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

Merge branch 'master' of https://github.com/ggerganov/llama.cpp

Browse source
This commit is contained in:
Jason Flax 2024-11-10 18:21:17 -05:00
commit 9c76cdba16
71 changed files with 17066 additions and 20938 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.github/workflows/build.yml vendored
View file

@ -92,7 +92,7 @@ jobs:
name: llama-bin-macos-arm64.zip
macOS-latest-cmake-x64:
runs-on: macos-12
runs-on: macos-13
steps:
- name: Clone

13
CMakePresets.json
View file

@ -48,10 +48,23 @@
}
},
{
"name": "arm64-apple-clang", "hidden": true,
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x64", "strategy": "external" },
"cacheVariables": {
"CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
}
},
{ "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug" , "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] },

27
Makefile
View file

@ -1,7 +1,6 @@
# Define the default target now so that it is always the first target
BUILD_TARGETS = \
libllava.a \
llama-baby-llama \
llama-batched \
llama-batched-bench \
llama-bench \
@ -34,6 +33,7 @@ BUILD_TARGETS = \
llama-save-load-state \
llama-server \
llama-simple \
llama-simple-chat \
llama-speculative \
llama-tokenize \
llama-vdot \
@ -55,7 +55,6 @@ TEST_TARGETS = \
tests/test-llama-grammar \
tests/test-log \
tests/test-model-load-cancel \
tests/test-opt \
tests/test-quantize-fns \
tests/test-quantize-perf \
tests/test-rope \
@ -63,6 +62,7 @@ TEST_TARGETS = \
tests/test-tokenizer-0 \
tests/test-tokenizer-1-bpe \
tests/test-tokenizer-1-spm
# tests/test-opt \
# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
@ -915,6 +915,7 @@ endif # GGML_METAL
OBJ_GGML += \
ggml/src/ggml.o \
ggml/src/ggml-cpu.o \
ggml/src/ggml-alloc.o \
ggml/src/ggml-backend.o \
ggml/src/ggml-quants.o \
@ -935,7 +936,6 @@ OBJ_COMMON = \
common/console.o \
common/ngram-cache.o \
common/sampling.o \
common/train.o \
common/build-info.o \
common/json-schema-to-grammar.o
@ -1047,6 +1047,12 @@ ggml/src/ggml.o: \
ggml/include/ggml.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-cpu.o: \
ggml/src/ggml-cpu.c \
ggml/include/ggml.h \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-alloc.o: \
ggml/src/ggml-alloc.c \
ggml/include/ggml.h \
@ -1212,11 +1218,6 @@ common/json-schema-to-grammar.o: \
common/json-schema-to-grammar.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/train.o: \
common/train.cpp \
common/train.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/ngram-cache.o: \
common/ngram-cache.cpp \
common/ngram-cache.h
@ -1287,6 +1288,11 @@ llama-simple: examples/simple/simple.cpp \
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-simple-chat: examples/simple-chat/simple-chat.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-tokenize: examples/tokenize/tokenize.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@ -1384,11 +1390,6 @@ llama-bench: examples/llama-bench/llama-bench.cpp \
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-baby-llama: examples/baby-llama/baby-llama.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-export-lora: examples/export-lora/export-lora.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

83
Package.swift
View file

@ -20,6 +20,7 @@ var ggmlSources = [
"src/ggml.c",
"src/ggml-alloc.c",
"src/ggml-backend.cpp",
"src/ggml-cpu.c",
"src/ggml-quants.c",
"src/ggml-aarch64.c"
]
@ -34,10 +35,30 @@ var cSettings: [CSetting] = [
.define("ACCELERATE_NEW_LAPACK"),
.define("ACCELERATE_LAPACK_ILP64"),
]
var sources = [
"src/llama.cpp",
"src/llama-vocab.cpp",
"src/llama-grammar.cpp",
"src/llama-sampling.cpp",
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-cpu.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.cpp",
"ggml/src/ggml-quants.c",
"ggml/src/ggml-aarch64.c",
"common/sampling.cpp",
"common/common.cpp",
"common/json-schema-to-grammar.cpp",
"common/log.cpp",
]
#if canImport(Darwin)
sources.append("ggml/src/ggml-metal.m")
ggmlSources.append("src/ggml-metal.m")
resources.append(.process("src/ggml-metal.metal"))
//resources.append(.process("src/ggml-metal.metal"))
resources.append(.process("ggml/src/ggml-metal.metal"))
linkerSettings.append(.linkedFramework("Accelerate"))
cSettings.append(
contentsOf: [
@ -67,30 +88,48 @@ let package = Package(
.package(url: "https://github.com/apple/swift-syntax.git", branch: "main")
],
targets: [
.target(name: "llama_cpp",
path: ".",
exclude: [
"cmake",
"examples",
"scripts",
"models",
"tests",
"CMakeLists.txt",
"Makefile",
"ggml"
],
sources: cppSources,
publicHeadersPath: "spm-headers",
cSettings: cSettings),
.target(
name: "llama",
dependencies: ["llama_cpp"],
path: "ggml",
sources: ggmlSources,
path: ".",
exclude: [
"cmake",
"examples",
"scripts",
"models",
"tests",
"CMakeLists.txt",
"Makefile"
],
sources: sources,
resources: resources,
publicHeadersPath: "include",
publicHeadersPath: "spm-headers",
cSettings: cSettings,
linkerSettings: linkerSettings),
linkerSettings: linkerSettings
),
// .target(name: "llama_cpp",
// path: ".",
// exclude: [
// "cmake",
// "examples",
// "scripts",
// "models",
// "tests",
// "CMakeLists.txt",
// "Makefile",
// "ggml"
// ],
// sources: cppSources,
// publicHeadersPath: "spm-headers",
// cSettings: cSettings),
// .target(
// name: "llama",
// dependencies: ["llama_cpp"],
// path: "ggml",
// sources: ggmlSources,
// resources: resources,
// publicHeadersPath: "include",
// cSettings: cSettings,
// linkerSettings: linkerSettings),
.target(name: "LlamaObjC",
dependencies: ["llama"],
path: "objc",

3
README.md
View file

@ -17,7 +17,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
## Hot topics
- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
----

164
ci/run.sh
View file

@ -326,36 +326,36 @@ function gg_run_open_llama_7b_v2 {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-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/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -460,34 +460,34 @@ function gg_run_pythia_1_4b {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --model ${model_f16} -ngl 99 -c 0 -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/llama-cli --model ${model_q8_0} -ngl 99 -c 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/llama-cli --model ${model_q4_0} -ngl 99 -c 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/llama-cli --model ${model_q4_1} -ngl 99 -c 0 -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/llama-cli --model ${model_q5_0} -ngl 99 -c 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/llama-cli --model ${model_q5_1} -ngl 99 -c 0 -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/llama-cli --model ${model_q2_k} -ngl 99 -c 0 -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/llama-cli --model ${model_q3_k} -ngl 99 -c 0 -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/llama-cli --model ${model_q4_k} -ngl 99 -c 0 -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/llama-cli --model ${model_q5_k} -ngl 99 -c 0 -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/llama-cli --model ${model_q6_k} -ngl 99 -c 0 -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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -591,36 +591,36 @@ function gg_run_pythia_2_8b {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --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/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -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/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-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/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-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/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -706,8 +706,8 @@ function gg_run_embd_bge_small {
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
set +e
}
@ -752,7 +752,7 @@ function gg_run_rerank_tiny {
model_f16="${path_models}/ggml-model-f16.gguf"
# for this model, the SEP token is "</s>"
(time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
(time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
# sample output
# rerank score 0: 0.029

16
cmake/arm64-apple-clang.cmake Normal file
View file

@ -0,0 +1,16 @@
set( CMAKE_SYSTEM_NAME Darwin )
set( CMAKE_SYSTEM_PROCESSOR arm64 )
set( target arm64-apple-darwin-macho )
set( CMAKE_C_COMPILER clang )
set( CMAKE_CXX_COMPILER clang++ )
set( CMAKE_C_COMPILER_TARGET ${target} )
set( CMAKE_CXX_COMPILER_TARGET ${target} )
set( arch_c_flags "-march=armv8.4-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function" )
set( CMAKE_C_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )
set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

2
common/CMakeLists.txt
View file

@ -66,8 +66,6 @@ add_library(${TARGET} STATIC
ngram-cache.h
sampling.cpp
sampling.h
train.cpp
train.h
)
if (BUILD_SHARED_LIBS)

2
common/common.cpp
View file

@ -1951,6 +1951,8 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
ggml_cpu_init(); // some ARM features are detected at runtime
const auto & sparams = params.sparams;
fprintf(stream, "build_commit: %s\n", LLAMA_COMMIT);

2
common/common.h
View file

@ -155,7 +155,7 @@ struct common_sampler_params {
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_ctx = 4096; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_ubatch = 512; // physical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt

1515
common/train.cpp
View file

File diff suppressed because it is too large Load diff

233
common/train.h
View file

@ -1,233 +0,0 @@
// Various helper functions and utilities for training
#pragma once
#include <string>
#include <random>
#include <vector>
#include "ggml.h"
#include "llama.h"
#define LLAMA_TRAIN_MAX_NODES 16384
typedef std::string mt19937_state;
struct train_state {
struct ggml_opt_context * opt;
uint64_t train_its;
uint64_t train_samples;
uint64_t train_tokens;
uint64_t train_epochs;
size_t shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
mt19937_state shuffle_rng_state_current;
mt19937_state shuffle_rng_state_next;
size_t shuffle_sample_count;
size_t shuffle_next_sample;
};
struct train_params_common {
const char * fn_train_data;
const char * fn_checkpoint_in;
const char * fn_checkpoint_out;
const char * pattern_fn_it;
const char * fn_latest;
bool print_usage;
int save_every;
uint32_t seed;
int n_ctx;
int n_threads;
int n_batch;
int n_gradient_accumulation;
int n_epochs;
int n_gpu_layers;
bool custom_n_ctx;
bool use_flash;
bool use_checkpointing;
std::string sample_start;
bool include_sample_start;
bool escape;
bool overlapping_samples;
bool fill_with_next_samples;
bool separate_with_eos;
bool separate_with_bos;
bool sample_random_offsets;
bool force_reshuffle;
int warmup;
int cos_decay_steps;
float cos_decay_restart;
float cos_decay_min;
bool enable_restart;
int opt_past;
float opt_delta;
int opt_max_no_improvement;
int adam_n_iter;
float adam_alpha;
float adam_min_alpha;
float adam_decay;
int adam_decay_min_ndim;
float adam_beta1;
float adam_beta2;
float adam_gclip;
float adam_eps_f;
};
typedef void (*save_train_files_callback)(void * data, struct train_state * train);
struct train_opt_callback_data {
struct train_params_common * params;
struct train_state * train;
save_train_files_callback save_cb;
void * save_data;
struct llama_context * lctx;
int last_save_iter;
llama_token * tokens_data;
size_t tokens_size;
size_t * samples_begin;
size_t * samples_size;
size_t * shuffled_samples_offs;
size_t * shuffled_samples_begin;
size_t * shuffled_samples_size;
size_t samples_count;
struct ggml_tensor * tokens_input;
struct ggml_tensor * target_probs;
int first_iter;
int first_epoch;
int iter_at_last_epoch;
int64_t last_time;
double millis_per_iter;
};
struct train_state * init_train_state();
void free_train_state(struct train_state * state);
struct train_params_common get_default_train_params_common();
void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
void finish_processing_train_args(struct train_params_common * params);
struct random_normal_distribution;
struct random_uniform_distribution;
struct random_normal_distribution * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
void free_random_normal_distribution (struct random_normal_distribution * rnd);
void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
// generate random float in interval [0,1)
float frand();
float frand_normal (struct random_normal_distribution * rnd);
float frand_uniform(struct random_uniform_distribution * rnd);
int clamp (const int v, const int min, const int max);
float fclamp(const float v, const float min, const float max);
void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
size_t tokenize_file(
struct llama_context * lctx,
const char * filename,
const std::string & sample_start,
bool include_sample_start,
bool overlapping_samples,
unsigned context_length,
std::vector<llama_token> & out_tokens,
std::vector<size_t> & out_samples_begin,
std::vector<size_t> & out_samples_size);
int64_t get_example_targets_batch(
struct llama_context * lctx,
struct ggml_tensor * tokens_input,
struct ggml_tensor * target_probs,
int64_t example_id,
const size_t * samples_offs,
const size_t * samples_begin,
const size_t * samples_size,
size_t samples_count,
const llama_token * train_data,
size_t n_train_data,
bool separate_with_eos,
bool separate_with_bos,
bool fill_with_next_samples,
bool sample_random_offsets);
void mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
mt19937_state mt19937_get_state(const std::mt19937& rng);
mt19937_state mt19937_seed_to_state(unsigned seed);
mt19937_state shuffle_samples(
const mt19937_state & rng_state,
size_t * shuffled_offs,
size_t * shuffled_begins,
size_t * shuffled_sizes,
const size_t * begins,
const size_t * sizes,
size_t count);
size_t hash_combine(size_t h1, size_t h2);
size_t compute_samples_hash(
const char* fn,
const size_t* samples_begin,
const size_t* samples_size,
size_t sample_count);
std::string replace_str(const char * s, const char * needle, const char * replacement);
void print_duration(double milliseconds);
float cosine_decay(
int64_t step,
int64_t decay_steps,
float minimum);
float cosine_decay_restart(
int64_t step,
int64_t decay_steps,
float minimum,
float restart_step_mult);
float learning_schedule(
int64_t step,
int64_t warmup_steps,
int64_t decay_steps,
float learning_rate,
float overall_minimum,
float cos_decay_minimum,
float cos_decay_restart_step_mult,
bool enable_restart);
void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

27
convert_hf_to_gguf.py
View file

@ -72,7 +72,8 @@ class Model:
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
use_temp_file: bool = False, eager: bool = False,
metadata_override: Path | None = None, model_name: str | None = None,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
if type(self) is Model:
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
@ -87,7 +88,7 @@ class Model:
self.is_safetensors = len(self.part_names) > 0
if not self.is_safetensors:
self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
self.hparams = Model.load_hparams(self.dir_model)
self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
self.tensor_names = None
@ -1541,6 +1542,17 @@ class LlamaModel(Model):
special_vocab._set_special_token("eot", 32010)
special_vocab.add_to_gguf(self.gguf_writer)
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 "add_prefix_space" in tokenizer_config_json:
self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
# Apply to granite small models only
if self.hparams.get("vocab_size", 32000) == 49152:
self.gguf_writer.add_add_bos_token(False)
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
@ -1557,17 +1569,6 @@ class LlamaModel(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
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 "add_prefix_space" in tokenizer_config_json:
self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
# Apply to granite small models only
if self.hparams.get("vocab_size", 32000) == 49152:
self.gguf_writer.add_add_bos_token(False)
@staticmethod
def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
if n_head_kv is not None and n_head != n_head_kv:

