vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'upstream/master' into deepseek-v2

Browse source
This commit is contained in:
Stanisław Szymczyk 2024-05-24 16:13:00 +02:00
commit a54685b98a
99 changed files with 24473 additions and 17181 deletions
Download patch file Download diff file Expand all files Collapse all files

1
.devops/nix/package.nix
View file

@ -214,7 +214,6 @@ effectiveStdenv.mkDerivation (
(cmakeBool "LLAMA_CUDA" useCuda)
(cmakeBool "LLAMA_HIPBLAS" useRocm)
(cmakeBool "LLAMA_METAL" useMetalKit)
(cmakeBool "LLAMA_MPI" useMpi)
(cmakeBool "LLAMA_VULKAN" useVulkan)
(cmakeBool "LLAMA_STATIC" enableStatic)
]

78
.github/labeler.yml vendored Normal file
View file

@ -0,0 +1,78 @@
# https://github.com/actions/labeler
SYCL:
- changed-files:
- any-glob-to-any-file:
- ggml-sycl.h
- ggml-sycl.cpp
- README-sycl.md
Nvidia GPU:
- changed-files:
- any-glob-to-any-file:
- ggml-cuda/**
Vulkan:
- changed-files:
- any-glob-to-any-file:
- ggml_vk_generate_shaders.py
- ggml-vulkan*
documentation:
- changed-files:
- any-glob-to-any-file:
- docs/**
- media/**
testing:
- changed-files:
- any-glob-to-any-file:
- tests/**
build:
- changed-files:
- any-glob-to-any-file:
- cmake/**
- CMakeLists.txt
- CMakePresets.json
- codecov.yml
examples:
- changed-files:
- any-glob-to-any-file: examples/**
devops:
- changed-files:
- any-glob-to-any-file:
- .devops/**
- .github/**
- ci/**
python:
- changed-files:
- any-glob-to-any-file:
- "**/*.py"
- requirements/**
- gguf-py/**
- .flake8
script:
- changed-files:
- any-glob-to-any-file:
- scripts/**
android:
- changed-files:
- any-glob-to-any-file:
- examples/llama.android/**
server:
- changed-files:
- any-glob-to-any-file:
- examples/server/**
ggml:
- changed-files:
- any-glob-to-any-file:
- ggml.c
- ggml.h
- ggml-*.c
- ggml-*.h
- ggml-cuda/**
nix:
- changed-files:
- any-glob-to-any-file:
- "**/*.nix"
- .github/workflows/nix-*.yml
- .devops/nix/nixpkgs-instances.nix
embedding:
- changed-files:
- any-glob-to-any-file: examples/embedding/

48
.github/workflows/build.yml vendored
View file

@ -271,49 +271,15 @@ jobs:
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
name: llama-bin-ubuntu-x64.zip
# ubuntu-latest-cmake-sanitizer:
# runs-on: ubuntu-latest
#
# continue-on-error: true
#
# strategy:
# matrix:
# sanitizer: [ADDRESS, THREAD, UNDEFINED]
# build_type: [Debug, Release]
#
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v4
#
# - name: Dependencies
# id: depends
# run: |
# sudo apt-get update
# sudo apt-get install build-essential
#
# - name: Build
# id: cmake_build
# run: |
# mkdir build
# cd build
# cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
# cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
#
# - name: Test
# id: cmake_test
# run: |
# cd build
# ctest -L main --verbose --timeout 900
ubuntu-latest-cmake-mpi:
ubuntu-latest-cmake-sanitizer:
runs-on: ubuntu-latest
continue-on-error: true
strategy:
matrix:
mpi_library: [mpich, libopenmpi-dev]
sanitizer: [ADDRESS, THREAD, UNDEFINED]
build_type: [Debug, Release]
steps:
- name: Clone
@ -324,21 +290,21 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential ${{ matrix.mpi_library }}
sudo apt-get install build-essential
- name: Build
id: cmake_build
run: |
mkdir build
cd build
cmake -DLLAMA_MPI=ON ..
cmake --build . --config Release -j $(nproc)
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
ctest -L main --verbose
ctest -L main --verbose --timeout 900
ubuntu-latest-cmake-rpc:
runs-on: ubuntu-latest

17
.github/workflows/labeler.yml vendored Normal file
View file

@ -0,0 +1,17 @@
name: "Pull Request Labeler"
on:
- pull_request_target
jobs:
labeler:
permissions:
contents: read
pull-requests: write
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
repository: "ggerganov/llama.cpp"
- uses: actions/labeler@v5
with:
configuration-path: '.github/labeler.yml'

8
.github/workflows/server.yml vendored
View file

@ -32,10 +32,8 @@ jobs:
strategy:
matrix:
# TODO: temporary disabled due to linux kernel issues
#sanitizer: [ADDRESS, THREAD, UNDEFINED]
sanitizer: [UNDEFINED]
build_type: [Debug]
sanitizer: [ADDRESS, THREAD, UNDEFINED]
build_type: [RelWithDebInfo]
include:
- build_type: Release
sanitizer: ""
@ -102,10 +100,8 @@ jobs:
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server
- name: Tests
id: server_integration_tests
if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
run: |
cd examples/server/tests
PORT=8888 ./tests.sh

29
.github/workflows/zig-build.yml vendored
View file

@ -1,29 +0,0 @@
name: Zig CI
on:
pull_request:
push:
branches:
- master
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
build:
strategy:
fail-fast: false
matrix:
runs-on: [ubuntu-latest, macos-latest, windows-latest]
runs-on: ${{ matrix.runs-on }}
steps:
- uses: actions/checkout@v4
with:
submodules: recursive
fetch-depth: 0
- uses: goto-bus-stop/setup-zig@v2
with:
version: 0.11.0
- name: Build Summary
run: zig build --summary all -freference-trace

69
CMakeLists.txt
View file

@ -1,4 +1,4 @@
cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
project("llama.cpp" C CXX)
include(CheckIncludeFileCXX)
@ -77,6 +77,7 @@ option(LLAMA_AVX2 "llama: enable AVX2"
option(LLAMA_AVX512 "llama: enable AVX512" OFF)
option(LLAMA_AVX512_VBMI "llama: enable AVX512-VBMI" OFF)
option(LLAMA_AVX512_VNNI "llama: enable AVX512-VNNI" OFF)
option(LLAMA_AVX512_BF16 "llama: enable AVX512-BF16" OFF)
option(LLAMA_FMA "llama: enable FMA" ${INS_ENB})
# in MSVC F16C is implied with AVX2/AVX512
if (NOT MSVC)
@ -122,9 +123,7 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
"llama: metal minimum macOS version")
set(LLAMA_METAL_STD "" CACHE STRING "llama: metal standard version (-std flag)")
option(LLAMA_KOMPUTE "llama: use Kompute" OFF)
option(LLAMA_MPI "llama: use MPI" OFF)
option(LLAMA_RPC "llama: use RPC" OFF)
option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF)
option(LLAMA_SYCL "llama: use SYCL" OFF)
option(LLAMA_SYCL_F16 "llama: use 16 bit floats for sycl calculations" OFF)
set(LLAMA_SYCL_TARGET "INTEL" CACHE STRING "llama: sycl target device")
@ -134,6 +133,8 @@ set(LLAMA_SCHED_MAX_COPIES "4" CACHE STRING "llama: max input copies for pipeli
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ON)
option(LLAMA_LASX "llama: enable lasx" ON)
option(LLAMA_LSX "llama: enable lsx" ON)
# add perf arguments
option(LLAMA_PERF "llama: enable perf" OFF)
@ -382,10 +383,6 @@ if (LLAMA_LLAMAFILE)
set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
endif()
if (LLAMA_QKK_64)
add_compile_definitions(GGML_QKK_64)
endif()
if (LLAMA_CUBLAS)
message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
set(LLAMA_CUDA ON)
@ -466,35 +463,6 @@ if (LLAMA_CUDA)
endif()
endif()
if (LLAMA_MPI)
cmake_minimum_required(VERSION 3.10)
find_package(MPI)
if (MPI_C_FOUND)
message(STATUS "MPI found")
set(GGML_HEADERS_MPI ggml-mpi.h)
set(GGML_SOURCES_MPI ggml-mpi.c)
add_compile_definitions(GGML_USE_MPI)
add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
if (NOT MSVC)
add_compile_options(-Wno-cast-qual)
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${MPI_C_LIBRARIES})
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
# Even if you're only using the C header, C++ programs may bring in MPI
# C++ functions, so more linkage is needed
if (MPI_CXX_FOUND)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${MPI_CXX_LIBRARIES})
endif()
else()
message(WARNING "MPI not found")
endif()
endif()
if (LLAMA_RPC)
add_compile_definitions(GGML_USE_RPC)
@ -532,6 +500,12 @@ if (LLAMA_VULKAN)
add_compile_definitions(GGML_USE_VULKAN)
# Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
# Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
endif()
if (LLAMA_VULKAN_CHECK_RESULTS)
add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
endif()
@ -581,7 +555,7 @@ if (LLAMA_HIPBLAS)
else()
# Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_ARGETS})
set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
endif()
cmake_minimum_required(VERSION 3.21)
enable_language(HIP)
@ -1090,6 +1064,10 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
endif()
if (LLAMA_AVX512_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
elseif (LLAMA_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (LLAMA_AVX)
@ -1121,6 +1099,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
if (LLAMA_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
if (LLAMA_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
message(STATUS "PowerPC detected")
@ -1130,6 +1111,17 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
#TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
message(STATUS "loongarch64 detected")
list(APPEND ARCH_FLAGS -march=loongarch64)
if (LLAMA_LASX)
list(APPEND ARCH_FLAGS -mlasx)
endif()
if (LLAMA_LSX)
list(APPEND ARCH_FLAGS -mlsx)
endif()
else()
message(STATUS "Unknown architecture")
endif()
@ -1218,7 +1210,6 @@ add_library(ggml OBJECT
${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
${GGML_SOURCES_RPC} ${GGML_HEADERS_RPC}
${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
${GGML_SOURCES_SYCL} ${GGML_HEADERS_SYCL}
@ -1306,7 +1297,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
"${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
"${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
"${GGML_HEADERS_METAL}" "${GGML_HEADERS_EXTRA}")
set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
install(TARGETS ggml PUBLIC_HEADER)

21
Makefile
View file

@ -379,15 +379,16 @@ ifneq ($(filter ppc64le%,$(UNAME_M)),)
CUDA_POWER_ARCH = 1
endif
ifneq ($(filter loongarch64%,$(UNAME_M)),)
MK_CFLAGS += -mlasx
MK_CXXFLAGS += -mlasx
endif
else
MK_CFLAGS += -march=rv64gcv -mabi=lp64d
MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
endif
ifdef LLAMA_QKK_64
MK_CPPFLAGS += -DGGML_QKK_64
endif
ifndef LLAMA_NO_ACCELERATE
# Mac OS - include Accelerate framework.
# `-framework Accelerate` works both with Apple Silicon and Mac Intel
@ -399,13 +400,6 @@ ifndef LLAMA_NO_ACCELERATE
endif
endif # LLAMA_NO_ACCELERATE
ifdef LLAMA_MPI
MK_CPPFLAGS += -DGGML_USE_MPI
MK_CFLAGS += -Wno-cast-qual
MK_CXXFLAGS += -Wno-cast-qual
OBJS += ggml-mpi.o
endif # LLAMA_MPI
ifdef LLAMA_OPENBLAS
MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas)
@ -629,11 +623,6 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
endif
endif # LLAMA_METAL
ifdef LLAMA_MPI
ggml-mpi.o: ggml-mpi.c ggml-mpi.h
$(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_MPI
ifndef LLAMA_NO_LLAMAFILE
sgemm.o: sgemm.cpp sgemm.h ggml.h
$(CXX) $(CXXFLAGS) -c $< -o $@

44
README.md
View file

@ -107,7 +107,6 @@ Typically finetunes of the base models below are supported as well.
- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
@ -128,6 +127,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
- [x] [OLMo](https://allenai.org/olmo)
- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
@ -141,6 +141,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
**HTTP server**
@ -301,7 +302,7 @@ cd llama.cpp
### Build
In order to build llama.cpp you have three different options.
In order to build llama.cpp you have four different options.
- Using `make`:
- On Linux or MacOS:
@ -382,45 +383,6 @@ To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or th
When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
argument.
### MPI Build
MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
- Using `make`:
```bash
make CC=mpicc CXX=mpicxx LLAMA_MPI=1
```
- Using `CMake`:
```bash
cmake -S . -B build -DLLAMA_MPI=ON
```
Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
Here is an example hostfile:
```
192.168.0.1:2
malvolio.local:1
```
The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
Finally, you're ready to run a computation using `mpirun`:
```bash
mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
```
### BLAS Build
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:

172
build.zig
View file

@ -1,172 +0,0 @@
// Compatible with Zig Version 0.11.0
const std = @import("std");
const ArrayList = std.ArrayList;
const Compile = std.Build.Step.Compile;
const ConfigHeader = std.Build.Step.ConfigHeader;
const Mode = std.builtin.Mode;
const CrossTarget = std.zig.CrossTarget;
const Maker = struct {
builder: *std.build.Builder,
target: CrossTarget,
optimize: Mode,
enable_lto: bool,
include_dirs: ArrayList([]const u8),
cflags: ArrayList([]const u8),
cxxflags: ArrayList([]const u8),
objs: ArrayList(*Compile),
fn addInclude(m: *Maker, dir: []const u8) !void {
try m.include_dirs.append(dir);
}
fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
}
fn addCFlag(m: *Maker, flag: []const u8) !void {
try m.cflags.append(flag);
}
fn addCxxFlag(m: *Maker, flag: []const u8) !void {
try m.cxxflags.append(flag);
}
fn addFlag(m: *Maker, flag: []const u8) !void {
try m.addCFlag(flag);
try m.addCxxFlag(flag);
}
fn init(builder: *std.build.Builder) !Maker {
const target = builder.standardTargetOptions(.{});
const zig_version = @import("builtin").zig_version_string;
const commit_hash = try std.ChildProcess.exec(
.{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
);
try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
\\int LLAMA_BUILD_NUMBER = {};
\\char const *LLAMA_COMMIT = "{s}";
\\char const *LLAMA_COMPILER = "Zig {s}";
\\char const *LLAMA_BUILD_TARGET = "{s}";
\\
, .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
var m = Maker{
.builder = builder,
.target = target,
.optimize = builder.standardOptimizeOption(.{}),
.enable_lto = false,
.include_dirs = ArrayList([]const u8).init(builder.allocator),
.cflags = ArrayList([]const u8).init(builder.allocator),
.cxxflags = ArrayList([]const u8).init(builder.allocator),
.objs = ArrayList(*Compile).init(builder.allocator),
};
try m.addCFlag("-std=c11");
try m.addCxxFlag("-std=c++11");
try m.addProjectInclude(&.{});
try m.addProjectInclude(&.{"common"});
return m;
}
fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
if (o.target.getAbi() != .msvc)
o.defineCMacro("_GNU_SOURCE", null);
if (std.mem.endsWith(u8, src, ".c")) {
o.addCSourceFiles(&.{src}, m.cflags.items);
o.linkLibC();
} else {
o.addCSourceFiles(&.{src}, m.cxxflags.items);
if (o.target.getAbi() == .msvc) {
o.linkLibC(); // need winsdk + crt
} else {
// linkLibCpp already add (libc++ + libunwind + libc)
o.linkLibCpp();
}
}
for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
o.want_lto = m.enable_lto;
return o;
}
fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
e.addCSourceFiles(&.{src}, m.cxxflags.items);
for (deps) |d| e.addObject(d);
for (m.objs.items) |o| e.addObject(o);
for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
// https://github.com/ziglang/zig/issues/15448
if (e.target.getAbi() == .msvc) {
e.linkLibC(); // need winsdk + crt
} else {
// linkLibCpp already add (libc++ + libunwind + libc)
e.linkLibCpp();
}
m.builder.installArtifact(e);
e.want_lto = m.enable_lto;
return e;
}
};
pub fn build(b: *std.build.Builder) !void {
var make = try Maker.init(b);
make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
const ggml = make.obj("ggml", "ggml.c");
const sgemm = make.obj("sgemm", "sgemm.cpp");
const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
const unicode = make.obj("unicode", "unicode.cpp");
const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
const llama = make.obj("llama", "llama.cpp");
const buildinfo = make.obj("common", "common/build-info.cpp");
const common = make.obj("common", "common/common.cpp");
const console = make.obj("console", "common/console.cpp");
const sampling = make.obj("sampling", "common/sampling.cpp");
const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
const train = make.obj("train", "common/train.cpp");
const clip = make.obj("clip", "examples/llava/clip.cpp");
const llava = make.obj("llava", "examples/llava/llava.cpp");
_ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
_ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
_ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
_ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
_ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
_ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
if (server.target.isWindows()) {
server.linkSystemLibrary("ws2_32");
}
const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
for (server_assets) |asset| {
const input_path = b.fmt("examples/server/public/{s}", .{asset});
const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
// Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
defer b.allocator.free(input);
var buf = std.ArrayList(u8).init(b.allocator);
defer buf.deinit();
for (input) |byte| {
try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
}
var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
defer b.allocator.free(name);
std.mem.replaceScalar(u8, name, '.', '_');
try std.fs.cwd().writeFile(output_path, b.fmt(
"unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
.{ name, buf.items, name, input.len },
));
std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
}
}

423
ci/run.sh
View file

@ -202,12 +202,15 @@ function gg_sum_test_scripts_release {
}
function gg_get_model {
local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
if [[ -s $gguf_3b ]]; then
echo -n "$gguf_3b"
elif [[ -s $gguf_7b ]]; then
echo -n "$gguf_7b"
local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
if [[ -s $gguf_0 ]]; then
echo -n "$gguf_0"
elif [[ -s $gguf_1 ]]; then
echo -n "$gguf_1"
elif [[ -s $gguf_2 ]]; then
echo -n "$gguf_2"
else
echo >&2 "No model found. Can't run gg_run_ctest_with_model."
exit 1
@ -256,139 +259,6 @@ function gg_sum_ctest_with_model_release {
gg_printf '```\n'
}
# open_llama_3b_v2
function gg_run_open_llama_3b_v2 {
cd ${SRC}
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
path_models="../models-mnt/open-llama/3B-v2"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert.py ${path_models}
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
model_q5_0="${path_models}/ggml-model-q5_0.gguf"
model_q5_1="${path_models}/ggml-model-q5_1.gguf"
model_q2_k="${path_models}/ggml-model-q2_k.gguf"
model_q3_k="${path_models}/ggml-model-q3_k.gguf"
model_q4_k="${path_models}/ggml-model-q4_k.gguf"
model_q5_k="${path_models}/ggml-model-q5_k.gguf"
model_q6_k="${path_models}/ggml-model-q6_k.gguf"
wiki_test_60="${path_wiki}/wiki.test-60.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/main --model ${model_f16} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$ppl"
return 0
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
set +e
}
function gg_sum_open_llama_3b_v2 {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'OpenLLaMA 3B-v2:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# open_llama_7b_v2
# requires: GG_BUILD_CUDA
@ -417,7 +287,7 @@ function gg_run_open_llama_7b_v2 {
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert.py ${path_models}
python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@ -526,6 +396,272 @@ function gg_sum_open_llama_7b_v2 {
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# pythia_1.4b
function gg_run_pythia_1_4b {
cd ${SRC}
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
path_models="../models-mnt/pythia/1.4B"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
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 ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
model_q5_0="${path_models}/ggml-model-q5_0.gguf"
model_q5_1="${path_models}/ggml-model-q5_1.gguf"
model_q2_k="${path_models}/ggml-model-q2_k.gguf"
model_q3_k="${path_models}/ggml-model-q3_k.gguf"
model_q4_k="${path_models}/ggml-model-q4_k.gguf"
model_q5_k="${path_models}/ggml-model-q5_k.gguf"
model_q6_k="${path_models}/ggml-model-q6_k.gguf"
wiki_test_60="${path_wiki}/wiki.test-60.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/main --model ${model_f16} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$ppl"
return 0
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
#check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
set +e
}
function gg_sum_pythia_1_4b {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Pythia 1.4B:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# pythia_2_8b
# requires: GG_BUILD_CUDA
function gg_run_pythia_2_8b {
cd ${SRC}
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
path_models="../models-mnt/pythia/2.8B"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
model_q5_0="${path_models}/ggml-model-q5_0.gguf"
model_q5_1="${path_models}/ggml-model-q5_1.gguf"
model_q2_k="${path_models}/ggml-model-q2_k.gguf"
model_q3_k="${path_models}/ggml-model-q3_k.gguf"
model_q4_k="${path_models}/ggml-model-q4_k.gguf"
model_q5_k="${path_models}/ggml-model-q5_k.gguf"
model_q6_k="${path_models}/ggml-model-q6_k.gguf"
wiki_test="${path_wiki}/wiki.test.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/main --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$ppl"
return 0
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
#check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
set +e
}
function gg_sum_pythia_2_8b {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Pythia 2.8B:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# bge-small
function gg_run_embd_bge_small {
@ -552,7 +688,7 @@ function gg_run_embd_bge_small {
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models}
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@ -606,9 +742,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
if [ -z ${GG_BUILD_CUDA} ]; then
test $ret -eq 0 && gg_run open_llama_3b_v2
test $ret -eq 0 && gg_run pythia_1_4b
else
test $ret -eq 0 && gg_run open_llama_7b_v2
test $ret -eq 0 && gg_run pythia_2_8b
#test $ret -eq 0 && gg_run open_llama_7b_v2
fi
test $ret -eq 0 && gg_run ctest_with_model_debug
test $ret -eq 0 && gg_run ctest_with_model_release

1308
common/common.cpp
View file

File diff suppressed because it is too large Load diff

89
common/common.h
View file

@ -27,7 +27,7 @@
#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
#define print_build_info() do { \
fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT); \
fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT); \
fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET); \
} while(0)
@ -35,14 +35,18 @@
// build info
extern int LLAMA_BUILD_NUMBER;
extern char const *LLAMA_COMMIT;
extern char const *LLAMA_COMPILER;
extern char const *LLAMA_BUILD_TARGET;
extern char const * LLAMA_COMMIT;
extern char const * LLAMA_COMPILER;
extern char const * LLAMA_BUILD_TARGET;
struct llama_control_vector_load_info;
int get_math_cpu_count();
int32_t get_num_physical_cores();
//
// CPU utils
//
int32_t cpu_get_num_physical_cores();
int32_t cpu_get_num_math();
//
// CLI argument parsing
@ -51,7 +55,7 @@ int32_t get_num_physical_cores();
struct gpt_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed
int32_t n_threads = get_math_cpu_count();
int32_t n_threads = cpu_get_num_math();
int32_t n_threads_draft = -1;
int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads)
int32_t n_threads_batch_draft = -1;
@ -179,33 +183,34 @@ struct gpt_params {
void gpt_params_handle_model_default(gpt_params & params);
bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
bool gpt_params_parse_ex (int argc, char ** argv, gpt_params & params);
bool gpt_params_parse (int argc, char ** argv, gpt_params & params);
bool gpt_params_find_arg (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
std::string get_system_info(const gpt_params & params);
std::string gpt_random_prompt(std::mt19937 & rng);
void process_escapes(std::string& input);
bool validate_file_name(const std::string & filename);
std::string gpt_params_get_system_info(const gpt_params & params);
//
// String utils
//
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
std::vector<std::string> string_split(std::string input, char separator);
std::string string_strip(const std::string & str);
std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
std::string string_get_sortable_timestamp();
std::string string_random_prompt(std::mt19937 & rng);
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
//
// Filesystem utils
//
bool fs_validate_filename(const std::string & filename);
bool fs_create_directory_with_parents(const std::string & path);
std::string fs_get_cache_directory();
//
// Model utils
@ -276,29 +281,15 @@ std::string llama_detokenize_bpe(
// defaults to true when model type is SPM, otherwise false.
bool llama_should_add_bos_token(const llama_model * model);
//
// YAML utils
//
bool create_directory_with_parents(const std::string & path);
void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
std::string get_sortable_timestamp();
void dump_non_result_info_yaml(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
//
// KV cache utils
//
// Dump the KV cache view with the number of sequences per cell.
void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
// Dump the KV cache view showing individual sequences in each cell (long output).
void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
//
// Embedding utils
@ -332,6 +323,20 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
//
// Split utils
//
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
//
// YAML utils
//
void yaml_dump_vector_float (FILE * stream, const char * prop_name, const std::vector<float> & data);
void yaml_dump_vector_int (FILE * stream, const char * prop_name, const std::vector<int> & data);
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
void yaml_dump_non_result_info(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

96
common/sampling.cpp
View file

@ -125,7 +125,7 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
std::string result = "CFG -> Penalties ";
if (params.mirostat == 0) {
for (auto sampler_type : params.samplers_sequence) {
const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
if (!sampler_type_name.empty()) {
result += "-> " + sampler_type_name + " ";
}
@ -137,6 +137,87 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
return result;
}
std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
switch (sampler_type) {
case llama_sampler_type::TOP_K: return "top_k";
case llama_sampler_type::TFS_Z: return "tfs_z";
case llama_sampler_type::TYPICAL_P: return "typical_p";
case llama_sampler_type::TOP_P: return "top_p";
case llama_sampler_type::MIN_P: return "min_p";
case llama_sampler_type::TEMPERATURE: return "temperature";
default : return "";
}
}
std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
{"top_k", llama_sampler_type::TOP_K},
{"top_p", llama_sampler_type::TOP_P},
{"typical_p", llama_sampler_type::TYPICAL_P},
{"min_p", llama_sampler_type::MIN_P},
{"tfs_z", llama_sampler_type::TFS_Z},
{"temperature", llama_sampler_type::TEMPERATURE}
};
// since samplers names are written multiple ways
// make it ready for both system names and input names
std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
{"top-k", llama_sampler_type::TOP_K},
{"top-p", llama_sampler_type::TOP_P},
{"nucleus", llama_sampler_type::TOP_P},
{"typical-p", llama_sampler_type::TYPICAL_P},
{"typical", llama_sampler_type::TYPICAL_P},
{"min-p", llama_sampler_type::MIN_P},
{"tfs-z", llama_sampler_type::TFS_Z},
{"tfs", llama_sampler_type::TFS_Z},
{"temp", llama_sampler_type::TEMPERATURE}
};
std::vector<llama_sampler_type> sampler_types;
sampler_types.reserve(names.size());
for (const auto & name : names)
{
auto sampler_item = sampler_canonical_name_map.find(name);
if (sampler_item != sampler_canonical_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
else
{
if (allow_alt_names)
{
sampler_item = sampler_alt_name_map.find(name);
if (sampler_item != sampler_alt_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
}
}
}
return sampler_types;
}
std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
std::unordered_map<char, llama_sampler_type> sampler_name_map {
{'k', llama_sampler_type::TOP_K},
{'p', llama_sampler_type::TOP_P},
{'y', llama_sampler_type::TYPICAL_P},
{'m', llama_sampler_type::MIN_P},
{'f', llama_sampler_type::TFS_Z},
{'t', llama_sampler_type::TEMPERATURE}
};
std::vector<llama_sampler_type> sampler_types;
sampler_types.reserve(names_string.size());
for (const auto & c : names_string) {
const auto sampler_item = sampler_name_map.find(c);
if (sampler_item != sampler_name_map.end()) {
sampler_types.push_back(sampler_item->second);
}
}
return sampler_types;
}
// no reasons to expose this function in header
static void sampler_queue(
struct llama_context * ctx_main,
@ -179,7 +260,7 @@ static llama_token llama_sampling_sample_impl(
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx,
bool is_resampling) { // Add a parameter to indicate if we are resampling
bool is_resampling) {
const llama_sampling_params & params = ctx_sampling->params;
const float temp = params.temp;
@ -188,8 +269,8 @@ static llama_token llama_sampling_sample_impl(
const float mirostat_eta = params.mirostat_eta;
std::vector<float> original_logits;
auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
if (!is_resampling) {
auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
if (ctx_sampling->grammar != NULL && !is_resampling) {
GGML_ASSERT(!original_logits.empty());
}
llama_token id = 0;
@ -252,7 +333,7 @@ static llama_token llama_sampling_sample_impl(
// Restore logits from the copy
std::copy(original_logits.begin(), original_logits.end(), logits);
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true); // Pass true for is_resampling
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
}
}
@ -285,7 +366,8 @@ static llama_token_data_array llama_sampling_prepare_impl(
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
if (apply_grammar && original_logits != NULL) {
if (ctx_sampling->grammar != NULL && !apply_grammar) {
GGML_ASSERT(original_logits != NULL);
// Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
*original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
}
@ -342,7 +424,7 @@ llama_token llama_sampling_sample(
struct llama_context * ctx_cfg,
const int idx) {
// Call the implementation function with is_resampling set to false by default
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
}
llama_token_data_array llama_sampling_prepare(

