ggml : fix ggml_flash_attn to use op_params (#2387)

* ggml : fix ggml_flash_attn to use op_params

.github/workflows/build.yml

@@ -197,6 +197,8 @@ jobs:
     strategy:
       matrix:
         include:
+          - build: 'noavx'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
           - build: 'avx2'
             defines: '-DLLAMA_BUILD_SERVER=ON'
           - build: 'avx'

.gitignore

@@ -16,6 +16,8 @@ build/
 build-em/
 build-debug/
 build-release/
+build-ci-debug/
+build-ci-release/
 build-static/
 build-cublas/
 build-opencl/
@@ -25,9 +27,10 @@ build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 out/
+tmp/
 
 models/*
-*.bin
+models-mnt
 
 /main
 /quantize
@@ -58,3 +61,18 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
+
+
+pyproject.toml
+poetry.lock
+poetry.toml
+
+# Test binaries
+tests/test-double-float
+tests/test-grad0
+tests/test-opt
+tests/test-quantize-fns
+tests/test-quantize-perf
+tests/test-sampling
+tests/test-tokenizer-0
+

CMakeLists.txt

@@ -512,6 +512,7 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
     target_link_libraries(ggml_shared PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+    install(TARGETS ggml_shared LIBRARY)
 endif()
 
 add_library(llama
@@ -533,8 +534,32 @@ if (BUILD_SHARED_LIBS)
     if (LLAMA_METAL)
         set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
     endif()
+    install(TARGETS llama LIBRARY)
 endif()
 
+include(GNUInstallDirs)
+install(
+    FILES convert.py
+    PERMISSIONS
+        OWNER_READ
+        OWNER_WRITE
+        OWNER_EXECUTE
+        GROUP_READ
+        GROUP_EXECUTE
+        WORLD_READ
+        WORLD_EXECUTE
+    DESTINATION ${CMAKE_INSTALL_BINDIR})
+install(
+    FILES convert-lora-to-ggml.py
+    PERMISSIONS
+        OWNER_READ
+        OWNER_WRITE
+        OWNER_EXECUTE
+        GROUP_READ
+        GROUP_EXECUTE
+        WORLD_READ
+        WORLD_EXECUTE
+    DESTINATION ${CMAKE_INSTALL_BINDIR})
 
 #
 # programs, examples and tests

Makefile

@@ -1,5 +1,8 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server libembdinput.so embd-input-test
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server embd-input-test
+
+# Binaries only useful for tests
+TEST_TARGETS = tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
 
 default: $(BUILD_TARGETS)
 
@@ -90,6 +93,28 @@ ifeq ($(UNAME_S),Haiku)
 	CXXFLAGS += -pthread
 endif
 
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# library name prefix
+ifneq ($(_WIN32),1)
+	LIB_PRE := lib
+endif
+
+# Dynamic Shared Object extension
+ifneq ($(_WIN32),1)
+	DSO_EXT := .so
+else
+	DSO_EXT := .dll
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
@@ -102,7 +127,7 @@ endif
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
 	CFLAGS   += -march=native -mtune=native
 	CXXFLAGS += -march=native -mtune=native
@@ -168,8 +193,12 @@ ifdef LLAMA_CUBLAS
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
-	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler
+ifdef LLAMA_CUDA_NVCC
+	NVCC = $(LLAMA_CUDA_NVCC)
+else
+	NVCC = nvcc
+endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
 else
@@ -198,19 +227,23 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
-
+ifdef LLAMA_CUDA_CCBIN
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
-	CFLAGS   += -DGGML_USE_CLBLAST
-	CXXFLAGS += -DGGML_USE_CLBLAST
+
+	CFLAGS   += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+	CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
 		LDFLAGS += -lclblast -framework OpenCL
 	else
-		LDFLAGS += -lclblast -lOpenCL
+		LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
 	endif
 	OBJS    += ggml-opencl.o
 
@@ -290,17 +323,20 @@ llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+grammar-parser.o: examples/grammar-parser.cpp examples/grammar-parser.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
+	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
@@ -324,15 +360,15 @@ embedding: examples/embedding/embedding.cpp                   build-info.h ggml.
 save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
-libembdinput.so: examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
 
 
-embd-input-test: libembdinput.so examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.so,$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
+embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -349,6 +385,8 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 # Tests
 #
 
+tests: $(TEST_TARGETS)
+
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	./$@
@@ -356,6 +394,23 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-.PHONY: tests clean
-tests:
-	bash ./tests/run-tests.sh
+tests/test-double-float: tests/test-double-float.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-grad0: tests/test-grad0.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-opt: tests/test-opt.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-0: tests/test-tokenizer-0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)

README.md

@@ -242,6 +242,23 @@ In order to build llama.cpp you have three different options.
     zig build -Doptimize=ReleaseFast
     ```
 
