Nomic Vulkan backend (#4456)

Signed-off-by: Jared Van Bortel <jared@nomic.ai>
Co-authored-by: niansa <anton-sa@web.de>
Co-authored-by: Adam Treat <treat.adam@gmail.com>
Co-authored-by: Aaron Miller <apage43@ninjawhale.com>
Co-authored-by: ToKiNoBug <tokinobug@163.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Jared Van Bortel 2024-01-29 15:50:50 -05:00 committed by GitHub
parent 2aed77eb06
commit fbf1ddec69
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
45 changed files with 4271 additions and 19 deletions

1
.ecrc
View File

@ -1,4 +1,5 @@
{ {
"Exclude": ["^\\.gitmodules$"],
"Disable": { "Disable": {
"IndentSize": true "IndentSize": true
} }

View File

@ -337,6 +337,7 @@ jobs:
OPENCL_VERSION: 2023.04.17 OPENCL_VERSION: 2023.04.17
CLBLAST_VERSION: 1.6.0 CLBLAST_VERSION: 1.6.0
SDE_VERSION: 9.33.0-2024-01-07 SDE_VERSION: 9.33.0-2024-01-07
VULKAN_VERSION: 1.3.261.1
strategy: strategy:
matrix: matrix:
@ -353,6 +354,8 @@ jobs:
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"' defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
- build: 'openblas' - build: 'openblas'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'kompute'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
steps: steps:
- name: Clone - name: Clone
@ -361,6 +364,12 @@ jobs:
with: with:
fetch-depth: 0 fetch-depth: 0
- name: Clone Kompute submodule
id: clone_kompute
if: ${{ matrix.build == 'kompute' }}
run: |
git submodule update --init kompute
- name: Download OpenCL SDK - name: Download OpenCL SDK
id: get_opencl id: get_opencl
if: ${{ matrix.build == 'clblast' }} if: ${{ matrix.build == 'clblast' }}
@ -395,6 +404,15 @@ jobs:
$lib = $(join-path $msvc 'bin\Hostx64\x64\lib.exe') $lib = $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
& $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
- name: Install Vulkan SDK
id: get_vulkan
if: ${{ matrix.build == 'kompute' }}
run: |
curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
& "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
- name: Build - name: Build
id: cmake_build id: cmake_build
run: | run: |
@ -432,7 +450,8 @@ jobs:
- name: Test - name: Test
id: cmake_test id: cmake_test
if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512 # not all machines have native AVX-512
if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
run: | run: |
cd build cd build
ctest -L main -C Release --verbose --timeout 900 ctest -L main -C Release --verbose --timeout 900

3
.gitmodules vendored Normal file
View File

@ -0,0 +1,3 @@
[submodule "kompute"]
path = kompute
url = https://github.com/nomic-ai/kompute.git

View File

@ -103,6 +103,7 @@ option(LLAMA_VULKAN "llama: use Vulkan"
option(LLAMA_METAL "llama: use Metal" ${LLAMA_METAL_DEFAULT}) option(LLAMA_METAL "llama: use Metal" ${LLAMA_METAL_DEFAULT})
option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" OFF) option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" OFF)
option(LLAMA_METAL_SHADER_DEBUG "llama: compile Metal with -fno-fast-math" OFF) option(LLAMA_METAL_SHADER_DEBUG "llama: compile Metal with -fno-fast-math" OFF)
option(LLAMA_KOMPUTE "llama: use Kompute" OFF)
option(LLAMA_MPI "llama: use MPI" OFF) option(LLAMA_MPI "llama: use MPI" OFF)
option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" 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 "llama: use SYCL" OFF)
@ -484,7 +485,6 @@ if (LLAMA_HIPBLAS)
endif() endif()
endif() endif()
if (LLAMA_SYCL) if (LLAMA_SYCL)
if ( NOT DEFINED ENV{ONEAPI_ROOT}) if ( NOT DEFINED ENV{ONEAPI_ROOT})
message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh") message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
@ -510,6 +510,160 @@ if (LLAMA_SYCL)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread) set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
endif() endif()
if (LLAMA_KOMPUTE)
add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
find_package(Vulkan COMPONENTS glslc REQUIRED)
find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
if (NOT glslc_executable)
message(FATAL_ERROR "glslc not found")
endif()
function(compile_shader)
set(options)
set(oneValueArgs)
set(multiValueArgs SOURCES)
cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(source ${compile_shader_SOURCES})
get_filename_component(filename ${source} NAME)
set(spv_file ${filename}.spv)
add_custom_command(
OUTPUT ${spv_file}
DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
COMMENT "Compiling ${source} to ${spv_file}"
)
get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
set(FILE_NAME "shader${RAW_FILE_NAME}")
string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
if(CMAKE_GENERATOR MATCHES "Visual Studio")
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
)
else()
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
)
endif()
endforeach()
endfunction()
if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
message(STATUS "Kompute found")
set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
add_subdirectory(kompute)
# Compile our shaders
compile_shader(SOURCES
kompute-shaders/op_scale.comp
kompute-shaders/op_scale_8.comp
kompute-shaders/op_add.comp
kompute-shaders/op_addrow.comp
kompute-shaders/op_mul.comp
kompute-shaders/op_silu.comp
kompute-shaders/op_relu.comp
kompute-shaders/op_gelu.comp
kompute-shaders/op_softmax.comp
kompute-shaders/op_norm.comp
kompute-shaders/op_rmsnorm.comp
kompute-shaders/op_diagmask.comp
kompute-shaders/op_mul_mat_mat_f32.comp
kompute-shaders/op_mul_mat_f16.comp
kompute-shaders/op_mul_mat_q8_0.comp
kompute-shaders/op_mul_mat_q4_0.comp
kompute-shaders/op_mul_mat_q4_1.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f16.comp
kompute-shaders/op_getrows_q4_0.comp
kompute-shaders/op_getrows_q4_1.comp
kompute-shaders/op_getrows_q6_k.comp
kompute-shaders/op_rope_f16.comp
kompute-shaders/op_rope_f32.comp
kompute-shaders/op_cpy_f16_f16.comp
kompute-shaders/op_cpy_f16_f32.comp
kompute-shaders/op_cpy_f32_f16.comp
kompute-shaders/op_cpy_f32_f32.comp
)
# Create a custom target for our generated shaders
add_custom_target(generated_shaders DEPENDS
shaderop_scale.h
shaderop_scale_8.h
shaderop_add.h
shaderop_addrow.h
shaderop_mul.h
shaderop_silu.h
shaderop_relu.h
shaderop_gelu.h
shaderop_softmax.h
shaderop_norm.h
shaderop_rmsnorm.h
shaderop_diagmask.h
shaderop_mul_mat_mat_f32.h
shaderop_mul_mat_f16.h
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f16.h
shaderop_getrows_q4_0.h
shaderop_getrows_q4_1.h
shaderop_getrows_q6_k.h
shaderop_rope_f16.h
shaderop_rope_f32.h
shaderop_cpy_f16_f16.h
shaderop_cpy_f16_f32.h
shaderop_cpy_f32_f16.h
shaderop_cpy_f32_f32.h
)
# Create a custom command that depends on the generated_shaders
add_custom_command(
OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
DEPENDS generated_shaders
COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
)
# Add the stamp to the main sources to ensure dependency tracking
set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
add_compile_definitions(GGML_USE_KOMPUTE)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
else()
message(WARNING "Kompute not found")
endif()
endif()
function(get_flags CCID CCVER) function(get_flags CCID CCVER)
set(C_FLAGS "") set(C_FLAGS "")
set(CXX_FLAGS "") set(CXX_FLAGS "")
@ -859,6 +1013,7 @@ add_library(ggml OBJECT
${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI} ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA} ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
${GGML_SOURCES_SYCL} ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL} ${GGML_HEADERS_SYCL}
${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
) )
target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES}) target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