5
common/sampling.h
View file

@ -116,6 +116,11 @@ std::string llama_sampling_print(const llama_sampling_params & params);
// Print sampling order into a string
std::string llama_sampling_order_print(const llama_sampling_params & params);
std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
// this is a common sampling function used across the examples for convenience
// it can serve as a starting point for implementing your own sampling function
// Note: When using multiple sequences, it is the caller's responsibility to call

2
common/train.cpp
View file

@ -1380,7 +1380,7 @@ bool consume_common_train_arg(
void finish_processing_train_args(struct train_params_common * params) {
if (params->escape) {
process_escapes(params->sample_start);
string_process_escapes(params->sample_start);
}
}

1
convert-hf-to-gguf-update.py
View file

@ -72,6 +72,7 @@ models = [
{"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
{"name": "stablelm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
{"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
{"name": "command-r", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
{"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },

338
convert-hf-to-gguf.py
View file

@ -14,6 +14,7 @@ from pathlib import Path
from hashlib import sha256
from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast
import math
import numpy as np
import torch
@ -446,6 +447,9 @@ class Model:
if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
# ref: https://huggingface.co/openai-community/gpt2
res = "gpt-2"
if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
# ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
res = "stablelm2"
if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
# ref: https://huggingface.co/smallcloudai/Refact-1_6-base
res = "refact"
@ -573,6 +577,10 @@ class Model:
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.IdToPiece(token_id)
text = piece.encode("utf-8")
@ -588,21 +596,23 @@ class Model:
elif tokenizer.IsByte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
tokens[token_id] = text
scores[token_id] = score
toktypes[token_id] = toktype
added_tokens_file = self.dir_model / 'added_tokens.json'
if added_tokens_file.is_file():
with open(added_tokens_file, "r", encoding="utf-8") as f:
added_tokens_json = json.load(f)
for key in added_tokens_json:
key = key.encode("utf-8")
if key not in tokens:
tokens.append(key)
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
token_id = added_tokens_json[key]
if (token_id >= vocab_size):
logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
continue
tokens[token_id] = key.encode("utf-8")
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
if vocab_size > len(tokens):
pad_count = vocab_size - len(tokens)
@ -612,8 +622,6 @@ class Model:
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.UNUSED)
assert len(tokens) == vocab_size
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_tokenizer_pre("default")
self.gguf_writer.add_token_list(tokens)
@ -665,6 +673,44 @@ class GPTNeoXModel(Model):
self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
tensors: list[tuple[str, Tensor]] = []
if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
# Map bloom-style qkv_linear to gpt-style qkv_linear
# bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252 # noqa
# gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312 # noqa
qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
data_torch = torch.cat(
(
qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
),
dim=0,
)
logger.info("re-format attention.linear_qkv.weight")
elif re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.bias", name):
qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
data_torch = torch.cat(
(
qkv_bias[:, 0, :].reshape((n_embed,)),
qkv_bias[:, 1, :].reshape((n_embed,)),
qkv_bias[:, 2, :].reshape((n_embed,)),
),
dim=0,
)
logger.info("re-format attention.linear_qkv.bias")
tensors.append((self.map_tensor_name(name), data_torch))
return tensors
@Model.register("BloomForCausalLM")
class BloomModel(Model):
@ -1141,45 +1187,6 @@ class RefactModel(Model):
return tensors
@Model.register("PersimmonForCausalLM")
class PersimmonModel(Model):
model_arch = gguf.MODEL_ARCH.PERSIMMON
def set_gguf_parameters(self):
block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
head_count = self.hparams["num_attention_heads"]
head_count_kv = head_count
hidden_size = self.hparams["hidden_size"]
self.gguf_writer.add_name('persimmon-8b-chat')
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hidden_size)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
# NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
# than the head size?
# ref: https://github.com/ggerganov/llama.cpp/pull/4889
# self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
self.gguf_writer.add_head_count(head_count)
self.gguf_writer.add_head_count_kv(head_count_kv)
self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
def set_vocab(self):
self._set_vocab_sentencepiece()
# self.gguf_writer.add_bos_token_id(71013)
# self.gguf_writer.add_eos_token_id(71013)
def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
del name, new_name, bid, n_dims # unused
# TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
return True
@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
class StableLMModel(Model):
model_arch = gguf.MODEL_ARCH.STABLELM
@ -1772,6 +1779,38 @@ class Phi3MiniModel(Model):
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
for token_id, foken_data in added_tokens_decoder.items():
token_id = int(token_id)
token = foken_data["content"].encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert tokens[token_id] == token
tokens[token_id] = token
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
if foken_data.get("special"):
toktypes[token_id] = SentencePieceTokenTypes.CONTROL
tokenizer_file = self.dir_model / 'tokenizer.json'
if tokenizer_file.is_file():
with open(tokenizer_file, "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
added_tokens = tokenizer_json.get("added_tokens", [])
for foken_data in added_tokens:
token_id = int(foken_data["id"])
token = foken_data["content"].encode("utf-8")
if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
assert tokens[token_id] == token
tokens[token_id] = token
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
if foken_data.get("special"):
toktypes[token_id] = SentencePieceTokenTypes.CONTROL
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_tokenizer_pre("default")
self.gguf_writer.add_token_list(tokens)
@ -1784,23 +1823,59 @@ class Phi3MiniModel(Model):
def set_gguf_parameters(self):
block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
rot_pct = 1.0
n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
rms_eps = self.find_hparam(["rms_norm_eps"])
max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
rope_dims = n_embd // n_head
self.gguf_writer.add_name("Phi3")
self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
self.gguf_writer.add_context_length(max_pos_embds)
self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(8192)
self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_head_count_kv(n_head_kv)
self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
self.gguf_writer.add_rope_dimension_count(rope_dims)
self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
self.gguf_writer.add_file_type(self.ftype)
# write rope scaling for long context (128k) model
rope_scaling = self.find_hparam(['rope_scaling'], True)
if (rope_scaling is None):
return
scale = max_pos_embds / orig_max_pos_embds
rope_scaling_type = rope_scaling.get('type', '').lower()
if len(rope_scaling_type) == 0:
raise KeyError('Missing the required key rope_scaling.type')
if rope_scaling_type == 'su':
attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
elif rope_scaling_type == 'yarn':
attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
else:
raise NotImplementedError(f'The rope scaling type {rope_scaling_type} is not supported yet')
self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
long_factors = rope_scaling.get('long_factor', None)
short_factors = rope_scaling.get('short_factor', None)
if long_factors is None or short_factors is None:
raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
@Model.register("PlamoForCausalLM")
class PlamoModel(Model):
@ -2391,6 +2466,157 @@ class JinaBertV2Model(BertModel):
self.gguf_writer.add_add_eos_token(True)
@Model.register("ArcticForCausalLM")
class ArcticModel(Model):
model_arch = gguf.MODEL_ARCH.ARCTIC
def set_vocab(self):
# The reason for using a custom implementation here is that the
# snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
# tokenizer.model and used them as BOS and EOS instead of adding new tokens.
from sentencepiece import SentencePieceProcessor
tokenizer_path = self.dir_model / 'tokenizer.model'
if not tokenizer_path.is_file():
logger.error(f'Error: Missing {tokenizer_path}')
sys.exit(1)
# Read the whole vocabulary from the tokenizer.model file
tokenizer = SentencePieceProcessor()
tokenizer.LoadFromFile(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.IdToPiece(token_id)
text = piece.encode("utf-8")
score = tokenizer.GetScore(token_id)
toktype = SentencePieceTokenTypes.NORMAL
if tokenizer.IsUnknown(token_id):
toktype = SentencePieceTokenTypes.UNKNOWN
elif tokenizer.IsControl(token_id):
toktype = SentencePieceTokenTypes.CONTROL
elif tokenizer.IsUnused(token_id):
toktype = SentencePieceTokenTypes.UNUSED
elif tokenizer.IsByte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens[token_id] = text
scores[token_id] = score
toktypes[token_id] = toktype
# Use the added_tokens_decoder field from tokeniser_config.json as the source
# of information about added/redefined tokens and modify them accordingly.
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
if "added_tokens_decoder" in tokenizer_config_json:
added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
for token_id, token_json in added_tokens_decoder.items():
token_id = int(token_id)
if (token_id >= vocab_size):
logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
continue
token_content = token_json["content"]
token_type = SentencePieceTokenTypes.USER_DEFINED
token_score = -10000.0
# Map unk_token to UNKNOWN, other special tokens to CONTROL
# Set the score to 0.0 as in the original tokenizer.model
if ("special" in token_json) and token_json["special"]:
if token_content == tokenizer_config_json["unk_token"]:
token_type = SentencePieceTokenTypes.UNKNOWN
else:
token_type = SentencePieceTokenTypes.CONTROL
token_score = 0.0
logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
tokens[token_id] = token_content.encode("utf-8")
toktypes[token_id] = token_type
scores[token_id] = token_score
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_tokenizer_pre("default")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
_experts: list[dict[str, Tensor]] | None = None
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
if name.endswith("q_proj.weight"):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith("k_proj.weight"):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
# process the experts separately
if name.find("block_sparse_moe.experts") != -1:
n_experts = self.hparams["num_local_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for wid in ["w1", "w2", "w3"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
def write_tensors(self):
super().write_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("DeepseekV2ForCausalLM")
class DeepseekV2Model(Model):
model_arch = gguf.MODEL_ARCH.DEEPSEEK2

143
convert-persimmon-to-gguf.py
View file

@ -1,143 +0,0 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
from pathlib import Path
from pprint import pprint
import torch
from sentencepiece import SentencePieceProcessor
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
logger = logging.getLogger("persimmon-to-gguf")
def _flatten_dict(dct, tensors, prefix=None):
assert isinstance(dct, dict)
for key in dct.keys():
new_prefix = prefix + '.' + key if prefix is not None else key
if isinstance(dct[key], torch.Tensor):
tensors[new_prefix] = dct[key]
elif isinstance(dct[key], dict):
_flatten_dict(dct[key], tensors, new_prefix)
else:
raise ValueError(type(dct[key]))
return None
def _get_sentencepiece_tokenizer_info(dir_model: Path):
tokenizer_path = dir_model / 'adept_vocab.model'
logger.info('getting sentencepiece tokenizer from', tokenizer_path)
tokenizer = SentencePieceProcessor(str(tokenizer_path))
logger.info('adding tokens')
tokens: list[bytes] = []
scores: list[float] = []
toktypes: list[int] = []
for i in range(tokenizer.vocab_size()):
text: bytes
score: float
piece = tokenizer.id_to_piece(i)
text = piece.encode("utf-8")
score = tokenizer.get_score(i)
toktype = 1
if tokenizer.is_unknown(i):
toktype = 2
if tokenizer.is_control(i):
toktype = 3
if tokenizer.is_unused(i):
toktype = 5
if tokenizer.is_byte(i):
toktype = 6
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
pass
return tokens, scores, toktypes
def main():
parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("--ckpt-path", type=Path, help="path to persimmon checkpoint .pt file")
parser.add_argument("--model-dir", type=Path, help="directory containing model e.g. 8b_chat_model_release")
parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
sys.path.append(str(args.adept_inference_dir))
persimmon_model = torch.load(args.ckpt_path)
hparams = persimmon_model['args']
pprint(hparams)
tensors: dict[str, torch.Tensor] = {}
_flatten_dict(persimmon_model['model'], tensors, None)
arch = gguf.MODEL_ARCH.PERSIMMON
gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
block_count = hparams.num_layers
head_count = hparams.num_attention_heads
head_count_kv = head_count
ctx_length = hparams.seq_length
hidden_size = hparams.hidden_size
gguf_writer.add_name('persimmon-8b-chat')
gguf_writer.add_context_length(ctx_length)
gguf_writer.add_embedding_length(hidden_size)
gguf_writer.add_block_count(block_count)
gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
# ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
gguf_writer.add_head_count(head_count)
gguf_writer.add_head_count_kv(head_count_kv)
gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
gguf_writer.add_tokenizer_model('llama')
gguf_writer.add_tokenizer_pre('default')
gguf_writer.add_token_list(tokens)
gguf_writer.add_token_scores(scores)
gguf_writer.add_token_types(toktypes)
gguf_writer.add_bos_token_id(71013)
gguf_writer.add_eos_token_id(71013)
tensor_map = gguf.get_tensor_name_map(arch, block_count)
logger.info(tensor_map)
for name in tensors.keys():
data_torch = tensors[name]
if name.endswith(".self_attention.rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
data = data_torch.to(torch.float32).squeeze().numpy()
new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
if new_name is None:
raise ValueError(f"Can not map tensor '{name}'")
n_dims = len(data.shape)
logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}")
gguf_writer.add_tensor(new_name, data)
logger.info("gguf: write header")
gguf_writer.write_header_to_file()
logger.info("gguf: write metadata")
gguf_writer.write_kv_data_to_file()
logger.info("gguf: write tensors")
gguf_writer.write_tensors_to_file()
gguf_writer.close()
logger.info(f"gguf: model successfully exported to '{args.outfile}'")
if __name__ == '__main__':
main()

2
examples/batched/batched.cpp
View file

@ -48,7 +48,7 @@ int main(int argc, char ** argv) {
params.prompt = "Hello my name is";
}
process_escapes(params.prompt);
string_process_escapes(params.prompt);
// init LLM

4
examples/embedding/embedding.cpp
View file

@ -80,7 +80,7 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
params.prompt = string_random_prompt(rng);
}
llama_backend_init();
@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
}
// split the prompt into lines

4
examples/eval-callback/eval-callback.cpp
View file

@ -152,7 +152,7 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
params.prompt = string_random_prompt(rng);
}
llama_backend_init();
@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
}
bool OK = run(ctx, params);

7
examples/finetune/finetune.cpp
View file

@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
// not capturing these, to silcence warnings
const int rope_mode = 0;
return ggml_rope_custom(ctx,
t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
return ggml_rope_ext(ctx,
t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0,
rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
);
};
@ -643,7 +643,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
struct ggml_tensor * t16;
if (enable_flash_attn) {
t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
//t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
} else {
struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

4
examples/imatrix/imatrix.cpp
View file

@ -598,7 +598,7 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
params.prompt = string_random_prompt(rng);
}
sparams.dataset = params.prompt_file;
@ -667,7 +667,7 @@ int main(int argc, char ** argv) {
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
}
bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);

16
examples/infill/infill.cpp
View file

@ -50,9 +50,9 @@ static void write_logfile(
return;
}
const std::string timestamp = get_sortable_timestamp();
const std::string timestamp = string_get_sortable_timestamp();
const bool success = create_directory_with_parents(params.logdir);
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
@ -70,7 +70,7 @@ static void write_logfile(
fprintf(logfile, "binary: infill\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
@ -78,8 +78,8 @@ static void write_logfile(
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
dump_string_yaml_multiline(logfile, "output", output.c_str());
dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_dump_timing_info_yaml(logfile, ctx);
fclose(logfile);
@ -236,7 +236,7 @@ int main(int argc, char ** argv) {
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", get_system_info(params).c_str());
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
}
const bool add_bos = llama_should_add_bos_token(model);
GGML_ASSERT(llama_add_eos_token(model) != 1);
@ -621,8 +621,8 @@ int main(int argc, char ** argv) {
if (params.escape) {
//process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
process_escapes(params.input_prefix);
process_escapes(params.input_suffix);
string_process_escapes(params.input_prefix);
string_process_escapes(params.input_suffix);
}
suff_rm_leading_spc = params.escape;
if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {

4
examples/llama-bench/llama-bench.cpp
View file

@ -195,12 +195,12 @@ static const cmd_params cmd_params_defaults = {
/* model */ {"models/7B/ggml-model-q4_0.gguf"},
/* n_prompt */ {512},
/* n_gen */ {128},
/* n_pg */ {{512, 128}},
/* n_pg */ {},
/* n_batch */ {2048},
/* n_ubatch */ {512},
/* type_k */ {GGML_TYPE_F16},
/* type_v */ {GGML_TYPE_F16},
/* n_threads */ {get_math_cpu_count()},
/* n_threads */ {cpu_get_num_math()},
/* n_gpu_layers */ {99},
/* split_mode */ {LLAMA_SPLIT_MODE_LAYER},
/* main_gpu */ {0},

21
examples/llama.android/app/src/main/cpp/CMakeLists.txt
View file

@ -12,15 +12,20 @@ cmake_minimum_required(VERSION 3.22.1)
# build script scope).
project("llama-android")
include(FetchContent)
FetchContent_Declare(
llama
GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
GIT_TAG master
)
## Fetch latest llama.cpp from GitHub
#include(FetchContent)
#FetchContent_Declare(
# llama
# GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
# GIT_TAG master
#)
#
## Also provides "common"
#FetchContent_MakeAvailable(llama)
# Also provides "common"
FetchContent_MakeAvailable(llama)
# llama.cpp CI uses the code from the current branch
# ref: https://github.com/ggerganov/llama.cpp/pull/7341#issuecomment-2117617700
add_subdirectory(../../../../../../ build-llama)
# Creates and names a library, sets it as either STATIC
# or SHARED, and provides the relative paths to its source code.

2
examples/llava/llava-cli.cpp
View file

@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
#endif // LOG_DISABLE_LOGS
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
gpt_print_usage(argc, argv, params);
gpt_params_print_usage(argc, argv, params);
show_additional_info(argc, argv);
return 1;
}

2
examples/lookahead/lookahead.cpp
View file

@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
}
// build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/

2
examples/lookup/lookup.cpp
View file

@ -121,7 +121,7 @@ int main(int argc, char ** argv){
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
}
// print current draft sequence

2
examples/main/README.md
View file

@ -325,3 +325,5 @@ These options provide extra functionality and customization when running the LLa
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
- `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable or in an OS-specific local cache.

26
examples/main/main.cpp
View file

@ -60,9 +60,9 @@ static void write_logfile(
return;
}
const std::string timestamp = get_sortable_timestamp();
const std::string timestamp = string_get_sortable_timestamp();
const bool success = create_directory_with_parents(params.logdir);
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
@ -80,7 +80,7 @@ static void write_logfile(
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
@ -88,8 +88,8 @@ static void write_logfile(
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
dump_string_yaml_multiline(logfile, "output", output.c_str());
dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_dump_timing_info_yaml(logfile, ctx);
fclose(logfile);
@ -181,7 +181,7 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
params.prompt = string_random_prompt(rng);
}
LOG("%s: llama backend init\n", __func__);
@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", get_system_info(params).c_str());
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
}
std::string path_session = params.path_prompt_cache;
@ -474,12 +474,12 @@ int main(int argc, char ** argv) {
LOG_TEE("\n\n");
if (params.interactive) {
const char *control_message;
const char * control_message;
if (params.multiline_input) {
control_message = " - To return control to LLaMa, end your input with '\\'.\n"
control_message = " - To return control to the AI, end your input with '\\'.\n"
" - To return control without starting a new line, end your input with '/'.\n";
} else {
control_message = " - Press Return to return control to LLaMa.\n"
control_message = " - Press Return to return control to the AI.\n"
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
@ -707,7 +707,7 @@ int main(int argc, char ** argv) {
const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
llama_sampling_accept(ctx_sampling, ctx, id, true);
llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
@ -728,7 +728,7 @@ int main(int argc, char ** argv) {
// push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
@ -879,7 +879,7 @@ int main(int argc, char ** argv) {
embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
}
if (params.escape) {
process_escapes(buffer);
string_process_escapes(buffer);
}
const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);

2
examples/parallel/parallel.cpp
View file

@ -210,7 +210,7 @@ int main(int argc, char ** argv) {
while (true) {
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
}
llama_batch_clear(batch);

5
examples/perplexity/README.md
View file

@ -42,10 +42,13 @@ In addition to the KL divergence the following statistics are calculated with `-
Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
Note: the FP16 logits used for the calculation of all metrics other than perplexity are stored in a binary file between runs.
In order to save space this file does **not** contain the exact same FP32 logits but instead casts them to 16 bit unsigned integers (with some scaling).
So the "f16" results are to be understood as the difference resulting only from this downcast.
| Quantization | imatrix | Model size [GiB] | PPL | ΔPPL | KLD | Mean Δp | RMS Δp |
|--------------|---------|------------------|------------------------|------------------------|-----------------------|-------------------|------------------|
| f16 | None | 14.97 | 6.233160 ± 0.037828 | - | - | - | - |
| f16 | None | 14.97 | 6.233160 ± 0.037828 | 0.001524 ± 0.000755 | 0.000551 ± 0.000002 | 0.001 ± 0.002 % | 0.787 ± 0.004 % |
| q8_0 | None | 7.96 | 6.234284 ± 0.037878 | 0.002650 ± 0.001006 | 0.001355 ± 0.000006 | -0.019 ± 0.003 % | 1.198 ± 0.007 % |
| q6_K | None | 6.14 | 6.253382 ± 0.038078 | 0.021748 ± 0.001852 | 0.005452 ± 0.000035 | -0.007 ± 0.006 % | 2.295 ± 0.019 % |
| q5_K_M | None | 5.33 | 6.288607 ± 0.038338 | 0.056974 ± 0.002598 | 0.010762 ± 0.000079 | -0.114 ± 0.008 % | 3.160 ± 0.031 % |

18
examples/perplexity/perplexity.cpp
View file

@ -44,9 +44,9 @@ static void write_logfile(
return;
}
const std::string timestamp = get_sortable_timestamp();
const std::string timestamp = string_get_sortable_timestamp();
const bool success = create_directory_with_parents(params.logdir);
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
@ -64,7 +64,7 @@ static void write_logfile(
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc);
yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
@ -72,9 +72,9 @@ static void write_logfile(
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
dump_vector_float_yaml(logfile, "logits", results.logits);
yaml_dump_vector_float(logfile, "logits", results.logits);
fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
dump_vector_float_yaml(logfile, "probs", results.probs);
yaml_dump_vector_float(logfile, "probs", results.probs);
llama_dump_timing_info_yaml(logfile, ctx);
fclose(logfile);
@ -1425,7 +1425,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
// Use all tasks
tasks.resize(n_task);
printf("%s: reading tasks", __func__);
int n_dot = n_task/100;
int n_dot = std::max((int) n_task/100, 1);
int i = 0;
for (auto& task : tasks) {
++i;
@ -1675,7 +1675,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
llama_batch_free(batch);
if (n_done < 100) return;
if (n_done < 100 && (params.multiple_choice_tasks != 0 && params.multiple_choice_tasks < (size_t)n_task)) return;
float p = 1.f*n_correct/n_done;
float sigma = sqrt(p*(1-p)/(n_done-1));
@ -2007,7 +2007,7 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
params.prompt = string_random_prompt(rng);
}
llama_backend_init();
@ -2035,7 +2035,7 @@ int main(int argc, char ** argv) {
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
}
struct results_perplexity results;

4
examples/quantize/quantize.cpp
View file

@ -259,7 +259,7 @@ int main(int argc, char ** argv) {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
@ -284,7 +284,7 @@ int main(int argc, char ** argv) {
} else {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--keep-split")) {
} else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
params.keep_split = true;
} else {
usage(argv[0]);

6
examples/quantize/tests.sh
View file

@ -41,8 +41,8 @@ $SPLIT --split-max-tensors 28 $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/g
echo PASS
echo
# 3. Requant model with '--keep_split'
$QUANTIZE --allow-requantize --keep_split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
# 3. Requant model with '--keep-split'
$QUANTIZE --allow-requantize --keep-split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
echo PASS
echo
@ -51,7 +51,7 @@ $MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt
echo PASS
echo
# 4. Requant mode without '--keep_split'
# 4. Requant mode without '--keep-split'
$QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
echo PASS
echo

4
examples/retrieval/retrieval.cpp
View file

@ -11,7 +11,7 @@ struct retrieval_params {
};
static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
gpt_print_usage(argc, argv, gpt_params);
gpt_params_print_usage(argc, argv, gpt_params);
printf("retrieval options:\n");
printf(" --context-file FNAME file containing context to embed.\n");
printf(" specify multiple files by providing --context-file option multiple times.\n");
@ -226,7 +226,7 @@ int main(int argc, char ** argv) {
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
}
// max batch size

6
examples/server/README.md
View file

@ -18,8 +18,8 @@ The project is under active development, and we are [looking for feedback and co
**Command line options:**
- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
- `-t N`, `--threads N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores).
- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`.
- `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
@ -48,7 +48,7 @@ The project is under active development, and we are [looking for feedback and co
- `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
- `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error.
- `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching). Default: disabled
- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.

52
examples/server/public_simplechat/index.html Normal file
View file

@ -0,0 +1,52 @@
<!DOCTYPE html>
<html lang="en">
<head>
<title>SimpleChat (LlamaCPP, ...) </title>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta name="message" content="Save Nature Save Earth" />
<meta name="description" content="SimpleChat: trigger LLM web service endpoints /chat/completions and /completions, single/multi chat sessions" />
<meta name="author" content="by Humans for All" />
<meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate" />
<script src="simplechat.js" defer></script>
<link rel="stylesheet" href="simplechat.css" />
</head>
<body>
<div class="samecolumn" id="fullbody">
<div class="sameline">
<p class="heading flex-grow" > <b> SimpleChat </b> </p>
<div class="sameline">
<label for="api-ep">Mode:</label>
<select name="api-ep" id="api-ep">
<option value="chat" selected>Chat</option>
<option value="completion">Completion</option>
</select>
</div>
</div>
<div id="sessions-div" class="sameline"></div>
<hr>
<div class="sameline">
<label for="system-in">System</label>
<input type="text" name="system" id="system-in" class="flex-grow"/>
</div>
<hr>
<div id="chat-div">
<p> Enter the system prompt above, before entering/submitting any user query.</p>
<p> Enter your text to the ai assistant below.</p>
<p> Use shift+enter for inserting enter.</p>
<p> Refresh the page to start over fresh.</p>
</div>
<hr>
<div class="sameline">
<textarea id="user-in" class="flex-grow" rows="3"></textarea>
<button id="user-btn">submit</button>
</div>
</div>
</body>
</html>

81
examples/server/public_simplechat/readme.md Normal file
View file

@ -0,0 +1,81 @@
# SimpleChat
by Humans for All.
## overview
This simple web frontend, allows triggering/testing the server's /completions or /chat/completions endpoints
in a simple way with minimal code from a common code base. Inturn additionally it tries to allow single or
multiple independent back and forth chatting to an extent, with the ai llm model at a basic level, with their
own system prompts.
The UI follows a responsive web design so that the layout can adapt to available display space in a usable
enough manner, in general.
NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
culling of old messages from the chat.
NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
the js file as needed.
## usage
One could run this web frontend directly using server itself or if anyone is thinking of adding a built in web
frontend to configure the server over http(s) or so, then run this web frontend using something like python's
http module.
### running using examples/server
bin/server -m path/model.gguf --path ../examples/server/public_simplechat [--port PORT]
### running using python3's server module
first run examples/server
* bin/server -m path/model.gguf
next run this web front end in examples/server/public_simplechat
* cd ../examples/server/public_simplechat
* python3 -m http.server PORT
### using the front end
Open this simple web front end from your local browser
* http://127.0.0.1:PORT/index.html
Once inside
* Select between chat and completion mode. By default it is set to chat mode.
* If you want to provide a system prompt, then ideally enter it first, before entering any user query.
* if chat.add_system_begin is used
* you cant change the system prompt, after it is has been submitted once along with user query.
* you cant set a system prompt, after you have submitted any user query
* if chat.add_system_anytime is used
* one can change the system prompt any time during chat, by changing the contents of system prompt.
* inturn the updated/changed system prompt will be inserted into the chat session.
* this allows for the subsequent user chatting to be driven by the new system prompt set above.
* Enter your query and either press enter or click on the submit button.
If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
* Wait for the logic to communicate with the server and get the response.
* the user is not allowed to enter any fresh query during this time.
* the user input box will be disabled and a working message will be shown in it.
* just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
* Using NewChat one can start independent chat sessions.
* two independent chat sessions are setup by default.
## Devel note
Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
may not be visible. Also remember that just refreshing/reloading page in browser or for that
matter clearing site data, dont directly override site caching in all cases. Worst case you may
have to change port. Or in dev tools of browser, you may be able to disable caching fully.
Concept of multiple chat sessions with different servers, as well as saving and restoring of
those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
its instances relatively easily, however given the current goal of keeping this simple, it has
not been added, for now.
By switching between chat.add_system_begin/anytime, one can control whether one can change
the system prompt, anytime during the conversation or only at the beginning.