+-   Using `gmake` (FreeBSD):
+
+    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
+    2. Add your user to **video** group
+    3. Install compilation dependencies.
+
+        ```bash
+        sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
+            opencl clblast openblas
+
+            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        ```
+
+    **Notes:** With this packages you can build llama.cpp with OPENBLAS and
+    CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read
+    the instructions for use and activate this options in this document below.
+
 ### Metal Build
 
 Using Metal allows the computation to be executed on the GPU for Apple devices:
@@ -384,7 +401,7 @@ Building the program with BLAS support may lead to some performance improvements
 
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 7.0/Turing/RTX 2000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
   | LLAMA_CUDA_MMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |

ci/README.md

@@ -0,0 +1,25 @@
+# CI
+
+In addition to [Github Actions](https://github.com/ggerganov/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
+
+https://github.com/ggml-org/ci
+
+It monitors the `master` branch for new commits and runs the
+[ci/run.sh](https://github.com/ggerganov/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
+to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
+to cover various hardware architectures, including GPU and Apple Silicon instances.
+
+Collaborators can optionally trigger the CI run by adding the `ggml-ci` keyword to their commit message.
+Only the branches of this repo are monitored for this keyword.
+
+It is a good practice, before publishing changes to execute the full CI locally on your machine:
+
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with CUDA support
+GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```

ci/run.sh

@@ -0,0 +1,409 @@
+#/bin/bash
+#
+# sample usage:
+#
+# mkdir tmp
+#
+# # CPU-only build
+# bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with CUDA support
+# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+
+if [ -z "$2" ]; then
+    echo "usage: $0 <output-dir> <mnt-dir>"
+    exit 1
+fi
+
+mkdir -p "$1"
+mkdir -p "$2"
+
+OUT=$(realpath "$1")
+MNT=$(realpath "$2")
+
+rm -v $OUT/*.log
+rm -v $OUT/*.exit
+rm -v $OUT/*.md
+
+sd=`dirname $0`
+cd $sd/../
+SRC=`pwd`
+
+## helpers
+
+# download a file if it does not exist or if it is outdated
+function gg_wget {
+    local out=$1
+    local url=$2
+
+    local cwd=`pwd`
+
+    mkdir -p $out
+    cd $out
+
+    # should not re-download if file is the same
+    wget -nv -N $url
+
+    cd $cwd
+}
+
+function gg_printf {
+    printf -- "$@" >> $OUT/README.md
+}
+
+function gg_run {
+    ci=$1
+
+    set -o pipefail
+    set -x
+
+    gg_run_$ci | tee $OUT/$ci.log
+    cur=$?
+    echo "$cur" > $OUT/$ci.exit
+
+    set +x
+    set +o pipefail
+
+    gg_sum_$ci
+
+    ret=$((ret | cur))
+}
+
+## ci
+
+# ctest_debug
+
+function gg_run_ctest_debug {
+    cd ${SRC}
+
+    rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    set +e
+}
+
+function gg_sum_ctest_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# ctest_release
+
+function gg_run_ctest_release {
+    cd ${SRC}
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    if [ -z ${GG_BUILD_LOW_PERF} ]; then
+        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    else
+        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    fi
+
+    set +e
+}
+
+function gg_sum_ctest_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+# open_llama_3b_v2
+
+function gg_run_open_llama_3b_v2 {
+    cd ${SRC}
+
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
+
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/open-llama/3B-v2"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.bin"
+    model_q8_0="${path_models}/ggml-model-q8_0.bin"
+    model_q4_0="${path_models}/ggml-model-q4_0.bin"
+    model_q4_1="${path_models}/ggml-model-q4_1.bin"
+    model_q5_0="${path_models}/ggml-model-q5_0.bin"
+    model_q5_1="${path_models}/ggml-model-q5_1.bin"
+    model_q2_k="${path_models}/ggml-model-q2_k.bin"
+    model_q3_k="${path_models}/ggml-model-q3_k.bin"
+    model_q4_k="${path_models}/ggml-model-q4_k.bin"
+    model_q5_k="${path_models}/ggml-model-q5_k.bin"
+    model_q6_k="${path_models}/ggml-model-q6_k.bin"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    set +e
+}
+
+function gg_sum_open_llama_3b_v2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'OpenLLaMA 3B-v2:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+}
+
+# open_llama_7b_v2
+# requires: GG_BUILD_CUDA
+
+function gg_run_open_llama_7b_v2 {
+    cd ${SRC}
+
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
+
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/open-llama/7B-v2"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.bin"
+    model_q8_0="${path_models}/ggml-model-q8_0.bin"
+    model_q4_0="${path_models}/ggml-model-q4_0.bin"
+    model_q4_1="${path_models}/ggml-model-q4_1.bin"
+    model_q5_0="${path_models}/ggml-model-q5_0.bin"
+    model_q5_1="${path_models}/ggml-model-q5_1.bin"
+    model_q2_k="${path_models}/ggml-model-q2_k.bin"
+    model_q3_k="${path_models}/ggml-model-q3_k.bin"
+    model_q4_k="${path_models}/ggml-model-q4_k.bin"
+    model_q5_k="${path_models}/ggml-model-q5_k.bin"
+    model_q6_k="${path_models}/ggml-model-q6_k.bin"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    set +e
+}
+
+function gg_sum_open_llama_7b_v2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'OpenLLaMA 7B-v2:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+}
+
+## main
+
+if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    rm -rf ${SRC}/models-mnt
+
+    mnt_models=${MNT}/models
+    mkdir -p ${mnt_models}
+    ln -sfn ${mnt_models} ${SRC}/models-mnt
+
+    python3 -m pip install -r ${SRC}/requirements.txt
+fi
+
+ret=0
+
+test $ret -eq 0 && gg_run ctest_debug
+test $ret -eq 0 && gg_run ctest_release
+
+if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    if [ -z ${GG_BUILD_CUDA} ]; then
+        test $ret -eq 0 && gg_run open_llama_3b_v2
+    else
+        test $ret -eq 0 && gg_run open_llama_7b_v2
+    fi
+fi
+
+exit $ret

convert-lora-to-ggml.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python
 import json
 import os
 import re

convert.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python
 import argparse
 import concurrent.futures
 import copy
@@ -141,9 +142,9 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
 @dataclass
 class Params:
     n_vocab: int
-    n_embd: int
-    n_mult: int
-    n_head: int
+    n_embd:  int
+    n_mult:  int
+    n_head:  int
     n_layer: int
 
     @staticmethod
@@ -166,11 +167,11 @@ class Params:
         n_head=n_embd // 128 # guessed
 
         return Params(
-            n_vocab=n_vocab,
-            n_embd=n_embd,
-            n_mult=256,
-            n_head=n_head,
-            n_layer=n_layer,
+            n_vocab = n_vocab,
+            n_embd  = n_embd,
+            n_mult  = 256,
+            n_head  = n_head,
+            n_layer = n_layer,
         )
 
     @staticmethod
@@ -178,28 +179,53 @@ class Params:
         config = json.load(open(config_path))
 
         n_vocab = config["vocab_size"];
-        n_embd = config["hidden_size"];
-        n_head = config["num_attention_heads"];
+        n_embd  = config["hidden_size"];
+        n_head  = config["num_attention_heads"];
         n_layer = config["num_hidden_layers"];
-        n_ff = config["intermediate_size"];
+        n_ff    = config["intermediate_size"];
 
         n_mult = find_n_mult(n_ff, n_embd);
 
         return Params(
-            n_vocab=n_vocab,
-            n_embd=n_embd,
-            n_mult=n_mult,
-            n_head=n_head,
-            n_layer=n_layer,
+            n_vocab = n_vocab,
+            n_embd  = n_embd,
+            n_mult  = n_mult,
+            n_head  = n_head,
+            n_layer = n_layer,
+        )
+
+    # LLaMA v2 70B params.json
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
+    @staticmethod
+    def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+        config = json.load(open(config_path))
+
+        n_vocab = config["vocab_size"];
+        n_embd  = config["dim"];
+        n_head  = config["n_heads"];
+        n_layer = config["n_layers"];
+        n_mult  = config["multiple_of"];
+
+        if n_vocab == -1:
+            n_vocab = model["tok_embeddings.weight"].shape[0]
+
+        return Params(
+            n_vocab = n_vocab,
+            n_embd  = n_embd,
+            n_mult  = n_mult,
+            n_head  = n_head,
+            n_layer = n_layer,
         )
 
     @staticmethod
     def load(model_plus: 'ModelPlus') -> 'Params':
+        hf_config_path   = model_plus.paths[0].parent / "config.json"
         orig_config_path = model_plus.paths[0].parent / "params.json"
-        hf_transformer_config_path = model_plus.paths[0].parent / "config.json"
 
-        if hf_transformer_config_path.exists():
-            params = Params.loadHFTransformerJson(model_plus.model, hf_transformer_config_path)
+        if hf_config_path.exists():
+            params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
+        elif orig_config_path.exists():
+            params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
         else:
             params = Params.guessed(model_plus.model)
 
@@ -208,14 +234,21 @@ class Params:
 
 
 class SentencePieceVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
-        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path], vocabtype: Optional[str]) -> None:
+        self.vocabtype = vocabtype
+        if self.vocabtype == "bpe":
+          self.sentencepiece_tokenizer = json.loads(open(str(fname_tokenizer)).read())
+        else:
+          self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
         added_tokens: Dict[str, int]
         if fname_added_tokens is not None:
             added_tokens = json.load(open(fname_added_tokens))
         else:
             added_tokens = {}
-        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+        if self.vocabtype == "bpe":
+          vocab_size: int = len(self.sentencepiece_tokenizer)
+        else:
+          vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
         expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids = sorted(added_tokens.values())
         if expected_ids != actual_ids:
@@ -229,22 +262,32 @@ class SentencePieceVocab:
 
     def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float]]:
         tokenizer = self.sentencepiece_tokenizer
-        for i in range(tokenizer.vocab_size()):
+        if self.vocabtype == "bpe":
+          from transformers.models.gpt2 import tokenization_gpt2
+          byte_encoder = tokenization_gpt2.bytes_to_unicode()
+          byte_decoder = {v: k for k, v in byte_encoder.items()}
+          for i, item in enumerate(tokenizer):
             text: bytes
-            if tokenizer.is_unknown(i):
-                text = " \u2047 ".encode("utf-8")
-            elif tokenizer.is_control(i):
-                text = b""
-            elif tokenizer.is_byte(i):
-                piece = tokenizer.id_to_piece(i)
-                if len(piece) != 6:
-                    raise Exception(f"Invalid token: {piece}")
-                byte_value = int(piece[3:-1], 16)
-                text = struct.pack("B", byte_value)
-            else:
-                text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
-            score: float = tokenizer.get_score(i)
+            text = b''.join([x.to_bytes(1, byteorder='big') for x in [byte_decoder[y] for y in item]])
+            score: float = -i
             yield text, score
+        else:
+          for i in range(tokenizer.vocab_size()):
+              text: bytes
+              if tokenizer.is_unknown(i):
+                  text = " \u2047 ".encode("utf-8")
+              elif tokenizer.is_control(i):
+                  text = b""
+              elif tokenizer.is_byte(i):
+                  piece = tokenizer.id_to_piece(i)
+                  if len(piece) != 6:
+                      raise Exception(f"Invalid token: {piece}")
+                  byte_value = int(piece[3:-1], 16)
+                  text = struct.pack("B", byte_value)
+              else:
+                  text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
+              score: float = tokenizer.get_score(i)
+              yield text, score
 
     def added_tokens(self) -> Iterable[Tuple[bytes, float]]:
         for text in self.added_tokens_list:
@@ -1035,8 +1078,7 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(fname_out: Path, vocab: Vocab) -> None:
         of = OutputFile(fname_out)
-        params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0,
-                        n_head=1, n_layer=0)
+        params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0, n_head=1, n_layer=0)
         of = OutputFile(fname_out)
         of.write_file_header(params, file_type=GGMLFileType.AllF32)
         of.write_vocab(vocab)
@@ -1171,14 +1213,18 @@ def filter_and_sort_tensors(model: LazyModel) -> LazyModel:
     return {name: model[name] for name in TENSORS_LIST if name in model}
 
 
-def load_vocab(path: Path) -> SentencePieceVocab:
+def load_vocab(path: Path, vocabtype: Optional[str]) -> SentencePieceVocab:
+    print(f"vocabtype: {vocabtype}")
     # Be extra-friendly and accept either a file or a directory.  Also, if it's
     # a directory, it might be the model directory, and tokenizer.model might
     # be in the parent of that.
     if path.is_dir():
-        path2 = path / "tokenizer.model"
+        vocab_file = "tokenizer.model"
+        if vocabtype == 'bpe':
+          vocab_file = "vocab.json"
+        path2 = path / vocab_file
         # Use `.parent` instead of /.. to handle the symlink case better.
-        path3 = path.parent / "tokenizer.model"
+        path3 = path.parent / vocab_file
         if path2.exists():
             path = path2
         elif path3.exists():
@@ -1189,7 +1235,8 @@ def load_vocab(path: Path) -> SentencePieceVocab:
                 "if it's in another directory, pass the directory as --vocab-dir")
     added_tokens_path = path.parent / "added_tokens.json"
     print(f"Loading vocab file {path}")
-    return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None,
+                              vocabtype)
 
 
 def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
@@ -1227,6 +1274,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
     parser.add_argument("model", type=Path,
                         help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--vocabtype", default='spm', choices=["spm", "bpe"], help="vocab format (default: spm)")
     args = parser.parse_args(args_in)
 
     vocab: Vocab
@@ -1234,7 +1282,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
     elif args.vocab_only:
-        vocab = load_vocab(args.vocab_dir or args.model)
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
         assert args.outfile, "need --outfile if using --vocab-only"
         outfile = args.outfile
         OutputFile.write_vocab_only(outfile, vocab)
@@ -1248,7 +1296,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
             vocab = model_plus.vocab
         else:
             vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-            vocab = load_vocab(vocab_dir)
+            vocab = load_vocab(vocab_dir, args.vocabtype)
         params = Params.load(model_plus)
         model = model_plus.model
         model = do_necessary_conversions(model, params)

examples/CMakeLists.txt

@@ -13,6 +13,8 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
+    grammar-parser.h
+    grammar-parser.cpp
     )
 
 if (BUILD_SHARED_LIBS)

examples/Miku.sh

@@ -2,21 +2,21 @@
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"
-MODEL="${MODEL:-./models/gpt4all-7B/gpt4all-lora-unfiltered-quantized.bin}"
+MODEL="${MODEL:-./models/llama-2-7b-chat.ggmlv3.q4_K_M.bin}"
 USER_NAME="${USER_NAME:-Anon}"
 
 # Uncomment and adjust to the number of CPU cores you want to use.
 #N_THREAD="${N_THREAD:-4}"
+CTX_SIZE="${CTX_SIZE:-4096}"
 N_PREDICTS="${N_PREDICTS:-4096}"
 
 GEN_OPTIONS=(--batch_size 1024
---ctx_size 2048
+--ctx_size "$CTX_SIZE"
 --keep -1
 --repeat_last_n 256
 --repeat_penalty 1.17647
---temp 0.7
---top_k 40
---top_p 0.5)
+--temp 0.6
+--mirostat 2)
 
 if [ -n "$N_THREAD" ]; then
     GEN_OPTIONS+=(--threads "$N_THREAD")
@@ -24,16 +24,17 @@ fi
 
 ./main "${GEN_OPTIONS[@]}" \
     --model "$MODEL" \
+    --in-prefix " " \
+    --in-suffix "${AI_NAME}:" \
     --n_predict "$N_PREDICTS" \
     --color --interactive \
     --reverse-prompt "${USER_NAME}:" \
-    --prompt "
-This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
+    --prompt "This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
 ${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
 ${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help.
 ${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad.
 ${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her.
-The conversation is only between ${USER_NAME} and ${AI_NAME}
+The conversation is only between ${USER_NAME} and ${AI_NAME}.
 The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
 ${AI_NAME} can only communicate through text, so she can't send images or videos.
 

examples/baby-llama/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(TARGET 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)

examples/baby-llama/baby-llama.cpp

@@ -8,6 +8,8 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static const float rms_norm_eps = 1e-6f;
+
 float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
@@ -562,7 +564,7 @@ struct ggml_tensor * forward(
         // norm
         {
             // cur shape [n_embd,N,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
             // cur = attention_norm*cur
             cur = ggml_mul(ctx0,
@@ -685,7 +687,7 @@ struct ggml_tensor * forward(
             // norm
             {
                 // cur shape [n_embd,N,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
 
                 // cur = ffn_norm*cur
                 // cur shape [n_embd,N,1,1]
@@ -729,7 +731,7 @@ struct ggml_tensor * forward(
     {
 
         // inpL shape [n_embd,N,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
         // inpL = norm*inpL
         // inpL shape [n_embd,N,1,1]
@@ -817,7 +819,7 @@ struct ggml_tensor * forward_batch(
         // norm
         {
             // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
             assert_shape_2d(cur, n_embd, N*n_batch);
 
             // cur = attention_norm*cur
@@ -981,7 +983,7 @@ struct ggml_tensor * forward_batch(
             // norm
             {
                 // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
                 assert_shape_2d(cur, n_embd, N*n_batch);
 
                 // cur = ffn_norm*cur
@@ -1034,7 +1036,7 @@ struct ggml_tensor * forward_batch(
     {
 
         // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
         assert_shape_2d(inpL, n_embd, N*n_batch);
 
         // inpL = norm*inpL
@@ -1104,7 +1106,7 @@ struct ggml_tensor * forward_lora(
         // norm
         {
             // cur shape [n_embd,N,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
             // cur = attention_norm*cur
             cur = ggml_mul(ctx0,
@@ -1251,7 +1253,7 @@ struct ggml_tensor * forward_lora(
             // norm
             {
                 // cur shape [n_embd,N,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
 
                 // cur = ffn_norm*cur
                 // cur shape [n_embd,N,1,1]
@@ -1295,7 +1297,7 @@ struct ggml_tensor * forward_lora(
     {
 
         // inpL shape [n_embd,N,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
         // inpL = norm*inpL
         // inpL shape [n_embd,N,1,1]

examples/benchmark/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/common.cpp

@@ -117,6 +117,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                params.n_threads = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -168,6 +171,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "-gqa" || arg == "--gqa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_gqa = std::stoi(argv[i]);
+        } else if (arg == "-eps" || arg == "--rms-norm-eps") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rms_norm_eps = std::stof(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -260,12 +275,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.cfg_scale = std::stof(argv[i]);
-        } else if (arg == "--cfg-smooth-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.cfg_smooth_factor = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -279,6 +288,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_keep = std::stoi(argv[i]);
