print("Start the server first: ./server.sh")
sys.exit(1)
lm_results = json.loads(lm_data)
print(" -> %d enriched requests" % len(lm_results))
# Phase 2: synth all in one GPU batch (send JSON array)
print("POST /synth (batch=%d, JSON array)..." % len(lm_results))
body = json.dumps(lm_results).encode()
req = urllib.request.Request(
URL + "/synth",
data=body,
headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(req) as resp:
resp_data = resp.read()
content_type = resp.headers.get("Content-Type", "")
if "multipart/mixed" in content_type:
parts = parse_multipart_mixed(resp_data, content_type)
for i, mp3_data in enumerate(parts):
path = "server-batch%d.mp3" % i
with open(path, "wb") as f:
f.write(mp3_data)
print(" -> %s (%d bytes)" % (path, len(mp3_data)))
else:
path = "server-batch0.mp3"
with open(path, "wb") as f:
f.write(resp_data)
print(" -> %s (%d bytes)" % (path, len(resp_data)))
print("Done: %d MP3(s)" % len(lm_results))
engine/examples/client-understand.sh
+54
View File
@@ -0,0 +1,54 @@
#!/bin/bash
# Roundtrip via ace-server: audio -> understand -> synth -> MP3
#
# Usage: ./client-understand.sh input.wav (or input.mp3)
#
# POST /understand (async job):
# input -> server-understand.json (audio codes + metadata)
#
# POST /synth (async job):
# server-understand.json + input -> server-understand.mp3
#
# Start the server first (./server.sh).
set -eu
if [ $# -lt 1 ]; then
echo "Usage: $0 <input.wav|input.mp3>"
exit 1
fi
HOST="http://127.0.0.1:8085"
input="$1"
# poll a job until done, exit 1 on failure
wait_job() {
local id="$1"
while true; do
status=$(curl -sf "${HOST}/job?id=${id}" | jq -r '.status')
case "$status" in
done) return 0 ;;
failed|cancelled) echo "Job ${id}: ${status}"; return 1 ;;
esac
sleep 2
done
}
# understand: submit, poll, fetch result
UND_ID=$(curl -sf "${HOST}/understand" \
-F "audio=@${input}" | jq -r '.id')
echo "Understand job: ${UND_ID}"
wait_job "${UND_ID}"
curl -sf "${HOST}/job?id=${UND_ID}&result=1" -o server-understand.json
sed -i \
-e 's/"audio_cover_strength": *[0-9.]*/"audio_cover_strength": 0.04/' \
server-understand.json
# synth: submit, poll, fetch result
SYNTH_ID=$(curl -sf "${HOST}/synth" \
-F "request=@server-understand.json" \
-F "audio=@${input}" | jq -r '.id')
echo "Synth job: ${SYNTH_ID}"
wait_job "${SYNTH_ID}"
curl -sf "${HOST}/job?id=${SYNTH_ID}&result=1" -o server-understand.mp3
engine/examples/client.sh
+36
View File
@@ -0,0 +1,36 @@
#!/bin/bash
# Test ace-server: LM enriches caption, synth renders to MP3.
# Start the server first (./server.sh), then run this.
set -eu
HOST="http://127.0.0.1:8085"
# poll a job until done, exit 1 on failure
wait_job() {
local id="$1"
while true; do
status=$(curl -sf "${HOST}/job?id=${id}" | jq -r '.status')
case "$status" in
done) return 0 ;;
failed|cancelled) echo "Job ${id}: ${status}"; return 1 ;;
esac
sleep 2
done
}
# LM: submit, poll, fetch result
LM_ID=$(curl -sf "${HOST}/lm" \
-H "Content-Type: application/json" \
-d @full-sft.json | jq -r '.id')
echo "LM job: ${LM_ID}"
wait_job "${LM_ID}"
curl -sf "${HOST}/job?id=${LM_ID}&result=1" | jq '.[0]' > server-lm0.json
# synth: submit, poll, fetch result
SYNTH_ID=$(curl -sf "${HOST}/synth" \
-H "Content-Type: application/json" \
-d @server-lm0.json | jq -r '.id')
echo "Synth job: ${SYNTH_ID}"
wait_job "${SYNTH_ID}"
curl -sf "${HOST}/job?id=${SYNTH_ID}&result=1" -o server0.mp3
engine/examples/dit-only-sft.json
+12
View File
@@ -0,0 +1,12 @@
{
"caption": "Ambient electronic soundscape with warm analog pads",
"lyrics": "",
"bpm": 90,
"duration": 180,
"keyscale": "C minor",
"timesignature": "4",
"vocal_language": "en",
"inference_steps": 50,
"guidance_scale": 1.0,
"shift": 1.0
}
engine/examples/dit-only-sft.sh
+9
View File
@@ -0,0 +1,9 @@
#!/bin/bash
set -eu
../build/ace-synth \
--request dit-only-sft.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/dit-only.json
+12
View File
@@ -0,0 +1,12 @@
{
"caption": "Ambient electronic soundscape with warm analog pads",
"lyrics": "",
"bpm": 90,
"duration": 180,
"keyscale": "C minor",
"timesignature": "4",
"vocal_language": "en",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0
}
engine/examples/dit-only.sh
+9
View File
@@ -0,0 +1,9 @@
#!/bin/bash
set -eu
../build/ace-synth \
--request dit-only.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/full-sft.json
+12
View File
@@ -0,0 +1,12 @@
{
"caption": "Upbeat French house with infectious disco-inspired bassline, crisp four-on-the-floor kick pattern, wah-wah filtered guitar riffs, retro synth stabs, soulful male lead vocals with gospel-style backing harmonies, smooth saxophone accents, warm vinyl crackle texture, bright summer vibe, polished modern mix with vintage analog warmth, driving yet laid-back energy perfect for rooftop parties and sunset drives",
"lyrics": "[Intro - Ligne de Basse Funk & Beat House]\n\n[Verse 1]\nSous le soleil de Paris, on danse sans fin\nLa nuit s'allume, le beat nous guide\nLes étoiles scintillent au rythme du kick\nUn sourire léger, tout est si vivant\n\n[Pre-Chorus]\nLaisse-toi porter par la musique qui chante\n\n[Chorus]\nOn danse sous le ciel étoilé\nLe monde s'arrête, on s'envole\nAvec ce groove qui nous emporte\nJusqu'au matin, on ne s'arrête pas\n\n[Verse 2]\nLa ville respire au son des cuivres légers\nLes mains en l'air, on oublie le temps\nLa basse funk nous secoue les pieds\nUn été éternel, rien ne peut nous briser\n\n[Chorus]\nOn danse sous le ciel étoilé\nLe monde s'arrête, on s'envole\nAvec ce groove qui nous emporte\nJusqu'au matin, on ne s'arrête pas\n\n[Guitar Solo - Wah-Wah Funk]\n\n[Bridge - Saxophone & Cordes]\nRespire profondément, l'univers t'appelle\n\n[Outro - Synth Fade avec Craquement Vinyle]",
"duration": 240,
"bpm": 124,
"vocal_language": "fr",
"keyscale": "F# major",
"timesignature": "4",
"inference_steps": 50,
"guidance_scale": 1.0,
"shift": 1.0
}
engine/examples/full-sft.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request full-sft.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request full-sft0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/full.json
+12
View File
@@ -0,0 +1,12 @@
{
"caption": "Upbeat French house with infectious disco-inspired bassline, crisp four-on-the-floor kick pattern, wah-wah filtered guitar riffs, retro synth stabs, soulful male lead vocals with gospel-style backing harmonies, smooth saxophone accents, warm vinyl crackle texture, bright summer vibe, polished modern mix with vintage analog warmth, driving yet laid-back energy perfect for rooftop parties and sunset drives",
"lyrics": "[Intro - Ligne de Basse Funk & Beat House]\n\n[Verse 1]\nSous le soleil de Paris, on danse sans fin\nLa nuit s'allume, le beat nous guide\nLes étoiles scintillent au rythme du kick\nUn sourire léger, tout est si vivant\n\n[Pre-Chorus]\nLaisse-toi porter par la musique qui chante\n\n[Chorus]\nOn danse sous le ciel étoilé\nLe monde s'arrête, on s'envole\nAvec ce groove qui nous emporte\nJusqu'au matin, on ne s'arrête pas\n\n[Verse 2]\nLa ville respire au son des cuivres légers\nLes mains en l'air, on oublie le temps\nLa basse funk nous secoue les pieds\nUn été éternel, rien ne peut nous briser\n\n[Chorus]\nOn danse sous le ciel étoilé\nLe monde s'arrête, on s'envole\nAvec ce groove qui nous emporte\nJusqu'au matin, on ne s'arrête pas\n\n[Guitar Solo - Wah-Wah Funk]\n\n[Bridge - Saxophone & Cordes]\nRespire profondément, l'univers t'appelle\n\n[Outro - Synth Fade avec Craquement Vinyle]",
"duration": 240,
"bpm": 124,
"vocal_language": "fr",
"keyscale": "F# major",
"timesignature": "4",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0
}
engine/examples/full.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request full.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request full0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/lego.json
+9
View File
@@ -0,0 +1,9 @@
{
"caption": "electric guitar riff, funk guitar, house music, instrumental",
"lyrics": "[Instrumental]",
"task_type": "lego",
"track": "guitar",
"inference_steps": 50,
"guidance_scale": 1.0,
"shift": 1.0
}
engine/examples/lego.sh
+33
View File
@@ -0,0 +1,33 @@
#!/bin/bash
# Generate a source track, then lego a guitar stem over it
#
# Note: lego requires acestep-v15-base; turbo/sft do not support it
#
# LM + DiT phase (source track):
# simple.json -> simple0.json -> simple00.wav
#
# Lego phase (guitar stem over source):
# lego.json + simple00.wav -> lego0.wav
set -eu
# Phase 1: generate a source track with the simple prompt
../build/ace-lm \
--request simple.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request simple0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf \
--format wav16
# Phase 2: lego guitar on the generated track (base model required)
../build/ace-synth \
--src-audio simple00.wav \
--request lego.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-base-Q8_0.gguf \
--vae ../models/vae-BF16.gguf \
--format wav16
engine/examples/partial-sft.json
+9
View File
@@ -0,0 +1,9 @@
{
"caption": "Hard-hitting hip hop track with deep 808 bass, crisp trap hi-hats, heavy snare rolls, dark piano melody, and confident aggressive vocal delivery",
"lyrics": "[Intro]\nYeah... c'est comme ca...\n\n[Verse 1]\nJe marche dans la ville quand le soleil se couche\nLes lumieres s'allument j'ai les mots dans la bouche\nOn m'a dit fais attention le monde est pas facile\nJ'ai repondu tranquille j'ai grandi dans la ville\nLes murs ont des oreilles les rues ont des histoires\nChaque coin chaque angle chaque bout de trottoir\nJ'ai vu des gens tomber j'ai vu des gens se lever\nMoi je reste debout j'ai pas le temps de plier\n\n[Chorus]\nOn avance on recule pas\nLa vie nous teste a chaque pas\nOn avance on recule pas\nRegarde devant oublie tout ca\n\n[Verse 2]\nLe reveil sonne tot le cafe brule les levres\nLe metro le boulot la routine la fievre\nMais le soir dans ma chambre je reprends mon stylo\nJe pose sur le papier tout ce que j'ai sur le dos\nMes reves sont plus grands que les murs de ma chambre\nPlus chauds que juillet plus forts que decembre\nOn m'a dit sois realiste range tes illusions\nJ'ai repondu ma vie c'est pas de la fiction\n\n[Chorus]\nOn avance on recule pas\nLa vie nous teste a chaque pas\nOn avance on recule pas\nRegarde devant oublie tout ca\n\n[Outro]\nYeah... on lache rien... jamais...",
"duration": 200,
"vocal_language": "fr",
"inference_steps": 50,
"guidance_scale": 1.0,
"shift": 1.0
}
engine/examples/partial-sft.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request partial-sft.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request partial-sft0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/partial.json
+9
View File
@@ -0,0 +1,9 @@
{
"caption": "Hard-hitting hip hop track with deep 808 bass, crisp trap hi-hats, heavy snare rolls, dark piano melody, and confident aggressive vocal delivery",
"lyrics": "[Intro]\nYeah... c'est comme ca...\n\n[Verse 1]\nJe marche dans la ville quand le soleil se couche\nLes lumieres s'allument j'ai les mots dans la bouche\nOn m'a dit fais attention le monde est pas facile\nJ'ai repondu tranquille j'ai grandi dans la ville\nLes murs ont des oreilles les rues ont des histoires\nChaque coin chaque angle chaque bout de trottoir\nJ'ai vu des gens tomber j'ai vu des gens se lever\nMoi je reste debout j'ai pas le temps de plier\n\n[Chorus]\nOn avance on recule pas\nLa vie nous teste a chaque pas\nOn avance on recule pas\nRegarde devant oublie tout ca\n\n[Verse 2]\nLe reveil sonne tot le cafe brule les levres\nLe metro le boulot la routine la fievre\nMais le soir dans ma chambre je reprends mon stylo\nJe pose sur le papier tout ce que j'ai sur le dos\nMes reves sont plus grands que les murs de ma chambre\nPlus chauds que juillet plus forts que decembre\nOn m'a dit sois realiste range tes illusions\nJ'ai repondu ma vie c'est pas de la fiction\n\n[Chorus]\nOn avance on recule pas\nLa vie nous teste a chaque pas\nOn avance on recule pas\nRegarde devant oublie tout ca\n\n[Outro]\nYeah... on lache rien... jamais...",
"duration": 200,
"vocal_language": "fr",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0
}
engine/examples/partial.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request partial.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request partial0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/simple-batch.json
+8
View File
@@ -0,0 +1,8 @@
{
"caption": "Upbeat pop rock anthem with driving electric guitars, punchy drums, catchy vocal hooks, and a singalong chorus",
"vocal_language": "fr",
"lm_batch_size": 2,
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0
}
engine/examples/simple-batch.sh
+23
View File
@@ -0,0 +1,23 @@
#!/bin/bash
# Generate 2 songs: LM produces 2 enriched requests (different codes/metas),
# DiT renders them in a single GPU batch.
#
# LM phase (lm_batch_size=2 in simple-batch.json):
# simple-batch.json -> simple-batch0.json, simple-batch1.json
#
# DiT phase (both requests in one batch):
# simple-batch0.json + simple-batch1.json -> simple-batch00.mp3, simple-batch11.mp3
set -eu
# Phase 1: LM generates 2 variations (different lyrics/codes/metas)
../build/ace-lm \
--request simple-batch.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
# Phase 2: DiT+VAE renders both in one GPU batch
../build/ace-synth \
--request simple-batch0.json simple-batch1.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/simple-sft.json
+7
View File
@@ -0,0 +1,7 @@
{
"caption": "Upbeat pop rock anthem with driving electric guitars, punchy drums, catchy vocal hooks, and a singalong chorus",
"vocal_language": "fr",
"inference_steps": 50,
"guidance_scale": 1.0,
"shift": 1.0
}
engine/examples/simple-sft.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request simple-sft.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request simple-sft0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-sft-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/examples/simple.cmd
+15
View File
@@ -0,0 +1,15 @@
@echo off
set PATH=%~dp0..\build\Release;%PATH%
ace-lm.exe ^
--request simple.json ^
--lm ..\models\acestep-5Hz-lm-4B-Q6_K.gguf
ace-synth.exe ^
--request simple0.json ^
--embedding ..\models\Qwen3-Embedding-0.6B-Q8_0.gguf ^
--dit ..\models\acestep-v15-turbo-Q6_K.gguf ^
--vae ..\models\vae-BF16.gguf
pause
engine/examples/simple.json
+7
View File
@@ -0,0 +1,7 @@
{
"caption": "Upbeat pop rock anthem with driving electric guitars, punchy drums, catchy vocal hooks, and a singalong chorus",
"vocal_language": "fr",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0
}
engine/examples/simple.sh
+13
View File
@@ -0,0 +1,13 @@
#!/bin/bash
set -eu
../build/ace-lm \
--request simple.json \
--lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf
../build/ace-synth \
--request simple0.json \
--embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
--dit ../models/acestep-v15-turbo-Q8_0.gguf \
--vae ../models/vae-BF16.gguf
engine/format.sh
+3
View File
@@ -0,0 +1,3 @@
#!/bin/bash
find . -name "*.cpp" -o -name "*.h" | grep -v -e build/ -e ggml/ -e vendor/ | xargs clang-format -i
engine/ggml/.editorconfig
+22
View File
@@ -0,0 +1,22 @@
# https://EditorConfig.org
# Top-most EditorConfig file
root = true
# Unix-style newlines with a newline ending every file, utf-8 charset
[*]
end_of_line = lf
insert_final_newline = true
trim_trailing_whitespace = true
charset = utf-8
indent_style = space
indent_size = 4
[*.md]
indent_size = 2
[Makefile]
indent_style = tab
[prompts/*.txt]
insert_final_newline = unset
engine/ggml/.github/pull_request_template.md
Vendored
+1
View File
@@ -0,0 +1 @@
*For changes to the core `ggml` library (including to the CMake build system), please open a PR in https://github.com/ggml-org/llama.cpp. Doing so will make your PR more visible, better tested and more likely to be reviewed.*
engine/ggml/.github/workflows/ci.yml
Vendored
+272
View File
@@ -0,0 +1,272 @@
name: CI
on:
push:
branches: [ master ]
pull_request:
branches: [ master ]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
build:
strategy:
matrix:
os: [ubuntu-latest, macos-latest, windows-latest]
libraries: [shared, static]
runs-on: ${{ matrix.os }}
steps:
- name: Clone
uses: actions/checkout@v6
- name: Dependencies for Ubuntu
if: matrix.os == 'ubuntu-latest'
run: |
sudo apt-get update
sudo apt-get install llvm
- name: Add msbuild to PATH
if: matrix.os == 'windows-latest'
uses: microsoft/setup-msbuild@v2
- name: Create Build Environment
run: mkdir build
- name: Configure CMake
working-directory: ./build
run: cmake ..
${{ contains(matrix.os, 'windows') && '-A x64' || '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++' }}
${{ matrix.libraries == 'static' && '-DBUILD_SHARED_LIBS=OFF' || '-DBUILD_SHARED_LIBS=ON' }}
-DCMAKE_INSTALL_PREFIX=${{ github.workspace }}/installed
-DGGML_METAL=OFF
- name: Build
working-directory: ./build
run: cmake --build . ${{ contains(matrix.os, 'windows') && '--config Release' || '' }}
- name: Test
working-directory: ./build
run: ctest --verbose --timeout 900 ${{ contains(matrix.os, 'windows') && '--build-config Release' || '' }}
- name: Install
working-directory: ./build
run: cmake --build . --target install ${{ contains(matrix.os, 'windows') && '--config Release' || '' }}
- name: Test CMake config
run: |
mkdir test-cmake
cmake -S examples/test-cmake -B test-cmake -DCMAKE_PREFIX_PATH=${{ github.workspace }}/installed ${{ contains(matrix.os, 'windows') && '-A x64' || '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++' }}
cmake --build test-cmake ${{ contains(matrix.os, 'windows') && '--config Release' || '' }}
# TODO: simplify the following workflows using a matrix
ggml-ci-x64-cpu-low-perf:
runs-on: ubuntu-22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-x64-cpu-low-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-low-perf:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-low-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-x64-cpu-high-perf:
runs-on: ubuntu-22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-x64-cpu-high-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-high-perf:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-high-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-high-perf-sve:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-high-perf-sve
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-x64-nvidia-cuda:
runs-on: [self-hosted, Linux, X64, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
nvidia-smi
GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/ggml /mnt/ggml
ggml-ci-x64-nvidia-vulkan-cm:
runs-on: [self-hosted, Linux, X64, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/ggml /mnt/ggml
ggml-ci-x64-nvidia-vulkan-cm2:
runs-on: [self-hosted, Linux, X64, NVIDIA, COOPMAT2]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/ggml /mnt/ggml
# TODO: provision AMX-compatible machine
#ggml-ci-x64-cpu-amx:
# runs-on: [self-hosted, Linux, X64, CPU, AMX]
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v6
# - name: Test
# id: ggml-ci
# run: |
# bash ./ci/run.sh ~/results/ggml /mnt/ggml
ggml-ci-mac-metal:
runs-on: [self-hosted, macOS, ARM64]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/ggml ~/mnt/ggml
ggml-ci-mac-vulkan:
runs-on: [self-hosted, macOS, ARM64]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/ggml ~/mnt/ggml
engine/ggml/.github/workflows/release.yml
Vendored
+27
View File
@@ -0,0 +1,27 @@
name: Release
on:
push:
tags:
- 'v*'
jobs:
release:
runs-on: ubuntu-latest
permissions:
contents: write
steps:
- name: Checkout code
uses: actions/checkout@v6
- name: Create Release
id: create_release
uses: ggml-org/action-create-release@v1
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
with:
tag_name: ${{ github.ref_name }}
release_name: ${{ github.ref }}
draft: false
prerelease: false
engine/ggml/.gitignore
+39
View File
@@ -0,0 +1,39 @@
build/
build-*/
out/
tmp/
models/
models-mnt
compile_commands.json
CMakeSettings.json
.vs/
.vscode/
.idea/
.clangd
.venv/
ggml_env/
.exrc
.cache
.DS_Store
.stablelm
.gpt-2
src/arm_neon.h
tests/arm_neon.h
zig-out/
zig-cache/
*.o
*.d
*.dot
*.sw?