47
convert_lora_to_gguf.py
View file

@ -12,6 +12,7 @@ import json
from math import prod
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
from transformers import AutoConfig
import torch
@ -256,8 +257,8 @@ def parse_args() -> argparse.Namespace:
help="only print out what will be done, without writing any new files",
)
parser.add_argument(
"--base", type=Path, required=True,
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required",
"--base", type=Path,
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
)
parser.add_argument(
"lora_path", type=Path,
@ -267,6 +268,12 @@ def parse_args() -> argparse.Namespace:
return parser.parse_args()
def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
# normally, adapter does not come with base model config, we need to load it from AutoConfig
config = AutoConfig.from_pretrained(hf_model_id)
return config.to_dict()
if __name__ == '__main__':
args = parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
@ -281,7 +288,7 @@ if __name__ == '__main__':
ftype = ftype_map[args.outtype]
dir_base_model: Path = args.base
dir_base_model: Path | None = args.base
dir_lora: Path = args.lora_path
lora_config = dir_lora / "adapter_config.json"
input_model = dir_lora / "adapter_model.safetensors"
@ -301,9 +308,29 @@ if __name__ == '__main__':
input_model = os.path.join(dir_lora, "adapter_model.bin")
lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
# load LoRA config
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
# load base model
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = Model.load_hparams(dir_base_model)
if dir_base_model is None:
if "base_model_name_or_path" in lparams:
model_id = lparams["base_model_name_or_path"]
logger.info(f"Loading base model from Hugging Face: {model_id}")
try:
hparams = load_hparams_from_hf(model_id)
except OSError as e:
logger.error(f"Failed to load base model config: {e}")
logger.error("Please try downloading the base model and add its path to --base")
sys.exit(1)
else:
logger.error("'base_model_name_or_path' is not found in adapter_config.json")
logger.error("Base model config is required. Please download the base model and add its path to --base")
sys.exit(1)
else:
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = Model.load_hparams(dir_base_model)
with torch.inference_mode():
try:
model_class = Model.from_model_architecture(hparams["architectures"][0])
@ -323,13 +350,15 @@ if __name__ == '__main__':
self.dir_model_card = dir_lora_model
self.lora_alpha = float(lora_alpha)
def set_vocab(self):
pass
def set_type(self):
self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
def set_gguf_parameters(self):
self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
super().set_gguf_parameters()
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
# Never add extra tensors (e.g. rope_freqs) for LoRA adapters
@ -350,7 +379,7 @@ if __name__ == '__main__':
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
logger.error("Hint: if you are using TRL, make sure not to call setup_chat_format()")
logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
sys.exit(1)
if base_name in tensor_map:
@ -384,9 +413,6 @@ if __name__ == '__main__':
yield (dest_name + ".lora_a", lora_a)
yield (dest_name + ".lora_b", lora_b)
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
alpha: float = lparams["lora_alpha"]
model_instance = LoraModel(
@ -399,6 +425,7 @@ if __name__ == '__main__':
dry_run=args.dry_run,
dir_lora_model=dir_lora,
lora_alpha=alpha,
hparams=hparams,
)
logger.info("Exporting model...")

2
examples/CMakeLists.txt
View file

@ -13,7 +13,6 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN)
else()
add_subdirectory(cvector-generator)
add_subdirectory(baby-llama)
add_subdirectory(batched-bench)
add_subdirectory(batched)
add_subdirectory(convert-llama2c-to-ggml)
@ -49,6 +48,7 @@ else()
endif()
add_subdirectory(save-load-state)
add_subdirectory(simple)
add_subdirectory(simple-chat)
add_subdirectory(speculative)
add_subdirectory(tokenize)
endif()

5
examples/baby-llama/CMakeLists.txt
View file

@ -1,5 +0,0 @@
set(TARGET llama-baby-llama)
add_executable(${TARGET} baby-llama.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

1639
examples/baby-llama/baby-llama.cpp
View file

File diff suppressed because it is too large Load diff

1
examples/llava/clip.cpp
View file

@ -4,6 +4,7 @@
// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
#include "clip.h"
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"

2
examples/rpc/rpc-server.cpp
View file

@ -1,3 +1,5 @@
#include "ggml-cpu.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif

11
examples/server/README.md
View file

@ -692,7 +692,10 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
### GET `/slots`: Returns the current slots processing state
This endpoint can be disabled with `--no-slots`
> [!WARNING]
> This endpoint is intended for debugging and may be modified in future versions. For security reasons, we strongly advise against enabling it in production environments.
This endpoint is disabled by default and can be enabled with `--slots`
If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
@ -709,6 +712,7 @@ Example:
"grammar": "",
"id": 0,
"ignore_eos": false,
"is_processing": false,
"logit_bias": [],
"min_p": 0.05000000074505806,
"mirostat": 0,
@ -741,7 +745,6 @@ Example:
"temperature"
],
"seed": 42,
"state": 1,
"stop": [
"\n"
],
@ -755,10 +758,6 @@ Example:
]
```
Possible values for `slot[i].state` are:
- `0`: SLOT_STATE_IDLE
- `1`: SLOT_STATE_PROCESSING
### GET `/metrics`: Prometheus compatible metrics exporter
This endpoint is only accessible if `--metrics` is set.

73
examples/server/server.cpp
View file

@ -247,6 +247,7 @@ struct server_slot {
if (is_processing()) {
SLT_INF(*this, "stop processing: n_past = %d, truncated = %d\n", n_past, truncated);
t_last_used = ggml_time_us();
t_token_generation = (ggml_time_us() - t_start_generation) / 1e3;
state = SLOT_STATE_IDLE;
callback_on_release(id);
@ -725,12 +726,12 @@ struct server_context {
return nullptr;
}
server_slot * get_available_slot(const std::string & prompt) {
server_slot * get_available_slot(const server_task & task) {
server_slot * ret = nullptr;
// find the slot that has at least n% prompt similarity
if (ret == nullptr && slot_prompt_similarity != 0.0f && !prompt.empty()) {
int max_lcp_len = 0;
if (ret == nullptr && slot_prompt_similarity != 0.0f) {
int lcs_len = 0;
float similarity = 0;
for (server_slot & slot : slots) {
@ -740,25 +741,26 @@ struct server_context {
}
// skip the slot if it does not contains cached tokens
if (slot.prompt_tokens.empty()) {
if (slot.cache_tokens.empty()) {
continue;
}
// length of the Longest Common Prefix between the current slot's prompt and the input prompt
int lcp_len = longest_common_prefix(slot.cache_tokens, slot.prompt_tokens);
// length of the Longest Common Subsequence between the current slot's prompt and the input prompt
int cur_lcs_len = longest_common_subsequence(slot.cache_tokens, task.prompt_tokens);
// fraction of the common substring length compared to the current slot's prompt length
similarity = static_cast<float>(lcp_len) / static_cast<int>(slot.prompt_tokens.size());
// fraction of the common subsequence length compared to the current slot's prompt length
float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(slot.cache_tokens.size());
// select the current slot if the criteria match
if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) {
max_lcp_len = lcp_len;
if (cur_lcs_len > lcs_len && cur_similarity > slot_prompt_similarity) {
lcs_len = cur_lcs_len;
similarity = cur_similarity;
ret = &slot;
}
}
if (ret != nullptr) {
SLT_DBG(*ret, "selected slot by lcp similarity, max_lcp_len = %d, similarity = %f\n", max_lcp_len, similarity);
SLT_DBG(*ret, "selected slot by lcs similarity, lcs_len = %d, similarity = %f\n", lcs_len, similarity);
}
}
@ -1514,18 +1516,7 @@ struct server_context {
{
const int id_slot = json_value(task.data, "id_slot", -1);
server_slot * slot;
if (id_slot != -1) {
slot = get_slot_by_id(id_slot);
} else {
std::string prompt;
if (task.data.contains("prompt") && task.data.at("prompt").is_string()) {
prompt = json_value(task.data, "prompt", std::string());
}
slot = get_available_slot(prompt);
}
server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task);
if (slot == nullptr) {
// if no slot is available, we defer this task for processing later
@ -1575,11 +1566,11 @@ struct server_context {
for (server_slot & slot : slots) {
json slot_data = get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["id_task"] = slot.id_task;
slot_data["state"] = slot.state;
slot_data["prompt"] = common_detokenize(ctx, slot.prompt_tokens);
slot_data["next_token"] = {
slot_data["id"] = slot.id;
slot_data["id_task"] = slot.id_task;
slot_data["is_processing"] = slot.is_processing();
slot_data["prompt"] = common_detokenize(ctx, slot.prompt_tokens);
slot_data["next_token"] = {
{"has_next_token", slot.has_next_token},
{"has_new_line", slot.has_new_line},
{"n_remain", slot.n_remaining},
@ -1590,10 +1581,10 @@ struct server_context {
{"stopping_word", slot.stopping_word},
};
if (slot_data["state"] == SLOT_STATE_IDLE) {
n_idle_slots++;
} else {
if (slot.is_processing()) {
n_processing_slots++;
} else {
n_idle_slots++;
}
slots_data.push_back(slot_data);
@ -2714,8 +2705,8 @@ int main(int argc, char ** argv) {
};
const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) {
if (ctx_server.params.embedding || ctx_server.params.reranking) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
if (ctx_server.params.embedding) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
return;
}
@ -2820,8 +2811,8 @@ int main(int argc, char ** argv) {
// TODO: maybe merge this function with "handle_completions_generic"
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) {
if (ctx_server.params.embedding || ctx_server.params.reranking) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
if (ctx_server.params.embedding) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
return;
}
@ -2946,11 +2937,6 @@ int main(int argc, char ** argv) {
};
const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
// TODO: somehow clean up this checks in the future
if (!ctx_server.params.embedding || ctx_server.params.reranking) {
res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings` and without `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
return;
}
const json body = json::parse(req.body);
bool is_openai = false;
@ -3002,10 +2988,11 @@ int main(int argc, char ** argv) {
};
const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
if (!ctx_server.params.reranking) {
res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
if (!ctx_server.params.reranking || ctx_server.params.embedding) {
res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
return;
}
const json body = json::parse(req.body);
// TODO: implement
@ -3259,7 +3246,7 @@ int main(int argc, char ** argv) {
ctx_server.queue_tasks.terminate();
};
LOG_INF("%s: server is listening on %s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
ctx_server.queue_tasks.start_loop();

10
examples/server/tests/features/steps/steps.py
View file

@ -260,13 +260,13 @@ async def step_wait_for_server_status(context, expecting_status: Literal['health
async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
match expected_slot_status_string:
case 'idle':
expected_slot_status = 0
expected_slot_status = False
case 'busy':
expected_slot_status = 1
expected_slot_status = True
case _:
assert False, "unknown status"
expected_slots = [{'id': slot_id, 'state': expected_slot_status}
expected_slots = [{'id': slot_id, 'is_processing': expected_slot_status}
for slot_id in range(context.n_slots)]
await request_slots_status(context, expected_slots)
@ -1354,8 +1354,8 @@ async def wait_for_slots_status(context,
if status_code == 503 and status_code == expected_http_status_code:
return
if status_code == 200 and status_code == expected_http_status_code:
n_slots_idle = sum(1 if slot["state"] == 0 else 0 for slot in slots)
n_slots_processing = sum(1 if slot["state"] != 0 else 0 for slot in slots)
n_slots_idle = sum(1 if not slot["is_processing"] else 0 for slot in slots)
n_slots_processing = sum(1 if slot["is_processing"] else 0 for slot in slots)
if ((slots_idle is None or slots_idle == n_slots_idle)
and (slots_processing is None or slots_processing == n_slots_processing)):
return

52
examples/server/utils.hpp
View file

@ -439,18 +439,60 @@ static std::string gen_chatcmplid() {
// other common utils
//
static size_t longest_common_prefix(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
static size_t longest_common_prefix(const llama_tokens & a, const llama_tokens & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
return i;
}
static size_t longest_common_prefix(const std::string & a, const std::string & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
static size_t longest_common_subsequence(const llama_tokens & a, const llama_tokens & b) {
// check for empty sequences
if (a.empty() || b.empty()) {
return 0;
}
return i;
// get the lengths of the input sequences
size_t a_len = a.size();
size_t b_len = b.size();
// initialize the maximum length of the longest common subsequence (LCS)
size_t max_length = 0;
// use two rows instead of a 2D matrix to optimize space
std::vector<size_t> prev_row(b_len + 1, 0);
std::vector<size_t> curr_row(b_len + 1, 0);
// iterate through the elements of a
for (size_t i = 1; i <= a_len; i++) {
// iterate through the elements of b
for (size_t j = 1; j <= b_len; j++) {
// if elements at the current positions match
if (a[i - 1] == b[j - 1]) {
// if it's the first element of either sequences, set LCS length to 1
if (i == 1 || j == 1) {
curr_row[j] = 1;
} else {
// increment LCS length by 1 compared to the previous element
curr_row[j] = prev_row[j - 1] + 1;
}
// update max_length if necessary
if (curr_row[j] > max_length) {
max_length = curr_row[j];
}
} else {
// reset LCS length if elements don't match
curr_row[j] = 0;
}
}
// update the previous row for the next iteration
prev_row = curr_row;
}
// return the maximum length of the LCS
return max_length;
}
static bool ends_with(const std::string & str, const std::string & suffix) {

5
examples/simple-chat/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET llama-simple-chat)
add_executable(${TARGET} simple-chat.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

7
examples/simple-chat/README.md Normal file
View file

@ -0,0 +1,7 @@
# llama.cpp/example/simple-chat
The purpose of this example is to demonstrate a minimal usage of llama.cpp to create a simple chat program using the chat template from the GGUF file.
```bash
./llama-simple-chat -m Meta-Llama-3.1-8B-Instruct.gguf -c 2048
...