+        } else if (arg == "--chunks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_chunks = std::stoi(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -387,6 +402,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } else if (arg == "--perplexity") {
             params.perplexity = true;
+        } else if (arg == "--perplexity-lines") {
+            params.perplexity_lines = true;
         } else if (arg == "--ignore-eos") {
             params.logit_bias[llama_token_eos()] = -INFINITY;
         } else if (arg == "--no-penalize-nl") {
@@ -415,6 +432,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
+        } else if (arg == "--in-prefix-bos") {
+            params.input_prefix_bos = true;
         } else if (arg == "--in-prefix") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -427,6 +446,28 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.input_suffix = argv[i];
+        } else if (arg == "--grammar") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.grammar = argv[i];
+        } else if (arg == "--grammar-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(params.grammar)
+            );
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
@@ -456,90 +497,96 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -i, --interactive     run in interactive mode\n");
-    fprintf(stderr, "  --interactive-first   run in interactive mode and wait for input right away\n");
-    fprintf(stderr, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    fprintf(stderr, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    fprintf(stderr, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stderr, "                        halt generation at PROMPT, return control in interactive mode\n");
-    fprintf(stderr, "                        (can be specified more than once for multiple prompts).\n");
-    fprintf(stderr, "  --color               colorise output to distinguish prompt and user input from generations\n");
-    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
-    fprintf(stderr, "                        prompt to start generation with (default: empty)\n");
-    fprintf(stderr, "  -e                    process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stderr, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    fprintf(stderr, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    fprintf(stderr, "                        not supported with --interactive or other interactive options\n");
-    fprintf(stderr, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    fprintf(stderr, "  --random-prompt       start with a randomized prompt.\n");
-    fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    fprintf(stderr, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    fprintf(stderr, "  -f FNAME, --file FNAME\n");
-    fprintf(stderr, "                        prompt file to start generation.\n");
-    fprintf(stderr, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
-    fprintf(stderr, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    fprintf(stderr, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    fprintf(stderr, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    fprintf(stderr, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    fprintf(stderr, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    fprintf(stderr, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
-    fprintf(stderr, "  --mirostat N          use Mirostat sampling.\n");
-    fprintf(stderr, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    fprintf(stderr, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    fprintf(stderr, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    fprintf(stderr, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
-    fprintf(stderr, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    fprintf(stderr, "                        modifies the likelihood of token appearing in the completion,\n");
-    fprintf(stderr, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    fprintf(stderr, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    fprintf(stderr, "  --cfg-negative-prompt PROMPT \n");
-    fprintf(stderr, "                        negative prompt to use for guidance. (default: empty)\n");
-    fprintf(stderr, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stderr, "  --cfg-smooth-factor N smooth factor between old and new logits (default: %f, 1.0 = no smoothing)\n", params.cfg_smooth_factor);
-    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
-    fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "usage: %s [options]\n", argv[0]);
+    fprintf(stdout, "\n");
+    fprintf(stdout, "options:\n");
+    fprintf(stdout, "  -h, --help            show this help message and exit\n");
+    fprintf(stdout, "  -i, --interactive     run in interactive mode\n");
+    fprintf(stdout, "  --interactive-first   run in interactive mode and wait for input right away\n");
+    fprintf(stdout, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    fprintf(stdout, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    fprintf(stdout, "  -r PROMPT, --reverse-prompt PROMPT\n");
+    fprintf(stdout, "                        halt generation at PROMPT, return control in interactive mode\n");
+    fprintf(stdout, "                        (can be specified more than once for multiple prompts).\n");
+    fprintf(stdout, "  --color               colorise output to distinguish prompt and user input from generations\n");
+    fprintf(stdout, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stdout, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stdout, "                        prompt to start generation with (default: empty)\n");
+    fprintf(stdout, "  -e                    process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    fprintf(stdout, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    fprintf(stdout, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    fprintf(stdout, "                        not supported with --interactive or other interactive options\n");
+    fprintf(stdout, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    fprintf(stdout, "  --random-prompt       start with a randomized prompt.\n");
+    fprintf(stdout, "  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
+    fprintf(stdout, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    fprintf(stdout, "  -f FNAME, --file FNAME\n");
+    fprintf(stdout, "                        prompt file to start generation.\n");
+    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
+    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stdout, "  -gqa N, --gqa N       grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
+    fprintf(stdout, "  -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
+    fprintf(stdout, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    fprintf(stdout, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    fprintf(stdout, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    fprintf(stdout, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    fprintf(stdout, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    fprintf(stdout, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stdout, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    fprintf(stdout, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    fprintf(stdout, "  --mirostat N          use Mirostat sampling.\n");
+    fprintf(stdout, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    fprintf(stdout, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    fprintf(stdout, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    fprintf(stdout, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    fprintf(stdout, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    fprintf(stdout, "                        modifies the likelihood of token appearing in the completion,\n");
+    fprintf(stdout, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
+    fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
+    fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
+    fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
+    fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
+    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    fprintf(stdout, "  --perplexity-lines    compute perplexity over each line of the prompt\n");
+    fprintf(stdout, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     if (llama_mlock_supported()) {
-        fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported()) {
-        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stderr, "  --numa                attempt optimizations that help on some NUMA systems\n");
-    fprintf(stderr, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    fprintf(stderr, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    fprintf(stdout, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    fprintf(stdout, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stderr, "                        number of layers to store in VRAM\n");
-    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
-    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
+    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stdout, "                        number of layers to store in VRAM\n");
+    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
-    fprintf(stderr, "  --mtest               compute maximum memory usage\n");
-    fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
-    fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stderr, "  -m FNAME, --model FNAME\n");
-    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "\n");
+    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
+    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
+    fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    fprintf(stdout, "  -m FNAME, --model FNAME\n");
+    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stdout, "\n");
 }
 
 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -575,18 +622,20 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
-    lparams.n_ctx        = params.n_ctx;
-    lparams.n_batch      = params.n_batch;
-    lparams.n_gpu_layers = params.n_gpu_layers;
-    lparams.main_gpu     = params.main_gpu;
-    memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
-    lparams.low_vram     = params.low_vram;
-    lparams.seed         = params.seed;
-    lparams.f16_kv       = params.memory_f16;
-    lparams.use_mmap     = params.use_mmap;
-    lparams.use_mlock    = params.use_mlock;
-    lparams.logits_all   = params.perplexity;
-    lparams.embedding    = params.embedding;
+    lparams.n_ctx           = params.n_ctx;
+    lparams.n_batch         = params.n_batch;
+    lparams.n_gqa           = params.n_gqa;
+    lparams.rms_norm_eps    = params.rms_norm_eps;
+    lparams.n_gpu_layers    = params.n_gpu_layers;
+    lparams.main_gpu        = params.main_gpu;
+    lparams.tensor_split    = params.tensor_split;
+    lparams.low_vram        = params.low_vram;
+    lparams.seed            = params.seed;
+    lparams.f16_kv          = params.memory_f16;
+    lparams.use_mmap        = params.use_mmap;
+    lparams.use_mlock       = params.use_mlock;
+    lparams.logits_all      = params.perplexity;
+    lparams.embedding       = params.embedding;
     lparams.rope_freq_base  = params.rope_freq_base;
     lparams.rope_freq_scale = params.rope_freq_scale;
 

examples/common.h

@@ -22,16 +22,19 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                           = -1;  // RNG seed
+    uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
-    int32_t n_predict                       = -1;  // new tokens to predict
-    int32_t n_ctx                           = 512; // context size
-    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
-    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
-    float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
-    int32_t n_probs                         = 0;   // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t n_predict                       = -1;   // new tokens to predict
+    int32_t n_ctx                           = 512;  // context size
+    int32_t n_batch                         = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_gqa                           = 1;    // grouped-query attention factor (TODO: move to hparams)
+    int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
+    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
+    int32_t n_gpu_layers                    = 0;    // number of layers to store in VRAM
+    int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
+    int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
+    float   rms_norm_eps                    = 1e-6; // rms norm epsilon
     float   rope_freq_base                  = 10000.0f; // RoPE base frequency
     float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
 
@@ -46,7 +49,7 @@ struct gpt_params {
     int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float   frequency_penalty = 0.00f; // 0.0 = disabled
     float   presence_penalty  = 0.00f; // 0.0 = disabled
-    int     mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
 
@@ -54,7 +57,6 @@ struct gpt_params {
     // https://arxiv.org/abs/2306.17806
     std::string cfg_negative_prompt;       // string to help guidance
     float       cfg_scale         = 1.f;   // How strong is guidance
-    float       cfg_smooth_factor = 1.f;   // Smooth factor between old and new logits
 
     std::string model             = "models/7B/ggml-model.bin"; // model path
     std::string model_alias       = "unknown"; // model alias
@@ -62,6 +64,7 @@ struct gpt_params {
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
     std::string lora_adapter = "";  // lora adapter path
@@ -79,9 +82,11 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
 
+    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool perplexity        = false; // compute perplexity over the prompt
+    bool perplexity_lines  = false; // compute perplexity over each line of the prompt
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage

examples/embd-input/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET embdinput)
 add_library(${TARGET} embd-input-lib.cpp embd-input.h)
+install(TARGETS ${TARGET} LIBRARY)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
@@ -8,6 +9,7 @@ endif()
 
 set(TARGET embd-input-test)
 add_executable(${TARGET} embd-input-test.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama embdinput ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/embd-input/minigpt4.py

@@ -64,7 +64,7 @@ class MiniGPT4(Blip2Base):
         self.max_txt_len = max_txt_len
         self.end_sym = end_sym
         self.model = MyModel(["main", *args])
-        # system promt
+        # system prompt
         self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
            "You will be able to see the image once I provide it to you. Please answer my questions."
            "###")

examples/embedding/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET embedding)
 add_executable(${TARGET} embedding.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/grammar-parser.cpp

@@ -0,0 +1,423 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from llama.cpp
+    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<llama_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<llama_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = LLAMA_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum llama_gretype type = last_sym_start < out_elements.size()
+                        ? LLAMA_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<llama_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<llama_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({LLAMA_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    bool is_char_element(llama_grammar_element elem) {
+        switch (elem.type) {
+            case LLAMA_GRETYPE_CHAR:           return true;
+            case LLAMA_GRETYPE_CHAR_NOT:       return true;
+            case LLAMA_GRETYPE_CHAR_ALT:       return true;
+            case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
+                case LLAMA_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case LLAMA_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case LLAMA_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                case LLAMA_GRETYPE_ALT:
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                case LLAMA_GRETYPE_CHAR_NOT:
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<llama_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            llama_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case LLAMA_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case LLAMA_GRETYPE_CHAR_ALT:
+                    case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, i, state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const llama_grammar_element *> parse_state::c_rules() {
+        std::vector<const llama_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

examples/grammar-parser.h

@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by llama.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "llama.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                 symbol_ids;
+        std::vector<std::vector<llama_grammar_element>> rules;
+
+        std::vector<const llama_grammar_element *> c_rules();
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}

examples/llama2-13b.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/13B/llama-2-13b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8

examples/llama2.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/7B/llama-2-7b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8

examples/llm.vim

@@ -0,0 +1,23 @@
+function! Llm()
+
+  let url = "http://127.0.0.1:8080/completion"
+
+  " Get the content of the current buffer
+  let buffer_content = join(getline(1, '$'), "\n")
+
+  " Create the JSON payload
+  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false}
+  let json_payload.prompt = buffer_content
+
+  " Define the curl command
+  let curl_command = 'curl -k -s -X POST -H "Content-Type: application/json" -d @- ' . url
+  let response = system(curl_command, json_encode(json_payload))
+
+  " Extract the content field from the response
+  let content = json_decode(response).content
+
+  " Insert the content at the cursor position
+  call setline(line('.'), getline('.') . content)
+endfunction
+
+command! Llm call Llm()

examples/main/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET main)
 add_executable(${TARGET} main.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/main/main.cpp

@@ -6,6 +6,7 @@
 #include "common.h"
 #include "llama.h"
 #include "build-info.h"
+#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -93,8 +94,8 @@ int main(int argc, char ** argv) {
     }
 
     if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified);"
-                " you are on your own\n", __func__, params.n_ctx);
+        // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
+        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
     } else if (params.n_ctx < 8) {
         fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
@@ -139,17 +140,14 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    // determine the maximum memory usage needed to do inference for the given n_batch and n_predict parameters
+    // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters
     // uncomment the "used_mem" line in llama.cpp to see the results
     if (params.mem_test) {
         {
-            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos());
-            llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
-        }
+            fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
 
-        {
-            const std::vector<llama_token> tmp = { 0, };
-            llama_eval(ctx, tmp.data(), tmp.size(), params.n_predict - 1, params.n_threads);
+            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos());
+            llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
         }
 
         llama_print_timings(ctx);
@@ -327,6 +325,10 @@ int main(int argc, char ** argv) {
             }
         }
 
+        if (params.input_prefix_bos) {
+            fprintf(stderr, "Input prefix with BOS\n");
+        }
+
         if (!params.input_prefix.empty()) {
             fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
         }
@@ -340,6 +342,31 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "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);
     fprintf(stderr, "\n\n");
 
+    grammar_parser::parse_state parsed_grammar;
+    llama_grammar *             grammar = NULL;