61
examples/server/public_simplechat/simplechat.css Normal file
View file

@ -0,0 +1,61 @@
/**
* the styling of the simplechat web frontend
* by Humans for All
*/
#fullbody {
height: 98vh;
}
.heading {
background-color: lightgray;
}
.session-selected {
background-color: lightblue;
}
.role-system {
background-color: lightblue;
}
.role-user {
background-color: lightgray;
}
.flex-grow {
flex-grow: 1;
}
.float-right {
float: right;
}
#chat-div {
overflow: scroll;
flex-grow: 1;
flex-shrink: 1;
min-height: 40vh;
}
button {
min-width: 8vw;
}
.sameline {
display: flex;
flex-direction: row;
}
.samecolumn {
display: flex;
flex-direction: column;
}
* {
margin: 0.6vmin;
}
@media print {
#fullbody {
height: auto;
}
}

478
examples/server/public_simplechat/simplechat.js Normal file
View file

@ -0,0 +1,478 @@
// @ts-check
// A simple completions and chat/completions test related web front end logic
// by Humans for All
class Roles {
static System = "system";
static User = "user";
static Assistant = "assistant";
}
class ApiEP {
static Chat = "chat";
static Completion = "completion";
}
let gUsageMsg = `
<p> Enter the system prompt above, before entering/submitting any user query.</p>
<p> Enter your text to the ai assistant below.</p>
<p> Use shift+enter for inserting enter.</p>
<p> Refresh the page to start over fresh.</p>
`;
class SimpleChat {
constructor() {
/**
* Maintain in a form suitable for common LLM web service chat/completions' messages entry
* @type {{role: string, content: string}[]}
*/
this.xchat = [];
this.iLastSys = -1;
}
/**
* Add an entry into xchat
* @param {string} role
* @param {string|undefined|null} content
*/
add(role, content) {
if ((content == undefined) || (content == null) || (content == "")) {
return false;
}
this.xchat.push( {role: role, content: content} );
if (role == Roles.System) {
this.iLastSys = this.xchat.length - 1;
}
return true;
}
/**
* Show the contents in the specified div
* @param {HTMLDivElement} div
* @param {boolean} bClear
*/
show(div, bClear=true) {
if (bClear) {
div.replaceChildren();
}
let last = undefined;
for(const x of this.xchat) {
let entry = document.createElement("p");
entry.className = `role-${x.role}`;
entry.innerText = `${x.role}: ${x.content}`;
div.appendChild(entry);
last = entry;
}
if (last !== undefined) {
last.scrollIntoView(false);
} else {
if (bClear) {
div.innerHTML = gUsageMsg;
}
}
}
/**
* Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
* Convert the json into string.
* @param {Object} obj
*/
request_jsonstr(obj) {
obj["temperature"] = 0.7;
return JSON.stringify(obj);
}
/**
* Return a string form of json object suitable for chat/completions
*/
request_messages_jsonstr() {
let req = {
messages: this.xchat,
}
return this.request_jsonstr(req);
}
/**
* Return a string form of json object suitable for /completions
*/
request_prompt_jsonstr() {
let prompt = "";
for(const chat of this.xchat) {
prompt += `${chat.role}: ${chat.content}\n`;
}
let req = {
prompt: prompt,
}
return this.request_jsonstr(req);
}
/**
* Allow setting of system prompt, but only at begining.
* @param {string} sysPrompt
* @param {string} msgTag
*/
add_system_begin(sysPrompt, msgTag) {
if (this.xchat.length == 0) {
if (sysPrompt.length > 0) {
return this.add(Roles.System, sysPrompt);
}
} else {
if (sysPrompt.length > 0) {
if (this.xchat[0].role !== Roles.System) {
console.error(`ERRR:SimpleChat:SC:${msgTag}:You need to specify system prompt before any user query, ignoring...`);
} else {
if (this.xchat[0].content !== sysPrompt) {
console.error(`ERRR:SimpleChat:SC:${msgTag}:You cant change system prompt, mid way through, ignoring...`);
}
}
}
}
return false;
}
/**
* Allow setting of system prompt, at any time.
* @param {string} sysPrompt
* @param {string} msgTag
*/
add_system_anytime(sysPrompt, msgTag) {
if (sysPrompt.length <= 0) {
return false;
}
if (this.iLastSys < 0) {
return this.add(Roles.System, sysPrompt);
}
let lastSys = this.xchat[this.iLastSys].content;
if (lastSys !== sysPrompt) {
return this.add(Roles.System, sysPrompt);
}
return false;
}
/**
* Retrieve the latest system prompt.
*/
get_system_latest() {
if (this.iLastSys == -1) {
return "";
}
let sysPrompt = this.xchat[this.iLastSys].content;
return sysPrompt;
}
}
let gBaseURL = "http://127.0.0.1:8080";
let gChatURL = {
'chat': `${gBaseURL}/chat/completions`,
'completion': `${gBaseURL}/completions`,
}
const gbCompletionFreshChatAlways = true;
/**
* Set the class of the children, based on whether it is the idSelected or not.
* @param {HTMLDivElement} elBase
* @param {string} idSelected
* @param {string} classSelected
* @param {string} classUnSelected
*/
function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
for(let child of elBase.children) {
if (child.id == idSelected) {
child.className = classSelected;
} else {
child.className = classUnSelected;
}
}
}
/**
* Create button and set it up.
* @param {string} id
* @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
* @param {string | undefined} name
* @param {string | undefined} innerText
*/
function el_create_button(id, callback, name=undefined, innerText=undefined) {
if (!name) {
name = id;
}
if (!innerText) {
innerText = id;
}
let btn = document.createElement("button");
btn.id = id;
btn.name = name;
btn.innerText = innerText;
btn.addEventListener("click", callback);
return btn;
}
class MultiChatUI {
constructor() {
/** @type {Object<string, SimpleChat>} */
this.simpleChats = {};
/** @type {string} */
this.curChatId = "";
// the ui elements
this.elInSystem = /** @type{HTMLInputElement} */(document.getElementById("system-in"));
this.elDivChat = /** @type{HTMLDivElement} */(document.getElementById("chat-div"));
this.elBtnUser = /** @type{HTMLButtonElement} */(document.getElementById("user-btn"));
this.elInUser = /** @type{HTMLInputElement} */(document.getElementById("user-in"));
this.elSelectApiEP = /** @type{HTMLSelectElement} */(document.getElementById("api-ep"));
this.elDivSessions = /** @type{HTMLDivElement} */(document.getElementById("sessions-div"));
this.validate_element(this.elInSystem, "system-in");
this.validate_element(this.elDivChat, "chat-div");
this.validate_element(this.elInUser, "user-in");
this.validate_element(this.elSelectApiEP, "api-ep");
this.validate_element(this.elDivChat, "sessions-div");
}
/**
* Check if the element got
* @param {HTMLElement | null} el
* @param {string} msgTag
*/
validate_element(el, msgTag) {
if (el == null) {
throw Error(`ERRR:SimpleChat:MCUI:${msgTag} element missing in html...`);
} else {
console.debug(`INFO:SimpleChat:MCUI:${msgTag} Id[${el.id}] Name[${el["name"]}]`);
}
}
/**
* Reset user input ui.
* * clear user input
* * enable user input
* * set focus to user input
*/
ui_reset_userinput() {
this.elInUser.value = "";
this.elInUser.disabled = false;
this.elInUser.focus();
}
/**
* Setup the needed callbacks wrt UI, curChatId to defaultChatId and
* optionally switch to specified defaultChatId.
* @param {string} defaultChatId
* @param {boolean} bSwitchSession
*/
setup_ui(defaultChatId, bSwitchSession=false) {
this.curChatId = defaultChatId;
if (bSwitchSession) {
this.handle_session_switch(this.curChatId);
}
this.elBtnUser.addEventListener("click", (ev)=>{
if (this.elInUser.disabled) {
return;
}
this.handle_user_submit(this.curChatId, this.elSelectApiEP.value).catch((/** @type{Error} */reason)=>{
let msg = `ERRR:SimpleChat\nMCUI:HandleUserSubmit:${this.curChatId}\n${reason.name}:${reason.message}`;
console.debug(msg.replace("\n", ":"));
alert(msg);
this.ui_reset_userinput();
});
});
this.elInUser.addEventListener("keyup", (ev)=> {
// allow user to insert enter into their message using shift+enter.
// while just pressing enter key will lead to submitting.
if ((ev.key === "Enter") && (!ev.shiftKey)) {
this.elBtnUser.click();
ev.preventDefault();
}
});
this.elInSystem.addEventListener("keyup", (ev)=> {
// allow user to insert enter into the system prompt using shift+enter.
// while just pressing enter key will lead to setting the system prompt.
if ((ev.key === "Enter") && (!ev.shiftKey)) {
let chat = this.simpleChats[this.curChatId];
chat.add_system_anytime(this.elInSystem.value, this.curChatId);
chat.show(this.elDivChat);
ev.preventDefault();
}
});
}
/**
* Setup a new chat session and optionally switch to it.
* @param {string} chatId
* @param {boolean} bSwitchSession
*/
new_chat_session(chatId, bSwitchSession=false) {
this.simpleChats[chatId] = new SimpleChat();
if (bSwitchSession) {
this.handle_session_switch(chatId);
}
}
/**
* Handle user query submit request, wrt specified chat session.
* @param {string} chatId
* @param {string} apiEP
*/
async handle_user_submit(chatId, apiEP) {
let chat = this.simpleChats[chatId];
chat.add_system_anytime(this.elInSystem.value, chatId);
let content = this.elInUser.value;
if (!chat.add(Roles.User, content)) {
console.debug(`WARN:SimpleChat:MCUI:${chatId}:HandleUserSubmit:Ignoring empty user input...`);
return;
}
chat.show(this.elDivChat);
let theBody;
let theUrl = gChatURL[apiEP]
if (apiEP == ApiEP.Chat) {
theBody = chat.request_messages_jsonstr();
} else {
theBody = chat.request_prompt_jsonstr();
}
this.elInUser.value = "working...";
this.elInUser.disabled = true;
console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:${theUrl}:ReqBody:${theBody}`);
let resp = await fetch(theUrl, {
method: "POST",
headers: {
"Content-Type": "application/json",
},
body: theBody,
});
let respBody = await resp.json();
console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
let assistantMsg;
if (apiEP == ApiEP.Chat) {
assistantMsg = respBody["choices"][0]["message"]["content"];
} else {
try {
assistantMsg = respBody["choices"][0]["text"];
} catch {
assistantMsg = respBody["content"];
}
}
chat.add(Roles.Assistant, assistantMsg);
if (chatId == this.curChatId) {
chat.show(this.elDivChat);
} else {
console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
}
// Purposefully clear at end rather than begin of this function
// so that one can switch from chat to completion mode and sequece
// in a completion mode with multiple user-assistant chat data
// from before to be sent/occur once.
if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
chat.xchat.length = 0;
}
this.ui_reset_userinput();
}
/**
* Show buttons for NewChat and available chat sessions, in the passed elDiv.
* If elDiv is undefined/null, then use this.elDivSessions.
* Take care of highlighting the selected chat-session's btn.
* @param {HTMLDivElement | undefined} elDiv
*/
show_sessions(elDiv=undefined) {
if (!elDiv) {
elDiv = this.elDivSessions;
}
elDiv.replaceChildren();
// Btn for creating new chat session
let btnNew = el_create_button("New CHAT", (ev)=> {
if (this.elInUser.disabled) {
console.error(`ERRR:SimpleChat:MCUI:NewChat:Current session [${this.curChatId}] awaiting response, ignoring request...`);
alert("ERRR:SimpleChat\nMCUI:NewChat\nWait for response to pending query, before starting new chat session");
return;
}
let chatId = `Chat${Object.keys(this.simpleChats).length}`;
let chatIdGot = prompt("INFO:SimpleChat\nMCUI:NewChat\nEnter id for new chat session", chatId);
if (!chatIdGot) {
console.error("ERRR:SimpleChat:MCUI:NewChat:Skipping based on user request...");
return;
}
this.new_chat_session(chatIdGot, true);
this.create_session_btn(elDiv, chatIdGot);
el_children_config_class(elDiv, chatIdGot, "session-selected", "");
});
elDiv.appendChild(btnNew);
// Btns for existing chat sessions
let chatIds = Object.keys(this.simpleChats);
for(let cid of chatIds) {
let btn = this.create_session_btn(elDiv, cid);
if (cid == this.curChatId) {
btn.className = "session-selected";
}
}
}
create_session_btn(elDiv, cid) {
let btn = el_create_button(cid, (ev)=>{
let target = /** @type{HTMLButtonElement} */(ev.target);
console.debug(`DBUG:SimpleChat:MCUI:SessionClick:${target.id}`);
if (this.elInUser.disabled) {
console.error(`ERRR:SimpleChat:MCUI:SessionClick:${target.id}:Current session [${this.curChatId}] awaiting response, ignoring switch...`);
alert("ERRR:SimpleChat\nMCUI:SessionClick\nWait for response to pending query, before switching");
return;
}
this.handle_session_switch(target.id);
el_children_config_class(elDiv, target.id, "session-selected", "");
});
elDiv.appendChild(btn);
return btn;
}
/**
* Switch ui to the specified chatId and set curChatId to same.
* @param {string} chatId
*/
async handle_session_switch(chatId) {
let chat = this.simpleChats[chatId];
if (chat == undefined) {
console.error(`ERRR:SimpleChat:MCUI:HandleSessionSwitch:${chatId} missing...`);
return;
}
this.elInSystem.value = chat.get_system_latest();
this.elInUser.value = "";
chat.show(this.elDivChat);
this.elInUser.focus();
this.curChatId = chatId;
console.log(`INFO:SimpleChat:MCUI:HandleSessionSwitch:${chatId} entered...`);
}
}
let gMuitChat;
const gChatIds = [ "Default", "Other" ];
function startme() {
console.log("INFO:SimpleChat:StartMe:Starting...");
gMuitChat = new MultiChatUI();
for (let cid of gChatIds) {
gMuitChat.new_chat_session(cid);
}
gMuitChat.setup_ui(gChatIds[0]);
gMuitChat.show_sessions();
}
document.addEventListener("DOMContentLoaded", startme);

16
examples/server/server.cpp
View file

@ -102,7 +102,6 @@ struct slot_params {
bool stream = true;
bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
uint32_t seed = -1; // RNG seed
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_discard = 0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
int32_t n_predict = -1; // new tokens to predict
@ -1020,7 +1019,7 @@ struct server_context {
sampler_names.emplace_back(sampler_name);
}
}
slot.sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
slot.sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
} else {
slot.sparams.samplers_sequence = default_sparams.samplers_sequence;
}
@ -1257,14 +1256,14 @@ struct server_context {
std::vector<std::string> samplers_sequence;
samplers_sequence.reserve(slot.sparams.samplers_sequence.size());
for (const auto & sampler_type : slot.sparams.samplers_sequence) {
samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
}
return json {
{"n_ctx", slot.n_ctx},
{"n_predict", slot.n_predict},
{"model", params.model_alias},
{"seed", slot.params.seed},
{"seed", slot.sparams.seed},
{"temperature", slot.sparams.temp},
{"dynatemp_range", slot.sparams.dynatemp_range},
{"dynatemp_exponent", slot.sparams.dynatemp_exponent},
@ -1982,8 +1981,7 @@ struct server_context {
slot.state = SLOT_STATE_PROCESSING;
slot.command = SLOT_COMMAND_NONE;
slot.release();
slot.print_timings();
send_final_response(slot);
send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
continue;
}
} else {
@ -2854,7 +2852,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
invalid_param = true;
break;
}
if (!parse_kv_override(argv[i], params.kv_overrides)) {
if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
invalid_param = true;
break;
@ -3312,7 +3310,7 @@ int main(int argc, char ** argv) {
const auto handle_slots_save = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
json request_data = json::parse(req.body);
std::string filename = request_data.at("filename");
if (!validate_file_name(filename)) {
if (!fs_validate_filename(filename)) {
res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
return;
}
@ -3342,7 +3340,7 @@ int main(int argc, char ** argv) {
const auto handle_slots_restore = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
json request_data = json::parse(req.body);
std::string filename = request_data.at("filename");
if (!validate_file_name(filename)) {
if (!fs_validate_filename(filename)) {
res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
return;
}

55
examples/server/tests/features/results.feature
View file

@ -13,6 +13,7 @@ Feature: Results
Scenario Outline: consistent results with same seed
Given <n_slots> slots
And 1.0 temperature
Then the server is starting
Then the server is healthy
@ -26,10 +27,12 @@ Feature: Results
Examples:
| n_slots |
| 1 |
| 2 |
# FIXME: unified KV cache nondeterminism
# | 2 |
Scenario Outline: different results with different seed
Given <n_slots> slots
And 1.0 temperature
Then the server is starting
Then the server is healthy
@ -70,12 +73,46 @@ Feature: Results
Then all predictions are equal
Examples:
| n_parallel | temp |
| 1 | 0.0 |
| 2 | 0.0 |
| 4 | 0.0 |
| 1 | 1.0 |
# FIXME: These tests fail on master. The problem seems to be the unified KV cache.
| 1 | 0.0 |
| 1 | 1.0 |
# FIXME: unified KV cache nondeterminism
# See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
# and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
# | 2 | 1.0 |
# | 4 | 1.0 |
# and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
# and https://github.com/ggerganov/llama.cpp/pull/7347 .
# | 2 | 0.0 |
# | 4 | 0.0 |
# | 2 | 1.0 |
# | 4 | 1.0 |
Scenario Outline: consistent token probs with same seed and prompt
Given <n_slots> slots
And <n_kv> KV cache size
And 1.0 temperature
And <n_predict> max tokens to predict
Then the server is starting
Then the server is healthy
Given 1 prompts "The meaning of life is" with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Given <n_parallel> prompts "The meaning of life is" with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Then all token probabilities are equal
Examples:
| n_slots | n_kv | n_predict | n_parallel |
| 4 | 1024 | 1 | 1 |
# FIXME: unified KV cache nondeterminism
# See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
# and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
# and https://github.com/ggerganov/llama.cpp/pull/7347 .
# | 4 | 1024 | 1 | 4 |
# | 4 | 1024 | 100 | 1 |
# This test still fails even the above patches; the first token probabilities are already different.
# | 4 | 1024 | 100 | 4 |

41
examples/server/tests/features/steps/steps.py
View file

@ -23,6 +23,7 @@ from prometheus_client import parser
def step_server_config(context, server_fqdn, server_port):
context.server_fqdn = server_fqdn
context.server_port = int(server_port)
context.n_threads = None
context.n_gpu_layer = None
if 'PORT' in os.environ:
context.server_port = int(os.environ['PORT'])
@ -109,6 +110,11 @@ def step_n_gpu_layer(context, ngl):
context.n_gpu_layer = ngl
@step('{n_threads:d} threads')
def step_n_threads(context, n_threads):
context.n_thread = n_threads
@step('{draft:d} as draft')
def step_draft(context, draft):
context.draft = draft
@ -193,7 +199,7 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status):
case 'ready' | 'idle':
await wait_for_health_status(context, context.base_url, 200, 'ok',
timeout=10,
timeout=30,
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=context.n_slots,
slots_processing=0,
@ -274,13 +280,22 @@ async def step_predictions_equal(context):
@step('all predictions are different')
@async_run_until_complete
async def step_predictions_equal(context):
async def step_predictions_different(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_predictions_different(context.tasks_result)
context.tasks_result = []
@step('all token probabilities are equal')
@async_run_until_complete
async def step_token_probabilities_equal(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_token_probabilities_equal(context.tasks_result)
context.tasks_result = []
@step('the completion is truncated')
def step_assert_completion_truncated(context):
step_assert_completion_truncated(context, '')
@ -868,7 +883,8 @@ async def request_completion(prompt,
"cache_prompt": cache_prompt,
"id_slot": id_slot,
"seed": seed if seed is not None else 42,
"temperature": temperature if temperature is not None else "0.8f",
"temperature": temperature if temperature is not None else 0.8,
"n_probs": 2,
},
headers=headers,
timeout=3600) as response:
@ -1123,6 +1139,23 @@ def assert_all_predictions_different(completion_responses):
assert content_i != content_j, "contents not different"
def assert_all_token_probabilities_equal(completion_responses):
n_predict = len(completion_responses[0]['completion_probabilities'])
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
for pos in range(n_predict):
for i, response_i in enumerate(completion_responses):
probs_i = response_i['completion_probabilities'][pos]['probs']
print(f"pos {pos}, probs {i}: {probs_i}")
for pos in range(n_predict):
for i, response_i in enumerate(completion_responses):
probs_i = response_i['completion_probabilities'][pos]['probs']
for j, response_j in enumerate(completion_responses):
if i == j:
continue
probs_j = response_j['completion_probabilities'][pos]['probs']
assert probs_i == probs_j, "contents not equal"
async def gather_tasks_results(context):
n_tasks = len(context.concurrent_tasks)
if context.debug:
@ -1261,6 +1294,8 @@ def start_server_background(context):
server_args.extend(['--batch-size', context.n_batch])
if context.n_ubatch:
server_args.extend(['--ubatch-size', context.n_ubatch])
if context.n_threads:
server_args.extend(['--threads', context.threads])
if context.n_gpu_layer:
server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
if context.draft is not None:

4
examples/sycl/win-build-sycl.bat
View file

@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
:: for FP16
:: faster for long-prompt inference
:: cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
:: cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
:: for FP32
cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release
cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
if %errorlevel% neq 0 goto ERROR
:: build example/main only
:: make main

7
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs(
// not capturing these, to silcence warnings
const int rope_mode = 0;
return ggml_rope_custom(
ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
return ggml_rope_ext(
ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
);
};
@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
struct ggml_tensor * t16;
if (enable_flash_attn) {
t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
//t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
} else {
struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

54
ggml-common.h
View file

@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
// QK = number of values after dequantization
// QK_K = super-block size
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif // GGML_QKK_64
#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
// QR = QK / number of values before dequantization
@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2;
#define QI4_NL (QK4_NL / (4*QR4_NL))
#define QR4_NL 2
#if QK_K == 64
#define QI4_XS QI4_NL
#define QR4_XS QR4_NL
#else
#define QI4_XS (QK_K / (4*QR4_XS))
#define QR4_XS 8
#endif
#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 3.4375 bits per weight
#ifdef GGML_QKK_64
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[2];
ggml_half d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
#else
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
@ -244,20 +225,11 @@ typedef struct {
ggml_half d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
#endif
// 4-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 4.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_half d[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
union {
struct {
@ -270,21 +242,11 @@ typedef struct {
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
#endif
// 5-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 5.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_half d; // super-block scale
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
union {
struct {
@ -298,7 +260,6 @@ typedef struct {
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
// 6-bit quantization
// weight is represented as x = a * q
@ -356,11 +317,7 @@ typedef struct {
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
// 3.4375 bpw
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
ggml_half d;
uint8_t qs[QK_K/4];
@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
typedef struct {
uint8_t qs[QK_K/8]; // grid index, low 8 bits
uint8_t qh[QK_K/16]; // grid index, high 3 bits + grid shift bit (for two groups of 8)
#if QK_K == 64
ggml_half d;
#endif
uint8_t scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
} block_iq1_m;
#if QK_K == 64
static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
#else
static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
#endif
// Used by IQ1_M quants
typedef union {
@ -406,9 +356,6 @@ typedef struct {
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
#else
typedef struct {
ggml_half d;
uint16_t scales_h;
@ -416,7 +363,6 @@ typedef struct {
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif
#endif // GGML_COMMON_DECL
#endif // GGML_COMMON_DECL

102
ggml-cuda.cu
View file

@ -43,19 +43,59 @@
#include <mutex>
#include <stdint.h>
#include <stdio.h>
#include <stdarg.h>
#include <stdlib.h>
#include <string>
#include <vector>
static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
GGML_UNUSED(level);
GGML_UNUSED(user_data);
fprintf(stderr, "%s", msg);
}
ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback;
void * ggml_cuda_log_user_data = NULL;
GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) {
ggml_cuda_log_callback = log_callback;
ggml_cuda_log_user_data = user_data;
}
#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
GGML_ATTRIBUTE_FORMAT(2, 3)
static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) {
if (ggml_cuda_log_callback != NULL) {
va_list args;
va_start(args, format);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data);
} else {
std::vector<char> buffer2(len + 1); // vsnprintf adds a null terminator
va_end(args);
va_start(args, format);
vsnprintf(&buffer2[0], buffer2.size(), format, args);
ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data);
}
va_end(args);
}
}
[[noreturn]]
void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
int id = -1; // in case cudaGetDevice fails
cudaGetDevice(&id);
fprintf(stderr, "CUDA error: %s\n", msg);
fprintf(stderr, " current device: %d, in function %s at %s:%d\n", id, func, file, line);
fprintf(stderr, " %s\n", stmt);
GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
GGML_CUDA_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, file, line);
GGML_CUDA_LOG_ERROR(" %s\n", stmt);
// abort with GGML_ASSERT to get a stack trace
GGML_ASSERT(!"CUDA error");
}
@ -91,7 +131,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
cudaError_t err = cudaGetDeviceCount(&info.device_count);
if (err != cudaSuccess) {
fprintf(stderr, "%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
return info;
}
@ -99,16 +139,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
int64_t total_vram = 0;
#if defined(GGML_CUDA_FORCE_MMQ)
fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__);
GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__);
#else
fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: no\n", __func__);
GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: no\n", __func__);
#endif
#if defined(CUDA_USE_TENSOR_CORES)
fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
#else
fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
#endif
fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
for (int id = 0; id < info.device_count; ++id) {
int device_vmm = 0;
@ -129,7 +169,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
cudaDeviceProp prop;
CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
fprintf(stderr, " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
GGML_CUDA_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
info.default_tensor_split[id] = total_vram;
total_vram += prop.totalGlobalMem;
@ -235,8 +275,8 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
*actual_size = look_ahead_size;
pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
(uint32_t)(max_size/1024/1024), (uint32_t)(pool_size/1024/1024), (uint32_t)(size/1024/1024));
GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
(uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
#endif
return ptr;
}
@ -250,7 +290,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
return;
}
}
fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
ggml_cuda_set_device(device);
CUDA_CHECK(cudaFree(ptr));
pool_size -= size;
@ -499,7 +539,9 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
void * dev_ptr;
cudaError_t err = cudaMalloc(&dev_ptr, size);
if (err != cudaSuccess) {
fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err));
// clear the error
cudaGetLastError();
GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
return nullptr;
}
@ -1002,8 +1044,8 @@ static void * ggml_cuda_host_malloc(size_t size) {
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size/1024.0/1024.0, cudaGetErrorString(err));
GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return nullptr;
}
@ -2246,7 +2288,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
break;
case GGML_OP_MUL_MAT:
if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
return false;
} else {
ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
@ -2300,7 +2342,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst));
GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
CUDA_CHECK(err);
}
@ -2476,7 +2518,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to GPU architecture\n", __func__);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
}
}
@ -2523,14 +2565,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to split buffer\n", __func__);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
}
@ -2539,7 +2581,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
// Changes in batch size or context size can cause changes to the grid size of some kernels.
use_cuda_graph = false;
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
}
@ -2567,7 +2609,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
}
}
@ -2605,7 +2647,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
if (!ok) {
fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
}
GGML_ASSERT(ok);
}
@ -2624,7 +2666,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
use_cuda_graph = false;
cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
#ifndef NDEBUG
fprintf(stderr, "%s: disabling CUDA graphs due to failed graph capture\n", __func__);
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
#endif
} else {
graph_evaluated_or_captured = true; // CUDA graph has been captured
@ -2691,7 +2733,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
if (stat == cudaErrorGraphExecUpdateFailure) {
#ifndef NDEBUG
fprintf(stderr, "%s: CUDA graph update failed\n", __func__);
GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
#endif
// The pre-existing graph exec cannot be updated due to violated constraints
// so instead clear error and re-instantiate
@ -2948,13 +2990,13 @@ static ggml_guid_t ggml_backend_cuda_guid() {
GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
return nullptr;
}
ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
if (ctx == nullptr) {
fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__);
return nullptr;
}
@ -2998,8 +3040,8 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
// clear the error
cudaGetLastError();
fprintf(stderr, "%s: warning: failed to register %.2f MiB of pinned memory: %s\n", __func__,
size/1024.0/1024.0, cudaGetErrorString(err));
GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return false;
}
return true;