__pycache__/
# Model files
ggml-model-f16.bin
*.bat
engine/ggml/.gitmodules
View File
engine/ggml/AUTHORS
+335
View File
@@ -0,0 +1,335 @@
# date: Tue Feb 4 13:03:51 EET 2025
# this file is auto-generated by scripts/gen-authors.sh
0cc4m <picard12@live.de>
65a <10104049+65a@users.noreply.github.com>
AT <manyoso@users.noreply.github.com>
Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
Adam Tazi <52357206+ad1tazi@users.noreply.github.com>
Adrien Gallouët <adrien@gallouet.fr>
Adrien Gallouët <angt@huggingface.co>
Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com>
AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
AidanBeltonS <aidan.belton@codeplay.com>
Akarshan Biswas <akarshan.biswas@gmail.com>
Akarshan Biswas <akarshanbiswas@fedoraproject.org>
Albert Jin <albert.jin@gmail.com>
Alberto Cabrera Pérez <alberto.cabrera@codeplay.com>
Alberto Cabrera Pérez <alberto.cabrera@intel.com>
Alex Azarov <alex@azarov.by>
Alex O'Connell <35843486+acon96@users.noreply.github.com>
Alex von Gluck IV <kallisti5@unixzen.com>
AmbientL <107641468+AmbientL@users.noreply.github.com>
AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
Ananta Bastola <anantarajbastola@gmail.com>
Andreas (Andi) Kunar <andreask@msn.com>
Andreas Kieslinger <47689530+aendk@users.noreply.github.com>
Andrei <abetlen@gmail.com>
Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com>
Andrii Ryzhkov <andriiryzhkov@users.noreply.github.com>
Arjun <ccldarjun@icloud.com>
Ashraful Islam <ashraful.meche@gmail.com>
Astariul <43774355+astariul@users.noreply.github.com>
AsukaMinato <asukaminato@nyan.eu.org>
Avi Lumelsky <avilume@gmail.com>
Bart Pelle <3662930+Velocity-@users.noreply.github.com>
Ben Ashbaugh <ben.ashbaugh@intel.com>
Bernhard M. Wiedemann <githubbmwprimary@lsmod.de>
Borislav Stanimirov <b.stanimirov@abv.bg>
Brad Ito <phlogisticfugu@users.noreply.github.com>
Brad Murray <59848399+bradmurray-dt@users.noreply.github.com>
Brian <mofosyne@gmail.com>
Bryan Lozano <b.lozano.havoc@gmail.com>
Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
CarterLi999 <664681047@qq.com>
Cebtenzzre <cebtenzzre@gmail.com>
Changyeon Kim <cyzero.kim@samsung.com>
Charles Xu <63788048+chaxu01@users.noreply.github.com>
Charles Xu <charles.xu@arm.com>
Chen Xi <xi2.chen@intel.com>
Chen Xi <xixichen08@foxmail.com>
Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
Chris Elrod <elrodc@gmail.com>
Christian Kastner <ckk@kvr.at>
Clint Herron <hanclinto@gmail.com>
Conrad Kramer <conrad@conradkramer.com>
Cordeiro <1471463+ocordeiro@users.noreply.github.com>
Cristiano Calcagno <cristianoc@users.noreply.github.com>
DAN™ <dranger003@gmail.com>
Dan Forbes <dan@danforbes.dev>
Dan Johansson <164997844+eddnjjn@users.noreply.github.com>
Dan Johansson <dan.johansson@arm.com>
Daniel Bevenius <daniel.bevenius@gmail.com>
Daniel Ziegenberg <daniel@ziegenberg.at>
Daniele <57776841+daniandtheweb@users.noreply.github.com>
Daulet Zhanguzin <daulet@users.noreply.github.com>
Dave <dave-fl@users.noreply.github.com>
Dave Airlie <airlied@gmail.com>
Dave Airlie <airlied@redhat.com>
David Miller <david@patagona.ca>
DavidKorczynski <david@adalogics.com>
Davidson Francis <davidsondfgl@gmail.com>
Dibakar Gope <dibakar.gope@arm.com>
Didzis Gosko <didzis@users.noreply.github.com>
Diego Devesa <slarengh@gmail.com>
Diogo <dgcruz983@gmail.com>
Djip007 <3705339+Djip007@users.noreply.github.com>
Djip007 <djip.perois@free.fr>
Dou Xinpeng <15529241576@163.com>
Dou Xinpeng <81913537+Dou-Git@users.noreply.github.com>
Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
Ebey Abraham <ebey97@gmail.com>
Eldar Yusupov <eyusupov@gmail.com>
Emmanuel Durand <emmanueldurand@protonmail.com>
Engininja2 <139037756+Engininja2@users.noreply.github.com>
Eric Zhang <34133756+EZForever@users.noreply.github.com>
Erik Scholz <Green-Sky@users.noreply.github.com>
Ettore Di Giacinto <mudler@users.noreply.github.com>
Eve <139727413+netrunnereve@users.noreply.github.com>
F1L1P <78918286+F1L1Pv2@users.noreply.github.com>
Faisal Zaghloul <quic_fzaghlou@quicinc.com>
FantasyGmm <16450052+FantasyGmm@users.noreply.github.com>
Felix <stenbackfelix@gmail.com>
Finn Voorhees <finnvoorhees@gmail.com>
FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
Frankie Robertson <frankier@users.noreply.github.com>
GainLee <perfecter.gen@gmail.com>
George Hindle <george@georgehindle.com>
Georgi Gerganov <ggerganov@gmail.com>
Gilad S <7817232+giladgd@users.noreply.github.com>
Gilad S <giladgd@users.noreply.github.com>
Gilad S. <7817232+giladgd@users.noreply.github.com>
Guillaume Wenzek <gwenzek@users.noreply.github.com>
Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
Haus1 <haus.xda@gmail.com>
Herman Semenov <GermanAizek@yandex.ru>
HimariO <dsfhe49854@gmail.com>
Hirochika Matsumoto <git@hkmatsumoto.com>
Hong Bo PENG <penghb@cn.ibm.com>
Hugo Rosenkranz-Costa <hugo.rosenkranz@gmail.com>
Hyunsung Lee <ita9naiwa@gmail.com>
IGUILIZ Salah-Eddine <76955987+salahiguiliz@users.noreply.github.com>
Ian Bull <irbull@eclipsesource.com>
Ihar Hrachyshka <ihrachys@redhat.com>
Ikko Eltociear Ashimine <eltociear@gmail.com>
Ivan <nekotekina@gmail.com>
Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
Ivan Stepanov <ivanstepanovftw@gmail.com>
Ivan Zdane <accounts@ivanzdane.com>
Jack Mousseau <jmousseau@users.noreply.github.com>
Jack Vial <vialjack@gmail.com>
JacobLinCool <jacoblincool@gmail.com>
Jakob Frick <jakob.maria.frick@gmail.com>
Jan Ploski <jpl@plosquare.com>
Jared Van Bortel <jared@nomic.ai>
Jeff Bolz <jbolz@nvidia.com>
Jeffrey Quesnelle <jquesnelle@gmail.com>
Jeroen Mostert <jeroen.mostert@cm.com>
Jiahao Li <liplus17@163.com>
JidongZhang-THU <1119708529@qq.com>
Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
Jo Liss <joliss42@gmail.com>
Joe Todd <joe.todd@codeplay.com>
Johannes Gäßler <johannesg@5d6.de>
John Balis <phobossystems@gmail.com>
Josh Bleecher Snyder <josharian@gmail.com>
Judd <foldl@users.noreply.github.com>
Jun Hee Yoo <contact.jhyoo@gmail.com>
Junil Kim <logyourself@gmail.com>
Justina Cho <justcho5@gmail.com>
Justine Tunney <jtunney@gmail.com>
Justine Tunney <jtunney@mozilla.com>
Karol Kontny <82021046+kkontny@users.noreply.github.com>
Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Kevin Gibbons <bakkot@gmail.com>
Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
Kylin <56434533+KyL0N@users.noreply.github.com>
LoganDark <git@logandark.mozmail.com>
LoganDark <github@logandark.mozmail.com>
LostRuins <39025047+LostRuins@users.noreply.github.com>
Lukas Möller <mail@lukas-moeller.ch>
M Refi D.A <24388107+refinism@users.noreply.github.com>
M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
Ma Mingfei <mingfei.ma@intel.com>
Mahesh Madhav <67384846+heshpdx@users.noreply.github.com>
MaiHD <maihd.dev@gmail.com>
Mark Zhuang <zhuangqiubin@gmail.com>
Markus Tavenrath <mtavenrath@users.noreply.github.com>
Masaya, Kato <62578291+msy-kato@users.noreply.github.com>
Mathieu Baudier <mbaudier@argeo.org>
Mathijs de Bruin <mathijs@mathijsfietst.nl>
Matt Stephenson <mstephenson6@users.noreply.github.com>
Max Krasnyansky <max.krasnyansky@gmail.com>
Max Krasnyansky <quic_maxk@quicinc.com>
Mayank Kumar Pal <mynkpl1998@gmail.com>
Meng, Hengyu <hengyu.meng@intel.com>
Mengqing Cao <cmq0113@163.com>
Metal Whale <45712559+metalwhale@users.noreply.github.com>
Michael Klimenko <mklimenko29@gmail.com>
Michael Podvitskiy <podvitskiymichael@gmail.com>
Michael Verrilli <msv@pobox.com>
Molly Sophia <mollysophia379@gmail.com>
Natsu <chino@hotococoa.moe>
Neo Zhang <14088817+arthw@users.noreply.github.com>
Neo Zhang Jianyu <jianyu.zhang@intel.com>
Neuman Vong <neuman.vong@gmail.com>
Nevin <nevinpuri1901@gmail.com>
Nicholai Tukanov <nicholaitukanov@gmail.com>
Nico Bosshard <nico@bosshome.ch>
Nicolò Scipione <nicolo.scipione@codeplay.com>
Nikita Sarychev <42014488+sARY77@users.noreply.github.com>
Nouamane Tazi <nouamane98@gmail.com>
Olivier Chafik <ochafik@google.com>
Olivier Chafik <ochafik@users.noreply.github.com>
Ondřej Čertík <ondrej@certik.us>
Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
PAB <pierreantoine.bannier@gmail.com>
Paul Tsochantaris <ptsochantaris@icloud.com>
Peter <peter277@users.noreply.github.com>
Philpax <me@philpax.me>
Pierre Alexandre SCHEMBRI <pa.schembri@gmail.com>
Plamen Minev <pacominev@gmail.com>
Playdev <josang1204@gmail.com>
Prashant Vithule <119530321+Vithulep@users.noreply.github.com>
Przemysław Pawełczyk <przemoc@gmail.com>
R0CKSTAR <xiaodong.ye@mthreads.com>
R0CKSTAR <yeahdongcn@gmail.com>
Radoslav Gerganov <rgerganov@gmail.com>
Radosław Gryta <radek.gryta@gmail.com>
Ravindra Marella <marella@users.noreply.github.com>
Ray Cromwell <cromwellian@gmail.com>
Reinforce-II <fate@eastal.com>
Rémy Oudompheng <oudomphe@phare.normalesup.org>
Reza Rezvan <reza@rezvan.xyz>
Rick G <26732651+TheFlipbook@users.noreply.github.com>
RiverZhou <riverzhou2000@gmail.com>
Robert Ormandi <52251610+ormandi@users.noreply.github.com>
Romain Biessy <romain.biessy@codeplay.com>
Ronsor <ronsor@ronsor.pw>
Rotem Dan <rotemdan@gmail.com>
Ryan Hitchman <hitchmanr@gmail.com>
SRHMorris <69468379+SRHMorris@users.noreply.github.com>
SXX <sxx1136965276@gmail.com>
Salvatore Mesoraca <s.mesoraca16@gmail.com>
Sam Spilsbury <smspillaz@gmail.com>
Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
Santtu Keskinen <santtu.keskinen@gmail.com>
Sergio López <slp@redhat.com>
Sergio López <slp@sinrega.org>
Shanshan Shen <467638484@qq.com>
Shijie <821898965@qq.com>
Shupei Fan <dymarkfan@outlook.com>
Siddharth Ramakrishnan <srr2141@columbia.edu>
Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Skyler Celestinian-Sterling <80314197+Celestinian@users.noreply.github.com>
Slava Primenko <primenko.s@gmail.com>
Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com>
Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
Supreet Sethi <supreet.sethi@gmail.com>
Takuya Takeuchi <takuya.takeuchi.dev@gmail.com>
Tamotsu Takahashi <ttakah+github@gmail.com>
Tanmay <tnmysachan@gmail.com>
Tanmay Sachan <tnmysachan@gmail.com>
Timothy Cronin <40186632+4imothy@users.noreply.github.com>
Tom Bailey <tombailey@users.noreply.github.com>
Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Tony Wasserka <4840017+neobrain@users.noreply.github.com>
Tristan Druyen <tristan@vault81.mozmail.com>
Tyé singwa <92231658+tye-singwa@users.noreply.github.com>
UEXTM.com <84163508+uextm@users.noreply.github.com>
WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com>
William Tambellini <william.tambellini@gmail.com>
William Tambellini <wtambellini@sdl.com>
XiaotaoChen <chenxiaotao1234@gmail.com>
Xinpeng Dou <81913537+Dou-Git@users.noreply.github.com>
Xuan Son Nguyen <thichthat@gmail.com>
Yavor Ivanov <yivanov@viewray.com>
YavorGIvanov <yivanov@viewray.com>
Yilong Guo <vfirst218@gmail.com>
Yilong Guo <yilong.guo@intel.com>
Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
Zhenwei Jin <109658203+kylo5aby@users.noreply.github.com>
Zhiyuan Li <lizhiyuan@uniartisan.com>
Zhiyuan Li <uniartisan2017@gmail.com>
a3sh <38979186+A3shTnT@users.noreply.github.com>
ag2s20150909 <19373730+ag2s20150909@users.noreply.github.com>
agray3 <agray3@users.noreply.github.com>
amd-dwang <dong.wang@amd.com>
amritahs-ibm <amritahs@linux.vnet.ibm.com>
apcameron <37645737+apcameron@users.noreply.github.com>
appvoid <78444142+appvoid@users.noreply.github.com>
ariez-xyz <41232910+ariez-xyz@users.noreply.github.com>
automaticcat <daogiatuank54@gmail.com>
bandoti <141645996+bandoti@users.noreply.github.com>
bmwl <brian.marshall@tolko.com>
bobqianic <129547291+bobqianic@users.noreply.github.com>
bssrdf <merlintiger@hotmail.com>
chengchi <davesjoewang@gmail.com>
compilade <113953597+compilade@users.noreply.github.com>
compilade <git@compilade.net>
ddpasa <112642920+ddpasa@users.noreply.github.com>
denersc <denerstassun@gmail.com>
dscripka <dscripka@users.noreply.github.com>
fitzsim <fitzsim@fitzsim.org>
fj-y-saito <85871716+fj-y-saito@users.noreply.github.com>
fraxy-v <65565042+fraxy-v@users.noreply.github.com>
gn64 <yukikaze.jp@gmail.com>
goerch <jhr.walter@t-online.de>
goldwaving <77494627+goldwaving@users.noreply.github.com>
haopeng <657407891@qq.com>
hidenorly <hidenorly@users.noreply.github.com>
hipudding <huafengchun@gmail.com>
hydai <z54981220@gmail.com>
issixx <46835150+issixx@users.noreply.github.com>
jaeminSon <woalsdnd@gmail.com>
jdomke <28772296+jdomke@users.noreply.github.com>
jiez <373447296@qq.com>
johnson442 <56517414+johnson442@users.noreply.github.com>
junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
k.h.lai <adrian.k.h.lai@outlook.com>
katsu560 <118887472+katsu560@users.noreply.github.com>
klosax <131523366+klosax@users.noreply.github.com>
kunnis <kunnis@users.noreply.github.com>
l3utterfly <gc.pthzfoldr@gmail.com>
le.chang <cljs118@126.com>
leejet <31925346+leejet@users.noreply.github.com>
leejet <leejet714@gmail.com>
leo-pony <nengjunma@outlook.com>
lhez <quic_lih@quicinc.com>
liuwei-git <14815172+liuwei-git@users.noreply.github.com>
luoyu-intel <yu.luo@intel.com>
magicse <magicse@users.noreply.github.com>
mahorozte <41834471+mahorozte@users.noreply.github.com>
mashizora <30516315+mashizora@users.noreply.github.com>
matt23654 <matthew.webber@protonmail.com>
matteo <matteogeniaccio@yahoo.it>
ochafik <ochafik@google.com>
otaGran <ujt2h8@gmail.com>
pengxin99 <pengxin.yuan@intel.com>
pikalover6 <49179590+pikalover6@users.noreply.github.com>
postmasters <namnguyen@google.com>
sjinzh <sjinzh@gmail.com>
skirodev <57715494+skirodev@users.noreply.github.com>
slaren <slarengh@gmail.com>
snadampal <87143774+snadampal@users.noreply.github.com>
someone13574 <81528246+someone13574@users.noreply.github.com>
stduhpf <stephduh@live.fr>
taher <8665427+nullhook@users.noreply.github.com>
texmex76 <40733439+texmex76@users.noreply.github.com>
the-crypt-keeper <84680712+the-crypt-keeper@users.noreply.github.com>
thewh1teagle <61390950+thewh1teagle@users.noreply.github.com>
ucag.li <ucag@qq.com>
ulatekh <ulatekh@yahoo.com>
uvos <devnull@uvos.xyz>
uvos <philipp@uvos.xyz>
wangshuai09 <391746016@qq.com>
woachk <24752637+woachk@users.noreply.github.com>
xctan <axunlei@gmail.com>
yangyaofei <yangyaofei@gmail.com>
yuri@FreeBSD <yuri@FreeBSD>
zhentaoyu <zhentao.yu@intel.com>
zhouwg <6889919+zhouwg@users.noreply.github.com>
zhouwg <zhouwg2000@gmail.com>
谢乃闻 <sienaiwun@users.noreply.github.com>
布客飞龙 <562826179@qq.com>
旺旺碎冰冰 <38837039+Cyberhan123@users.noreply.github.com>
engine/ggml/CMakeLists.txt
+505
View File
@@ -0,0 +1,505 @@
cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
project("ggml" C CXX ASM)
### GGML Version
set(GGML_VERSION_MAJOR 0)
set(GGML_VERSION_MINOR 15)
set(GGML_VERSION_PATCH 2)
set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
if(GIT_EXE)
# Get current git commit hash
execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
OUTPUT_VARIABLE GGML_BUILD_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
ERROR_QUIET
)
# Check if the working directory is dirty (i.e., has uncommitted changes)
execute_process(COMMAND ${GIT_EXE} diff-index --quiet HEAD -- .
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
RESULT_VARIABLE GGML_GIT_DIRTY
ERROR_QUIET
)
endif()
set(GGML_VERSION "${GGML_VERSION_BASE}")
if(NOT GGML_BUILD_COMMIT)
set(GGML_BUILD_COMMIT "unknown")
endif()
# Build the commit string with optional dirty flag
if(DEFINED GGML_GIT_DIRTY AND GGML_GIT_DIRTY EQUAL 1)
set(GGML_BUILD_COMMIT "${GGML_BUILD_COMMIT}-dirty")
endif()
include(CheckIncludeFileCXX)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
endif()
if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
set(GGML_STANDALONE ON)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
# configure project version
# TODO
else()
set(GGML_STANDALONE OFF)
if (NOT CMAKE_RUNTIME_OUTPUT_DIRECTORY)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
endif()
endif()
if (EMSCRIPTEN)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
option(GGML_WASM_SINGLE_FILE "ggml: embed WASM inside the generated ggml.js" ON)
else()
if (MINGW)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
else()
set(BUILD_SHARED_LIBS_DEFAULT ON)
endif()
endif()
# remove the lib prefix on win32 mingw
if (WIN32)
set(CMAKE_STATIC_LIBRARY_PREFIX "")
set(CMAKE_SHARED_LIBRARY_PREFIX "")
set(CMAKE_SHARED_MODULE_PREFIX "")
endif()
option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
option(GGML_BACKEND_DL "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
#
# option list
#
# TODO: mark all options as advanced when not GGML_STANDALONE
if (APPLE)
set(GGML_METAL_DEFAULT ON)
set(GGML_BLAS_DEFAULT ON)
set(GGML_BLAS_VENDOR_DEFAULT "Apple")
else()
set(GGML_METAL_DEFAULT OFF)
set(GGML_BLAS_DEFAULT OFF)
set(GGML_BLAS_VENDOR_DEFAULT "Generic")
endif()
if (CMAKE_CROSSCOMPILING OR DEFINED ENV{SOURCE_DATE_EPOCH})
message(STATUS "Setting GGML_NATIVE_DEFAULT to OFF")
set(GGML_NATIVE_DEFAULT OFF)
else()
set(GGML_NATIVE_DEFAULT ON)
endif()
# defaults
if (NOT GGML_LLAMAFILE_DEFAULT)
set(GGML_LLAMAFILE_DEFAULT OFF)
endif()
if (NOT GGML_CUDA_GRAPHS_DEFAULT)
set(GGML_CUDA_GRAPHS_DEFAULT OFF)
endif()
# general
option(GGML_STATIC "ggml: static link libraries" OFF)
option(GGML_NATIVE "ggml: optimize the build for the current system" ${GGML_NATIVE_DEFAULT})
option(GGML_LTO "ggml: enable link time optimization" OFF)
option(GGML_CCACHE "ggml: use ccache if available" ON)
# debug
option(GGML_ALL_WARNINGS "ggml: enable all compiler warnings" ON)
option(GGML_ALL_WARNINGS_3RD_PARTY "ggml: enable all compiler warnings in 3rd party libs" OFF)
option(GGML_GPROF "ggml: enable gprof" OFF)
# build
option(GGML_FATAL_WARNINGS "ggml: enable -Werror flag" OFF)
# sanitizers
option(GGML_SANITIZE_THREAD "ggml: enable thread sanitizer" OFF)
option(GGML_SANITIZE_ADDRESS "ggml: enable address sanitizer" OFF)
option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
# instruction set specific
if (GGML_NATIVE OR NOT GGML_NATIVE_DEFAULT)
set(INS_ENB OFF)
else()
set(INS_ENB ON)
endif()
message(DEBUG "GGML_NATIVE : ${GGML_NATIVE}")
message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
message(DEBUG "INS_ENB : ${INS_ENB}")
option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
option(GGML_CPU_REPACK "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable" OFF)
option(GGML_SSE42 "ggml: enable SSE 4.2" ${INS_ENB})
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF)
option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB})
option(GGML_BMI2 "ggml: enable BMI2" ${INS_ENB})
option(GGML_AVX512 "ggml: enable AVX512F" OFF)
option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF)
option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF)
option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16" OFF)
if (NOT MSVC)
# in MSVC F16C and FMA is implied with AVX2/AVX512
option(GGML_FMA "ggml: enable FMA" ${INS_ENB})
option(GGML_F16C "ggml: enable F16C" ${INS_ENB})
# MSVC does not seem to support AMX
option(GGML_AMX_TILE "ggml: enable AMX-TILE" OFF)
option(GGML_AMX_INT8 "ggml: enable AMX-INT8" OFF)
option(GGML_AMX_BF16 "ggml: enable AMX-BF16" OFF)
endif()
option(GGML_LASX "ggml: enable lasx" ON)
option(GGML_LSX "ggml: enable lsx" ON)
option(GGML_RVV "ggml: enable rvv" ON)
option(GGML_RV_ZFH "ggml: enable riscv zfh" ON)
option(GGML_RV_ZVFH "ggml: enable riscv zvfh" ON)
option(GGML_RV_ZICBOP "ggml: enable riscv zicbop" ON)
option(GGML_RV_ZIHINTPAUSE "ggml: enable riscv zihintpause" ON)
option(GGML_RV_ZVFBFWMA "ggml: enable riscv zvfbfwma" OFF)
option(GGML_XTHEADVECTOR "ggml: enable xtheadvector" OFF)
option(GGML_VXE "ggml: enable vxe" ${GGML_NATIVE})
option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
# ggml core
set(GGML_SCHED_MAX_COPIES "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
option(GGML_CPU "ggml: enable CPU backend" ON)
option(GGML_SCHED_NO_REALLOC "ggml: disallow reallocations in ggml-alloc (for debugging)" OFF)
# 3rd party libs / backends
option(GGML_ACCELERATE "ggml: enable Accelerate framework" ON)
option(GGML_BLAS "ggml: use BLAS" ${GGML_BLAS_DEFAULT})
set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
"ggml: BLAS library vendor")
option(GGML_LLAMAFILE "ggml: use LLAMAFILE" ${GGML_LLAMAFILE_DEFAULT})
option(GGML_CUDA "ggml: use CUDA" OFF)
option(GGML_MUSA "ggml: use MUSA" OFF)
option(GGML_CUDA_FORCE_MMQ "ggml: use mmq kernels instead of cuBLAS" OFF)
option(GGML_CUDA_FORCE_CUBLAS "ggml: always use cuBLAS instead of mmq kernels" OFF)
set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"ggml: max. batch size for using peer access")
option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF)
option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF)
option(GGML_CUDA_FA "ggml: compile ggml FlashAttention CUDA kernels" ON)
option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF)
option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT})
option(GGML_CUDA_NCCL "ggml: use NVIDIA Collective Comm. Library" ON)
set (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING
"ggml: cuda link binary compression mode; requires cuda 12.8+")
set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size")
option(GGML_HIP "ggml: use HIP" OFF)
option(GGML_HIP_GRAPHS "ggml: use HIP graph" ON)
option(GGML_HIP_RCCL "ggml: use ROCm Collective Comm. Library" OFF)
option(GGML_HIP_NO_VMM "ggml: do not try to use HIP VMM" ON)
option(GGML_HIP_ROCWMMA_FATTN "ggml: enable rocWMMA for FlashAttention" OFF)
option(GGML_HIP_MMQ_MFMA "ggml: enable MFMA MMA for CDNA in MMQ" ON)
option(GGML_HIP_EXPORT_METRICS "ggml: enable kernel perf metrics output" OFF)
option(GGML_MUSA_GRAPHS "ggml: use MUSA graph, experimental, unstable" OFF)
option(GGML_MUSA_MUDNN_COPY "ggml: enable muDNN for accelerated copy" OFF)
option(GGML_VULKAN "ggml: use Vulkan" OFF)
option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF)
option(GGML_VULKAN_DEBUG "ggml: enable Vulkan debug output" OFF)
option(GGML_VULKAN_MEMORY_DEBUG "ggml: enable Vulkan memory debug output" OFF)
option(GGML_VULKAN_SHADER_DEBUG_INFO "ggml: enable Vulkan shader debug info" OFF)
option(GGML_VULKAN_VALIDATE "ggml: enable Vulkan validation" OFF)
option(GGML_VULKAN_RUN_TESTS "ggml: run Vulkan tests" OFF)
option(GGML_WEBGPU "ggml: use WebGPU" OFF)
option(GGML_WEBGPU_DEBUG "ggml: enable WebGPU debug output" OFF)
option(GGML_WEBGPU_CPU_PROFILE "ggml: enable WebGPU profiling (CPU)" OFF)
option(GGML_WEBGPU_GPU_PROFILE "ggml: enable WebGPU profiling (GPU)" OFF)
option(GGML_WEBGPU_JSPI "ggml: use JSPI for WebGPU" ON)
option(GGML_ZDNN "ggml: use zDNN" OFF)
option(GGML_VIRTGPU "ggml: use the VirtGPU/Virglrenderer API Remoting frontend" OFF)
option(GGML_VIRTGPU_BACKEND "ggml: build the VirtGPU/Virglrenderer API Remoting backend" OFF)
option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT})
option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF)
option(GGML_METAL_SHADER_DEBUG "ggml: compile Metal with -fno-fast-math" OFF)
option(GGML_METAL_EMBED_LIBRARY "ggml: embed Metal library" ${GGML_METAL})
set (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
"ggml: metal minimum macOS version")
set (GGML_METAL_STD "" CACHE STRING "ggml: metal standard version (-std flag)")
option(GGML_OPENMP "ggml: use OpenMP" ON)
option(GGML_RPC "ggml: use RPC" OFF)
option(GGML_SYCL "ggml: use SYCL" OFF)
option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl calculations" OFF)
option(GGML_SYCL_GRAPH "ggml: enable graphs in the SYCL backend" ON)
option(GGML_SYCL_HOST_MEM_FALLBACK "ggml: allow host memory fallback in SYCL reorder (requires kernel 6.8+)" ON)
option(GGML_SYCL_SUPPORT_LEVEL_ZERO_API "ggml: use Level Zero API in SYCL backend" ON)
option(GGML_SYCL_DNN "ggml: enable oneDNN in the SYCL backend" ON)
set (GGML_SYCL_TARGET "INTEL" CACHE STRING
"ggml: sycl target device")
set (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
"ggml: sycl device architecture")
option(GGML_OPENVINO "ggml: use OPENVINO" OFF)
option(GGML_OPENCL "ggml: use OpenCL" OFF)
option(GGML_OPENCL_PROFILING "ggml: use OpenCL profiling (increases overhead)" OFF)
option(GGML_OPENCL_EMBED_KERNELS "ggml: embed kernels" ON)
option(GGML_OPENCL_USE_ADRENO_KERNELS "ggml: use optimized kernels for Adreno" ON)
set (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
"ggml: OpenCL API version to target")
option(GGML_HEXAGON "ggml: enable Hexagon backend" OFF)
set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml: quantize group size (32, 64, or 128)")
# toolchain for vulkan-shaders-gen
set (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
option(GGML_ZENDNN "ggml: use ZenDNN" OFF)
option(ZENDNN_ROOT "ggml: path to ZenDNN installation" "")
# extra artifacts
option(GGML_BUILD_TESTS "ggml: build tests" ${GGML_STANDALONE})
option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
#
# dependencies
#
set(CMAKE_C_STANDARD 11)
set(CMAKE_C_STANDARD_REQUIRED true)
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED true)
set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
include(GNUInstallDirs)
#
# build the library
#
add_subdirectory(src)
#
# tests and examples
#
if (GGML_BUILD_TESTS)
enable_testing()
add_subdirectory(tests)
endif ()
if (GGML_BUILD_EXAMPLES)
add_subdirectory(examples)
endif ()
#
# install
#
include(CMakePackageConfigHelpers)
# all public headers
set(GGML_PUBLIC_HEADERS
include/ggml.h
include/ggml-cpu.h
include/ggml-alloc.h
include/ggml-backend.h
include/ggml-blas.h
include/ggml-cann.h
include/ggml-cpp.h
include/ggml-cuda.h
include/ggml-opt.h
include/ggml-metal.h
include/ggml-rpc.h
include/ggml-virtgpu.h
include/ggml-sycl.h
include/ggml-vulkan.h
include/ggml-webgpu.h
include/ggml-zendnn.h
include/ggml-openvino.h
include/gguf.h)
set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
#if (GGML_METAL)
# set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
#endif()
install(TARGETS ggml LIBRARY PUBLIC_HEADER)
install(TARGETS ggml-base LIBRARY)
if (GGML_STANDALONE)
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
@ONLY)
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
endif()
#
# Create CMake package
#
# Capture variables prefixed with GGML_.
set(variable_set_statements
"
####### Expanded from @GGML_VARIABLES_EXPANED@ by configure_package_config_file() #######
####### Any changes to this file will be overwritten by the next CMake run #######
")
set(GGML_SHARED_LIB ${BUILD_SHARED_LIBS})
get_cmake_property(all_variables VARIABLES)
foreach(variable_name IN LISTS all_variables)
if(variable_name MATCHES "^GGML_")
string(REPLACE ";" "\\;"
variable_value "${${variable_name}}")
set(variable_set_statements
"${variable_set_statements}set(${variable_name} \"${variable_value}\")\n")
endif()
endforeach()
set(GGML_VARIABLES_EXPANDED ${variable_set_statements})
# Create the CMake package and set install location.
set(GGML_INSTALL_VERSION ${GGML_VERSION})
set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header files")
set(GGML_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
set(GGML_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")
configure_package_config_file(
${CMAKE_CURRENT_SOURCE_DIR}/cmake/ggml-config.cmake.in
${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml
PATH_VARS GGML_INCLUDE_INSTALL_DIR
GGML_LIB_INSTALL_DIR
GGML_BIN_INSTALL_DIR)
write_basic_package_version_file(
${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
VERSION ${GGML_INSTALL_VERSION}
COMPATIBILITY SameMajorVersion)
target_compile_definitions(ggml-base PRIVATE
GGML_VERSION="${GGML_INSTALL_VERSION}"
GGML_COMMIT="${GGML_BUILD_COMMIT}"
)
message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
message(STATUS "ggml commit: ${GGML_BUILD_COMMIT}")
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
if (MSVC)
set(MSVC_WARNING_FLAGS
/wd4005 # Macro redefinition
/wd4244 # Conversion from one type to another type, possible loss of data
/wd4267 # Conversion from 'size_t' to a smaller type, possible loss of data
/wd4305 # Conversion from 'type1' to 'type2', possible loss of data
/wd4566 # Conversion from 'char' to 'wchar_t', possible loss of data
/wd4996 # Disable POSIX deprecation warnings
/wd4702 # Unreachable code warnings
)
set(MSVC_COMPILE_OPTIONS
"$<$<COMPILE_LANGUAGE:C>:/utf-8>"
"$<$<COMPILE_LANGUAGE:CXX>:/utf-8>"
)
function(configure_msvc_target target_name)
if(TARGET ${target_name})
target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
target_compile_options(${target_name} PRIVATE ${MSVC_COMPILE_OPTIONS})
endif()
endfunction()
configure_msvc_target(ggml-base)
configure_msvc_target(ggml)
configure_msvc_target(ggml-cpu)
configure_msvc_target(ggml-cpu-x64)
configure_msvc_target(ggml-cpu-sse42)
configure_msvc_target(ggml-cpu-sandybridge)
# __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
# skipping ggml-cpu-ivybridge
# skipping ggml-cpu-piledriver
configure_msvc_target(ggml-cpu-haswell)
configure_msvc_target(ggml-cpu-skylakex)
configure_msvc_target(ggml-cpu-cannonlake)
configure_msvc_target(ggml-cpu-cascadelake)
configure_msvc_target(ggml-cpu-icelake)
# MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
# https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
# https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
# skipping ggml-cpu-cooperlake
# skipping ggml-cpu-zen4
configure_msvc_target(ggml-cpu-alderlake)
# MSVC doesn't support AMX
# skipping ggml-cpu-sapphirerapids
if (GGML_BUILD_EXAMPLES)
configure_msvc_target(common-ggml)
configure_msvc_target(common)
configure_msvc_target(mnist-common)
configure_msvc_target(mnist-eval)
configure_msvc_target(mnist-train)
configure_msvc_target(gpt-2-ctx)
configure_msvc_target(gpt-2-alloc)
configure_msvc_target(gpt-2-backend)
configure_msvc_target(gpt-2-sched)
configure_msvc_target(gpt-2-quantize)
configure_msvc_target(gpt-2-batched)
configure_msvc_target(gpt-j)
configure_msvc_target(gpt-j-quantize)
configure_msvc_target(magika)
configure_msvc_target(yolov3-tiny)
configure_msvc_target(sam)
configure_msvc_target(simple-ctx)
configure_msvc_target(simple-backend)
endif()
if (GGML_BUILD_TESTS)
configure_msvc_target(test-mul-mat)
configure_msvc_target(test-arange)
configure_msvc_target(test-backend-ops)
configure_msvc_target(test-cont)
configure_msvc_target(test-conv-transpose)
configure_msvc_target(test-conv-transpose-1d)
configure_msvc_target(test-conv1d)
configure_msvc_target(test-conv2d)
configure_msvc_target(test-conv2d-dw)
configure_msvc_target(test-customop)
configure_msvc_target(test-dup)
configure_msvc_target(test-opt)
configure_msvc_target(test-pool)
endif ()
endif()
engine/ggml/CONTRIBUTING.md
+3
View File
@@ -0,0 +1,3 @@
Please use [llama.cpp's contribution guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md) for this project.