View File

@ -373,6 +373,11 @@ GGML_CALL static void ggml_backend_registry_init(void) {
extern GGML_CALL int ggml_backend_vk_reg_devices(void); extern GGML_CALL int ggml_backend_vk_reg_devices(void);
ggml_backend_vk_reg_devices(); ggml_backend_vk_reg_devices();
#endif #endif
#ifdef GGML_USE_KOMPUTE
extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
ggml_backend_kompute_reg_devices();
#endif
} }
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) { GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {

1990
ggml-kompute.cpp Normal file

File diff suppressed because it is too large Load Diff

46
ggml-kompute.h Normal file
View File

@ -0,0 +1,46 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_vk_device {
int index;
int type; // same as VkPhysicalDeviceType
size_t heapSize;
const char * name;
const char * vendor;
int subgroupSize;
uint64_t bufferAlignment;
uint64_t maxAlloc;
};
struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
bool ggml_vk_has_vulkan(void);
bool ggml_vk_has_device(void);
struct ggml_vk_device ggml_vk_current_device(void);
//
// backend API
//
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
#ifdef __cplusplus
}
#endif

1
kompute Submodule

@ -0,0 +1 @@
Subproject commit 4565194ed7c32d1d2efa32ceab4d3c6cae006306

102
kompute-shaders/common.comp Normal file
View File

@ -0,0 +1,102 @@
#extension GL_EXT_shader_16bit_storage: require
#extension GL_EXT_shader_8bit_storage: require
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
#extension GL_EXT_control_flow_attributes: enable
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
#define QK4_0 32
#define QK4_1 32
#define GELU_COEF_A 0.044715
#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
#define TWOPI_F 6.283185307179586f
#define QK_K 256
#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
#define sizeof_block_q4_0 0x12
struct block_q4_0 {
float16_t d;
uint8_t qs[QK4_0 / 2];
};
mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float md = -8.f * xb.d;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
}
return reg;
}
#define sizeof_block_q4_1 0x14
struct block_q4_1 {
float16_t d;
float16_t m;
uint8_t qs[QK4_1 / 2];
};
mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float m = xb.m;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
}
return reg;
}
#define sizeof_block_q6_k 210
struct block_q6_k {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
float16_t d; // super-block scale
};
mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
const float16_t d_all = xb.d;
const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
const uint qhIndex = 32*(il/8) + 16*(il&1);
float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
const uint16_t kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? uint8_t(0xF0) : uint8_t(0x0F);
const float16_t coef = il>1 ? float16_t(1.f/16.f) : float16_t(1.f);
const float16_t ml = float16_t(d_all * sc * 32.f);
const float16_t dl = float16_t(d_all * sc * coef);
mat4 reg;
for (int i = 0; i < 16; ++i) {
const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
: ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
return reg;
}
#define QK8_0 32
// struct block_q8_0 {
// float16_t d; // delta
// int8_t qs[QK8_0]; // quants
// };
#define sizeof_block_q8_0 34

View File

@ -0,0 +1,58 @@
#version 450
#include "common.comp"
layout(local_size_x = 1024) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb00;
int nb01;
int nb02;
int nb03;
int ne10;
int ne11;
int ne12;
int ne13;
int nb10;
int nb11;
int nb12;
int nb13;
int ne0;
int nb0;
int nb1;
int nb2;
int nb3;
//int offs; // TODO: needed for GGML_OP_ACC, see metal code
} pcs;
// general-purpose kernel for addition of two tensors
// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
// cons: not very efficient
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint i13 = i03 % pcs.ne13;
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
int offs = 0; // TMP (see above)
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11 ) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1 + offs) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
}
}