1
ggml-cuda.h
View file

@ -38,6 +38,7 @@ GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t *
GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
#ifdef __cplusplus
}
#endif

25
ggml-cuda/common.cuh
View file

@ -315,6 +315,20 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
#endif
return c;
}
#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
// __shfl_xor() for half2 was added in ROCm 5.6
static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
typedef union half2_b32 {
half2 val;
int b32;
} half2_b32_t;
half2_b32_t tmp;
tmp.val = var;
tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
return tmp.val;
}
#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
#endif // defined(GGML_USE_HIPBLAS)
#define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
@ -463,6 +477,17 @@ static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -
typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
static __device__ __forceinline__ float get_alibi_slope(
const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
) {
if (max_bias <= 0.0f) {
return 1.0f;
}
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
return powf(base, exph);
}
//////////////////////

138
ggml-cuda/convert.cu
View file

@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
const block_q2_K * x = (const block_q2_K *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t n = tid/32;
const int64_t l = tid - 32*n;
const int64_t is = 8*n + l/16;
@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
#else
const int64_t is = tid/16; // 0 or 1
const int64_t il = tid%16; // 0...15
const uint8_t q = x[i].qs[il] >> (2*is);
dst_t * y = yy + i*QK_K + 16*is + il;
float dall = __low2half(x[i].dm);
float dmin = __high2half(x[i].dm);
y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
#endif
}
template<typename dst_t>
@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
const int64_t i = blockIdx.x;
const block_q3_K * x = (const block_q3_K *) vx;
#if QK_K == 256
const int64_t r = threadIdx.x/4;
const int64_t tid = r/2;
const int64_t is0 = r%2;
@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
const uint8_t * hm = x[i].hmask;
for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
#else
const int64_t tid = threadIdx.x;
const int64_t is = tid/16; // 0 or 1
const int64_t il = tid%16; // 0...15
const int64_t im = il/8; // 0...1
const int64_t in = il%8; // 0...7
dst_t * y = yy + i*QK_K + 16*is + il;
const uint8_t q = x[i].qs[il] >> (2*is);
const uint8_t h = x[i].hmask[in] >> (2*is + im);
const float d = (float)x[i].d;
if (is == 0) {
y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
} else {
y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
}
#endif
}
#if QK_K == 256
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
if (j < 4) {
d = q[j] & 63; m = q[j + 4] & 63;
@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
}
}
#endif
template<typename dst_t>
static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
const int64_t i = blockIdx.x;
#if QK_K == 256
// assume 32 threads
const int64_t tid = threadIdx.x;
const int64_t il = tid/8;
@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
y[l + 0] = d1 * (q[l] & 0xF) - m1;
y[l +32] = d2 * (q[l] >> 4) - m2;
}
#else
const int64_t tid = threadIdx.x;
const uint8_t * q = x[i].qs;
dst_t * y = yy + i*QK_K;
const float d = (float)x[i].dm[0];
const float m = (float)x[i].dm[1];
y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4);
#endif
}
template<typename dst_t>
@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
const int64_t i = blockIdx.x;
#if QK_K == 256
// assume 64 threads - this is very slightly better than the one below
const int64_t tid = threadIdx.x;
const int64_t il = tid/16; // il is in 0...3
@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
hm <<= 1;
y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2;
y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2;
#else
const int64_t tid = threadIdx.x;
const uint8_t q = x[i].qs[tid];
const int64_t im = tid/8; // 0...3
const int64_t in = tid%8; // 0...7
const int64_t is = tid/16; // 0 or 1
const uint8_t h = x[i].qh[in] >> im;
const float d = x[i].d;
dst_t * y = yy + i*QK_K + tid;
y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16));
#endif
}
template<typename dst_t>
@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
const block_q6_K * x = (const block_q6_K *) vx;
const int64_t i = blockIdx.x;
#if QK_K == 256
// assume 64 threads - this is very slightly better than the one below
const int64_t tid = threadIdx.x;
@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
#else
// assume 32 threads
const int64_t tid = threadIdx.x;
const int64_t ip = tid/16; // 0 or 1
const int64_t il = tid - 16*ip; // 0...15
dst_t * y = yy + i*QK_K + 16*ip + il;
const float d = x[i].d;
const uint8_t ql = x[i].ql[16*ip + il];
const uint8_t qh = x[i].qh[il] >> (2*ip);
const int8_t * sc = x[i].scales;
y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32);
#endif
}
template<typename dst_t>
@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
const block_iq2_xxs * x = (const block_iq2_xxs *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
const block_iq2_xs * x = (const block_iq2_xs *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
const block_iq2_s * x = (const block_iq2_s *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
const block_iq3_xxs * x = (const block_iq3_xxs *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
const block_iq3_s * x = (const block_iq3_s *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
const block_iq1_s * x = (const block_iq1_s *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
const block_iq1_m * x = (const block_iq1_m *) vx;
const int64_t tid = threadIdx.x;
#if QK_K == 256
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
#else
NO_DEVICE_CODE;
#endif
}
template<typename dst_t>
static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
#if QK_K != 64
template<typename dst_t>
static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
#endif
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
template<typename dst_t>
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
#else
dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}
template<typename dst_t>
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
#else
dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}
template<typename dst_t>
@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k
template<typename dst_t>
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
#else
dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}
template<typename dst_t>
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
#else
dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
#endif
}
template<typename dst_t>
@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t
template<typename dst_t>
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
#if QK_K == 64
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
#else
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
#endif
}
template <typename src_t, typename dst_t>

151
ggml-cuda/dmmv.cu
View file

@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
tmp += dall * sum1 - dmin * sum2;
}
#else
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3
const int offset = tid * K_QUANTS_PER_ITERATION;
uint32_t uaux[2];
const uint8_t * d = (const uint8_t *)uaux;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + offset;
const uint8_t * q = x[i].qs + offset;
const uint32_t * s = (const uint32_t *)x[i].scales;
uaux[0] = s[0] & 0x0f0f0f0f;
uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
const float2 dall = __half22float2(x[i].dm);
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
const uint8_t ql = q[l];
sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
+ y[l+16] * d[1] * ((ql >> 2) & 3)
+ y[l+32] * d[2] * ((ql >> 4) & 3)
+ y[l+48] * d[3] * ((ql >> 6) & 3);
sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
}
tmp += dall.x * sum1 - dall.y * sum2;
}
#endif
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
tmp += d * sum;
}
#else
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3
const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14
const int in = offset/8; // 0 or 1
const int im = offset%8; // 0...7
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + offset;
const uint8_t * q = x[i].qs + offset;
const uint8_t * s = x[i].scales;
const float dall = (float)x[i].d;
float sum = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
const uint8_t hl = x[i].hmask[im+l] >> in;
const uint8_t ql = q[l];
sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
+ y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
+ y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
+ y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
const block_q4_K * x = (const block_q4_K *)vx + ib0;
#if QK_K == 256
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
#endif
}
#else
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
const int step = tid * K_QUANTS_PER_ITERATION;
uint16_t aux16[2];
const uint8_t * s = (const uint8_t *)aux16;
float tmp = 0;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const uint8_t * q = x[i].qs + step;
const float * y = yy + i*QK_K + step;
const uint16_t * a = (const uint16_t *)x[i].scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const float d = (float)x[i].dm[0];
const float m = (float)x[i].dm[1];
float sum = 0.f;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
+ y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
+ y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3])
+ y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]);
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
}
#else
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
const int step = tid * K_QUANTS_PER_ITERATION;
const int im = step/8;
const int in = step%8;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const uint8_t * q = x[i].qs + step;
const int8_t * s = x[i].scales;
const float * y = yy + i*QK_K + step;
const float d = x[i].d;
float sum = 0.f;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
const uint8_t h = x[i].qh[in+j] >> im;
sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
+ y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
+ y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16))
+ y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16));
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
const block_q6_K * x = (const block_q6_K *)vx + ib0;
#if QK_K == 256
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
}
#else
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3
const int step = tid * K_QUANTS_PER_ITERATION;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + step;
const uint8_t * ql = x[i].ql + step;
const uint8_t * qh = x[i].qh + step;
const int8_t * s = x[i].scales;
const float d = x[i+0].d;
float sum = 0;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
+ y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
+ y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32)
+ y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32);
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);

115
ggml-cuda/fattn-common.cuh
View file

@ -1,7 +1,44 @@
#include "common.cuh"
#include <cstdint>
#define FATTN_KQ_STRIDE 256
#define HALF_MAX_HALF __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
#define SOFTMAX_FTZ_THRESHOLD -20.0f // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
typedef void (* fattn_kernel_t)(
const char * __restrict__ Q,
const char * __restrict__ K,
const char * __restrict__ V,
const char * __restrict__ mask,
float * __restrict__ dst,
float2 * __restrict__ dst_meta,
const float scale,
const float max_bias,
const float m0,
const float m1,
const uint32_t n_head_log2,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int nb31,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int ne0,
const int ne1,
const int ne2,
const int ne3);
template<int D, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
@ -45,3 +82,81 @@ static __global__ void flash_attn_combine_results(
dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
}
template <int D, int parallel_blocks>
void launch_fattn(ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, int nwarps, int cols_per_block) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
const ggml_tensor * mask = dst->src[3];
ggml_tensor * KQV = dst;
GGML_ASSERT(Q->type == GGML_TYPE_F32);
GGML_ASSERT(K->type == GGML_TYPE_F16);
GGML_ASSERT(V->type == GGML_TYPE_F16);
GGML_ASSERT(KQV->type == GGML_TYPE_F32);
GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
"the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
ggml_cuda_pool & pool = ctx.pool();
cudaStream_t main_stream = ctx.stream();
ggml_cuda_pool_alloc<float> dst_tmp(pool);
ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
}
const dim3 block_dim(WARP_SIZE, nwarps, 1);
const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
const int shmem = 0;
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
const uint32_t n_head = Q->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
(const char *) Q->data,
(const char *) K->data,
(const char *) V->data,
mask ? ((const char *) mask->data) : nullptr,
(parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
scale, max_bias, m0, m1, n_head_log2,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
K->nb[1], K->nb[2], K->nb[3],
KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
);
CUDA_CHECK(cudaGetLastError());
if ((parallel_blocks) == 1) {
return;
}
const dim3 block_dim_combine(D, 1, 1);
const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
const int shmem_combine = 0;
flash_attn_combine_results<D, parallel_blocks>
<<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
(dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
CUDA_CHECK(cudaGetLastError());
}

316
ggml-cuda/fattn-tile-f16.cu Normal file
View file

@ -0,0 +1,316 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-tile-f16.cuh"
#define FATTN_KQ_STRIDE_TILE_F16 64
template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,
const char * __restrict__ V,
const char * __restrict__ mask,
float * __restrict__ dst,
float2 * __restrict__ dst_meta,
const float scale,
const float max_bias,
const float m0,
const float m1,
const uint32_t n_head_log2,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int nb31,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int ne0,
const int ne1,
const int ne2,
const int ne3) {
#if FP16_AVAILABLE
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.y + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.y / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const int stride_KV2 = nb11 / sizeof(half2);
const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
__shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
half2 * KQ2 = (half2 *) KQ;
__shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
half kqmax[ncols/nwarps];
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
kqmax[j0/nwarps] = -HALF_MAX_HALF;
}
half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
// Convert Q to half2 and store in registers:
__shared__ half2 Q_h2[ncols][D/2];
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
}
}
__syncthreads();
const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
// Calculate KQ tile and keep track of new maximum KQ values:
half kqmax_new[ncols/nwarps];
#pragma unroll
for (int j = 0; j < ncols/nwarps; ++j) {
kqmax_new[j] = kqmax[j];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
const int i_KQ = i_KQ_0 + threadIdx.y;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
const int k_KQ = k_KQ_0 + threadIdx.x;
KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
}
}
__syncthreads();
half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
#pragma unroll
for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
half2 Q_k[ncols/nwarps];
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
const int i_KQ = i_KQ_0 + threadIdx.x;
K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
}
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
const int j_KQ = j_KQ_0 + threadIdx.y;
Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
}
}
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
const int i_KQ = i_KQ_0 + threadIdx.x;
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
const int j_KQ = j_KQ_0 + threadIdx.y;
half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
}
}
__syncthreads();
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
#pragma unroll
for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
const half2 val = h2exp(diff);
kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
}
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
}
}
__syncthreads();
#pragma unroll
for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
const int k = k0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
}
}
__syncthreads();
#pragma unroll
for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
half2 V_k[(D/2)/WARP_SIZE][2];
half2 KQ_k[ncols/nwarps];
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
}
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
}
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
}
}
}
__syncthreads();
}
#pragma unroll
for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
const int j_VKQ = j_VKQ_0 + threadIdx.y;
if (ic0 + j_VKQ >= ne01) {
return;
}
half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
kqsum_j = warp_reduce_sum(kqsum_j);
#pragma unroll
for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
const int i0 = i00 + 2*threadIdx.x;
half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
if (parallel_blocks == 1) {
dst_val /= __half2half2(kqsum_j);
}
const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = __low2float(dst_val);
dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
}
if (parallel_blocks != 1 && threadIdx.x == 0) {
dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
}
}
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
template <int cols_per_block, int parallel_blocks>
void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
switch (Q->ne[0]) {
case 64: {
constexpr int D = 64;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
default: {
GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
}
void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * KQV = dst;
const ggml_tensor * Q = dst->src[0];
const int32_t precision = KQV->op_params[2];
GGML_ASSERT(precision == GGML_PREC_DEFAULT);
if (Q->ne[1] <= 16) {
constexpr int cols_per_block = 16;
constexpr int parallel_blocks = 4;
launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 32) {
constexpr int cols_per_block = 32;
constexpr int parallel_blocks = 4;
launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
constexpr int cols_per_block = 32;
constexpr int parallel_blocks = 1;
launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
}

3
ggml-cuda/fattn-tile-f16.cuh Normal file
View file

@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

309
ggml-cuda/fattn-tile-f32.cu Normal file
View file

@ -0,0 +1,309 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-tile-f32.cuh"
#define FATTN_KQ_STRIDE_TILE_F32 32
template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
const char * __restrict__ Q,
const char * __restrict__ K,
const char * __restrict__ V,
const char * __restrict__ mask,
float * __restrict__ dst,
float2 * __restrict__ dst_meta,
const float scale,
const float max_bias,
const float m0,
const float m1,
const uint32_t n_head_log2,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int nb31,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int ne0,
const int ne1,
const int ne2,
const int ne3) {
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.y + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.y / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const int stride_KV2 = nb11 / sizeof(half2);
const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
__shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
__shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
float2 * KV_tmp2 = (float2 *) KV_tmp;
float kqmax[ncols/nwarps];
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
kqmax[j0/nwarps] = -FLT_MAX/2.0f;
}
float kqsum[ncols/nwarps] = {0.0f};
float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
// Convert Q to half2 and store in registers:
__shared__ float Q_f[ncols][D];
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
}
}
__syncthreads();
const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
// Calculate KQ tile and keep track of new maximum KQ values:
float kqmax_new[ncols/nwarps];
#pragma unroll
for (int j = 0; j < ncols/nwarps; ++j) {
kqmax_new[j] = kqmax[j];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
const int i_KQ = i_KQ_0 + threadIdx.y;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] = __low2float(tmp);
KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
}
}
__syncthreads();
float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
#pragma unroll
for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
float Q_k[ncols/nwarps];
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
const int i_KQ = i_KQ_0 + threadIdx.x;
K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
}
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
const int j_KQ = j_KQ_0 + threadIdx.y;
Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
}
}
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
const int i_KQ = i_KQ_0 + threadIdx.x;
#pragma unroll
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
const int j_KQ = j_KQ_0 + threadIdx.y;
sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
}
}
__syncthreads();
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
float kqsum_add = 0.0f;
#pragma unroll
for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
const float val = expf(diff);
kqsum_add += val;
KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
}
kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
}
}
__syncthreads();
#pragma unroll
for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
const int k = k0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
KV_tmp2[k*(D/2) + i].x = __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
}
}
__syncthreads();
#pragma unroll
for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
float2 V_k[(D/2)/WARP_SIZE];
float KQ_k[ncols/nwarps];
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
}
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
}
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
}
}
}
__syncthreads();
}
#pragma unroll
for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
const int j_VKQ = j_VKQ_0 + threadIdx.y;
if (ic0 + j_VKQ >= ne01) {
return;
}
float kqsum_j = kqsum[j_VKQ_0/nwarps];
kqsum_j = warp_reduce_sum(kqsum_j);
#pragma unroll
for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
const int i0 = i00 + 2*threadIdx.x;
float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
if (parallel_blocks == 1) {
dst_val.x /= kqsum_j;
dst_val.y /= kqsum_j;
}
const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
}
if (parallel_blocks != 1 && threadIdx.x == 0) {
dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
}
}
}
template <int cols_per_block, int parallel_blocks>
void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
switch (Q->ne[0]) {
case 64: {
constexpr int D = 64;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
default: {
GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
}
void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
if (Q->ne[1] <= 16) {
constexpr int cols_per_block = 16;
constexpr int parallel_blocks = 4;
launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 32) {
constexpr int cols_per_block = 32;
constexpr int parallel_blocks = 4;
launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
constexpr int cols_per_block = 32;
constexpr int parallel_blocks = 1;
launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
}

3
ggml-cuda/fattn-tile-f32.cuh Normal file
View file

@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

224
ggml-cuda/fattn-vec-f16.cu
View file

@ -53,17 +53,8 @@ static __global__ void flash_attn_vec_ext_f16(
const int stride_KV = nb11 / sizeof(half);
const int stride_KV2 = nb11 / sizeof(half2);
half slopeh = __float2half(1.0f);
// ALiBi
if (max_bias > 0.0f) {
const int h = blockIdx.y;
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slopeh = __float2half(powf(base, exph));
}
const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
constexpr int nwarps = D / WARP_SIZE;
@ -103,7 +94,7 @@ static __global__ void flash_attn_vec_ext_f16(
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
}
}
@ -221,6 +212,10 @@ static __global__ void flash_attn_vec_ext_f16(
#pragma unroll
for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
if (ncols > 2 && ic0 + j_VKQ >= ne01) {
break;
}
kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
@ -232,199 +227,104 @@ static __global__ void flash_attn_vec_ext_f16(
dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
}
if (parallel_blocks != 1 && tid != 0) {
#pragma unroll
for (int j = 0; j < ncols; ++j) {
dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
}
if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
}
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
ggml_cuda_pool & pool, cudaStream_t main_stream
) {
ggml_cuda_pool_alloc<float> dst_tmp(pool);
ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
}
constexpr int nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
const dim3 block_dim(WARP_SIZE, nwarps, 1);
const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
const int shmem = 0;
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
const uint32_t n_head = Q->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
<<<blocks_num, block_dim, shmem, main_stream>>> (
(const char *) Q->data,
(const char *) K->data,
(const char *) V->data,
mask ? ((const char *) mask->data) : nullptr,
parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
scale, max_bias, m0, m1, n_head_log2,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
K->nb[1], K->nb[2], K->nb[3],
KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
);
CUDA_CHECK(cudaGetLastError());
if (parallel_blocks == 1) {
return;
}
const dim3 block_dim_combine(D, 1, 1);
const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
const int shmem_combine = 0;
flash_attn_combine_results<D, parallel_blocks>
<<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
(dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
CUDA_CHECK(cudaGetLastError());
}
void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
const ggml_tensor * mask = dst->src[3];
ggml_tensor * KQV = dst;
ggml_tensor * Q = dst->src[0];
const int32_t precision = KQV->op_params[2];
GGML_ASSERT(precision == GGML_PREC_DEFAULT);
constexpr int cols_per_block = 1;
constexpr int cols_per_block = 1;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 256:
launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 64: {
constexpr int D = 64;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 256: {
constexpr int D = 256;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
default:
GGML_ASSERT(false);
break;
}
}
template <int cols_per_block, int parallel_blocks>
void launch_fattn_vec_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
switch (Q->ne[0]) {
case 64: {
constexpr int D = 64;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
default: {
GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
}
void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
const ggml_tensor * mask = dst->src[3];
ggml_tensor * KQV = dst;
const ggml_tensor * KQV = dst;
const ggml_tensor * Q = dst->src[0];
const int32_t precision = KQV->op_params[2];
GGML_ASSERT(precision == GGML_PREC_DEFAULT);
GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
if (Q->ne[1] == 1) {
constexpr int cols_per_block = 1;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
return;
}
if (Q->ne[1] == 2) {
constexpr int cols_per_block = 2;
constexpr int cols_per_block = 2;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 4) {
constexpr int cols_per_block = 4;
constexpr int cols_per_block = 4;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 8) {
constexpr int cols_per_block = 8;
constexpr int cols_per_block = 8;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
constexpr int cols_per_block = 8;
constexpr int cols_per_block = 8;
constexpr int parallel_blocks = 1;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
}

179
ggml-cuda/fattn-vec-f32.cu
View file

@ -52,17 +52,7 @@ static __global__ void flash_attn_vec_ext_f32(
const int stride_KV = nb11 / sizeof(half);
const int stride_KV2 = nb11 / sizeof(half2);
float slope = 1.0f;
// ALiBi
if (max_bias > 0.0f) {
const int h = blockIdx.y;
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = powf(base, exph);
}
const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
constexpr int nwarps = D / WARP_SIZE;
@ -101,7 +91,7 @@ static __global__ void flash_attn_vec_ext_f32(
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
Q_h2[j][i0/WARP_SIZE] = Q_f2[j*(nb01/sizeof(float2)) + i];
Q_h2[j][i0/WARP_SIZE] = ncols <= 2 || ic0 + j ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
Q_h2[j][i0/WARP_SIZE].x *= scale;
Q_h2[j][i0/WARP_SIZE].y *= scale;
}
@ -210,6 +200,10 @@ static __global__ void flash_attn_vec_ext_f32(
#pragma unroll
for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
if (ncols > 2 && ic0 + j_VKQ >= ne01) {
break;
}
kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
@ -221,164 +215,65 @@ static __global__ void flash_attn_vec_ext_f32(
dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
}
if (parallel_blocks != 1 && tid != 0) {
#pragma unroll
for (int j = 0; j < ncols; ++j) {
dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
}
if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
}
}
template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
ggml_cuda_pool & pool, cudaStream_t main_stream
) {
ggml_cuda_pool_alloc<float> dst_tmp(pool);
ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
template <int cols_per_block, int parallel_blocks>
void launch_fattn_vec_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
switch (Q->ne[0]) {
case 64: {
constexpr int D = 64;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = D/WARP_SIZE;
fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
} break;
default: {
GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
} break;
}
constexpr int nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
const dim3 block_dim(WARP_SIZE, nwarps, 1);
const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
const int shmem = 0;
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
const uint32_t n_head = Q->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
<<<blocks_num, block_dim, shmem, main_stream>>> (
(const char *) Q->data,
(const char *) K->data,
(const char *) V->data,
mask ? ((const char *) mask->data) : nullptr,
parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
scale, max_bias, m0, m1, n_head_log2,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
K->nb[1], K->nb[2], K->nb[3],
KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
);
CUDA_CHECK(cudaGetLastError());
if (parallel_blocks == 1) {
return;
}
const dim3 block_dim_combine(D, 1, 1);
const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
const int shmem_combine = 0;
flash_attn_combine_results<D, parallel_blocks>
<<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
(dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
CUDA_CHECK(cudaGetLastError());
}
void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
const ggml_tensor * mask = dst->src[3];
ggml_tensor * KQV = dst;
GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
if (Q->ne[1] == 1) {
constexpr int cols_per_block = 1;
constexpr int cols_per_block = 1;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] == 2) {
constexpr int cols_per_block = 2;
constexpr int cols_per_block = 2;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 4) {
constexpr int cols_per_block = 4;
constexpr int cols_per_block = 4;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
if (Q->ne[1] <= 8) {
constexpr int cols_per_block = 8;
constexpr int cols_per_block = 8;
constexpr int parallel_blocks = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
return;
}
constexpr int cols_per_block = 8;
constexpr int cols_per_block = 8;
constexpr int parallel_blocks = 1;
switch (Q->ne[0]) {
case 64:
launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
case 128:
launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
break;
default:
GGML_ASSERT(false);
break;
}
launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
}

201
ggml-cuda/fattn.cu
View file

@ -1,5 +1,7 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-tile-f16.cuh"
#include "fattn-tile-f32.cuh"
#include "fattn-vec-f16.cuh"
#include "fattn-vec-f32.cuh"
#include "fattn.cuh"
@ -83,19 +85,9 @@ static __global__ void flash_attn_ext_f16(
const int stride_Q = nb01 / sizeof(float);
const int stride_KV = nb11 / sizeof(half);
half slopeh = __float2half(1.0f);
half2 slope2 = make_half2(1.0f, 1.0f);
// ALiBi
if (max_bias > 0.0f) {
const int h = blockIdx.y;
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slopeh = __float2half(powf(base, exph));
slope2 = make_half2(slopeh, slopeh);
}
const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
const half2 slope2 = make_half2(slopef, slopef);
frag_b Q_b[D/16][ncols/frag_n];
@ -437,117 +429,64 @@ static_assert(get_VKQ_stride( 80, 1, 16) == 16, "Test failed.");
static_assert(get_VKQ_stride( 80, 2, 16) == 16, "Test failed.");
static_assert(get_VKQ_stride( 80, 4, 16) == 16, "Test failed.");
template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename KQ_acc_t> void launch_fattn_f16_impl(
const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
ggml_cuda_pool & pool, cudaStream_t main_stream
) {
ggml_cuda_pool_alloc<float> dst_tmp(pool);
ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
template <int D, int cols_per_block, int nwarps, typename KQ_acc_t>
void launch_fattn_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
if (parallel_blocks > 1) {
dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
}
constexpr int frag_m = (cols_per_block) == 8 && (D) % 32 == 0 ? 32 : 16;
const dim3 block_dim(WARP_SIZE, nwarps, 1);
const dim3 blocks_num(parallel_blocks*(Q->ne[1] + cols_per_block - 1) / cols_per_block, Q->ne[2], Q->ne[3]);
const int shmem = 0;
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
const uint32_t n_head = Q->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>
<<<blocks_num, block_dim, shmem, main_stream>>> (
(const char *) Q->data,
(const char *) K->data,
(const char *) V->data,
mask ? ((const char *) mask->data) : nullptr,
(parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
scale, max_bias, m0, m1, n_head_log2,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
K->nb[1], K->nb[2], K->nb[3],
KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
);
CUDA_CHECK(cudaGetLastError());
if ((parallel_blocks) == 1) {
return;
}
const dim3 block_dim_combine(D, 1, 1);
const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
const int shmem_combine = 0;
flash_attn_combine_results<D, parallel_blocks>
<<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
(dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
CUDA_CHECK(cudaGetLastError());
}
template <int D, int cols_per_block, int nwarps, typename KQ_acc_t> void launch_fattn_f16(
const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
const int nsm, ggml_cuda_pool & pool, cudaStream_t main_stream
) {
constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
if (4*blocks_num_pb1 < 2*nsm) {
launch_fattn_f16_impl<D, cols_per_block, nwarps, 4, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
constexpr int parallel_blocks = 4;
fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
return;
}
if (2*blocks_num_pb1 < 2*nsm) {
launch_fattn_f16_impl<D, cols_per_block, nwarps, 2, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
constexpr int parallel_blocks = 2;
fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
return;
}
launch_fattn_f16_impl<D, cols_per_block, nwarps, 1, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
constexpr int parallel_blocks = 1;
fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
}
void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
const ggml_tensor * mask = dst->src[3];
ggml_tensor * KQV = dst;
GGML_ASSERT(Q->type == GGML_TYPE_F32);
GGML_ASSERT(K->type == GGML_TYPE_F16);
GGML_ASSERT(V->type == GGML_TYPE_F16);
GGML_ASSERT(KQV->type == GGML_TYPE_F32);
GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
"the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
const ggml_tensor * KQV = dst;
const ggml_tensor * Q = dst->src[0];
ggml_cuda_set_device(ctx.device);
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const int32_t precision = KQV->op_params[2];
// On AMD the tile kernels perform poorly, use the vec kernel instead:
if (cc >= CC_OFFSET_AMD) {
if (precision == GGML_PREC_DEFAULT) {
ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
} else {
ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
}
return;
}
if (!fast_fp16_available(cc)) {
ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
if (Q->ne[1] <= 8) {
ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
} else {
ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
}
return;
}
if (!fp16_mma_available(cc)) {
ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
if (Q->ne[1] <= 8) {
ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
} else {
ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
}
return;
}
@ -562,22 +501,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
constexpr int nwarps = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
break;
case 80:
launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
break;
case 96:
launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
break;
case 112:
launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
break;
case 128:
launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
break;
case 256:
launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
break;
default:
GGML_ASSERT(false);
@ -588,22 +527,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
constexpr int nwarps = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
break;
case 80:
launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
break;
case 96:
launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
break;
case 112:
launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
break;
case 128:
launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
break;
// case 256:
// launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
// launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
// break;
default:
GGML_ASSERT(false);
@ -623,16 +562,16 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
constexpr int nwarps = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
break;
case 96:
launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
break;
case 128:
launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
break;
case 256:
launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
break;
default:
GGML_ASSERT(false);
@ -646,22 +585,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
constexpr int nwarps = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
break;
case 80:
launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
break;
case 96:
launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
break;
case 112:
launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
break;
case 128:
launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
break;
case 256:
launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
break;
default:
GGML_ASSERT(false);
@ -674,22 +613,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
constexpr int nwarps = 4;
switch (Q->ne[0]) {
case 64:
launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
break;
case 80:
launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
break;
case 96:
launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
break;
case 112:
launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
break;
case 128:
launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
break;
case 256:
launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
break;
default:
GGML_ASSERT(false);