*For changes to the core `ggml` library (including to the CMake build system), please open a PR in https://github.com/ggml-org/llama.cpp. Doing so will make your PR more visible, better tested and more likely to be reviewed.*
engine/ggml/LICENSE
+21
View File
@@ -0,0 +1,21 @@
MIT License
Copyright (c) 2023-2026 The ggml authors
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
engine/ggml/README.md
+50
View File
@@ -0,0 +1,50 @@
# ggml
[Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205)
Tensor library for machine learning
***Note that this project is under active development. \
Some of the development is currently happening in the [llama.cpp](https://github.com/ggerganov/llama.cpp) and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) repos***
## Features
- Low-level cross-platform implementation
- Integer quantization support
- Broad hardware support
- Automatic differentiation
- ADAM and L-BFGS optimizers
- No third-party dependencies
- Zero memory allocations during runtime
## Build
```bash
git clone https://github.com/ggml-org/ggml
cd ggml
# install python dependencies in a virtual environment
python3.10 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# build the examples
mkdir build && cd build
cmake ..
cmake --build . --config Release -j 8
```
## GPT inference (example)
```bash
# run the GPT-2 small 117M model
../examples/gpt-2/download-ggml-model.sh 117M
./bin/gpt-2-backend -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
```
For more information, checkout the corresponding programs in the [examples](examples) folder.
## Resources
- [Introduction to ggml](https://huggingface.co/blog/introduction-to-ggml)
- [The GGUF file format](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md)
engine/ggml/ci/run.sh
+395
View File
@@ -0,0 +1,395 @@
#/bin/bash
#
# sample usage:
#
# mkdir tmp
#
# # CPU-only build
# bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# # with CUDA support
# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# # With SYCL support
# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
if [ -z "$2" ]; then
echo "usage: $0 <output-dir> <mnt-dir>"
exit 1
fi
mkdir -p "$1"
mkdir -p "$2"
OUT=$(realpath "$1")
MNT=$(realpath "$2")
rm -v $OUT/*.log
rm -v $OUT/*.exit
rm -v $OUT/*.md
sd=`dirname $0`
cd $sd/../
SRC=`pwd`
CMAKE_EXTRA=""
CTEST_EXTRA=""
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
if command -v nvidia-smi >/dev/null 2>&1; then
CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
else
echo "Warning: Using fallback CUDA architectures"
CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
fi
else
echo "Error: nvidia-smi not found, cannot build with CUDA"
exit 1
fi
fi
if [ ! -z ${GG_BUILD_ROCM} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
fi
if [ ! -z ${GG_BUILD_SYCL} ]; then
if [ -z ${ONEAPI_ROOT} ]; then
echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
echo "source /opt/intel/oneapi/setvars.sh"
exit 1
fi
# Use only main GPU
export ONEAPI_DEVICE_SELECTOR="level_zero:0"
# Enable sysman for correct memory reporting
export ZES_ENABLE_SYSMAN=1
# to circumvent precision issues on CPY operations
export SYCL_PROGRAM_COMPILE_OPTIONS="-cl-fp32-correctly-rounded-divide-sqrt"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
fi
if [ ! -z ${GG_BUILD_VULKAN} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
# if on Mac, disable METAL
if [[ "$OSTYPE" == "darwin"* ]]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
fi
fi
if [ ! -z ${GG_BUILD_WEBGPU} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
fi
if [ ! -z ${GG_BUILD_MUSA} ]; then
# Use qy1 by default (MTT S80)
MUSA_ARCH=${MUSA_ARCH:-21}
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
fi
if [ ! -z ${GG_BUILD_NO_SVE} ]; then
# arm 9 and newer enables sve by default, adjust these flags depending on the cpu used
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
fi
## helpers
# download a file if it does not exist or if it is outdated
function gg_wget {
local out=$1
local url=$2
local cwd=`pwd`
mkdir -p $out
cd $out
# should not re-download if file is the same
wget -nv -N $url
cd $cwd
}
function gg_printf {
printf -- "$@" >> $OUT/README.md
}
function gg_run {
ci=$1
set -o pipefail
set -x
gg_run_$ci | tee $OUT/$ci.log
cur=$?
echo "$cur" > $OUT/$ci.exit
set +x
set +o pipefail
gg_sum_$ci
ret=$((ret | cur))
}
## ci
# ctest_debug
function gg_run_ctest_debug {
cd ${SRC}
rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
set -e
(time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
(time ctest ${CTEST_EXTRA} --output-on-failure -E "test-opt|test-backend-ops" ) 2>&1 | tee -a $OUT/${ci}-ctest.log
set +e
}
function gg_sum_ctest_debug {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Runs ctest in debug mode\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
gg_printf '```\n'
gg_printf '\n'
}
# ctest_release
function gg_run_ctest_release {
cd ${SRC}
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
if [ -z $GG_BUILD_LOW_PERF ]; then
(time ctest ${CTEST_EXTRA} --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
else
(time ctest ${CTEST_EXTRA} --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
fi
set +e
}
function gg_sum_ctest_release {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Runs ctest in release mode\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
gg_printf '```\n'
}
# gpt_2
function gg_run_gpt_2 {
cd ${SRC}
gg_wget models-mnt/gpt-2 https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
cd build-ci-release
set -e
model="../models-mnt/gpt-2/ggml-model-gpt-2-117M.bin"
prompts="../examples/prompts/gpt-2.txt"
(time ./bin/gpt-2-backend --model ${model} -s 1234 -n 64 -tt ${prompts} ) 2>&1 | tee -a $OUT/${ci}-tg.log
(time ./bin/gpt-2-backend --model ${model} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
(time ./bin/gpt-2-sched --model ${model} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
(time ./bin/gpt-2-batched --model ${model} -s 1234 -n 64 -np 8 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
set +e
}
function gg_sum_gpt_2 {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Runs short GPT-2 text generation\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-tg.log)"
gg_printf '```\n'
}
# TODO: update
## mnist
#
#function gg_run_mnist {
# cd ${SRC}
#
# cd build-ci-release
#
# set -e
#
# mkdir -p models/mnist
# python3 ../examples/mnist/convert-h5-to-ggml.py ../examples/mnist/models/mnist/mnist_model.state_dict
#
# model_f32="./models/mnist/ggml-model-f32.bin"
# samples="../examples/mnist/models/mnist/t10k-images.idx3-ubyte"
#
# # first command runs and exports "mnist.ggml", the second command runs the exported model
#
# (time ./bin/mnist ${model_f32} ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
# (time ./bin/mnist-cpu ./mnist.ggml ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
#
# set +e
#}
#
#function gg_sum_mnist {
# gg_printf '### %s\n\n' "${ci}"
#
# gg_printf 'MNIST\n'
# gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
# gg_printf '```\n'
# gg_printf '%s\n' "$(cat $OUT/${ci}-mnist.log)"
# gg_printf '```\n'
#}
# sam
function gg_run_sam {
cd ${SRC}
gg_wget models-mnt/sam/ https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
gg_wget models-mnt/sam/ https://raw.githubusercontent.com/YavorGIvanov/sam.cpp/ceafb7467bff7ec98e0c4f952e58a9eb8fd0238b/img.jpg
cd build-ci-release
set -e
path_models="../models-mnt/sam/"
model_f16="${path_models}/ggml-model-f16.bin"
img_0="${path_models}/img.jpg"
python3 ../examples/sam/convert-pth-to-ggml.py ${path_models}/sam_vit_b_01ec64.pth ${path_models}/ 1
# Test default parameters
(time ./bin/sam -m ${model_f16} -i ${img_0} -st 0.925 ) 2>&1 | tee -a $OUT/${ci}-main.log
grep -q "point prompt" $OUT/${ci}-main.log
grep -q "bbox (371, 436), (144, 168)" $OUT/${ci}-main.log ||
grep -q "bbox (370, 439), (144, 168)" $OUT/${ci}-main.log
# Test box prompt and single mask output
(time ./bin/sam -m ${model_f16} -i ${img_0} -st 0.925 -b 368,144,441,173 -sm) 2>&1 | tee -a $OUT/${ci}-main.log
grep -q "box prompt" $OUT/${ci}-main.log
grep -q "bbox (370, 439), (144, 169)" $OUT/${ci}-main.log ||
grep -q "bbox (370, 439), (144, 168)" $OUT/${ci}-main.log
set +e
}
function gg_sum_sam {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Run SAM\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-main.log)"
gg_printf '```\n'
}
# yolo
function gg_run_yolo {
cd ${SRC}
gg_wget models-mnt/yolo/ https://huggingface.co/ggml-org/models/resolve/main/yolo/yolov3-tiny.weights
gg_wget models-mnt/yolo/ https://huggingface.co/ggml-org/models/resolve/main/yolo/dog.jpg
cd build-ci-release
cp -r ../examples/yolo/data .
set -e
path_models="../models-mnt/yolo/"
python3 ../examples/yolo/convert-yolov3-tiny.py ${path_models}/yolov3-tiny.weights
(time ./bin/yolov3-tiny -m yolov3-tiny.gguf -i ${path_models}/dog.jpg ) 2>&1 | tee -a $OUT/${ci}-main.log
grep -qE "dog: (55|56|57|58|59)%" $OUT/${ci}-main.log
grep -qE "car: (50|51|52|53|54)%" $OUT/${ci}-main.log
grep -qE "truck: (54|55|56|57|58)%" $OUT/${ci}-main.log
grep -qE "bicycle: (57|58|59|60|61)%" $OUT/${ci}-main.log
set +e
}
function gg_sum_yolo {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Run YOLO\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-main.log)"
gg_printf '```\n'
}
## main
if true ; then
# Create symlink: ./ggml/models-mnt -> $MNT/models/models-mnt
rm -rf ${SRC}/models-mnt
mnt_models=${MNT}/models
mkdir -p ${mnt_models}
ln -sfn ${mnt_models} ${SRC}/models-mnt
# Create a fresh python3 venv and enter it
if ! python3 -m venv "$MNT/venv"; then
echo "Error: Failed to create Python virtual environment at $MNT/venv."
exit 1
fi
source "$MNT/venv/bin/activate"
pip install -r ${SRC}/requirements.txt --disable-pip-version-check
fi
ret=0
test $ret -eq 0 && gg_run ctest_debug
test $ret -eq 0 && gg_run ctest_release
test $ret -eq 0 && gg_run gpt_2
#test $ret -eq 0 && gg_run mnist
test $ret -eq 0 && gg_run sam
test $ret -eq 0 && gg_run yolo
if [ -z $GG_BUILD_LOW_PERF ]; then
# run tests meant for low-perf runners
date
fi
cat $OUT/README.md
exit $ret
engine/ggml/cmake/FindNCCL.cmake
+36
View File
@@ -0,0 +1,36 @@
# cmake/FindNCCL.cmake
# NVIDIA does not distribute CMake files with NCCl, therefore use this file to find it instead.
find_path(NCCL_INCLUDE_DIR
NAMES nccl.h
HINTS ${NCCL_ROOT} $ENV{NCCL_ROOT} $ENV{CUDA_HOME} /usr/local/cuda
PATH_SUFFIXES include
)
find_library(NCCL_LIBRARY
NAMES nccl
HINTS ${NCCL_ROOT} $ENV{NCCL_ROOT} $ENV{CUDA_HOME} /usr/local/cuda
PATH_SUFFIXES lib lib64
)
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(NCCL
DEFAULT_MSG
NCCL_LIBRARY NCCL_INCLUDE_DIR
)
if(NCCL_FOUND)
set(NCCL_LIBRARIES ${NCCL_LIBRARY})
set(NCCL_INCLUDE_DIRS ${NCCL_INCLUDE_DIR})
if(NOT TARGET NCCL::NCCL)
add_library(NCCL::NCCL UNKNOWN IMPORTED)
set_target_properties(NCCL::NCCL PROPERTIES
IMPORTED_LOCATION "${NCCL_LIBRARY}"
INTERFACE_INCLUDE_DIRECTORIES "${NCCL_INCLUDE_DIR}"
)
endif()
endif()
mark_as_advanced(NCCL_INCLUDE_DIR NCCL_LIBRARY)
engine/ggml/cmake/GitVars.cmake
+22
View File
@@ -0,0 +1,22 @@
find_package(Git)
# the commit's SHA1
execute_process(COMMAND
"${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_SHA1
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
# the date of the commit
execute_process(COMMAND
"${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_DATE
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
# the subject of the commit
execute_process(COMMAND
"${GIT_EXECUTABLE}" log -1 --format=%s
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
engine/ggml/cmake/common.cmake
+50
View File
@@ -0,0 +1,50 @@
function(ggml_get_flags CCID CCVER)
set(C_FLAGS "")
set(CXX_FLAGS "")
if (CCID MATCHES "Clang")
set(C_FLAGS -Wunreachable-code-break -Wunreachable-code-return)
set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
if (
(CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
(CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
)
list(APPEND C_FLAGS -Wdouble-promotion)
endif()
elseif (CCID STREQUAL "GNU")
set(C_FLAGS -Wdouble-promotion)
set(CXX_FLAGS -Wno-array-bounds)
if (CCVER VERSION_GREATER_EQUAL 8.1.0)
list(APPEND CXX_FLAGS -Wextra-semi)
endif()
endif()
set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE)
set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
endfunction()
function(ggml_get_system_arch)
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR
CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
set(GGML_SYSTEM_ARCH "ARM" PARENT_SCOPE)
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
set(GGML_SYSTEM_ARCH "loongarch64" PARENT_SCOPE)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
set(GGML_SYSTEM_ARCH "riscv64" PARENT_SCOPE)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
set(GGML_SYSTEM_ARCH "s390x" PARENT_SCOPE)
else()
set(GGML_SYSTEM_ARCH "UNKNOWN" PARENT_SCOPE)
endif()
endfunction()
engine/ggml/cmake/ggml-config.cmake.in
+201
View File
@@ -0,0 +1,201 @@
@PACKAGE_INIT@
@GGML_VARIABLES_EXPANDED@
# Find all dependencies before creating any target.
include(CMakeFindDependencyMacro)
find_dependency(Threads)
if (NOT GGML_SHARED_LIB)
set(GGML_BASE_INTERFACE_LINK_LIBRARIES "")
set(GGML_CPU_INTERFACE_LINK_LIBRARIES "")
set(GGML_CPU_INTERFACE_LINK_OPTIONS "")
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if(NOT ACCELERATE_FRAMEWORK)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${ACCELERATE_FRAMEWORK})
endif()
if (GGML_OPENMP_ENABLED)
find_dependency(OpenMP)
set(GGML_OPENMP_INTERFACE_LINK_LIBRARIES "")
if (TARGET OpenMP::OpenMP_C)
list(APPEND GGML_OPENMP_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_C)
endif()
if (TARGET OpenMP::OpenMP_CXX)
list(APPEND GGML_OPENMP_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_CXX)
endif()
list(APPEND GGML_BASE_INTERFACE_LINK_LIBRARIES ${GGML_OPENMP_INTERFACE_LINK_LIBRARIES})
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${GGML_OPENMP_INTERFACE_LINK_LIBRARIES})
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind)
if(NOT memkind)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES memkind)
endif()
if (GGML_BLAS)
find_dependency(BLAS)
list(APPEND GGML_BLAS_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
list(APPEND GGML_BLAS_INTERFACE_LINK_OPTIONS ${BLAS_LINKER_FLAGS})
endif()
if (GGML_CUDA)
set(GGML_CUDA_INTERFACE_LINK_LIBRARIES "")
find_dependency(CUDAToolkit)
if (GGML_STATIC)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cudart_static>)
if (WIN32)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas> $<LINK_ONLY:CUDA::cublasLt>)
else()
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas_static> $<LINK_ONLY:CUDA::cublasLt_static>)
endif()
endif()
if (NOT GGML_CUDA_NO_VMM)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cuda_driver>)
endif()
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation)
find_library(METAL_FRAMEWORK Metal)
find_library(METALKIT_FRAMEWORK MetalKit)
if(NOT FOUNDATION_LIBRARY OR NOT METAL_FRAMEWORK OR NOT METALKIT_FRAMEWORK)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
set(GGML_METAL_INTERFACE_LINK_LIBRARIES
${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
endif()
if (GGML_OPENCL)
find_dependency(OpenCL)
set(GGML_OPENCL_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:OpenCL::OpenCL>)
endif()
if (GGML_VULKAN)
find_dependency(Vulkan)
set(GGML_VULKAN_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:Vulkan::Vulkan>)
endif()
if (GGML_HIP)
find_dependency(hip)
find_dependency(hipblas)
find_dependency(rocblas)
set(GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
endif()
if (GGML_SYCL)
set(GGML_SYCL_INTERFACE_LINK_LIBRARIES "")
find_package(DNNL)
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES DNNL::dnnl)
endif()
if (WIN32)
find_dependency(IntelSYCL)
find_dependency(MKL)
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
endif()
endif()
endif()
set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
if(NOT TARGET ggml::ggml)
find_package(Threads REQUIRED)
find_library(GGML_LIBRARY ggml
REQUIRED
HINTS ${GGML_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH)
add_library(ggml::ggml UNKNOWN IMPORTED)
set_target_properties(ggml::ggml
PROPERTIES
IMPORTED_LOCATION "${GGML_LIBRARY}")
find_library(GGML_BASE_LIBRARY ggml-base
REQUIRED
HINTS ${GGML_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH)
add_library(ggml::ggml-base UNKNOWN IMPORTED)
set_target_properties(ggml::ggml-base
PROPERTIES
IMPORTED_LOCATION "${GGML_BASE_LIBRARY}"
INTERFACE_LINK_LIBRARIES "${GGML_BASE_INTERFACE_LINK_LIBRARIES}")
set(_ggml_all_targets "")
if (NOT GGML_BACKEND_DL)
foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
REQUIRED
HINTS ${GGML_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH)
message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
set_target_properties(ggml::${_ggml_backend}
PROPERTIES
INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
INTERFACE_COMPILE_FEATURES c_std_90
POSITION_INDEPENDENT_CODE ON)
string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
if(is_cpu_variant)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
set_target_properties(ggml::${_ggml_backend}
PROPERTIES
INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
if(GGML_CPU_INTERFACE_LINK_OPTIONS)
set_target_properties(ggml::${_ggml_backend}
PROPERTIES
INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
endif()
else()
list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
set_target_properties(ggml::${_ggml_backend}
PROPERTIES
INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
set_target_properties(ggml::${_ggml_backend}
PROPERTIES
INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
endif()
endif()
list(APPEND _ggml_all_targets ggml::${_ggml_backend})
endforeach()
endif()
list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
set_target_properties(ggml::ggml
PROPERTIES
INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
add_library(ggml::all INTERFACE IMPORTED)
set_target_properties(ggml::all
PROPERTIES
INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
endif()
check_required_components(ggml)
engine/ggml/docs/gguf.md
+828
View File
@@ -0,0 +1,828 @@
# GGUF
GGUF is a file format for storing models for inference with GGML and executors based on GGML. GGUF is a binary format that is designed for fast loading and saving of models, and for ease of reading. Models are traditionally developed using PyTorch or another framework, and then converted to GGUF for use in GGML.
It is a successor file format to GGML, GGMF and GGJT, and is designed to be unambiguous by containing all the information needed to load a model. It is also designed to be extensible, so that new information can be added to models without breaking compatibility.
For more information about the motivation behind GGUF, see [Historical State of Affairs](#historical-state-of-affairs).
## Specification
GGUF is a format based on the existing GGJT, but makes a few changes to the format to make it more extensible and easier to use. The following features are desired:
- Single-file deployment: they can be easily distributed and loaded, and do not require any external files for additional information.
- Extensible: new features can be added to GGML-based executors/new information can be added to GGUF models without breaking compatibility with existing models.
- `mmap` compatibility: models can be loaded using `mmap` for fast loading and saving.
- Easy to use: models can be easily loaded and saved using a small amount of code, with no need for external libraries, regardless of the language used.
- Full information: all information needed to load a model is contained in the model file, and no additional information needs to be provided by the user.
The key difference between GGJT and GGUF is the use of a key-value structure for the hyperparameters (now referred to as metadata), rather than a list of untyped values. This allows for new metadata to be added without breaking compatibility with existing models, and to annotate the model with additional information that may be useful for inference or for identifying the model.
### GGUF Naming Convention
GGUF follow a naming convention of `[<Sidecar>]<BaseName><SizeLabel><FineTune><Version><Encoding><Type><Shard>.gguf` where each component is delimitated by a `-` if present. Ultimately this is intended to make it easier for humans to at a glance get the most important details of a model. It is not intended to be perfectly parsable in the field due to the diversity of existing gguf filenames.
The components are:
1. **Sidecar**: (Optional) Prefix marking the file as an auxiliary module loaded alongside a base model, rather than a standalone model. When present, sits at the very front of the filename followed by `-`. Lowercase by convention.
- `mmproj` : Multimodal projector (vision/audio encoder and projection layer for use with a base LLM)
- `mtp` : Multi-Token Prediction heads (speculative-decoding draft module, intended to be loaded alongside a base model of matching architecture and version). Note that oftentimes the MTP weights can be distributed inside the base model, in which case there is no separate `mtp-` sidecar file.
1. **BaseName**: A descriptive name for the model base type or architecture.
- This can be derived from gguf metadata `general.basename` substituting spaces for dashes.
1. **SizeLabel**: Parameter weight class (useful for leader boards) represented as `<expertCount>x<count><scale-prefix>`
- This can be derived from gguf metadata `general.size_label` if available or calculated if missing.
- Rounded decimal point is supported in count with a single letter scale prefix to assist in floating point exponent shown below
- `Q`: Quadrillion parameters.
- `T`: Trillion parameters.
- `B`: Billion parameters.
- `M`: Million parameters.
- `K`: Thousand parameters.
- Additional `-<attributes><count><scale-prefix>` can be appended as needed to indicate other attributes of interest
1. **FineTune**: A descriptive name for the model fine tuning goal (e.g. Chat, Instruct, etc...)
- This can be derived from gguf metadata `general.finetune` substituting spaces for dashes.
1. **Version**: (Optional) Denotes the model version number, formatted as `v<Major>.<Minor>`
- If model is missing a version number then assume `v1.0` (First Public Release)
- This can be derived from gguf metadata `general.version`
1. **Encoding**: Indicates the weights encoding scheme that was applied to the model. Content, type mixture and arrangement however are determined by user code and can vary depending on project needs.
1. **Type**: Indicates the kind of gguf file and the intended purpose for it
- If missing, then file is by default a typical gguf tensor model file
- `LoRA` : GGUF file is a LoRA adapter
- `vocab` : GGUF file with only vocab data and metadata
1. **Shard**: (Optional) Indicates and denotes that the model has been split into multiple shards, formatted as `<ShardNum>-of-<ShardTotal>`.
- *ShardNum* : Shard position in this model. Must be 5 digits padded by zeros.
- Shard number always starts from `00001` onwards (e.g. First shard always starts at `00001-of-XXXXX` rather than `00000-of-XXXXX`).
- *ShardTotal* : Total number of shards in this model. Must be 5 digits padded by zeros.
#### Validating Above Naming Convention
At a minimum all model files should have at least BaseName, SizeLabel, Version, in order to be easily validated as a file that is keeping with the GGUF Naming Convention. An example of this issue is that it is easy for Encoding to be mistaken as a FineTune if Version is omitted.
To validate you can use this regular expression `^(?:(?<Sidecar>mmproj|mtp)-)?(?<BaseName>[A-Za-z0-9\s]*(?:(?:-(?:(?:[A-Za-z\s][A-Za-z0-9\s]*)|(?:[0-9\s]*)))*))-(?:(?<SizeLabel>(?:\d+x)?(?:\d+\.)?\d+[A-Za-z](?:-[A-Za-z]+(\d+\.)?\d+[A-Za-z]+)?)(?:-(?<FineTune>[A-Za-z0-9\s-]+))?)?-(?:(?<Version>v\d+(?:\.\d+)*))(?:-(?<Encoding>(?!LoRA|vocab)[\w_]+))?(?:-(?<Type>LoRA|vocab))?(?:-(?<Shard>\d{5}-of-\d{5}))?\.gguf$` which will check that you got the minimum BaseName, SizeLabel and Version present in the correct order.