View File

@ -0,0 +1,25 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
uint row;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
}
}

View File

@ -0,0 +1,52 @@
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

View File

@ -0,0 +1,52 @@
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

View File

@ -0,0 +1,52 @@
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

View File

@ -0,0 +1,52 @@
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

View File

@ -0,0 +1,30 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint n_past;
int ne00;
int ne01;
} pcs;
void main() {
const uint i02 = gl_WorkGroupID.z;
const uint i01 = gl_WorkGroupID.y;
const uint i00 = gl_WorkGroupID.x;
const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
if (i00 > pcs.n_past + i01) {
out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
} else {
out_[index + pcs.outOff] = in_[index + pcs.inOff];
}
}

View File

@ -0,0 +1,22 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 8;
for (uint x = 0; x < 8; x++) {
const uint i = baseIndex + x;
const float y = in_[i + pcs.inOff];
out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
}
}

View File

@ -0,0 +1,17 @@
void main() {
const uint i = gl_WorkGroupID.x;
const int r = inB[i + pcs.inBOff];
int z = 0;
for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
const mat4 result = dequantize_block(inIndex, ind%NL);
for (uint j = 0; j < 4; ++j) {
for (uint k = 0; k < 4; ++k) {
const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
out_[outIndex] = result[j][k];
++z;
}
}
}
}

View File

@ -0,0 +1,31 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
for (int j = 0; j < k; j++) {
out_[y + j] = inA[x + j];
}
}
void main() {
const uint i = gl_WorkGroupID.x;
const int r = inB[i + pcs.inBOff];
dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
}

View File

@ -0,0 +1,38 @@
#version 450
#include "common.comp"
#define NL 2
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q4_0
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q4_0 get_unaligned_block_q4_0(uint index) {
block_q4_0 fres;
fres.d = u8BufToFloat16(inA, index);
[[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
fres.qs[it] = inA[index+2+it];
}
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q4_0 block = get_unaligned_block_q4_0(index);
return dequantize_q4_0(block, il);
}
#include "op_getrows.comp"

View File

@ -0,0 +1,39 @@
#version 450
#include "common.comp"
#define NL 2
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q4_1
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q4_1 get_unaligned_block_q4_1(uint index) {
block_q4_1 fres;
fres.d = u8BufToFloat16(inA, index);
fres.m = u8BufToFloat16(inA, index+2);
[[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
fres.qs[it] = inA[index+4+it];
}
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q4_1 block = get_unaligned_block_q4_1(index);
return dequantize_q4_1(block, il);
}
#include "op_getrows.comp"

View File

@ -0,0 +1,44 @@
#version 450
#include "common.comp"
#define NL 16
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q6_k
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q6_k get_unaligned_block_q6_k(uint index) {
block_q6_k fres;
[[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
fres.ql[it] = inA[index + it];
}
[[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
fres.qh[it] = inA[index + QK_K/2 + it];
}
[[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
}
fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q6_k block = get_unaligned_block_q6_k(index);
return dequantize_q6_k(block, il);
}
#include "op_getrows.comp"

View File

@ -0,0 +1,52 @@
#version 450
#include "common.comp"
layout(local_size_x = 1024) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb00;
int nb01;
int nb02;
int nb03;
int ne10;
int ne11;
int ne12;
int ne13;
int nb10;
int nb11;
int nb12;
int nb13;
int ne0;
int nb0;
int nb1;
int nb2;
int nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint i13 = i03 % pcs.ne13;
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
}
}

View File

@ -0,0 +1,67 @@
#version 450
#include "common.comp"
#extension GL_KHR_shader_subgroup_arithmetic : require
layout(local_size_x_id = 0) in;
layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
int ne10;
int ne11;
int ne12;
uint nb10;
uint nb11;
uint nb12;
int ne0;
int ne1;
uint r2;
uint r3;
} pcs;
#define N_F16_F32 4
void main() {
const uint r0 = gl_WorkGroupID.x;
const uint rb = gl_WorkGroupID.y*N_F16_F32;
const uint im = gl_WorkGroupID.z;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02;
const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
for (uint row = 0; row < N_F16_F32; ++row) {
uint r1 = rb + row;
if (r1 >= pcs.ne11) {
break;
}
const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB
float sumf = 0;
for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
sumf += float(inA[x+i]) * float(inB[y+i]);
}
const float all_sum = subgroupAdd(sumf);
if (subgroupElect()) {
out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
}
}
}

View File

@ -0,0 +1,51 @@
#version 450
#include "common.comp"
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
// device subgroup size
layout (local_size_x_id = 0) in;
layout(binding = 0) readonly buffer tensorInA { float inA[]; };
layout(binding = 1) readonly buffer tensorInB { float inB[]; };
layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
layout(push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
int ne11;
int ne12;
uint nb01;
uint nb02;
uint nb11;
uint nb12;
uint nb1;
uint nb2;
}
pcs;
void main() {
uvec3 gid = gl_WorkGroupID;
uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
float sum = 0.0f;
for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
sum += float(inA[x+i]) * float(inB[y+i]);
}
const float all_sum = subgroupAdd(sum);
if (subgroupElect()) {
out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
}
}

View File

@ -0,0 +1,33 @@
#version 450
#include "common.comp"
#define BLOCKS_IN_QUANT QK4_0
#define SIZE_OF_BLOCK sizeof_block_q4_0
#define N_ROWS 4
#include "op_mul_mv_q_n_pre.comp"
// The q4_0 version of this function
float block_q_n_dot_y(uint block_index, uint yb, uint il) {
vec2 acc = vec2(0.0, 0.0);
const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
float d = float(u8BufToFloat16(inA, index));
float sumy = 0.0f;
for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
const float yl0 = inB[yb + i];
const float yl1 = inB[yb + i + 1];
const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
sumy += yl0 + yl1 + yl8 + yl9;
acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
}
return d * (sumy * -8.f + acc[0] + acc[1]);
}
#include "op_mul_mv_q_n.comp"