1253
ggml-cuda/mmq.cu
View file

File diff suppressed because it is too large Load diff

72
ggml-cuda/rope.cu
View file

@ -58,10 +58,10 @@ static __global__ void rope(
dst[i + 1] = x0*sin_theta + x1*cos_theta;
}
template<typename T, bool has_pos>
template<typename T, bool has_pos, bool has_freq_facs>
static __global__ void rope_neox(
const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, const float * freq_factors
) {
const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
@ -88,7 +88,9 @@ static __global__ void rope_neox(
float cur_rot = inv_ndims * ic - ib;
const int p = has_pos ? pos[i2] : 0;
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f;
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f)/freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -164,7 +166,7 @@ static void rope_cuda(
template<typename T>
static void rope_neox_cuda(
const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
) {
GGML_ASSERT(ncols % 2 == 0);
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@ -175,15 +177,29 @@ static void rope_neox_cuda(
const float inv_ndims = -1.0f / n_dims;
if (pos == nullptr) {
rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims
);
if (freq_factors == nullptr) {
rope_neox<T, false, false><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
} else {
rope_neox<T, false, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
}
} else {
rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims
);
if (freq_factors == nullptr) {
rope_neox<T, true, false><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
} else {
rope_neox<T, true, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
);
}
}
}
@ -214,24 +230,27 @@ static void rope_cuda_f32(
static void rope_neox_cuda_f16(
const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
}
static void rope_neox_cuda_f32(
const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
) {
rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
}
void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
@ -241,7 +260,6 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne2 = dst->ne[2];
const int64_t nrows = ggml_nrows(src0);
//const int n_past = ((int32_t *) dst->op_params)[0];
@ -259,16 +277,22 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const float * freq_factors = nullptr;
const int32_t * pos = nullptr;
if ((mode & 1) == 0) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
GGML_ASSERT(src1->ne[0] == ne2);
pos = (const int32_t *) src1_d;
}
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
pos = (const int32_t *) src1_d;
if (is_neox) {
if (src2 != nullptr) {
freq_factors = (const float *) src2->data;
}
} else {
GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox");
}
rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
@ -280,12 +304,12 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
if (src0->type == GGML_TYPE_F32) {
rope_neox_cuda_f32(
(const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, stream
attn_factor, corr_dims, freq_factors, stream
);
} else if (src0->type == GGML_TYPE_F16) {
rope_neox_cuda_f16(
(const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, stream
attn_factor, corr_dims, freq_factors, stream
);
} else {
GGML_ASSERT(false);

13
ggml-cuda/softmax.cu
View file

@ -1,3 +1,4 @@
#include "common.cuh"
#include "softmax.cuh"
template <typename T>
@ -23,17 +24,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, float * dst
const int warp_id = threadIdx.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
float slope = 1.0f;
// ALiBi
if (max_bias > 0.0f) {
const int h = rowx/nrows_y; // head index
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = powf(base, exph);
}
const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
extern __shared__ float data_soft_max_f32[];
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication

126
ggml-cuda/vecdotq.cuh
View file

@ -712,7 +712,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#ifndef GGML_QKK_64
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
int v[2];
@ -754,58 +753,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
}
return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
#else
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
float sumf_d = 0.0f;
float sumf_m = 0.0f;
uint16_t aux16[2];
const uint8_t * s = (const uint8_t *)aux16;
const uint16_t * a = (const uint16_t *)bq4_K->scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const float dall = bq4_K->dm[0];
const float dmin = bq4_K->dm[1];
const float d8_1 = __low2float(bq8_1[0].ds);
const float d8_2 = __low2float(bq8_1[1].ds);
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * q4 = (const int *)bq4_K->qs + (iqs/2);
const int v1 = q4[0];
const int v2 = q4[4];
const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
return dall * sumf_d - dmin * sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
#endif
}
static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#ifndef GGML_QKK_64
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
int vl[2];
@ -847,48 +799,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
}
return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
#else
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
const int8_t * s = bq5_K->scales;
const float d = bq5_K->d;
const float d8_1 = __low2half(bq8_1[0].ds);
const float d8_2 = __low2half(bq8_1[1].ds);
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * ql = (const int *)bq5_K->qs + (iqs/2);
const int vl1 = ql[0];
const int vl2 = ql[4];
const int step = 4 * (iqs/2); // 0, 4, 8, 12
const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
const int in = step%8; // 0, 4, 0, 4
const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
+ d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
return d * sumf_d;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
#endif
}
static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
@ -919,7 +829,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
#if QR2_XXS == 8
@ -960,15 +869,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
}
return d * (sumi1 + sumi2);
#endif
#else
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
const int ib32 = iqs;
@ -1002,17 +907,12 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
#else
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
}
// TODO
static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
const int ib32 = iqs;
@ -1048,16 +948,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
#else
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
const int ib32 = iqs;
@ -1082,16 +977,12 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
#else
NO_DEVICE_CODE;
#endif
#else
NO_DEVICE_CODE;
#endif
}
// TODO: don't use lookup table for signs
static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
const int ib32 = iqs;
@ -1114,14 +1005,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
#else
NO_DEVICE_CODE;
#endif
#else
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
const int ib32 = iqs;
@ -1149,14 +1036,10 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
const float d = d1q * __low2float (bq8_1[ib32].ds);
const float m = d1q * __high2float(bq8_1[ib32].ds);
return d * sumi + m * delta;
#else
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
const int ib32 = iqs;
@ -1192,9 +1075,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
#else
NO_DEVICE_CODE;
#endif
}
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
@ -1250,9 +1130,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
@ -1270,10 +1148,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
sumi2 = __dp4a(v2, q8[j+4], sumi2);
}
return d * (sumi1 + sumi2);
#else
NO_DEVICE_CODE;
#endif
#else
return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
#endif

40
ggml-impl.h
View file

@ -17,6 +17,18 @@
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#if defined(_WIN32)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
/**
* Converts brain16 to float32.
*
@ -443,6 +455,34 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__loongarch_asx)
typedef union {
int32_t i;
float f;
} ft_union;
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
}
static __m256 __lasx_xvreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER

4
ggml-kompute.cpp
View file

@ -1677,6 +1677,10 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
} break;
case GGML_OP_ROPE:
{
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
GGML_ASSERT(ne10 == ne02);
GGML_ASSERT(src0t == dstt);
// const int n_past = ((int32_t *) dst->op_params)[0];

138
ggml-metal.m
View file

@ -381,10 +381,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
// dictionary of preprocessor macros
NSMutableDictionary * prep = [NSMutableDictionary dictionary];
#ifdef GGML_QKK_64
prep[@"GGML_QKK_64"] = @(1);
#endif
MTLCompileOptions* options = [MTLCompileOptions new];
options.preprocessorMacros = prep;
@ -927,22 +923,32 @@ static enum ggml_status ggml_metal_graph_compute(
const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
const int64_t ne13 = src1 ? src1->ne[3] : 0;
const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
const uint64_t nb13 = src1 ? src1->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne21 = src2 ? src2->ne[1] : 0;
const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
@ -1763,11 +1769,7 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -1785,16 +1787,6 @@ static enum ggml_status ggml_metal_graph_compute(
const int n_as = src0->ne[2];
// src2 = ids
const int64_t ne20 = src2->ne[0];
const int64_t ne21 = src2->ne[1];
const int64_t ne22 = src2->ne[2]; GGML_UNUSED(ne22);
const int64_t ne23 = src2->ne[3]; GGML_UNUSED(ne23);
const uint64_t nb20 = src2->nb[0]; GGML_UNUSED(nb20);
const uint64_t nb21 = src2->nb[1];
const uint64_t nb22 = src2->nb[2]; GGML_UNUSED(nb22);
const uint64_t nb23 = src2->nb[3]; GGML_UNUSED(nb23);
const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
GGML_ASSERT(src2t == GGML_TYPE_I32);
@ -2018,12 +2010,7 @@ static enum ggml_status ggml_metal_graph_compute(
{
nth0 = 4;
nth1 = 16;
#if QK_K == 64
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
#else
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
#endif
} break;
default:
{
@ -2088,11 +2075,7 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -2244,7 +2227,13 @@ static enum ggml_status ggml_metal_graph_compute(
// skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
float freq_base;
float freq_scale;
float ext_factor;
float attn_factor;
float beta_fast;
float beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
@ -2252,6 +2241,15 @@ static enum ggml_status ggml_metal_graph_compute(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
GGML_ASSERT(!is_glm && "GLM RoPE not implemented in Metal");
if (!is_neox) {
GGML_ASSERT(id_src2 == nil && "TODO: freq_factors not implemented for !is_neox");
}
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
@ -2263,33 +2261,38 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:14];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&n_past length:sizeof( int) atIndex:19];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:20];
[encoder setBytes:&mode length:sizeof( int) atIndex:21];
[encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:22];
[encoder setBytes:&freq_base length:sizeof( float) atIndex:23];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:24];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:25];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:26];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:27];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:28];
if (id_src2 != nil) {
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:2];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&n_past length:sizeof( int) atIndex:20];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:21];
[encoder setBytes:&mode length:sizeof( int) atIndex:22];
[encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:23];
[encoder setBytes:&freq_base length:sizeof( float) atIndex:24];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:25];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:26];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:27];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:28];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:29];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
@ -2535,11 +2538,6 @@ static enum ggml_status ggml_metal_graph_compute(
GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
"the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
//const int64_t ne31 = src3 ? src3->ne[1] : 0;
const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);

431
ggml-metal.metal
View file

@ -1640,6 +1640,7 @@ static void rope_yarn_corr_dims(
typedef void (rope_t)(
device const void * src0,
device const int32_t * src1,
device const float * src2,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
@ -1675,6 +1676,7 @@ template<typename T>
kernel void kernel_rope(
device const void * src0,
device const int32_t * src1,
device const float * src2,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
@ -1744,8 +1746,10 @@ kernel void kernel_rope(
// simplified from `(ib * n_dims + ic) * inv_ndims`
const float cur_rot = inv_ndims*ic - ib;
const float freq_factor = src2 != src0 ? src2[ic/2] : 1.0f;
const float theta = theta_0 * pow(freq_base, cur_rot) / freq_factor;
const float theta = theta_0 * pow(freq_base, cur_rot);
float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -2204,11 +2208,7 @@ kernel void kernel_flash_attn_ext_f16(
// pointer to the mask
device const half * mp = (device const half *) (mask + iq1*nb31);
// prepare diagonal scale matrix
simdgroup_float8x8 mscale(scale);
// prepare diagonal slope matrix
simdgroup_float8x8 mslope(1.0f);
float slope = 1.0f;
// ALiBi
if (max_bias > 0.0f) {
@ -2217,7 +2217,7 @@ kernel void kernel_flash_attn_ext_f16(
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
mslope = simdgroup_float8x8(pow(base, exph));
slope = pow(base, exph);
}
// loop over the KV cache
@ -2242,18 +2242,20 @@ kernel void kernel_flash_attn_ext_f16(
simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
}
simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
const short tx = tiisg%4;
const short ty = tiisg/4;
if (mask != q) {
// mqk = mqk*scale + mask*slope
simdgroup_half8x8 mm;
simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
simdgroup_multiply(mm, mslope, mm);
simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
ss[8*cc + ty*TF + 2*tx + 0] = scale*ss[8*cc + ty*TF + 2*tx + 0] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
ss[8*cc + ty*TF + 2*tx + 1] = scale*ss[8*cc + ty*TF + 2*tx + 1] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
} else {
// mqk = mqk*scale
simdgroup_multiply(mqk, mscale, mqk);
ss[8*cc + ty*TF + 2*tx + 0] *= scale;
ss[8*cc + ty*TF + 2*tx + 1] *= scale;
}
simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
}
}
@ -2816,8 +2818,7 @@ kernel void kernel_cpy_f32_f16(
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
// TODO: is there a better way to handle -INFINITY?
dst_data[i00] = src[0] == -INFINITY ? -MAXHALF : src[0];
dst_data[i00] = src[0];
}
}
@ -3385,7 +3386,6 @@ void kernel_mul_mv_q2_K_f32_impl(
const int step = sizeof(block_q2_K) * nb;
#if QK_K == 256
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int iq = it/4; // 0 or 1
@ -3437,57 +3437,6 @@ void kernel_mul_mv_q2_K_f32_impl(
y4 += 4 * QK_K;
}
#else
const int ix = tiisg/2; // 0...15
const int it = tiisg%2; // 0...1
device const float * y4 = y + ix * QK_K + 8 * it;
for (int ib = ix; ib < nb; ib += 16) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
}
device const uint8_t * sc = (device const uint8_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
qs += step/2;
sc += step;
dh += step/2;
}
y4 += 16 * QK_K;
}
#endif
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
@ -3525,7 +3474,6 @@ kernel void kernel_mul_mv_q2_K_f32(
kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
}
#if QK_K == 256
void kernel_mul_mv_q3_K_f32_impl(
device const void * src0,
device const float * src1,
@ -3684,84 +3632,6 @@ void kernel_mul_mv_q3_K_f32_impl(
}
}
}
#else
void kernel_mul_mv_q3_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne10,
constant int64_t & ne12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
const int row = 2 * r0 + sgitg;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
const int ix = tiisg/4;
const int il = 4 * (tiisg%4);// 0, 4, 8, 12
const int iq = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
float2 sum = {0.f, 0.f};
for (int i = ix; i < nb; i += 8) {
const float d_all = (float)(x[i].d);
device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
device const uint16_t * s = (device const uint16_t *)(x[i].scales);
device const float * y = yy + i * QK_K + il;
const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
for (int l = 0; l < 4; l += 2) {
const uint16_t hm = h[l/2] >> iq;
sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 : 4))
+ y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
+ y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
+ y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
+ y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
+ y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
+ y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
}
}
const float sumf = sum[0] + sum[1] * 1.f/256.f;
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + row] = tot;
}
}
#endif
[[host_name("kernel_mul_mv_q3_K_f32")]]
kernel void kernel_mul_mv_q3_K_f32(
@ -3791,7 +3661,6 @@ kernel void kernel_mul_mv_q3_K_f32(
kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
}
#if QK_K == 256
void kernel_mul_mv_q4_K_f32_impl(
device const void * src0,
device const float * src1,
@ -3905,103 +3774,6 @@ void kernel_mul_mv_q4_K_f32_impl(
}
}
}
#else
void kernel_mul_mv_q4_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne10,
constant int64_t & ne12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int ix = tiisg/4; // 0...7
const int it = tiisg%4; // 0...3
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = r0 * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[8];
float yh[8];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q4_K) * nb / 2;
device const float * y4 = y + ix * QK_K + 8 * it;
uint16_t sc16[4];
for (int ib = ix; ib < nb; ib += 8) {
float2 sumy = {0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i] = y4[i+ 0]; sumy[0] += yl[i];
yh[i] = y4[i+32]; sumy[1] += yh[i];
}
device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = x[ib].d;
for (int row = 0; row < N_DST; row++) {
sc16[0] = sc[0] & 0x000f;
sc16[1] = sc[0] & 0x0f00;
sc16[2] = sc[0] & 0x00f0;
sc16[3] = sc[0] & 0xf000;
float2 acc1 = {0.f, 0.f};
float2 acc2 = {0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
(acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
qs += step;
sc += step;
dh += step;
}
y4 += 8 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
#endif
[[host_name("kernel_mul_mv_q4_K_f32")]]
kernel void kernel_mul_mv_q4_K_f32(
@ -4069,8 +3841,6 @@ void kernel_mul_mv_q5_K_f32_impl(
const int step = sizeof(block_q5_K) * nb;
#if QK_K == 256
#
float yl[16], yh[16];
const uint16_t kmask1 = 0x3f3f;
@ -4153,54 +3923,6 @@ void kernel_mul_mv_q5_K_f32_impl(
y1 += 4 * QK_K;
}
#else
float yl[8], yh[8];
const int il = 4 * (tiisg/8); // 0, 4, 8, 12
const int ix = tiisg%8;
const int iq = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
device const float * y = yy + ix*QK_K + il;
for (int i = ix; i < nb; i += 8) {
for (int l = 0; l < 4; ++l) {
yl[l+0] = y[l+ 0];
yl[l+4] = y[l+16];
yh[l+0] = y[l+32];
yh[l+4] = y[l+48];
}
device const half * dh = &x[i].d;
device const uint8_t * q = x[i].qs + il;
device const uint8_t * h = x[i].qh + in;
device const int8_t * s = x[i].scales;
for (int row = 0; row < 2; ++row) {
const float d = dh[0];
float2 acc = {0.f, 0.f};
for (int l = 0; l < 4; ++l) {
const uint8_t hl = h[l] >> iq;
acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
+ yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
+ yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
}
sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
q += step;
h += step;
s += step;
dh += step/2;
}
y += 8 * QK_K;
}
#endif
for (int row = 0; row < 2; ++row) {
const float tot = simd_sum(sumf[row]);
@ -4279,7 +4001,6 @@ void kernel_mul_mv_q6_K_f32_impl(
float sumf = 0;
#if QK_K == 256
const int tid = tiisg/2;
const int ix = tiisg%2;
const int ip = tid/8; // 0 or 1
@ -4315,30 +4036,6 @@ void kernel_mul_mv_q6_K_f32_impl(
}
#else
const int ix = tiisg/4;
const int il = 4*(tiisg%4);
for (int i = ix; i < nb; i += 8) {
device const float * y = yy + i * QK_K + il;
device const uint8_t * ql = x[i].ql + il;
device const uint8_t * qh = x[i].qh + il;
device const int8_t * s = x[i].scales;
const float d = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 4; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >> 4) | ((qh[l] & kmask3) >> 0)) - 32);
sums[3] += y[l+48] * ((int8_t)((ql[l+16] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
}
#endif
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + row] = tot;
@ -5172,9 +4869,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
device const float * y4 = y + 32 * ix;
#if QK_K != 64
iq1m_scale_t scale;
#endif
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
@ -5195,10 +4890,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
device const uint16_t * sc = (device const uint16_t *)xr->scales;
for (int row = 0; row < N_DST; row++) {
#if QK_K != 64
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
#endif
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
@ -5214,14 +4906,9 @@ void kernel_mul_mv_iq1_m_f32_impl(
}
const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
#if QK_K == 64
const float d = (float) *((device const half *)(sc - 1));
sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) +
(sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1));
#else
sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
(sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
#endif
sc += nb*sizeof(block_iq1_m)/2;
qs += nb*sizeof(block_iq1_m);
@ -5333,7 +5020,6 @@ void kernel_mul_mv_iq4_nl_f32_impl(
}
}
#if QK_K != 64
void kernel_mul_mv_iq4_xs_f32_impl(
device const void * src0,
device const float * src1,
@ -5428,7 +5114,6 @@ void kernel_mul_mv_iq4_xs_f32_impl(
}
}
}
#endif
[[host_name("kernel_mul_mv_iq1_s_f32")]]
kernel void kernel_mul_mv_iq1_s_f32(
@ -5541,11 +5226,7 @@ kernel void kernel_mul_mv_iq4_xs_f32(
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
#if QK_K == 64
kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
#else
kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
#endif
}
//============================= templates and their specializations =============================
@ -5671,10 +5352,9 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
float dl, ml;
uint8_t sc = xb->scales[il];
#if QK_K == 256
q = q + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
#endif
half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uchar mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
@ -5690,7 +5370,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
device const uint8_t * h = (device const uint8_t *)xb->hmask;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
q = q + 32 * (il/8) + 16 * (il&1);
h = h + 16 * (il&1);
uint8_t m = 1 << (il/2);
@ -5711,17 +5390,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
}
#else
float kcoef = il&1 ? 1.f/16.f : 1.f;
uint16_t kmask = il&1 ? 0xF0 : 0x0F;
float dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
float coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
uint8_t m = 1<<(il*2);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
}
#endif
}
static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
@ -5733,7 +5401,6 @@ template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
device const uchar * q = xb->qs;
#if QK_K == 256
short is = (il/4) * 2;
q = q + (il/4) * 32 + 16 * (il&1);
il = il & 3;
@ -5742,16 +5409,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
#else
(void) get_scale_min_k4_just2;
q = q + 16 * (il&1);
device const uint8_t * s = xb->scales;
device const half2 * dh = (device const half2 *)xb->d;
const float2 d = (float2)dh[0];
const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
const float ml = il<2 ? d[1] * (s[0]>>4) : d[1] * (s[1]>>4);
#endif
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
@ -5763,7 +5421,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
device const uint8_t * q = xb->qs;
device const uint8_t * qh = xb->qh;
#if QK_K == 256
short is = (il/4) * 2;
q = q + 32 * (il/4) + 16 * (il&1);
qh = qh + 16 * (il&1);
@ -5780,17 +5437,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
}
#else
q = q + 16 * (il&1);
device const int8_t * s = xb->scales;
const float dl = xb->d * s[il];
uint8_t m = 1<<(il*2);
const float coef = il<2 ? 1.f : 1.f/16.f;
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
}
#endif
}
template <typename type4x4>
@ -5800,15 +5446,11 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
device const uint8_t * qh = (device const uint8_t *)xb->qh;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
qh = qh + 32*(il/8) + 16*(il&1);
float sc = scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
#else
ql = ql + 16 * (il&1);
float sc = scales[il];
#endif
const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F;
const float coef = il>1 ? 1.f/16.f : 1.f;
@ -5965,20 +5607,15 @@ void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 &
const int ib32 = il/2;
il = il%2;
device const uint16_t * sc = (device const uint16_t *)xb->scales;
#if QK_K == 64
const float d = xb->d;
#else
iq1m_scale_t scale;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = scale.f16;
#endif
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * qh = xb->qh + 2*ib32 + il;
#if QK_K == 64
const float dl = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1);
#else
const float dl = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
#endif
const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
@ -6008,9 +5645,6 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4
template <typename type4x4>
void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
#if QK_K == 64
dequantize_iq4_nl(xb, il, reg);
#else
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
@ -6027,7 +5661,6 @@ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
#endif
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
@ -6532,11 +6165,7 @@ kernel void kernel_mul_mm_id(
sgitg);
}
#if QK_K == 256
#define QK_NL 16
#else
#define QK_NL 4
#endif
//
// get rows
@ -6576,11 +6205,7 @@ template [[host_name("kernel_get_rows_iq2_s")]] kernel get_rows_t kernel_get_r
template [[host_name("kernel_get_rows_iq1_s")]] kernel get_rows_t kernel_get_rows<block_iq1_s, QK_NL, dequantize_iq1_s>;
template [[host_name("kernel_get_rows_iq1_m")]] kernel get_rows_t kernel_get_rows<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_get_rows_iq4_nl")]] kernel get_rows_t kernel_get_rows<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
//
// matrix-matrix multiplication
@ -6608,11 +6233,7 @@ template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mat_mm_t kernel_mul_m
template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_s, QK_NL, dequantize_iq1_s>;
template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
//
// indirect matrix-matrix multiplication
@ -6640,11 +6261,7 @@ template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mat_mm_id_t kernel
template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s, QK_NL, dequantize_iq1_s>;
template [[host_name("kernel_mul_mm_id_iq1_m_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
//
// matrix-vector multiplication
@ -6853,7 +6470,5 @@ template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t
template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
#if QK_K != 64
template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
#endif

216
ggml-mpi.c
View file

@ -1,216 +0,0 @@
#include "ggml-mpi.h"
#include "ggml.h"
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define UNUSED GGML_UNUSED
struct ggml_mpi_context {
int rank;
int size;
};
void ggml_mpi_backend_init(void) {
MPI_Init(NULL, NULL);
}
void ggml_mpi_backend_free(void) {
MPI_Finalize();
}
struct ggml_mpi_context * ggml_mpi_init(void) {
struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
return ctx;
}
void ggml_mpi_free(struct ggml_mpi_context * ctx) {
free(ctx);
}
int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
return ctx->rank;
}
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads) {
UNUSED(ctx_mpi);
// synchronize the worker node parameters with the root node
MPI_Barrier(MPI_COMM_WORLD);
MPI_Bcast(n_tokens, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_past, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
}
static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
if (t == NULL) {
fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
return -1;
}
for (int i = 0; i < gf->n_nodes; i++) {
if (gf->nodes[i] == t) {
return i;
}
}
fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
return -1;
}
static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
MPI_Datatype mpi_type;
switch (t->type) {
case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
case GGML_TYPE_F32: mpi_type = MPI_FLOAT; break;
default: GGML_ASSERT(false && "not implemented");
}
const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
GGML_ASSERT(retval == MPI_SUCCESS);
}
static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
MPI_Datatype mpi_type;
switch (t->type) {
case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
case GGML_TYPE_F32: mpi_type = MPI_FLOAT; break;
default: GGML_ASSERT(false && "not implemented");
}
MPI_Status status; UNUSED(status);
const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
GGML_ASSERT(retval == MPI_SUCCESS);
}
// TODO: there are many improvements that can be done to this implementation
void ggml_mpi_graph_compute_pre(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers) {
const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size;
struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
if (inp_tokens == NULL) {
fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
return;
}
struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
if (inp0 == NULL) {
fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
return;
}
GGML_ASSERT(inp0 == gf->nodes[0]);
// distribute the compute graph into slices across the MPI nodes
//
// the main node (0) processes the last layers + the remainder of the compute graph
// and is responsible to pass the input tokens to the first node (1)
//
// node 1: [( 0) * n_per_node, ( 1) * n_per_node)
// node 2: [( 1) * n_per_node, ( 2) * n_per_node)
// ...
// node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
// node 0: [(n-1) * n_per_node, n_nodes)
//
if (mpi_rank > 0) {
if (mpi_rank == 1) {
// the first node (1) receives the input tokens from the main node (0)
ggml_mpi_tensor_recv(inp_tokens, 0);
} else {
// recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
}
} else if (mpi_size > 1) {
// node 0 sends the input tokens to node 1
ggml_mpi_tensor_send(inp_tokens, 1);
// recv the output data from the last node
ggml_mpi_tensor_recv(inp0, mpi_size - 1);
}
{
const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
const int il0 = (mpi_idx + 0) * n_per_node;
const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
char name_l0[GGML_MAX_NAME];
char name_l1[GGML_MAX_NAME];
snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
const int idx_l0 = ggml_graph_get_node_idx(gf, name_l0);
const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
if (idx_l0 < 0 || idx_l1 < 0) {
fprintf(stderr, "%s: layer input nodes not found\n", __func__);
return;
}
// attach the input data to all nodes that need it
// TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
for (int i = idx_l0; i < idx_l1; i++) {
if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[0] = inp0;
}
if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[1] = inp0;
}
}
// TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
for (int i = 1; i < idx_l1 - idx_l0; i++) {
gf->nodes[i] = gf->nodes[idx_l0 + i];
gf->grads[i] = gf->grads[idx_l0 + i];
}
// the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
if (mpi_idx != 0) {
gf->nodes[0]->op = GGML_OP_NONE;
}
gf->n_nodes = idx_l1 - idx_l0;
//fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
}
}
void ggml_mpi_graph_compute_post(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers) {
UNUSED(n_layers);
const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size;
// send the output data to the next node
if (mpi_rank > 0) {
ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
}
}