For example:
* `Mixtral-8x7B-v0.1-KQ2.gguf`:
- Model Name: Mixtral
- Expert Count: 8
- Parameter Count: 7B
- Version Number: v0.1
- Weight Encoding Scheme: KQ2
* `Hermes-2-Pro-Llama-3-8B-F16.gguf`:
- Model Name: Hermes 2 Pro Llama 3
- Expert Count: 0
- Parameter Count: 8B
- Version Number: v1.0
- Weight Encoding Scheme: F16
- Shard: N/A
* `Grok-100B-v1.0-Q4_0-00003-of-00009.gguf`
- Model Name: Grok
- Expert Count: 0
- Parameter Count: 100B
- Version Number: v1.0
- Weight Encoding Scheme: Q4_0
- Shard: 3 out of 9 total shards
* `mtp-Qwen3-27B-v1.0-Q4_K_M.gguf`
- Sidecar: mtp (Multi-Token Prediction draft module)
- Model Name: Qwen3
- Expert Count: 0
- Parameter Count: 27B (of the main model — sidecar tensors are smaller)
- Version Number: v1.0
- Weight Encoding Scheme: Q4_K_M
* `mmproj-Qwen2-VL-7B-v1.0-F16.gguf`
- Sidecar: mmproj (multimodal projector)
- Model Name: Qwen2-VL
- Expert Count: 0
- Parameter Count: 7B (of the main model — sidecar tensors are smaller)
- Version Number: v1.0
- Weight Encoding Scheme: F16
<details><summary>Example Node.js Regex Function</summary>
```js
#!/usr/bin/env node
const ggufRegex = /^(?:(?<Sidecar>mmproj|mtp)-)?(?<BaseName>[A-Za-z0-9\s]*(?:(?:-(?:(?:[A-Za-z\s][A-Za-z0-9\s]*)|(?:[0-9\s]*)))*))-(?:(?<SizeLabel>(?:\d+x)?(?:\d+\.)?\d+[A-Za-z](?:-[A-Za-z]+(\d+\.)?\d+[A-Za-z]+)?)(?:-(?<FineTune>[A-Za-z0-9\s-]+))?)?-(?:(?<Version>v\d+(?:\.\d+)*))(?:-(?<Encoding>(?!LoRA|vocab)[\w_]+))?(?:-(?<Type>LoRA|vocab))?(?:-(?<Shard>\d{5}-of-\d{5}))?\.gguf$/;
function parseGGUFFilename(filename) {
const match = ggufRegex.exec(filename);
if (!match)
return null;
const {Sidecar = null, BaseName = null, SizeLabel = null, FineTune = null, Version = "v1.0", Encoding = null, Type = null, Shard = null} = match.groups;
return {Sidecar: Sidecar, BaseName: BaseName, SizeLabel: SizeLabel, FineTune: FineTune, Version: Version, Encoding: Encoding, Type: Type, Shard: Shard};
}
const testCases = [
{filename: 'Mixtral-8x7B-v0.1-KQ2.gguf', expected: { Sidecar: null, BaseName: 'Mixtral', SizeLabel: '8x7B', FineTune: null, Version: 'v0.1', Encoding: 'KQ2', Type: null, Shard: null}},
{filename: 'Grok-100B-v1.0-Q4_0-00003-of-00009.gguf', expected: { Sidecar: null, BaseName: 'Grok', SizeLabel: '100B', FineTune: null, Version: 'v1.0', Encoding: 'Q4_0', Type: null, Shard: "00003-of-00009"}},
{filename: 'Hermes-2-Pro-Llama-3-8B-v1.0-F16.gguf', expected: { Sidecar: null, BaseName: 'Hermes-2-Pro-Llama-3', SizeLabel: '8B', FineTune: null, Version: 'v1.0', Encoding: 'F16', Type: null, Shard: null}},
{filename: 'Phi-3-mini-3.8B-ContextLength4k-instruct-v1.0.gguf', expected: { Sidecar: null, BaseName: 'Phi-3-mini', SizeLabel: '3.8B-ContextLength4k', FineTune: 'instruct', Version: 'v1.0', Encoding: null, Type: null, Shard: null}},
{filename: 'mtp-Qwen3-27B-v1.0-Q4_K_M.gguf', expected: { Sidecar: 'mtp', BaseName: 'Qwen3', SizeLabel: '27B', FineTune: null, Version: 'v1.0', Encoding: 'Q4_K_M', Type: null, Shard: null}},
{filename: 'mmproj-Qwen2-VL-7B-v1.0-F16.gguf', expected: { Sidecar: 'mmproj', BaseName: 'Qwen2-VL', SizeLabel: '7B', FineTune: null, Version: 'v1.0', Encoding: 'F16', Type: null, Shard: null}},
{filename: 'not-a-known-arrangement.gguf', expected: null},
];
testCases.forEach(({ filename, expected }) => {
const result = parseGGUFFilename(filename);
const passed = JSON.stringify(result) === JSON.stringify(expected);
console.log(`${filename}: ${passed ? "PASS" : "FAIL"}`);
if (!passed) {
console.log(result);
console.log(expected);
}
});
```
</details>
### File Structure
![image](https://github.com/ggerganov/ggml/assets/1991296/c3623641-3a1d-408e-bfaf-1b7c4e16aa63)
*diagram by [@mishig25](https://github.com/mishig25) (GGUF v3)*
GGUF files are structured as follows. They use a global alignment specified in the `general.alignment` metadata field, referred to as `ALIGNMENT` below. Where required, the file is padded with `0x00` bytes to the next multiple of `general.alignment`.
Fields, including arrays, are written sequentially without alignment unless otherwise specified.
Models are little-endian by default. They can also come in big-endian for use with big-endian computers; in this case, all values (including metadata values and tensors) will also be big-endian. At the time of writing, there is no way to determine if a model is big-endian; this may be rectified in future versions. If no additional information is provided, assume the model is little-endian.
```c
enum ggml_type: uint32_t {
GGML_TYPE_F32 = 0,
GGML_TYPE_F16 = 1,
GGML_TYPE_Q4_0 = 2,
GGML_TYPE_Q4_1 = 3,
// GGML_TYPE_Q4_2 = 4, support has been removed
// GGML_TYPE_Q4_3 = 5, support has been removed
GGML_TYPE_Q5_0 = 6,
GGML_TYPE_Q5_1 = 7,
GGML_TYPE_Q8_0 = 8,
GGML_TYPE_Q8_1 = 9,
GGML_TYPE_Q2_K = 10,
GGML_TYPE_Q3_K = 11,
GGML_TYPE_Q4_K = 12,
GGML_TYPE_Q5_K = 13,
GGML_TYPE_Q6_K = 14,
GGML_TYPE_Q8_K = 15,
GGML_TYPE_IQ2_XXS = 16,
GGML_TYPE_IQ2_XS = 17,
GGML_TYPE_IQ3_XXS = 18,
GGML_TYPE_IQ1_S = 19,
GGML_TYPE_IQ4_NL = 20,
GGML_TYPE_IQ3_S = 21,
GGML_TYPE_IQ2_S = 22,
GGML_TYPE_IQ4_XS = 23,
GGML_TYPE_I8 = 24,
GGML_TYPE_I16 = 25,
GGML_TYPE_I32 = 26,
GGML_TYPE_I64 = 27,
GGML_TYPE_F64 = 28,
GGML_TYPE_IQ1_M = 29,
GGML_TYPE_BF16 = 30,
// GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
// GGML_TYPE_Q4_0_4_8 = 32,
// GGML_TYPE_Q4_0_8_8 = 33,
GGML_TYPE_TQ1_0 = 34,
GGML_TYPE_TQ2_0 = 35,
// GGML_TYPE_IQ4_NL_4_4 = 36,
// GGML_TYPE_IQ4_NL_4_8 = 37,
// GGML_TYPE_IQ4_NL_8_8 = 38,
GGML_TYPE_MXFP4 = 39, // MXFP4 (1 block)
GGML_TYPE_COUNT = 40,
};
enum gguf_metadata_value_type: uint32_t {
// The value is a 8-bit unsigned integer.
GGUF_METADATA_VALUE_TYPE_UINT8 = 0,
// The value is a 8-bit signed integer.
GGUF_METADATA_VALUE_TYPE_INT8 = 1,
// The value is a 16-bit unsigned little-endian integer.
GGUF_METADATA_VALUE_TYPE_UINT16 = 2,
// The value is a 16-bit signed little-endian integer.
GGUF_METADATA_VALUE_TYPE_INT16 = 3,
// The value is a 32-bit unsigned little-endian integer.
GGUF_METADATA_VALUE_TYPE_UINT32 = 4,
// The value is a 32-bit signed little-endian integer.
GGUF_METADATA_VALUE_TYPE_INT32 = 5,
// The value is a 32-bit IEEE754 floating point number.
GGUF_METADATA_VALUE_TYPE_FLOAT32 = 6,
// The value is a boolean.
// 1-byte value where 0 is false and 1 is true.
// Anything else is invalid, and should be treated as either the model being invalid or the reader being buggy.
GGUF_METADATA_VALUE_TYPE_BOOL = 7,
// The value is a UTF-8 non-null-terminated string, with length prepended.
GGUF_METADATA_VALUE_TYPE_STRING = 8,
// The value is an array of other values, with the length and type prepended.
///
// Arrays can be nested, and the length of the array is the number of elements in the array, not the number of bytes.
GGUF_METADATA_VALUE_TYPE_ARRAY = 9,
// The value is a 64-bit unsigned little-endian integer.
GGUF_METADATA_VALUE_TYPE_UINT64 = 10,
// The value is a 64-bit signed little-endian integer.
GGUF_METADATA_VALUE_TYPE_INT64 = 11,
// The value is a 64-bit IEEE754 floating point number.
GGUF_METADATA_VALUE_TYPE_FLOAT64 = 12,
};
// A string in GGUF.
struct gguf_string_t {
// The length of the string, in bytes.
uint64_t len;
// The string as a UTF-8 non-null-terminated string.
char string[len];
};
union gguf_metadata_value_t {
uint8_t uint8;
int8_t int8;
uint16_t uint16;
int16_t int16;
uint32_t uint32;
int32_t int32;
float float32;
uint64_t uint64;
int64_t int64;
double float64;
bool bool_;
gguf_string_t string;
struct {
// Any value type is valid, including arrays.
gguf_metadata_value_type type;
// Number of elements, not bytes
uint64_t len;
// The array of values.
gguf_metadata_value_t array[len];
} array;
};
struct gguf_metadata_kv_t {
// The key of the metadata. It is a standard GGUF string, with the following caveats:
// - It must be a valid ASCII string.
// - It must be a hierarchical key, where each segment is `lower_snake_case` and separated by a `.`.
// - It must be at most 2^16-1/65535 bytes long.
// Any keys that do not follow these rules are invalid.
gguf_string_t key;
// The type of the value.
// Must be one of the `gguf_metadata_value_type` values.
gguf_metadata_value_type value_type;
// The value.
gguf_metadata_value_t value;
};
struct gguf_header_t {
// Magic number to announce that this is a GGUF file.
// Must be `GGUF` at the byte level: `0x47` `0x47` `0x55` `0x46`.
// Your executor might do little-endian byte order, so it might be
// check for 0x46554747 and letting the endianness cancel out.
// Consider being *very* explicit about the byte order here.
uint32_t magic;
// The version of the format implemented.
// Must be `3` for version described in this spec, which introduces big-endian support.
//
// This version should only be increased for structural changes to the format.
// Changes that do not affect the structure of the file should instead update the metadata
// to signify the change.
uint32_t version;
// The number of tensors in the file.
// This is explicit, instead of being included in the metadata, to ensure it is always present
// for loading the tensors.
uint64_t tensor_count;
// The number of metadata key-value pairs.
uint64_t metadata_kv_count;
// The metadata key-value pairs.
gguf_metadata_kv_t metadata_kv[metadata_kv_count];
};
uint64_t align_offset(uint64_t offset) {
return offset + (ALIGNMENT - (offset % ALIGNMENT)) % ALIGNMENT;
}
struct gguf_tensor_info_t {
// The name of the tensor. It is a standard GGUF string, with the caveat that
// it must be at most 64 bytes long.
gguf_string_t name;
// The number of dimensions in the tensor.
// Currently at most 4, but this may change in the future.
uint32_t n_dimensions;
// The dimensions of the tensor.
uint64_t dimensions[n_dimensions];
// The type of the tensor.
ggml_type type;
// The offset of the tensor's data in this file in bytes.
//
// This offset is relative to `tensor_data`, not to the start
// of the file, to make it easier for writers to write the file.
// Readers should consider exposing this offset relative to the
// file to make it easier to read the data.
//
// Must be a multiple of `ALIGNMENT`. That is, `align_offset(offset) == offset`.
uint64_t offset;
};
struct gguf_file_t {
// The header of the file.
gguf_header_t header;
// Tensor infos, which can be used to locate the tensor data.
gguf_tensor_info_t tensor_infos[header.tensor_count];
// Padding to the nearest multiple of `ALIGNMENT`.
//
// That is, if `sizeof(header) + sizeof(tensor_infos)` is not a multiple of `ALIGNMENT`,
// this padding is added to make it so.
//
// This can be calculated as `align_offset(position) - position`, where `position` is
// the position of the end of `tensor_infos` (i.e. `sizeof(header) + sizeof(tensor_infos)`).
uint8_t _padding[];
// Tensor data.
//
// This is arbitrary binary data corresponding to the weights of the model. This data should be close
// or identical to the data in the original model file, but may be different due to quantization or
// other optimizations for inference. Any such deviations should be recorded in the metadata or as
// part of the architecture definition.
//
// Each tensor's data must be stored within this array, and located through its `tensor_infos` entry.
// The offset of each tensor's data must be a multiple of `ALIGNMENT`, and the space between tensors
// should be padded to `ALIGNMENT` bytes.
uint8_t tensor_data[];
};
```
## Standardized key-value pairs
The following key-value pairs are standardized. This list may grow in the future as more use cases are discovered. Where possible, names are shared with the original model definitions to make it easier to map between the two.
Not all of these are required, but they are all recommended. Keys that are required are bolded. For omitted pairs, the reader should assume that the value is unknown and either default or error as appropriate.
The community can develop their own key-value pairs to carry additional data. However, these should be namespaced with the relevant community name to avoid collisions. For example, the `rustformers` community might use `rustformers.` as a prefix for all of their keys.
If a particular community key is widely used, it may be promoted to a standardized key.
By convention, most counts/lengths/etc are `uint64` unless otherwise specified. This is to allow for larger models to be supported in the future. Some models may use `uint32` for their values; it is recommended that readers support both.
### General
#### Required
- **`general.architecture: string`**: describes what architecture this model implements. All lowercase ASCII, with only `[a-z0-9]+` characters allowed. Known values include:
- `llama`
- `mpt`
- `gptneox`
- `gptj`
- `gpt2`
- `bloom`
- `falcon`
- `mamba`
- `rwkv`
- **`general.quantization_version: uint32`**: The version of the quantization format. Not required if the model is not quantized (i.e. no tensors are quantized). If any tensors are quantized, this _must_ be present. This is separate to the quantization scheme of the tensors itself; the quantization version may change without changing the scheme's name (e.g. the quantization scheme is Q5_K, and the quantization version is 4).
- **`general.alignment: uint32`**: the global alignment to use, as described above. This can vary to allow for different alignment schemes, but it must be a multiple of 8. Some writers may not write the alignment. If the alignment is **not** specified, assume it is `32`.
#### General metadata
- `general.name: string`: The name of the model. This should be a human-readable name that can be used to identify the model. It should be unique within the community that the model is defined in.
- `general.author: string`: The author of the model.
- `general.version: string`: The version of the model.
- `general.organization: string`: The organization of the model.
- `general.basename: string`: The base model name / architecture of the model
- `general.finetune: string`: What has the base model been optimized toward.
- `general.description: string`: free-form description of the model including anything that isn't covered by the other fields
- `general.quantized_by: string`: The name of the individual who quantized the model
- `general.size_label: string`: Size class of the model, such as number of weights and experts. (Useful for leader boards)
- `general.license: string`: License of the model, expressed as a [SPDX license expression](https://spdx.github.io/spdx-spec/v2-draft/SPDX-license-expressions/) (e.g. `"MIT OR Apache-2.0`). Do not include any other information, such as the license text or the URL to the license.
- `general.license.name: string`: Human friendly license name
- `general.license.link: string`: URL to the license.
- `general.url: string`: URL to the model's homepage. This can be a GitHub repo, a paper, etc.
- `general.doi: string`: Digital Object Identifier (DOI) https://www.doi.org/
- `general.uuid: string`: [Universally unique identifier](https://en.wikipedia.org/wiki/Universally_unique_identifier)
- `general.repo_url: string`: URL to the model's repository such as a GitHub repo or HuggingFace repo
- `general.tags: string[]`: List of tags that can be used as search terms for a search engine or social media
- `general.languages: string[]`: What languages can the model speak. Encoded as [ISO 639](https://en.wikipedia.org/wiki/List_of_ISO_639_language_codes) two letter codes
- `general.datasets: string[]`: Links or references to datasets that the model was trained upon
- `general.file_type: uint32`: An enumerated value describing the type of the majority of the tensors in the file. Optional; can be inferred from the tensor types.
- `ALL_F32 = 0`
- `MOSTLY_F16 = 1`
- `MOSTLY_Q4_0 = 2`
- `MOSTLY_Q4_1 = 3`
- `MOSTLY_Q4_1_SOME_F16 = 4`
- `MOSTLY_Q4_2 = 5` (support removed)
- `MOSTLY_Q4_3 = 6` (support removed)
- `MOSTLY_Q8_0 = 7`
- `MOSTLY_Q5_0 = 8`
- `MOSTLY_Q5_1 = 9`
- `MOSTLY_Q2_K = 10`
- `MOSTLY_Q3_K_S = 11`
- `MOSTLY_Q3_K_M = 12`
- `MOSTLY_Q3_K_L = 13`
- `MOSTLY_Q4_K_S = 14`
- `MOSTLY_Q4_K_M = 15`
- `MOSTLY_Q5_K_S = 16`
- `MOSTLY_Q5_K_M = 17`
- `MOSTLY_Q6_K = 18`
#### Source metadata
Information about where this model came from. This is useful for tracking the provenance of the model, and for finding the original source if the model is modified. For a model that was converted from GGML, for example, these keys would point to the model that was converted from.
- `general.source.url: string`: URL to the source of the model's homepage. This can be a GitHub repo, a paper, etc.
- `general.source.doi: string`: Source Digital Object Identifier (DOI) https://www.doi.org/
- `general.source.uuid: string`: Source [Universally unique identifier](https://en.wikipedia.org/wiki/Universally_unique_identifier)
- `general.source.repo_url: string`: URL to the source of the model's repository such as a GitHub repo or HuggingFace repo
- `general.base_model.count: uint32`: Number of parent models
- `general.base_model.{id}.name: string`: The name of the parent model.
- `general.base_model.{id}.author: string`: The author of the parent model.
- `general.base_model.{id}.version: string`: The version of the parent model.
- `general.base_model.{id}.organization: string`: The organization of the parent model.
- `general.base_model.{id}.url: string`: URL to the source of the parent model's homepage. This can be a GitHub repo, a paper, etc.
- `general.base_model.{id}.doi: string`: Parent Digital Object Identifier (DOI) https://www.doi.org/
- `general.base_model.{id}.uuid: string`: Parent [Universally unique identifier](https://en.wikipedia.org/wiki/Universally_unique_identifier)
- `general.base_model.{id}.repo_url: string`: URL to the source of the parent model's repository such as a GitHub repo or HuggingFace repo
### LLM
In the following, `[llm]` is used to fill in for the name of a specific LLM architecture. For example, `llama` for LLaMA, `mpt` for MPT, etc. If mentioned in an architecture's section, it is required for that architecture, but not all keys are required for all architectures. Consult the relevant section for more information.
- `[llm].context_length: uint64`: Also known as `n_ctx`. length of the context (in tokens) that the model was trained on. For most architectures, this is the hard limit on the length of the input. Architectures, like RWKV, that are not reliant on transformer-style attention may be able to handle larger inputs, but this is not guaranteed.
- `[llm].embedding_length: uint64`: Also known as `n_embd`. Embedding layer size.
- `[llm].block_count: uint64`: The number of blocks of attention+feed-forward layers (i.e. the bulk of the LLM). Does not include the input or embedding layers.
- `[llm].feed_forward_length: uint64`: Also known as `n_ff`. The length of the feed-forward layer.
- `[llm].use_parallel_residual: bool`: Whether or not the parallel residual logic should be used.
- `[llm].tensor_data_layout: string`: When a model is converted to GGUF, tensors may be rearranged to improve performance. This key describes the layout of the tensor data. This is not required; if not present, it is assumed to be `reference`.
- `reference`: tensors are laid out in the same order as the original model
- further options can be found for each architecture in their respective sections
- `[llm].expert_count: uint32`: Number of experts in MoE models (optional for non-MoE arches).
- `[llm].expert_used_count: uint32`: Number of experts used during each token token evaluation (optional for non-MoE arches).
#### Attention
- `[llm].attention.head_count: uint64`: Also known as `n_head`. Number of attention heads.
- `[llm].attention.head_count_kv: uint64`: The number of heads per group used in Grouped-Query-Attention. If not present or if present and equal to `[llm].attention.head_count`, the model does not use GQA.
- `[llm].attention.max_alibi_bias: float32`: The maximum bias to use for ALiBI.
- `[llm].attention.clamp_kqv: float32`: Value (`C`) to clamp the values of the `Q`, `K`, and `V` tensors between (`[-C, C]`).
- `[llm].attention.layer_norm_epsilon: float32`: Layer normalization epsilon.
- `[llm].attention.layer_norm_rms_epsilon: float32`: Layer RMS normalization epsilon.
- `[llm].attention.key_length: uint32`: The optional size of a key head, $d_k$. If not specified, it will be `n_embd / n_head`.
- `[llm].attention.value_length: uint32`: The optional size of a value head, $d_v$. If not specified, it will be `n_embd / n_head`.
#### RoPE
- `[llm].rope.dimension_count: uint64`: The number of rotary dimensions for RoPE.
- `[llm].rope.freq_base: float32`: The base frequency for RoPE.
##### Scaling
The following keys describe RoPE scaling parameters:
- `[llm].rope.scaling.type: string`: Can be `none`, `linear`, or `yarn`.
- `[llm].rope.scaling.factor: float32`: A scale factor for RoPE to adjust the context length.
- `[llm].rope.scaling.original_context_length: uint32_t`: The original context length of the base model.
- `[llm].rope.scaling.finetuned: bool`: True if model has been finetuned with RoPE scaling.
Note that older models may not have these keys, and may instead use the following key:
- `[llm].rope.scale_linear: float32`: A linear scale factor for RoPE to adjust the context length.
It is recommended that models use the newer keys if possible, as they are more flexible and allow for more complex scaling schemes. Executors will need to support both indefinitely.
#### SSM
- `[llm].ssm.conv_kernel: uint32`: The size of the rolling/shift state.
- `[llm].ssm.inner_size: uint32`: The embedding size of the states.
- `[llm].ssm.state_size: uint32`: The size of the recurrent state.
- `[llm].ssm.time_step_rank: uint32`: The rank of time steps.
#### Models
The following sections describe the metadata for each model architecture. Each key specified _must_ be present.
##### LLaMA
- `llama.context_length`
- `llama.embedding_length`
- `llama.block_count`
- `llama.feed_forward_length`
- `llama.rope.dimension_count`
- `llama.attention.head_count`
- `llama.attention.layer_norm_rms_epsilon`
###### Optional
- `llama.rope.scale`
- `llama.attention.head_count_kv`
- `llama.tensor_data_layout`:
- `Meta AI original pth`:
```python
def permute(weights: NDArray, n_head: int) -> NDArray:
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape))
```
- `llama.expert_count`
- `llama.expert_used_count`
##### MPT
- `mpt.context_length`
- `mpt.embedding_length`
- `mpt.block_count`
- `mpt.attention.head_count`
- `mpt.attention.alibi_bias_max`
- `mpt.attention.clip_kqv`
- `mpt.attention.layer_norm_epsilon`
##### GPT-NeoX
- `gptneox.context_length`
- `gptneox.embedding_length`
- `gptneox.block_count`
- `gptneox.use_parallel_residual`
- `gptneox.rope.dimension_count`
- `gptneox.attention.head_count`
- `gptneox.attention.layer_norm_epsilon`
###### Optional
- `gptneox.rope.scale`
##### GPT-J
- `gptj.context_length`
- `gptj.embedding_length`
- `gptj.block_count`
- `gptj.rope.dimension_count`
- `gptj.attention.head_count`
- `gptj.attention.layer_norm_epsilon`
###### Optional
- `gptj.rope.scale`
##### GPT-2
- `gpt2.context_length`
- `gpt2.embedding_length`
- `gpt2.block_count`
- `gpt2.attention.head_count`
- `gpt2.attention.layer_norm_epsilon`
##### BLOOM
- `bloom.context_length`
- `bloom.embedding_length`
- `bloom.block_count`
- `bloom.feed_forward_length`
- `bloom.attention.head_count`
- `bloom.attention.layer_norm_epsilon`
##### Falcon
- `falcon.context_length`
- `falcon.embedding_length`
- `falcon.block_count`
- `falcon.attention.head_count`
- `falcon.attention.head_count_kv`
- `falcon.attention.use_norm`
- `falcon.attention.layer_norm_epsilon`
###### Optional
- `falcon.tensor_data_layout`:
- `jploski` (author of the original GGML implementation of Falcon):
```python
# The original query_key_value tensor contains n_head_kv "kv groups",
# each consisting of n_head/n_head_kv query weights followed by one key
# and one value weight (shared by all query heads in the kv group).
# This layout makes it a big pain to work with in GGML.
# So we rearrange them here,, so that we have n_head query weights
# followed by n_head_kv key weights followed by n_head_kv value weights,
# in contiguous fashion.
if "query_key_value" in src:
qkv = model[src].view(
n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
model[src] = torch.cat((q,k,v)).reshape_as(model[src])
```
##### Mamba
- `mamba.context_length`
- `mamba.embedding_length`
- `mamba.block_count`
- `mamba.ssm.conv_kernel`
- `mamba.ssm.inner_size`
- `mamba.ssm.state_size`
- `mamba.ssm.time_step_rank`
- `mamba.attention.layer_norm_rms_epsilon`
##### RWKV
The vocabulary size is the same as the number of rows in the `head` matrix.
- `rwkv.architecture_version: uint32`: The only allowed value currently is 4. Version 5 is expected to appear some time in the future.
- `rwkv.context_length: uint64`: Length of the context used during training or fine-tuning. RWKV is able to handle larger context than this limit, but the output quality may suffer.
- `rwkv.block_count: uint64`
- `rwkv.embedding_length: uint64`
- `rwkv.feed_forward_length: uint64`
##### Whisper
Keys that do not have types defined should be assumed to share definitions with `llm.` keys.
(For example, `whisper.context_length` is equivalent to `llm.context_length`.)
This is because they are both transformer models.
- `whisper.encoder.context_length`
- `whisper.encoder.embedding_length`
- `whisper.encoder.block_count`
- `whisper.encoder.mels_count: uint64`
- `whisper.encoder.attention.head_count`
- `whisper.decoder.context_length`
- `whisper.decoder.embedding_length`
- `whisper.decoder.block_count`
- `whisper.decoder.attention.head_count`
#### Prompting
**TODO**: Include prompt format, and/or metadata about how it should be used (instruction, conversation, autocomplete, etc).
### LoRA
**TODO**: Figure out what metadata is needed for LoRA. Probably desired features:
- match an existing model exactly, so that it can't be misapplied
- be marked as a LoRA so executors won't try to run it by itself
Should this be an architecture, or should it share the details of the original model with additional fields to mark it as a LoRA?
### Tokenizer
The following keys are used to describe the tokenizer of the model. It is recommended that model authors support as many of these as possible, as it will allow for better tokenization quality with supported executors.
#### GGML
GGML supports an embedded vocabulary that enables inference of the model, but implementations of tokenization using this vocabulary (i.e. `llama.cpp`'s tokenizer) may have lower accuracy than the original tokenizer used for the model. When a more accurate tokenizer is available and supported, it should be used instead.
It is not guaranteed to be standardized across models, and may change in the future. It is recommended that model authors use a more standardized tokenizer if possible.
- `tokenizer.ggml.model: string`: The name of the tokenizer model.
- `llama`: Llama style SentencePiece (tokens and scores extracted from HF `tokenizer.model`)
- `replit`: Replit style SentencePiece (tokens and scores extracted from HF `spiece.model`)
- `gpt2`: GPT-2 / GPT-NeoX style BPE (tokens extracted from HF `tokenizer.json`)
- `rwkv`: RWKV tokenizer
- `tokenizer.ggml.tokens: array[string]`: A list of tokens indexed by the token ID used by the model.
- `tokenizer.ggml.scores: array[float32]`: If present, the score/probability of each token. If not present, all tokens are assumed to have equal probability. If present, it must have the same length and index as `tokens`.
- `tokenizer.ggml.token_type: array[int32]`: The token type (1=normal, 2=unknown, 3=control, 4=user defined, 5=unused, 6=byte). If present, it must have the same length and index as `tokens`.
- `tokenizer.ggml.merges: array[string]`: If present, the merges of the tokenizer. If not present, the tokens are assumed to be atomic.
- `tokenizer.ggml.added_tokens: array[string]`: If present, tokens that were added after training.
##### Special tokens
- `tokenizer.ggml.bos_token_id: uint32`: Beginning of sequence marker
- `tokenizer.ggml.eos_token_id: uint32`: End of sequence marker
- `tokenizer.ggml.unknown_token_id: uint32`: Unknown token
- `tokenizer.ggml.separator_token_id: uint32`: Separator token
- `tokenizer.ggml.padding_token_id: uint32`: Padding token
#### Hugging Face
Hugging Face maintains their own `tokenizers` library that supports a wide variety of tokenizers. If your executor uses this library, it may be able to use the model's tokenizer directly.
- `tokenizer.huggingface.json: string`: the entirety of the HF `tokenizer.json` for a given model (e.g. <https://huggingface.co/mosaicml/mpt-7b-instruct/blob/main/tokenizer.json>). Included for compatibility with executors that support HF tokenizers directly.
#### Other
Other tokenizers may be used, but are not necessarily standardized. They may be executor-specific. They will be documented here as they are discovered/further developed.