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        fprintf(stderr, "%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        fprintf(stderr, "\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos());
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                fprintf(stderr,
+                    "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
     // TODO: replace with ring-buffer
     std::vector<llama_token> last_n_tokens(n_ctx);
     std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
@@ -557,7 +584,7 @@ int main(int argc, char ** argv) {
                 llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
                 if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale, params.cfg_smooth_factor);
+                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
                 }
 
                 // Apply penalties
@@ -573,6 +600,10 @@ int main(int argc, char ** argv) {
                     logits[llama_token_nl()] = nl_logit;
                 }
 
+                if (grammar != NULL) {
+                    llama_sample_grammar(ctx, &candidates_p, grammar);
+                }
+
                 if (temp <= 0) {
                     // Greedy sampling
                     id = llama_sample_token_greedy(ctx, &candidates_p);
@@ -598,20 +629,14 @@ int main(int argc, char ** argv) {
                 }
                 // printf("`%d`", candidates_p.size);
 
+                if (grammar != NULL) {
+                    llama_grammar_accept_token(ctx, grammar, id);
+                }
+
                 last_n_tokens.erase(last_n_tokens.begin());
                 last_n_tokens.push_back(id);
             }
 
-            // replace end of text token with newline token when in interactive mode
-            if (id == llama_token_eos() && params.interactive && !params.instruct) {
-                id = llama_token_newline.front();
-                if (params.antiprompt.size() != 0) {
-                    // tokenize and inject first reverse prompt
-                    const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
-                    embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
-                }
-            }
-
             // add it to the context
             embd.push_back(id);
 
@@ -677,11 +702,34 @@ int main(int argc, char ** argv) {
                 }
             }
 
+            // deal with end of text token in interactive mode
+            if (last_n_tokens.back() == llama_token_eos()) {
+                if (params.interactive) {
+                    if (params.antiprompt.size() != 0) {
+                        // tokenize and inject first reverse prompt
+                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
+                        embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
+                        is_antiprompt = true;
+                    }
+
+                    is_interacting = true;
+                    printf("\n");
+                    console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                    fflush(stdout);
+                } else if (params.instruct) {
+                    is_interacting = true;
+                }
+            }
+
             if (n_past > 0 && is_interacting) {
                 if (params.instruct) {
                     printf("\n> ");
                 }
 
+                if (params.input_prefix_bos) {
+                    embd_inp.push_back(llama_token_bos());
+                }
+
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
                     buffer += params.input_prefix;
@@ -728,18 +776,26 @@ int main(int argc, char ** argv) {
             }
 
             if (n_past > 0) {
+                if (is_interacting) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(
+                            parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos()) {
-            if (params.instruct) {
-                is_interacting = true;
-            } else {
-                fprintf(stderr, " [end of text]\n");
-                break;
-            }
+        if (!embd.empty() && embd.back() == llama_token_eos() && !(params.instruct || params.interactive)) {
+            fprintf(stderr, " [end of text]\n");
+            break;
         }
 
         // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
@@ -759,6 +815,9 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
     llama_backend_free();
 
     return 0;

examples/make-ggml.py

@@ -0,0 +1,92 @@
+"""
+This script converts Hugging Face llama models to GGML and quantizes them.
+
+Usage:
+python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+
+Arguments:
+- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
+- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
+- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
+- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
+
+Quant types:
+- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
+- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
+- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
+- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+- Q2_K: smallest, extreme quality loss - not recommended
+- Q3_K: alias for Q3_K_M
+- Q3_K_S: very small, very high quality loss
+- Q3_K_M: very small, very high quality loss
+- Q3_K_L: small, substantial quality loss
+- Q4_K: alias for Q4_K_M
+- Q4_K_S: small, significant quality loss
+- Q4_K_M: medium, balanced quality - recommended
+- Q5_K: alias for Q5_K_M
+- Q5_K_S: large, low quality loss - recommended
+- Q5_K_M: large, very low quality loss - recommended
+- Q6_K: very large, extremely low quality loss
+- Q8_0: very large, extremely low quality loss - not recommended
+- F16: extremely large, virtually no quality loss - not recommended
+- F32: absolutely huge, lossless - not recommended
+"""
+import subprocess
+subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True)
+
+import argparse
+import os
+from huggingface_hub import snapshot_download
+
+def main(model, outname, outdir, quants, keep_fp16):
+    ggml_version = "v3"
+
+    if not os.path.isdir(model):
+        print(f"Model not found at {model}. Downloading...")
+        try:
+            if outname is None:
+                outname = model.split('/')[-1]
+            model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache')
+        except Exception as e:
+            raise Exception(f"Could not download the model: {e}")
+
+    if outdir is None:
+        outdir = f'../models/{outname}'
+
+    if not os.path.isfile(f"{model}/config.json"):
+        raise Exception(f"Could not find config.json in {model}")
+
+    os.makedirs(outdir, exist_ok=True)
+
+    print("Building llama.cpp")
+    subprocess.run(f"cd .. && make quantize", shell=True, check=True)
+
+    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+
+    print(f"Making unquantised GGML at {fp16}")
+    if not os.path.isfile(fp16):
+        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+    else:
+        print(f"Unquantised GGML already exists at: {fp16}")
+
+    print("Making quants")
+    for type in quants:
+        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        print(f"Making {type} : {outfile}")
+        subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
+
+    if not keep_fp16:
+        os.remove(fp16)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--outname', default=None, help='Output model(s) name')
+    parser.add_argument('--outdir', default=None, help='Output directory')
+    parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
+    parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False)
+
+    args = parser.parse_args()
+
+    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)

examples/metal/CMakeLists.txt

@@ -1,3 +1,4 @@
 set(TEST_TARGET metal)
 add_executable(${TEST_TARGET} metal.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/perplexity/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET perplexity)
 add_executable(${TARGET} perplexity.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/perplexity/perplexity.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <ctime>
+#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -32,13 +33,15 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     // BOS tokens will be added for each chunk before eval
     auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    int count   = 0;
+    const int n_chunk_max = tokens.size() / params.n_ctx;
 
-    const int n_chunk = tokens.size() / params.n_ctx;
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_vocab = llama_n_vocab(ctx);
     const int n_batch = params.n_batch;
 
+    int count = 0;
     double nll = 0.0;
+
     fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
     for (int i = 0; i < n_chunk; ++i) {
@@ -118,6 +121,77 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     printf("\n");
 }
 
+void perplexity_lines(llama_context * ctx, const gpt_params & params) {
+    // Calculates perplexity over each line of the prompt
+
+    std::vector<std::string> prompt_lines;
+    std::istringstream strstream(params.prompt);
+    std::string line;
+
+    while (std::getline(strstream,line,'\n')) {
+        prompt_lines.push_back(line);
+    }
+
+    const int n_vocab = llama_n_vocab(ctx);
+
+    int counttotal   = 0;
+    size_t n_lines = prompt_lines.size();
+
+    double nll = 0.0;
+
+    fprintf(stderr, "%s: calculating perplexity over %lu lines\n", __func__, n_lines);
+
+    printf("\nLine\tPPL line\tPPL cumulative\n");
+
+    for (size_t i = 0; i < n_lines; ++i) {
+
+        // Tokenize and insert BOS at start
+        std::vector<int> batch_embd = ::llama_tokenize(ctx, prompt_lines[i], true);
+
+        size_t batch_size  = batch_embd.size();
+
+        // Stop if line is too long
+        if( batch_size > (size_t)params.n_ctx ) {
+            fprintf(stderr, "%s : tokens in line %lu > n_ctxl\n", __func__, i);
+            return;
+        }
+
+        if (llama_eval(ctx, batch_embd.data(), batch_size, 0, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return;
+        }
+
+        const auto batch_logits = llama_get_logits(ctx);
+        std::vector<float> logits;
+        logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+
+        double nllline = 0.0;
+        int countline = 0;
+
+        // Perplexity over second half of the line
+        for (size_t j = batch_size/2; j < batch_size - 1; ++j) {
+            // Calculate probability of next token, given the previous ones.
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+
+            const float prob = softmax(tok_logits)[batch_embd[ j + 1]];
+
+            nllline += -std::log(prob);
+            ++countline;
+        }
+
+        nll += nllline;
+        counttotal += countline;
+
+        // perplexity is e^(average negative log-likelihood)
+        printf("%lu\t%.8lf\t%.8lf\n", i + 1, std::exp(nllline/countline), std::exp(nll / counttotal) );
+        fflush(stdout);
+    }
+
+    printf("\n");
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -166,7 +240,11 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    perplexity(ctx, params);
+    if (params.perplexity_lines) {
+        perplexity_lines(ctx, params);
+    } else {
+        perplexity(ctx, params);
+    }
 
     llama_print_timings(ctx);
     llama_free(ctx);

examples/quantize-stats/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/quantize/quantize.cpp

@@ -14,103 +14,27 @@ struct quant_option {
 };
 
 static const std::vector<struct quant_option> QUANT_OPTIONS = {
-    {
-        "Q4_0",
-        LLAMA_FTYPE_MOSTLY_Q4_0,
-        " 3.50G, +0.2499 ppl @ 7B - small, very high quality loss - legacy, prefer using Q3_K_M",
-    },
-    {
-        "Q4_1",
-        LLAMA_FTYPE_MOSTLY_Q4_1,
-        " 3.90G, +0.1846 ppl @ 7B - small, substantial quality loss - legacy, prefer using Q3_K_L",
-    },
-    {
-        "Q5_0",
-        LLAMA_FTYPE_MOSTLY_Q5_0,
-        " 4.30G, +0.0796 ppl @ 7B - medium, balanced quality - legacy, prefer using Q4_K_M",
-    },
-    {
-        "Q5_1",
-        LLAMA_FTYPE_MOSTLY_Q5_1,
-        " 4.70G, +0.0415 ppl @ 7B - medium, low quality loss - legacy, prefer using Q5_K_M",
-    },
+    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.50G, +0.2499 ppl @ 7B", },
+    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1846 ppl @ 7B", },
+    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.30G, +0.0796 ppl @ 7B", },
+    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0415 ppl @ 7B", },
 #ifdef GGML_USE_K_QUANTS
-    {
-        "Q2_K",
-        LLAMA_FTYPE_MOSTLY_Q2_K,
-        " 2.67G, +0.8698 ppl @ 7B - smallest, extreme quality loss - not recommended",
-    },
-    {
-        "Q3_K",
-        LLAMA_FTYPE_MOSTLY_Q3_K_M,
-        "alias for Q3_K_M"
-    },
-    {
-        "Q3_K_S",
-        LLAMA_FTYPE_MOSTLY_Q3_K_S,
-        " 2.75G, +0.5505 ppl @ 7B - very small, very high quality loss",
-    },
-    {
-        "Q3_K_M",
-        LLAMA_FTYPE_MOSTLY_Q3_K_M,
-        " 3.06G, +0.2437 ppl @ 7B - very small, very high quality loss",
-    },
-    {
-        "Q3_K_L",
-        LLAMA_FTYPE_MOSTLY_Q3_K_L,
-        " 3.35G, +0.1803 ppl @ 7B - small, substantial quality loss",
-    },
-    {
-        "Q4_K",
-        LLAMA_FTYPE_MOSTLY_Q4_K_M,
-        "alias for Q4_K_M",
-    },
-    {
-        "Q4_K_S",
-        LLAMA_FTYPE_MOSTLY_Q4_K_S,
-        " 3.56G, +0.1149 ppl @ 7B - small, significant quality loss",
-    },
-    {
-        "Q4_K_M",
-        LLAMA_FTYPE_MOSTLY_Q4_K_M,
-        " 3.80G, +0.0535 ppl @ 7B - medium, balanced quality - *recommended*",
-    },
-    {
-        "Q5_K",
-        LLAMA_FTYPE_MOSTLY_Q5_K_M,
-        "alias for Q5_K_M",
-    },
-    {
-        "Q5_K_S",
-        LLAMA_FTYPE_MOSTLY_Q5_K_S,
-        " 4.33G, +0.0353 ppl @ 7B - large, low quality loss - *recommended*",
-    },
-    {
-        "Q5_K_M",
-        LLAMA_FTYPE_MOSTLY_Q5_K_M,
-        " 4.45G, +0.0142 ppl @ 7B - large, very low quality loss - *recommended*",
-    },
-    {
-        "Q6_K",
-        LLAMA_FTYPE_MOSTLY_Q6_K,
-        " 5.15G, +0.0044 ppl @ 7B - very large, extremely low quality loss",
-    },
+    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.67G, +0.8698 ppl @ 7B", },
+    { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
+    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5505 ppl @ 7B", },
+    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.06G, +0.2437 ppl @ 7B", },
+    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1803 ppl @ 7B", },
+    { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
+    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.56G, +0.1149 ppl @ 7B", },
+    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0535 ppl @ 7B", },
+    { "Q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
+    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0353 ppl @ 7B", },
+    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0142 ppl @ 7B", },
+    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0044 ppl @ 7B", },
 #endif
-    {
-        "Q8_0",
-        LLAMA_FTYPE_MOSTLY_Q8_0,
-        " 6.70G, +0.0004 ppl @ 7B - very large, extremely low quality loss - not recommended",
-    },
-    {
-        "F16",
-        LLAMA_FTYPE_MOSTLY_F16,
-        "13.00G              @ 7B - extremely large, virtually no quality loss - not recommended",
-    },
-    {
-        "F32",
-        LLAMA_FTYPE_ALL_F32,
-        "26.00G              @ 7B - absolutely huge, lossless - not recommended",
-    },
+    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ 7B", },
+    { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
+    { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
 };
 
 

examples/save-load-state/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET save-load-state)
 add_executable(${TARGET} save-load-state.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/server/CMakeLists.txt

@@ -2,10 +2,14 @@ set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 add_executable(${TARGET} server.cpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/server/public/index.html

@@ -73,6 +73,37 @@
       margin: 0;
     }
 
+    fieldset.two {
+      display: grid;
+      grid-template: "a a";
+      gap: 1em;
+    }
+
+    fieldset.three {
+      display: grid;
+      grid-template: "a a a";
+      gap: 1em;
+    }
+
+    details {
+      border: 1px solid #aaa;
+      border-radius: 4px;
+      padding: 0.5em 0.5em 0;
+      margin-top: 0.5em;
+    }
+
+    summary {
+      font-weight: bold;
+      margin: -0.5em -0.5em 0;
+      padding: 0.5em;
+      cursor: pointer;
+    }
+
+    details[open] {
+      padding: 0.5em;
+    }
+
+
     textarea {
       padding: 5px;
       flex-grow: 1;
@@ -125,10 +156,17 @@
     const params = signal({
       n_predict: 400,
       temperature: 0.7,
-      repeat_last_n: 256,
-      repeat_penalty: 1.18,
-      top_k: 40,
-      top_p: 0.5,
+      repeat_last_n: 256, // 0 = disable penalty, -1 = context size
+      repeat_penalty: 1.18, // 1.0 = disabled
+      top_k: 40, // <= 0 to use vocab size
+      top_p: 0.5, // 1.0 = disabled
+      tfs_z: 1.0, // 1.0 = disabled
+      typical_p: 1.0, // 1.0 = disabled
+      presence_penalty: 0.0, // 0.0 = disabled
+      frequency_penalty: 0.0, // 0.0 = disabled
+      mirostat: 0, // 0/1/2
+      mirostat_tau: 5, // target entropy
+      mirostat_eta: 0.1, // learning rate
     })
 
     const llamaStats = signal(null)
@@ -264,6 +302,27 @@
       const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
       const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
       const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
+      const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
+
+      const FloatField = ({label, max, min, name, step, value}) => {
+        return html`
+          <div>
+            <label for="${name}">${label}</label>
+            <input type="range" id="${name}" min="${min}" max="${max}" step="${step}" name="${name}" value="${value}" oninput=${updateParamsFloat} />
+            <span>${value}</span>
+          </div>
+        `
+      };
+
+      const IntField = ({label, max, min, name, value}) => {
+        return html`
+          <div>
+            <label for="${name}">${label}</label>
+            <input type="range" id="${name}" min="${min}" max="${max}" name="${name}" value="${value}" oninput=${updateParamsInt} />
+            <span>${value}</span>
+          </div>
+        `
+      };
 
       return html`
         <form>
@@ -272,7 +331,9 @@
               <label for="prompt">Prompt</label>
               <textarea type="text" name="prompt" value="${session.value.prompt}" rows=4 oninput=${updateSession}/>
             </div>
+          </fieldset>
 
+          <fieldset class="two">
             <div>
               <label for="user">User name</label>
               <input type="text" name="user" value="${session.value.user}" oninput=${updateSession} />
@@ -282,7 +343,9 @@
               <label for="bot">Bot name</label>
               <input type="text" name="char" value="${session.value.char}" oninput=${updateSession} />
             </div>
+          </fieldset>
 
+          <fieldset>
             <div>
               <label for="template">Prompt template</label>
               <textarea id="template" name="template" value="${session.value.template}" rows=4 oninput=${updateSession}/>
@@ -292,38 +355,44 @@
               <label for="template">Chat history template</label>
               <textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/>
             </div>
-
-            <div>
-              <label for="temperature">Temperature</label>
-              <input type="range" id="temperature" min="0.0" max="1.0" step="0.01" name="temperature" value="${params.value.temperature}" oninput=${updateParamsFloat} />
-              <span>${params.value.temperature}</span>
-            </div>
-
-            <div>
-              <label for="nPredict">Predictions</label>
-              <input type="range" id="nPredict" min="1" max="2048" step="1" name="n_predict" value="${params.value.n_predict}" oninput=${updateParamsFloat} />
-              <span>${params.value.n_predict}</span>
-            </div>
-
-            <div>
-              <label for="repeat_penalty">Penalize repeat sequence</label>
-              <input type="range" id="repeat_penalty" min="0.0" max="2.0" step="0.01" name="repeat_penalty" value="${params.value.repeat_penalty}" oninput=${updateParamsFloat} />
-              <span>${params.value.repeat_penalty}</span>
-            </div>
-
-            <div>
-              <label for="repeat_last_n">Consider N tokens for penalize</label>
-              <input type="range" id="repeat_last_n" min="0.0" max="2048" name="repeat_last_n" value="${params.value.repeat_last_n}" oninput=${updateParamsFloat} />
-              <span>${params.value.repeat_last_n}</span>