197
examples/simple-chat/simple-chat.cpp Normal file
View file

@ -0,0 +1,197 @@
#include "llama.h"
#include <cstdio>
#include <cstring>
#include <iostream>
#include <string>
#include <vector>
static void print_usage(int, char ** argv) {
printf("\nexample usage:\n");
printf("\n %s -m model.gguf [-c context_size] [-ngl n_gpu_layers]\n", argv[0]);
printf("\n");
}
int main(int argc, char ** argv) {
std::string model_path;
int ngl = 99;
int n_ctx = 2048;
// parse command line arguments
for (int i = 1; i < argc; i++) {
try {
if (strcmp(argv[i], "-m") == 0) {
if (i + 1 < argc) {
model_path = argv[++i];
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-c") == 0) {
if (i + 1 < argc) {
n_ctx = std::stoi(argv[++i]);
} else {
print_usage(argc, argv);
return 1;
}
} else if (strcmp(argv[i], "-ngl") == 0) {
if (i + 1 < argc) {
ngl = std::stoi(argv[++i]);
} else {
print_usage(argc, argv);
return 1;
}
} else {
print_usage(argc, argv);
return 1;
}
} catch (std::exception & e) {
fprintf(stderr, "error: %s\n", e.what());
print_usage(argc, argv);
return 1;
}
}
if (model_path.empty()) {
print_usage(argc, argv);
return 1;
}
// only print errors
llama_log_set([](enum ggml_log_level level, const char * text, void * /* user_data */) {
if (level >= GGML_LOG_LEVEL_ERROR) {
fprintf(stderr, "%s", text);
}
}, nullptr);
// initialize the model
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = ngl;
llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
if (!model) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
// initialize the context
llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = n_ctx;
ctx_params.n_batch = n_ctx;
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
if (!ctx) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
// initialize the sampler
llama_sampler * smpl = llama_sampler_chain_init(llama_sampler_chain_default_params());
llama_sampler_chain_add(smpl, llama_sampler_init_min_p(0.05f, 1));
llama_sampler_chain_add(smpl, llama_sampler_init_temp(0.8f));
llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
// helper function to evaluate a prompt and generate a response
auto generate = [&](const std::string & prompt) {
std::string response;
// tokenize the prompt
const int n_prompt_tokens = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
std::vector<llama_token> prompt_tokens(n_prompt_tokens);
if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
GGML_ABORT("failed to tokenize the prompt\n");
}
// prepare a batch for the prompt
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
llama_token new_token_id;
while (true) {
// check if we have enough space in the context to evaluate this batch
int n_ctx = llama_n_ctx(ctx);
int n_ctx_used = llama_get_kv_cache_used_cells(ctx);
if (n_ctx_used + batch.n_tokens > n_ctx) {
printf("\033[0m\n");
fprintf(stderr, "context size exceeded\n");
exit(0);
}
if (llama_decode(ctx, batch)) {
GGML_ABORT("failed to decode\n");
}
// sample the next token
new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id)) {
break;
}
// convert the token to a string, print it and add it to the response
char buf[256];
int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
GGML_ABORT("failed to convert token to piece\n");
}
std::string piece(buf, n);
printf("%s", piece.c_str());
fflush(stdout);
response += piece;
// prepare the next batch with the sampled token
batch = llama_batch_get_one(&new_token_id, 1);
}
return response;
};
std::vector<llama_chat_message> messages;
std::vector<char> formatted(llama_n_ctx(ctx));
int prev_len = 0;
while (true) {
// get user input
printf("\033[32m> \033[0m");
std::string user;
std::getline(std::cin, user);
if (user.empty()) {
break;
}
// add the user input to the message list and format it
messages.push_back({"user", strdup(user.c_str())});
int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
if (new_len > (int)formatted.size()) {
formatted.resize(new_len);
new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
}
if (new_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
return 1;
}
// remove previous messages to obtain the prompt to generate the response
std::string prompt(formatted.begin() + prev_len, formatted.begin() + new_len);
// generate a response
printf("\033[33m");
std::string response = generate(prompt);
printf("\n\033[0m");
// add the response to the messages
messages.push_back({"assistant", strdup(response.c_str())});
prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, nullptr, 0);
if (prev_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
return 1;
}
}
// free resources
for (auto & msg : messages) {
free(const_cast<char *>(msg.content));
}
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
return 0;
}

20
flake.lock generated
View file

@ -5,11 +5,11 @@
"nixpkgs-lib": "nixpkgs-lib"
},
"locked": {
"lastModified": 1727826117,
"narHash": "sha256-K5ZLCyfO/Zj9mPFldf3iwS6oZStJcU4tSpiXTMYaaL0=",
"lastModified": 1730504689,
"narHash": "sha256-hgmguH29K2fvs9szpq2r3pz2/8cJd2LPS+b4tfNFCwE=",
"owner": "hercules-ci",
"repo": "flake-parts",
"rev": "3d04084d54bedc3d6b8b736c70ef449225c361b1",
"rev": "506278e768c2a08bec68eb62932193e341f55c90",
"type": "github"
},
"original": {
@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1729665710,
"narHash": "sha256-AlcmCXJZPIlO5dmFzV3V2XF6x/OpNWUV8Y/FMPGd8Z4=",
"lastModified": 1730200266,
"narHash": "sha256-l253w0XMT8nWHGXuXqyiIC/bMvh1VRszGXgdpQlfhvU=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "2768c7d042a37de65bb1b5b3268fc987e534c49d",
"rev": "807e9154dcb16384b1b765ebe9cd2bba2ac287fd",
"type": "github"
},
"original": {
@ -36,14 +36,14 @@
},
"nixpkgs-lib": {
"locked": {
"lastModified": 1727825735,
"narHash": "sha256-0xHYkMkeLVQAMa7gvkddbPqpxph+hDzdu1XdGPJR+Os=",
"lastModified": 1730504152,
"narHash": "sha256-lXvH/vOfb4aGYyvFmZK/HlsNsr/0CVWlwYvo2rxJk3s=",
"type": "tarball",
"url": "https://github.com/NixOS/nixpkgs/archive/fb192fec7cc7a4c26d51779e9bab07ce6fa5597a.tar.gz"
"url": "https://github.com/NixOS/nixpkgs/archive/cc2f28000298e1269cea6612cd06ec9979dd5d7f.tar.gz"
},
"original": {
"type": "tarball",
"url": "https://github.com/NixOS/nixpkgs/archive/fb192fec7cc7a4c26d51779e9bab07ce6fa5597a.tar.gz"
"url": "https://github.com/NixOS/nixpkgs/archive/cc2f28000298e1269cea6612cd06ec9979dd5d7f.tar.gz"
}
},
"root": {

19
ggml/include/ggml-backend.h
View file

@ -305,27 +305,10 @@ extern "C" {
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
//
// CPU backend
//
GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
// Create a backend buffer from an existing pointer
// CPU buffer types are always available
GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
#ifdef __cplusplus
}
#endif

38
ggml/include/ggml-cpp.h Normal file
View file

@ -0,0 +1,38 @@
#pragma once
#ifndef __cplusplus
#error "This header is for C++ only"
#endif
#include "ggml.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include <memory>
// Smart pointers for ggml types
// ggml
struct ggml_context_deleter { void operator()(ggml_context * ctx) { ggml_free(ctx); } };
struct gguf_context_deleter { void operator()(gguf_context * ctx) { gguf_free(ctx); } };
typedef std::unique_ptr<ggml_context, ggml_context_deleter> ggml_context_ptr;
typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
// ggml-alloc
struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
// ggml-backend
struct ggml_backend_deleter { void operator()(ggml_backend_t backend) { ggml_backend_free(backend); } };
struct ggml_backend_buffer_deleter { void operator()(ggml_backend_buffer_t buffer) { ggml_backend_buffer_free(buffer); } };
struct ggml_backend_event_deleter { void operator()(ggml_backend_event_t event) { ggml_backend_event_free(event); } };
struct ggml_backend_sched_deleter { void operator()(ggml_backend_sched_t sched) { ggml_backend_sched_free(sched); } };
typedef std::unique_ptr<ggml_backend, ggml_backend_deleter> ggml_backend_ptr;
typedef std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> ggml_backend_buffer_ptr;
typedef std::unique_ptr<ggml_backend_event, ggml_backend_event_deleter> ggml_backend_event_ptr;
typedef std::unique_ptr<ggml_backend_sched, ggml_backend_sched_deleter> ggml_backend_sched_ptr;

150
ggml/include/ggml-cpu.h Normal file
View file

@ -0,0 +1,150 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
// Scheduling priorities
enum ggml_sched_priority {
GGML_SCHED_PRIO_NORMAL,
GGML_SCHED_PRIO_MEDIUM,
GGML_SCHED_PRIO_HIGH,
GGML_SCHED_PRIO_REALTIME
};
// Threadpool params
// Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
struct ggml_threadpool_params {
bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
int n_threads; // number of threads
enum ggml_sched_priority prio; // thread priority
uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling)
bool strict_cpu; // strict cpu placement
bool paused; // start in paused state
};
struct ggml_threadpool; // forward declaration, see ggml.c
typedef struct ggml_threadpool * ggml_threadpool_t;
// the compute plan that needs to be prepared for ggml_graph_compute()
// since https://github.com/ggerganov/ggml/issues/287
struct ggml_cplan {
size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
int n_threads;
struct ggml_threadpool * threadpool;
// abort ggml_graph_compute when true
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
// numa strategies
enum ggml_numa_strategy {
GGML_NUMA_STRATEGY_DISABLED = 0,
GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
GGML_NUMA_STRATEGY_ISOLATE = 2,
GGML_NUMA_STRATEGY_NUMACTL = 3,
GGML_NUMA_STRATEGY_MIRROR = 4,
GGML_NUMA_STRATEGY_COUNT
};
GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan(
const struct ggml_cgraph * cgraph,
int n_threads, /* = GGML_DEFAULT_N_THREADS */
struct ggml_threadpool * threadpool /* = NULL */ );
GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
// TODO: move to backend interface
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_matmul_int8(void);
// get the sve vector length in bytes
GGML_API int ggml_cpu_get_sve_cnt(void);
// Internal types and functions exposed for tests and benchmarks
typedef void (*ggml_from_float_to_mat_t)
(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
const void * GGML_RESTRICT y, size_t by, int nrc);
typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
struct ggml_type_traits_cpu {
ggml_from_float_to_mat_t from_float_to_mat;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
int64_t nrows; // number of rows to process simultaneously
int64_t ncols; // number of columns to process simultaneously
ggml_gemv_t gemv;
ggml_gemm_t gemm;
};
GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
GGML_API void ggml_cpu_init(void);
//
// CPU backend
//
GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
#ifdef __cplusplus
}
#endif

171
ggml/include/ggml.h
View file

@ -217,7 +217,6 @@
#define GGML_MAX_DIMS 4
#define GGML_MAX_PARAMS 2048
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_SRC 10
#define GGML_MAX_N_THREADS 512
#define GGML_MAX_OP_PARAMS 64
@ -559,10 +558,10 @@ extern "C" {
enum ggml_log_level {
GGML_LOG_LEVEL_NONE = 0,
GGML_LOG_LEVEL_INFO = 1,
GGML_LOG_LEVEL_WARN = 2,
GGML_LOG_LEVEL_ERROR = 3,
GGML_LOG_LEVEL_DEBUG = 4,
GGML_LOG_LEVEL_DEBUG = 1,
GGML_LOG_LEVEL_INFO = 2,
GGML_LOG_LEVEL_WARN = 3,
GGML_LOG_LEVEL_ERROR = 4,
GGML_LOG_LEVEL_CONT = 5, // continue previous log
};
@ -574,6 +573,13 @@ extern "C" {
GGML_TENSOR_FLAG_LOSS = 8, // ...defines loss for numerical optimization (multiple loss tensors add up)
};
struct ggml_init_params {
// memory pool
size_t mem_size; // bytes
void * mem_buffer; // if NULL, memory will be allocated internally
bool no_alloc; // don't allocate memory for the tensor data
};
// n-dimensional tensor
struct ggml_tensor {
enum ggml_type type;
@ -619,66 +625,6 @@ extern "C" {
// If it returns true, the computation is aborted
typedef bool (*ggml_abort_callback)(void * data);
// Scheduling priorities
enum ggml_sched_priority {
GGML_SCHED_PRIO_NORMAL,
GGML_SCHED_PRIO_MEDIUM,
GGML_SCHED_PRIO_HIGH,
GGML_SCHED_PRIO_REALTIME
};
// Threadpool params
// Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
struct ggml_threadpool_params {
bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
int n_threads; // number of threads
enum ggml_sched_priority prio; // thread priority
uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling)
bool strict_cpu; // strict cpu placement
bool paused; // start in paused state
};
struct ggml_threadpool; // forward declaration, see ggml.c
typedef struct ggml_threadpool * ggml_threadpool_t;
// the compute plan that needs to be prepared for ggml_graph_compute()
// since https://github.com/ggerganov/ggml/issues/287
struct ggml_cplan {
size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
int n_threads;
struct ggml_threadpool * threadpool;
// abort ggml_graph_compute when true
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
// scratch buffer
struct ggml_scratch {
size_t offs;
size_t size;
void * data;
};
struct ggml_init_params {
// memory pool
size_t mem_size; // bytes
void * mem_buffer; // if NULL, memory will be allocated internally
bool no_alloc; // don't allocate memory for the tensor data
};
// numa strategies
enum ggml_numa_strategy {
GGML_NUMA_STRATEGY_DISABLED = 0,
GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
GGML_NUMA_STRATEGY_ISOLATE = 2,
GGML_NUMA_STRATEGY_NUMACTL = 3,
GGML_NUMA_STRATEGY_MIRROR = 4,
GGML_NUMA_STRATEGY_COUNT
};
//
// GUID
@ -701,9 +647,6 @@ extern "C" {
// accepts a UTF-8 path, even on Windows
GGML_API FILE * ggml_fopen(const char * fname, const char * mode);
GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_API void ggml_print_object (const struct ggml_object * obj);
GGML_API void ggml_print_objects(const struct ggml_context * ctx);
@ -760,12 +703,12 @@ extern "C" {
// main
GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
GGML_API void ggml_free(struct ggml_context * ctx);
GGML_API struct ggml_context * ggml_init (struct ggml_init_params params);
GGML_API void ggml_reset(struct ggml_context * ctx);
GGML_API void ggml_free (struct ggml_context * ctx);
GGML_API size_t ggml_used_mem(const struct ggml_context * ctx);
GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx);
GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
@ -805,8 +748,7 @@ extern "C" {
int64_t ne2,
int64_t ne3);
GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
GGML_API void * ggml_new_buffer(struct ggml_context * ctx, size_t nbytes);
GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
@ -816,35 +758,25 @@ extern "C" {
GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
// Converts a flat index into coordinates
GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API void * ggml_get_data (const struct ggml_tensor * tensor);
GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name);
GGML_ATTRIBUTE_FORMAT(2, 3)
GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...);
// Tensor flags
GGML_API void ggml_set_input(struct ggml_tensor * tensor);
GGML_API void ggml_set_output(struct ggml_tensor * tensor);
GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
//
// operations on tensors with backpropagation
//
@ -2060,9 +1992,6 @@ extern "C" {
// automatic differentiation
//
GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
@ -2094,27 +2023,6 @@ extern "C" {
GGML_API size_t ggml_graph_overhead(void);
GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan(
const struct ggml_cgraph * cgraph,
int n_threads, /* = GGML_DEFAULT_N_THREADS */
struct ggml_threadpool * threadpool /* = NULL */ );
GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
@ -2285,6 +2193,8 @@ extern "C" {
} lbfgs;
};
GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
// optimize the function defined by the tensor f
@ -2316,12 +2226,6 @@ extern "C" {
ggml_opt_callback callback,
void * callback_data);
//
// tensor flags
//
GGML_API void ggml_set_input(struct ggml_tensor * tensor);
GGML_API void ggml_set_output(struct ggml_tensor * tensor);
//
// quantization
//
@ -2490,8 +2394,6 @@ extern "C" {
GGML_API int ggml_cpu_has_avx512_bf16(void);
GGML_API int ggml_cpu_has_amx_int8 (void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (void);
GGML_API int ggml_cpu_has_f16c (void);
@ -2508,17 +2410,9 @@ extern "C" {
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);
GGML_API int ggml_cpu_has_matmul_int8(void);
GGML_API int ggml_cpu_has_cann (void);
GGML_API int ggml_cpu_has_llamafile (void);
// get the sve vector length in bytes
GGML_API int ggml_cpu_get_sve_cnt(void);
//
// Internal types and functions exposed for tests and benchmarks
//
#ifdef __cplusplus
// restrict not standard in C++
#define GGML_RESTRICT
@ -2527,14 +2421,6 @@ extern "C" {
#endif
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_from_float_to_mat_t)
(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
const void * GGML_RESTRICT y, size_t by, int nrc);
typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
struct ggml_type_traits {
const char * type_name;
@ -2545,13 +2431,6 @@ extern "C" {
ggml_to_float_t to_float;
ggml_from_float_t from_float;
ggml_from_float_t from_float_ref;
ggml_from_float_to_mat_t from_float_to_mat;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
int64_t nrows; // number of rows to process simultaneously
int64_t ncols; // number of columns to process simultaneously
ggml_gemv_t gemv;
ggml_gemm_t gemm;
};
GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);