View File

@ -0,0 +1,35 @@
#version 450
#include "common.comp"
#define BLOCKS_IN_QUANT QK4_1
#define SIZE_OF_BLOCK sizeof_block_q4_1
#define N_ROWS 4
#include "op_mul_mv_q_n_pre.comp"
// The q4_1 version of this function
float block_q_n_dot_y(uint block_index, uint yb, uint il) {
vec2 acc = vec2(0.0, 0.0);
const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
float d = float(u8BufToFloat16(inA, index));
float m = float(u8BufToFloat16(inA, index+2));
float sumy = 0.0f;
for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
const float yl0 = inB[yb + i];
const float yl1 = inB[yb + i + 1];
const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
sumy += yl0 + yl1 + yl8 + yl9;
acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
}
return d * (acc[0] + acc[1]) + sumy * m;
}
#include "op_mul_mv_q_n.comp"

View File

@ -0,0 +1,94 @@
#version 450
#include "common.comp"
#define SIZE_OF_BLOCK sizeof_block_q6_k
layout(local_size_x_id = 0) in;
layout(local_size_y_id = 1) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne10;
int ne0;
int ne1;
int ne01;
int gqa;
} pcs;
void main() {
const uint8_t kmask1 = uint8_t(0x03);
const uint8_t kmask2 = uint8_t(0x0C);
const uint8_t kmask3 = uint8_t(0x30);
const uint8_t kmask4 = uint8_t(0xC0);
const uint nb = pcs.ne00/QK_K;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint r2 = gl_WorkGroupID.z;
const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0);
const uint x = row * nb + offset0; // Based from inA without base offset
const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
float sumf = 0;
// bits of invocation ID for gl_SubgroupSize=32:
// x x x x x
// 4 3 2 1 0
// ( tid ) ix
// ip ( il )
const uint block_stride = gl_SubgroupSize / 16; // number of blocks each subgroup processes
const uint tid = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
const uint ix = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
const uint ip = tid/8; // first or second half of block (0 or 1)
const uint il = tid%8; // each half has 8 parts, one per scale
const uint n = 4; // 4 scales at a time (and 4 sums)
const uint l0 = n*il; // offset into half-block, 0..28
const uint is = 8*ip + l0/16; // 0, 1, 8, 9
const uint y_offset = 128*ip + l0;
const uint q_offset_l = 64*ip + l0;
const uint q_offset_h = 32*ip + l0;
for (uint i = ix; i < nb; i += block_stride) {
const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
const uint qlIndex = q_offset_l;
const uint q2Index = qlIndex + QK_K/8;
const uint qhIndex = q_offset_h;
const uint y = yy + i * QK_K + y_offset;
float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
for (uint l = 0; l < n; ++l) {
const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
sums[2] += inB[y+l+64] * (int8_t((currentQ1 >> 4) | ((currentQh & kmask3) << 0)) - 32);
sums[3] += inB[y+l+96] * (int8_t((currentQ2 >> 4) | ((currentQh & kmask4) >> 2)) - 32);
}
float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
}
const float tot = subgroupAdd(sumf);
if (subgroupElect()) {
out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
}
}

View File

@ -0,0 +1,73 @@
#version 450
#include "common.comp"
#include "op_mul_mv_q_n_pre.comp"
#define SIZE_OF_D 2
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
#define NB_Q8_0 8
void main() {
// NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
if (gl_SubgroupInvocationID > 31)
return;
const int nr = N_DST;
const int nsg = N_SIMDGROUP;
const int nw = N_SIMDWIDTH;
const int nb = pcs.ne00/QK8_0;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
float yl[NB_Q8_0];
float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
const uint ix = gl_SubgroupInvocationID.x/4;
const uint il = gl_SubgroupInvocationID.x%4;
uint yb = y + ix * QK8_0 + NB_Q8_0*il;
// each thread in a SIMD group deals with NB_Q8_0 quants at a time
for (uint ib = ix; ib < nb; ib += nw/4) {
for (int i = 0; i < NB_Q8_0; ++i) {
yl[i] = inB[yb + i];
}
for (int row = 0; row < nr; row++) {
const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
float sumq = 0.f;
for (int iq = 0; iq < NB_Q8_0; ++iq) {
const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
sumq += qs_iq * yl[iq];
}
const float16_t d = u8BufToFloat16(inA, x + block_offset);
sumf[row] += sumq*d;
}
yb += NB_Q8_0 * nw;
}
for (int row = 0; row < nr; ++row) {
const float tot = subgroupAdd(sumf[row]);
if (subgroupElect() && first_row + row < pcs.ne01) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
}
}
}

View File

@ -0,0 +1,48 @@
void main() {
// NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
if (gl_SubgroupInvocationID > 31)
return;
const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
const uint x = offset0; // Based from inA without base offset
const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
const uint ix = gl_SubgroupInvocationID/2;
const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
uint yb = y + ix * BLOCKS_IN_QUANT + il;
//debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
// gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
// gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
for (uint ib = ix; ib < nb; ib += 16) {
for (int row = 0; row < N_ROWS; row++) {
const uint block_index = x + ib + row * nb;
sumf[row] += block_q_n_dot_y(block_index, yb, il);
}
yb += BLOCKS_IN_QUANT * 16;
}
for (int row = 0; row < N_ROWS; ++row) {
const float tot = subgroupAdd(sumf[row]);
if (first_row + row < pcs.ne01 && subgroupElect()) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
}
}
}

View File

@ -0,0 +1,22 @@
layout(local_size_x_id = 0) in;
layout(local_size_y = 1) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
int ne10;
int ne12;
int ne0;
int ne1;
uint r2;
uint r3;
} pcs;