-            </div>
-
           </fieldset>
+
+          <fieldset class="two">
+            ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})}
+            ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
+            ${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})}
+            ${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})}
+            ${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})}
+            ${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})}
+          </fieldset>
+          <details>
+            <summary>More options</summary>
+            <fieldset class="two">
+              ${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})}
+              ${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})}
+              ${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})}
+              ${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})}
+            </fieldset>
+            <hr />
+            <fieldset class="three">
+              <div>
+                <label><input type="radio" name="mirostat" value="0" checked=${params.value.mirostat == 0} oninput=${updateParamsInt} /> no Mirostat</label>
+                <label><input type="radio" name="mirostat" value="1" checked=${params.value.mirostat == 1} oninput=${updateParamsInt} /> Mirostat v1</label>
+                <label><input type="radio" name="mirostat" value="2" checked=${params.value.mirostat == 2} oninput=${updateParamsInt} /> Mirostat v2</label>
+              </div>
+              ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
+              ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
+            </fieldset>
+          </details>
         </form>
       `
     }
     // poor mans markdown replacement
     const Markdownish = (params) => {
       const md = params.text
+        .replace(/&/g, '&amp;')
+        .replace(/</g, '&lt;')
+        .replace(/>/g, '&gt;')
         .replace(/^#{1,6} (.*)$/gim, '<h3>$1</h3>')
         .replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
         .replace(/__(.*?)__/g, '<strong>$1</strong>')

examples/server/server.cpp

@@ -601,47 +601,49 @@ struct llama_server_context
 static void server_print_usage(const char *argv0, const gpt_params &params,
                                const server_params &sparams)
 {
-    fprintf(stderr, "usage: %s [options]\n", argv0);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    fprintf(stdout, "usage: %s [options]\n", argv0);
+    fprintf(stdout, "\n");
+    fprintf(stdout, "options:\n");
+    fprintf(stdout, "  -h, --help            show this help message and exit\n");
+    fprintf(stdout, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stdout, "  -gqa N, --gqa N       grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
+    fprintf(stdout, "  -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
     {
-        fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported())
     {
-        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stderr, "                        number of layers to store in VRAM\n");
-    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stderr, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stdout, "                        number of layers to store in VRAM\n");
+    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
 #endif
-    fprintf(stderr, "  -m FNAME, --model FNAME\n");
-    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
-    fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
-    fprintf(stderr, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    fprintf(stderr, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    fprintf(stderr, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    fprintf(stderr, "\n");
+    fprintf(stdout, "  -m FNAME, --model FNAME\n");
+    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stdout, "  -a ALIAS, --alias ALIAS\n");
+    fprintf(stdout, "                        set an alias for the model, will be added as `model` field in completion response\n");
+    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    fprintf(stdout, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    fprintf(stdout, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    fprintf(stdout, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    fprintf(stdout, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    fprintf(stdout, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    fprintf(stdout, "\n");
 }
 
 static void server_params_parse(int argc, char **argv, server_params &sparams,
@@ -724,9 +726,27 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_ctx = std::stoi(argv[i]);
         }
+        else if (arg == "-gqa" || arg == "--gqa")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_gqa = std::stoi(argv[i]);
+        }
+        else if (arg == "-eps" || arg == "--rms-norm-eps") {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.rms_norm_eps = std::stof(argv[i]);
+        }
         else if (arg == "--rope-freq-base")
         {
-            if (++i >= argc) {
+            if (++i >= argc)
+            {
                 invalid_param = true;
                 break;
             }
@@ -734,7 +754,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         }
         else if (arg == "--rope-freq-scale")
         {
-            if (++i >= argc) {
+            if (++i >= argc)
+            {
                 invalid_param = true;
                 break;
             }

examples/simple/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(TARGET simple)
 add_executable(${TARGET} simple.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/train-text-from-scratch/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(TARGET train-text-from-scratch)
 add_executable(${TARGET} train-text-from-scratch.cpp)
+install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -16,6 +16,8 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static const float rms_norm_eps = 1e-6f;
+
 struct random_normal_distribution {
     std::mt19937 gen;
     std::normal_distribution<float> rd;
@@ -439,7 +441,7 @@ struct ggml_tensor * forward(
         // norm
         {
             // cur shape [n_embd,N,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
             // cur = attention_norm*cur
             cur = ggml_mul(ctx0,
@@ -562,7 +564,7 @@ struct ggml_tensor * forward(
             // norm
             {
                 // cur shape [n_embd,N,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
 
                 // cur = ffn_norm*cur
                 // cur shape [n_embd,N,1,1]
@@ -606,7 +608,7 @@ struct ggml_tensor * forward(
     {
 
         // inpL shape [n_embd,N,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
 
         // inpL = norm*inpL
         // inpL shape [n_embd,N,1,1]
@@ -694,7 +696,7 @@ struct ggml_tensor * forward_batch(
         // norm
         {
             // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
             assert_shape_2d(cur, n_embd, N*n_batch);
 
             // cur = attention_norm*cur
@@ -857,7 +859,7 @@ struct ggml_tensor * forward_batch(
             // norm
             {
                 // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
                 assert_shape_2d(cur, n_embd, N*n_batch);
 
                 // cur = ffn_norm*cur
@@ -910,7 +912,7 @@ struct ggml_tensor * forward_batch(
     {
 
         // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
         assert_shape_2d(inpL, n_embd, N*n_batch);
 
         // inpL = norm*inpL
@@ -979,7 +981,7 @@ struct ggml_tensor * forward_batch_wo_cache(
         // norm
         {
             // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
             assert_shape_2d(cur, n_embd, N*n_batch);
 
             // cur = attention_norm*cur
@@ -1085,7 +1087,7 @@ struct ggml_tensor * forward_batch_wo_cache(
             // norm
             {
                 // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
                 assert_shape_2d(cur, n_embd, N*n_batch);
 
                 // cur = ffn_norm*cur
@@ -1138,7 +1140,7 @@ struct ggml_tensor * forward_batch_wo_cache(
     {
 
         // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
         assert_shape_2d(inpL, n_embd, N*n_batch);
 
         // inpL = norm*inpL
@@ -1203,7 +1205,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
 
         // norm
         {
-            cur = ggml_rms_norm(ctx0, inpL);
+            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
             assert_shape_2d(cur, n_embd, N*n_batch);
 
             // cur = attention_norm*cur
@@ -1267,7 +1269,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
         {
             // norm
             {
-                cur = ggml_rms_norm(ctx0, inpFF);
+                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
                 assert_shape_2d(cur, n_embd, N*n_batch);
 
                 // cur = ffn_norm*cur
@@ -1311,7 +1313,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
     // norm
     {
 
-        inpL = ggml_rms_norm(ctx0, inpL);
+        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
         assert_shape_2d(inpL, n_embd, N*n_batch);
 
         // inpL = norm*inpL
@@ -1434,7 +1436,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
     gf->perf_time_us = 0;
 
     const auto & hparams = model->hparams;
-    //const int n_ctx      = hparams.n_ctx;
+    const int n_ctx      = hparams.n_ctx;
     const int n_vocab    = hparams.n_vocab;
     const int n_embd     = hparams.n_embd;
     const int n_layer    = hparams.n_layer;
@@ -1603,7 +1605,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         struct my_llama_layer & layer = model->layers[il];
         // tensors with values necessary for backward pass are in persistent buf(-1)
         // other tensors with buf(0) and buf(1) are only temporary needed, and their memory reused after layer is completed.
-        use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm     (ctx0, cur));                                    assert_shape_2d(t02, n_embd, N*n_batch);
+        use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm     (ctx0, cur, rms_norm_eps));                      assert_shape_2d(t02, n_embd, N*n_batch);
         use_buf( 0); struct ggml_tensor * t03 = expand(gf, ggml_repeat       (ctx0, layer.attention_norm, t02));              assert_shape_2d(t03, n_embd, N*n_batch);
         use_buf(-1); struct ggml_tensor * t04 = expand(gf, ggml_mul          (ctx0, t02, t03));                               assert_shape_2d(t04, n_embd, N*n_batch);
         use_buf(-1); struct ggml_tensor * t05 = expand(gf, ggml_mul_mat      (ctx0, layer.wq, t04));                          assert_shape_2d(t05, n_embd, N*n_batch);
@@ -1623,7 +1625,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         use_buf(-1); struct ggml_tensor * t19 = expand(gf, ggml_reshape_2d   (ctx0, t18, n_embd, N*n_batch));                 assert_shape_2d(t19, n_embd, N*n_batch);
         use_buf( 0); struct ggml_tensor * t20 = expand(gf, ggml_mul_mat      (ctx0, layer.wo, t19));                          assert_shape_2d(t20, n_embd, N*n_batch);
         use_buf(-1); struct ggml_tensor * t21 = expand(gf, ggml_add          (ctx0, t20, cur));                               assert_shape_2d(t21, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm     (ctx0, t21));                                    assert_shape_2d(t22, n_embd, N*n_batch);
+        use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm     (ctx0, t21, rms_norm_eps));                      assert_shape_2d(t22, n_embd, N*n_batch);
         use_buf( 0); struct ggml_tensor * t23 = expand(gf, ggml_repeat       (ctx0, layer.ffn_norm, t22));                    assert_shape_2d(t23, n_embd, N*n_batch);
         use_buf(-1); struct ggml_tensor * t24 = expand(gf, ggml_mul          (ctx0, t23, t22));                               assert_shape_2d(t24, n_embd, N*n_batch);
         use_buf(-1); struct ggml_tensor * t25 = expand(gf, ggml_mul_mat      (ctx0, layer.w3, t24));                          assert_shape_2d(t25, n_ff, N*n_batch);
@@ -1666,7 +1668,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
     }
     clr_buf(0);
     use_buf(0);
-    struct ggml_tensor * t31   = expand(gf, ggml_rms_norm  (ctx0, cur));                       assert_shape_2d(t31, n_embd, N*n_batch);
+    struct ggml_tensor * t31   = expand(gf, ggml_rms_norm  (ctx0, cur, rms_norm_eps));         assert_shape_2d(t31, n_embd, N*n_batch);
     struct ggml_tensor * t32   = expand(gf, ggml_repeat    (ctx0, model->norm, t31));          assert_shape_2d(t32, n_embd, N*n_batch);
     struct ggml_tensor * t33   = expand(gf, ggml_mul       (ctx0, t32, t31));                  assert_shape_2d(t33, n_embd, N*n_batch);
     use_buf(-1);
@@ -1863,10 +1865,10 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1));                                            assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
         t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd));                 assert_shape_2d(t11->grad, N*n_batch, n_embd);
         t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3));                                            assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
-        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
+        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
         t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch));                                  assert_shape_2d(t08->grad, n_embd, N*n_batch);
         t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3));                                            assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
-        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
+        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
         t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch));                                  assert_shape_2d(t05->grad, n_embd, N*n_batch);
         t04->grad = expand(gb, ggml_add_inplace(ctx0,
                         ggml_add_inplace(ctx0,

flake.nix

@@ -6,56 +6,68 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
-        inherit (pkgs.stdenv) isAarch64 isDarwin;
-        inherit (pkgs.lib) optionals;
-        isM1 = isAarch64 && isDarwin;
-        osSpecific = if isM1 then
-          with pkgs.darwin.apple_sdk_11_0.frameworks; [
-            Accelerate
-            MetalKit
-            MetalPerformanceShaders
-            MetalPerformanceShadersGraph
-          ]
-        else if isDarwin then
-          with pkgs.darwin.apple_sdk.frameworks; [
-            Accelerate
-            CoreGraphics
-            CoreVideo
-          ]
-        else
-          [ ];
+        inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
+        buildInputs = with pkgs; [ openmpi ];
+        osSpecific = with pkgs; buildInputs ++
+        (
+          if isAarch64 && isDarwin then
+            with pkgs.darwin.apple_sdk_11_0.frameworks; [
+              Accelerate
+              MetalKit
+              MetalPerformanceShaders
+              MetalPerformanceShadersGraph
+            ]
+          else if isAarch32 && isDarwin then
+            with pkgs.darwin.apple_sdk.frameworks; [
+              Accelerate
+              CoreGraphics
+              CoreVideo
+            ]
+          else
+            with pkgs; [ openblas ]
+        );
         pkgs = import nixpkgs { inherit system; };
+        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
         llama-python =
-          pkgs.python310.withPackages (ps: with ps; [ numpy sentencepiece ]);
+          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
+        postPatch = ''
+          substituteInPlace ./ggml-metal.m \
+            --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+          substituteInPlace ./*.py --replace '/usr/bin/env python' '${llama-python}/bin/python'
+        '';
+        postInstall = ''
+          mv $out/bin/main $out/bin/llama
+          mv $out/bin/server $out/bin/llama-server
+        '';
+        cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
       in {
         packages.default = pkgs.stdenv.mkDerivation {
           name = "llama.cpp";
           src = ./.;
-          postPatch = if isM1 then ''
-            substituteInPlace ./ggml-metal.m \
-              --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-          '' else
-            "";
-          nativeBuildInputs = with pkgs; [ cmake ];
+          postPatch = postPatch;
+          nativeBuildInputs = nativeBuildInputs;
           buildInputs = osSpecific;
-          cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" ] ++ (optionals isM1 [
-            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-            "-DLLAMA_METAL=ON"
+          cmakeFlags = cmakeFlags
+            ++ (if isAarch64 && isDarwin then [
+              "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
+              "-DLLAMA_METAL=ON"
+            ] else [
+              "-DLLAMA_BLAS=ON"
+              "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          installPhase = ''
-            runHook preInstall
-
-            mkdir -p $out/bin
-            mv bin/* $out/bin/
-            mv $out/bin/main $out/bin/llama
-            mv $out/bin/server $out/bin/llama-server
-
-            echo "#!${llama-python}/bin/python" > $out/bin/convert.py
-            cat ${./convert.py} >> $out/bin/convert.py
-            chmod +x $out/bin/convert.py
-
-            runHook postInstall
-          '';
+          postInstall = postInstall;
+          meta.mainProgram = "llama";
+        };
+        packages.opencl = pkgs.stdenv.mkDerivation {
+          name = "llama.cpp";
+          src = ./.;
+          postPatch = postPatch;
+          nativeBuildInputs = nativeBuildInputs;
+          buildInputs = with pkgs; buildInputs ++ [ clblast ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_CLBLAST=ON"
+          ];
+          postInstall = postInstall;
           meta.mainProgram = "llama";
         };
         apps.llama-server = {
@@ -72,7 +84,7 @@
         };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
-          packages = with pkgs; [ cmake llama-python ] ++ osSpecific;
+          packages = nativeBuildInputs ++ osSpecific;
         };
       });
 }

ggml-cuda.cu

@@ -220,7 +220,7 @@ typedef struct {
 static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
 
 #define WARP_SIZE 32
-#define MATRIX_ROW_PADDING 256 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
 #define CUDA_ADD_BLOCK_SIZE 256
 #define CUDA_MUL_BLOCK_SIZE 256
@@ -332,12 +332,10 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     }
 }
 
-static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols) {
+static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
-    const float eps = 1e-6f;
-
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += WARP_SIZE) {
@@ -935,12 +933,18 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
     uint16_t aux[4];
     const uint8_t * sc = (const uint8_t *)aux;
 
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
     float tmp = 0; // partial sum for thread in warp
 
     for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
-        const uint8_t * q1 = x[i].qs + q_offset;
-        const uint8_t * q2 = q1 + 64;
         const float   * y1 = yy + i*QK_K + y_offset;
         const float   * y2 = y1 + 128;
 
@@ -953,14 +957,41 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
         aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
 
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
         float4 s = {0.f, 0.f, 0.f, 0.f};
         float smin = 0;
-        for (int l = 0; l < n; ++l) {
-            s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
-            s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
+        for (int l = 0; l < 4; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
+            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
             smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
         }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
 
     }
 #else
@@ -1040,10 +1071,12 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
     uint16_t aux[4];
     const uint8_t * sc = (const uint8_t *)aux;
 
+    uint16_t q16[8];
+    const uint8_t * q4 = (const uint8_t *)q16;
+
     for (int i = ix; i < num_blocks_per_row; i += 2) {
 
         const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * ql2 = ql1 + 64;
         const uint8_t * qh  = x[i].qh + l0;
         const float   * y1  = yy + i*QK_K + y_offset;
         const float   * y2  = y1 + 128;
@@ -1059,15 +1092,25 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 
         float4 sum = {0.f, 0.f, 0.f, 0.f};
         float smin = 0;
+        const uint16_t * q1 = (const uint16_t *)ql1;