- `tokenizer.rwkv.world: string`: a RWKV World tokenizer, like [this](https://github.com/BlinkDL/ChatRWKV/blob/main/tokenizer/rwkv_vocab_v20230424.txt). This text file should be included verbatim.
- `tokenizer.chat_template : string`: a Jinja template that specifies the input format expected by the model. For more details see: <https://huggingface.co/docs/transformers/main/en/chat_templating>
### Computation graph
This is a future extension and still needs to be discussed, and may necessitate a new GGUF version. At the time of writing, the primary blocker is the stabilization of the computation graph format.
A sample computation graph of GGML nodes could be included in the model itself, allowing an executor to run the model without providing its own implementation of the architecture. This would allow for a more consistent experience across executors, and would allow for more complex architectures to be supported without requiring the executor to implement them.
## Standardized tensor names
To minimize complexity and maximize compatibility, it is recommended that models using the transformer architecture use the following naming convention for their tensors:
### Base layers
`AA.weight` `AA.bias`
where `AA` can be:
- `token_embd`: Token embedding layer
- `pos_embd`: Position embedding layer
- `output_norm`: Output normalization layer
- `output`: Output layer
### Attention and feed-forward layer blocks
`blk.N.BB.weight` `blk.N.BB.bias`
where N signifies the block number a layer belongs to, and where `BB` could be:
- `attn_norm`: Attention normalization layer
- `attn_norm_2`: Attention normalization layer
- `attn_qkv`: Attention query-key-value layer
- `attn_q`: Attention query layer
- `attn_k`: Attention key layer
- `attn_v`: Attention value layer
- `attn_output`: Attention output layer
- `ffn_norm`: Feed-forward network normalization layer
- `ffn_up`: Feed-forward network "up" layer
- `ffn_gate`: Feed-forward network "gate" layer
- `ffn_down`: Feed-forward network "down" layer
- `ffn_gate_inp`: Expert-routing layer for the Feed-forward network in MoE models
- `ffn_gate_exp`: Feed-forward network "gate" layer per expert in MoE models
- `ffn_down_exp`: Feed-forward network "down" layer per expert in MoE models
- `ffn_up_exp`: Feed-forward network "up" layer per expert in MoE models
- `ssm_in`: State space model input projections layer
- `ssm_conv1d`: State space model rolling/shift layer
- `ssm_x`: State space model selective parametrization layer
- `ssm_a`: State space model state compression layer
- `ssm_d`: State space model skip connection layer
- `ssm_dt`: State space model time step layer
- `ssm_out`: State space model output projection layer
## Version History
This document is actively updated to describe the current state of the metadata, and these changes are not tracked outside of the commits.
However, the format _itself_ has changed. The following sections describe the changes to the format itself.
### v3
Adds big-endian support.
### v2
Most countable values (lengths, etc) were changed from `uint32` to `uint64` to allow for larger models to be supported in the future.
### v1
Initial version.
## Historical State of Affairs
The following information is provided for context, but is not necessary to understand the rest of this document.
### Overview
At present, there are three GGML file formats floating around for LLMs:
- **GGML** (unversioned): baseline format, with no versioning or alignment.
- **GGMF** (versioned): the same as GGML, but with versioning. Only one version exists.
- **GGJT**: Aligns the tensors to allow for use with `mmap`, which requires alignment. v1, v2 and v3 are identical, but the latter versions use a different quantization scheme that is incompatible with previous versions.
GGML is primarily used by the examples in `ggml`, while GGJT is used by `llama.cpp` models. Other executors may use any of the three formats, but this is not 'officially' supported.
These formats share the same fundamental structure:
- a magic number with an optional version number
- model-specific hyperparameters, including
- metadata about the model, such as the number of layers, the number of heads, etc.
- a `ftype` that describes the type of the majority of the tensors,
- for GGML files, the quantization version is encoded in the `ftype` divided by 1000
- an embedded vocabulary, which is a list of strings with length prepended. The GGMF/GGJT formats embed a float32 score next to the strings.
- finally, a list of tensors with their length-prepended name, type, and (aligned, in the case of GGJT) tensor data
Notably, this structure does not identify what model architecture the model belongs to, nor does it offer any flexibility for changing the structure of the hyperparameters. This means that the only way to add new hyperparameters is to add them to the end of the list, which is a breaking change for existing models.
### Drawbacks
Unfortunately, over the last few months, there are a few issues that have become apparent with the existing models:
- There's no way to identify which model architecture a given model is for, because that information isn't present
- Similarly, existing programs cannot intelligently fail upon encountering new architectures
- Adding or removing any new hyperparameters is a breaking change, which is impossible for a reader to detect without using heuristics
- Each model architecture requires its own conversion script to their architecture's variant of GGML
- Maintaining backwards compatibility without breaking the structure of the format requires clever tricks, like packing the quantization version into the ftype, which are not guaranteed to be picked up by readers/writers, and are not consistent between the two formats
### Why not other formats?
There are a few other formats that could be used, but issues include:
- requiring additional dependencies to load or save the model, which is complicated in a C environment
- limited or no support for 4-bit quantization
- existing cultural expectations (e.g. whether or not the model is a directory or a file)
- lack of support for embedded vocabularies
- lack of control over direction of future development
Ultimately, it is likely that GGUF will remain necessary for the foreseeable future, and it is better to have a single format that is well-documented and supported by all executors than to contort an existing format to fit the needs of GGML.
engine/ggml/examples/CMakeLists.txt
+34
View File
@@ -0,0 +1,34 @@
if (GGML_ALL_WARNINGS)
if (NOT MSVC)
set(cxx_flags
# TODO(marella): Add other warnings.
-Wpedantic
-Wunused-variable
-Wno-unused-function
-Wno-multichar
)
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>")
endif()
endif()
add_library(common STATIC common.cpp)
target_include_directories(common PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
add_library(common-ggml STATIC common-ggml.cpp)
target_link_libraries(common-ggml PRIVATE ggml)
target_include_directories(common-ggml PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
add_subdirectory(yolo)
if (NOT GGML_BACKEND_DL)
add_subdirectory(gpt-2)
add_subdirectory(gpt-j)
add_subdirectory(mnist)
add_subdirectory(sam)
add_subdirectory(simple)
add_subdirectory(magika)
endif()
if (GGML_METAL)
add_subdirectory(perf-metal)
endif()
engine/ggml/examples/common-ggml.cpp
+244
View File
@@ -0,0 +1,244 @@
#include "common-ggml.h"
#include <regex>
#include <map>
static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
{"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
{"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
{"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
{"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
{"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
{"q2_k", GGML_FTYPE_MOSTLY_Q2_K},
{"q3_k", GGML_FTYPE_MOSTLY_Q3_K},
{"q4_k", GGML_FTYPE_MOSTLY_Q4_K},
{"q5_k", GGML_FTYPE_MOSTLY_Q5_K},
{"q6_k", GGML_FTYPE_MOSTLY_Q6_K},
};
void ggml_print_ftypes(FILE * fp) {
for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
fprintf(fp, " type = \"%s\" or %d\n", it->first.c_str(), it->second);
}
}
enum ggml_ftype ggml_parse_ftype(const char * str) {
enum ggml_ftype ftype;
if (str[0] == 'q') {
const auto it = GGML_FTYPE_MAP.find(str);
if (it == GGML_FTYPE_MAP.end()) {
fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
return GGML_FTYPE_UNKNOWN;
}
ftype = it->second;
} else {
ftype = (enum ggml_ftype) atoi(str);
}
return ftype;
}
bool ggml_common_quantize_0(
std::ifstream & finp,
std::ofstream & fout,
const ggml_ftype ftype,
const std::vector<std::string> & to_quant,
const std::vector<std::string> & to_skip) {
ggml_type qtype = GGML_TYPE_F32;
switch (ftype) {
case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
case GGML_FTYPE_MOSTLY_Q2_K: qtype = GGML_TYPE_Q2_K; break;
case GGML_FTYPE_MOSTLY_Q3_K: qtype = GGML_TYPE_Q3_K; break;
case GGML_FTYPE_MOSTLY_Q4_K: qtype = GGML_TYPE_Q4_K; break;
case GGML_FTYPE_MOSTLY_Q5_K: qtype = GGML_TYPE_Q5_K; break;
case GGML_FTYPE_MOSTLY_Q6_K: qtype = GGML_TYPE_Q6_K; break;
case GGML_FTYPE_UNKNOWN:
case GGML_FTYPE_ALL_F32:
case GGML_FTYPE_MOSTLY_F16:
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
case GGML_FTYPE_MOSTLY_IQ2_XXS:
case GGML_FTYPE_MOSTLY_IQ2_XS:
case GGML_FTYPE_MOSTLY_IQ2_S:
case GGML_FTYPE_MOSTLY_IQ3_XXS:
case GGML_FTYPE_MOSTLY_IQ3_S:
case GGML_FTYPE_MOSTLY_IQ1_S:
case GGML_FTYPE_MOSTLY_IQ4_NL:
case GGML_FTYPE_MOSTLY_IQ4_XS:
case GGML_FTYPE_MOSTLY_IQ1_M:
case GGML_FTYPE_MOSTLY_BF16:
case GGML_FTYPE_MOSTLY_MXFP4:
case GGML_FTYPE_MOSTLY_NVFP4:
case GGML_FTYPE_MOSTLY_Q1_0:
{
fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
return false;
}
};
if (!ggml_is_quantized(qtype)) {
fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
return false;
}
size_t total_size_org = 0;
size_t total_size_new = 0;
std::vector<float> work;
std::vector<uint8_t> data_u8;
std::vector<ggml_fp16_t> data_f16;
std::vector<float> data_f32;
while (true) {
int32_t n_dims;
int32_t length;
int32_t ttype;
finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
finp.read(reinterpret_cast<char *>(&length), sizeof(length));
finp.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
if (finp.eof()) {
break;
}
int32_t nelements = 1;
int32_t ne[4] = { 1, 1, 1, 1 };
for (int i = 0; i < n_dims; ++i) {
finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
nelements *= ne[i];
}
std::string name(length, 0);
finp.read (&name[0], length);
printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
bool quantize = false;
// check if we should quantize this tensor
for (const auto & s : to_quant) {
if (std::regex_match(name, std::regex(s))) {
quantize = true;
break;
}
}
// check if we should skip this tensor
for (const auto & s : to_skip) {
if (std::regex_match(name, std::regex(s))) {
quantize = false;
break;
}
}
// quantize only 2D tensors
quantize &= (n_dims == 2);
if (quantize) {
if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
return false;
}
if (ttype == GGML_TYPE_F16) {
data_f16.resize(nelements);
finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
data_f32.resize(nelements);
for (int i = 0; i < nelements; ++i) {
data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
}
} else {
data_f32.resize(nelements);
finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
}
ttype = qtype;
} else {
const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
data_u8.resize(nelements*bpe);
finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
}
fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
fout.write(reinterpret_cast<char *>(&length), sizeof(length));
fout.write(reinterpret_cast<char *>(&ttype), sizeof(ttype));
for (int i = 0; i < n_dims; ++i) {
fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
}
fout.write(&name[0], length);
if (quantize) {
work.resize(nelements); // for quantization
size_t cur_size = 0;
switch ((ggml_type) ttype) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
{
cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], nullptr);
} break;
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
case GGML_TYPE_I32:
case GGML_TYPE_I64:
case GGML_TYPE_F64:
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q8_K:
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_BF16:
case GGML_TYPE_TQ1_0:
case GGML_TYPE_TQ2_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_NVFP4:
case GGML_TYPE_Q1_0:
case GGML_TYPE_COUNT:
{
fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
return false;
}
}
fout.write(reinterpret_cast<char *>(work.data()), cur_size);
total_size_new += cur_size;
printf("size = %8.2f MB -> %8.2f MB\n", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
} else {
printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
total_size_new += data_u8.size();
}
total_size_org += nelements * sizeof(float);
}
printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
printf("%s: quant size = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
return true;
}
engine/ggml/examples/common-ggml.h
+18
View File
@@ -0,0 +1,18 @@
#pragma once
#include "ggml.h"
#include <fstream>
#include <vector>
#include <string>
enum ggml_ftype ggml_parse_ftype(const char * str);
void ggml_print_ftypes(FILE * fp = stderr);
bool ggml_common_quantize_0(
std::ifstream & finp,
std::ofstream & fout,
const ggml_ftype ftype,
const std::vector<std::string> & to_quant,
const std::vector<std::string> & to_skip);
engine/ggml/examples/common.cpp
+675
View File
@@ -0,0 +1,675 @@
#define _USE_MATH_DEFINES // for M_PI
#include "common.h"
#include <cmath>
#include <codecvt>
#include <cstring>
#include <fstream>
#include <locale>
#include <regex>
#include <sstream>
// Function to check if the next argument exists
static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
if (i + 1 < argc && argv[i + 1][0] != '-') {
return argv[++i];
} else {
fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-p" || arg == "--prompt") {
params.prompt = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-n" || arg == "--n_predict") {
params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-np" || arg == "--n_parallel") {
params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_k") {
params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_p") {
params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--temp") {
params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-last-n") {
params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-penalty") {
params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-b" || arg == "--batch_size") {
params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-c" || arg == "--context") {
params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--ignore-eos") {
params.ignore_eos = true;
} else if (arg == "-m" || arg == "--model") {
params.model = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-i" || arg == "--interactive") {
params.interactive = true;
} else if (arg == "-ip" || arg == "--interactive-port") {
params.interactive = true;
params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, params);
exit(0);
} else if (arg == "-f" || arg == "--file") {
get_next_arg(i, argc, argv, arg, params);
std::ifstream file(argv[i]);
if (!file) {
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
break;
}
std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
if (params.prompt.back() == '\n') {
params.prompt.pop_back();
}
} else if (arg == "-tt" || arg == "--token_test") {
params.token_test = get_next_arg(i, argc, argv, arg, params);
}
else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
return true;
}
void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
fprintf(stderr, " prompt to start generation with (default: random)\n");
fprintf(stderr, " -f FNAME, --file FNAME\n");
fprintf(stderr, " load prompt from a file\n");
fprintf(stderr, " -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
fprintf(stderr, " test tokenization\n");
fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d)\n", params.n_predict);
fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k);
fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", params.top_p);
fprintf(stderr, " --temp N temperature (default: %.1f)\n", params.temp);
fprintf(stderr, " --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
fprintf(stderr, " --repeat-penalty N penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stderr, " -c N, --context N context / KV cache size (default: %d)\n", params.n_ctx);
fprintf(stderr, " --ignore-eos ignore EOS token during generation\n");
fprintf(stderr, " -ngl N, --gpu-layers N number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, "\n");
}
std::string gpt_random_prompt(std::mt19937 & rng) {
const int r = rng() % 10;
switch (r) {
case 0: return "So";
case 1: return "Once upon a time";
case 2: return "When";
case 3: return "The";
case 4: return "After";
case 5: return "If";
case 6: return "import";
case 7: return "He";
case 8: return "She";
case 9: return "They";
}
return "The";
}
std::string trim(const std::string & s) {
std::regex e("^\\s+|\\s+$");
return std::regex_replace(s, e, "");
}
std::string replace(const std::string & s, const std::string & from, const std::string & to) {
std::string result = s;
size_t pos = 0;
while ((pos = result.find(from, pos)) != std::string::npos) {
result.replace(pos, from.length(), to);
pos += to.length();
}
return result;
}
void gpt_vocab::add_special_token(const std::string & token) {
special_tokens.push_back(token);
}
std::map<std::string, int32_t> json_parse(const std::string & fname) {
std::map<std::string, int32_t> result;
// read file into string
std::string json;
{
std::ifstream ifs(fname);
if (!ifs) {
fprintf(stderr, "Failed to open %s\n", fname.c_str());
exit(1);
}
json = std::string((std::istreambuf_iterator<char>(ifs)),
(std::istreambuf_iterator<char>()));
}
if (json[0] != '{') {
return result;
}
// parse json
{
bool has_key = false;
bool in_token = false;
std::string str_key = "";
std::string str_val = "";
int n = json.size();
for (int i = 1; i < n; ++i) {
if (!in_token) {
if (json[i] == ' ') continue;
if (json[i] == '"') {
in_token = true;
continue;
}
} else {
if (json[i] == '\\' && i+1 < n) {
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
++i;
} else if (json[i] == '"') {
if (has_key == false) {
has_key = true;
++i;
while (json[i] == ' ') ++i;
++i; // :
while (json[i] == ' ') ++i;
if (json[i] != '\"') {
while (json[i] != ',' && json[i] != '}') {
str_val += json[i++];
}
has_key = false;
} else {
in_token = true;
continue;
}
} else {
has_key = false;
}
str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
str_key = ::replace(str_key, "\\\"", "\""); // \\\" -> "
try {
result[str_key] = std::stoi(str_val);
} catch (...) {
//fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
}
str_key = "";
str_val = "";
in_token = false;
continue;
}
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
}
}
}
return result;
}
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;
}
static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
std::vector<gpt_vocab::id> output;
std::stringstream ss(input);
std::string token;
while (std::getline(ss, token, delimiter)) {
output.push_back(std::stoi(token));
}
return output;
}
static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
if (fpath_test.empty()){
fprintf(stderr, "%s : No test file found.\n", __func__);
return std::map<std::string, std::vector<gpt_vocab::id>>();
}
std::map<std::string, std::vector<gpt_vocab::id>> tests;
auto fin = std::ifstream(fpath_test, std::ios_base::in);
const char * delimeter = " => ";
const char del_tok = ',';
std::string line;
while (std::getline(fin, line)) {
size_t delimiterPos = line.find(delimeter);
if (delimiterPos != std::string::npos) {
std::string text = line.substr(0, delimiterPos);
std::string s_tokens = line.substr(delimiterPos + std::strlen(delimeter));
tests[text] = parse_tokens_from_string(s_tokens, del_tok);
}
}
return tests;
}
void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
std::map<std::string, std::vector<gpt_vocab::id>> tests = extract_tests_from_file(fpath_test);
size_t n_fails = 0;
for (const auto & test : tests) {
std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, test.first);
if (tokens != test.second){
n_fails++;
// print out failure cases
fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
fprintf(stderr, "%s : tokens in hf: ", __func__);
for (const auto & t : test.second) {
fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
}
fprintf(stderr, "\n");
fprintf(stderr, "%s : tokens in ggml: ", __func__);
for (const auto & t : tokens) {
fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
}
fprintf(stderr, "\n");
}
}
fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
}
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
vocab.token_to_id = ::json_parse(fname);
for (const auto & kv : vocab.token_to_id) {
vocab.id_to_token[kv.second] = kv.first;
}
printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
// print the vocabulary
//for (auto kv : vocab.token_to_id) {
// printf("'%s' -> %d\n", kv.first.data(), kv.second);
//}
return true;
}
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;
}
gpt_vocab::id gpt_sample_top_k_top_p_repeat(
const gpt_vocab & vocab,
const float * logits,
const int32_t * last_n_tokens_data,
size_t last_n_tokens_data_size,
int top_k,
double top_p,
double temp,
int repeat_last_n,
float repeat_penalty,
std::mt19937 & rng) {
int n_logits = vocab.id_to_token.size();
const auto * plogits = logits;
const auto last_n_tokens = std::vector<int32_t>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_data_size);
if (temp <= 0) {
// select the token with the highest logit directly
float max_logit = plogits[0];
gpt_vocab::id max_id = 0;
for (int i = 1; i < n_logits; ++i) {
if (plogits[i] > max_logit) {
max_logit = plogits[i];
max_id = i;
}
}
return max_id;
}
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
logits_id.reserve(n_logits);
{
const float scale = 1.0f/temp;
for (int i = 0; i < n_logits; ++i) {
// repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
// credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
if (repeat_last_n > 0 && std::find(last_n_tokens.end()-repeat_last_n, last_n_tokens.end(), i) != last_n_tokens.end()) {
// if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
if (plogits[i] < 0.0f) {
logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
} else {
logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
}
} else {
logits_id.push_back(std::make_pair(plogits[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++) {
// for (int i = 0; i < 10; i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
// }
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}
void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
const float rc = 1.0f / (2.0f * M_PI * cutoff);
const float dt = 1.0f / sample_rate;
const float alpha = dt / (rc + dt);
float y = data[0];
for (size_t i = 1; i < data.size(); i++) {
y = alpha * (y + data[i] - data[i - 1]);
data[i] = y;
}
}
bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
const int n_samples = pcmf32.size();
const int n_samples_last = (sample_rate * last_ms) / 1000;
if (n_samples_last >= n_samples) {
// not enough samples - assume no speech
return false;
}
if (freq_thold > 0.0f) {
high_pass_filter(pcmf32, freq_thold, sample_rate);
}
float energy_all = 0.0f;
float energy_last = 0.0f;
for (int i = 0; i < n_samples; i++) {
energy_all += fabsf(pcmf32[i]);
if (i >= n_samples - n_samples_last) {
energy_last += fabsf(pcmf32[i]);
}
}
energy_all /= n_samples;
energy_last /= n_samples_last;
if (verbose) {
fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
}
if (energy_last > vad_thold*energy_all) {
return false;
}
return true;
}
float similarity(const std::string & s0, const std::string & s1) {
const size_t len0 = s0.size() + 1;
const size_t len1 = s1.size() + 1;
std::vector<int> col(len1, 0);
std::vector<int> prevCol(len1, 0);
for (size_t i = 0; i < len1; i++) {
prevCol[i] = i;
}
for (size_t i = 0; i < len0; i++) {
col[0] = i;
for (size_t j = 1; j < len1; j++) {
col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (i > 0 && s0[i - 1] == s1[j - 1] ? 0 : 1));
}
col.swap(prevCol);
}
const float dist = prevCol[len1 - 1];
return 1.0f - (dist / std::max(s0.size(), s1.size()));
}
bool is_file_exist(const char * filename) {
std::ifstream infile(filename);
return infile.good();
}
engine/ggml/examples/common.h
+322
View File
@@ -0,0 +1,322 @@
// Various helper functions and utilities
#pragma once
#include <string>
#include <map>
#include <vector>
#include <random>
#include <thread>
#include <ctime>
#include <fstream>
#include <sstream>
//
// GPT CLI argument parsing
//
struct gpt_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
int32_t n_predict = 200; // new tokens to predict
int32_t n_parallel = 1; // number of parallel streams
int32_t n_batch = 32; // batch size for prompt processing
int32_t n_ctx = 2048; // context size (this is the KV cache max size)
int32_t n_gpu_layers = 0; // number of layers to offlload to the GPU
bool ignore_eos = false; // ignore EOS token when generating text
// sampling parameters
int32_t top_k = 40;
float top_p = 0.9f;
float temp = 0.9f;
int32_t repeat_last_n = 64;
float repeat_penalty = 1.00f;
std::string model = "models/gpt-2-117M/ggml-model.bin"; // model path
std::string prompt = "";
std::string token_test = "";
bool interactive = false;
int32_t interactive_port = -1;
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
std::string gpt_random_prompt(std::mt19937 & rng);
//
// Vocab utils
//
std::string trim(const std::string & s);
std::string replace(
const std::string & s,
const std::string & from,
const std::string & to);
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);
};
// poor-man's JSON parsing
std::map<std::string, int32_t> json_parse(const std::string & fname);
std::string convert_to_utf8(const std::wstring & input);
std::wstring convert_to_wstring(const std::string & input);
void gpt_split_words(std::string str, std::vector<std::string>& words);
// split text into tokens
//
// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
//
// Regex (Python):
// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
//
// Regex (C++):
// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
//
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
// test outputs of gpt_tokenize
//
// - compare with tokens generated by the huggingface tokenizer
// - test cases are chosen based on the model's main language (under 'prompt' directory)
// - if all sentences are tokenized identically, print 'All tests passed.'
// - otherwise, print sentence, huggingface tokens, ggml tokens
//
void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test);
// load the tokens from encoder.json
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
// sample next token given probabilities for each embedding
//
// - consider only the top K tokens
// - from them, consider only the top tokens with cumulative probability > P
//
// TODO: not sure if this implementation is correct
// TODO: temperature is not implemented
//
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);
gpt_vocab::id gpt_sample_top_k_top_p_repeat(
const gpt_vocab & vocab,
const float * logits,
const int32_t * last_n_tokens_data,
size_t last_n_tokens_data_size,
int top_k,
double top_p,
double temp,
int repeat_last_n,
float repeat_penalty,
std::mt19937 & rng);
//
// Audio utils
//
// Write PCM data into WAV audio file
class wav_writer {
private:
std::ofstream file;
uint32_t dataSize = 0;
std::string wav_filename;
bool write_header(const uint32_t sample_rate,
const uint16_t bits_per_sample,
const uint16_t channels) {
file.write("RIFF", 4);
file.write("\0\0\0\0", 4); // Placeholder for file size
file.write("WAVE", 4);
file.write("fmt ", 4);
const uint32_t sub_chunk_size = 16;
const uint16_t audio_format = 1; // PCM format
const uint32_t byte_rate = sample_rate * channels * bits_per_sample / 8;
const uint16_t block_align = channels * bits_per_sample / 8;
file.write(reinterpret_cast<const char *>(&sub_chunk_size), 4);
file.write(reinterpret_cast<const char *>(&audio_format), 2);
file.write(reinterpret_cast<const char *>(&channels), 2);
file.write(reinterpret_cast<const char *>(&sample_rate), 4);
file.write(reinterpret_cast<const char *>(&byte_rate), 4);
file.write(reinterpret_cast<const char *>(&block_align), 2);
file.write(reinterpret_cast<const char *>(&bits_per_sample), 2);
file.write("data", 4);
file.write("\0\0\0\0", 4); // Placeholder for data size
return true;
}
// It is assumed that PCM data is normalized to a range from -1 to 1
bool write_audio(const float * data, size_t length) {
for (size_t i = 0; i < length; ++i) {
const int16_t intSample = int16_t(data[i] * 32767);
file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
dataSize += sizeof(int16_t);
}
if (file.is_open()) {
file.seekp(4, std::ios::beg);
uint32_t fileSize = 36 + dataSize;
file.write(reinterpret_cast<char *>(&fileSize), 4);
file.seekp(40, std::ios::beg);
file.write(reinterpret_cast<char *>(&dataSize), 4);
file.seekp(0, std::ios::end);
}
return true;
}
bool open_wav(const std::string & filename) {
if (filename != wav_filename) {
if (file.is_open()) {
file.close();
}
}
if (!file.is_open()) {
file.open(filename, std::ios::binary);
wav_filename = filename;
dataSize = 0;
}
return file.is_open();
}
public:
bool open(const std::string & filename,
const uint32_t sample_rate,
const uint16_t bits_per_sample,
const uint16_t channels) {
if (open_wav(filename)) {
write_header(sample_rate, bits_per_sample, channels);
} else {
return false;
}
return true;
}
bool close() {
file.close();
return true;
}
bool write(const float * data, size_t length) {
return write_audio(data, length);
}
~wav_writer() {
if (file.is_open()) {
file.close();
}
}
};
// Apply a high-pass frequency filter to PCM audio
// Suppresses frequencies below cutoff Hz
void high_pass_filter(
std::vector<float> & data,
float cutoff,
float sample_rate);
// Basic voice activity detection (VAD) using audio energy adaptive threshold
bool vad_simple(
std::vector<float> & pcmf32,
int sample_rate,
int last_ms,
float vad_thold,
float freq_thold,
bool verbose);
// compute similarity between two strings using Levenshtein distance
float similarity(const std::string & s0, const std::string & s1);
//
// Terminal utils
//
#define SQR(X) ((X) * (X))
#define UNCUBE(x) x < 48 ? 0 : x < 115 ? 1 : (x - 35) / 40
/**
* Quantizes 24-bit RGB to xterm256 code range [16,256).