View File

@ -0,0 +1,84 @@
#version 450
#include "common.comp"
layout(local_size_x = 256) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint ne00;
uint nb01;
float eps;
} pcs;
shared float sum[gl_WorkGroupSize.x];
void main() {
const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
// MEAN
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += in_[x+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float mean = sum[0];
// recenter
const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] = in_[x+i00] - mean;
}
// VARIANCE
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float variance = sum[0];
const float scale = 1.0f/sqrt(variance + pcs.eps);
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] *= scale;
}
}

View File

@ -0,0 +1,21 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
}
}

View File

@ -0,0 +1,53 @@
#version 450
#include "common.comp"
layout(local_size_x = 512) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint ne00;
uint nb01;
float eps;
} pcs;
shared float sum[gl_WorkGroupSize.x];
void main() {
const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] = in_[x+i00] * scale;
}
}

View File

@ -0,0 +1,73 @@
#version 450
#include "rope_common.comp"
layout(binding = 0) buffer restrict readonly tensorInA { float16_t inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };
void main() {
const uint i3 = gl_WorkGroupID.z;
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_orig_ctx, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
const int p = inB[pcs.inBOff + i2];
float theta = float(p);
if (!is_neox) {
for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
float cos_theta, sin_theta;
rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
theta *= theta_scale;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_
const float x0 = float(inA[src]);
const float x1 = float(inA[src+1]);
out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta);
out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
}
} else {
const float inv_ndims = -1.f/pcs.n_dims;
for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
const uint cur_rot = ic;
float cos_theta, sin_theta;
rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
theta *= theta_scale;
const uint i0 = ic/2;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_
const float x0 = float(inA[src]);
const float x1 = float(inA[src+pcs.n_dims/2]);
out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta);
out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
}
for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
const uint i0 = ic;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_
out_[dst_data + 0] = inA[src + 0];
out_[dst_data + 1] = inA[src + 1];
}
}
}

View File

@ -0,0 +1,73 @@
#version 450
#include "rope_common.comp"
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
void main() {
const uint i3 = gl_WorkGroupID.z;
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_orig_ctx, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
const int p = inB[pcs.inBOff + i2];
float theta = float(p);
if (!is_neox) {
for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
float cos_theta, sin_theta;
rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
theta *= theta_scale;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
const float x0 = inA[src];
const float x1 = inA[src+1];
out_[dst_data] = x0*cos_theta - x1*sin_theta;
out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
}
} else {
const float inv_ndims = -1.f/pcs.n_dims;
for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
const uint cur_rot = ic;
float cos_theta, sin_theta;
rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
theta *= theta_scale;
const uint i0 = ic/2;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
const float x0 = inA[src];
const float x1 = inA[src+pcs.n_dims/2];
out_[dst_data] = x0*cos_theta - x1*sin_theta;
out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
}
for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
const uint i0 = ic;
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
out_[dst_data + 0] = inA[src + 0];
out_[dst_data + 1] = inA[src + 1];
}
}
}

View File

@ -0,0 +1,19 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
float scale;
} pcs;
void main() {
const uint i = gl_WorkGroupID.x;
out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
}

View File

@ -0,0 +1,23 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
float scale;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 8;
for (uint x = 0; x < 8; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
}
}

View File

@ -0,0 +1,22 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
const float y = in_[i + pcs.inOff];
out_[i + pcs.outOff] = y / (1.0 + exp(-y));
}
}

View File

@ -0,0 +1,56 @@
// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
#version 450
#include "common.comp"
layout(local_size_x_id = 0) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
float scale;
int mask;
} pcs;
void main() {
if (gl_SubgroupInvocationID > 31)
return;
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
const uint pdst = extra_off + pcs.outOff; // Based from out_
// parallel max
float localMax = uintBitsToFloat(0xFF800000);
for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f));
}
float max_ = subgroupMax(localMax);
// parallel sum
float localSum = 0.0f;
for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_);
localSum += exp_psrc0;
out_[pdst + i00] = exp_psrc0;
}
const float sum = subgroupAdd(localSum);
for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
out_[pdst + i00] /= sum;
}
}

View File

@ -0,0 +1,67 @@
#include "common.comp"
// TODO: use a local size of 32 or more (Metal uses 1024)
layout(local_size_x = 1) in;
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int n_dims;
int mode;
int n_orig_ctx;
float freq_base;
float freq_scale;
float ext_factor;
float attn_factor;
float beta_fast;
float beta_slow;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
float rope_yarn_ramp(const float low, const float high, const float i0) {
const float y = (i0 / 2 - low) / max(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
void rope_yarn(
float theta_extrap, float freq_scale, float corr_dims[2], float i0, float ext_factor, float mscale,
out float cos_theta, out float sin_theta
) {
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = freq_scale * theta_extrap;
float theta = theta_interp;
if (ext_factor != 0.0f) {
float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
// Get n-d magnitude scaling corrected for interpolation
mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
}
cos_theta = cos(theta) * mscale;
sin_theta = sin(theta) * mscale;
}
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
return n_dims * log(n_orig_ctx / (n_rot * TWOPI_F)) / (2 * log(base));
}
void rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, out float dims[2]
) {
// start and end correction dims
dims[0] = max(0.0f, floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
}