9
ggml/src/CMakeLists.txt
View file

@ -800,6 +800,7 @@ if (GGML_KOMPUTE)
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_q4_k.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f32.comp
kompute-shaders/op_getrows_f16.comp
@ -833,6 +834,7 @@ if (GGML_KOMPUTE)
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q4_k.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f32.h
shaderop_getrows_f16.h
@ -1364,9 +1366,12 @@ endif()
add_library(ggml
../include/ggml.h
../include/ggml-cpu.h
../include/ggml-alloc.h
../include/ggml-backend.h
../include/ggml-cpp.h
ggml.c
ggml-cpu.c
ggml-alloc.c
ggml-backend.cpp
ggml-quants.c
@ -1391,7 +1396,7 @@ if (EMSCRIPTEN)
endif()
target_compile_definitions(ggml PUBLIC ${GGML_CDEF_PUBLIC})
target_include_directories(ggml PUBLIC ../include)
target_include_directories(ggml PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
target_link_directories (ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
target_compile_features (ggml PRIVATE c_std_11) # don't bump
@ -1400,7 +1405,7 @@ list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
find_library(MATH_LIBRARY m)
if (MATH_LIBRARY)
if (NOT WIN32 OR NOT GGML_SYCL)
if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT})
list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
endif()
endif()

1
ggml/src/ggml-aarch64.c
View file

@ -7,6 +7,7 @@
#include "ggml-quants.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-impl.h"
#include <math.h>

1239
ggml/src/ggml-backend.cpp
View file

File diff suppressed because it is too large Load diff

1
ggml/src/ggml-cann.cpp
View file

@ -1227,7 +1227,6 @@ static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggm
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
return buffer;

13720
ggml/src/ggml-cpu.c Normal file
View file

File diff suppressed because it is too large Load diff

12
ggml/src/ggml-cuda.cu
View file

@ -1297,11 +1297,17 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
if (err != cudaErrorPeerAccessAlreadyEnabled) {
CUDA_CHECK(err);
} else {
// reset the error
cudaGetLastError();
}
} else {
cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
if (err != cudaErrorPeerAccessNotEnabled) {
CUDA_CHECK(err);
} else {
// reset the error
cudaGetLastError();
}
}
}
@ -3107,18 +3113,20 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
}
return false;
} break;
case GGML_OP_NORM:
case GGML_OP_RMS_NORM:
return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
case GGML_OP_NORM:
case GGML_OP_ADD:
case GGML_OP_ADD1:
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_RMS_NORM:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:

87
ggml/src/ggml-impl.h
View file

@ -8,6 +8,7 @@
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
#include <stdbool.h>
#include <stdint.h>
#include <string.h>
#ifdef __cplusplus
extern "C" {
@ -36,6 +37,20 @@ extern "C" {
#endif
#endif
static inline int ggml_up32(int n) {
return (n + 31) & ~31;
}
//static inline int ggml_up64(int n) {
// return (n + 63) & ~63;
//}
static inline int ggml_up(int n, int m) {
// assert m is a power of 2
GGML_ASSERT((m & (m - 1)) == 0);
return (n + m - 1) & ~(m - 1);
}
//
// logging
//
@ -51,6 +66,74 @@ void ggml_log_callback_default(enum ggml_log_level level, const char * text, voi
#define GGML_LOG_DEBUG(...) ggml_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
#define GGML_LOG_CONT(...) ggml_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__)
#define GGML_DEBUG 0
#if (GGML_DEBUG >= 1)
#define GGML_PRINT_DEBUG(...) GGML_LOG_DEBUG(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG(...)
#endif
#if (GGML_DEBUG >= 5)
#define GGML_PRINT_DEBUG_5(...) GGML_LOG_DEBUG(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG_5(...)
#endif
#if (GGML_DEBUG >= 10)
#define GGML_PRINT_DEBUG_10(...) GGML_LOG_DEBUG(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG_10(...)
#endif
// tensor params
static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
assert(params_size <= GGML_MAX_OP_PARAMS);
memcpy(tensor->op_params, params, params_size);
}
static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
return ((const int32_t *)(tensor->op_params))[i];
}
static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t i) {
assert(i < GGML_MAX_OP_PARAMS / sizeof(float));
return ((const float *)(tensor->op_params))[i];
}
static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
((int32_t *)(tensor->op_params))[i] = value;
}
static void ggml_set_op_params_f32(struct ggml_tensor * tensor, uint32_t i, float value) {
assert(i < GGML_MAX_OP_PARAMS / sizeof(float));
((float *)(tensor->op_params))[i] = value;
}
struct ggml_map_custom1_op_params {
ggml_custom1_op_t fun;
int n_tasks;
void * userdata;
};
struct ggml_map_custom2_op_params {
ggml_custom2_op_t fun;
int n_tasks;
void * userdata;
};
struct ggml_map_custom3_op_params {
ggml_custom3_op_t fun;
int n_tasks;
void * userdata;
};
// bitset
typedef uint32_t ggml_bitset_t;
@ -204,6 +287,10 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
void * ggml_aligned_malloc(size_t size);
void ggml_aligned_free(void * ptr, size_t size);
// TODO: move to threading file
void ggml_critical_section_start(void);
void ggml_critical_section_end(void);
#ifdef __cplusplus
}
#endif

42
ggml/src/ggml-kompute.cpp
View file

@ -20,6 +20,7 @@
#include "shaderop_mul_mat_q8_0.h"
#include "shaderop_mul_mat_q4_0.h"
#include "shaderop_mul_mat_q4_1.h"
#include "shaderop_mul_mat_q4_k.h"
#include "shaderop_mul_mat_q6_k.h"
#include "shaderop_mul_mat_mat_f32.h"
#include "shaderop_getrows_f32.h"
@ -1067,6 +1068,40 @@ static void ggml_vk_mul_mat_q8_0(Args&&... args) {
ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
static void ggml_vk_mul_mat_q4_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10,
int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0,
int32_t ne1, int32_t r2, int32_t r3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3;
} pushConsts {
0, 0, 0,
ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_q6_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
@ -1384,6 +1419,7 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_K:
return true;
default:
;
@ -1635,6 +1671,12 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q4_K:
ggml_vk_mul_mat_q4_k(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03
);
break;
case GGML_TYPE_Q6_K:
ggml_vk_mul_mat_q6_k(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,

7
ggml/src/ggml-metal.m
View file

@ -450,7 +450,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
#if !__has_feature(objc_arc)
[options release];
#endif
}
#if GGML_METAL_EMBED_LIBRARY
[src release];
#endif // GGML_METAL_EMBED_LIBRARY
}
}

885
ggml/src/ggml-metal.metal
View file

@ -12,6 +12,436 @@ using namespace metal;
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
constexpr constant static float kvalues_iq4nl_f[16] = {
-127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
};
// NOTE: this is not dequantizing - we are simply fitting the template
template <typename type4x4>
void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
reg = (type4x4)(*src);
}
template <typename type4x4>
void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
reg = (type4x4)(*src);
}
template <typename type4x4>
void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 1);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float md = -8.h * xb->d;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
}
}
template <typename type4x4>
void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 2);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float m = xb->m;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
}
}
template <typename type4x4>
void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 3);
const float d = xb->d;
const float md = -16.h * xb->d;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + md;
reg[i/2][2*(i%2)+1] = d * x1 + md;
}
}
template <typename type4x4>
void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 4);
const float d = xb->d;
const float m = xb->m;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + m;
reg[i/2][2*(i%2)+1] = d * x1 + m;
}
}
template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
device const int8_t * qs = ((device const int8_t *)xb->qs);
const half d = xb->d;
for (int i = 0; i < 16; i++) {
reg[i/4][i%4] = (qs[i + 16*il] * d);
}
}
template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
const float d = xb->d;
const float min = xb->dmin;
device const uint8_t * q = (device const uint8_t *)xb->qs;
float dl, ml;
uint8_t sc = xb->scales[il];
q = q + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
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);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * q = (device const uint8_t *)xb->qs;
device const uint8_t * h = (device const uint8_t *)xb->hmask;
device const int8_t * scales = (device const int8_t *)xb->scales;
q = q + 32 * (il/8) + 16 * (il&1);
h = h + 16 * (il&1);
uint8_t m = 1 << (il/2);
uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
((il/4)>0 ? 12 : 3);
uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
: (scale_2&kmask2) | ((scale_1&kmask1) << 4);
float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
const float ml = 4.f * dl;
il = (il/2) & 3;
const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
const uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl *= coef;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
}
}
static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
: uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
}
template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
device const uchar * q = xb->qs;
short is = (il/4) * 2;
q = q + (il/4) * 32 + 16 * (il&1);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.h;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
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;
short is = (il/4) * 2;
q = q + 32 * (il/4) + 16 * (il&1);
qh = qh + 16 * (il&1);
uint8_t ul = 1 << (il/2);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.f;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
const float qh_val = il<2 ? 16.f : 256.f;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
}
}
template <typename type4x4>
void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * ql = (device const uint8_t *)xb->ql;
device const uint8_t * qh = (device const uint8_t *)xb->qh;
device const int8_t * scales = (device const int8_t *)xb->scales;
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;
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;
const float ml = d_all * sc * 32.f;
const float dl = d_all * sc * coef;
for (int i = 0; i < 16; ++i) {
const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
: ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
}
template <typename type4x4>
void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
// each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
device const uint16_t * q2 = xb->qs + 4*ib32;
const uint32_t aux32_g = q2[0] | (q2[1] << 16);
const uint32_t aux32_s = q2[2] | (q2[3] << 16);
thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint16_t * q2 = xb->qs + 4*ib32;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
signs = ksigns_iq2xs[q2[2*il+1] >> 9];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * q3 = xb->qs + 8*ib32;
device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
const uint32_t aux32 = gas[0] | (gas[1] << 16);
const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 8*ib32;
device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
}
grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
}
}
template <typename type4x4>
void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * signs = qs + QK_K/8;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
for (int i = 0; i < 8; ++i) {
reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
}
}
template <typename type4x4>
void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
const float d = xb->d;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint16_t * qh = xb->qh;
const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
const uint16_t h = qh[ib32] >> 6*il;
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml;
reg[1][i] = dl * (grid1[i] >> 4) + ml;
reg[2][i] = dl * (grid2[i] & 0xf) + ml;
reg[3][i] = dl * (grid2[i] >> 4) + ml;
}
}
template <typename type4x4>
void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
device const uint16_t * sc = (device const uint16_t *)xb->scales;
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;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * qh = xb->qh + 2*ib32 + il;
const float dl = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
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)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
reg[1][i] = dl * (grid1[i] >> 4) + ml1;
reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
reg[3][i] = dl * (grid2[i] >> 4) + ml2;
}
}
template <typename type4x4>
void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
device const uint16_t * q4 = (device const uint16_t *)xb->qs;
const float d = xb->d;
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
}
template <typename type4x4>
void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
const float d = (float)xb->d * (ls - 32);
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
}
enum ggml_sort_order {
GGML_SORT_ORDER_ASC,
GGML_SORT_ORDER_DESC,
@ -2776,11 +3206,11 @@ kernel void kernel_flash_attn_ext_vec_f16(
const short iv3 = iq3 / rv3;
// load the queries from shared memory into local memory
float4 mq[D4];
float4 mq[D4/NW];
for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
mq[i] = (float4) sq4[i];
mq[ii/NW] = (float4) sq4[i];
}
// pointer to the mask
@ -2812,7 +3242,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
mk[2] = (float4) pk4[i + 2*(nb11/8)];
mk[3] = (float4) pk4[i + 3*(nb11/8)];
mqk += (float4) (mq[i] * mk);
mqk += (float4) (mq[ii/NW] * mk);
}
// reduce the results from the threads in the simdgroup
@ -2857,8 +3287,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
// O = diag(ms)*O
#pragma unroll
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;
lo[i/NW] *= ms;
lo[ii/NW] *= ms;
}
}
@ -2872,10 +3301,10 @@ kernel void kernel_flash_attn_ext_vec_f16(
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;
lo[i/NW] += pv4[i + 0*(nb21/8)] * ss[4*cc + 0];
lo[i/NW] += pv4[i + 1*(nb21/8)] * ss[4*cc + 1];
lo[i/NW] += pv4[i + 2*(nb21/8)] * ss[4*cc + 2];
lo[i/NW] += pv4[i + 3*(nb21/8)] * ss[4*cc + 3];
lo[ii/NW] += pv4[i + 0*(nb21/8)] * ss[4*cc + 0];
lo[ii/NW] += pv4[i + 1*(nb21/8)] * ss[4*cc + 1];
lo[ii/NW] += pv4[i + 2*(nb21/8)] * ss[4*cc + 2];
lo[ii/NW] += pv4[i + 3*(nb21/8)] * ss[4*cc + 3];
}
}
}
@ -3340,10 +3769,6 @@ static inline int best_index_int8(int n, constant float * val, float x) {
return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
}
constexpr constant static float kvalues_iq4nl_f[16] = {
-127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
};
kernel void kernel_cpy_f32_iq4_nl(
device const float * src0,
device void * dst,
@ -5458,440 +5883,6 @@ kernel void kernel_mul_mv_iq4_xs_f32(
kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
//============================= templates and their specializations =============================
// NOTE: this is not dequantizing - we are simply fitting the template
template <typename type4x4>
void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
float4x4 temp = *(((device float4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
half4x4 temp = *(((device half4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 1);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float md = -8.h * xb->d;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
}
}
template <typename type4x4>
void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 2);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float m = xb->m;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
}
}
template <typename type4x4>
void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 3);
const float d = xb->d;
const float md = -16.h * xb->d;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + md;
reg[i/2][2*(i%2)+1] = d * x1 + md;
}
}
template <typename type4x4>
void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 4);
const float d = xb->d;
const float m = xb->m;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + m;
reg[i/2][2*(i%2)+1] = d * x1 + m;
}
}
template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
device const int8_t * qs = ((device const int8_t *)xb->qs);
const half d = xb->d;
for (int i = 0; i < 16; i++) {
reg[i/4][i%4] = (qs[i + 16*il] * d);
}
}
template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
const float d = xb->d;
const float min = xb->dmin;
device const uint8_t * q = (device const uint8_t *)xb->qs;
float dl, ml;
uint8_t sc = xb->scales[il];
q = q + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
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);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * q = (device const uint8_t *)xb->qs;
device const uint8_t * h = (device const uint8_t *)xb->hmask;
device const int8_t * scales = (device const int8_t *)xb->scales;
q = q + 32 * (il/8) + 16 * (il&1);
h = h + 16 * (il&1);
uint8_t m = 1 << (il/2);
uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
((il/4)>0 ? 12 : 3);
uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
: (scale_2&kmask2) | ((scale_1&kmask1) << 4);
float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
const float ml = 4.f * dl;
il = (il/2) & 3;
const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
const uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl *= coef;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
}
}
static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
: uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
}
template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
device const uchar * q = xb->qs;
short is = (il/4) * 2;
q = q + (il/4) * 32 + 16 * (il&1);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.h;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
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;
short is = (il/4) * 2;
q = q + 32 * (il/4) + 16 * (il&1);
qh = qh + 16 * (il&1);
uint8_t ul = 1 << (il/2);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.f;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
const float qh_val = il<2 ? 16.f : 256.f;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
}
}
template <typename type4x4>
void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * ql = (device const uint8_t *)xb->ql;
device const uint8_t * qh = (device const uint8_t *)xb->qh;
device const int8_t * scales = (device const int8_t *)xb->scales;
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;
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;
const float ml = d_all * sc * 32.f;
const float dl = d_all * sc * coef;
for (int i = 0; i < 16; ++i) {
const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
: ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
}
template <typename type4x4>
void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
// each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
device const uint16_t * q2 = xb->qs + 4*ib32;
const uint32_t aux32_g = q2[0] | (q2[1] << 16);
const uint32_t aux32_s = q2[2] | (q2[3] << 16);
thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint16_t * q2 = xb->qs + 4*ib32;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
signs = ksigns_iq2xs[q2[2*il+1] >> 9];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * q3 = xb->qs + 8*ib32;
device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
const uint32_t aux32 = gas[0] | (gas[1] << 16);
const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 8*ib32;
device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
}
grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
}
}
template <typename type4x4>
void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * signs = qs + QK_K/8;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
for (int i = 0; i < 8; ++i) {
reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
}
}
template <typename type4x4>
void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
const float d = xb->d;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint16_t * qh = xb->qh;
const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
const uint16_t h = qh[ib32] >> 6*il;
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml;
reg[1][i] = dl * (grid1[i] >> 4) + ml;
reg[2][i] = dl * (grid2[i] & 0xf) + ml;
reg[3][i] = dl * (grid2[i] >> 4) + ml;
}
}
template <typename type4x4>
void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
device const uint16_t * sc = (device const uint16_t *)xb->scales;
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;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * qh = xb->qh + 2*ib32 + il;
const float dl = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
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)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
reg[1][i] = dl * (grid1[i] >> 4) + ml1;
reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
reg[3][i] = dl * (grid2[i] >> 4) + ml2;
}
}
template <typename type4x4>
void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
device const uint16_t * q4 = (device const uint16_t *)xb->qs;
const float d = xb->d;
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
}
template <typename type4x4>
void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
const float d = (float)xb->d * (ls - 32);
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
kernel void kernel_get_rows_q(
device const void * src0,