*/
static int rgb2xterm256(int r, int g, int b) {
unsigned char cube[] = {0, 0137, 0207, 0257, 0327, 0377};
int av, ir, ig, ib, il, qr, qg, qb, ql;
av = r * .299 + g * .587 + b * .114 + .5;
ql = (il = av > 238 ? 23 : (av - 3) / 10) * 10 + 8;
qr = cube[(ir = UNCUBE(r))];
qg = cube[(ig = UNCUBE(g))];
qb = cube[(ib = UNCUBE(b))];
if (SQR(qr - r) + SQR(qg - g) + SQR(qb - b) <=
SQR(ql - r) + SQR(ql - g) + SQR(ql - b))
return ir * 36 + ig * 6 + ib + 020;
return il + 0350;
}
static std::string set_xterm256_foreground(int r, int g, int b) {
int x = rgb2xterm256(r, g, b);
std::ostringstream oss;
oss << "\033[38;5;" << x << "m";
return oss.str();
}
// Lowest is red, middle is yellow, highest is green. Color scheme from
// Paul Tol; it is colorblind friendly https://sronpersonalpages.nl/~pault
const std::vector<std::string> k_colors = {
set_xterm256_foreground(220, 5, 12),
set_xterm256_foreground(232, 96, 28),
set_xterm256_foreground(241, 147, 45),
set_xterm256_foreground(246, 193, 65),
set_xterm256_foreground(247, 240, 86),
set_xterm256_foreground(144, 201, 135),
set_xterm256_foreground( 78, 178, 101),
};
// ANSI formatting codes
static std::string set_inverse() {
return "\033[7m";
}
static std::string set_underline() {
return "\033[4m";
}
static std::string set_dim() {
return "\033[2m";
}
// Style scheme for different confidence levels
const std::vector<std::string> k_styles = {
set_inverse(), // Low confidence - inverse (highlighted)
set_underline(), // Medium confidence - underlined
set_dim(), // High confidence - dim
};
//
// Other utils
//
// check if file exists using ifstream
bool is_file_exist(const char * filename);
engine/ggml/examples/gpt-2/CMakeLists.txt
+32
View File
@@ -0,0 +1,32 @@
#
# gpt-2
set(TEST_TARGET gpt-2-ctx)
add_executable(${TEST_TARGET} main-ctx.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
set(TEST_TARGET gpt-2-alloc)
add_executable(${TEST_TARGET} main-alloc.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
set(TEST_TARGET gpt-2-backend)
add_executable(${TEST_TARGET} main-backend.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
set(TEST_TARGET gpt-2-sched)
add_executable(${TEST_TARGET} main-sched.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
#
# gpt-2-quantize
set(TEST_TARGET gpt-2-quantize)
add_executable(${TEST_TARGET} quantize.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
#
# gpt-2-batched
set(TEST_TARGET gpt-2-batched)
add_executable(${TEST_TARGET} main-batched.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
engine/ggml/examples/gpt-2/README.md
+225
View File
@@ -0,0 +1,225 @@
# gpt-2
This is a C++ example running GPT-2 inference using the [ggml](https://github.com/ggerganov/ggml) library.
The program runs on the CPU - no video card is required.
The [Cerebras-GPT](https://huggingface.co/cerebras) models are also supported.
The example supports the following GPT-2 models:
| Model | Description | Disk Size |
| --- | --- | --- |
| 117M | Small model | 240 MB |
| 345M | Medium model | 680 MB |
| 774M | Large model | 1.5 GB |
| 1558M | XL model | 3.0 GB |
Sample performance on MacBook M1 Pro:
| Model | Size | Time / Token |
| --- | --- | --- |
| GPT-2 | 117M | 5 ms |
| GPT-2 | 345M | 12 ms |
| GPT-2 | 774M | 23 ms |
| GPT-2 | 1558M | 42 ms |
*TODO: add tables for Cerebras-GPT models*
Sample output:
```bash
$ ./bin/gpt-2 -h
usage: ./bin/gpt-2 [options]
options:
-h, --help show this help message and exit
-s SEED, --seed SEED RNG seed (default: -1)
-t N, --threads N number of threads to use during computation (default: 8)
-p PROMPT, --prompt PROMPT
prompt to start generation with (default: random)
-n N, --n_predict N number of tokens to predict (default: 200)
--top_k N top-k sampling (default: 40)
--top_p N top-p sampling (default: 0.9)
--temp N temperature (default: 1.0)
-b N, --batch_size N batch size for prompt processing (default: 8)
-m FNAME, --model FNAME
model path (default: models/gpt-2-117M/ggml-model.bin)
$ ./bin/gpt-2
gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
gpt2_model_load: n_vocab = 50257
gpt2_model_load: n_ctx = 1024
gpt2_model_load: n_embd = 768
gpt2_model_load: n_head = 12
gpt2_model_load: n_layer = 12
gpt2_model_load: f16 = 1
gpt2_model_load: ggml ctx size = 311.12 MB
gpt2_model_load: memory size = 72.00 MB, n_mem = 12288
gpt2_model_load: model size = 239.08 MB
main: number of tokens in prompt = 1
So this is going to be the end of the line for us.
If the Dolphins continue to do their business, it's possible that the team could make a bid to bring in new defensive coordinator Scott Linehan.
Linehan's job is a little daunting, but he's a great coach and an excellent coach. I don't believe we're going to make the playoffs.
We're going to have to work hard to keep our heads down and get ready to go.<|endoftext|>
main: mem per token = 2048612 bytes
main: load time = 106.32 ms
main: sample time = 7.10 ms
main: predict time = 506.40 ms / 5.06 ms per token
main: total time = 629.84 ms
```
## Downloading and converting the original models (GPT-2)
You can download the original model files using the [download-model.sh](download-model.sh) Bash script. The models are
in Tensorflow format, so in order to use them with ggml, you need to convert them to appropriate format. This is done
via the [convert-ckpt-to-ggml.py](convert-ckpt-to-ggml.py) python script.
Here is the entire process for the GPT-2 117M model (download from official site + conversion):
```bash
cd ggml/build
../examples/gpt-2/download-model.sh 117M
Downloading model 117M ...
models/gpt-2-117M/checkpoint 100%[=============================>] 77 --.-KB/s in 0s
models/gpt-2-117M/encoder.json 100%[=============================>] 1018K 1.20MB/s in 0.8s
models/gpt-2-117M/hparams.json 100%[=============================>] 90 --.-KB/s in 0s
models/gpt-2-117M/model.ckpt.data-00000-of-00001 100%[=============================>] 474.70M 1.21MB/s in 8m 39s
models/gpt-2-117M/model.ckpt.index 100%[=============================>] 5.09K --.-KB/s in 0s
models/gpt-2-117M/model.ckpt.meta 100%[=============================>] 460.11K 806KB/s in 0.6s
models/gpt-2-117M/vocab.bpe 100%[=============================>] 445.62K 799KB/s in 0.6s
Done! Model '117M' saved in 'models/gpt-2-117M/'
Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.
python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/ 1
```
This conversion requires that you have python and Tensorflow installed on your computer. Still, if you want to avoid
this, you can download the already converted ggml models as described below.
## Downloading and converting the original models (Cerebras-GPT)
Clone the respective repository from here: https://huggingface.co/cerebras
Use the [convert-cerebras-to-ggml.py](convert-cerebras-to-ggml.py) script to convert the model to `ggml` format:
```bash
cd ggml/build
git clone https://huggingface.co/cerebras/Cerebras-GPT-111M models/
python ../examples/gpt-2/convert-cerebras-to-ggml.py models/Cerebras-GPT-111M/
```
## Downloading the ggml model directly (GPT-2)
For convenience, I will be hosting the converted ggml model files in order to make it easier to run the examples. This
way, you can directly download a single binary file and start using it. No python or Tensorflow is required.
Here is how to get the 117M ggml model:
```bash
cd ggml/build
../examples/gpt-2/download-ggml-model.sh 117M
Downloading ggml model 117M ...
models/gpt-2-117M/ggml-model.bin 100%[===============================>] 239.58M 8.52MB/s in 28s
Done! Model '117M' saved in 'models/gpt-2-117M/ggml-model.bin'
You can now use it like this:
$ ./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
```
At some point, I might decide to stop hosting these models. So in that case, simply revert to the manual process above.
## Quantizing the models
You can also try to quantize the `ggml` models via 4-bit integer quantization.
Keep in mind that for smaller models, this will render them completely useless.
You generally want to quantize larger models.
```bash
# quantize GPT-2 F16 to Q4_0 (faster but less precise)
./bin/gpt-2-quantize models/gpt-2-1558M/ggml-model-f16.bin models/gpt-2-1558M/ggml-model-q4_0.bin 2
./bin/gpt-2 -m models/gpt-2-1558M/ggml-model-q4_0.bin -p "This is an example"
# quantize Cerebras F16 to Q4_1 (slower but more precise)
./bin/gpt-2-quantize models/Cerebras-GPT-6.7B/ggml-model-f16.bin models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin 3
./bin/gpt-2 -m models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin -p "This is an example"
```
## Batched generation example
You can try the batched generation from a given prompt using the gpt-2-batched binary.
Sample output:
```bash
$ gpt-2-batched -np 5 -m models/gpt-2-117M/ggml-model.bin -p "Hello my name is" -n 50
main: seed = 1697037431
gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
gpt2_model_load: n_vocab = 50257
gpt2_model_load: n_ctx = 1024
gpt2_model_load: n_embd = 768
gpt2_model_load: n_head = 12
gpt2_model_load: n_layer = 12
gpt2_model_load: ftype = 1
gpt2_model_load: qntvr = 0
gpt2_model_load: ggml tensor size = 320 bytes
gpt2_model_load: backend buffer size = 312.72 MB
ggml_init_cublas: found 1 CUDA devices:
Device 0: NVIDIA GeForce GTX 1660, compute capability 7.5
gpt2_model_load: using CPU backend
gpt2_model_load: memory size = 72.00 MB, n_mem = 12288
gpt2_model_load: model size = 239.08 MB
extract_tests_from_file : No test file found.
test_gpt_tokenizer : 0 tests failed out of 0 tests.
main: compute buffer size: 3.26 MB
main: generating 5 sequences ...
main: prompt: 'Hello my name is'
main: number of tokens in prompt = 4, first 8 tokens: 15496 616 1438 318
sequence 0:
Hello my name is John. You can call me any way you want, if you want, but for my very first date, I will be on the phone with you. We're both in our early 20s, but I feel like it's all
sequence 1:
Hello my name is Robert, and I want to say that we're proud to have your company here on the world's largest platform for sharing your stories with us. This is a huge opportunity for our community. We have hundreds of people on this team and
sequence 2:
Hello my name is Jack. I'm the one who created you.
Jack is a boy with a big smile and a big heart. He is a handsome guy. He loves the outdoors and loves the people he meets. He wants to be a
sequence 3:
Hello my name is John. I am a Canadian citizen with a large number of family in Quebec and I am interested in studying. My aim is to take up a post in the Journal of the International Academy of Sciences of Canada which I am currently finishing.
sequence 4:
Hello my name is Dan. I am an entrepreneur. I am a great father. I am a great husband. I am a great husband. I am a great dad. And I am a great husband.
I love my life. I love
main: load time = 880.80 ms
main: sample time = 91.43 ms
main: predict time = 2518.29 ms
main: total time = 3544.32 ms
```
engine/ggml/examples/gpt-2/convert-cerebras-to-ggml.py
+183
View File
@@ -0,0 +1,183 @@
# Convert Cerebras models to ggml format
#
# ref: https://www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/
#
import sys
import struct
import json
import torch
import numpy as np
import re
from transformers import AutoModelForCausalLM
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
cs = [chr(n) for n in cs]
return dict(zip(bs, cs))
if len(sys.argv) < 2:
print("Usage: convert-cerebras-to-ggml.py dir-model [use-f32]\n")
sys.exit(1)
# output in the same directory as the model
dir_model = sys.argv[1]
fname_out = sys.argv[1] + "/ggml-model-f16.bin"
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
encoder = json.load(f)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
# use 16-bit or 32-bit floats
use_f16 = True
if len(sys.argv) > 2:
use_f16 = False
fname_out = sys.argv[1] + "/ggml-model-f32.bin"
model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
#print (model)
list_vars = model.state_dict()
#print (list_vars)
print(hparams)
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["vocab_size"]))
fout.write(struct.pack("i", hparams["n_positions"]))
fout.write(struct.pack("i", hparams["n_embd"]))
fout.write(struct.pack("i", hparams["n_head"]))
fout.write(struct.pack("i", hparams["n_layer"]))
fout.write(struct.pack("i", use_f16))
byte_encoder = bytes_to_unicode()
byte_decoder = {v:k for k, v in byte_encoder.items()}
fout.write(struct.pack("i", len(encoder)))
for key in encoder:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for name in list_vars.keys():
data = list_vars[name].squeeze().numpy()
print("Processing variable: " + name + " with shape: ", data.shape)
# rename headers to keep compatibility
if name == "transformer.ln_f.weight":
name = "model/ln_f/g"
elif name == "transformer.ln_f.bias":
name = "model/ln_f/b"
elif name == "transformer.wte.weight":
name = "model/wte"
elif name == "transformer.wpe.weight":
name = "model/wpe"
elif name == "lm_head.weight":
name = "model/lm_head"
elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_1/g"
elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_1/b"
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_attn/w"
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_attn/b"
elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_proj/w"
elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_proj/b"
elif re.match(r"transformer.h.\d+.ln_2.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_2/g"
elif re.match(r"transformer.h.\d+.ln_2.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_2/b"
elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_fc/w"
elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_fc/b"
elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_proj/w"
elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_proj/b"
else:
print("Unrecognized variable name. %s", name)
# we don't need these
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
print(" Skipping variable: " + name)
continue
n_dims = len(data.shape);
# ftype == 0 -> float32, ftype == 1 -> float16
ftype = 0;
if use_f16:
if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
print(" Converting to float16")
data = data.astype(np.float16)
ftype = 1
else:
print(" Converting to float32")
data = data.astype(np.float32)
ftype = 0
# for efficiency - transpose the projection matrices
# "model/h.*/attn/c_attn/w"
# "model/h.*/attn/c_proj/w"
# "model/h.*/mlp/c_fc/w"
# "model/h.*/mlp/c_proj/w"
if name[-14:] == "/attn/c_attn/w" or \
name[-14:] == "/attn/c_proj/w" or \
name[-11:] == "/mlp/c_fc/w" or \
name[-13:] == "/mlp/c_proj/w":
print(" Transposing")
data = data.transpose()
# header
str = name.encode('utf-8')
fout.write(struct.pack("iii", n_dims, len(str), ftype))
for i in range(n_dims):
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
fout.write(str);
# data
data.tofile(fout)
fout.close()
print("Done. Output file: " + fname_out)
print("")
engine/ggml/examples/gpt-2/convert-ckpt-to-ggml.py
+159
View File
@@ -0,0 +1,159 @@
# Convert a model checkpoint to a ggml compatible file
#
# Load the model using TensorFlow.
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
# - Number of dimensions (int)
# - Name length (int)
# - Dimensions (int[n_dims])
# - Name (char[name_length])
# - Data (float[n_dims])
#
# By default, the bigger matrices are converted to 16-bit floats.
# This can be disabled by adding the "use-f32" CLI argument.
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#
import sys
import json
import struct
import numpy as np
import tensorflow as tf
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
cs = [chr(n) for n in cs]
return dict(zip(bs, cs))
# helper method to convert a numpy array to different float types
def convert_to_ftype(data, ftype):
# fp16
if ftype == 1:
return data.astype(np.float16)
assert False, "Invalid ftype: " + str(ftype)
if len(sys.argv) < 3:
print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
print(" ftype == 0 -> float32")
print(" ftype == 1 -> float16")
sys.exit(1)
# output in the same directory as the model
dir_model = sys.argv[1]
fname_out = sys.argv[1] + "/ggml-model.bin"
with open(dir_model + "/encoder.json", "r", encoding="utf-8") as f:
encoder = json.load(f)
with open(dir_model + "/hparams.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
# possible data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if len(sys.argv) > 2:
ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1:
print("Invalid ftype: " + str(ftype))
sys.exit(1)
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
list_vars = tf.train.list_variables(dir_model)
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["n_vocab"]))
fout.write(struct.pack("i", hparams["n_ctx"]))
fout.write(struct.pack("i", hparams["n_embd"]))
fout.write(struct.pack("i", hparams["n_head"]))
fout.write(struct.pack("i", hparams["n_layer"]))
fout.write(struct.pack("i", ftype))
byte_encoder = bytes_to_unicode()
byte_decoder = {v:k for k, v in byte_encoder.items()}
fout.write(struct.pack("i", len(encoder)))
for key in encoder:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for name, shape in list_vars:
print("Processing variable: " + name + " with shape: ", shape)
data = tf.train.load_variable(dir_model, name).squeeze()
n_dims = len(data.shape);
# for efficiency - transpose the projection matrices
# "model/h.*/attn/c_attn/w"
# "model/h.*/attn/c_proj/w"
# "model/h.*/mlp/c_fc/w"
# "model/h.*/mlp/c_proj/w"
if name[-14:] == "/attn/c_attn/w" or \
name[-14:] == "/attn/c_proj/w" or \
name[-11:] == "/mlp/c_fc/w" or \
name[-13:] == "/mlp/c_proj/w":
print(" Transposing")
data = data.transpose()
dshape = data.shape
ftype_cur = 0
if ftype != 0:
# match name:
# "model/wte"
# "model/h.*/attn/c_attn/w"
# "model/h.*/attn/c_proj/w"
# "model/h.*/mlp/c_fc/w"
# "model/h.*/mlp/c_proj/w"
if name == "model/wte" or name[-2:] == "/w":
print(" Converting to " + ftype_str[ftype])
data = convert_to_ftype(data, ftype)
ftype_cur = ftype
else:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
# header
str = name.encode('utf-8')
fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
for i in range(n_dims):
fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
fout.write(str);
# data
data.tofile(fout)
fout.close()
print("Done. Output file: " + fname_out)
print("")
engine/ggml/examples/gpt-2/convert-h5-to-ggml.py
+195
View File
@@ -0,0 +1,195 @@
# Convert GPT-2 h5 transformer model to ggml format
#
# Load the model using GPT2Model.
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
# - Number of dimensions (int)
# - Name length (int)
# - Dimensions (int[n_dims])
# - Name (char[name_length])
# - Data (float[n_dims])
#
# By default, the bigger matrices are converted to 16-bit floats.
# This can be disabled by adding the "use-f32" CLI argument.
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#
import sys
import struct
import json
import numpy as np
import re
from transformers import GPT2Model
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
cs = [chr(n) for n in cs]
return dict(zip(bs, cs))
if len(sys.argv) < 2:
print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
sys.exit(1)
# output in the same directory as the model
dir_model = sys.argv[1]
fname_out = sys.argv[1] + "/ggml-model.bin"
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
encoder = json.load(f)
with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
encoder_added = json.load(f)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
# use 16-bit or 32-bit floats
use_f16 = True
if len(sys.argv) > 2:
use_f16 = False
fname_out = sys.argv[1] + "/ggml-model-f32.bin"
model = GPT2Model.from_pretrained(dir_model, low_cpu_mem_usage=True)
#print (model)
list_vars = model.state_dict()
#print (list_vars)
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["vocab_size"]))
fout.write(struct.pack("i", hparams["n_positions"]))
fout.write(struct.pack("i", hparams["n_embd"]))
fout.write(struct.pack("i", hparams["n_head"]))
fout.write(struct.pack("i", hparams["n_layer"]))
#fout.write(struct.pack("i", hparams["rotary_dim"]))
fout.write(struct.pack("i", use_f16))
byte_encoder = bytes_to_unicode()
byte_decoder = {v:k for k, v in byte_encoder.items()}
fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
for key in encoder:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for key in encoder_added:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for name in list_vars.keys():
data = list_vars[name].squeeze().numpy()
print("Processing variable: " + name + " with shape: ", data.shape)
# we don't need these
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
print(" Skipping variable: " + name)
continue
n_dims = len(data.shape);
# ftype == 0 -> float32, ftype == 1 -> float16
ftype = 0;
if use_f16:
if name[-7:] == ".weight" and n_dims == 2:
print(" Converting to float16")
data = data.astype(np.float16)
ftype = 1
else:
print(" Converting to float32")
data = data.astype(np.float32)
ftype = 0
# for efficiency - transpose these matrices:
# "transformer.h.*.mlp.c_proj.weight
if name.endswith(".mlp.c_proj.weight"):
print(" Transposing")
data = data.transpose()
# rename headers to keep compatibility
if name == "ln_f.weight":
name = "model/ln_f/g"
elif name == "ln_f.bias":
name = "model/ln_f/b"
elif name == "wte.weight":
name = "model/wte"
elif name == "wpe.weight":
name = "model/wpe"
elif re.match(r"h\.\d+\.ln_1\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_1/g"
elif re.match(r"h\.\d+\.ln_1\.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_1/b"
elif re.match(r"h\.\d+\.attn\.c_attn\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_attn/w"
elif re.match(r"h\.\d+\.attn\.c_attn\.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_attn/b"
elif re.match(r"h\.\d+\.attn\.c_proj\.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_proj/w"
elif re.match(r"h.\d+.attn.c_proj.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/attn/c_proj/b"
elif re.match(r"h.\d+.ln_2.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_2/g"
elif re.match(r"h.\d+.ln_2.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/ln_2/b"
elif re.match(r"h.\d+.mlp.c_fc.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_fc/w"
elif re.match(r"h.\d+.mlp.c_fc.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_fc/b"
elif re.match(r"h.\d+.mlp.c_proj.weight", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_proj/w"
elif re.match(r"h.\d+.mlp.c_proj.bias", name):
i = re.findall("\d+", name)[0]
name = f"model/h{i}/mlp/c_proj/b"
else:
print("Unrecognized variable name. %s", name)
str = name.encode('utf-8')
fout.write(struct.pack("iii", n_dims, len(str), ftype))
for i in range(n_dims):
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
fout.write(str);
# data
data.tofile(fout)
fout.close()
print("Done. Output file: " + fname_out)
print("")