View File

@ -15,6 +15,8 @@
# include "ggml-vulkan.h" # include "ggml-vulkan.h"
#elif defined(GGML_USE_SYCL) #elif defined(GGML_USE_SYCL)
# include "ggml-sycl.h" # include "ggml-sycl.h"
#elif defined(GGML_USE_KOMPUTE)
# include "ggml-kompute.h"
#endif #endif
#ifdef GGML_USE_METAL #ifdef GGML_USE_METAL
@ -1313,6 +1315,11 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
buft = ggml_backend_sycl_buffer_type(gpu); buft = ggml_backend_sycl_buffer_type(gpu);
#elif defined(GGML_USE_CLBLAST) #elif defined(GGML_USE_CLBLAST)
buft = ggml_backend_opencl_buffer_type(); buft = ggml_backend_opencl_buffer_type();
#elif defined(GGML_USE_KOMPUTE)
buft = ggml_backend_kompute_buffer_type(gpu);
if (buft == nullptr) {
LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
}
#endif #endif
if (buft == nullptr) { if (buft == nullptr) {
@ -4107,7 +4114,7 @@ static bool llm_load_tensors(
} }
// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
try { try {
llama_model_loader ml(fname, params.use_mmap, params.kv_overrides); llama_model_loader ml(fname, params.use_mmap, params.kv_overrides);
@ -4128,6 +4135,22 @@ static int llama_model_load(const std::string & fname, llama_model & model, cons
return 0; return 0;
} }
#ifdef GGML_USE_KOMPUTE
if (ggml_vk_has_device() && params.n_gpu_layers > 0 && (
!(model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON)
|| !(
model.ftype == LLAMA_FTYPE_ALL_F32 ||
model.ftype == LLAMA_FTYPE_MOSTLY_F16 ||
model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1
)
)) {
// disable Vulkan due to unsupported model architecture or quantization type
// TODO(cebtenzzre): propagate this error outside of llama_load_model_from_file
params.n_gpu_layers = 0;
}
#endif
if (!llm_load_tensors( if (!llm_load_tensors(
ml, model, params.n_gpu_layers, params.split_mode, params.main_gpu, params.tensor_split, params.use_mlock, ml, model, params.n_gpu_layers, params.split_mode, params.main_gpu, params.tensor_split, params.use_mlock,
params.progress_callback, params.progress_callback_user_data params.progress_callback, params.progress_callback_user_data
@ -10259,6 +10282,16 @@ struct llama_context * llama_new_context_with_model(
} }
ctx->backends.push_back(backend); ctx->backends.push_back(backend);
} }
#elif defined(GGML_USE_KOMPUTE)
if (model->n_gpu_layers > 0) {
auto * backend = ggml_backend_kompute_init(model->main_gpu);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
#endif #endif
ctx->backend_cpu = ggml_backend_cpu_init(); ctx->backend_cpu = ggml_backend_cpu_init();
if (ctx->backend_cpu == nullptr) { if (ctx->backend_cpu == nullptr) {

View File

@ -49,7 +49,8 @@
#define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
#define LLAMA_SESSION_VERSION 4 #define LLAMA_SESSION_VERSION 4
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL) #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU. // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
#define LLAMA_SUPPORTS_GPU_OFFLOAD #define LLAMA_SUPPORTS_GPU_OFFLOAD
#endif #endif