89
ggml/src/ggml-quants.c
View file

@ -4,7 +4,7 @@
#include "ggml-quants.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu.h"
#include <math.h>
#include <string.h>
@ -9104,10 +9104,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i m3 = _mm_set1_epi8(3);
const __m128i m32s = _mm_set1_epi8(32);
const __m128i m2 = _mm_set1_epi8(2);
const __m128i m15 = _mm_set1_epi8(15);
__m256 acc = _mm256_setzero_ps();
@ -9119,12 +9117,20 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
// handle the q6_k -32 offset separately using bsums
const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
__m128i shuffle = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
int is = 0;
for (int j = 0; j < QK_K/128; ++j) {
const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
@ -9132,26 +9138,26 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 2), m3), 4);
const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 2), m3), 4);
const __m128i q4h_4 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 4), m3), 4);
const __m128i q4h_5 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 4), m3), 4);
const __m128i q4h_6 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 6), m3), 4);
const __m128i q4h_7 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 6), m3), 4);
const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m4), q4h_0);
const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m4), q4h_1);
const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m4), q4h_2);
const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m4), q4h_3);
const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m4), q4h_4);
const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m4), q4h_5);
const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m4), q4h_6);
const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m4), q4h_7);
const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
@ -9162,15 +9168,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i q8s_0 = _mm_maddubs_epi16(m32s, q8_0);
__m128i q8s_1 = _mm_maddubs_epi16(m32s, q8_1);
__m128i q8s_2 = _mm_maddubs_epi16(m32s, q8_2);
__m128i q8s_3 = _mm_maddubs_epi16(m32s, q8_3);
__m128i q8s_4 = _mm_maddubs_epi16(m32s, q8_4);
__m128i q8s_5 = _mm_maddubs_epi16(m32s, q8_5);
__m128i q8s_6 = _mm_maddubs_epi16(m32s, q8_6);
__m128i q8s_7 = _mm_maddubs_epi16(m32s, q8_7);
__m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
__m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
__m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
@ -9180,32 +9177,20 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
__m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
__m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
p16_0 = _mm_sub_epi16(p16_0, q8s_0);
p16_1 = _mm_sub_epi16(p16_1, q8s_1);
p16_2 = _mm_sub_epi16(p16_2, q8s_2);
p16_3 = _mm_sub_epi16(p16_3, q8s_3);
p16_4 = _mm_sub_epi16(p16_4, q8s_4);
p16_5 = _mm_sub_epi16(p16_5, q8s_5);
p16_6 = _mm_sub_epi16(p16_6, q8s_6);
p16_7 = _mm_sub_epi16(p16_7, q8s_7);
const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_2 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_3 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
is += 4;
p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_2, scale_2)), p16_5);
p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_3, scale_3)), p16_7);
p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
@ -9214,8 +9199,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
}
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
}
*s = hsum_float_8(acc);

9
ggml/src/ggml-rpc.cpp
View file

@ -1296,13 +1296,6 @@ static ggml_backend_buffer_type_t ggml_backend_rpc_device_get_buffer_type(ggml_b
UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_rpc_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
UNUSED(dev);
UNUSED(max_tensor_size);
}
static bool ggml_backend_rpc_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
UNUSED(dev);
UNUSED(op);
@ -1328,7 +1321,7 @@ static const struct ggml_backend_device_i ggml_backend_rpc_device_i = {
/* .init_backend = */ ggml_backend_rpc_device_init,
/* .get_buffer_type = */ ggml_backend_rpc_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_rpc_device_buffer_from_ptr,
/* .buffer_from_host_ptr = */ NULL,
/* .supports_op = */ ggml_backend_rpc_device_supports_op,
/* .supports_buft = */ ggml_backend_rpc_device_supports_buft,
/* .offload_op = */ NULL,

5
ggml/src/ggml-vulkan.cpp
View file

@ -1047,7 +1047,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
return buf;
}
buf->size = size;
vk::BufferCreateInfo buffer_create_info{
vk::BufferCreateFlags(),
size,
@ -1075,7 +1074,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
if (memory_type_index == UINT32_MAX) {
device->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw vk::OutOfDeviceMemoryError("No suitable memory type found");
}
@ -1092,13 +1090,11 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
}
catch (const vk::SystemError& e) {
device->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw e;
}
} else {
// Out of Host/Device memory, clean up buffer
device->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw e;
}
}
@ -1111,6 +1107,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0);
buf->device = device;
buf->size = size;
#ifdef GGML_VULKAN_MEMORY_DEBUG
device->memory_logger->log_allocation(buf, size);

15449
ggml/src/ggml.c
View file

File diff suppressed because it is too large Load diff

9
ggml/src/kompute-shaders/common.comp
View file

@ -15,6 +15,7 @@
#define TWOPI_F 6.283185307179586f
#define QK_K 256
#define K_SCALE_SIZE 12
#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
@ -64,6 +65,14 @@ mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
return reg;
}
#define sizeof_block_q4_k 144
struct block_q4_k {
float16_t d;
float16_t dmin;
uint8_t scales[K_SCALE_SIZE];
uint8_t qs[QK_K/2];
};
#define sizeof_block_q6_k 210
struct block_q6_k {
uint8_t ql[QK_K/2]; // quants, lower 4 bits

133
ggml/src/kompute-shaders/op_mul_mat_q4_k.comp Normal file
View file

@ -0,0 +1,133 @@
#version 450
#include "common.comp"
#define N_DST 4
#define SIZE_OF_BLOCK sizeof_block_q4_k
layout(local_size_x = 4) in;
layout(local_size_y = 8) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { block_q4_k 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 ne02;
int ne12;
int r2;
int r3;
} pcs;
void main() {
const uint16_t kmask1 = uint16_t(0x3f3f);
const uint16_t kmask2 = uint16_t(0x0f0f);
const uint16_t kmask3 = uint16_t(0xc0c0);
const uint ix = gl_SubgroupInvocationID/8; // 0...3
const uint it = gl_SubgroupInvocationID%8; // 0...7
const uint iq = it/4; // 0 or 1
const uint ir = it%4; // 0...3
const uint nb = pcs.ne00/QK_K;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = r0 * N_DST;
const uint ib_row = first_row * nb;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
const uint xblk = ib_row + offset0 + pcs.inAOff;
const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff;
float yl[16];
float yh[16];
float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
float all_sum = 0.f;
uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
for (uint ib = ix; ib < nb; ib += 4) {
const uint blk_idx = ib + xblk;
float sumy[4] = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+0] = inB[y4+i+ 0]; sumy[0] += yl[i+0];
yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
}
for (int row = 0; row < N_DST; row++) {
uint row_idx = row * nb;
uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
uint16_t sc16[4];
sc16[0] = sc_0 & kmask1;
sc16[1] = sc_2 & kmask1;
sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
float acc1[4] = {0.f, 0.f, 0.f, 0.f};
float acc2[4] = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
acc1[0] += yl[i+0] * (q1 & 0x000F);
acc1[1] += yl[i+1] * (q1 & 0x0F00);
acc1[2] += yl[i+8] * (q1 & 0x00F0);
acc1[3] += yl[i+9] * (q1 & 0xF000);
acc2[0] += yh[i+0] * (q2 & 0x000F);
acc2[1] += yh[i+1] * (q2 & 0x0F00);
acc2[2] += yh[i+8] * (q2 & 0x00F0);
acc2[3] += yh[i+9] * (q2 & 0xF000);
}
uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
float dall = float(inA[blk_idx + row_idx].d);
float dmin = float(inA[blk_idx + row_idx].dmin);
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
(acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
(acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
(acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = subgroupAdd(sumf[row]);
if (subgroupElect()) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
}
}
}

1
include/llama.h
View file

@ -2,6 +2,7 @@
#define LLAMA_H
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-backend.h"
#include <stddef.h>

5
objc/GGMLThreadpool.mm
View file

@ -16,6 +16,11 @@
return self->threadpool;
}
- (void)dealloc
{
ggml_threadpool_free(threadpool);
[super dealloc];
}
@end
@implementation GGMLThreadpoolParams {

6
objc/GPTSampler.mm
View file

@ -46,4 +46,10 @@
common_sampler_reset(sampler);
}
- (void)dealloc
{
common_sampler_free(sampler);
[super dealloc];
}
@end

2
objc/LlamaContext.mm
View file

@ -20,8 +20,8 @@
- (void)dealloc
{
[super dealloc];
llama_free(ctx);
[super dealloc];
}
- (void)attachThreadpool:(GGMLThreadpool *)threadpool

6
objc/LlamaModel.mm
View file

@ -24,12 +24,8 @@
- (void)dealloc
{
[super dealloc];
llama_free_model(model);
}
- (LlamaContext *)context:(LlamaContextParams *)params {
return nil;
[super dealloc];
}
- (BOOL)addBOSToken {