engine/ggml/examples/gpt-2/download-ggml-model.sh
Executable
+69
View File
@@ -0,0 +1,69 @@
#!/bin/bash
# This script downloads GPT-2 model files that have already been converted to ggml format.
# This way you don't have to convert them yourself.
#
# If you want to download the original GPT-2 model files, use the "download-model.sh" script instead.
#src="https://ggml.ggerganov.com"
#pfx="ggml-model-gpt-2"
src="https://huggingface.co/ggerganov/ggml"
pfx="resolve/main/ggml-model-gpt-2"
ggml_path=$(dirname $(realpath $0))
# GPT-2 models
models=( "117M" "345M" "774M" "1558M" )
# list available models
function list_models {
printf "\n"
printf " Available models:"
for model in "${models[@]}"; do
printf " $model"
done
printf "\n\n"
}
if [ "$#" -ne 1 ]; then
printf "Usage: $0 <model>\n"
list_models
exit 1
fi
model=$1
if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
printf "Invalid model: $model\n"
list_models
exit 1
fi
# download ggml model
printf "Downloading ggml model $model ...\n"
mkdir -p models/gpt-2-$model
if [ -x "$(command -v wget)" ]; then
wget --quiet --show-progress -O models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
elif [ -x "$(command -v curl)" ]; then
curl -L --output models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
else
printf "Either wget or curl is required to download models.\n"
exit 1
fi
if [ $? -ne 0 ]; then
printf "Failed to download ggml model $model \n"
printf "Please try again later or download the original GPT-2 model files and convert them yourself.\n"
exit 1
fi
printf "Done! Model '$model' saved in 'models/gpt-2-$model/ggml-model.bin'\n"
printf "You can now use it like this:\n\n"
printf " $ ./bin/gpt-2 -m models/gpt-2-$model/ggml-model.bin -p \"This is an example\"\n"
printf "\n"
engine/ggml/examples/gpt-2/download-model.sh
Executable
+48
View File
@@ -0,0 +1,48 @@
#!/bin/bash
ggml_path=$(dirname $(realpath $0))
# GPT-2 models
models=( "117M" "345M" "774M" "1558M" )
# list available models
function list_models {
printf "\n"
printf " Available models:"
for model in "${models[@]}"; do
printf " $model"
done
printf "\n\n"
}
if [ "$#" -ne 1 ]; then
printf "Usage: $0 <model>\n"
list_models
exit 1
fi
model=$1
if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
printf "Invalid model: $model\n"
list_models
exit 1
fi
# download model
printf "Downloading model $model ...\n"
mkdir -p models/gpt-2-$model
for file in checkpoint encoder.json hparams.json model.ckpt.data-00000-of-00001 model.ckpt.index model.ckpt.meta vocab.bpe; do
wget --quiet --show-progress -O models/gpt-2-$model/$file https://openaipublic.blob.core.windows.net/gpt-2/models/$model/$file
done
printf "Done! Model '$model' saved in 'models/gpt-2-$model/'\n\n"
printf "Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.\n"
printf "\n"
printf " python $ggml_path/convert-ckpt-to-ggml.py models/gpt-2-$model/\n"
printf "\n"
engine/ggml/examples/gpt-2/main-alloc.cpp
+880
View File
@@ -0,0 +1,880 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
int32_t n_ctx = 1024;
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t ftype = 1;
float eps = 1e-5f;
};
struct gpt2_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 gpt2_model {
gpt2_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // token embedding
struct ggml_tensor * wpe; // position embedding
struct ggml_tensor * lm_head; // language model head
std::vector<gpt2_layer> layers;
// key + value memory
struct ggml_tensor * memory_k;
struct ggml_tensor * memory_v;
//
struct ggml_context * ctx_w;
std::map<std::string, struct ggml_tensor *> tensors;
};
// load the model's weights from a file
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
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;
}
}
// 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_w;
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 += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // wte
ctx_size += ggml_row_size(GGML_TYPE_F32 , n_ctx*n_embd); // wpe
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // lm_head
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
ctx_size += n_layer*(ggml_row_size(wtype, 3*n_embd*n_embd)); // c_attn_attn_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd)); // c_attn_attn_b
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_attn_proj_b
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_proj_b
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
ctx_size += (6 + 12*n_layer)*512; // object overhead
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
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_w = ggml_init(params);
if (!model.ctx_w) {
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;
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, 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.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["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) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
return false;
}
auto tensor = model.tensors[name];
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
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 [%d, %d], expected [%d, %d]\n",
__func__, name.c_str(), (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.c_str(), 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.c_str(), 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();
return true;
}
// build the computation graph
struct ggml_cgraph * gpt2_graph(
const gpt2_model & model,
const int n_past,
const int n_tokens) {
const int N = n_tokens;
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;
// since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_cgraph * gf = ggml_new_graph(ctx);
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
// at this point, the tensor data is not allocated yet and cannot be set
// we will find the tensor after the graph is allocated by its name, and set the data then
ggml_set_name(embd, "embd");
// setting a tensor as an input will ensure that it is allocated at the beginning of the graph
// this is important to ensure that the input tensors are not overwritten before they are used
ggml_set_input(embd);
struct ggml_tensor * position = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
ggml_set_name(position, "position");
ggml_set_input(position);
// wte + wpe
struct ggml_tensor * inpL =
ggml_add(ctx,
ggml_get_rows(ctx, model.wte, embd),
ggml_get_rows(ctx, model.wpe, position));
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
// [ 768, N]
cur = ggml_norm(ctx, inpL, hparams.eps);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
cur = ggml_add(ctx,
ggml_mul(ctx,
ggml_repeat(ctx, model.layers[il].ln_1_g, cur),
cur),
ggml_repeat(ctx, 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(ctx,
model.layers[il].c_attn_attn_w,
cur);
cur = ggml_add(ctx,
ggml_repeat(ctx, model.layers[il].c_attn_attn_b, cur),
cur);
}
// self-attention
{
struct ggml_tensor * Qcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_tensor * Kcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_tensor * Vcur = ggml_view_2d(ctx, 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(ctx, 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(ctx, 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(ctx, Kcur, k));
ggml_build_forward_expand(gf, ggml_cpy(ctx, 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(ctx,
ggml_cont_3d(ctx, Qcur, 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(ctx,
ggml_reshape_3d(ctx,
ggml_view_1d(ctx, 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);
// 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(ctx, K, Q);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx,
KQ,
1.0f/sqrtf(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(ctx, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx, 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_cont_3d(ctx,
ggml_permute(ctx,
ggml_reshape_3d(ctx,
ggml_view_1d(ctx, 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),
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(ctx, V_trans, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
// [64, 12, N]
struct ggml_tensor * KQV_merged = ggml_permute(ctx, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
// [768, N]
cur = ggml_cont_2d(ctx, KQV_merged, 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(ctx,
model.layers[il].c_attn_proj_w,
cur);
cur = ggml_add(ctx,
ggml_repeat(ctx, model.layers[il].c_attn_proj_b, cur),
cur);
}
// add the input
cur = ggml_add(ctx, cur, inpL);
struct ggml_tensor * inpFF = cur;
// feed-forward network
{
// norm
{
cur = ggml_norm(ctx, inpFF, hparams.eps);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
cur = ggml_add(ctx,
ggml_mul(ctx,
ggml_repeat(ctx, model.layers[il].ln_2_g, cur),
cur),
ggml_repeat(ctx, 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(ctx,
model.layers[il].c_mlp_fc_w,
cur);
cur = ggml_add(ctx,
ggml_repeat(ctx, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
// [3072, N]
cur = ggml_gelu(ctx, 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(ctx,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx,
ggml_repeat(ctx, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// input for next layer
inpL = ggml_add(ctx, cur, inpFF);
}
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx, inpL, hparams.eps);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
inpL = ggml_add(ctx,
ggml_mul(ctx,
ggml_repeat(ctx, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx, model.ln_f_b, inpL));
}
// inpL = WTE * inpL
// [ 768, 50257] - model.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx, model.lm_head, inpL);
ggml_set_name(inpL, "logits");
// setting a tensor as the output will ensure that it is not overwritten by subsequent operations
ggml_set_output(inpL);
// logits -> probs
//inpL = ggml_soft_max(ctx0, inpL);
ggml_build_forward_expand(gf, inpL);
ggml_free(ctx);
return gf;
}
// evaluate the transformer
//
// - model: the model
// - allocr: ggml_gallocr to use to allocate the compute buffer
// - 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 gpt2_eval(
const gpt2_model & model,
ggml_gallocr_t allocr,
const int n_threads,
const int n_past,
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
const int n_vocab = hparams.n_vocab;
struct ggml_cgraph * gf = gpt2_graph(model, n_past, embd_inp.size());
// allocate the graph tensors
ggml_gallocr_alloc_graph(allocr, gf);
// set the graph inputs
struct ggml_tensor * embd = ggml_graph_get_tensor(gf, "embd");
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
struct ggml_tensor * position = ggml_graph_get_tensor(gf, "position");
for (int i = 0; i < N; ++i) {
((int32_t *) position->data)[i] = n_past + i;
}
// run the computation
struct ggml_cplan plan = ggml_graph_plan(gf, n_threads, nullptr);
static std::vector<uint8_t> work_buffer;
work_buffer.resize(plan.work_size);
plan.work_data = work_buffer.data();
ggml_graph_compute(gf, &plan);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
//}
// get the graph outputs
struct ggml_tensor * logits = ggml_graph_get_tensor(gf, "logits");
//embd_w.resize(n_vocab*N);
//memcpy(embd_w.data(), ggml_get_data(logits), 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(logits) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
return true;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-2-117M/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
params.prompt = gpt_random_prompt(rng);
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gpt2_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
test_gpt_tokenizer(vocab, params.token_test);
}
ggml_gallocr_t allocr = NULL;
// allocate the compute buffer
{
allocr = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
// create the worst case graph for memory usage estimation
int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
int n_past = model.hparams.n_ctx - n_tokens;
struct ggml_cgraph * gf = gpt2_graph(model, n_past, n_tokens);
// pre-allocate the compute buffer for the worst case (optional)
ggml_gallocr_reserve(allocr, gf);
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
printf("%d ", embd_inp[i]);
}
printf("\n\n");
// submit the input prompt token-by-token
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
std::vector<gpt_vocab::id> embd;
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gpt2_eval(model, allocr, params.n_threads, n_past, embd, logits)) {
printf("Failed to predict\n");
return 1;
}
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 = params.top_k;
const float top_p = params.top_p;
const float temp = params.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);
} else {
// if here, it means we are still processing the input prompt
for (size_t k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (int32_t(embd.size()) >= params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx_w);
return 0;
}
engine/ggml/examples/gpt-2/main-backend.cpp
+946
View File
@@ -0,0 +1,946 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define GPT2_MAX_NODES 4096
static void ggml_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);
fflush(stderr);
}
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
int32_t n_ctx = 1024;
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t ftype = 1;
float eps = 1e-5f;
};
struct gpt2_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 gpt2_model {
gpt2_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // token embedding
struct ggml_tensor * wpe; // position embedding
struct ggml_tensor * lm_head; // language model head
std::vector<gpt2_layer> layers;
// key + value memory
struct ggml_tensor * memory_k;
struct ggml_tensor * memory_v;
//
struct ggml_context * ctx_w;
struct ggml_context * ctx_kv;
ggml_backend_t backend = NULL;
ggml_backend_buffer_t buffer_w;
ggml_backend_buffer_t buffer_kv;
std::map<std::string, struct ggml_tensor *> tensors;
};
// load the model's weights from a file
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, int n_ctx, int n_gpu_layers) {
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;
}
}
// 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;
}
ggml_log_set(ggml_log_callback_default, nullptr);
auto & ctx = model.ctx_w;
// create the ggml context
{
size_t n_tensors = 2 + 6 + 12*model.hparams.n_layer;
struct ggml_init_params params = {
/*.mem_size =*/ ggml_tensor_overhead() * n_tensors,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx = ggml_init(params);
if (!ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false;
}
}
// initialize the backend
#ifdef GGML_USE_CUDA
if (n_gpu_layers > 0) {
fprintf(stderr, "%s: using CUDA backend\n", __func__);
model.backend = ggml_backend_cuda_init(0);
if (!model.backend) {
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
}
}
#endif
#ifdef GGML_USE_METAL
if (n_gpu_layers > 0) {
fprintf(stderr, "%s: using Metal backend\n", __func__);
model.backend = ggml_backend_metal_init();
if (!model.backend) {
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
}
}
#endif
if (!model.backend) {
// fallback to CPU backend
fprintf(stderr, "%s: using CPU backend\n", __func__);
model.backend = ggml_backend_cpu_init();
}
if (!model.backend) {
fprintf(stderr, "%s: ggml_backend_cpu_init() failed\n", __func__);
return false;
}
// create the tensors for the model
{
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;
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, 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.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["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;
}
}
// allocate the model tensors in a backend buffer
model.buffer_w = ggml_backend_alloc_ctx_tensors(ctx, model.backend);
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
printf("%s: backend buffer size = %6.2f MB\n", __func__, ggml_backend_buffer_get_size(model.buffer_w)/(1024.0*1024.0));
// override the default training context with the user-provided
model.hparams.n_ctx = n_ctx;
// key + value memory
{
auto * ctx = model.ctx_kv;
// create the ggml context
{
size_t n_tensors = 2;
struct ggml_init_params params = {
/*.mem_size =*/ ggml_tensor_overhead() * n_tensors,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx = ggml_init(params);
if (!ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false;
}
}
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;
// k and v here can also be GGML_TYPE_F16 to save memory and speed up the computation
// if backend supports it
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);
// allocate the KV memory in a backend buffer
model.buffer_kv = ggml_backend_alloc_ctx_tensors(ctx, model.backend);
const size_t memory_size = ggml_backend_buffer_get_size(model.buffer_kv);
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;
std::vector<char> read_buf;
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) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
return false;
}
auto tensor = model.tensors[name];
ggml_set_name(tensor, name.c_str());
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
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 [%d, %d], expected [%d, %d]\n",
__func__, name.c_str(), (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.c_str(), 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.c_str(), ggml_nbytes(tensor), nelements*bpe);
return false;
}
if (ggml_backend_buffer_is_host(model.buffer_w)) {
// for some backends such as CPU and Metal, the tensor data is in system memory and we can read directly into it
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
} else {
// read into a temporary buffer first, then copy to device memory
read_buf.resize(ggml_nbytes(tensor));
fin.read(read_buf.data(), ggml_nbytes(tensor));
ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor));
}
// GPT-2 models share the WTE tensor as the LM head
if (name == "model/wte" && has_lm_head == false) {
//ggml_backend_tensor_copy(tensor, model.lm_head);
model.lm_head = 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();
return true;
}
// build the computation graph
struct ggml_cgraph * gpt2_graph(
const gpt2_model & model,
const int n_past,
const int n_tokens) {
const int N = n_tokens;
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;
// since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
static size_t buf_size = ggml_tensor_overhead()*GPT2_MAX_NODES + ggml_graph_overhead_custom(GPT2_MAX_NODES, false);
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx, GPT2_MAX_NODES, false);
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
// at this point, the tensor data is not allocated yet and cannot be set
// we will find the tensor after the graph is allocated by its name, and set the data then
ggml_set_name(embd, "embd");
// setting a tensor as an input will ensure that it is allocated at the beginning of the graph
// this is important to ensure that the input tensors are not overwritten before they are used
ggml_set_input(embd);
struct ggml_tensor * position = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
ggml_set_name(position, "position");
ggml_set_input(position);
// wte + wpe
struct ggml_tensor * inpL =
ggml_add(ctx,
ggml_get_rows(ctx, model.wte, embd),
ggml_get_rows(ctx, model.wpe, position));
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
// [ 768, N]
cur = ggml_norm(ctx, inpL, hparams.eps);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
cur = ggml_add(ctx,
ggml_mul(ctx,
cur,
model.layers[il].ln_1_g),
model.layers[il].ln_1_b);
}
// 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(ctx,
model.layers[il].c_attn_attn_w,
cur);
cur = ggml_add(ctx,
cur,
model.layers[il].c_attn_attn_b);
}
// self-attention
{
struct ggml_tensor * Qcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_tensor * Kcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_tensor * Vcur = ggml_view_2d(ctx, 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(ctx, 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(ctx, 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(ctx, Kcur, k));
ggml_build_forward_expand(gf, ggml_cpy(ctx, 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(ctx,
ggml_cont_3d(ctx, Qcur, 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(ctx,
ggml_reshape_3d(ctx,
ggml_view_1d(ctx, 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);
// 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(ctx, K, Q);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx,
KQ,
1.0f/sqrtf(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(ctx, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx, 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_cont_3d(ctx,
ggml_permute(ctx,
ggml_reshape_3d(ctx,
ggml_view_1d(ctx, 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),
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(ctx, V_trans, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
// [64, 12, N]
struct ggml_tensor * KQV_merged = ggml_permute(ctx, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
// [768, N]
cur = ggml_cont_2d(ctx, KQV_merged, 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(ctx,
model.layers[il].c_attn_proj_w,
cur);
cur = ggml_add(ctx,
cur,
model.layers[il].c_attn_proj_b);
}
// add the input
cur = ggml_add(ctx, cur, inpL);
struct ggml_tensor * inpFF = cur;
// feed-forward network
{
// norm
{
cur = ggml_norm(ctx, inpFF, hparams.eps);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
cur = ggml_add(ctx,
ggml_mul(ctx,
cur,
model.layers[il].ln_2_g),
model.layers[il].ln_2_b);
}
// 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(ctx,
model.layers[il].c_mlp_fc_w,
cur);
cur = ggml_add(ctx,
cur,
model.layers[il].c_mlp_fc_b);
// GELU activation
// [3072, N]
cur = ggml_gelu(ctx, 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(ctx,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx,
cur,
model.layers[il].c_mlp_proj_b);
}
// input for next layer
inpL = ggml_add(ctx, cur, inpFF);
}
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx, inpL, hparams.eps);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
inpL = ggml_add(ctx,
ggml_mul(ctx,
inpL,
model.ln_f_g),
model.ln_f_b);
}
// inpL = WTE * inpL
// [ 768, 50257] - model.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx, model.lm_head, inpL);
ggml_set_name(inpL, "logits");
// setting a tensor as the output will ensure that it is not overwritten by subsequent operations
ggml_set_output(inpL);
// logits -> probs
//inpL = ggml_soft_max(ctx0, inpL);
ggml_build_forward_expand(gf, inpL);
ggml_free(ctx);
return gf;
}
// evaluate the transformer
//
// - model: the model
// - allocr: ggml_gallocr to use to allocate the compute buffer
// - 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 gpt2_eval(
const gpt2_model & model,
ggml_gallocr_t allocr,
const int n_threads,
const int n_past,
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
const int n_vocab = hparams.n_vocab;
struct ggml_cgraph * gf = gpt2_graph(model, n_past, embd_inp.size());
// allocate the graph tensors
ggml_gallocr_alloc_graph(allocr, gf);
// set the graph inputs
struct ggml_tensor * embd = ggml_graph_get_tensor(gf, "embd");
ggml_backend_tensor_set(embd, embd_inp.data(), 0, N*ggml_element_size(embd));
struct ggml_tensor * position = ggml_graph_get_tensor(gf, "position");
for (int i = 0; i < N; ++i) {
int32_t v = n_past + i;
ggml_backend_tensor_set(position, &v, i*sizeof(int32_t), sizeof(v));
}
// set backend options
if (ggml_backend_is_cpu(model.backend)) {
ggml_backend_cpu_set_n_threads(model.backend, n_threads);
}
// run the computation
ggml_backend_graph_compute(model.backend, gf);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
//}
// get the graph outputs
struct ggml_tensor * logits = ggml_graph_get_tensor(gf, "logits");
//embd_w.resize(n_vocab*N);
//ggml_backend_tensor_get(logits, embd_w.data(), 0, sizeof(float)*n_vocab*N);
// return result just for the last token
embd_w.resize(n_vocab);
ggml_backend_tensor_get(logits, embd_w.data(), (n_vocab*(N-1))*sizeof(float), sizeof(float)*n_vocab);
return true;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-2-117M/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
params.prompt = gpt_random_prompt(rng);
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gpt2_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(params.model, model, vocab, params.n_ctx, params.n_gpu_layers)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
test_gpt_tokenizer(vocab, params.token_test);
}
ggml_gallocr_t allocr = NULL;
// allocate the compute buffer
{
// create a graph allocator with the backend's default buffer type
allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
// create the worst case graph for memory usage estimation
int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
int n_past = model.hparams.n_ctx - n_tokens;
struct ggml_cgraph * gf = gpt2_graph(model, n_past, n_tokens);
// pre-allocate the compute buffer for the worst case (optional)
ggml_gallocr_reserve(allocr, gf);
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
printf("%d ", embd_inp[i]);
}
printf("\n\n");
// submit the input prompt token-by-token
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
std::vector<gpt_vocab::id> embd;
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gpt2_eval(model, allocr, params.n_threads, n_past, embd, logits)) {
printf("Failed to predict\n");
return 1;
}
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 = params.top_k;
const float top_p = params.top_p;
const float temp = params.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);
} else {
// if here, it means we are still processing the input prompt
for (size_t k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (int32_t(embd.size()) >= params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (!params.ignore_eos && embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx_w);
ggml_gallocr_free(allocr);
ggml_backend_buffer_free(model.buffer_w);
ggml_backend_buffer_free(model.buffer_kv);
ggml_backend_free(model.backend);
return 0;
}
engine/ggml/examples/gpt-2/main-batched.cpp
+1210
View File
File diff suppressed because it is too large Load Diff
engine/ggml/examples/gpt-2/main-ctx.cpp
+840
View File
@@ -0,0 +1,840 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
int32_t n_ctx = 1024;
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t ftype = 1;
float eps = 1e-5f;
};
struct gpt2_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 gpt2_model {
gpt2_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // token embedding
struct ggml_tensor * wpe; // position embedding
struct ggml_tensor * lm_head; // language model head
std::vector<gpt2_layer> layers;
// key + value memory
struct ggml_tensor * memory_k;
struct ggml_tensor * memory_v;
//
struct ggml_context * ctx_w;
std::map<std::string, struct ggml_tensor *> tensors;
};
// load the model's weights from a file
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
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;
}
}
// 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_w;
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 += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // wte
ctx_size += ggml_row_size(GGML_TYPE_F32, n_ctx*n_embd); // wpe
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // lm_head
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
ctx_size += n_layer*(ggml_row_size(wtype, 3*n_embd*n_embd)); // c_attn_attn_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd)); // c_attn_attn_b
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_attn_proj_b
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_proj_b
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
ctx_size += (6 + 12*n_layer)*512; // object overhead
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
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_w = ggml_init(params);
if (!model.ctx_w) {
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;
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, 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.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["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) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
return false;
}
auto tensor = model.tensors[name];
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
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 [%d, %d], expected [%d, %d]\n",
__func__, name.c_str(), (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.c_str(), 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.c_str(), 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();
return true;
}
// 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 gpt2_eval(
const gpt2_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 = 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%s: 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) {
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 = ggml_new_graph(ctx0);
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;
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL, hparams.eps);
// 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);
// 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);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, 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;
// feed-forward network
{
// norm
{
cur = ggml_norm(ctx0, inpFF, hparams.eps);
// 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);
}
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL, hparams.eps);
// 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));
}
// 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;
}
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
ggml_free(ctx0);
return true;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-2-117M/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
params.prompt = gpt_random_prompt(rng);
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gpt2_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
test_gpt_tokenizer(vocab, params.token_test);
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
printf("%d ", embd_inp[i]);
}
printf("\n\n");
// submit the input prompt token-by-token
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
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 = params.top_k;
const float top_p = params.top_p;
const float temp = params.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);
} else {
// if here, it means we are still processing the input prompt
for (size_t k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (int32_t(embd.size()) >= params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx_w);
return 0;
}
engine/ggml/examples/gpt-2/main-sched.cpp
+1079
View File
File diff suppressed because it is too large Load Diff
engine/ggml/examples/gpt-2/quantize.cpp
+184
View File
@@ -0,0 +1,184 @@
#include "ggml.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
int32_t n_ctx = 1024;
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t ftype = 1;
};
// quantize a model
bool gpt2_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
gpt_vocab vocab;
printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
auto finp = std::ifstream(fname_inp, std::ios::binary);
if (!finp) {
fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
return false;
}
auto fout = std::ofstream(fname_out, std::ios::binary);
if (!fout) {
fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
return false;
}
// verify magic
{
uint32_t magic;
finp.read((char *) &magic, sizeof(magic));
if (magic != GGML_FILE_MAGIC) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
return false;
}
fout.write((char *) &magic, sizeof(magic));
}
gpt2_hparams hparams;
// load hparams
{
finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
finp.read((char *) &hparams.ftype, sizeof(hparams.ftype));
const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
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 (src) = %d\n", __func__, hparams.ftype);
printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fout.write((char *) &ftype_dst, sizeof(ftype_dst));
}
// load vocab
{
int32_t n_vocab = 0;
finp.read ((char *) &n_vocab, sizeof(n_vocab));
fout.write((char *) &n_vocab, sizeof(n_vocab));
if (n_vocab != hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
return false;
}
std::string word;
for (int i = 0; i < n_vocab; i++) {
uint32_t len;
finp.read ((char *) &len, sizeof(len));
fout.write((char *) &len, sizeof(len));
word.resize(len);
finp.read ((char *) word.data(), len);
fout.write((char *) word.data(), len);
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
}
}
// regexes of tensor names to be quantized
const std::vector<std::string> to_quant = {
"model/wte",
"model/lm_head",
"model/h.*/attn/c_attn/w",
"model/h.*/attn/c_proj/w",
"model/h.*/mlp/c_fc/w",
"model/h.*/mlp/c_proj/w",
};
if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
return false;
}
finp.close();
fout.close();
return true;
}
// usage:
// ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
ggml_print_ftypes(stderr);
return 1;
}
// needed to initialize f16 tables
{
struct ggml_init_params params = { 0, NULL, false };
struct ggml_context * ctx = ggml_init(params);
ggml_free(ctx);
}
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
const int64_t t_main_start_us = ggml_time_us();
int64_t t_quantize_us = 0;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
return 1;
}
t_quantize_us = ggml_time_us() - t_start_us;
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n");
printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
return 0;
}
engine/ggml/examples/gpt-j/CMakeLists.txt
+13
View File
@@ -0,0 +1,13 @@
#
# gpt-j
set(TEST_TARGET gpt-j)
add_executable(${TEST_TARGET} main.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
#
# gpt-j-quantize
set(TEST_TARGET gpt-j-quantize)
add_executable(${TEST_TARGET} quantize.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
engine/ggml/examples/gpt-j/README.md
+239
View File
@@ -0,0 +1,239 @@
# gpt-j
Local GPT-J inference on your computer using C/C++
No video card required. You just need to have 16 GB of RAM.
## Motivation
The GPT-J 6B model is the open-source alternative to OpenAI's GPT-3. It's basically a neural network that allows you to
generate coherent, human-like text given a certain context (prompt).
The GPT-J model is quite big - the compact version of the model uses 16-bit floating point representation of the weights
and is still 12 GB big. This means that in order to run inference on your computer, you would need to have a video card
with at least 12 GB of video RAM. Alternatively, you can try to run the python implementations on the CPU, but that
would probably not be very efficient as they are primarily optimized for running on a GPU (or at least this is my guess -
I don't have much experience with python).
I wanted to try and run the model on my MacBook, so I decided to implement the model inference from scratch using my own
custom build tensor library. The tensor library (called [ggml](https://github.com/ggerganov/ggml), written in C) is in
early development stage, but it already allows me to run the GPT-J model.
On my 32GB MacBook M1 Pro, I achieve an inference speed of about `125 ms/token` or about ~6 words per second (1 word
typically consists of 1 or 2 tokens).
Here is a sample run with prompt `int main(int argc, char ** argv) {`:
```bash
$ time ./bin/gpt-j -p "int main(int argc, char ** argv) {"
gptj_model_load: loading model from 'models/gpt-j-6B/ggml-model.bin' - please wait ...
gptj_model_load: n_vocab = 50400
gptj_model_load: n_ctx = 2048
gptj_model_load: n_embd = 4096
gptj_model_load: n_head = 16
gptj_model_load: n_layer = 28
gptj_model_load: n_rot = 64
gptj_model_load: f16 = 1
gptj_model_load: ggml ctx size = 13334.86 MB
gptj_model_load: memory_size = 1792.00 MB, n_mem = 57344
gptj_model_load: ................................... done
gptj_model_load: model size = 11542.79 MB / num tensors = 285
main: number of tokens in prompt = 13
int main(int argc, char ** argv) {
(void)argc;
(void)argv;
{
struct sockaddr_in addr;
int addrlen;
char * ip = "192.168.1.4";
int i;
if ( (addrlen = sizeof(addr)) == -1 )
return -1;
for (i = 0; i < 10; ++i) {
addr.sin_family = AF_INET;
addr.sin_addr.s_addr = inet_addr(ip);
main: mem per token = 16430420 bytes
main: load time = 6211.48 ms
main: sample time = 13.74 ms
main: predict time = 26420.34 ms / 124.62 ms per token
main: total time = 33035.37 ms
real 0m33.171s
user 3m32.269s
sys 0m3.686s
$
```
It took ~6.2 seconds to load the model to memory. After that, it took ~26.4 seconds to generate 200 tokens of what
looks like to be the beginning of a networking program in C. Pretty cool!
Here is another run, just for fun:
```bash
time ./bin/gpt-j -n 500 -t 8 -p "Ask HN: Inherited the worst code and tech team I have ever seen. How to fix it?
"
gptj_model_load: loading model from 'models/gpt-j-6B/ggml-model.bin' - please wait ...
gptj_model_load: n_vocab = 50400
gptj_model_load: n_ctx = 2048
gptj_model_load: n_embd = 4096
gptj_model_load: n_head = 16
gptj_model_load: n_layer = 28
gptj_model_load: n_rot = 64
gptj_model_load: f16 = 1
gptj_model_load: ggml ctx size = 13334.86 MB
gptj_model_load: memory_size = 1792.00 MB, n_mem = 57344
gptj_model_load: ................................... done
gptj_model_load: model size = 11542.79 MB / num tensors = 285
main: number of tokens in prompt = 24
Ask HN: Inherited the worst code and tech team I have ever seen. How to fix it?
I've inherited a team with some very strange and un-documented practices, one of them is that they use an old custom
application with a very slow tech stack written in Python that the team doesn't want to touch but also doesn't want to
throw away as it has some "legacy" code in it.
The problem is, the tech stack is very very slow.
They have a single web server on a VM that is slow.
The server is a little bit busy (not very busy though) and they have a lot of processes (30+ that are constantly being
spawned by the application)
They have an application that is single threaded and was written in Python and the team don't want to touch this, and
the application is very slow.
My task as a new member of the team is to fix this.
I'm a senior dev on the team (3 years on the project) and have been told that I will take the lead on this task. I know
next to nothing about Python. So here is what I have so far.
What I have done is I've been trying to debug the processes with the "ps" command. This way I can see what is running
and where. From what I see, the application spawns 10 processes a minute and some of them are used for nothing.
I have also started to look for the code. The application source is not in GitHub or any other repository, it is only on
our internal GitLab.
What I've found so far:
The application uses a custom SQLAlchemy implementation to interact with the data. I've looked at the source, it looks
like an object cache or something like that. But from what I've seen, the cache gets full every 20 minutes and then gets
cleared with a special command.
Another strange thing is that the application creates a file for every entry in the database (even if the entry already
exists). I've looked at the file to see if it contains something, but it seems to be a JSON file with lots of records.
The other strange thing is that I can only find the database tables in the GitLab repository and not the code. So I
can't really understand how the application is supposed to interact with the database.
I also found a "log" directory, but the code is encrypted with AES. From what I've found, it is in
main: mem per token = 16430420 bytes
main: load time = 3900.10 ms
main: sample time = 32.58 ms
main: predict time = 68049.91 ms / 130.11 ms per token
main: total time = 73020.05 ms
real 1m13.156s
user 9m1.328s
sys. 0m7.103s
```
## Implementation details
The high level implementation of the model is contained in the [main.cpp](main.cpp) file. The core computations are
performed by the [ggml](https://github.com/ggerganov/ggml/blob/master/include/ggml.h) library.
#### Matrix multiplication
The most performance critical part of the implementation is of course the matrix multiplication routine. 99% of the time
is spent here, so it was important to optimize this as much as possible.
On Arm64, I utilize the 128-bit NEON intrinsics for 16-bit floating point operations:
https://github.com/ggerganov/ggml/blob/fb558f78d905f85c54813602649ddd628ffe0f3a/src/ggml.c#L187-L243
These instructions allow each core to operate simultaneously on 64 16-bit floats. I'm no expert in SIMD, but after quite
some trials this was the most efficient code for dot product of a row and column that I could come up with. Combined
with the parallel computation on 8 CPU threads, I believe I'm close to the maximum performance that one could possibly
get on the M1 CPU. Still, I'm curious to know if there is a more efficient way to implement this.
#### Attempt to use the M1 GPU
One interesting property of the GPT-J transformer architecture is that it allows you to perform part of the inference in
parallel - i.e. the Feed-forward network can be computed in parallel to the Self-attention layer:
https://github.com/ggerganov/ggml/blob/fb558f78d905f85c54813602649ddd628ffe0f3a/examples/gpt-j/main.cpp#L507-L531
So I thought why not try and bring in the M1 GPU to compute half of the neural network in parallel to the CPU and
potentially gain some extra performance. Thanks to the M1's shared memory model, it was relatively easy to offload part
of the computation to the GPU using Apple's [Metal Performance
Shaders](https://developer.apple.com/documentation/metalperformanceshaders). The GPU shares the host memory, so there is
no need to copy the data back and forth as you would normally do with Cuda or OpenCL. The weight matrices are directly
available to be used by the GPU.
However, to my surprise, using MPS together with the CPU did not lead to any performance improvement at all. My
conclusion was that the 8-thread NEON CPU computation is already saturating the memory bandwidth of the M1 and since
the CPU and the GPU on the MacBook are sharing that bandwidth, it does not help to offload the computation to the GPU.