39
ggml-mpi.h
View file

@ -1,39 +0,0 @@
#pragma once
struct ggml_context;
struct ggml_tensor;
struct ggml_cgraph;
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_mpi_context;
void ggml_mpi_backend_init(void);
void ggml_mpi_backend_free(void);
struct ggml_mpi_context * ggml_mpi_init(void);
void ggml_mpi_free(struct ggml_mpi_context * ctx);
int ggml_mpi_rank(struct ggml_mpi_context * ctx);
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads);
void ggml_mpi_graph_compute_pre(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers);
void ggml_mpi_graph_compute_post(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers);
#ifdef __cplusplus
}
#endif

5
ggml-opencl.cpp
View file

@ -1835,7 +1835,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
}
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
int64_t i12 = i02 * r2;
int64_t e12 = i12 + r2;
events.reserve(e12 - i12);
for (; i12 < e12; i12++) {
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();

4117
ggml-quants.c
View file

File diff suppressed because it is too large Load diff

229
ggml-rpc.cpp
View file

@ -56,6 +56,7 @@ struct socket_t {
};
// ggml_tensor is serialized into rpc_tensor
#pragma pack(push, 1)
struct rpc_tensor {
uint64_t id;
uint32_t type;
@ -71,6 +72,7 @@ struct rpc_tensor {
uint64_t data;
char name[GGML_MAX_NAME];
};
#pragma pack(pop)
// RPC commands
enum rpc_cmd {
@ -340,23 +342,6 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
return result;
}
static ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
result->nb[i] = tensor->nb[i];
}
result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
result->op = (ggml_op) tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
result->op_params[i] = tensor->op_params[i];
}
result->flags = tensor->flags;
result->data = reinterpret_cast<void *>(tensor->data);
ggml_set_name(result, tensor->name);
return result;
}
GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
UNUSED(buffer);
if (ggml_is_quantized(tensor->type)) {
@ -465,13 +450,15 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer
memcpy(&remote_ptr, output.data(), sizeof(remote_ptr));
size_t remote_size;
memcpy(&remote_size, output.data() + sizeof(uint64_t), sizeof(remote_size));
ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
ggml_backend_rpc_buffer_interface,
new ggml_backend_rpc_buffer_context{buft_ctx->sock, {}, remote_ptr, "RPC"},
remote_size);
return buffer;
if (remote_ptr != 0) {
ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
ggml_backend_rpc_buffer_interface,
new ggml_backend_rpc_buffer_context{buft_ctx->sock, {}, remote_ptr, "RPC"},
remote_size);
return buffer;
} else {
return nullptr;
}
}
static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
@ -658,7 +645,7 @@ GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
}
}
#endif
GGML_PRINT_DEBUG("Connecting to %s\n", endpoint);
fprintf(stderr, "Connecting to %s\n", endpoint);
std::string host;
int port;
if (!parse_endpoint(endpoint, host, port)) {
@ -731,22 +718,61 @@ GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint
// RPC server-side implementation
static void rpc_alloc_buffer(ggml_backend_t backend, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
class rpc_server {
public:
rpc_server(ggml_backend_t backend) : backend(backend) {}
~rpc_server();
bool alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
void get_alignment(std::vector<uint8_t> & output);
void get_max_size(std::vector<uint8_t> & output);
bool buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool free_buffer(const std::vector<uint8_t> & input);
bool buffer_clear(const std::vector<uint8_t> & input);
bool set_tensor(const std::vector<uint8_t> & input);
bool get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
private:
ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
ggml_tensor * create_node(uint64_t id,
struct ggml_context * ctx,
const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map);
ggml_backend_t backend;
std::unordered_set<ggml_backend_buffer_t> buffers;
};
bool rpc_server::alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | size (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t size;
memcpy(&size, input.data(), sizeof(size));
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
uint64_t remote_ptr = reinterpret_cast<uint64_t>(buffer);
uint64_t remote_size = buffer->size;
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, size, remote_ptr, remote_size);
uint64_t remote_ptr = 0;
uint64_t remote_size = 0;
if (buffer != nullptr) {
remote_ptr = reinterpret_cast<uint64_t>(buffer);
remote_size = buffer->size;
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, size, remote_ptr, remote_size);
buffers.insert(buffer);
} else {
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, size);
}
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
output.resize(2*sizeof(uint64_t), 0);
memcpy(output.data(), &remote_ptr, sizeof(remote_ptr));
memcpy(output.data() + sizeof(uint64_t), &remote_size, sizeof(remote_size));
return true;
}
static void rpc_get_alignment(ggml_backend_t backend, std::vector<uint8_t> & output) {
void rpc_server::get_alignment(std::vector<uint8_t> & output) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t alignment = ggml_backend_buft_get_alignment(buft);
GGML_PRINT_DEBUG("[%s] alignment: %lu\n", __func__, alignment);
@ -755,7 +781,7 @@ static void rpc_get_alignment(ggml_backend_t backend, std::vector<uint8_t> & out
memcpy(output.data(), &alignment, sizeof(alignment));
}
static void rpc_get_max_size(ggml_backend_t backend, std::vector<uint8_t> & output) {
void rpc_server::get_max_size(std::vector<uint8_t> & output) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t max_size = ggml_backend_buft_get_max_size(buft);
GGML_PRINT_DEBUG("[%s] max_size: %lu\n", __func__, max_size);
@ -764,41 +790,90 @@ static void rpc_get_max_size(ggml_backend_t backend, std::vector<uint8_t> & outp
memcpy(output.data(), &max_size, sizeof(max_size));
}
static void rpc_buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
bool rpc_server::buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
void * base = ggml_backend_buffer_get_base(buffer);
// output serialization format: | base_ptr (8 bytes) |
uint64_t base_ptr = reinterpret_cast<uint64_t>(base);
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &base_ptr, sizeof(base_ptr));
return true;
}
static void rpc_free_buffer(const std::vector<uint8_t> & input) {
bool rpc_server::free_buffer(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
ggml_backend_buffer_free(buffer);
buffers.erase(buffer);
return true;
}
static void rpc_buffer_clear(const std::vector<uint8_t> & input) {
bool rpc_server::buffer_clear(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) | value (1 byte) |
if (input.size() != sizeof(uint64_t) + sizeof(uint8_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
uint8_t value;
memcpy(&value, input.data() + sizeof(uint64_t), sizeof(value));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, remote_ptr, value);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
ggml_backend_buffer_clear(buffer, value);
return true;
}
static void rpc_set_tensor(const std::vector<uint8_t> & input) {
ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
result->nb[i] = tensor->nb[i];
}
result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
return nullptr;
}
result->op = (ggml_op) tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
result->op_params[i] = tensor->op_params[i];
}
result->flags = tensor->flags;
result->data = reinterpret_cast<void *>(tensor->data);
ggml_set_name(result, tensor->name);
return result;
}
bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
// serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
if (input.size() < sizeof(rpc_tensor) + sizeof(uint64_t)) {
return false;
}
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
@ -811,14 +886,23 @@ static void rpc_set_tensor(const std::vector<uint8_t> & input) {
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
if (tensor == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
ggml_backend_tensor_set(tensor, data, offset, size);
ggml_free(ctx);
return true;
}
static void rpc_get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
if (input.size() != sizeof(rpc_tensor) + 2*sizeof(uint64_t)) {
return false;
}
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
@ -832,15 +916,24 @@ static void rpc_get_tensor(const std::vector<uint8_t> & input, std::vector<uint8
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
if (tensor == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
// output serialization format: | data (size bytes) |
output.resize(size, 0);
ggml_backend_tensor_get(tensor, output.data(), offset, size);
ggml_free(ctx);
return true;
}
static void rpc_copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
bool rpc_server::copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor src | rpc_tensor dst |
if (input.size() != 2*sizeof(rpc_tensor)) {
return false;
}
const rpc_tensor * rpc_src = (const rpc_tensor *)input.data();
const rpc_tensor * rpc_dst = (const rpc_tensor *)(input.data() + sizeof(rpc_src));
@ -852,18 +945,24 @@ static void rpc_copy_tensor(const std::vector<uint8_t> & input, std::vector<uint
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * src = deserialize_tensor(ctx, rpc_src);
ggml_tensor * dst = deserialize_tensor(ctx, rpc_dst);
if (src == nullptr || dst == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensors\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
bool result = ggml_backend_buffer_copy_tensor(src, dst);
// output serialization format: | result (1 byte) |
output.resize(1, 0);
output[0] = result;
ggml_free(ctx);
return true;
}
static struct ggml_tensor * create_node(uint64_t id,
struct ggml_context * ctx,
const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
ggml_tensor * rpc_server::create_node(uint64_t id,
struct ggml_context * ctx,
const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
if (id == 0) {
return nullptr;
}
@ -872,6 +971,9 @@ static struct ggml_tensor * create_node(uint64_t id,
}
const rpc_tensor * tensor = tensor_ptrs.at(id);
struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
if (result == nullptr) {
return nullptr;
}
tensor_map[id] = result;
for (int i = 0; i < GGML_MAX_SRC; i++) {
result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
@ -881,14 +983,23 @@ static struct ggml_tensor * create_node(uint64_t id,
return result;
}
static void rpc_graph_compute(ggml_backend_t backend, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format:
// | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
if (input.size() < sizeof(uint32_t)) {
return false;
}
uint32_t n_nodes;
memcpy(&n_nodes, input.data(), sizeof(n_nodes));
if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t)) {
return false;
}
const uint64_t * nodes = (const uint64_t *)(input.data() + sizeof(n_nodes));
uint32_t n_tensors;
memcpy(&n_tensors, input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t), sizeof(n_tensors));
if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t) + n_tensors*sizeof(rpc_tensor)) {
return false;
}
const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
@ -914,9 +1025,17 @@ static void rpc_graph_compute(ggml_backend_t backend, const std::vector<uint8_t>
output.resize(1, 0);
output[0] = status;
ggml_free(ctx);
return true;
}
rpc_server::~rpc_server() {
for (auto buffer : buffers) {
ggml_backend_buffer_free(buffer);
}
}
static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
rpc_server server(backend);
while (true) {
uint8_t cmd;
if (!recv_data(sockfd, &cmd, 1)) {
@ -932,45 +1051,46 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
if (!recv_data(sockfd, input.data(), input_size)) {
break;
}
bool ok = true;
switch (cmd) {
case ALLOC_BUFFER: {
rpc_alloc_buffer(backend, input, output);
ok = server.alloc_buffer(input, output);
break;
}
case GET_ALIGNMENT: {
rpc_get_alignment(backend, output);
server.get_alignment(output);
break;
}
case GET_MAX_SIZE: {
rpc_get_max_size(backend, output);
server.get_max_size(output);
break;
}
case BUFFER_GET_BASE: {
rpc_buffer_get_base(input, output);
ok = server.buffer_get_base(input, output);
break;
}
case FREE_BUFFER: {
rpc_free_buffer(input);
ok = server.free_buffer(input);
break;
}
case BUFFER_CLEAR: {
rpc_buffer_clear(input);
ok = server.buffer_clear(input);
break;
}
case SET_TENSOR: {
rpc_set_tensor(input);
ok = server.set_tensor(input);
break;
}
case GET_TENSOR: {
rpc_get_tensor(input, output);
ok = server.get_tensor(input, output);
break;
}
case COPY_TENSOR: {
rpc_copy_tensor(input, output);
ok = server.copy_tensor(input, output);
break;
}
case GRAPH_COMPUTE: {
rpc_graph_compute(backend, input, output);
ok = server.graph_compute(input, output);
break;
}
case GET_DEVICE_MEMORY: {
@ -982,9 +1102,12 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
}
default: {
fprintf(stderr, "Unknown command: %d\n", cmd);
return;
ok = false;
}
}
if (!ok) {
break;
}
uint64_t output_size = output.size();
if (!send_data(sockfd, &output_size, sizeof(output_size))) {
break;

540
ggml-sycl.cpp
View file

@ -3847,21 +3847,27 @@ static void concat_f32(const float *x,const float *y, float *dst, const int ne
}
}
static void upscale_f32(const float *x, float *dst, const int ne00, const int nb02, const int scale_factor,
const sycl::nd_item<3> &item_ct1) {
int ne0 = ne00 * scale_factor;
int nidx = item_ct1.get_local_id(2) +
item_ct1.get_group(2) * item_ct1.get_local_range(2);
if (nidx >= ne0) {
static void upscale_f32(const float *x, float *dst, const int nb00, const int nb01,
const int nb02, const int nb03, const int ne10, const int ne11,
const int ne12, const int ne13, const float sf0, const float sf1,
const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) {
int index = item_ct1.get_local_id(0) +
item_ct1.get_group(0) * item_ct1.get_local_range(0);
if (index >= ne10 * ne11 * ne12 * ne13) {
return;
}
// operation
int i00 = nidx / scale_factor;
int i01 = item_ct1.get_group(1) / scale_factor;
int offset_src = i00 + i01 * ne00 + item_ct1.get_group(0) * nb02;
int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
dst[offset_dst] = x[offset_src];
int i10 = index % ne10;
int i11 = (index / ne10) % ne11;
int i12 = (index / (ne10 * ne11)) % ne12;
int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
int i00 = i10 / sf0;
int i01 = i11 / sf1;
int i02 = i12 / sf2;
int i03 = i13 / sf3;
dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
}
static void pad_f32(const float *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
@ -4191,7 +4197,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri
const block_q2_K * x = (const block_q2_K *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int n = tid/32;
const int l = tid - 32*n;
const int is = 8*n + l/16;
@ -4205,18 +4210,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri
y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
#else
const int is = tid/16; // 0 or 1
const int il = tid%16; // 0...15
const uint8_t q = x[i].qs[il] >> (2*is);
dst_t * y = yy + i*QK_K + 16*is + il;
float dall = x[i].dm[0];
float dmin = x[i].dm[1];
y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
#endif
}
template<typename dst_t>
@ -4226,7 +4219,6 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri
const int i = item_ct1.get_group(2);
const block_q3_K * x = (const block_q3_K *) vx;
#if QK_K == 256
const int r = item_ct1.get_local_id(2) / 4;
const int tid = r/2;
const int is0 = r%2;
@ -4250,31 +4242,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri
const uint8_t * hm = x[i].hmask;
for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
#else
const int tid = item_ct1.get_local_id(2);
const int is = tid/16; // 0 or 1
const int il = tid%16; // 0...15
const int im = il/8; // 0...1
const int in = il%8; // 0...7
dst_t * y = yy + i*QK_K + 16*is + il;
const uint8_t q = x[i].qs[il] >> (2*is);
const uint8_t h = x[i].hmask[in] >> (2*is + im);
const float d = (float)x[i].d;
if (is == 0) {
y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
} else {
y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
}
#endif
}
#if QK_K == 256
static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
if (j < 4) {
d = q[j] & 63; m = q[j + 4] & 63;
@ -4283,7 +4252,6 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8
m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
}
}
#endif
template<typename dst_t>
static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
@ -4292,7 +4260,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
const int i = item_ct1.get_group(2);
#if QK_K == 256
// assume 32 threads
const int tid = item_ct1.get_local_id(2);
const int il = tid/8;
@ -4316,15 +4283,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
y[l + 0] = d1 * (q[l] & 0xF) - m1;
y[l +32] = d2 * (q[l] >> 4) - m2;
}
#else
const int tid = item_ct1.get_local_id(2);
const uint8_t * q = x[i].qs;
dst_t * y = yy + i*QK_K;
const float d = (float)x[i].dm[0];
const float m = (float)x[i].dm[1];
y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4);
#endif
}
template<typename dst_t>
@ -4334,7 +4292,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri
const int i = item_ct1.get_group(2);
#if QK_K == 256
// assume 64 threads - this is very slightly better than the one below
const int tid = item_ct1.get_local_id(2);
const int il = tid/16; // il is in 0...3
@ -4361,18 +4318,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri
hm <<= 1;
y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2;
y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2;
#else
const int tid = item_ct1.get_local_id(2);
const uint8_t q = x[i].qs[tid];
const int im = tid/8; // 0...3
const int in = tid%8; // 0...7
const int is = tid/16; // 0 or 1
const uint8_t h = x[i].qh[in] >> im;
const float d = x[i].d;
dst_t * y = yy + i*QK_K + tid;
y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16));
#endif
}
template<typename dst_t>
@ -4381,7 +4326,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
const block_q6_K * x = (const block_q6_K *) vx;
const int i = item_ct1.get_group(2);
#if QK_K == 256
// assume 64 threads - this is very slightly better than the one below
const int tid = item_ct1.get_local_id(2);
@ -4401,24 +4345,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
#else
// assume 32 threads
const int tid = item_ct1.get_local_id(2);
const int ip = tid/16; // 0 or 1
const int il = tid - 16*ip; // 0...15
dst_t * y = yy + i*QK_K + 16*ip + il;
const float d = x[i].d;
const uint8_t ql = x[i].ql[16*ip + il];
const uint8_t qh = x[i].qh[il] >> (2*ip);
const int8_t * sc = x[i].scales;
y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32);
#endif
}
template<typename dst_t>
@ -4432,7 +4358,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res
const block_iq2_xxs * x = (const block_iq2_xxs *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4443,10 +4368,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f);
#else
assert(false);
#endif
}
template<typename dst_t>
@ -4460,7 +4381,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest
const block_iq2_xs * x = (const block_iq2_xs *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4469,10 +4389,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
#else
assert(false);
#endif
}
template <typename dst_t>
@ -4484,7 +4400,6 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
const block_iq2_s * x = (const block_iq2_s *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4492,13 +4407,9 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
#pragma unroll
for (int j = 0; j < 8; ++j)
for (int j = 0; j < 8; ++j) {
y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
#else
assert(false);
#endif
}
}
template<typename dst_t>
@ -4512,7 +4423,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res
const block_iq3_xxs * x = (const block_iq3_xxs *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4527,10 +4437,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
#else
assert(false);
#endif
}
template <typename dst_t>
@ -4543,7 +4449,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
const block_iq3_s * x = (const block_iq3_s *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4557,10 +4462,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
#else
assert(false);
#endif
}
template <typename dst_t>
@ -4573,7 +4474,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
const block_iq1_s * x = (const block_iq1_s *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4587,10 +4487,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
#else
assert(false);
#endif
}
template <typename dst_t>
@ -4603,7 +4499,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
const block_iq1_m * x = (const block_iq1_m *) vx;
const int tid = item_ct1.get_local_id(2);
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@ -4621,10 +4516,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
#else
assert(false);
#endif
}
template <typename dst_t>
@ -4698,7 +4589,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const int tid =
item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15
const int ix =
@ -4749,42 +4639,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx,
tmp += dall * sum1 - dmin * sum2;
}
#else
const int tid = item_ct1.get_local_id(2) /
(2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7
const int ix = item_ct1.get_local_id(2) %
(2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3
const int offset = tid * K_QUANTS_PER_ITERATION;
uint32_t uaux[2];
const uint8_t * d = (const uint8_t *)uaux;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + offset;
const uint8_t * q = x[i].qs + offset;
const uint32_t * s = (const uint32_t *)x[i].scales;
uaux[0] = s[0] & 0x0f0f0f0f;
uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
const sycl::float2 dall =
x[i].dm.convert<float, sycl::rounding_mode::automatic>();
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
const uint8_t ql = q[l];
sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
+ y[l+16] * d[1] * ((ql >> 2) & 3)
+ y[l+32] * d[2] * ((ql >> 4) & 3)
+ y[l+48] * d[3] * ((ql >> 6) & 3);
sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
}
tmp += dall.x() * sum1 - dall.y() * sum2;
}
#endif
// sum up partial sums and write back result
#pragma unroll
@ -4822,8 +4676,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
@ -4876,34 +4728,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx,
tmp += d * sum;
}
#else
const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7
const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3
const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14
const int in = offset/8; // 0 or 1
const int im = offset%8; // 0...7
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + offset;
const uint8_t * q = x[i].qs + offset;
const uint8_t * s = x[i].scales;
const float dall = (float)x[i].d;
float sum = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
const uint8_t hl = x[i].hmask[im+l] >> in;
const uint8_t ql = q[l];
sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
+ y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
+ y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
+ y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
#pragma unroll
@ -4938,7 +4762,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
const block_q4_K * x = (const block_q4_K *)vx + ib0;
#if QK_K == 256
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
@ -5027,36 +4850,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
#endif
}
#else
const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15
const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
const int step = tid * K_QUANTS_PER_ITERATION;
uint16_t aux16[2];
const uint8_t * s = (const uint8_t *)aux16;
float tmp = 0;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const uint8_t * q = x[i].qs + step;
const float * y = yy + i*QK_K + step;
const uint16_t * a = (const uint16_t *)x[i].scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const float d = (float)x[i].dm[0];
const float m = (float)x[i].dm[1];
float sum = 0.f;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
+ y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
+ y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3])
+ y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]);
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
#pragma unroll
@ -5091,7 +4884,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx,
float tmp = 0; // partial sum for thread in warp
#if QK_K == 256
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
@ -5168,30 +4960,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx,
dmin * smin;
}
#else
const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15
const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
const int step = tid * K_QUANTS_PER_ITERATION;
const int im = step/8;
const int in = step%8;
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const uint8_t * q = x[i].qs + step;
const int8_t * s = x[i].scales;
const float * y = yy + i*QK_K + step;
const float d = x[i].d;
float sum = 0.f;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
const uint8_t h = x[i].qh[in+j] >> im;
sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
+ y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
+ y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16))
+ y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16));
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
@ -5218,8 +4986,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
const block_q6_K * x = (const block_q6_K *)vx + ib0;
#if QK_K == 256
const int tid =
item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix =
@ -5276,37 +5042,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
}
#else
const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...7
const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0...3
const int step = tid * K_QUANTS_PER_ITERATION;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + step;
const uint8_t * ql = x[i].ql + step;
const uint8_t * qh = x[i].qh + step;
const int8_t * s = x[i].scales;
const float d = x[i+0].d;
float sum = 0;
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
+ y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
+ y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32)
+ y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32);
}
tmp += sum;
}
#endif
// sum up partial sums and write back result
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
@ -6851,7 +6586,6 @@ static __dpct_inline__ float
vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
#ifndef GGML_QKK_64
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
int v[2];
@ -6893,52 +6627,6 @@ vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
}
return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
#else
#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
float sumf_d = 0.0f;
float sumf_m = 0.0f;
uint16_t aux16[2];
const uint8_t * s = (const uint8_t *)aux16;
const uint16_t * a = (const uint16_t *)bq4_K->scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const float dall = bq4_K->dm[0];
const float dmin = bq4_K->dm[1];
const float d8_1 = bq8_1[0].ds[0];
const float d8_2 = bq8_1[1].ds[1];
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * q4 = (const int *)bq4_K->qs + (iqs/2);
const int v1 = q4[0];
const int v2 = q4[4];
const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0));
const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0));
const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0));
sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
return dall * sumf_d - dmin * sumf_m;
#else
bad_arch();
#endif // __SYCL_ARCH__ >= VER_4VEC
#endif
}
template <int mmq_y>
@ -6997,11 +6685,7 @@ load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql,
const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
#if QK_K == 256
x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
#else
x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
#endif
}
#pragma unroll
@ -7044,7 +6728,6 @@ static __dpct_inline__ float
vec_dot_q5_K_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
#ifndef GGML_QKK_64
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
int vl[2];
@ -7086,48 +6769,6 @@ vec_dot_q5_K_q8_1(const void *__restrict__ vbq,
}
return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
#else
#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
const int8_t * s = bq5_K->scales;
const float d = bq5_K->d;
const float d8_1 = bq8_1[0].ds[0];
const float d8_2 = bq8_1[1].ds[1];
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * ql = (const int *)bq5_K->qs + (iqs/2);
const int vl1 = ql[0];
const int vl2 = ql[4];
const int step = 4 * (iqs/2); // 0, 4, 8, 12
const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
const int in = step%8; // 0, 4, 0, 4
const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1])
+ d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]);
return d * sumf_d;
#else
bad_arch();
#endif // __SYCL_ARCH__ >= VER_4VEC
#endif
}
template <int mmq_y>
@ -7199,9 +6840,7 @@ load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql,
const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
#if QK_K == 256
x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
#endif
}
#pragma unroll
@ -7381,7 +7020,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs,
const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs,
const uint8_t *kmask_iq2xs) {
#if QK_K == 256
const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
#if QR2_XXS == 8
@ -7422,10 +7060,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq,
}
return d * (sumi1 + sumi2);
#endif
#else
assert(false);
return 0.f;
#endif
}
static __dpct_inline__ float
@ -7434,7 +7068,6 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq,
const uint64_t *iq2xs_grid, const uint64_t *ksigns64) {
#if DPCT_COMPATIBILITY_TEMP >= \
MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
const int ib32 = iqs;
@ -7472,16 +7105,11 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq,
assert(false);
return 0.f;
#endif
#else
assert(false);
return 0.f;
#endif
}
static __dpct_inline__ float
vec_dot_iq2_s_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
#if QK_K == 256
const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
const int ib32 = iqs;
@ -7525,9 +7153,6 @@ vec_dot_iq2_s_q8_1(const void *__restrict__ vbq,
}
const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f;
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
#else
assert(false);
#endif
}
static __dpct_inline__ float
@ -7536,7 +7161,6 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) {
#if DPCT_COMPATIBILITY_TEMP >= \
MIN_CC_DP4A // lowest compute capability for integer intrinsics
#if QK_K == 256
const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
const int ib32 = iqs;
@ -7564,17 +7188,12 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
assert(false);
return 0.f;
#endif
#else
assert(false);
return 0.f;
#endif
}
static __dpct_inline__ float
vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs,
const uint32_t *iq3s_grid) {
#if QK_K == 256
const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
const int ib32 = iqs;
@ -7603,16 +7222,12 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
(1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) *
bq8_1[ib32].ds[0];
return d * sumi;
#else
assert(false);
#endif
}
static __dpct_inline__ float
vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs,
const uint32_t *iq1s_grid_gpu) {
#if QK_K == 256
const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
const int ib32 = iqs;
@ -7631,15 +7246,11 @@ vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
const float d = d1q * bq8_1[ib32].ds[0];
const float m = d1q * bq8_1[ib32].ds[1];
return d * sumi + m * delta;
#else
assert(false);
#endif
}
static __dpct_inline__ float
vec_dot_iq1_m_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
#if QK_K == 256
const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
const int ib32 = iqs;
@ -7664,9 +7275,6 @@ vec_dot_iq1_m_q8_1(const void *__restrict__ vbq,
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = (float)scale.f16 * bq8_1[ib32].ds[0];
return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
#else
assert(false);
#endif
}
static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4,
@ -7714,7 +7322,6 @@ static __dpct_inline__ float
vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq,
const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
#if QK_K == 256
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
@ -7732,9 +7339,6 @@ vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq,
sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2);
}
return d * (sumi1 + sumi2);
#else
assert(false);
#endif
}
template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x,
@ -10085,18 +9689,17 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst,
});
}
static void upscale_f32_sycl(const float *x, float *dst, const int ne00,
const int ne01, const int ne02,
const int scale_factor, dpct::queue_ptr stream) {
int ne0 = (ne00 * scale_factor);
int num_blocks = (ne0 + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
sycl::range<3> gridDim(ne02, (ne01 * scale_factor), num_blocks);
static void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01,
const int nb02, const int nb03, const int ne10, const int ne11,
const int ne12, const int ne13, const float sf0, const float sf1,
const float sf2, const float sf3, dpct::queue_ptr stream) {
int dst_size = ne10 * ne11 * ne12 * ne13;
int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<3>(gridDim *
sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE),
sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)),
[=](sycl::nd_item<3> item_ct1) {
upscale_f32(x, dst, ne00, ne00 * ne01, scale_factor, item_ct1);
sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
});
}
@ -10198,7 +9801,6 @@ template <typename dst_t>
static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k,
dpct::queue_ptr stream) {
const int nb = k / QK_K;
#if QK_K == 256
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
@ -10210,27 +9812,12 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k,
dequantize_block_q2_K(vx, y, item_ct1);
});
}
#else
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
sycl::range<3>(1, 1, 32),
sycl::range<3>(1, 1, 32)),
[=](sycl::nd_item<3> item_ct1) {
dequantize_block_q2_K(vx, y, item_ct1);
});
}
#endif
}
template <typename dst_t>
static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k,
dpct::queue_ptr stream) {
const int nb = k / QK_K;
#if QK_K == 256
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
@ -10242,19 +9829,6 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k,
dequantize_block_q3_K(vx, y, item_ct1);
});
}
#else
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
sycl::range<3>(1, 1, 32),
sycl::range<3>(1, 1, 32)),
[=](sycl::nd_item<3> item_ct1) {
dequantize_block_q3_K(vx, y, item_ct1);
});
}
#endif
}
template <typename dst_t>
@ -10315,7 +9889,6 @@ template <typename dst_t>
static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k,
dpct::queue_ptr stream) {
const int nb = k / QK_K;
#if QK_K == 256
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
@ -10327,27 +9900,12 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k,
dequantize_block_q5_K(vx, y, item_ct1);
});
}
#else
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
sycl::range<3>(1, 1, 32),
sycl::range<3>(1, 1, 32)),
[=](sycl::nd_item<3> item_ct1) {
dequantize_block_q5_K(vx, y, item_ct1);
});
}
#endif
}
template <typename dst_t>
static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k,
dpct::queue_ptr stream) {
const int nb = k / QK_K;
#if QK_K == 256
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
@ -10359,20 +9917,6 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k,
dequantize_block_q6_K(vx, y, item_ct1);
});
}
#else
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
sycl::range<3>(1, 1, 32),
sycl::range<3>(1, 1, 32)),
[=](sycl::nd_item<3> item_ct1) {
dequantize_block_q6_K(vx, y, item_ct1);
});
}
#endif
}
template <typename dst_t>
@ -10524,9 +10068,6 @@ template <typename dst_t>
static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k,
dpct::queue_ptr stream) {
const int nb = (k + QK_K - 1) / QK_K;
#if QK_K == 64
dequantize_row_iq4_nl_sycl(vx, y, k, stream);
#else
{
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
@ -10541,7 +10082,6 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k,
});
});
}
#endif
}
@ -12046,8 +11586,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
const int nrows_y, const int nrows_dst,
dpct::queue_ptr stream) try {
#if QK_K == 256
int id;
SYCL_CHECK(
CHECK_TRY_ERROR(id = get_current_device_id()));
@ -12162,7 +11700,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
});
}
}
#endif
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@ -13985,15 +13522,15 @@ inline void ggml_sycl_op_upscale(const ggml_tensor *src0,
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
#pragma message("TODO: generalize upscale operator")
#pragma message(" https://github.com/ggerganov/ggml/pull/814")
GGML_ASSERT(false && "TODO: generalize upscale operator");
const float sf0 = (float)dst->ne[0]/src0->ne[0];
const float sf1 = (float)dst->ne[1]/src0->ne[1];
const float sf2 = (float)dst->ne[2]/src0->ne[2];
const float sf3 = (float)dst->ne[3]/src0->ne[3];
const int scale_factor = dst->op_params[0];
upscale_f32_sycl(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
upscale_f32_sycl(src0_dd, dst_dd, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
main_stream);
(void) src1;
(void) dst;
@ -14449,6 +13986,9 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const dpct::queue_ptr &main_stream) {
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);