153
objc/LlamaSession.mm
View file

@ -72,6 +72,18 @@
[outputCondition unlock];
return line;
}
- (void)dealloc
{
[outputCondition unlock];
[inputCondition unlock];
[inputCondition dealloc];
[outputCondition dealloc];
[log dealloc];
[inputQueue dealloc];
[outputQueue dealloc];
[super dealloc];
}
@end
@implementation LlamaSession {
@ -107,6 +119,9 @@
BOOL need_insert_eot;
int n_ctx;
os_log_t os_log_inst;
GGMLThreadpool *threadpool;
GGMLThreadpool *threadpool_batch;
}
- (NSString *)chat_add_and_format:(std::vector<common_chat_msg> &) chat_msgs role:(const std::string &) role content:(const std::string &) content {
@ -135,6 +150,7 @@ static BOOL file_is_empty(NSString *path) {
self = [super init];
self->_params = [params copy];
self->_mutableLastOutput = [[NSMutableString alloc] init];
self->_queue = [BlockingLineQueue new];
if (params.logging) {
os_log_inst = OS_LOG_DEFAULT;
} else {
@ -186,7 +202,6 @@ static BOOL file_is_empty(NSString *path) {
set_process_priority(ggml_sched_priority(params.cpuParams.priority));
GGMLThreadpool *threadpool_batch;
if (tpp != tpp_batch) {
threadpool_batch = [tpp_batch threadpool];
if (!threadpool_batch) {
@ -198,7 +213,7 @@ static BOOL file_is_empty(NSString *path) {
tpp.paused = true;
}
GGMLThreadpool *threadpool = [tpp threadpool];
threadpool = [tpp threadpool];
if (!threadpool) {
[NSException raise:@"ThreadpoolFailure"
format:@"threadpool create failed"];
@ -299,28 +314,28 @@ static BOOL file_is_empty(NSString *path) {
n_matching_session_tokens++;
}
if ([params.prompt length] == 0 && n_matching_session_tokens == embd_inp.size()) {
// LOG_INF("%s: using full prompt from session file\n", __func__);
os_log_info(os_log_inst, "%s: using full prompt from session file\n", __func__);
} else if (n_matching_session_tokens >= embd_inp.size()) {
// LOG_INF("%s: session file has exact match for prompt!\n", __func__);
os_log_info(os_log_inst, "%s: session file has exact match for prompt!\n", __func__);
} else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
// LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
// __func__, n_matching_session_tokens, embd_inp.size());
os_log_error(os_log_inst, "%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
__func__, n_matching_session_tokens, embd_inp.size());
} else {
// LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
// __func__, n_matching_session_tokens, embd_inp.size());
os_log_info(os_log_inst, "%s: session file matches %zu / %zu tokens of prompt\n",
__func__, n_matching_session_tokens, embd_inp.size());
}
// remove any "future" tokens that we might have inherited from the previous session
llama_kv_cache_seq_rm([self.ctx cContext], -1, n_matching_session_tokens, -1);
}
//
// os_log_debug(os_log_inst, "recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
// embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
//
os_log_debug(os_log_inst, "recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
// if we will use the cache for the full prompt without reaching the end of the cache, force
// reevaluation of the last token to recalculate the cached logits
if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
// os_log_debug(os_log_inst, "recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
os_log_debug(os_log_inst, "recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
session_tokens.resize(embd_inp.size() - 1);
}
@ -344,37 +359,21 @@ static BOOL file_is_empty(NSString *path) {
if (params.verbosePrompt) {
os_log_info(os_log_inst,
"%s: prompt: '%s'\n", __func__, [params.prompt cStringUsingEncoding:NSUTF8StringEncoding]);
// LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
os_log_info(os_log_inst, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
os_log_info(os_log_inst, "%6d -> '%s'\n", embd_inp[i],
[[self.ctx tokenToPiece:embd_inp[i]] cStringUsingEncoding:NSUTF8StringEncoding]);
}
if (params.nKeep > addBOS) {
// LOG_INF("%s: static prompt based on n_keep: '", __func__);
os_log_info(os_log_inst, "%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.nKeep; i++) {
os_log_debug(os_log_inst, "%s",
[[self.ctx tokenToPiece:embd_inp[i]] cStringUsingEncoding:NSUTF8StringEncoding]);
}
}
}
//
// // ctrl+C handling
// {
//#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
// struct sigaction sigint_action;
// sigint_action.sa_handler = sigint_handler;
// sigemptyset (&sigint_action.sa_mask);
// sigint_action.sa_flags = 0;
// sigaction(SIGINT, &sigint_action, NULL);
//#elif defined (_WIN32)
// auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
// return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
// };
// SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
//#endif
// }
//
if (params.interactive) {
os_log_info(os_log_inst, "%s: interactive mode on.\n", __func__);
@ -427,9 +426,8 @@ static BOOL file_is_empty(NSString *path) {
os_log_info(os_log_inst, "sampler seed: %u\n", [_smpl seed]);
// LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
// LOG_INF("sampler chain: %s\n", gpt_sampler_print(smpl).c_str());
//
// LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
//
os_log_info(os_log_inst, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.nBatch, params.nPredict, params.nKeep);
// group-attention state
// number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
@ -439,8 +437,8 @@ static BOOL file_is_empty(NSString *path) {
if (ga_n != 1) {
GGML_ASSERT(ga_n > 0 && "grp_attn_n must be positive"); // NOLINT
GGML_ASSERT(ga_w % ga_n == 0 && "grp_attn_w must be a multiple of grp_attn_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT
GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
os_log_info(os_log_inst, "self-extend: n_ctx_train = %d, grp_attn_n = %ld, grp_attn_w = %ld\n", n_ctx_train, static_cast<long>(ga_n), static_cast<long>(ga_w));
}
@ -493,7 +491,7 @@ static BOOL file_is_empty(NSString *path) {
return [_mutableLastOutput copy];
}
- (void)start:(BlockingLineQueue *)queue {
- (void)start {
while ((n_remain != 0 && !is_antiprompt) || _params.interactive) {
// predict
if (!embd.empty()) {
@ -506,9 +504,7 @@ static BOOL file_is_empty(NSString *path) {
const int skipped_tokens = (int) embd.size() - max_embd_size;
embd.resize(max_embd_size);
// console::set_display(console::error);
os_log_error(os_log_inst, "<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
// console::set_display(console::reset);
}
if (_params.grpAttnN == 1) {
@ -626,8 +622,9 @@ static BOOL file_is_empty(NSString *path) {
// optionally save the session on first sample (for faster prompt loading next time)
if ([pathSession length] > 0 && need_to_save_session && !_params.promptCacheRO) {
need_to_save_session = false;
[self.ctx saveStateFile:pathSession tokens:session_tokens.data() nTokenCount:session_tokens.size()];
// llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
[self.ctx saveStateFile:pathSession
tokens:session_tokens.data()
nTokenCount:session_tokens.size()];
os_log_debug(os_log_inst, "saved session to %s\n", [pathSession cStringUsingEncoding:NSUTF8StringEncoding]);
}
@ -636,8 +633,6 @@ static BOOL file_is_empty(NSString *path) {
[_smpl accept:idToken acceptGrammar:true];
// os_log_debug(os_log_inst, "last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(idToken);
// echo this to console
@ -666,7 +661,6 @@ static BOOL file_is_empty(NSString *path) {
// display text
if (input_echo && display) {
// std::cout<< "DISPLAYING TEXT" << std::endl;
for (auto idToken : embd) {
NSString *token_str = [self.ctx tokenToPiece:idToken special:_params.special];
@ -685,24 +679,13 @@ static BOOL file_is_empty(NSString *path) {
output_tokens.push_back(idToken);
output_ss << [token_str cStringUsingEncoding:NSUTF8StringEncoding];
last_output_ss << [token_str cStringUsingEncoding:NSUTF8StringEncoding];
NSLog(@"Generated %s", last_output_ss.str().c_str());
[self willChangeValueForKey:@"lastOutput"];
[_mutableLastOutput appendString:token_str];
[self didChangeValueForKey:@"lastOutput"];
}
}
if (!last_output_ss.str().empty()) {
// queue->addOutputLine(last_output_ss.str());
}
}
// reset color to default if there is no pending user input
if (input_echo && (int) embd_inp.size() == n_consumed) {
if (!last_output_ss.str().empty()) {
// queue->addOutputLine(last_output_ss.str());
}
// console::set_display(console::reset);
display = true;
}
// if not currently processing queued inputs;
@ -769,7 +752,6 @@ static BOOL file_is_empty(NSString *path) {
content:assistant_ss.str()];
}
isInteracting = true;
// LOG("\n");
}
}
@ -797,39 +779,23 @@ static BOOL file_is_empty(NSString *path) {
os_log_info(os_log_inst, "%s", [_params.inputPrefix cStringUsingEncoding:NSUTF8StringEncoding]);
}
// color user input only
// console::set_display(console::user_input);
display = _params.displayPrompt;
std::string line;
// bool another_line = true;
static int read_one = 0;
// if (!read_one) {
// do {
// another_line = false;// console::readline(line, params.multiline_input);
// buffer += "What is the weather in New York?";//line;
// } while (another_line);
// read_one++;
// }
// else {
if (!last_output_ss.str().empty()) {
auto str = last_output_ss.str();
last_output_ss.str("");
[queue addOutputLine:[NSString stringWithCString:str.c_str() encoding:NSUTF8StringEncoding]];
[_queue addOutputLine:[NSString stringWithCString:str.c_str() encoding:NSUTF8StringEncoding]];
[self willChangeValueForKey:@"lastOutput"];
_mutableLastOutput = [[NSMutableString alloc] init];
[self didChangeValueForKey:@"lastOutput"];
}
buffer = [[queue inputLine] cStringUsingEncoding:NSUTF8StringEncoding];
// do {
// another_line = console::readline(line, params.multiline_input);
// buffer += line;
// } while (another_line);
// }
// done taking input, reset color
// console::set_display(console::reset);
buffer = [[_queue inputLine] cStringUsingEncoding:NSUTF8StringEncoding];
if ([_queue isClosed]) {
break;
}
display = true;
// Add tokens to embd only if the input buffer is non-empty
@ -881,7 +847,6 @@ static BOOL file_is_empty(NSString *path) {
output_tokens.push_back(token);
output_ss << [[self.ctx tokenToPiece:token] cStringUsingEncoding:NSUTF8StringEncoding];
}
// reset assistant message
assistant_ss.str("");
@ -915,6 +880,30 @@ static BOOL file_is_empty(NSString *path) {
isInteracting = true;
}
}
NSLog(@"Loop over");
[_queue addOutputLine:@""];
}
- (void)stop {
isInteracting = false;
_params.interactive = false;
_queue.isClosed = YES;
[_queue addInputLine:@""];
[_queue outputLine];
}
- (void)dealloc
{
[_queue dealloc];
[self.smpl dealloc];
[self.ctx dealloc];
[self.model dealloc];
llama_backend_free();
[threadpool dealloc];
[threadpool_batch dealloc];
[super dealloc];
}
@end

1
objc/include/LlamaModel.h
View file

@ -19,7 +19,6 @@ typedef int32_t LlamaToken;
@interface LlamaModel : NSObject
- (LlamaContext *)context:(LlamaContextParams *)params;
- (LlamaToken)tokenBOS;
- (LlamaToken)tokenEOT;
- (LlamaToken)tokenEOS;

6
objc/include/LlamaSession.h
View file

@ -12,6 +12,8 @@
- (void)addOutputLine:(NSString *)line;
- (NSString *)outputLine;
@property (nonatomic, assign) BOOL isClosed;
@end
NS_REFINED_FOR_SWIFT @interface LlamaSession : NSObject
@ -19,9 +21,11 @@ NS_REFINED_FOR_SWIFT @interface LlamaSession : NSObject
@property (nonatomic, strong) LlamaModel *model;
@property (nonatomic, strong) LlamaContext *ctx;
@property (nonatomic, strong, readonly) NSString *lastOutput;
@property (nonatomic, strong) BlockingLineQueue *queue;
- (instancetype)initWithParams:(GPTParams *)params;
- (void)start:(BlockingLineQueue *)queue;
- (void)start;
- (void)stop;
@end

3
pocs/vdot/q8dot.cpp
View file

@ -11,6 +11,7 @@
#include <type_traits>
#include <ggml.h>
#include <ggml-cpu.h>
constexpr int kVecSize = 1 << 16;
@ -136,7 +137,7 @@ int main(int argc, char** argv) {
auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
const auto * funcs = ggml_get_type_traits(ggml_type);
const auto * funcs = ggml_get_type_traits_cpu(ggml_type);
Stat simple, ggml;

10
pocs/vdot/vdot.cpp
View file

@ -9,6 +9,7 @@
#include <array>
#include <ggml.h>
#include <ggml-cpu.h>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
@ -236,7 +237,8 @@ int main(int argc, char** argv) {
int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
const auto * funcs = useQ4_1 ? ggml_get_type_traits(GGML_TYPE_Q4_1) : ggml_get_type_traits(GGML_TYPE_Q4_0);
const auto * funcs = ggml_get_type_traits(useQ4_1 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q4_0);
const auto * funcs_cpu = ggml_get_type_traits_cpu(useQ4_1 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q4_0);
std::vector<block_q4_0> q40;
std::vector<block_q4_1> q41;
@ -282,10 +284,10 @@ int main(int argc, char** argv) {
dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
}
else {
const auto * vdot = ggml_get_type_traits(funcs->vec_dot_type);
const auto * vdot = ggml_get_type_traits(funcs_cpu->vec_dot_type);
vdot->from_float(y1.data(), q8.data(), kVecSize);
if (useQ4_1) funcs->vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
else funcs->vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
if (useQ4_1) funcs_cpu->vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
else funcs_cpu->vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
}
sumq += result;
t2 = std::chrono::high_resolution_clock::now();

2
scripts/sync-ggml.last
View file

@ -1 +1 @@
162e232411ee98ceb0cccfa84886118d917d2123
a099cb514d6687e436a5a423d1fb0448be0feb20