Another observation was that the MPS GPU matrix multiplication using 16-bit floats had the same performance as the
8-thread NEON CPU implementation. Again, I explain this with a saturated memory channel. But of course, my explanation
could be totally wrong and somehow the implementation wasn't utilizing the resources correctly.
In the end, I decided to not use MPS or the GPU all together.
### Zero memory allocations
Another property of my implementation is that it does not perform any memory allocations once the model is loaded into
memory. All required memory is allocated at the start of the program with a single `malloc` (technically 2 calls, but
that is not important).
## Usage
If you want to give this a try and you are on Linux or Mac OS, simply follow these instructions:
```bash
# Download the ggml-compatible GPT-J 6B model (requires 12GB disk space)
../examples/gpt-j/download-ggml-model.sh 6B
# Run the inference (requires 16GB of CPU RAM)
./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin -p "This is an example"
# Input prompt through pipe and run the inference.
echo "This is an example" > prompt.txt
cat prompt.txt | ./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin
```
To run the `gpt-j` tool, you need the 12GB `ggml-model.bin` file which contains the GPT-J model in
[ggml](https://github.com/ggerganov/ggml) compatible format. In the instructions above, the binary file
is downloaded from my repository on Hugging Face using the [download-ggml-model.sh](download-ggml-model.sh) script.
You can also, download the file manually from this link:
https://huggingface.co/ggerganov/ggml/tree/main
---
Alternatively, if you don't want to download the 12GB ggml model file, you can perform the conversion yourself using
python.
First, you need to download the full GPT-J model from here: https://huggingface.co/EleutherAI/gpt-j-6B
Note that the full model is quite big - about 72 GB. After you download it, you need to convert it to ggml format using
the [convert-h5-to-ggml.py](convert-h5-to-ggml.py) script. This will generate the `ggml-model.bin` file, which you can
then use with the `gpt-j` program.
## GPT-2
I also implemented a tool for CPU inference using the smaller GPT-2 models. They have worse quality compared to GPT-J,
but are much faster to execute.
For example, the Small GPT-2 model is only 240 MB big and the inference speed on my MacBook is about 200 tokens/sec.
For more details, checkout the GPT-2 example here: [gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)
engine/ggml/examples/gpt-j/convert-h5-to-ggml.py
+173
View File
@@ -0,0 +1,173 @@
# Convert GPT-J-6B h5 transformer model to ggml format
#
# Load the model using GPTJForCausalLM.
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
# - Number of dimensions (int)
# - Name length (int)
# - Dimensions (int[n_dims])
# - Name (char[name_length])
# - Data (float[n_dims])
#
# By default, the bigger matrices are converted to 16-bit floats.
# This can be disabled by adding the "use-f32" CLI argument.
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#
import sys
import struct
import json
import torch
import numpy as np
from transformers import GPTJForCausalLM
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
cs = [chr(n) for n in cs]
return dict(zip(bs, cs))
if len(sys.argv) < 3:
print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
print(" ftype == 0 -> float32")
print(" ftype == 1 -> float16")
sys.exit(1)
# output in the same directory as the model
dir_model = sys.argv[1]
fname_out = sys.argv[1] + "/ggml-model.bin"
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
encoder = json.load(f)
with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
encoder_added = json.load(f)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
# possible data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if len(sys.argv) > 2:
ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1:
print("Invalid ftype: " + str(ftype))
sys.exit(1)
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
model = GPTJForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
#print (model)
list_vars = model.state_dict()
#print (list_vars)
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["vocab_size"]))
fout.write(struct.pack("i", hparams["n_positions"]))
fout.write(struct.pack("i", hparams["n_embd"]))
fout.write(struct.pack("i", hparams["n_head"]))
fout.write(struct.pack("i", hparams["n_layer"]))
fout.write(struct.pack("i", hparams["rotary_dim"]))
fout.write(struct.pack("i", ftype))
byte_encoder = bytes_to_unicode()
byte_decoder = {v:k for k, v in byte_encoder.items()}
fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
for key in encoder:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for key in encoder_added:
text = bytearray([byte_decoder[c] for c in key])
fout.write(struct.pack("i", len(text)))
fout.write(text)
for name in list_vars.keys():
data = list_vars[name].squeeze().numpy()
print("Processing variable: " + name + " with shape: ", data.shape)
# we don't need these
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
print(" Skipping variable: " + name)
continue
n_dims = len(data.shape);
# ftype == 0 -> float32, ftype == 1 -> float16
ftype_cur = 0;
if ftype != 0:
if name[-7:] == ".weight" and n_dims == 2:
print(" Converting to float16")
data = data.astype(np.float16)
ftype_cur = 1
else:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
else:
if data.dtype != np.float32:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
# for efficiency - transpose these matrices:
# (note - with latest ggml this is no longer more efficient, so disabling it)
# "transformer.h.*.mlp.fc_in.weight"
# "transformer.h.*.attn.out_proj.weight"
# "transformer.h.*.attn.q_proj.weight"
# "transformer.h.*.attn.k_proj.weight"
# "transformer.h.*.attn.v_proj.weight"
#if name.endswith(".mlp.fc_in.weight") or \
# name.endswith(".attn.out_proj.weight") or \
# name.endswith(".attn.q_proj.weight") or \
# name.endswith(".attn.k_proj.weight") or \
# name.endswith(".attn.v_proj.weight"):
# print(" Transposing")
# data = data.transpose()
# header
str = name.encode('utf-8')
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);
# data
data.tofile(fout)
fout.close()
print("Done. Output file: " + fname_out)
print("")
engine/ggml/examples/gpt-j/download-ggml-model.sh
Executable
+69
View File
@@ -0,0 +1,69 @@
#!/bin/bash
# This script downloads GPT-J model files that have already been converted to ggml format.
# This way you don't have to convert them yourself.
#
# If you want to download the original GPT-J model files, use the "download-model.sh" script instead.
#src="https://ggml.ggerganov.com"
#pfx="ggml-model-gpt-j"
src="https://huggingface.co/ggerganov/ggml"
pfx="resolve/main/ggml-model-gpt-j"
ggml_path=$(dirname $(realpath $0))
# GPT-J models
models=( "6B" )
# list available models
function list_models {
printf "\n"
printf " Available models:"
for model in "${models[@]}"; do
printf " $model"
done
printf "\n\n"
}
if [ "$#" -ne 1 ]; then
printf "Usage: $0 <model>\n"
list_models
exit 1
fi
model=$1
if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
printf "Invalid model: $model\n"
list_models
exit 1
fi
# download ggml model
printf "Downloading ggml model $model ...\n"
mkdir -p models/gpt-j-$model
if [ -x "$(command -v wget)" ]; then
wget --quiet --show-progress -O models/gpt-j-$model/ggml-model.bin $src/$pfx-$model.bin
elif [ -x "$(command -v curl)" ]; then
curl -L --output models/gpt-j-$model/ggml-model.bin $src/$pfx-$model.bin
else
printf "Either wget or curl is required to download models.\n"
exit 1
fi
if [ $? -ne 0 ]; then
printf "Failed to download ggml model $model \n"
printf "Please try again later or download the original GPT-J model files and convert them yourself.\n"
exit 1
fi
printf "Done! Model '$model' saved in 'models/gpt-j-$model/ggml-model.bin'\n"
printf "You can now use it like this:\n\n"
printf " $ ./bin/gpt-j -m models/gpt-j-$model/ggml-model.bin -p \"This is an example\"\n"
printf "\n"
engine/ggml/examples/gpt-j/download-model.sh
Executable
+11
View File
@@ -0,0 +1,11 @@
#!/bin/bash
printf "To obtain the GPT-J 6B model files, please visit: https://huggingface.co/EleutherAI/gpt-j-6B\n\n"
printf "The model is very big. For example, the reposirory above is 72GB in size.\n"
printf "If you are sure that you want to clone it, simply run the following command:\n\n"
printf " $ git clone https://huggingface.co/EleutherAI/gpt-j-6B models/gpt-j-6B\n\n"
printf "Alternatively, use the 'download-ggml-model.sh' script to download a 12GB ggml version of the model.\n"
printf "This version is enough to run inference using the ggml library.\n\n"
engine/ggml/examples/gpt-j/main.cpp
+755
View File
@@ -0,0 +1,755 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
// 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;
float eps = 1e-5f;
};
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; // token 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;
};
// load the model's weights from a file
bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
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.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: 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;
}
}
// 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;
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
ctx_size += ggml_row_size(wtype, n_embd*n_vocab); // wte
ctx_size += ggml_row_size(wtype, n_embd*n_vocab); // lmh_g
ctx_size += ggml_row_size(GGML_TYPE_F32, n_vocab); // lmh_b
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_q_proj_w
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_k_proj_w
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_v_proj_w
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_mlp_proj_b
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F16, n_embd); // memory_k
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F16, n_embd); // memory_v
ctx_size += (5 + 10*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_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);
printf("%s: memory_size = %8.2f MB, n_mem = %d\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) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
return false;
}
auto tensor = model.tensors[name];
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
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 [%d, %d], expected [%d, %d]\n",
__func__, name.c_str(), (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.c_str(), 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.c_str(), 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.c_str(), 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(" done\n");
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
}
fin.close();
return true;
}
// 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%s: 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) {
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 = ggml_new_graph(ctx0);
// KQ_pos - contains the positions
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
int * data = (int *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}
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, hparams.eps);
// 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), KQ_pos, n_rot, 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), KQ_pos, n_rot, 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,
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, hparams.eps);
// 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;
}
int main(int argc, char ** argv) {
ggml_time_init();
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-j-6B/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
params.prompt = gpt_random_prompt(rng);
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gptj_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gptj_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
test_gpt_tokenizer(vocab, params.token_test);
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> logits;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
printf("\n");
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
gptj_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
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 = params.top_k;
const float top_p = params.top_p;
const float temp = params.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);
} else {
// if here, it means we are still processing the input prompt
for (size_t k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (int32_t(embd.size()) > params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx);
return 0;
}
engine/ggml/examples/gpt-j/quantize.cpp
+182
View File
@@ -0,0 +1,182 @@
#include "ggml.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
// 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;
};
// quantize a model
bool gptj_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
gpt_vocab vocab;
printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
auto finp = std::ifstream(fname_inp, std::ios::binary);
if (!finp) {
fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
return false;
}
auto fout = std::ofstream(fname_out, std::ios::binary);
if (!fout) {
fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
return false;
}
// verify magic
{
uint32_t magic;
finp.read((char *) &magic, sizeof(magic));
if (magic != GGML_FILE_MAGIC) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
return false;
}
fout.write((char *) &magic, sizeof(magic));
}
gptj_hparams hparams;
// load hparams
{
finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
finp.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
finp.read((char *) &hparams.ftype, sizeof(hparams.ftype));
const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
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 (src) = %d\n", __func__, hparams.ftype);
printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fout.write((char *) &hparams.n_rot, sizeof(hparams.n_rot));
fout.write((char *) &ftype_dst, sizeof(ftype_dst));
}
// load vocab
{
int32_t n_vocab = 0;
finp.read ((char *) &n_vocab, sizeof(n_vocab));
fout.write((char *) &n_vocab, sizeof(n_vocab));
if (n_vocab != hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
return false;
}
std::string word;
for (int i = 0; i < n_vocab; i++) {
uint32_t len;
finp.read ((char *) &len, sizeof(len));
fout.write((char *) &len, sizeof(len));
word.resize(len);
finp.read ((char *) word.data(), len);
fout.write((char *) word.data(), len);
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
}
}
// regexes of tensor names to be quantized
const std::vector<std::string> to_quant = {
".*weight",
};
if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
return false;
}
finp.close();
fout.close();
return true;
}
// usage:
// ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
//
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
ggml_print_ftypes(stderr);
return 1;
}
// needed to initialize f16 tables
{
struct ggml_init_params params = { 0, NULL, false };
struct ggml_context * ctx = ggml_init(params);
ggml_free(ctx);
}
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
const int64_t t_main_start_us = ggml_time_us();
int64_t t_quantize_us = 0;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gptj_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
return 1;
}
t_quantize_us = ggml_time_us() - t_start_us;
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n");
printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
return 0;
}
engine/ggml/examples/magika/CMakeLists.txt
+17
View File
@@ -0,0 +1,17 @@
#
# magika
set(TEST_TARGET magika)
add_executable(${TEST_TARGET} main.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
#
# For GPU offloading
if (GGML_CUDA)
add_compile_definitions(GGML_USE_CUDA)
endif()
if (GGML_METAL)
add_compile_definitions(GGML_USE_METAL)
endif()
engine/ggml/examples/magika/README.md
+23
View File
@@ -0,0 +1,23 @@
# Google Magika inference
Simple example that shows how to use GGML for inference with the [Google Magika](https://github.com/google/magika) file type detection model.
### Usage
- Obtain the Magika model in H5 format
- Pinned version: https://github.com/google/magika/blob/4460acb5d3f86807c3b53223229dee2afa50c025/assets_generation/models/standard_v1/model.h5
- Use `convert.py` to convert the model to gguf format:
```bash
$ python examples/magika/convert.py /path/to/model.h5
```
- Invoke the program with the model file and a list of files to identify:
```bash
$ build/bin/magika model.h5.gguf examples/sam/example.jpg examples/magika/convert.py README.md src/ggml.c /bin/gcc write.exe jfk.wav
examples/sam/example.jpg : jpeg (100.00%) pptx (0.00%) smali (0.00%) shell (0.00%) sevenzip (0.00%)
examples/magika/convert.py : python (99.99%) javascript (0.00%) txt (0.00%) asm (0.00%) scala (0.00%)
README.md : markdown (100.00%) txt (0.00%) yaml (0.00%) ppt (0.00%) shell (0.00%)
src/ggml.c : c (99.95%) txt (0.04%) asm (0.01%) yaml (0.00%) html (0.00%)
/bin/gcc : elf (99.98%) odex (0.02%) pptx (0.00%) smali (0.00%) shell (0.00%)
write.exe : pebin (100.00%) ppt (0.00%) smali (0.00%) shell (0.00%) sevenzip (0.00%)
jfk.wav : wav (100.00%) ppt (0.00%) shell (0.00%) sevenzip (0.00%) scala (0.00%)
```
engine/ggml/examples/magika/convert.py
+32
View File
@@ -0,0 +1,32 @@
import sys
from tensorflow import keras
import gguf
def convert(model_name):
model = keras.models.load_model(model_name, compile=False)
gguf_model_name = model_name + ".gguf"
gguf_writer = gguf.GGUFWriter(gguf_model_name, "magika")
for layer in model.layers:
# export layers with weights
if layer.weights:
for weight in layer.weights:
print(f" [{weight.name}] {weight.shape} {weight.dtype}")
weight_data = weight.numpy()
gguf_writer.add_tensor(weight.name, weight_data.T)
gguf_writer.write_header_to_file()
gguf_writer.write_kv_data_to_file()
gguf_writer.write_tensors_to_file()
gguf_writer.close()
print("Model converted and saved to '{}'".format(gguf_model_name))
if __name__ == '__main__':
if len(sys.argv) > 1:
model_file = sys.argv[1]
else:
model_file = "model.h5"
convert(model_file)
engine/ggml/examples/magika/main.cpp
+374
View File
@@ -0,0 +1,374 @@
#include "ggml.h"
#include "gguf.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include <algorithm>
#include <cmath>
#include <numeric>
#include <stdexcept>
#include <string>
#include <vector>
static const char * magika_labels[] = {
"ai", "apk", "appleplist", "asm", "asp",
"batch", "bmp", "bzip", "c", "cab",
"cat", "chm", "coff", "crx", "cs",
"css", "csv", "deb", "dex", "dmg",
"doc", "docx", "elf", "emf", "eml",
"epub", "flac", "gif", "go", "gzip",
"hlp", "html", "ico", "ini", "internetshortcut",
"iso", "jar", "java", "javabytecode", "javascript",
"jpeg", "json", "latex", "lisp", "lnk",
"m3u", "macho", "makefile", "markdown", "mht",
"mp3", "mp4", "mscompress", "msi", "mum",
"odex", "odp", "ods", "odt", "ogg",
"outlook", "pcap", "pdf", "pebin", "pem",
"perl", "php", "png", "postscript", "powershell",
"ppt", "pptx", "python", "pythonbytecode", "rar",
"rdf", "rpm", "rst", "rtf", "ruby",
"rust", "scala", "sevenzip", "shell", "smali",
"sql", "squashfs", "svg", "swf", "symlinktext",
"tar", "tga", "tiff", "torrent", "ttf",
"txt", "unknown", "vba", "wav", "webm",
"webp", "winregistry", "wmf", "xar", "xls",
"xlsb", "xlsx", "xml", "xpi", "xz",
"yaml", "zip", "zlibstream"
};
struct magika_hparams {
const int block_size = 4096;
const int beg_size = 512;
const int mid_size = 512;
const int end_size = 512;
const int min_file_size_for_dl = 16;
const int n_label = 113;
const float f_norm_eps = 0.001f;
const int padding_token = 256;
};
struct magika_model {
~magika_model() {
ggml_backend_buffer_free(buf_w);
ggml_backend_free(backend);
ggml_free(ctx_w);
}
magika_hparams hparams;
struct ggml_tensor * dense_w;
struct ggml_tensor * dense_b;
struct ggml_tensor * layer_norm_gamma;
struct ggml_tensor * layer_norm_beta;
struct ggml_tensor * dense_1_w;
struct ggml_tensor * dense_1_b;
struct ggml_tensor * dense_2_w;
struct ggml_tensor * dense_2_b;
struct ggml_tensor * layer_norm_1_gamma;
struct ggml_tensor * layer_norm_1_beta;
struct ggml_tensor * target_label_w;
struct ggml_tensor * target_label_b;
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_backend_buffer_t buf_w = nullptr;
struct ggml_context * ctx_w = nullptr;
};
struct ggml_tensor * checked_get_tensor(struct ggml_context * ctx, const char * name) {
struct ggml_tensor * tensor = ggml_get_tensor(ctx, name);
if (!tensor) {
fprintf(stderr, "%s: tensor '%s' not found\n", __func__, name);
throw std::runtime_error("ggml_get_tensor() failed");
}
return tensor;
}
bool magika_model_load(const std::string & fname, magika_model & model) {
auto & ctx = model.ctx_w;
struct gguf_init_params params = {
/*.no_alloc =*/ true,
/*.ctx =*/ &ctx,
};
struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params);
if (!ctx_gguf) {
fprintf(stderr, "%s: gguf_init_from_file() failed\n", __func__);
return false;
}
model.buf_w = ggml_backend_alloc_ctx_tensors(ctx, model.backend);
if (!model.buf_w) {
fprintf(stderr, "%s: ggml_backend_alloc_ctx_tensors() failed\n", __func__);
gguf_free(ctx_gguf);
return false;
}
try {
model.dense_w = checked_get_tensor(ctx, "dense/kernel:0");
model.dense_b = checked_get_tensor(ctx, "dense/bias:0");
model.layer_norm_gamma = checked_get_tensor(ctx, "layer_normalization/gamma:0");
model.layer_norm_beta = checked_get_tensor(ctx, "layer_normalization/beta:0");
model.dense_1_w = checked_get_tensor(ctx, "dense_1/kernel:0");
model.dense_1_b = checked_get_tensor(ctx, "dense_1/bias:0");
model.dense_2_w = checked_get_tensor(ctx, "dense_2/kernel:0");
model.dense_2_b = checked_get_tensor(ctx, "dense_2/bias:0");
model.layer_norm_1_gamma = checked_get_tensor(ctx, "layer_normalization_1/gamma:0");
model.layer_norm_1_beta = checked_get_tensor(ctx, "layer_normalization_1/beta:0");
model.target_label_w = checked_get_tensor(ctx, "target_label/kernel:0");
model.target_label_b = checked_get_tensor(ctx, "target_label/bias:0");
} catch (const std::exception & e) {
fprintf(stderr, "%s: %s\n", __func__, e.what());
gguf_free(ctx_gguf);
return false;
}
FILE * f = fopen(fname.c_str(), "rb");
if (!f) {
fprintf(stderr, "%s: fopen() failed\n", __func__);
gguf_free(ctx_gguf);
return false;
}
const int n_tensors = gguf_get_n_tensors(ctx_gguf);
for (int i = 0; i < n_tensors; i++) {
const char * name = gguf_get_tensor_name(ctx_gguf, i);
struct ggml_tensor * tensor = ggml_get_tensor(ctx, name);
size_t offs = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i);
//printf("%-30s: [%3ld, %3ld, %3ld, %3ld] %s\n",
// name,
// tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3],
// ggml_type_name(tensor->type));
std::vector<uint8_t> buf(ggml_nbytes(tensor));
if (fseek(f, offs, SEEK_SET) != 0) {
fprintf(stderr, "%s: fseek() failed\n", __func__);
gguf_free(ctx_gguf);
fclose(f);
return false;
}
if (fread(buf.data(), 1, buf.size(), f) != buf.size()) {
fprintf(stderr, "%s: fread() failed\n", __func__);
gguf_free(ctx_gguf);
fclose(f);
return false;
}
ggml_backend_tensor_set(tensor, buf.data(), 0, buf.size());
}
fclose(f);
gguf_free(ctx_gguf);
return true;
}
struct ggml_cgraph * magika_graph(
const magika_model & model,
const int n_files) {
const auto & hparams = model.hparams;
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_cgraph * gf = ggml_new_graph(ctx);
struct ggml_tensor * input = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 257, 1536, n_files); // one-hot
ggml_set_name(input, "input");
ggml_set_input(input);
struct ggml_tensor * cur;
// dense
cur = ggml_mul_mat(ctx, model.dense_w, input);
cur = ggml_add(ctx, cur, model.dense_b); // [128, 1536, n_files]
cur = ggml_gelu(ctx, cur);
// reshape
cur = ggml_reshape_3d(ctx, cur, 512, 384, n_files); // [384, 512, n_files]
cur = ggml_cont(ctx, ggml_transpose(ctx, cur));
// layer normalization
cur = ggml_norm(ctx, cur, hparams.f_norm_eps);
cur = ggml_mul(ctx, cur, model.layer_norm_gamma); // [384, 512, n_files]
cur = ggml_add(ctx, cur, model.layer_norm_beta); // [384, 512, n_files]
// dense_1
cur = ggml_cont(ctx, ggml_transpose(ctx, cur));
cur = ggml_mul_mat(ctx, model.dense_1_w, cur);
cur = ggml_add(ctx, cur, model.dense_1_b); // [256, 384, n_files]
cur = ggml_gelu(ctx, cur);
// dense_2
cur = ggml_mul_mat(ctx, model.dense_2_w, cur);
cur = ggml_add(ctx, cur, model.dense_2_b); // [256, 384, n_files]
cur = ggml_gelu(ctx, cur);
// global_max_pooling1d
cur = ggml_cont(ctx, ggml_transpose(ctx, cur)); // [384, 256, n_files]
cur = ggml_pool_1d(ctx, cur, GGML_OP_POOL_MAX, 384, 384, 0); // [1, 256, n_files]
cur = ggml_reshape_2d(ctx, cur, 256, n_files); // [256, n_files]
// layer normalization 1
cur = ggml_norm(ctx, cur, hparams.f_norm_eps);
cur = ggml_mul(ctx, cur, model.layer_norm_1_gamma); // [256, n_files]
cur = ggml_add(ctx, cur, model.layer_norm_1_beta); // [256, n_files]
// target_label
cur = ggml_mul_mat(ctx, model.target_label_w, cur);
cur = ggml_add(ctx, cur, model.target_label_b); // [n_label, n_files]
cur = ggml_soft_max(ctx, cur); // [n_label, n_files]
ggml_set_name(cur, "target_label_probs");
ggml_set_output(cur);
ggml_build_forward_expand(gf, cur);
return gf;
}
bool magika_eval(
struct magika_model & model,
const std::vector<std::string> & fnames) {
const auto & hparams = model.hparams;
static ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
struct ggml_cgraph * gf = magika_graph(model, fnames.size());
if (!ggml_gallocr_alloc_graph(alloc, gf)) {
fprintf(stderr, "%s: ggml_gallocr_alloc_graph() failed\n", __func__);
return false;
}
struct ggml_tensor * input = ggml_graph_get_tensor(gf, "input");
for (size_t i = 0; i < fnames.size(); i++) {
FILE * f = fopen(fnames[i].c_str(), "rb");
if (!f) {
fprintf(stderr, "%s: fopen() failed\n", __func__);
return false;
}
fseek(f, 0, SEEK_END);
long fsize = ftell(f);
// the buffer is padded with the padding_token if the file is smaller than the block size
std::vector<int> buf(1536, hparams.padding_token);
std::vector<uint8_t> read_buf(std::max(hparams.beg_size, std::max(hparams.mid_size, hparams.end_size)));
// read beg
fseek(f, 0, SEEK_SET);
int n_read = fread(read_buf.data(), 1, hparams.beg_size, f);
for (int j = 0; j < n_read; j++) {
// pad at the end
buf[j] = read_buf[j];
}
// read mid
long mid_offs = std::max(0L, (fsize - hparams.mid_size) / 2);
fseek(f, mid_offs, SEEK_SET);
n_read = fread(read_buf.data(), 1, hparams.mid_size, f);
for (int j = 0; j < n_read; j++) {
// pad at both ends
long mid_idx = hparams.beg_size + (hparams.mid_size / 2) - n_read / 2 + j;
buf[mid_idx] = read_buf[j];
}
// read end
long end_offs = std::max(0L, fsize - hparams.end_size);
fseek(f, end_offs, SEEK_SET);
n_read = fread(read_buf.data(), 1, hparams.end_size, f);
for (int j = 0; j < n_read; j++) {
// pad at the beginning
int end_idx = hparams.beg_size + hparams.mid_size + hparams.end_size - n_read + j;
buf[end_idx] = read_buf[j];
}
fclose(f);
const size_t inp_bytes = hparams.beg_size + hparams.mid_size + hparams.end_size;
// convert to one-hot
std::vector<float> one_hot(257*inp_bytes);
for (size_t j = 0; j < inp_bytes; j++) {
one_hot[257*j + buf[j]] = 1.0f;
}
ggml_backend_tensor_set(input, one_hot.data(), 257*inp_bytes*i*sizeof(float), 257*inp_bytes*sizeof(float));
}
if (ggml_backend_graph_compute(model.backend, gf) != GGML_STATUS_SUCCESS) {
fprintf(stderr, "%s: ggml_backend_graph_compute() failed\n", __func__);
return false;
}
struct ggml_tensor * target_label_probs = ggml_graph_get_tensor(gf, "target_label_probs");
// print probabilities for the top labels of each file
for (size_t i = 0; i < fnames.size(); i++) {
std::vector<float> probs(hparams.n_label);
ggml_backend_tensor_get(target_label_probs, probs.data(), hparams.n_label*i*sizeof(float), hparams.n_label*sizeof(float));
// sort the probabilities
std::vector<int> idx(hparams.n_label);
std::iota(idx.begin(), idx.end(), 0);
std::sort(idx.begin(), idx.end(), [&probs](int i1, int i2) { return probs[i1] > probs[i2]; });
// print the top labels
const int top_n = 5;
printf("%-30s: ", fnames[i].c_str());
for (int j = 0; j < top_n; j++) {
printf("%s (%.2f%%) ", magika_labels[idx[j]], probs[idx[j]]*100);
}
printf("\n");
}
return true;
}
int main(int argc, const char ** argv) {
if (argc < 3) {
fprintf(stderr, "usage: %s <model> <file1> [<file2> ...]\n", argv[0]);
return 1;
}
const char * model_fname = argv[1];
std::vector<std::string> fnames;
for (int i = 2; i < argc; i++) {
fnames.push_back(argv[i]);
}
magika_model model;
if (!magika_model_load(model_fname, model)) {
fprintf(stderr, "magika_model_load() failed\n");
return 1;
}
magika_eval(model, fnames);
return 0;
}
engine/ggml/examples/mnist/.gitignore
+3
View File
@@ -0,0 +1,3 @@
data/
*.gguf
*.ggml

Some files were not shown because too many files have changed in this diff Show More