+        const uint16_t * q2 = q1 + 32;
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[8] & 0x0f0f;
+        q16[2] = (q1[0] >> 4) & 0x0f0f;
+        q16[3] = (q1[8] >> 4) & 0x0f0f;
+        q16[4] = q2[0] & 0x0f0f;
+        q16[5] = q2[8] & 0x0f0f;
+        q16[6] = (q2[0] >> 4) & 0x0f0f;
+        q16[7] = (q2[8] >> 4) & 0x0f0f;
         for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
             smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
                   + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
         }
@@ -1521,33 +1564,95 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
-    const int bq8_offset = QR4_K * (iqs / QI8_1);
-
     float sumf_d = 0.0f;
     float sumf_m = 0.0f;
 
+#ifndef GGML_QKK_64
+
+    // iqs is in 0...15. bq8_offset = 2 * (iqs/4) -> bq8_offset = 0, 2, 4, 6
+    const int bq8_offset = QR4_K * (iqs / (QI8_1/2));
+
     const float    d = bq4_K->d;
     const float dmin = bq4_K->dmin;
 
-    const int v = *((int *) &bq4_K->qs[sizeof(int) * iqs]);
+    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
+    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
+    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
+    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
+
+    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * (iqs%4));
+    const int v1 = q4[0];
+    const int v2 = q4[4];
+
+    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
 
     for (int i = 0; i < QR4_K; ++i) {
-        const int isc = bq8_offset + i;
-
-        uint8_t sc, m;
-        get_scale_min_k4(isc, bq4_K->scales, sc, m);
 
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        const int ui = *((int*) &bq8i->qs[sizeof(int) * (iqs % QI8_1)]);
         const float d8i = bq8i->d;
+        const int * q8 = (const int *)bq8i->qs + (iqs%4);
+        const int ui1 = q8[0];
+        const int ui2 = q8[4];
 
-        const int vi = (v >> (4*i)) & 0x0F0F0F0F;
+        const int vi1 = (v1 >> (4*i)) & 0x0F0F0F0F;
+        const int vi2 = (v2 >> (4*i)) & 0x0F0F0F0F;
 
-        sumf_d += d8i * (__dp4a(vi,         ui, 0) * sc); // SIMD dot product
-        sumf_m += d8i * (__dp4a(0x01010101, ui, 0) * m);  // multiply constant part of q4_K with sum of q8_1 values
+        const int dot1 = __dp4a(vi2, ui2, __dp4a(vi1, ui1, 0)); // SIMD dot product
+        const int dot2 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+
+        sumf_d += d8i * (dot1 * sc[i]);
+        sumf_m += d8i * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
     }
 
     return d*sumf_d - dmin*sumf_m;
+
+#else
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    const uint16_t * a = (const uint16_t *)bq4_K->scales;
+    aux16[0] = a[0] & 0x0f0f;
+    aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+    const float dall = bq4_K->d[0];
+    const float dmin = bq4_K->d[1];
+
+    const float d8_1 = bq8_1[0].d;
+    const float d8_2 = bq8_1[1].d;
+
+    const int ui1 = *((const int *)bq8_1[0].qs + iqs);
+    const int ui2 = *((const int *)bq8_1[0].qs + iqs + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + iqs);
+    const int ui4 = *((const int *)bq8_1[1].qs + iqs + 4);
+
+    const int * q4 = (const int *)bq4_K->qs + iqs;
+    const int v1 = q4[0];
+    const int v2 = q4[4];
+
+    const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
+    const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
+    const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+    const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
+
+    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
+    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
+
+    return dall * sumf_d - dmin * sumf_m;
+
+#endif
+
 #else
     return 0.0f; // only to satisfy the compiler
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
@@ -1559,7 +1664,11 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
     const block_q5_K * bq5_K = (const block_q5_K *) vbq;
 
-    const int bq8_offset = QR5_K * (iqs / QI8_1);
+#ifndef GGML_QKK_64
+
+    const int bq8_offset = QR5_K * (iqs / (QI8_1/2));
+    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * (iqs%4));
+    const int * qh = (const int *)(bq5_K->qh + 4 * (iqs%4));
 
     float sumf_d = 0.0f;
     float sumf_m = 0.0f;
@@ -1567,31 +1676,87 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     const float    d = bq5_K->d;
     const float dmin = bq5_K->dmin;
 
-    const int vl = *((int *) &bq5_K->qs[sizeof(int) * iqs]);
+    const int vl1 = ql[0];
+    const int vl2 = ql[4];
 
-    const int vh = (*((int *) &bq5_K->qh[sizeof(int) * (iqs % (QI5_K/4))])) >> bq8_offset;
+    const int vh1 = qh[0] >> bq8_offset;
+    const int vh2 = qh[4] >> bq8_offset;
+
+    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
 
     for (int i = 0; i < QR5_K; ++i) {
-        const int isc = bq8_offset + i;
-
-        uint8_t sc, m;
-        get_scale_min_k4(isc, bq5_K->scales, sc, m);
 
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        const int ui = *((int*) &bq8i->qs[sizeof(int) * (iqs % QI8_1)]);
         const float d8i = bq8i->d;
+        const int * q8 = (const int *)bq8i->qs + (iqs%4);
+        const int ui1 = q8[0];
+        const int ui2 = q8[4];
 
-        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
+        const int vil1 = (vl1 >> (4*i)) & 0x0F0F0F0F;
+        const int vil2 = (vl2 >> (4*i)) & 0x0F0F0F0F;
 
-        const int vih = ((vh >> i) << 4) & 0x10101010;
+        const int vih1 = ((vh1 >> i) << 4) & 0x10101010;
+        const int vih2 = ((vh2 >> i) << 4) & 0x10101010;
 
-        const int vi = vil | vih;
+        const int vi1 = vil1 | vih1;
+        const int vi2 = vil2 | vih2;
+
+        const int dot1 = __dp4a(vi2, ui2, __dp4a(vi1, ui1, 0)); // SIMD dot product
+        const int dot2 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+
+        sumf_d += d8i * (dot1 * sc[i]);
+        sumf_m += d8i * (dot2 * m[i]);
 
-        sumf_d += d8i * (__dp4a(vi,         ui, 0) * sc); // SIMD dot product
-        sumf_m += d8i * (__dp4a(0x01010101, ui, 0) * m);  // multiply constant part of q5_K with sum of q8_1 values
     }
 
     return d*sumf_d - dmin*sumf_m;
+
+#else
+
+    const int8_t * s = bq5_K->scales;
+
+    const float d = bq5_K->d;
+
+    const float d8_1 = bq8_1[0].d;
+    const float d8_2 = bq8_1[1].d;
+
+    const int ui1 = *((const int *)bq8_1[0].qs + iqs);
+    const int ui2 = *((const int *)bq8_1[0].qs + iqs + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + iqs);
+    const int ui4 = *((const int *)bq8_1[1].qs + iqs + 4);
+
+    const int * ql = (const int *)bq5_K->qs + iqs;
+    const int vl1 = ql[0];
+    const int vl2 = ql[4];
+
+    const int step = 4 * iqs; // 0, 4, 8, 12
+    const int im = step/8; // = 0 for iqs = 0, 1, = 1 for iqs = 2, 3
+    const int in = step%8; // 0, 4, 0, 4
+    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
+
+    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
+    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
+    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
+    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
+
+    const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
+                       + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
+
+    return d * sumf_d;
+
+#endif
+
 #else
     return 0.0f; // only to satisfy the compiler
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
@@ -1745,11 +1910,15 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
     }
 }
 
-static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nchannels_x) {
+static __global__ void mul_mat_p021_f16_f32(
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
+
     const half * x = (const half *) vx;
 
     const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
     const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / (nchannels_y / nchannels_x);
 
     const int nrows_y = ncols_x;
     const int nrows_dst = nrows_x;
@@ -1765,7 +1934,7 @@ static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const
         }
 
         // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel*ncols_x + col_x;
+        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
         const float xi = __half2float(x[ix]);
 
         const int row_y = col_x;
@@ -1793,12 +1962,13 @@ static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const
 
 static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x) {
+    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
 
     const half * x = (const half *) vx;
 
     const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
     const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / channel_x_divisor;
 
     const int nrows_y = ncols_x;
     const int nrows_dst = nrows_x;
@@ -1815,7 +1985,7 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
             break;
         }
 
-        const int ix = channel*channel_stride_x + row_x*row_stride_x + col_x;
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
         const float xi = __half2float(x[ix]);
 
         const int row_y = col_x;
@@ -2027,10 +2197,10 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
-static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     const dim3 block_dims(WARP_SIZE, 1, 1);
-    rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
 }
 
 static void quantize_row_q8_1_cuda(const float * x, void * vy, const int ndata, const int k, cudaStream_t stream) {
@@ -2259,7 +2429,10 @@ static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float *
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI4_K, block_q4_K, vec_dot_q4_K_q8_1>
+    // Note: we use QI4_K/2 instead of QI4_K to make the dot product template require 4 groups of quants to be processed per
+    //       kernel call instead of 2. This results in a better perfmance because the cost of computing the k-quant scales
+    //       is better amortized.
+    mul_mat_vec_q<QK_K, QI4_K/2, block_q4_K, vec_dot_q4_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
@@ -2268,7 +2441,10 @@ static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float *
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI5_K, block_q5_K, vec_dot_q5_K_q8_1>
+    // Note: we use QI5_K/2 instead of QI5_K to make the dot product template require 4 groups of quants to be processed per
+    //       kernel call instead of 2. This results in a better perfmance because the cost of computing the k-quant scales
+    //       is better amortized.
+    mul_mat_vec_q<QK_K, QI5_K/2, block_q5_K, vec_dot_q5_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
@@ -2324,20 +2500,23 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     }
 }
 
-static void ggml_mul_mat_p021_f16_f32_cuda(const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x, cudaStream_t stream) {
-    const dim3 block_nums(1, nrows_x, nchannels_x);
+static void ggml_mul_mat_p021_f16_f32_cuda(
+    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
+    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
+
+    const dim3 block_nums(1, nrows_x, nchannels_y);
     const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x);
+    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
 }
 
 static void ggml_mul_mat_vec_nc_f16_f32_cuda(
     const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int channel_stride_x, cudaStream_t stream) {
+    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
 
-    const dim3 block_nums(1, nrows_x, nchannels_x);
+    const dim3 block_nums(1, nrows_x, nchannels_y);
     const dim3 block_dims(WARP_SIZE, 1, 1);
     mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x);
+        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
 }
 
 static void ggml_cpy_f32_f32_cuda(
@@ -2423,20 +2602,53 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
-
+#ifdef DEBUG_CUDA_MALLOC
+    int nnz = 0;
+    size_t max_size = 0, tot_size = 0;
+#endif
+    size_t best_diff = 1ull << 36;
+    int ibest = -1;
     for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
         cuda_buffer& b = g_cuda_buffer_pool[id][i];
-        if (b.size >= size && b.ptr != nullptr) {
-            void * ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
-            return ptr;
+        if (b.ptr != nullptr) {
+#ifdef DEBUG_CUDA_MALLOC
+            ++nnz;
+            tot_size += b.size;
+            if (b.size > max_size) max_size = b.size;
+#endif
+            if (b.size >= size) {
+                size_t diff = b.size - size;
+                if (diff < best_diff) {
+                    best_diff = diff;
+                    ibest = i;
+                    if (!best_diff) {
+                        void * ptr = b.ptr;
+                        *actual_size = b.size;
+                        b.ptr = nullptr;
+                        b.size = 0;
+                        return ptr;
+                    }
+                }
+            }
         }
     }
+    if (ibest >= 0) {
+        cuda_buffer& b = g_cuda_buffer_pool[id][ibest];
+        void * ptr = b.ptr;
+        *actual_size = b.size;
+        b.ptr = nullptr;
+        b.size = 0;
+        return ptr;
+    }
+#ifdef DEBUG_CUDA_MALLOC
+    fprintf(stderr, "%s: %d buffers, max_size = %u MB, tot_size = %u MB, requested %u MB\n", __func__, nnz,
+            (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024));
+#endif
     void * ptr;
-    CUDA_CHECK(cudaMalloc((void **) &ptr, size));
-    *actual_size = size;
+    size_t look_ahead_size = (size_t) (1.05 * size);
+    look_ahead_size = 256 * ((look_ahead_size + 255)/256);
+    CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
+    *actual_size = look_ahead_size;
     return ptr;
 }
 
@@ -2464,7 +2676,9 @@ static size_t g_scratch_offset = 0;
 
 static int g_device_count = -1;
 static int g_main_device = 0;
+#ifndef GGML_CUDA_FORCE_DMMV
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
+#endif
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
@@ -2487,7 +2701,9 @@ void ggml_init_cublas() {
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
 
+#ifndef GGML_CUDA_FORCE_DMMV
             g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
+#endif
         }
         for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
@@ -2512,6 +2728,9 @@ void ggml_init_cublas() {
 }
 
 void ggml_cuda_set_tensor_split(const float * tensor_split) {
+    if (tensor_split == nullptr) {
+        return;
+    }
     bool all_zero = true;
     for (int i = 0; i < g_device_count; ++i) {
         if (tensor_split[i] != 0.0f) {
@@ -2652,6 +2871,7 @@ inline void ggml_cuda_op_mul(
     (void) dst;
     (void) src0_ddq_i;
     (void) i02;
+    (void) i1;
 }
 
 inline void ggml_cuda_op_gelu(
@@ -2731,8 +2951,11 @@ inline void ggml_cuda_op_rms_norm(
     const int64_t ne00 = src0->ne[0];
     const int64_t i01_diff = i01_high - i01_low;
 
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
     // compute
-    rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
+    rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, eps, cudaStream_main);
 
     (void) src1;
     (void) dst;
@@ -2779,8 +3002,8 @@ inline void ggml_cuda_op_mul_mat_vec(
 #endif
 
     if (use_mul_mat_vec_q) {
-        int64_t padded_row_size = ne00 + MATRIX_ROW_PADDING - 1;
-        padded_row_size -= padded_row_size % MATRIX_ROW_PADDING;
+        const int64_t padded_row_size = ne00 % MATRIX_ROW_PADDING == 0 ?
+            ne00 : ne00 - ne00 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
         size_t as;
         void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*sizeof(block_q8_1)/QK8_1, &as);
         quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne00, padded_row_size, cudaStream_main);
@@ -2947,13 +3170,18 @@ inline void ggml_cuda_op_rope(
     const int64_t ne00 = src0->ne[0];
     const int64_t i01_diff = i01_high - i01_low;
 
-    const int n_past = ((int32_t *) src1->data)[0];
-    const int n_dims = ((int32_t *) src1->data)[1];
-    const int mode   = ((int32_t *) src1->data)[2];
-    const int n_ctx  = ((int32_t *) src1->data)[3];
+    const int n_past = ((int32_t *) dst->op_params)[0];
+    const int n_dims = ((int32_t *) dst->op_params)[1];
+    const int mode   = ((int32_t *) dst->op_params)[2];
+    const int n_ctx  = ((int32_t *) dst->op_params)[3];
+    // RoPE alteration for extended context
 
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
-    const float p = ((mode & 1) == 0 ? n_past + i02 : i02);
+    float freq_base, freq_scale;
+    memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float p = (((mode & 1) == 0 ? n_past + i02 : i02)) * freq_scale;
 
     bool is_glm = mode & 4;
 
@@ -2966,6 +3194,7 @@ inline void ggml_cuda_op_rope(
         rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p, theta_scale, cudaStream_main);
     }
 
+    (void) src1;
     (void) dst;
     (void) src0_ddq_i;
     (void) src1_ddf_i;
@@ -2984,11 +3213,12 @@ inline void ggml_cuda_op_diag_mask_inf(
     const int64_t ne01 = src0->ne[1];
     const int64_t i01_diff = i01_high - i01_low;
 
-    const int n_past = ((int32_t *) src1->data)[0];
+    const int n_past = ((int32_t *) dst->op_params)[0];
 
     // compute
     diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main);
 
+    (void) src1;
     (void) dst;
     (void) src0_ddq_i;
     (void) src1_ddf_i;
@@ -3056,6 +3286,9 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     const int64_t ne11 = use_src1 ? src1->ne[1] : 1;
     const int64_t ne12 = use_src1 ? src1->ne[2] : 1;
     const int64_t ne13 = use_src1 ? src1->ne[3] : 1;
+    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
+
+    GGML_ASSERT(ne03 == ne13);
 
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
@@ -3067,12 +3300,19 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
 
     // strides for iteration over dims 3 and 2
-    const int64_t num_iters = flatten_rows ? 1 : ne02 * ne03;
-    const int64_t stride_mod = flatten_rows ? ne02 * ne03 : 1;
+    const int64_t num_iters_0 = ne02 >= ne12 ? ne02*ne03 : ne12*ne13;
+    const int64_t num_iters = flatten_rows ? 1 : num_iters_0;
+    const int64_t stride_mod = flatten_rows ? num_iters_0 : 1;
     const int64_t src0_stride = ne00 * ne01 * stride_mod;
     const int64_t src1_stride = ne10 * ne11 * stride_mod;
     const int64_t dst_stride = ne0 * ne1 * stride_mod;
 
+    const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;
+    const int64_t i03_max = flatten_rows ? 1 : ne03;
+    const int64_t i02_max = flatten_rows ? 1 : (ne02 >= ne12 ? ne02 : ne12);
+    const int64_t i02_divisor = ne02 >= ne12 ? 1 : ne12 / ne02;
+    GGML_ASSERT(!(flatten_rows && ne02 < ne12));
+
     const size_t src0_ts = ggml_type_size(src0->type);
     const size_t src0_bs = ggml_blck_size(src0->type);
 
@@ -3089,6 +3329,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
         dst->op == GGML_OP_SCALE || dst->op == GGML_OP_DIAG_MASK_INF || dst->op == GGML_OP_ROPE);
 
     const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+    GGML_ASSERT(!(split && ne02 < ne12));
 
     const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
 
@@ -3125,7 +3366,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
             row_high = id == g_device_count - 1 ? nrows0 : nrows0*g_tensor_split[id + 1];
         } else {
             row_low = 0;
-            row_high = nrows0;
+            row_high = nrows0*i02_divisor;
         }
         if (row_low == row_high) {
             continue;
@@ -3173,16 +3414,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
             dst_ddf[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_asf[id]);
         }
 
-        const int64_t i03_max = flatten_rows ? 1 : ne03;
-        const int64_t i02_max = flatten_rows ? 1 : ne02;
-        const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;
-
         for (int64_t i03 = 0; i03 < i03_max; i03++) {
             const int64_t i13 = i03 % ne13;
             for (int64_t i02 = 0; i02 < i02_max; i02++) {
                 const int64_t i12 = i02 % ne12;
 
-                const int64_t i0 = i03*ne02 + i02;
+                const int64_t i0 = i03*i02_max + i02;
 
                 // i0 values that contain the lower/upper rows for a split tensor when using multiple GPUs
                 const int64_t i0_offset_low = row_low/rows_per_iter;
@@ -3216,10 +3453,10 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                 const int64_t i11 = i13*ne12 + i12;
 
                 // for split tensors the data begins at i0 == i0_offset_low
-                char  * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs;
-                float * src0_ddf_i = src0_ddf[id] + (i0 - i0_offset_low)*src0_stride;
+                char  * src0_ddq_i = src0_ddq[id] + (i0/i02_divisor - i0_offset_low)*src0_stride*src0_ts/src0_bs;
+                float * src0_ddf_i = src0_ddf[id] + (i0/i02_divisor - i0_offset_low)*src0_stride;