1
spm-headers/ggml-cpp.h Symbolic link
View file

@ -0,0 +1 @@
../ggml/include/ggml-cpp.h

1
spm-headers/ggml-cpu.h Symbolic link
View file

@ -0,0 +1 @@
../ggml/include/ggml-cpu.h

670
src/llama.cpp
View file

File diff suppressed because it is too large Load diff

306
swift/LlamaKit/LlamaKit.swift
View file

@ -2,100 +2,45 @@ import Foundation
@_exported import JSONSchema
@_exported import LlamaObjC
public protocol DynamicCallable: Sendable {
@discardableResult
func dynamicallyCall(withKeywordArguments args: [String: Any]) async throws -> String
}
public enum AnyDecodable: Decodable {
case string(String)
case int(Int)
case double(Double)
case bool(Bool)
case null
// Add other cases as needed
// Initializers for each type
init(_ value: String) {
self = .string(value)
}
init(_ value: Int) {
self = .int(value)
}
init(_ value: Double) {
self = .double(value)
}
init(_ value: Bool) {
self = .bool(value)
}
init() {
self = .null
}
// Decodable conformance
public init(from decoder: Decoder) throws {
let container = try decoder.singleValueContainer()
if container.decodeNil() {
self = .null
} else if let intValue = try? container.decode(Int.self) {
self = .int(intValue)
} else if let doubleValue = try? container.decode(Double.self) {
self = .double(doubleValue)
} else if let boolValue = try? container.decode(Bool.self) {
self = .bool(boolValue)
} else if let stringValue = try? container.decode(String.self) {
self = .string(stringValue)
} else {
let context = DecodingError.Context(
codingPath: decoder.codingPath,
debugDescription: "Cannot decode AnyDecodable"
)
throw DecodingError.typeMismatch(AnyDecodable.self, context)
}
}
}
struct ToolCall: Decodable {
let id: Int
let name: String
let arguments: [String: AnyDecodable]
}
struct ToolResponse<T: Encodable>: Encodable {
let id: Int
let result: T
}
// MARK: LlamaChatSession
/// Standard chat session for a given LLM.
/// An actor that manages a standard chat session with a given Large Language Model (LLM).
///
/// `LlamaChatSession` provides methods to interact with the LLM, allowing for synchronous
/// and asynchronous message inference. It uses a `BlockingLineQueue` to manage the input
/// and output lines for communication with the LLM backend.
///
/// ### Key Responsibilities:
/// - Initialize the session with the specified parameters.
/// - Send messages to the LLM and receive responses.
/// - Provide an asynchronous stream for real-time inference results.
public actor LlamaChatSession {
private let queue = BlockingLineQueue()
private let session: __LlamaSession
/// The underlying session object interfacing with the LLM backend.
internal let session: __LlamaSession
/// Initialize the session
/// - parameter params: common parameters to initialize the session
/// - parameter flush: whether or not to flush the initial prompt, reading initial output
public init(params: GPTParams, flush: Bool = true) async throws {
self.session = __LlamaSession(params: params)
Task.detached { [session, queue] in
session.start(queue)
Task.detached { [session] in
session.start()
}
// flush
guard flush else { return }
_ = queue.outputLine()
_ = session.queue.outputLine()
}
/// Create a new inference stream for a given message
/// - parameter message: The message to receive an inference for.
/// - returns: A stream of output from the LLM.
/// Creates an asynchronous stream for inference results based on a given message.
///
/// This method sends a message to the LLM and returns an `AsyncStream` that yields
/// output as it becomes available, allowing for real-time streaming of inference results.
///
/// - Parameter message: The message to send to the LLM for inference.
/// - Returns: An `AsyncStream` of `String` values representing incremental output from the LLM.
public func inferenceStream(message: String) async -> AsyncStream<String> {
queue.addInputLine(message)
session.queue.addInputLine(message)
var observationToken: NSKeyValueObservation?
return AsyncStream { stream in
observationToken = self.session.observe(\.lastOutput, options: [.new, .old]) { session, change in
@ -119,17 +64,49 @@ public actor LlamaChatSession {
}
}
}
/// Sends a message to the LLM and returns the response.
///
/// This method sends a message to the LLM and waits for the complete response.
///
/// - Parameter message: The message to send to the LLM for inference.
/// - Returns: A `String` containing the LLM's response to the message.
public func infer(message: String) async -> String {
queue.addInputLine(message)
return queue.outputLine()
session.queue.addInputLine(message)
return session.queue.outputLine()
}
deinit {
session.stop()
}
}
// MARK: LlamaGrammarSession
/// An actor that manages a chat session with the LLM, enforcing a grammar defined by a JSON schema.
///
/// `LlamaSession` allows you to interact with the LLM while constraining the output to match a specified JSON schema.
/// It uses the `JSONSchema` and `JSONSchemaConvertible` protocols to define the expected output format,
/// and applies a grammar to the LLM's sampler parameters.
///
/// ### Key Responsibilities:
/// - Initialize the session with a specified JSON schema, converting it to a grammar for the LLM.
/// - Send messages to the LLM and decode the responses into a specified type `T`.
///
/// ### Type Parameters:
/// - `T`: A type that conforms to `JSONSchemaConvertible` and represents the expected output format.
public actor LlamaSession<T: JSONSchemaConvertible> {
/// The underlying chat session used to interact with the LLM.
private let session: LlamaChatSession
/// Initializes a new grammar-constrained session with the LLM.
///
/// - Parameters:
/// - params: The parameters used to initialize the session, such as model settings.
/// - flush: A boolean indicating whether to flush the initial prompt and read initial output.
/// Defaults to `true`.
/// - Throws: An error if the session fails to initialize.
public init(params: GPTParams, flush: Bool = true) async throws {
let converter = SchemaConverter(propOrder: [])
_ = converter.visit(schema: T.jsonSchema, name: nil)
@ -137,163 +114,16 @@ public actor LlamaSession<T: JSONSchemaConvertible> {
session = try await LlamaChatSession(params: params, flush: flush)
}
public func chat(message: String) async throws -> T {
/// Sends a message to the LLM and decodes the response into type `T`.
///
/// This method sends a message to the LLM, receives the output, and attempts to decode it into the specified type `T`.
/// It enforces that the LLM's output conforms to the JSON schema associated with `T`.
///
/// - Parameter message: The message to send to the LLM.
/// - Returns: An instance of type `T` decoded from the LLM's response.
/// - Throws: An error if the decoding fails.
public func infer(message: String) async throws -> T {
let output = await session.infer(message: message).data(using: .utf8)!
return try JSONDecoder().decode(T.self, from: output)
}
}
// MARK: LlamaToolSession
public actor LlamaToolSession {
private let session: LlamaChatSession
private struct GetIpAddress: DynamicCallable {
func dynamicallyCall(withKeywordArguments args: [String : Any]) async throws -> String {
getIPAddress()
}
}
internal static func getIPAddress() -> String {
var address: String!
// Get list of all interfaces on the local machine:
var ifaddr: UnsafeMutablePointer<ifaddrs>? = nil
if getifaddrs(&ifaddr) == 0 {
// Loop through linked list of interfaces
var ptr = ifaddr
while ptr != nil {
let interface = ptr!.pointee
// Check if the interface is IPv4 or IPv6:
let addrFamily = interface.ifa_addr.pointee.sa_family
if addrFamily == UInt8(AF_INET) || addrFamily == UInt8(AF_INET6) {
// Convert interface name to String:
let name = String(cString: interface.ifa_name)
// Only consider non-loopback interfaces (e.g., "en0" for Wi-Fi)
if name == "en0" { // Typically en0 is the Wi-Fi interface
// Convert the address to a readable format:
var hostname = [CChar](repeating: 0, count: Int(NI_MAXHOST))
if getnameinfo(interface.ifa_addr, socklen_t(interface.ifa_addr.pointee.sa_len),
&hostname, socklen_t(hostname.count),
nil, socklen_t(0), NI_NUMERICHOST) == 0 {
address = String(cString: hostname)
}
}
}
ptr = interface.ifa_next
}
freeifaddrs(ifaddr)
}
return address
}
public private(set) var tools: [String: (DynamicCallable, _JSONFunctionSchema)]
public init(params: GPTParams,
tools: [String: (DynamicCallable, _JSONFunctionSchema)]) async throws {
self.tools = tools
let ipFnSchema = _JSONFunctionSchema(name: "getIpAddress", description: "Get the IP Address for this system", parameters: _JSONFunctionSchema.Parameters(properties: [:], required: []))
self.tools["getIpAddress"] = (GetIpAddress(), ipFnSchema)
let encoded = try JSONEncoder().encode(self.tools.values.map(\.1))
let prompt = """
\(params.prompt ?? "")
You are a function calling AI model. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
<tool_call>
{"name": <function-name>,"arguments": <args-dict>}
</tool_call>
Feel free to chain tool calls, e.g., if you need the user's location to find points of interest near them, fetch the user's location first.
The first call you will be asked to warm up is to get the user's IP address. Here are the available tools:
<tools> \(String(data: encoded, encoding: .utf8)!) </tools><|eot_id|>
"""
params.prompt = prompt
params.interactive = true
params.antiPrompts.append("<|eot_id|>");
params.inputPrefix = "<|start_header_id|>user<|end_header_id|>";
params.inputSuffix = "<|eot_id|><|start_header_id|>assistant<|end_header_id|>";
session = try await LlamaChatSession(params: params, flush: false)
let fn = await session.infer(message: "What is my IP address?")
let toolCall = try JSONDecoder().decode(ToolCall.self, from: fn.data(using: .utf8)!)
guard let tool = self.tools[toolCall.name] else {
fatalError()
}
let resp = try await tool.0.dynamicallyCall(withKeywordArguments: toolCall.arguments)
print(resp)
let output = await session.infer(message: """
<tool_response>
{"id": \(toolCall.id), result: \(resp)}
</tool_response>
""")
print(output)
}
private func callTool(_ call: String) async -> String? {
var nxt: String?
do {
let toolCall = try JSONDecoder().decode(ToolCall.self, from: call.data(using: .utf8)!)
guard let tool = tools[toolCall.name] else {
fatalError()
}
// TODO: tool call decode is allowed to fail but the code below is not
let callable = tool.0
do {
let response = try await callable.dynamicallyCall(withKeywordArguments: toolCall.arguments)
print("tool response: \(response)")
nxt = await session.infer(message: """
<tool_response>
{"id": \(toolCall.id), result: \(response)}
</tool_response>
""")
// TODO: If this decodes correctly, we should tail this into this method
// TODO: so that we do not decode twice
if let _ = try? JSONDecoder().decode(ToolCall.self, from: nxt!.data(using: .utf8)!) {
return await callTool(nxt!)
}
} catch {
nxt = await session.infer(message: """
<tool_response>
{"id": \(toolCall.id), result: "The tool call has unfortunately failed."}
</tool_response>
""")
}
print(nxt ?? "nil")
} catch {}
return nxt
}
public func infer(message: String) async throws -> String {
let output = await session.infer(message: message)
guard let output = await callTool(output) else {
return output
}
return output
}
}
public protocol LlamaActor: Actor {
static func tools(_ self: Self) -> [String: (DynamicCallable, _JSONFunctionSchema)]
var session: LlamaToolSession! { get }
}
public extension LlamaActor {
func chat(_ message: String) async throws -> String {
try await session.infer(message: message)
}
}
@attached(member, names: arbitrary)
@attached(extension, conformances: LlamaActor, names: arbitrary)
public macro llamaActor() = #externalMacro(module: "LlamaKitMacros",
type: "LlamaActorMacro")
@attached(body)
public macro Tool() = #externalMacro(module: "LlamaKitMacros",
type: "ToolMacro")

402
swift/LlamaKit/LlamaToolSession.swift Normal file
View file

@ -0,0 +1,402 @@
import Foundation
public protocol DynamicCallable: Sendable {
@discardableResult
func dynamicallyCall(withKeywordArguments args: [String: Any]) async throws -> String
}
public enum AnyDecodable: Decodable {
case string(String)
case int(Int)
case double(Double)
case bool(Bool)
case null
// Add other cases as needed
// Initializers for each type
init(_ value: String) {
self = .string(value)
}
init(_ value: Int) {
self = .int(value)
}
init(_ value: Double) {
self = .double(value)
}
init(_ value: Bool) {
self = .bool(value)
}
init() {
self = .null
}
// Decodable conformance
public init(from decoder: Decoder) throws {
let container = try decoder.singleValueContainer()
if container.decodeNil() {
self = .null
} else if let intValue = try? container.decode(Int.self) {
self = .int(intValue)
} else if let doubleValue = try? container.decode(Double.self) {
self = .double(doubleValue)
} else if let boolValue = try? container.decode(Bool.self) {
self = .bool(boolValue)
} else if let stringValue = try? container.decode(String.self) {
self = .string(stringValue)
} else {
let context = DecodingError.Context(
codingPath: decoder.codingPath,
debugDescription: "Cannot decode AnyDecodable"
)
throw DecodingError.typeMismatch(AnyDecodable.self, context)
}
}
}
struct ToolCall: Decodable {
let id: Int
let name: String
let arguments: [String: AnyDecodable]
}
struct ToolResponse<T: Encodable>: Encodable {
let id: Int
let result: T
}
// MARK: LlamaToolSession
/// An actor that manages tool calls within a chat session using Llama language models.
///
/// `LlamaToolSession` is responsible for handling interactions between the language model and dynamically callable tools.
/// It maintains a dictionary of tools that can be invoked based on the language model's output.
/// With the integration of the `@llamaActor` and `@Tool` macros, users can easily define custom tools within their own actors,
/// which are then registered and managed by `LlamaToolSession`.
///
/// ### Key Responsibilities:
/// - Initializes the language model session with the provided tools and prompt.
/// - Parses the language model's output to detect tool calls and invokes the corresponding tools.
/// - Sends tool responses back to the language model to generate coherent replies.
/// - Supports both synchronous and asynchronous inference, including streaming results.
///
/// ### Integration with Macros:
/// - The `@llamaActor` macro simplifies the creation of actors with tools by automatically generating the necessary code to integrate with `LlamaToolSession`.
/// - The `@Tool` macro allows functions within the actor to be marked as tools, which are registered and made available to the language model.
///
/// ### Usage:
/// ```swift
/// @llamaActor actor MyLlama {
/// /// Gets the user's favorite season.
/// @Tool public func getFavoriteSeason() async throws -> String {
/// return "autumn"
/// }
///
/// /// Gets the user's favorite animal.
/// @Tool public func getFavoriteAnimal() async throws -> String {
/// return "cat"
/// }
/// }
///
/// let params = GPTParams(prompt: "Your initial prompt")
/// let myLlama = try await MyLlama(params: params)
/// let response = try await myLlama.session.infer(message: "What is my favorite season?")
/// print(response)
/// ```
///
/// ### Architecture Diagram:
/// ```
/// +-------------------+
/// | MyLlama Actor |
/// | (User-Defined) |
/// | |
/// | @llamaActor |
/// | +---------------+ |
/// | | getFavorite...|<------------------------------+
/// | | (Tool funcs) | |
/// +---------+-------+ |
/// | |
/// | Uses |
/// v |
/// +-------------------+ +-----------------+ +------------------+
/// | LlamaToolSession | | LlamaChatSession| | __LlamaSession |
/// | (Actor) | | (Actor) | | (Objective-C) |
/// +---------+---------+ +---------+-------+ +---------+--------+
/// | ^ ^
/// | | |
/// | | Observes |
/// v | |
/// +-------------------+ +-----------------+ +------------------+
/// | BlockingLineQueue |---------->| LlamaSession |<------| Tools (Macros) |
/// | (Objective-C) | Input/ | (Objective-C) | | (DynamicCallable)|
/// +-------------------+ Output +----------------+ +------------------+
/// ```
///
/// ### Notes:
/// - **Beta Feature**: This feature is currently in beta and may undergo significant changes.
/// - **Template Format**: `LlamaToolSession` currently only works with the llama chat template format.
/// - **Thread Safety**: `LlamaToolSession` is an actor to ensure thread-safe operations in a concurrent environment.
/// - **Extensibility**: Users can define custom tools using the `@Tool` macro within an actor annotated with `@llamaActor`.
/// - **Error Handling**: The session handles tool call failures gracefully, allowing the language model to continue processing.
///
/// ### See Also:
/// - `LlamaChatSession`
/// - `@llamaActor` Macro
/// - `@Tool` Macro
public actor LlamaToolSession {
private let session: LlamaChatSession
public private(set) var tools: [String: (DynamicCallable, _JSONFunctionSchema)]
public init(params: GPTParams,
tools: [String: (DynamicCallable, _JSONFunctionSchema)]) async throws {
self.tools = tools
let encoded = try JSONEncoder().encode(self.tools.values.map(\.1))
let prompt = """
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
\(params.prompt ?? "")
You are a function calling AI model. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
<tool_call>
{"name": <function-name>,"arguments": <args-dict>}
</tool_call>
Feel free to chain tool calls, e.g., if you need the user's location to find points of interest near them, fetch the user's location first.
<tools> \(String(data: encoded, encoding: .utf8)!) </tools>
<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"""
params.prompt = prompt
params.interactive = true
params.antiPrompts.append("<|eot_id|>")
params.inputPrefix = "<|start_header_id|>user<|end_header_id|>"
params.inputSuffix = "<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
session = try await LlamaChatSession(params: params, flush: false)
guard let line = await session.session.queue.outputLine() else {
return
}
_ = await callTool(line)
}
private func callTool(_ call: ToolCall) async throws -> String {
guard let tool = tools[call.name] else {
fatalError()
}
// TODO: tool call decode is allowed to fail but the code below is not
let callable = tool.0
return try await callable.dynamicallyCall(withKeywordArguments: call.arguments)
}
private func callTool(_ call: String) async -> String? {
var nxt: String?
do {
let toolCall = try JSONDecoder().decode(ToolCall.self, from: call.data(using: .utf8)!)
guard let tool = tools[toolCall.name] else {
fatalError()
}
// TODO: tool call decode is allowed to fail but the code below is not
let callable = tool.0
do {
let response = try await callable.dynamicallyCall(withKeywordArguments: toolCall.arguments)
print("tool response: \(response)")
nxt = await session.infer(message: """
<tool_response>
{"id": \(toolCall.id), result: \(response)}
</tool_response>
""")
// TODO: If this decodes correctly, we should tail this into this method
// TODO: so that we do not decode twice
if let _ = try? JSONDecoder().decode(ToolCall.self, from: nxt!.data(using: .utf8)!) {
return await callTool(nxt!)
}
} catch {
nxt = await session.infer(message: """
<tool_response>
{"id": \(toolCall.id), result: "The tool call has unfortunately failed."}
</tool_response>
""")
}
print(nxt ?? "nil")
} catch {}
return nxt
}
public func infer(message: String) async throws -> String {
let output = await session.infer(message: message)
guard let output = await callTool(output) else {
return output
}
return output
}
public func inferenceStream(message: String) async -> AsyncStream<String> {
let underlyingSession = await session.session
await session.session.queue.addInputLine(message)
var observationToken: NSKeyValueObservation?
actor StreamProcessor {
private(set) var buffer = ""
private(set) var totalBuffer = ""
private(set) var isProcessingToolCall = false
func setBuffer(_ value: String) {
buffer = value
}
func appendBuffer(_ value: String) {
buffer.append(value)
totalBuffer.append(value)
}
func setTotalBuffer(_ value: String) {
totalBuffer = value
}
func setIsProcessingToolCall(_ value: Bool) {
isProcessingToolCall = value
}
}
let streamProcessor = StreamProcessor()
return AsyncStream { continuation in
observationToken = underlyingSession.observe(\.lastOutput, options: [.new, .old]) { session, change in
Task {
guard let newValue = change.newValue,
let oldValue = change.oldValue else {
continuation.finish()
return
}
var delta = ""
for change in newValue!.difference(from: oldValue!) {
switch change {
case .remove(_, _, _):
if await streamProcessor.isProcessingToolCall {
let call = try await JSONDecoder().decode(ToolCall.self, from: streamProcessor.totalBuffer.data(using: .utf8)!)
await streamProcessor.setBuffer("")
await streamProcessor.setTotalBuffer("")
let response = try await self.callTool(call)
return await self.session.session.queue.addInputLine("""
<tool_response>
{"id": \(call.id), result: \(response)}
</tool_response>
""")
} else {
continuation.finish()
}
return
case .insert(_, let element, _):
delta.append(element)
}
}
await streamProcessor.appendBuffer(delta)
// Check if buffer contains a complete tool call
if await streamProcessor.totalBuffer.starts(with: "\n\n\n{") {
await streamProcessor.setIsProcessingToolCall(true)
} else {
await streamProcessor.setIsProcessingToolCall(false)
// If not a tool call, yield the delta
continuation.yield(delta)
await streamProcessor.setBuffer("")
}
// Else, continue buffering until we have a complete tool call
}
}
continuation.onTermination = { [observationToken] _ in
observationToken?.invalidate()
}
}
}
}
public protocol LlamaActor: Actor {
static func tools(_ self: Self) -> [String: (DynamicCallable, _JSONFunctionSchema)]
var session: LlamaToolSession! { get }
}
public extension LlamaActor {
func infer(_ message: String) async throws -> String {
try await session.infer(message: message)
}
func inferenceStream(message: String) async -> AsyncStream<String> {
await session.inferenceStream(message: message)
}
}
// MARK: @llamaActor Macro
/// A macro that transforms an actor into a Llama tool actor, automatically integrating with `LlamaToolSession`.
///
/// The `@llamaActor` macro simplifies the process of creating an actor with tools that can be used by the Llama language model.
/// It automatically generates the necessary code to initialize a `LlamaToolSession`, register the tools, and conform to the `LlamaActor` protocol.
///
/// ### Usage:
/// ```swift
/// @llamaActor actor MyLlama {
/// /// Gets the user's favorite season.
/// @Tool public func getFavoriteSeason() async throws -> String {
/// return "autumn"
/// }
///
/// /// Gets the user's favorite animal.
/// @Tool public func getFavoriteAnimal() async throws -> String {
/// return "cat"
/// }
/// }
/// ```
///
/// ### Macro Details:
/// - **Attached To**: Actor declarations.
/// - **Produces**:
/// - Member variables and initializers required for tool integration.
/// - An extension conforming the actor to `LlamaActor`, providing necessary functionalities.
/// - **Parameters**: None.
///
/// ### Notes:
/// - The macro processes functions marked with `@Tool` within the actor to generate dynamic callable tools.
/// - It collects the tools and their schemas to register them with `LlamaToolSession`.
///
/// ### See Also:
/// - `@Tool` Macro
/// - `LlamaToolSession`
@attached(member, names: arbitrary)
@attached(extension, conformances: LlamaActor, names: arbitrary)
public macro llamaActor() = #externalMacro(module: "LlamaKitMacros",
type: "LlamaActorMacro")
/// A macro that marks a function within an actor as a tool callable by the Llama language model.
///
/// The `@Tool` macro indicates that a function should be exposed as a tool to the language model.
/// It processes the function to generate a dynamically callable structure, registers it with `LlamaToolSession`,
/// and includes the tool's metadata in the model's prompt.
///
/// ### Usage:
/// ```swift
/// @llamaActor actor MyLlama {
/// /// Gets the user's favorite animal.
/// @Tool public func getFavoriteAnimal() async throws -> String {
/// return "cat"
/// }
/// }
/// ```
///
/// ### Macro Details:
/// - **Attached To**: Function declarations within an actor marked with `@llamaActor`.
/// - **Produces**:
/// - A dynamically callable structure that wraps the function.
/// - Registers the tool with its name, description, and parameters.
/// - **Parameters**: None.
///
/// ### Notes:
/// - The function's documentation comment is used as the tool's description.
/// - Parameter comments are used to describe the tool's parameters.
/// - Supports functions with parameters and return values that conform to `Codable`.
///
/// ### See Also:
/// - `@llamaActor` Macro
/// - `LlamaToolSession`
@attached(body)
public macro Tool() = #externalMacro(module: "LlamaKitMacros",
type: "ToolMacro")