15596
ggml-vulkan-shaders.hpp
View file

File diff suppressed because it is too large Load diff

431
ggml-vulkan.cpp
View file

File diff suppressed because it is too large Load diff

1128
ggml.c
View file

File diff suppressed because it is too large Load diff

70
ggml.h
View file

@ -481,9 +481,7 @@ extern "C" {
GGML_OP_ARGSORT,
GGML_OP_LEAKY_RELU,
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_ATTN_EXT,
GGML_OP_FLASH_FF,
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_SSM_CONV,
GGML_OP_SSM_SCAN,
@ -1460,11 +1458,12 @@ extern "C" {
struct ggml_tensor * b);
// rotary position embedding
// if mode & 1 == 1, skip n_past elements (DEPRECATED)
// if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
// if mode & 2 == 1, GPT-NeoX style
// if mode & 4 == 1, ChatGLM style
//
// b is an int32 vector with size a->ne[2], it contains the positions
// c is freq factors (e.g. phi3-128k), (optional)
GGML_API struct ggml_tensor * ggml_rope(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1483,10 +1482,11 @@ extern "C" {
int n_ctx);
// custom RoPE
GGML_API struct ggml_tensor * ggml_rope_custom(
GGML_API struct ggml_tensor * ggml_rope_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
@ -1499,7 +1499,23 @@ extern "C" {
float beta_slow);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
@ -1512,20 +1528,28 @@ extern "C" {
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
float beta_slow),
"use ggml_rope_ext instead");
// compute correction dims for YaRN RoPE scaling
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// xPos RoPE, in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
float base,
bool down);
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow),
"use ggml_rope_ext_inplace instead");
// compute correction dims for YaRN RoPE scaling
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// rotary position embedding backward, i.e compute dx from dy
// a - dy
@ -1533,6 +1557,7 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
@ -1734,13 +1759,6 @@ extern "C" {
struct ggml_tensor * a,
int k);
GGML_API struct ggml_tensor * ggml_flash_attn(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
bool masked);
#define GGML_KQ_MASK_PAD 32
// q: [n_embd, n_batch, n_head, 1]
@ -1761,6 +1779,7 @@ extern "C" {
struct ggml_tensor * a,
enum ggml_prec prec);
// TODO: needs to be adapted to ggml_flash_attn_ext
GGML_API struct ggml_tensor * ggml_flash_attn_back(
struct ggml_context * ctx,
struct ggml_tensor * q,
@ -1769,14 +1788,6 @@ extern "C" {
struct ggml_tensor * d,
bool masked);
GGML_API struct ggml_tensor * ggml_flash_ff(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b0,
struct ggml_tensor * b1,
struct ggml_tensor * c0,
struct ggml_tensor * c1);
GGML_API struct ggml_tensor * ggml_ssm_conv(
struct ggml_context * ctx,
struct ggml_tensor * s,
@ -2390,6 +2401,7 @@ extern "C" {
GGML_API int ggml_cpu_has_avx512 (void);
GGML_API int ggml_cpu_has_avx512_vbmi(void);
GGML_API int ggml_cpu_has_avx512_vnni(void);
GGML_API int ggml_cpu_has_avx512_bf16(void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_arm_fma (void);

102
ggml_vk_generate_shaders.py
View file

@ -2432,7 +2432,6 @@ layout (push_constant) uniform parameter
{
uint KX;
uint KY;
uint KZ;
float scale;
float max_bias;
float m0;
@ -2449,8 +2448,7 @@ layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) readonly buffer Z {C_TYPE data_c[];};
layout (binding = 3) buffer D {D_TYPE data_d[];};
layout (binding = 2) buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE vals[BLOCK_SIZE];
@ -2459,7 +2457,7 @@ void main() {
const uint rowx = gl_WorkGroupID.x;
const uint rowy = rowx % p.KY;
float slope = 0.0f;
float slope = 1.0f;
// ALiBi
if (p.max_bias > 0.0f) {
@ -2472,11 +2470,18 @@ void main() {
}
// Find max
vals[tid] = uintBitsToFloat(0xFF800000);
FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
vals[tid] = max(vals[tid], FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) + (p.KZ > 0 ? slope * FLOAT_TYPE(data_c[col]) : 0.0f));
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
}
vals[tid] = max_val;
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
@ -2486,15 +2491,21 @@ void main() {
barrier();
}
const FLOAT_TYPE max_val = vals[0];
max_val = vals[0];
barrier();
// Sum up values
vals[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
const uint i = rowx * p.KX + col;
const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
vals[tid] += val;
data_d[i] = D_TYPE(val);
}
@ -2509,7 +2520,13 @@ void main() {
const D_TYPE divisor = D_TYPE(vals[0]);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
data_d[rowx*p.KX + col] /= divisor;
}
}
@ -2592,7 +2609,8 @@ layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {int data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
layout (binding = 2) readonly buffer Z {float data_freq_factors[];};
layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
@ -2605,6 +2623,7 @@ layout (push_constant) uniform parameter {
float corr_dims[4];
float theta_scale;
float inv_ndims;
uint has_freq_facs;
} p;
float rope_yarn_ramp(const float low, const float high, const uint i0) {
@ -2654,7 +2673,8 @@ void main() {
const float cur_rot = p.inv_ndims * ic - ib;
const int pos = data_b[i2];
const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f);
const float freq_factor = p.has_freq_facs != 0 ? data_freq_factors[ic/2] : 1.0f;
const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f) / freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, uint(cur_rot), cos_theta, sin_theta);
@ -2672,20 +2692,26 @@ argsort_src = """
#extension GL_EXT_shader_16bit_storage : require
layout(local_size_x = 1024, local_size_y = 1, local_size_z = 1) in;
#define BLOCK_SIZE 1024
#define ASC 0
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
bool ascending;
uint ncols_pad;
uint order;
} p;
shared int dst_row[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = data_d[idx0];
data_d[idx0] = data_d[idx1];
data_d[idx1] = tmp;
int tmp = dst_row[idx0];
dst_row[idx0] = dst_row[idx1];
dst_row[idx1] = tmp;
}
void main() {
@ -2693,36 +2719,45 @@ void main() {
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
if (col >= p.ncols) {
if (col >= p.ncols_pad) {
return;
}
const uint a_idx = row * p.ncols;
const uint d_idx = row * p.ncols;
const uint row_offset = row * p.ncols;
// initialize indices
if (col < p.ncols) {
data_d[col] = col;
}
dst_row[col] = col;
barrier();
for (uint k = 2; k <= p.ncols; k *= 2) {
for (uint k = 2; k <= p.ncols_pad; k *= 2) {
for (uint j = k / 2; j > 0; j /= 2) {
const uint ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
if (p.ascending ? data_a[a_idx + data_d[d_idx + col]] > data_a[a_idx + data_d[d_idx + ixj]] : data_a[a_idx + data_d[d_idx + col]] < data_a[a_idx + data_d[d_idx + ixj]]) {
swap(d_idx + col, d_idx + ixj);
if (dst_row[col] >= p.ncols ||
(dst_row[ixj] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
} else {
if (p.ascending ? data_a[a_idx + data_d[d_idx + col]] < data_a[a_idx + data_d[d_idx + ixj]] : data_a[a_idx + data_d[d_idx + col]] > data_a[a_idx + data_d[d_idx + ixj]]) {
swap(d_idx + col, d_idx + ixj);
if (dst_row[ixj] >= p.ncols ||
(dst_row[col] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
}
}
barrier();
}
}
if (col < p.ncols) {
data_d[row_offset + col] = dst_row[col];
}
}
"""
@ -2845,13 +2880,16 @@ async def main():
stream.clear()
stream.extend((mulmat_head, shader_float_type, mulmat_body1, mulmat_load_scalar, mulmat_body2))
tasks.append(string_to_spv("matmul_f32", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_f16", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float16_t", "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float16_t", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_0_defines, mulmat_body1, mulmat_load_q4_0, mulmat_body2))

65
gguf-py/gguf/constants.py
View file

@ -63,13 +63,13 @@ class Keys:
KV_LORA_RANK = "{arch}.attention.kv_lora_rank"
class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count"
FREQ_BASE = "{arch}.rope.freq_base"
SCALING_TYPE = "{arch}.rope.scaling.type"
SCALING_FACTOR = "{arch}.rope.scaling.factor"
SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier"
DIMENSION_COUNT = "{arch}.rope.dimension_count"
FREQ_BASE = "{arch}.rope.freq_base"
SCALING_TYPE = "{arch}.rope.scaling.type"
SCALING_FACTOR = "{arch}.rope.scaling.factor"
SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor"
SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
class SSM:
CONV_KERNEL = "{arch}.ssm.conv_kernel"
@ -122,7 +122,6 @@ class MODEL_ARCH(IntEnum):
GPTNEOX = auto()
MPT = auto()
STARCODER = auto()
PERSIMMON = auto()
REFACT = auto()
BERT = auto()
NOMIC_BERT = auto()
@ -146,6 +145,7 @@ class MODEL_ARCH(IntEnum):
COMMAND_R = auto()
DBRX = auto()
OLMO = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
@ -157,6 +157,8 @@ class MODEL_TENSOR(IntEnum):
OUTPUT = auto()
OUTPUT_NORM = auto()
ROPE_FREQS = auto()
ROPE_FACTORS_LONG = auto()
ROPE_FACTORS_SHORT = auto()
ATTN_Q = auto()
ATTN_K = auto()
ATTN_V = auto()
@ -173,6 +175,7 @@ class MODEL_TENSOR(IntEnum):
FFN_DOWN = auto()
FFN_UP = auto()
FFN_ACT = auto()
FFN_NORM_EXP = auto()
FFN_GATE_EXP = auto()
FFN_DOWN_EXP = auto()
FFN_UP_EXP = auto()
@ -207,7 +210,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GPTNEOX: "gptneox",
MODEL_ARCH.MPT: "mpt",
MODEL_ARCH.STARCODER: "starcoder",
MODEL_ARCH.PERSIMMON: "persimmon",
MODEL_ARCH.REFACT: "refact",
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
@ -231,6 +233,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.ARCTIC: "arctic",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
}
@ -242,6 +245,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ROPE_FACTORS_LONG: "rope_factors_long",
MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
@ -263,6 +268,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.FFN_DOWN_SHEXP: "blk.{bid}.ffn_down_shexp",
MODEL_TENSOR.FFN_UP_SHEXP: "blk.{bid}.ffn_up_shexp",
MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn",
MODEL_TENSOR.FFN_NORM_EXP: "blk.{bid}.ffn_norm_exps",
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
@ -447,20 +453,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.PERSIMMON: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.REFACT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -764,6 +756,27 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.ARCTIC: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_NORM_EXP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.DEEPSEEK2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -804,9 +817,6 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.PERSIMMON: [
MODEL_TENSOR.ROPE_FREQS,
],
MODEL_ARCH.QWEN: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
@ -971,9 +981,8 @@ class GGUFValueType(IntEnum):
raise ValueError(f"Unknown type: {type(val)}")
# Note: Does not support GGML_QKK_64
QK_K = 256
# Items here are (block size, type size)
QK_K = 256
GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
GGMLQuantizationType.F32: (1, 4),
GGMLQuantizationType.F16: (1, 2),

3
gguf-py/gguf/gguf_writer.py
View file

@ -451,6 +451,9 @@ class GGUFWriter:
def add_rope_scaling_factor(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_attn_factors(self, value: Sequence[float]) -> None:
self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)

17
gguf-py/gguf/tensor_mapping.py
View file

@ -244,6 +244,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.fc11", # nomic-bert
"model.layers.{bid}.mlp.c_fc", # starcoder2
"encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2
"model.layers.{bid}.residual_mlp.w3", # arctic
),
MODEL_TENSOR.FFN_UP_EXP: (
@ -273,6 +274,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
"encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2
"transformer.h.{bid}.mlp.linear_1", # refact
"model.layers.{bid}.residual_mlp.w1", # arctic
),
MODEL_TENSOR.FFN_GATE_EXP: (
@ -308,6 +310,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.fc2", # nomic-bert
"model.layers.{bid}.mlp.c_proj", # starcoder2
"encoder.layer.{bid}.mlp.wo", # jina-bert-v2
"model.layers.{bid}.residual_mlp.w2", # arctic
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@ -409,6 +412,18 @@ class TensorNameMap:
),
}
# architecture-specific block mappings
arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
MODEL_ARCH.ARCTIC: {
MODEL_TENSOR.FFN_NORM: (
"model.layers.{bid}.residual_layernorm",
),
MODEL_TENSOR.FFN_NORM_EXP: (
"model.layers.{bid}.post_attention_layernorm",
),
},
}
mapping: dict[str, tuple[MODEL_TENSOR, str]]
def __init__(self, arch: MODEL_ARCH, n_blocks: int):
@ -420,6 +435,8 @@ class TensorNameMap:
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
self.mapping[key] = (tensor, tensor_name)
if arch in self.arch_block_mappings_cfg:
self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch])
for bid in range(n_blocks):
for tensor, keys in self.block_mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:

1396
llama.cpp
View file

File diff suppressed because it is too large Load diff

13
llama.h
View file

@ -81,9 +81,10 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_GPT2 = 7,
LLAMA_VOCAB_PRE_TYPE_REFACT = 8,
LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9,
LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10,
LLAMA_VOCAB_PRE_TYPE_OLMO = 11,
LLAMA_VOCAB_PRE_TYPE_DBRX = 12,
LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10,
LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11,
LLAMA_VOCAB_PRE_TYPE_OLMO = 12,
LLAMA_VOCAB_PRE_TYPE_DBRX = 13,
};
// note: these values should be synchronized with ggml_rope
@ -758,6 +759,12 @@ extern "C" {
// n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
// Get the number of threads used for generation of a single token.
LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
// Get the number of threads used for prompt and batch processing (multiple token).
LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
// Set whether to use causal attention or not
// If set to true, the model will only attend to the past tokens
LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);

1
requirements.txt
View file

@ -9,4 +9,3 @@
-r ./requirements/requirements-convert-hf-to-gguf.txt
-r ./requirements/requirements-convert-hf-to-gguf-update.txt
-r ./requirements/requirements-convert-llama-ggml-to-gguf.txt
-r ./requirements/requirements-convert-persimmon-to-gguf.txt

2
requirements/requirements-convert-persimmon-to-gguf.txt
View file

@ -1,2 +0,0 @@
-r ./requirements-convert.txt
torch~=2.1.1

5
scripts/LlamaConfig.cmake.in
View file

@ -5,7 +5,6 @@ set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(LLAMA_BLAS @LLAMA_BLAS@)
set(LLAMA_CUDA @LLAMA_CUDA@)
set(LLAMA_METAL @LLAMA_METAL@)
set(LLAMA_MPI @LLAMA_MPI@)
set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
@ -37,10 +36,6 @@ if (LLAMA_METAL)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
endif()
if (LLAMA_MPI)
find_package(MPI REQUIRED)
endif()
if (LLAMA_CLBLAST)
find_package(CLBlast REQUIRED)
endif()

166
scripts/gen-unicode-data.py
View file

@ -1,64 +1,134 @@
import regex
import ctypes
import unicodedata
def get_matches(regex_expr):
regex_expr_compiled = regex.compile(regex_expr)
unicode_ranges = []
current_range = None
for codepoint in range(0x110000):
char = chr(codepoint)
if regex_expr_compiled.match(char):
if current_range is None:
current_range = [codepoint, codepoint]
else:
current_range[1] = codepoint
elif current_range is not None:
unicode_ranges.append(tuple(current_range))
current_range = None
if current_range is not None:
unicode_ranges.append(tuple(current_range))
return unicode_ranges
class CoodepointFlags (ctypes.Structure):
_fields_ = [ # see definition in unicode.h
("is_undefined", ctypes.c_uint16, 1),
("is_number", ctypes.c_uint16, 1), # regex: \p{N}
("is_letter", ctypes.c_uint16, 1), # regex: \p{L}
("is_separator", ctypes.c_uint16, 1), # regex: \p{Z}
("is_accent_mark", ctypes.c_uint16, 1), # regex: \p{M}
("is_punctuation", ctypes.c_uint16, 1), # regex: \p{P}
("is_symbol", ctypes.c_uint16, 1), # regex: \p{S}
("is_control", ctypes.c_uint16, 1), # regex: \p{C}
]
def print_cat(mode, cat, ranges):
if mode == "range":
print("const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_{} = {{".format(cat)) # noqa: NP100
if mode == "map":
print("const std::map<uint32_t, uint32_t> unicode_map_{} = {{".format(cat)) # noqa: NP100
for i, values in enumerate(ranges):
end = ",\n" if (i % 4 == 3 or i + 1 == len(ranges)) else ", "
values = ["0x%08X" % value for value in values]
print("{" + ", ".join(values) + "}", end=end) # noqa: NP100
print("};") # noqa: NP100
print("") # noqa: NP100
assert (ctypes.sizeof(CoodepointFlags) == 2)
print_cat("range", "number", get_matches(r'\p{N}'))
print_cat("range", "letter", get_matches(r'\p{L}'))
print_cat("range", "separator", get_matches(r'\p{Z}'))
print_cat("range", "accent_mark", get_matches(r'\p{M}'))
print_cat("range", "punctuation", get_matches(r'\p{P}'))
print_cat("range", "symbol", get_matches(r'\p{S}'))
print_cat("range", "control", get_matches(r'\p{C}'))
MAX_CODEPOINTS = 0x110000
print_cat("range", "whitespace", get_matches(r'\s'))
regex_number = regex.compile(r'\p{N}')
regex_letter = regex.compile(r'\p{L}')
regex_separator = regex.compile(r'\p{Z}')
regex_accent_mark = regex.compile(r'\p{M}')
regex_punctuation = regex.compile(r'\p{P}')
regex_symbol = regex.compile(r'\p{S}')
regex_control = regex.compile(r'\p{C}')
regex_whitespace = regex.compile(r'\s')
codepoint_flags = (CoodepointFlags * MAX_CODEPOINTS)()
table_whitespace = []
table_lowercase = []
table_uppercase = []
table_nfd = []
map_lowercase = []
map_uppercase = []
for codepoint in range(0x110000):
for codepoint in range(MAX_CODEPOINTS):
# convert codepoint to unicode character
char = chr(codepoint)
# regex categories
flags = codepoint_flags[codepoint]
flags.is_number = bool(regex_number.match(char))
flags.is_letter = bool(regex_letter.match(char))
flags.is_separator = bool(regex_separator.match(char))
flags.is_accent_mark = bool(regex_accent_mark.match(char))
flags.is_punctuation = bool(regex_punctuation.match(char))
flags.is_symbol = bool(regex_symbol.match(char))
flags.is_control = bool(regex_control.match(char))
flags.is_undefined = bytes(flags)[0] == 0
assert (not flags.is_undefined)
# whitespaces
if bool(regex_whitespace.match(char)):
table_whitespace.append(codepoint)
# lowercase conversion
lower = ord(char.lower()[0])
upper = ord(char.upper()[0])
if codepoint != lower:
map_lowercase.append((codepoint, lower))
table_lowercase.append((codepoint, lower))
# uppercase conversion
upper = ord(char.upper()[0])
if codepoint != upper:
map_uppercase.append((codepoint, upper))
print_cat("map", "lowercase", map_lowercase)
print_cat("map", "uppercase", map_uppercase)
table_uppercase.append((codepoint, upper))
# NFD normalization
norm = ord(unicodedata.normalize('NFD', char)[0])
if codepoint != norm:
table_nfd.append((codepoint, norm))
# TODO: generate unicode_map_nfd
# group ranges with same flags
ranges_flags = [(0, codepoint_flags[0])] # start, flags
for codepoint, flags in enumerate(codepoint_flags):
if bytes(flags) != bytes(ranges_flags[-1][1]):
ranges_flags.append((codepoint, flags))
ranges_flags.append((MAX_CODEPOINTS, CoodepointFlags()))
# group ranges with same nfd
ranges_nfd = [(0, 0, 0)] # start, last, nfd
for codepoint, norm in table_nfd:
start = ranges_nfd[-1][0]
if ranges_nfd[-1] != (start, codepoint - 1, norm):
ranges_nfd.append(None)
start = codepoint
ranges_nfd[-1] = (start, codepoint, norm)
# Generate 'unicode-data.cpp'
def out(line=""):
print(line, end='\n') # noqa
out("""\
// generated with scripts/gen-unicode-data.py
#include "unicode-data.h"
#include <cstdint>
#include <vector>
#include <unordered_map>
#include <unordered_set>
""")
out("const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags = { // start, flags // last=next_start-1")
for codepoint, flags in ranges_flags:
flags = int.from_bytes(bytes(flags), "little")
out("{0x%06X, 0x%04X}," % (codepoint, flags))
out("};\n")
out("const std::unordered_set<uint32_t> unicode_set_whitespace = {")
out(", ".join("0x%06X" % cpt for cpt in table_whitespace))
out("};\n")
out("const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase = {")
for tuple in table_lowercase:
out("{0x%06X, 0x%06X}," % tuple)
out("};\n")
out("const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase = {")
for tuple in table_uppercase:
out("{0x%06X, 0x%06X}," % tuple)
out("};\n")
out("const std::vector<range_nfd> unicode_ranges_nfd = { // start, last, nfd")
for triple in ranges_nfd:
out("{0x%06X, 0x%06X, 0x%06X}," % triple)
out("};\n")

75
tests/test-backend-ops.cpp
View file

@ -16,6 +16,7 @@
#include <thread>
#include <vector>
static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
// static RNG initialization (revisit if n_threads stops being constant)
static const size_t n_threads = std::thread::hardware_concurrency();
@ -49,6 +50,22 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
t.join();
}
#if 0
const char * val_str = getenv("GGML_TEST_EPS");
float val = 1e-9f;
if (val_str != nullptr) {
val = std::stof(val_str);
printf("GGML_TEST_EPS=%e\n", val);
}
// test quantization with very small values that may result in nan scales due to division by zero
if (ggml_is_quantized(tensor->type)) {
for (int i = 0; i < 256; i++) {
data[i] = val;
}
}
#endif
if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
} else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
@ -64,6 +81,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
}
}
ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], im);
GGML_ASSERT(ggml_validate_row_data(tensor->type, dataq.data(), dataq.size()));
ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
} else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
// This is going to create some weird integers though.
@ -1124,20 +1142,22 @@ struct test_rope : public test_case {
int n_dims;
int mode;
int n_ctx;
bool ff;
std::string vars() override {
return VARS_TO_STR5(type, ne, n_dims, mode, n_ctx);
return VARS_TO_STR6(type, ne, n_dims, mode, n_ctx, ff);
}
test_rope(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 1},
int n_dims = 10, int mode = 0, int n_ctx = 512)
: type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx) {}
int n_dims = 10, int mode = 0, int n_ctx = 512, bool ff = false)
: type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx), ff(ff) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
ggml_tensor * out = ggml_rope(ctx, a, pos, n_dims, mode, n_ctx);
ggml_tensor * freq = ff ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2) : nullptr;
ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
return out;
}
@ -1151,7 +1171,12 @@ struct test_rope : public test_case {
}
ggml_backend_tensor_set(t, data.data(), 0, ne[2] * sizeof(int));
} else {
init_tensor_uniform(t);
if (t->ne[0] == n_dims/2) {
// frequency factors in the range [0.9f, 1.1f]
init_tensor_uniform(t, 0.9f, 1.1f);
} else {
init_tensor_uniform(t);
}
}
}
}
@ -1745,14 +1770,14 @@ struct test_llama : public test_llm {
struct ggml_tensor * Kcur = ggml_mul_mat(ctx, wk, cur);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx, wv, cur);
Qcur = ggml_rope_custom(
ctx, ggml_reshape_3d(ctx, Qcur, hp.n_embd_head, hp.n_head, hp.n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx, ggml_reshape_3d(ctx, Qcur, hp.n_embd_head, hp.n_head, hp.n_tokens), inp_pos, nullptr,
hp.n_rot, 0, 0, hp.n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_custom(
ctx, ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx, ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens), inp_pos, nullptr,
hp.n_rot, 0, 0, hp.n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -1871,13 +1896,13 @@ struct test_falcon : public test_llm {
Kcur = ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens);
// using mode = 2 for neox mode
Qcur = ggml_rope_custom(
ctx, Qcur, inp_pos, hp.n_rot, 2, 0, hp.n_orig_ctx,
Qcur = ggml_rope_ext(
ctx, Qcur, inp_pos, nullptr, hp.n_rot, 2, 0, hp.n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_custom(
ctx, Kcur, inp_pos, hp.n_rot, 2, 0, hp.n_orig_ctx,
Kcur = ggml_rope_ext(
ctx, Kcur, inp_pos, nullptr, hp.n_rot, 2, 0, hp.n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
@ -2170,16 +2195,20 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f));
for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512)); // llama 7B
test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512)); // llama 13B
test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512)); // llama 30B
test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512)); // llama 65B
test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512)); // neox (stablelm)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512)); // neox (phi-2)
// TODO: ff not supported yet for !neox
test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, false)); // llama 7B
test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, false)); // llama 13B
test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, false)); // llama 30B
test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, false)); // llama 65B
for (bool ff : {false, true}) { // freq_factors
test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, ff)); // neox (stablelm)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, ff)); // neox (phi-2)
}
}
test_cases.emplace_back(new test_concat(GGML_TYPE_F32));