View File

@ -370,12 +370,15 @@ struct test_case {
printf(" %s(%s): ", op_desc(out).c_str(), vars().c_str()); printf(" %s(%s): ", op_desc(out).c_str(), vars().c_str());
fflush(stdout); fflush(stdout);
// check if backends support op // check if the backends support the ops
bool supported = true; bool supported = true;
for (ggml_backend_t backend : {backend1, backend2}) { for (ggml_backend_t backend : {backend1, backend2}) {
if (!ggml_backend_supports_op(backend, out)) { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (!ggml_backend_supports_op(backend, t)) {
printf("not supported [%s] ", ggml_backend_name(backend)); printf("not supported [%s] ", ggml_backend_name(backend));
supported = false; supported = false;
break;
}
} }
} }
if (!supported) { if (!supported) {
@ -626,6 +629,13 @@ struct test_unary : public test_case {
ggml_tensor * out = ggml_unary(ctx, in, op); ggml_tensor * out = ggml_unary(ctx, in, op);
return out; return out;
} }
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
// test extended range of values to check for NaNs in GELU
init_tensor_uniform(t, -150.f, 150.f);
}
}
}; };
// GGML_OP_GET_ROWS // GGML_OP_GET_ROWS
@ -1066,18 +1076,24 @@ struct test_diag_mask_inf : public test_case {
struct test_soft_max : public test_case { struct test_soft_max : public test_case {
const ggml_type type; const ggml_type type;
const std::array<int64_t, 4> ne; const std::array<int64_t, 4> ne;
const float scale;
const bool mask;
std::string vars() override { std::string vars() override {
return VARS_TO_STR2(type, ne); return VARS_TO_STR4(type, ne, scale, mask);
} }
test_soft_max(ggml_type type = GGML_TYPE_F32, test_soft_max(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 10}) std::array<int64_t, 4> ne = {10, 10, 10, 10},
: type(type), ne(ne) {} float scale = 1.0f,
bool mask = false)
: type(type), ne(ne), scale(scale), mask(mask) {}
ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * out = ggml_soft_max(ctx, a); ggml_tensor * b = nullptr;
if (mask) { b = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]); }
ggml_tensor * out = ggml_soft_max_ext(ctx, a, b, scale);
return out; return out;
} }
}; };
@ -1474,6 +1490,393 @@ struct test_moe : public test_case {
} }
}; };
enum llm_norm_type {
LLM_NORM,
LLM_NORM_RMS,
};
struct llama_hparams {
uint32_t n_vocab;
uint32_t n_embd;
uint32_t n_head;
uint32_t n_head_kv;
static constexpr uint32_t n_layer = 1;
uint32_t n_rot;
uint32_t n_embd_head; // dimension of values (d_v)
uint32_t n_ff;
float f_norm_eps;
float f_norm_rms_eps;
// cparams
static constexpr uint32_t n_ctx = 512; // user-specified context size
static constexpr uint32_t n_orig_ctx = n_ctx;
// batch
int32_t n_tokens;
// llm_build_context
static constexpr int32_t n_kv = 32; // size of KV cache to consider (n_kv <= n_ctx
static constexpr int32_t kv_head = 1; // index of where we store new KV data in the cache
uint32_t n_embd_gqa() const { // dimension of key embeddings across all k-v heads
return n_embd_head * n_head_kv;
}
};
// LLM base class
struct test_llm : public test_case {
llama_hparams hp;
protected:
test_llm(llama_hparams hp)
: hp(std::move(hp)) {
}
public:
struct ggml_tensor * llm_build_norm(
struct ggml_context * ctx,
struct ggml_tensor * cur,
struct ggml_tensor * mw,
struct ggml_tensor * mb,
llm_norm_type type) {
switch (type) {
case LLM_NORM: cur = ggml_norm (ctx, cur, hp.f_norm_eps); break;
case LLM_NORM_RMS: cur = ggml_rms_norm(ctx, cur, hp.f_norm_rms_eps); break;
}
cur = ggml_mul(ctx, cur, mw);
if (mb) {
cur = ggml_add(ctx, cur, mb);
}
return cur;
}
void llm_build_kv_store(
struct ggml_context * ctx,
struct ggml_tensor * k_l,
struct ggml_tensor * v_l,
struct ggml_tensor * k_cur,
struct ggml_tensor * v_cur) {
// compute the transposed [n_tokens, n_embd] V matrix
struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, hp.n_embd_gqa(), hp.n_tokens));
struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, k_l, hp.n_tokens*hp.n_embd_gqa(),
(ggml_row_size(k_l->type, hp.n_embd_gqa()))*hp.kv_head);
struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, v_l, hp.n_tokens, hp.n_embd_gqa(),
( hp.n_ctx)*ggml_element_size(v_l),
(hp.kv_head)*ggml_element_size(v_l));
// important: storing RoPE-ed version of K in the KV cache!
ggml_cpy(ctx, k_cur, k_cache_view);
ggml_cpy(ctx, v_cur_t, v_cache_view);
}
// if max_alibi_bias > 0 then apply ALiBi
struct ggml_tensor * llm_build_kqv(
struct ggml_context * ctx,
struct ggml_tensor * k_l,
struct ggml_tensor * v_l,
struct ggml_tensor * q_cur,
struct ggml_tensor * kq_mask,
float kq_scale) {
struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
struct ggml_tensor * k =
ggml_view_3d(ctx, k_l,
hp.n_embd_head, hp.n_kv, hp.n_head_kv,
ggml_row_size(k_l->type, hp.n_embd_gqa()),
ggml_row_size(k_l->type, hp.n_embd_head),
0);
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
// split cached v into n_head heads
struct ggml_tensor * v =
ggml_view_3d(ctx, v_l,
hp.n_kv, hp.n_embd_head, hp.n_head_kv,
ggml_element_size(v_l)*hp.n_ctx,
ggml_element_size(v_l)*hp.n_ctx*hp.n_embd_head,
0);
struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq);
struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, hp.n_embd_head*hp.n_head, hp.n_tokens);
struct ggml_tensor * wo = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_embd);
cur = ggml_mul_mat(ctx, wo, cur);
return cur;
}
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->type == GGML_TYPE_I32) {
// pos
std::vector<int> data(hp.n_tokens);
for (int i = 0; i < hp.n_tokens; i++) {
data[i] = rand() % hp.n_ctx;
}
ggml_backend_tensor_set(t, data.data(), 0, hp.n_tokens * sizeof(int));
} else {
init_tensor_uniform(t);
}
}
}
};
// Llama
struct test_llama : public test_llm {
static constexpr float freq_base = 10000.0f;
static constexpr float freq_scale = 1.0f;
static constexpr float ext_factor = 0.0f;
static constexpr float attn_factor = 1.0f;
static constexpr float beta_fast = 32.0f;
static constexpr float beta_slow = 1.0f;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "LLAMA";
}
std::string vars() override {
auto n_tokens = hp.n_tokens;
return VARS_TO_STR1(n_tokens);
}
double max_nmse_err() override {
return 2e-3;
}
test_llama(int n_tokens = 1)
: test_llm({
/*n_vocab =*/ 32000,
/*n_embd =*/ 3200,
/*n_head =*/ 32,
/*n_head_kv =*/ 32,
/*n_rot =*/ 100,
/*n_embd_head =*/ 100,
/*n_ff =*/ 8640,
/*f_norm_eps =*/ 0.f,
/*f_norm_rms_eps =*/ 1e-5f,
/*n_tokens =*/ n_tokens,
}) {
}
ggml_tensor * build_graph(ggml_context * ctx) override {
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hp.n_embd, hp.n_tokens);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1);
ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