2
swift/main/main.swift
View file

@ -57,7 +57,7 @@ while true {
break
} else {
print("🧔🏽‍♂️: \(userInput)")
let response = try await llama.chat(userInput)
let response = try await llama.infer(userInput)
print("🤖: \(response)")
}
} else {

172
swift/test/LlamaKitTests.swift
View file

@ -4,6 +4,18 @@ import Testing
import JSONSchema
import OSLog
@llamaActor actor MyLlama {
/// Get the user's favorite season
@Tool public func getFavoriteSeason() async throws -> String {
return "autumn"
}
/// Get the user's favorite animal.
@Tool public func getFavoriteAnimal() async throws -> String {
return "cat"
}
}
// MARK: LlamaGrammarSession Suite
@Suite("LlamaSession Suite")
struct LlamaSessionSuite {
@ -43,7 +55,7 @@ struct LlamaSessionSuite {
func baseParams(url: String, to path: String) async throws -> GPTParams {
let params = GPTParams()
params.modelPath = try await downloadFile(url: url, to: path)
params.nPredict = 512
params.nPredict = 4096
params.nCtx = 4096
params.cpuParams.nThreads = 8
params.cpuParamsBatch.nThreads = 8
@ -89,7 +101,7 @@ struct LlamaSessionSuite {
"""
let session = try await LlamaSession<Trip>(params: params)
await #expect(throws: Never.self) {
let trip = try await session.chat(message: "Please create a trip for me to New York City that starts two weeks from now. The duration of the trip MUST be 3 days long.")
let trip = try await session.infer(message: "Please create a trip for me to New York City that starts two weeks from now. The duration of the trip MUST be 3 days long.")
#expect(trip.location.contains("New York"))
// TODO: Testing the other fields is difficult considering model size
// TODO: so for now, we are just asserting the grammar works
@ -100,83 +112,93 @@ struct LlamaSessionSuite {
let isSpellingCorrect: Bool
}
// MARK: Grammar Test
@Test func llamaSimpleGrammarSession() async throws {
let params = try await baseParams(url: "https://huggingface.co/RichardErkhov/openfoodfacts_-_spellcheck-mistral-7b-gguf/resolve/main/spellcheck-mistral-7b.Q8_0.gguf?download=true",
to: "spellcheck_q8.gguf")
let params = try await baseParams(url: "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/resolve/main/mistral-7b-instruct-v0.2.Q5_K_S.gguf?download=true",
to: "spellcheck_mistral.gguf")
params.prompt = """
###You are a spell checker. I will provide you with the word 'strawberry'. If the spelling of the given word is correct, please respond {"isCorrect": true} else respond {"isCorrect": false}.\n
You are a spell checker. I will provide you with versions of the word 'strawberry'. Tell me if the spelling is correct or not.
"""
params.inputPrefix = "<s>[INST]"
params.inputSuffix = "[/INST]"
params.antiPrompts.append("</s>")
let session = try await LlamaSession<IsCorrect>(params: params)
for _ in 0..<10 {
var output = try await session.chat(message: "###strawberry\n")
#expect(output.isSpellingCorrect)
output = try await session.chat(message: "###strawberrry\n")
#expect(!output.isSpellingCorrect)
var output = try await session.infer(message: "strawberry\n")
#expect(output.isSpellingCorrect)
output = try await session.infer(message: "st44rawberrry\n")
#expect(!output.isSpellingCorrect)
}
// MARK: Tool Test
@Test func llamaToolSession() async throws {
let params = try await baseParams(url: "https://huggingface.co/bartowski/Llama-3-Groq-8B-Tool-Use-GGUF/resolve/main/Llama-3-Groq-8B-Tool-Use-Q8_0.gguf?download=true", to: "llama_tools.gguf")
let fm = FileManager.default
let tmpDir = fm.temporaryDirectory
let cacheURL = tmpDir.appending(path: "cache")
params.pathPromptCache = cacheURL.path()
params.logging = true
defer { try? fm.removeItem(at: cacheURL) }
let llama = try await MyLlama(params: params)
var output = try await llama.infer("What's my favorite animal?")
#expect(output.contains("cat"))
output = try await llama.infer("What's my favorite season?")
#expect(output.contains("autumn"))
}
// MARK: Session dealloc Test
@Test func llamaToolSessionDealloc() async throws {
let params = try await baseParams(url: "https://huggingface.co/bartowski/Llama-3-Groq-8B-Tool-Use-GGUF/resolve/main/Llama-3-Groq-8B-Tool-Use-Q8_0.gguf?download=true", to: "llama_tools.gguf")
func reportMemoryUsage() -> UInt64? {
var info = mach_task_basic_info()
var count = mach_msg_type_number_t(MemoryLayout.size(ofValue: info)) / 4
let kerr = withUnsafeMutablePointer(to: &info) {
$0.withMemoryRebound(to: integer_t.self, capacity: 1) {
task_info(mach_task_self_, task_flavor_t(MACH_TASK_BASIC_INFO), $0, &count)
}
}
guard kerr == KERN_SUCCESS else {
print("Error with task_info(): \(kerr)")
return nil
}
return info.resident_size // Memory in bytes
}
var memPostAlloc: UInt64!
try await Task {
let llama = try await MyLlama(params: params)
memPostAlloc = reportMemoryUsage()! / 1024 / 1024
#expect(memPostAlloc > 500) // we are estimating here
var output = try await llama.infer("What's my favorite animal?")
print(output)
output = try await llama.infer("What question did i just ask you?")
print(output)
}.value
sleep(1)
var memDealloc = reportMemoryUsage()! / 1024 / 1024
#expect(memDealloc < 200)
try await Task {
let llama = try await MyLlama(params: params)
memPostAlloc = reportMemoryUsage()! / 1024 / 1024
#expect(memPostAlloc > 500)
_ = try await llama.infer("What was the first question I asked you?")
}.value
sleep(1)
memDealloc = reportMemoryUsage()! / 1024 / 1024
#expect(memDealloc < 200)
}
// MARK: Stream Test
@Test func llamaToolSessionStream() async throws {
let params = try await baseParams(url: "https://huggingface.co/bartowski/Llama-3-Groq-8B-Tool-Use-GGUF/resolve/main/Llama-3-Groq-8B-Tool-Use-Q8_0.gguf?download=true", to: "llama_tools.gguf")
params.prompt = "You are an AI tool assistant that can chain tool calls together."
let llama = try await MyLlama(params: params)
var buffer = ""
for await output in await llama.inferenceStream(message: "What's my favorite animal and season?") {
buffer += output
print(output, terminator: "")
}
#expect(buffer.contains("cat") && buffer.contains("autumn"))
}
}
import WeatherKit
import CoreLocation
func downloadFile() async throws -> String {
let fm = FileManager.default
let tmpDir = fm.temporaryDirectory
let destinationURL = tmpDir.appending(path: "llama_groq_gguf.gguf")
guard !fm.fileExists(atPath: destinationURL.path()) else {
return destinationURL.path()
}
print("Downloading Llama Tools, this may take a while...")
// Define the URL
guard let url = URL(string: "https://huggingface.co/bartowski/Llama-3-Groq-8B-Tool-Use-GGUF/resolve/main/Llama-3-Groq-8B-Tool-Use-Q5_K_M.gguf?download=true") else {
print("Invalid URL.")
throw URLError(.badURL)
}
// Start the async download
let (tempURL, _) = try await URLSession.shared.download(from: url)
// Define the destination path in the documents directory
// Move the downloaded file to the destination
try fm.moveItem(at: tempURL, to: destinationURL)
print("File downloaded to: \(destinationURL.path())")
return destinationURL.path()
}
@llamaActor actor MyLlama {
struct CurrentWeather: Codable {
let temperature: Double
let condition: WeatherCondition
}
/// Get the current weather in a given location.
/// - parameter location: The city and state, e.g. San Francisco, CA
/// - parameter unit: The unit of temperature
@Tool public func getCurrentWeather(location: String, unit: String) async throws -> CurrentWeather {
let weather = try await WeatherService().weather(for: CLGeocoder().geocodeAddressString(location)[0].location!)
var temperature = weather.currentWeather.temperature
temperature.convert(to: .fahrenheit)
return CurrentWeather(temperature: temperature.value,
condition: weather.currentWeather.condition)
}
}
@Test func llamaToolSession() async throws {
let params = GPTParams()
params.modelPath = try await downloadFile()
params.nPredict = 512
params.nCtx = 4096
params.cpuParams.nThreads = 8
params.cpuParamsBatch.nThreads = 8
params.nBatch = 1024
params.nGpuLayers = 1024
let llama = try await MyLlama(params: params)
let currentWeather = try await llama.getCurrentWeather(location: "San Francisco, CA", unit: "farenheit")
let output = try await llama.chat("What's the weather (in farenheit) in San Francisco, CA?")
#expect(output.contains(String(format: "%d", Int(currentWeather.temperature))))
// #expect(output.contains(currentWeather.condition.rawValue))
}

1
tests/test-backend-ops.cpp
View file

@ -16,6 +16,7 @@
#include <ggml.h>
#include <ggml-cpu.h>
#include <ggml-alloc.h>
#include <ggml-backend.h>

1
tests/test-barrier.cpp
View file

@ -1,4 +1,5 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-backend.h"
#include <chrono>

1
tests/test-grad0.cpp
View file

@ -1,5 +1,6 @@
#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
#include "ggml.h"
#include "ggml-cpu.h"
#include <cfloat>
#include <cmath>

10
tests/test-quantize-fns.cpp
View file

@ -1,6 +1,7 @@
// Unit tests for quantization specific functions - quantize, dequantize and dot product
#include "ggml.h"
#include "ggml-cpu.h"
#undef NDEBUG
#include <assert.h>
@ -78,18 +79,18 @@ static float dot_product(const float * a1, const float * a2, size_t test_size) {
// Total dot product error
static float dot_product_error(
const ggml_type_traits * qfns, size_t test_size, const float * test_data1, const float *test_data2
const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data1, const float *test_data2
) {
std::vector<uint8_t> tmp_q1(2*test_size);
std::vector<uint8_t> tmp_q2(2*test_size);
const auto * vdot = ggml_get_type_traits(qfns->vec_dot_type);
const auto * vdot = ggml_get_type_traits(qfns_cpu->vec_dot_type);
qfns->from_float(test_data1, tmp_q1.data(), test_size);
vdot->from_float(test_data2, tmp_q2.data(), test_size);
float result = INFINITY;
qfns->vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
qfns_cpu->vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
const float dot_ref = dot_product(test_data1, test_data2, test_size);
@ -132,6 +133,7 @@ int main(int argc, char * argv[]) {
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
ggml_type type = (ggml_type) i;
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
// deprecated - skip
if (qfns->blck_size == 0) {
@ -166,7 +168,7 @@ int main(int argc, char * argv[]) {
printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
}
const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
const float vec_dot_error = dot_product_error(qfns, qfns_cpu, test_size, test_data.data(), test_data2.data());
const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
? MAX_DOT_PRODUCT_ERROR_LOWBIT

6
tests/test-quantize-perf.cpp
View file

@ -1,6 +1,7 @@
// Benchmark quantization specific functions on synthetic data
#include "ggml.h"
#include "ggml-cpu.h"
#undef NDEBUG
#include <algorithm>
@ -271,6 +272,7 @@ int main(int argc, char * argv[]) {
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
ggml_type type = (ggml_type) i;
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (!params.include_types.empty() && ggml_type_name(type) && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
continue;
}
@ -328,7 +330,7 @@ int main(int argc, char * argv[]) {
for (size_t size : params.test_sizes) {
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
auto quantize_fn = [&](void) -> float {
const auto * vdot = ggml_get_type_traits(qfns->vec_dot_type);
const auto * vdot = ggml_get_type_traits(qfns_cpu->vec_dot_type);
vdot->from_float(test_data1, test_q1, size);
return test_q1[0];
};
@ -346,7 +348,7 @@ int main(int argc, char * argv[]) {
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
auto quantize_fn = [&](void) -> float {
float result;
qfns->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
qfns_cpu->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
return result;
};
size_t quantized_size = ggml_row_size(type, size);

1
tests/test-rope.cpp
View file

@ -1,4 +1,5 @@
#include "ggml.h"
#include "ggml-cpu.h"
#include <cmath>
#include <cstdio>