20
tests/test-chat-template.cpp
View file

@ -49,8 +49,14 @@ int main(void) {
"{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}",
// Llama-3
"{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
// Phi-3
"{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + ' ' + message['content'] + '<|end|> ' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|> ' }}{% else %}{{ eos_token }}{% endif %}"
//Phi-3-mini
"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
//Phi-3-small
"{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
//Phi-3-medium
"{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
//Phi-3-vision
"{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}"
};
std::vector<std::string> expected_output = {
// teknium/OpenHermes-2.5-Mistral-7B
@ -79,8 +85,14 @@ int main(void) {
"<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>Hi there<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Who are you<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I am an assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Another question<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
// Llama 3
"<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi there<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI am an assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nAnother question<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
// Phi 3
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\nI am an assistant<|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
//Phi-3-mini
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n I am an assistant <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
//Phi-3-small
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n I am an assistant <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
//Phi-3-medium
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n I am an assistant <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
//Phi-3-vision
"<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n I am an assistant <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
};
std::vector<char> formatted_chat(1024);
int32_t res;

116
tests/test-grad0.cpp
View file

@ -1515,90 +1515,50 @@ int main(int argc, const char ** argv) {
}
// flash_attn f32
{
srand(seed);
const int nargs = 3;
// TODO: adapt to ggml_flash_attn_ext() changes
//{
// srand(seed);
// const int nargs = 3;
int64_t ne2[4];
// int64_t ne2[4];
get_random_dims(ne2, 4);
int64_t D = ne2[0];
int64_t N = ne2[1];
int64_t M = ne2[2] + N;
int64_t B = ne2[3];
// get_random_dims(ne2, 4);
// int64_t D = ne2[0];
// int64_t N = ne2[1];
// int64_t M = ne2[2] + N;
// int64_t B = ne2[3];
for (int masked = 0; masked <= 1; ++masked) {
for (int ndims = 2; ndims <= 4; ++ndims) {
int max_nrep = (ndims >= 3) ? 2 : 1;
for (int nrep = 1; nrep < max_nrep; ++nrep) {
int64_t neq[4] = { D, N, B*nrep, ne[3] };
int64_t nek[4] = { D, M, B, ne[3] };
int64_t nev[4] = { M, D, B, ne[3] };
if (ndims == 2) {
neq[2] = 1; neq[3] = 1;
nek[2] = 1; nek[3] = 1;
nev[2] = 1; nev[3] = 1;
} else if (ndims == 3) {
neq[3] = 1;
nek[3] = 1;
nev[3] = 1;
}
x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
ggml_set_param(ctx0, x[0]);
ggml_set_param(ctx0, x[1]);
ggml_set_param(ctx0, x[2]);
// for (int masked = 0; masked <= 1; ++masked) {
// for (int ndims = 2; ndims <= 4; ++ndims) {
// int max_nrep = (ndims >= 3) ? 2 : 1;
// for (int nrep = 1; nrep < max_nrep; ++nrep) {
// int64_t neq[4] = { D, N, B*nrep, ne[3] };
// int64_t nek[4] = { D, M, B, ne[3] };
// int64_t nev[4] = { M, D, B, ne[3] };
// if (ndims == 2) {
// neq[2] = 1; neq[3] = 1;
// nek[2] = 1; nek[3] = 1;
// nev[2] = 1; nev[3] = 1;
// } else if (ndims == 3) {
// neq[3] = 1;
// nek[3] = 1;
// nev[3] = 1;
// }
// x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
// x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
// x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
// ggml_set_param(ctx0, x[0]);
// ggml_set_param(ctx0, x[1]);
// ggml_set_param(ctx0, x[2]);
struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
// struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
}
}
}
}
// check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
// }
// }
// }
//}
// flash_attn f16, not yet fully implemented
if(0)
{
srand(seed);
const int nargs = 3;
int64_t ne2[4];
get_random_dims(ne2, 4);
int64_t D = ne2[0];
int64_t N = ne2[1];
int64_t M = ne2[2] + N;
int64_t B = ne2[3];
for (int masked = 0; masked <= 1; ++masked) {
for (int ndims = 2; ndims <= 4; ++ndims) {
int64_t neq[4] = { D, N, B, ne[3] };
int64_t nek[4] = { D, M, B, ne[3] };
int64_t nev[4] = { M, D, B, ne[3] };
if (ndims == 2) {
neq[2] = 1; neq[3] = 1;
nek[2] = 1; nek[3] = 1;
nev[2] = 1; nev[3] = 1;
} else if (ndims == 3) {
neq[3] = 1;
nek[3] = 1;
nev[3] = 1;
}
x[0] = get_random_tensor_f16(ctx0, ndims, neq, -0.1250f, 0.1250f);
x[1] = get_random_tensor_f16(ctx0, ndims, nek, -0.1250f, 0.1250f);
x[2] = get_random_tensor_f16(ctx0, ndims, nev, -0.1250f, 0.1250f);
ggml_set_param(ctx0, x[0]);
ggml_set_param(ctx0, x[1]);
ggml_set_param(ctx0, x[2]);
struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
}
}
}
ggml_free(ctx0);
}

9
tests/test-tokenizer-0.sh
View file

@ -17,10 +17,15 @@ make -j tests/test-tokenizer-0
printf "Testing %s on %s ...\n" $name $input
python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
set -e
printf "Tokenizing using (py) Python AutoTokenizer ...\n"
python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
printf "Tokenizing using (cpp) llama.cpp ...\n"
./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"
diff $input.tok $input.tokcpp > /dev/null 2>&1

146
tests/test-tokenizer-random.py
View file

@ -1,5 +1,5 @@
# Test libllama tokenizer == AutoTokenizer.
# Brute force random tokens/text generation.
# Brute force random words/text generation.
#
# Sample usage:
#
@ -12,10 +12,10 @@ import argparse
import subprocess
import random
from typing import Iterator
from typing import Callable, Iterator
import cffi
from transformers import AutoTokenizer, PreTrainedTokenizerBase
from transformers import AutoTokenizer
logger = logging.getLogger("test-tokenizer-random-bpe")
@ -145,28 +145,50 @@ def generator_custom_text() -> Iterator[str]:
def generator_custom_text_edge_cases() -> Iterator[str]:
"""Edge cases found while debugging"""
yield from [
'\x1f-a', # unicode_ranges_control, {0x00001C, 0x00001F}
'¼-a', # unicode_ranges_digit, 0x00BC
'½-a', # unicode_ranges_digit, 0x00BD
'¾-a', # unicode_ranges_digit, 0x00BE
'a 〇b', # unicode_ranges_digit, 0x3007
'Ⅵ-a', # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
'\uFEFF//', # unicode_ranges_control, 0xFEFF (BOM)
'<s>a' # TODO: Phi-3 fail
'\x1f-a', # unicode_ranges_control, {0x00001C, 0x00001F}
'¼-a', # unicode_ranges_digit, 0x00BC
'½-a', # unicode_ranges_digit, 0x00BD
'¾-a', # unicode_ranges_digit, 0x00BE
'a 〇b', # unicode_ranges_digit, 0x3007
'Ⅵ-a', # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
'\uFEFF//', # unicode_ranges_control, 0xFEFF (BOM)
'Cửa Việt', # llama-3, ignore_merges = true
'<s>a', # Phi-3 fail
'<unk><|endoftext|><s>', # Phi-3 fail
'a\na', # TODO: Bert fail
]
def generator_random_chars(iterations = 100) -> Iterator[str]:
def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]:
special_tokens = set(tokenizer.all_special_tokens)
special_tokens.update([" ", "\n", "\t", "-", "!", "one", "1", "<s>", "</s>"])
special_tokens = list(sorted(special_tokens))
rand = random.Random()
for m in range(iterations):
rand.seed(m)
words = rand.choices(special_tokens, k=500)
if tokenizer.add_bos_token: # skip spam warning of double BOS
while words and words[0] == tokenizer.bos_token:
words.pop(0)
yield "".join(words)
def generator_vocab_words(vocab: list[str]) -> Iterator[str]:
"""Brute force check all vocab words"""
yield from vocab
def generator_random_chars(iterations=100) -> Iterator[str]:
"""Brute force random text with simple characters"""
WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
CHARS = list(set("""
CHARS = list(sorted(set("""
ABCDEFGHIJKLMNOPQRSTUVWXYZ
abcdefghijklmnopqrstuvwxyz
ÁÉÍÓÚÀÈÌÒÙÂÊÎÔÛÄËÏÖÜ
áéíóúàèìòùâêîôûäëïöü
.-,*/-+ª!"·$%&/()=?¿[]{}<>\\|@#~½¬~;:_
"""))
""")))
rand = random.Random()
for m in range(iterations):
@ -181,13 +203,13 @@ def generator_random_chars(iterations = 100) -> Iterator[str]:
yield "".join(text)
def generator_random_vocab_chars(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
def generator_random_vocab_chars(vocab: list[str], iterations=100) -> Iterator[str]:
"""Brute force random text with vocab characters"""
vocab_ids = list(tokenizer.vocab.values())
vocab_text = tokenizer.decode(vocab_ids, skip_special_tokens=True)
vocab_chars = list(set(vocab_text))
del vocab_ids, vocab_text
vocab_chars = set()
for word in vocab:
vocab_chars.update(word)
vocab_chars = list(sorted(vocab_chars))
rand = random.Random()
for m in range(iterations):
@ -196,19 +218,11 @@ def generator_random_vocab_chars(tokenizer: PreTrainedTokenizerBase, iterations
yield "".join(text)
def generator_random_vocab_tokens(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
"""Brute force random text from vocab tokens"""
def generator_random_vocab_words(vocab: list[str], iterations=100) -> Iterator[str]:
"""Brute force random text from vocab words"""
space_id = tokenizer.encode(" ", add_special_tokens=False)[0]
vocab_ids = list(tokenizer.vocab.values())
vocab_ids = list(sorted(vocab_ids + vocab_ids))
for i in range(1, len(vocab_ids), 2):
vocab_ids[i] = space_id
vocab_tokens = tokenizer.decode(vocab_ids, skip_special_tokens=True)
vocab_tokens = vocab_tokens.split(" ")
del vocab_ids
yield from vocab_tokens
vocab = [w.strip() for w in vocab]
yield from vocab
rand = random.Random()
for m in range(iterations):
@ -217,14 +231,13 @@ def generator_random_vocab_tokens(tokenizer: PreTrainedTokenizerBase, iterations
num_words = rand.randint(300, 400)
for i in range(num_words):
k = rand.randint(1, 3)
tokens = rand.choices(vocab_tokens, k=k)
tokens = [t.strip(" \n\r\t") for t in tokens]
words = rand.choices(vocab, k=k)
sep = rand.choice(" \n\r\t")
text.append("".join(tokens) + sep)
text.append("".join(words) + sep)
yield "".join(text)
def generator_random_bytes(iterations = 100) -> Iterator[str]:
def generator_random_bytes(iterations=100) -> Iterator[str]:
"""Brute force random bytes"""
WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
@ -242,10 +255,10 @@ def generator_random_bytes(iterations = 100) -> Iterator[str]:
yield "".join(text)
def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerBase, generator: Iterator[str]):
def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]):
def find_first_mismatch(ids1: list[int], ids2: list[int]):
for i, (a,b) in enumerate(zip(ids1, ids2)):
for i, (a, b) in enumerate(zip(ids1, ids2)):
if a != b:
return i
if len(ids1) == len(ids2):
@ -255,15 +268,12 @@ def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerB
t0 = time.perf_counter()
logger.info("%s: %s" % (generator.__name__, "ini"))
for text in generator:
ids1 = model.tokenize(text, add_special=False, parse_special=False)
ids2 = tokenizer.encode(text, add_special_tokens=False)
ids1 = func_tokenize1(text)
ids2 = func_tokenize2(text)
if ids1 != ids2:
i = find_first_mismatch(ids1, ids2)
ids1 = list(ids1)[max(0, i - 2) : i + 2 + 1]
ids2 = list(ids2)[max(0, i - 2) : i + 2 + 1]
text2 = tokenizer.decode(ids2, skip_special_tokens=True)
assert (text2 in text)
logger.info(" Text: " + repr(text2))
logger.info(" TokenIDs: " + str(ids1))
logger.info(" Expected: " + str(ids2))
raise Exception()
@ -271,25 +281,55 @@ def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerB
logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0))
if __name__ == "__main__":
def main(argv: list[str] = None):
parser = argparse.ArgumentParser()
parser.add_argument("vocab_file", help="path to vocab 'gguf' file")
parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args()
args = parser.parse_args(argv)
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=2048))
model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
test_compare_tokenizer(model, tokenizer, generator_custom_text())
test_compare_tokenizer(model, tokenizer, generator_custom_text_edge_cases())
test_compare_tokenizer(model, tokenizer, generator_random_chars(10_000))
test_compare_tokenizer(model, tokenizer, generator_random_vocab_chars(tokenizer, 10_000))
test_compare_tokenizer(model, tokenizer, generator_random_vocab_tokens(tokenizer, 10_000))
# test_compare_tokenizer(model, tokenizer, generator_random_bytes(10_000)) # FAIL
tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", True)
tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", False)
def func_tokenize1(text: str):
return model.tokenize(text, add_special=True, parse_special=True)
def func_tokenize2(text: str):
return tokenizer.encode(text, add_special_tokens=True)
vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_special_tokens(tokenizer, 10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000))
# test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_bytes(10_000)) # FAIL
model.free()
if __name__ == "__main__":
# main()
path_tokenizers = "./models/tokenizers/"
path_vocab_format = "./models/ggml-vocab-%s.gguf"
# import os
# tokenizers = os.listdir(path_tokenizers)
tokenizers = [
"llama-spm", # SPM
"phi-3", # SPM
]
for tokenizer in tokenizers:
print("\n" + "=" * 50 + "\n" + tokenizer + "\n") # noqa
vocab_file = path_vocab_format % tokenizer
dir_tokenizer = path_tokenizers + "/" + tokenizer
main([vocab_file, dir_tokenizer, "--verbose"])

9138
unicode-data.cpp
View file

File diff suppressed because it is too large Load diff

27
unicode-data.h
View file

@ -1,17 +1,20 @@
#pragma once
#include <cstdint>
#include <map>
#include <utility>
#include <vector>
#include <unordered_map>
#include <unordered_set>
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
extern const std::multimap<uint32_t, uint32_t> unicode_map_nfd;
extern const std::map<char32_t, char32_t> unicode_map_lowercase;
struct range_nfd {
uint32_t first;
uint32_t last;
uint32_t nfd;
};
static const uint32_t MAX_CODEPOINTS = 0x110000;
extern const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags;
extern const std::unordered_set<uint32_t> unicode_set_whitespace;
extern const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase;
extern const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase;
extern const std::vector<range_nfd> unicode_ranges_nfd;

200
unicode.cpp
View file

@ -1,4 +1,4 @@
#include "unicode.h"
#include "unicode.h"
#include "unicode-data.h"
#include <cassert>
@ -109,57 +109,49 @@ static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset)
// return result;
//}
static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
std::unordered_map<uint32_t, int> cpt_types;
for (auto p : unicode_ranges_number) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_NUMBER;
static std::vector<codepoint_flags> unicode_cpt_flags_array() {
std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
assert (unicode_ranges_flags.front().first == 0);
assert (unicode_ranges_flags.back().first == MAX_CODEPOINTS);
for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
const auto range_ini = unicode_ranges_flags[i-1]; // codepoint_ini, flags
const auto range_end = unicode_ranges_flags[i]; // codepoint_end, flags
for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
cpt_flags[cpt] = range_ini.second;
}
}
for (auto p : unicode_ranges_letter) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_LETTER;
}
for (auto cpt : unicode_set_whitespace) {
cpt_flags[cpt].is_whitespace = true;
}
for (auto p : unicode_ranges_separator) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_SEPARATOR;
}
for (auto p : unicode_map_lowercase) {
cpt_flags[p.second].is_lowercase = true;
}
for (auto p : unicode_ranges_accent_mark) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
}
for (auto p : unicode_map_uppercase) {
cpt_flags[p.second].is_uppercase = true;
}
for (auto p : unicode_ranges_punctuation) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_PUNCTUATION;
}
for (auto &range : unicode_ranges_nfd) { // start, last, nfd
cpt_flags[range.nfd].is_nfd = true;
}
for (auto p : unicode_ranges_symbol) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_SYMBOL;
}
}
for (auto p : unicode_ranges_control) {
for (auto i = p.first; i <= p.second; ++i) {
cpt_types[i] = CODEPOINT_TYPE_CONTROL;
}
}
return cpt_types;
return cpt_flags;
}
static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
std::unordered_map<uint8_t, std::string> map;
for (int ch = u'!'; ch <= u'~'; ++ch) {
for (int ch = 0x21; ch <= 0x7E; ++ch) { // u'!' to u'~'
assert(0 <= ch && ch < 256);
map[ch] = unicode_cpt_to_utf8(ch);
}
for (int ch = u'¡'; ch <= u'¬'; ++ch) {
for (int ch = 0xA1; ch <= 0xAC; ++ch) { // u'¡' to u'¬'
assert(0 <= ch && ch < 256);
map[ch] = unicode_cpt_to_utf8(ch);
}
for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
for (int ch = 0xAE; ch <= 0xFF; ++ch) { // u'®' to u'ÿ'
assert(0 <= ch && ch < 256);
map[ch] = unicode_cpt_to_utf8(ch);
}
@ -175,15 +167,15 @@ static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
std::unordered_map<std::string, uint8_t> map;
for (int ch = u'!'; ch <= u'~'; ++ch) {
for (int ch = 0x21; ch <= 0x7E; ++ch) { // u'!' to u'~'
assert(0 <= ch && ch < 256);
map[unicode_cpt_to_utf8(ch)] = ch;
}
for (int ch = u'¡'; ch <= u'¬'; ++ch) {
for (int ch = 0xA1; ch <= 0xAC; ++ch) { // u'¡' to u'¬'
assert(0 <= ch && ch < 256);
map[unicode_cpt_to_utf8(ch)] = ch;
}
for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
for (int ch = 0xAE; ch <= 0xFF; ++ch) { // u'®' to u'ÿ'
assert(0 <= ch && ch < 256);
map[unicode_cpt_to_utf8(ch)] = ch;
}
@ -238,8 +230,9 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
};
auto _get_cpt_type = [&] (const size_t pos) -> int {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
};
size_t _prev_end = offset_ini;
@ -261,7 +254,7 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
const char32_t cpt = _get_cpt(pos);
const int cpt_type = _get_cpt_type(pos);
const auto flags = _get_flags(pos);
// regex: 's|'t|'re|'ve|'m|'ll|'d
if (cpt == '\'' && pos+1 < offset_end) {
@ -281,39 +274,37 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
}
}
char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
// regex: <space>?\p{L}+
if (cpt2_type == CODEPOINT_TYPE_LETTER) {
if (flags2.is_letter) {
pos += (cpt == ' ');
while (cpt2_type == CODEPOINT_TYPE_LETTER) {
cpt2_type = _get_cpt_type(++pos);
while (flags2.is_letter) {
flags2 = _get_flags(++pos);
}
_add_token(pos);
continue;
}
// regex: <space>?\p{N}+
if (cpt2_type == CODEPOINT_TYPE_NUMBER) {
if (flags2.is_number) {
pos += (cpt == ' ');
while (cpt2_type == CODEPOINT_TYPE_NUMBER) {
cpt2_type = _get_cpt_type(++pos);
while (flags2.is_number) {
flags2 = _get_flags(++pos);
}
_add_token(pos);
continue;
}
// regex: <space>?[^\s\p{L}\p{N}]+
if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
pos += (cpt == ' ');
while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
cpt2_type = _get_cpt_type(++pos);
cpt2 = _get_cpt(pos);
while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
flags2 = _get_flags(++pos);
}
_add_token(pos);
continue;
}
size_t num_whitespaces = 0;
while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
while (_get_flags(pos+num_whitespaces).is_whitespace) {
num_whitespaces++;
}
@ -357,8 +348,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
};
auto _get_cpt_type = [&] (const size_t pos) -> int {
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
};
size_t _prev_end = offset_ini;
@ -380,7 +372,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
const char32_t cpt = _get_cpt(pos);
const int cpt_type = _get_cpt_type(pos);
const auto flags = _get_flags(pos);
// regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
if (cpt == '\'' && pos+1 < offset_end) {
@ -401,10 +393,10 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
}
// regex: [^\r\n\p{L}\p{N}]?\p{L}+ //####FIXME: the first \p{L} is correct?
if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) {
if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) { // one or more letters
if (!(cpt == '\r' || cpt == '\n' || /*flags.is_letter |*/ flags.is_number)) {
if (flags.is_letter || _get_flags(pos+1).is_letter) { // one or more letters
pos++;
while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) {
while (_get_flags(pos).is_letter) {
pos++;
}
_add_token(pos);
@ -413,9 +405,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
}
// regex: \p{N}{1,3}
if (cpt_type == CODEPOINT_TYPE_NUMBER) {
if (flags.is_number) {
size_t ini = pos;
while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) {
while (_get_flags(pos).is_number) {
if (++pos - ini >= 3 ) {
_add_token(pos);
ini = pos;
@ -426,14 +418,13 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
}
// regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
pos += (cpt == ' ');
while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
cpt2_type = _get_cpt_type(++pos);
cpt2 = _get_cpt(pos);
while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
flags2 = _get_flags(++pos);
}
char32_t cpt2 = _get_cpt(pos);
while (cpt2 == '\r' || cpt2 == '\n') {
cpt2 = _get_cpt(++pos);
}
@ -443,7 +434,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
size_t num_whitespaces = 0;
size_t last_end_r_or_n = 0;
while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
while (_get_flags(pos+num_whitespaces).is_whitespace) {
char32_t cpt2 = _get_cpt(pos+num_whitespaces);
if (cpt2 == '\r' || cpt2 == '\n') {
last_end_r_or_n = pos + num_whitespaces + 1;
@ -589,15 +580,14 @@ std::string unicode_cpt_to_utf8(uint32_t cp) {
}
std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
std::vector<uint32_t> result;
result.reserve(cpts.size());
auto comp = [] (const uint32_t cpt, const range_nfd & range) {
return cpt < range.first;
};
std::vector<uint32_t> result(cpts.size());
for (size_t i = 0; i < cpts.size(); ++i) {
auto it = unicode_map_nfd.find(cpts[i]);
if (it == unicode_map_nfd.end()) {
result.push_back(cpts[i]);
} else {
result.push_back(it->second);
}
const uint32_t cpt = cpts[i];
auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1;
result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
}
return result;
}
@ -611,31 +601,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
return result;
}
int unicode_cpt_type(uint32_t cp) {
static std::unordered_map<uint32_t, int> cpt_types = unicode_cpt_type_map();
const auto it = cpt_types.find(cp);
return it == cpt_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
codepoint_flags unicode_cpt_flags(const uint32_t cp) {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
static const auto cpt_flags = unicode_cpt_flags_array();
return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
}
int unicode_cpt_type(const std::string & utf8) {
if (utf8.length() == 0) {
return CODEPOINT_TYPE_UNIDENTIFIED;
codepoint_flags unicode_cpt_flags(const std::string & utf8) {
static const codepoint_flags undef(codepoint_flags::UNDEFINED);
if (utf8.empty()) {
return undef; // undefined
}
size_t offset = 0;
return unicode_cpt_type(unicode_cpt_from_utf8(utf8, offset));
}
bool unicode_cpt_is_whitespace(uint32_t cp) {
static const std::unordered_set<uint32_t> is_whitespace = [] {
std::unordered_set<uint32_t> is_whitespace;
for (auto p : unicode_ranges_whitespace) {
for (auto i = p.first; i <= p.second; ++i) {
is_whitespace.insert(i);
}
}
return is_whitespace;
}();
return (bool)is_whitespace.count(cp);
return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
}
std::string unicode_byte_to_utf8(uint8_t byte) {
@ -656,21 +634,21 @@ char32_t unicode_tolower(char32_t cp) {
std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
// unicode categories
static const std::map<std::string, int> k_ucat_enum = {
{ "\\p{N}", CODEPOINT_TYPE_NUMBER },
{ "\\p{L}", CODEPOINT_TYPE_LETTER },
{ "\\p{P}", CODEPOINT_TYPE_PUNCTUATION },
{ "\\p{N}", codepoint_flags::NUMBER },
{ "\\p{L}", codepoint_flags::LETTER },
{ "\\p{P}", codepoint_flags::PUNCTUATION },
};
static const std::map<int, int> k_ucat_cpt = {
{ CODEPOINT_TYPE_NUMBER, 0xD1 },
{ CODEPOINT_TYPE_LETTER, 0xD2 },
{ CODEPOINT_TYPE_PUNCTUATION, 0xD3 },
{ codepoint_flags::NUMBER, 0xD1 },
{ codepoint_flags::LETTER, 0xD2 },
{ codepoint_flags::PUNCTUATION, 0xD3 },
};
static const std::map<int, std::string> k_ucat_map = {
{ CODEPOINT_TYPE_NUMBER, "\x30-\x39" }, // 0-9
{ CODEPOINT_TYPE_LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
{ CODEPOINT_TYPE_PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
{ codepoint_flags::NUMBER, "\x30-\x39" }, // 0-9
{ codepoint_flags::LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
{ codepoint_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
};
// compute collapsed codepoints only if needed by at least one regex
@ -701,10 +679,10 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
continue;
}
const int cpt_type = unicode_cpt_type(cpts[i]);
const int cpt_flag = unicode_cpt_flags(cpts[i]).category_flag();
if (k_ucat_cpt.find(cpt_type) != k_ucat_cpt.end()) {
text_collapsed[i] = k_ucat_cpt.at(cpt_type);
if (k_ucat_cpt.find(cpt_flag) != k_ucat_cpt.end()) {
text_collapsed[i] = k_ucat_cpt.at(cpt_flag);
} else {
text_collapsed[i] = (char) 0xD0; // fallback
}

56
unicode.h
View file

@ -4,24 +4,56 @@
#include <string>
#include <vector>
#define CODEPOINT_TYPE_UNIDENTIFIED 0
#define CODEPOINT_TYPE_NUMBER 1
#define CODEPOINT_TYPE_LETTER 2
#define CODEPOINT_TYPE_SEPARATOR 3
#define CODEPOINT_TYPE_ACCENT_MARK 4
#define CODEPOINT_TYPE_PUNCTUATION 5
#define CODEPOINT_TYPE_SYMBOL 6
#define CODEPOINT_TYPE_CONTROL 7
struct codepoint_flags {
enum {
UNDEFINED = 0x0001,
NUMBER = 0x0002, // regex: \p{N}
LETTER = 0x0004, // regex: \p{L}
SEPARATOR = 0x0008, // regex: \p{Z}
ACCENT_MARK = 0x0010, // regex: \p{M}
PUNCTUATION = 0x0020, // regex: \p{P}
SYMBOL = 0x0040, // regex: \p{S}
CONTROL = 0x0080, // regex: \p{C}
MASK_CATEGORIES = 0x00FF,
};
// codepoint type
uint16_t is_undefined : 1;
uint16_t is_number : 1; // regex: \p{N}
uint16_t is_letter : 1; // regex: \p{L}
uint16_t is_separator : 1; // regex: \p{Z}
uint16_t is_accent_mark : 1; // regex: \p{M}
uint16_t is_punctuation : 1; // regex: \p{P}
uint16_t is_symbol : 1; // regex: \p{S}
uint16_t is_control : 1; // regex: \p{C}
// helper flags
uint16_t is_whitespace : 1; // regex: \s
uint16_t is_lowercase : 1;
uint16_t is_uppercase : 1;
uint16_t is_nfd : 1;
// decode from uint16
inline codepoint_flags(const uint16_t flags=0) {
*reinterpret_cast<uint16_t*>(this) = flags;
}
inline uint16_t as_uint() const {
return *reinterpret_cast<const uint16_t*>(this);
}
inline uint16_t category_flag() const {
return this->as_uint() & MASK_CATEGORIES;
}
};
std::string unicode_cpt_to_utf8(uint32_t cp);
std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
int unicode_cpt_type(uint32_t cp);
int unicode_cpt_type(const std::string & utf8);
bool unicode_cpt_is_whitespace(uint32_t cp);
codepoint_flags unicode_cpt_flags(const uint32_t cp);
codepoint_flags unicode_cpt_flags(const std::string & utf8);
std::string unicode_byte_to_utf8(uint8_t byte);
uint8_t unicode_utf8_to_byte(const std::string & utf8);