for (uint32_t il = 0; il < hp.n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
// norm
ggml_tensor * attn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
cur = llm_build_norm(ctx, inpL, attn_norm, nullptr, LLM_NORM_RMS);
// self-attention
{
ggml_tensor * wq = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_embd);
ggml_tensor * wk = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_embd_gqa());
ggml_tensor * wv = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_embd_gqa());
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx, wq, cur);
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,
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,
hp.n_rot, 0, 0, hp.n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
llm_build_kv_store(ctx, k_l, v_l, Kcur, Vcur);
cur = llm_build_kqv(ctx, k_l, v_l, Qcur, KQ_mask, 1.0f/sqrtf(float(hp.n_embd_head)));
}
struct ggml_tensor * ffn_inp = ggml_add(ctx, cur, inpSA);
// feed-forward network
ggml_tensor * ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
cur = llm_build_norm(ctx, ffn_inp, ffn_norm, nullptr, LLM_NORM_RMS);
ggml_tensor * ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_ff);
ggml_tensor * ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_ff, hp.n_embd);
ggml_tensor * ffn_up = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_ff);
struct ggml_tensor * tmp = ggml_mul_mat(ctx, ffn_up, cur);
cur = ggml_mul_mat(ctx, ffn_gate, cur);
cur = ggml_silu(ctx, cur);
cur = ggml_mul(ctx, cur, tmp);
cur = ggml_mul_mat(ctx, ffn_down, cur);
cur = ggml_add(ctx, cur, ffn_inp);
// input for next layer
inpL = cur;
}
cur = inpL;
ggml_tensor * output_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
cur = llm_build_norm(ctx, cur, output_norm, nullptr, LLM_NORM_RMS);
// lm_head
ggml_tensor * output = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_vocab);
cur = ggml_mul_mat(ctx, output, cur);
return cur;
}
};
// Falcon
struct test_falcon : public test_llm {
static constexpr float freq_base = 10000.0f;
static constexpr float freq_scale = 1.0f;
static constexpr float ext_factor = 0.0f;
static constexpr float attn_factor = 1.0f;
static constexpr float beta_fast = 32.0f;
static constexpr float beta_slow = 1.0f;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "FALCON";
}
std::string vars() override {
auto n_tokens = hp.n_tokens;
return VARS_TO_STR1(n_tokens);
}
double max_nmse_err() override {
return 2e-3;
}
test_falcon(int n_tokens = 1)
: test_llm({
/*n_vocab =*/ 32000,
/*n_embd =*/ 3200,
/*n_head =*/ 50,
/*n_head_kv =*/ 1,
/*n_rot =*/ 64,
/*n_embd_head =*/ 64,
/*n_ff =*/ 8640,
/*f_norm_eps =*/ 1e-5f,
/*f_norm_rms_eps =*/ 0.f,
/*n_tokens =*/ n_tokens,
}) {
}
ggml_tensor * build_graph(ggml_context * ctx) override {
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hp.n_embd, hp.n_tokens);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1);
ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
for (uint32_t il = 0; il < hp.n_layer; ++il) {
// norm
ggml_tensor * attn_norm_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
ggml_tensor * attn_norm_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
ggml_tensor * attn_norm = llm_build_norm(ctx, inpL, attn_norm_w, attn_norm_b, LLM_NORM);
// self-attention
{
cur = attn_norm;
ggml_tensor * wqkv = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_embd + 2*hp.n_embd_gqa());
cur = ggml_mul_mat(ctx, wqkv, cur);
struct ggml_tensor * Qcur = ggml_cont(ctx, ggml_view_2d(ctx, cur, hp.n_embd, hp.n_tokens, cur->nb[1], 0*sizeof(float)*(hp.n_embd)));
struct ggml_tensor * Kcur = ggml_cont(ctx, ggml_view_2d(ctx, cur, hp.n_embd_gqa(), hp.n_tokens, cur->nb[1], 1*sizeof(float)*(hp.n_embd)));
struct ggml_tensor * Vcur = ggml_cont(ctx, ggml_view_2d(ctx, cur, hp.n_embd_gqa(), hp.n_tokens, cur->nb[1], 1*sizeof(float)*(hp.n_embd + hp.n_embd_gqa())));
Qcur = ggml_reshape_3d(ctx, Qcur, hp.n_embd_head, hp.n_head, hp.n_tokens);
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,
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,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
llm_build_kv_store(ctx, k_l, v_l, Kcur, Vcur);
cur = llm_build_kqv(ctx, k_l, v_l, Qcur, KQ_mask, 1.0f/sqrtf(float(hp.n_embd_head)));
}
struct ggml_tensor * ffn_inp = cur;
// feed forward
{
ggml_tensor * ffn_up = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_embd, hp.n_ff);
ggml_tensor * ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, hp.n_ff, hp.n_embd);
cur = attn_norm;
cur = ggml_mul_mat(ctx, ffn_up, cur);
cur = ggml_gelu(ctx, cur);
cur = ggml_mul_mat(ctx, ffn_down, cur);
}
cur = ggml_add(ctx, cur, ffn_inp);
cur = ggml_add(ctx, cur, inpL);
// input for next layer
inpL = cur;
}
cur = inpL;
ggml_tensor * output_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
ggml_tensor * output_norm_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hp.n_embd);
cur = llm_build_norm(ctx, cur, output_norm, output_norm_b, LLM_NORM);
// lm_head
ggml_tensor * output = ggml_new_tensor_2d(ctx, GGML_TYPE_Q8_0, hp.n_embd, hp.n_vocab);
cur = ggml_mul_mat(ctx, output, cur);
return cur;
}
};
static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) { static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
std::vector<std::unique_ptr<test_case>> test_cases; std::vector<std::unique_ptr<test_case>> test_cases;
std::default_random_engine rng(0); std::default_random_engine rng(0);
@ -1626,6 +2029,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
exponent <<= 1; exponent <<= 1;
} }
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, 0.1f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, 0.1f, true));
for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) { 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, 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, 40, 10, 1}, 128, 0, 512)); // llama 13B
@ -1662,6 +2068,14 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
//test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336)); //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
#endif #endif
// these tests are disabled to save execution time, but they can be handy for debugging
#if 0
test_cases.emplace_back(new test_llama(1));
test_cases.emplace_back(new test_llama(2));
test_cases.emplace_back(new test_falcon(1));
test_cases.emplace_back(new test_falcon(2));
#endif
// run tests // run tests
if (mode == MODE_TEST) { if (mode == MODE_TEST) {
ggml_backend_t backend_cpu = ggml_backend_cpu_init(); ggml_backend_t backend_cpu = ggml_backend_cpu_init();

View File

@ -1,3 +1,7 @@
#include "llama.h" #include "llama.h"
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
int main(void) {} int main(void) {}