                 float * src1_ddf_i = src1_ddf[id] + i11*src1_stride;
-                float * dst_ddf_i  =  dst_ddf[id] + (i0 - i0_offset_low)*dst_stride;
+                float * dst_ddf_i  =  dst_ddf[id] + (i0             - i0_offset_low)*dst_stride;
 
                 // for split tensors the data pointer needs to be rounded down
                 // to the bin edge for i03, i02 bins beyond the first
@@ -3258,11 +3495,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     }
                 }
 
-                if (!src0_on_device || !src0_is_contiguous) {
+                if ((!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
                     if (src0_is_f32) {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
+                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
                     } else {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
+                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
                     }
                 }
 
@@ -3416,6 +3653,8 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     CUDA_CHECK(cudaSetDevice(g_main_device));
     cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
 
@@ -3428,7 +3667,7 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, cudaStream_main);
+    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, cudaStream_main);
 }
 
 void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
@@ -3442,6 +3681,8 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     const int64_t nb01 = src0->nb[1];
     const int64_t nb02 = src0->nb[2];
 
@@ -3460,7 +3701,7 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int row_stride_x = nb01 / sizeof(half);
     const int channel_stride_x = nb02 / sizeof(half);
 
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, channel_stride_x, cudaStream_main);
+    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, cudaStream_main);
 }
 
 void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3537,6 +3778,11 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
     (void) dst;
 }
 
+void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_cpy(src0, dst, nullptr);
+    (void) src1;
+}
+
 void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
     ggml_cuda_op(src0, src1, dst, ggml_cuda_op_diag_mask_inf, true, true);
@@ -3596,7 +3842,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         size_t size = ggml_nbytes_split(tensor, nrows_split);
         const size_t original_size = size;
 
-        // pad last row to a multiple of 256 elements to avoid out-of-bounds memory accesses
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
         if (ne0 % MATRIX_ROW_PADDING != 0) {
             size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
                 * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
@@ -3612,7 +3858,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         }
 
 
-        CUDA_CHECK(cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice));
+        CUDA_CHECK(cudaMemcpy(buf, buf_host, original_size, cudaMemcpyHostToDevice));
 
         extra->data_device[id] = buf;
 
@@ -3670,7 +3916,7 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
     // recursively assign CUDA buffers until a compute tensor is found
     if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
         const ggml_op src0_op = tensor->src[0]->op;
-        if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW) {
+        if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) {
             ggml_cuda_assign_buffers_impl(tensor->src[0], scratch, force_inplace);
         }
     }
@@ -3692,7 +3938,7 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
-            memcpy(&offset, tensor->src[2]->data, sizeof(size_t));
+            memcpy(&offset, tensor->op_params, sizeof(size_t));
         }
         extra = ggml_cuda_alloc_temp_tensor_extra();
         extra->data_device[g_main_device] = src0_ddc + offset;
@@ -3776,6 +4022,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
 
     switch (tensor->op) {
+        case GGML_OP_DUP:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cuda_dup;
+            break;
         case GGML_OP_ADD:
             if (!any_on_device) {
                 return false;
@@ -3788,18 +4040,23 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             }
             func = ggml_cuda_mul;
             break;
-        case GGML_OP_GELU:
-            if (!any_on_device) {
-                return false;
-            }
-            func = ggml_cuda_gelu;
-            break;
-        case GGML_OP_SILU:
-            if (!any_on_device) {
-                return false;
-            }
-            func = ggml_cuda_silu;
-            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_GELU:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cuda_gelu;
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cuda_silu;
+                    break;
+                default:
+                    return false;
+            } break;
         case GGML_OP_NORM:
             if (!any_on_device) {
                 return false;
@@ -3830,6 +4087,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             }
             func = ggml_cuda_cpy;
             break;
+        case GGML_OP_CONT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cuda_dup;
+            break;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:

ggml-metal.h

@@ -61,6 +61,13 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 // get data from the device into host memory
 void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
 
+// try to find operations that can be run concurrently in the graph
+// you should run it again if the topology of your graph changes
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+// if the graph has been optimized for concurrently dispatch
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

ggml-metal.m

@@ -36,12 +36,16 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
+    int concur_list[GGML_MAX_NODES];
+    int concur_list_len;
+
     // custom kernels
 #define GGML_METAL_DECL_KERNEL(name) \
     id<MTLFunction>             function_##name; \
     id<MTLComputePipelineState> pipeline_##name
 
     GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
@@ -97,6 +101,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
+    ctx->concur_list_len = 0;
 
     // determine if we can use MPS
     if (MPSSupportsMTLDevice(ctx->device)) {
@@ -157,6 +162,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
 
         GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(add_row);
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
@@ -215,6 +221,13 @@ void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
     ctx->n_cb = n_cb;
 }
 
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+    if (ctx->concur_list_len) {
+        return true;
+    }
+    return false;
+}
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -353,11 +366,98 @@ void ggml_metal_get_tensor(
     memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
 }
 
+void ggml_metal_graph_find_concurrency(
+        struct ggml_metal_context * ctx,
+        struct ggml_cgraph * gf) {
+    int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
+    int nodes_unused[GGML_MAX_NODES];
+
+    for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;}
+    for (int i = 0; i < gf->n_nodes; i++) {nodes_unused[i] = 1;}
+    ctx->concur_list_len = 0;
+
+    int n_left = gf->n_nodes;
+    int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0;  // at ctx->concur_list, the last layer (level) ends at level_pos
+
+    while (n_left > 0) {
+        // number of nodes at a layer (that can be issued concurrently)
+        int concurrency = 0;
+        for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
+            if (nodes_unused[i]) {
+                // if the requirements for gf->nodes[i] are satisfied
+                int exe_flag=1;
+                // scan all srcs
+                for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
+                    struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
+                    if (src_cur) {
+                        // if is leaf nodes it's satisfied.
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;}
+
+                        // otherwise this src should be the output from previous nodes.
+                        int is_found = 0;
+                        // scan 2*search_depth back because we inserted barrier.
+                        for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                            if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;}
+                        }
+                        if (is_found == 0) {exe_flag = 0; break;}
+                    }
+                }
+                if (exe_flag) {
+                    // check if nodes[i]'s data will be overwritten by a node before nodes[i].
+                    // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
+                    int64_t data_start = (int64_t) gf->nodes[i]->data;
+                    int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    for (int j = n_start; j < i; j++) {
+                        if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
+                                            && gf->nodes[j]->op != GGML_OP_VIEW \
+                                            && gf->nodes[j]->op != GGML_OP_TRANSPOSE \
+                                            && gf->nodes[j]->op != GGML_OP_PERMUTE) {
+                            if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
+                                ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
+                                continue;
+                            } else {
+                                exe_flag = 0;
+                            }
+                        }
+                    }
+                }
+                if (exe_flag) {
+                    ctx->concur_list[level_pos + concurrency] = i;
+                    nodes_unused[i] = 0;
+                    concurrency++;
+                    ctx->concur_list_len++;
+                }
+            }
+        }
+        n_left -= concurrency;
+        // adding a barrier different layer
+        ctx->concur_list[level_pos + concurrency] = -1;
+        ctx->concur_list_len++;
+        // jump all sorted nodes at nodes_bak
+        while (!nodes_unused[n_start]) {n_start++;}
+        level_pos += concurrency + 1;
+    }
+
+    if (ctx->concur_list_len > GGML_MAX_NODES) {
+        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+    }
+}
+
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
     metal_printf("%s: evaluating graph\n", __func__);
 
+    // if there is ctx->concur_list, dispatch concurrently
+    // else fallback to serial dispatch
+    MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
+
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES;
+
+    const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
+    edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
+
     // create multiple command buffers and enqueue them
     // then, we encode the graph into the command buffers in parallel
 
@@ -376,7 +476,7 @@ void ggml_metal_graph_compute(
     dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
 
     for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
+        const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
 
         dispatch_async(queue, ^{
             size_t offs_src0 = 0;
@@ -387,10 +487,21 @@ void ggml_metal_graph_compute(
 
             id<MTLComputeCommandEncoder> encoder = nil;
 
-            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+
+            for (int ind = node_start; ind < node_end; ++ind) {
+                const int i = has_concur ? ctx->concur_list[ind] : ind;
+
+                if (i == -1) {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                        continue;
+                    }
+                    [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
+                    continue;
+                }
 
-            for (int i = node_start; i < node_end; ++i) {
                 metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
@@ -461,13 +572,19 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ADD:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_add];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
 
                             const int64_t n = ggml_nelements(dst);
 
@@ -476,7 +593,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             if (ggml_nelements(src1) == ne10) {
@@ -497,7 +614,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_SCALE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float scale = *(const float *) src1->data;
@@ -511,52 +628,60 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
-                    case GGML_OP_SILU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
+                    case GGML_OP_UNARY:
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
 
-                            [encoder setComputePipelineState:ctx->pipeline_silu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                            const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst);
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_RELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_relu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_GELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_gelu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ASSERT(false);
+                                }
                         } break;
-                    case GGML_OP_RELU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_relu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_GELU:
-                    {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_gelu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                    } break;
                     case GGML_OP_SOFT_MAX:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -574,10 +699,10 @@ void ggml_metal_graph_compute(
                     case GGML_OP_DIAG_MASK_INF:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *)(dst->op_params))[0];
 
                             [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -637,7 +762,7 @@ void ggml_metal_graph_compute(
                                 }
                             } else {
                                 if (encoder == nil) {
-                                    encoder = [command_buffer computeCommandEncoder];
+                                    encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                                 }
 
                                 int nth0 = 32;
@@ -676,8 +801,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
@@ -685,8 +810,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
@@ -694,8 +819,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -703,8 +828,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
@@ -712,8 +837,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
                                         } break;
                                     default:
@@ -739,16 +864,22 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
-                                else if (src0t == GGML_TYPE_Q2_K ||
-                                         src0t == GGML_TYPE_Q3_K ||
-                                         src0t == GGML_TYPE_Q4_K ||
-                                         src0t == GGML_TYPE_Q5_K ||
-                                         src0t == GGML_TYPE_Q6_K) {
-                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                else if (src0t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+                                }
+                                else if (src0t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -758,7 +889,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_GET_ROWS:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             switch (src0->type) {
@@ -787,12 +918,13 @@ void ggml_metal_graph_compute(
                     case GGML_OP_RMS_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const float eps = 1e-6f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 256;
+                            const int nth = 512;
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -800,7 +932,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
                             [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -809,7 +941,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float eps = 1e-5f;
@@ -831,14 +963,15 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ALIBI:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
-                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
-                            const int   n_head   = ((int32_t *) src1->data)[1];
-                            const float max_bias = ((float *)   src1->data)[2];
+                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            const int n_head = ((int32_t *) dst->op_params)[1];
+                            float max_bias;
+                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
                             if (__builtin_popcount(n_head) != 1) {
                                 GGML_ASSERT(false && "only power-of-two n_head implemented");
@@ -873,18 +1006,17 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ROPE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_dims = ((int32_t *) src1->data)[1];
-                            const int mode   = ((int32_t *) src1->data)[2];
-
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *) dst->op_params)[0];
+                            const int n_dims = ((int32_t *) dst->op_params)[1];
+                            const int mode   = ((int32_t *) dst->op_params)[2];
 
                             float freq_base;
                             float freq_scale;
-                            memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
+                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -913,10 +1045,12 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
+                    case GGML_OP_DUP:
                     case GGML_OP_CPY:
+                    case GGML_OP_CONT:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -963,8 +1097,10 @@ void ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     default:
-                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                        GGML_ASSERT(false);
+                        {
+                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_ASSERT(false);
+                        }
                 }
             }
 

ggml.h

@@ -199,6 +199,7 @@
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          48
+#define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
 
@@ -329,16 +330,6 @@ extern "C" {
         GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
-        GGML_OP_ABS,
-        GGML_OP_SGN,
-        GGML_OP_NEG,
-        GGML_OP_STEP,
-        GGML_OP_TANH,
-        GGML_OP_ELU,
-        GGML_OP_RELU,
-        GGML_OP_GELU,
-        GGML_OP_GELU_QUICK,
-        GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
@@ -377,6 +368,8 @@ extern "C" {
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
 
+        GGML_OP_UNARY,
+
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
@@ -390,6 +383,18 @@ extern "C" {
         GGML_OP_COUNT,
     };
 
+    enum ggml_unary_op {
+        GGML_UNARY_OP_ABS,
+        GGML_UNARY_OP_SGN,
+        GGML_UNARY_OP_NEG,
+        GGML_UNARY_OP_STEP,
+        GGML_UNARY_OP_TANH,
+        GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_SILU,
+    };
 
     // ggml object
     struct ggml_object {
@@ -418,6 +423,9 @@ extern "C" {
         // compute data
         enum ggml_op op;
 
+        // op params - allocated as int32_t for alignment
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(uint32_t)];
+
         bool is_param;
 
         struct ggml_tensor * grad;
@@ -434,7 +442,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -455,6 +463,11 @@ extern "C" {
         void * abort_callback_data;
     };
 
+    // next prime after GGML_MAX_NODES
+    // #define GGML_GRAPH_HASHTABLE_SIZE 4099
+    // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
+    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -464,6 +477,8 @@ extern "C" {
         struct ggml_tensor * grads[GGML_MAX_NODES];
         struct ggml_tensor * leafs[GGML_MAX_NODES];
 
+        void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+
         // performance
         int     perf_runs;
         int64_t perf_cycles;
@@ -531,6 +546,7 @@ extern "C" {
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -554,6 +570,7 @@ extern "C" {
     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);
 
     GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
@@ -613,9 +630,11 @@ extern "C" {
     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 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_API struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...);
+    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_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
     // operations on tensors with backpropagation
@@ -625,6 +644,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_dup_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_add(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -849,14 +873,17 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     // a - x
     // b - dy
+    // TODO: update with configurable eps
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -948,11 +975,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // a -> b, in-place, return view(b)
+    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, in-place
+    GGML_API struct ggml_tensor * ggml_cont_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1128,9 +1166,9 @@ extern "C" {
             int                   n_past,
             int                   n_dims,
             int                   mode,
+            int                   n_ctx,
             float                 freq_base,
-            float                 freq_scale,
-            int                   n_ctx);
+            float                 freq_scale);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
@@ -1139,7 +1177,8 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_dims,
-            int                   mode);
+            int                   mode,
+            int                   n_ctx);
 
     // alibi position embedding
     // in-place, returns view(a)
@@ -1263,6 +1302,16 @@ extern "C" {
     typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
     typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
 
+    GGML_API struct ggml_tensor * ggml_unary(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_unary_op op);
+
+    GGML_API struct ggml_tensor * ggml_unary_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op op);
+
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,

grammars/arithmetic.gbnf

@@ -0,0 +1,6 @@
+root  ::= (expr "=" ws term "\n")+
+expr  ::= term ([-+*/] term)*
+term  ::= ident | num | "(" ws expr ")" ws
+ident ::= [a-z] [a-z0-9_]* ws
+num   ::= [0-9]+ ws
+ws    ::= [ \t\n]*

grammars/chess.gbnf

@@ -0,0 +1,13 @@
+# Specifies chess moves as a list in algebraic notation, using PGN conventions
+
+# Force first move to "1. ", then any 1-2 digit number after, relying on model to follow the pattern
+root    ::= "1. " move " " move "\n" ([1-9] [0-9]? ". " move " " move "\n")+
+move    ::= (pawn | nonpawn | castle) [+#]?
+
+# piece type, optional file/rank, optional capture, dest file & rank
+nonpawn ::= [NBKQR] [a-h]? [1-8]? "x"? [a-h] [1-8]
+
+# optional file & capture, dest file & rank, optional promotion
+pawn    ::= ([a-h] "x")? [a-h] [1-8] ("=" [NBKQR])?
+
+castle  ::= "O-O" "-O"?

grammars/japanese.gbnf

@@ -0,0 +1,7 @@
+# A probably incorrect grammar for Japanese
+root        ::= jp-char+ ([ \t\n] jp-char+)*
+jp-char     ::= hiragana | katakana | punctuation | cjk
+hiragana    ::= [ぁ-ゟ]
+katakana    ::= [ァ-ヿ]
+punctuation ::= [、-〾]
+cjk         ::= [一-鿿]

grammars/json.gbnf

@@ -0,0 +1,29 @@
+# Grammar for subset of JSON - doesn't support full string or number syntax
+
+root  ::= object
+value ::= object | array | string | number | boolean | "null"
+
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}"
+
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]"
+
+string  ::=
+  "\"" (
+    [^"\\] |
+    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
+  )* "\"" ws
+
+# Only plain integers currently
+number  ::= "-"? [0-9]+ ws
+boolean ::= ("true" | "false") ws
+
+# Optional space: by convention, applied in this grammar after literal chars when allowed
+ws ::= ([ \t\n] ws)?

grammars/list.gbnf

@@ -0,0 +1,4 @@
+root ::= item+
+
+# Excludes various line break characters
+item ::= "- " [^\r\n\x0b\x0c\x85\u2028\u2029]+ "\n"

k_quants.c

@@ -1666,6 +1666,62 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t ud, um;
+    const uint8_t * restrict db = (const uint8_t *)&ud;
+    const uint8_t * restrict mb = (const uint8_t *)&um;
+
+    float summs = 0;
+
+    // TODO: optimize this
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+        ud = (sc[0] >> 0) & 0x0f0f0f0f;
+        um = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
+        summs += dmin * smin;
+
+        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
+        const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+        const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+        const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+        const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
 #else
 
     float sumf = 0;
@@ -2295,6 +2351,93 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint64_t aux64;
+
+    uint16_t aux16[2];
+    const int8_t * aux8 = (const int8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
+        const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
+        const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
+        const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
+
+        memcpy(&aux64, x[i].hmask, 8);
+
+        __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
+        __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
+        __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
+        __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
+        q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
+        q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
+        q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
+        q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
+
+        // load low 2 bits
+        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
+
+        // prepare low and high bits
+        const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
+        const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
+        const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
+        const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3);
+
+        // load Q8 quants
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16,
+        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+        // and 2 if the high bit was set)
+        const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
+
+        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);
+
+        // multiply with scales
+        p16_0 = _mm_madd_epi16(scale_0, p16_0);
+        p16_1 = _mm_madd_epi16(scale_1, p16_1);
+        p16_2 = _mm_madd_epi16(scale_2, p16_2);
+        p16_3 = _mm_madd_epi16(scale_3, p16_3);
+
+        p16_0 = _mm_add_epi32(p16_0, p16_2);
+        p16_1 = _mm_add_epi32(p16_1, p16_3);
+        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+
+        // multiply with block scale and accumulate
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2781,6 +2924,60 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) - summs;
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    uint16_t aux16[2];
+    const uint8_t * scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
+        const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
+        const __m256 vd = _mm256_set1_ps(d);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
+        const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
+        const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
+        const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
+        const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
+        const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
+        const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+
+        const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
+        const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+
+    }
+
+    *s = hsum_float_8(acc) - summs;
+
 #else
 
     uint8_t aux8[QK_K];
@@ -3295,10 +3492,66 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mone  = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
+
+        const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
+        const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
+        const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
+        const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
+
+        int64_t aux64;
+        memcpy(&aux64, x[i].qh, 8);
+        const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
+        const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
+
+        const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
+        const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
+        const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
+        const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
+
+        const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
+        const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
+        const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
+        const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
+        const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
+
+        const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
+        const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
-
-    uint8_t aux8[QK_K];
+    int8_t aux8[QK_K];
     int16_t aux16[16];
     float   sums [8];
     memset(sums, 0, 8*sizeof(float));
@@ -3308,7 +3561,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const uint8_t * restrict hm = x[i].qh;
         const  int8_t * restrict q8 = y[i].qs;
-        uint8_t * restrict a = aux8;
+        int8_t * restrict a = aux8;
         for (int l = 0; l < 32; ++l) {
             a[l+ 0] = q4[l] & 0xF;
             a[l+32] = q4[l]  >> 4;
@@ -3858,6 +4111,77 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
+        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
+        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
+        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
+        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
+
+        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
+
+        const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
+        const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
+        const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
+        const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
+
+        const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
+        const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
+        const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
+        const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
+        __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
+        __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
+        __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        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_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_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);
+
+        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));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];

llama.h

@@ -83,12 +83,15 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
    struct llama_context_params {
-        uint32_t seed;                         // RNG seed, -1 for random
-        int32_t  n_ctx;                        // text context
-        int32_t  n_batch;                      // prompt processing batch size
-        int32_t  n_gpu_layers;                 // number of layers to store in VRAM
-        int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
-        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        uint32_t seed;         // RNG seed, -1 for random
+        int32_t  n_ctx;        // text context
+        int32_t  n_batch;      // prompt processing batch size
+        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
+        float    rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
+        int32_t  n_gpu_layers; // number of layers to store in VRAM
+        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
         float    rope_freq_base;  // RoPE base frequency
@@ -139,6 +142,40 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
     } llama_model_quantize_params;
 
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
     // performance timing information
     struct llama_timings {
         double t_start_ms;
@@ -153,6 +190,8 @@ extern "C" {
         int32_t n_eval;
     };
 
+    LLAMA_API int llama_max_devices();
+
     LLAMA_API struct llama_context_params llama_context_default_params();
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
@@ -329,6 +368,15 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos();  // end-of-sentence
     LLAMA_API llama_token llama_token_nl();   // next-line
 
+    // Grammar
+    //
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
     // Sampling functions
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
@@ -341,13 +389,11 @@ extern "C" {
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
     /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
     /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    /// @params smooth_factor Smooth factor between guidance logits and original logits. 1.0f means only use guidance logits. 0.0f means only original logits.
     LLAMA_API void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale,
-                             float   smooth_factor);
+                             float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
@@ -365,6 +411,9 @@ extern "C" {
     LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
     LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
 
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -386,6 +435,9 @@ extern "C" {
     /// @details Randomly selects a token from the candidates based on their probabilities.
     LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);

scripts/build-info.sh

@@ -16,7 +16,8 @@ fi
 echo "#ifndef BUILD_INFO_H"
 echo "#define BUILD_INFO_H"
 echo ""
-echo "#define BUILD_NUMBER $BUILD_NUMBER"
-echo "#define BUILD_COMMIT \"$BUILD_COMMIT\""
+echo "#define BUILD_NUMBER $BUILD_NUMBER" | tr -d '\n'
+echo ""
+echo "#define BUILD_COMMIT \"$BUILD_COMMIT\"" | tr -d '\n'
 echo ""
 echo "#endif // BUILD_INFO_H"

tests/CMakeLists.txt

@@ -1,6 +1,7 @@
 function(llama_add_test source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_executable(${TEST_TARGET} ${source})
+    install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE llama)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()

tests/test-grad0.c

@@ -64,7 +64,7 @@ void get_random_dims(int64_t * dims, int ndims) {
     }
 }
 
-struct ggml_tensor * get_random_tensor(
+struct ggml_tensor * get_random_tensor_f32(
         struct ggml_context * ctx0,
         int ndims,
         int64_t ne[],
@@ -112,7 +112,55 @@ struct ggml_tensor * get_random_tensor(
     return result;
 }
 
-struct ggml_tensor * get_random_tensor_int(
+struct ggml_tensor * get_random_tensor_f16(
+        struct ggml_context * ctx0,
+        int ndims,
+        int64_t ne[],
+        float fmin,
+        float fmax) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F16, ndims, ne);
+
+    switch (ndims) {
+        case 1:
+            for (int i0 = 0; i0 < ne[0]; i0++) {
+                ((ggml_fp16_t *)result->data)[i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < ne[1]; i1++) {
+                for (int i0 = 0; i0 < ne[0]; i0++) {
+                    ((ggml_fp16_t *)result->data)[i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < ne[2]; i2++) {
+                for (int i1 = 0; i1 < ne[1]; i1++) {
+                    for (int i0 = 0; i0 < ne[0]; i0++) {
+                        ((ggml_fp16_t *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < ne[3]; i3++) {
+                for (int i2 = 0; i2 < ne[2]; i2++) {
+                    for (int i1 = 0; i1 < ne[1]; i1++) {
+                        for (int i0 = 0; i0 < ne[0]; i0++) {
+                            ((ggml_fp16_t *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            assert(false);
+    };
+
+    return result;
+}
+
+struct ggml_tensor * get_random_tensor_i32(
         struct ggml_context * ctx0,
         int ndims,
         int64_t ne[],
@@ -160,23 +208,6 @@ struct ggml_tensor * get_random_tensor_int(
     return result;
 }
 
-float get_element(const struct ggml_tensor * t, int idx) {
-    if (t->type == GGML_TYPE_F32) {
-        return ((float *)t->data)[idx];
-    }
-
-    if (t->type == GGML_TYPE_I32) {
-        return ((int32_t *)t->data)[idx];
-    }
-
-    assert(false);
-    return INFINITY;
-}
-
-void set_element(struct ggml_tensor * t, int idx, float value) {
-    ((float *)t->data)[idx] = value;
-}
-
 void print_elements(const char* label, const struct ggml_tensor * t) {
     if (!t) {
         printf("%s: %s = null\n", __func__, label);
@@ -186,7 +217,7 @@ void print_elements(const char* label, const struct ggml_tensor * t) {
     printf("%s: %s = [", __func__, label);
     for (int k = 0; k < nelements; ++k) {
         if (k > 0) { printf(", "); }
-        printf("%.5f", get_element(t, k));
+        printf("%.5f", ggml_get_f32_1d(t, k));
     }
     printf("] shape: [");
     for (int k = 0; k < t->n_dims; ++k) {
@@ -237,23 +268,23 @@ bool check_gradient(
         const int nelements = ggml_nelements(x[i]);
         for (int k = 0; k < nelements; ++k) {
             // compute gradient using finite differences
-            const float x0 = get_element(x[i], k);
+            const float x0 = ggml_get_f32_1d(x[i], k);
             const float xm = x0 - eps;
             const float xp = x0 + eps;
-            set_element(x[i], k, xp);
+            ggml_set_f32_1d(x[i], k, xp);
 
             ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
             const float f0 = ggml_get_f32_1d(f, 0);
 
-            set_element(x[i], k, xm);
+            ggml_set_f32_1d(x[i], k, xm);
 
             ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
             const float f1 = ggml_get_f32_1d(f, 0);
             const float g0 = (f0 - f1)/(2.0f*eps);
 
-            set_element(x[i], k, x0);
+            ggml_set_f32_1d(x[i], k, x0);
 
             // compute gradient using backward graph
             ggml_graph_reset  (&gf);
@@ -261,7 +292,7 @@ bool check_gradient(
 
             ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
 
-            const float g1 = get_element(x[i]->grad, k);
+            const float g1 = ggml_get_f32_1d(x[i]->grad, k);
 
             const float error_abs = fabsf(g0 - g1);
             const float error_rel = g0 != 0 ? fabsf(g0 - g1)/fabsf(g0) : 0;
@@ -392,19 +423,35 @@ int main(int argc, const char ** argv) {
 
         struct ggml_tensor * x[MAX_NARGS];
 
-        // add
+        // add f32
         {
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
 
-                check_gradient("add", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f);
+                check_gradient("add f32", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f);
+            }
+        }
+
+        // add f16
+        {
+            const int nargs = 2;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
+
+                check_gradient("add f16", ctx0, x, f, ndims, nargs, 1e-1f, 2e-1f, 2e-1f);
             }
         }
 
@@ -414,7 +461,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -430,7 +477,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -446,7 +493,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 0.5f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 0.5f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -462,7 +509,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -478,7 +525,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -494,7 +541,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -510,7 +557,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -527,7 +574,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -537,6 +584,40 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // mean, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_mean(ctx0, x[0]));
+
+                check_gradient("mean", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // argmax
+        if (0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_argmax(ctx0, x[0]));
+
+                check_gradient("argmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // repeat
         {
             int64_t ne2[4];
@@ -549,15 +630,36 @@ int main(int argc, const char ** argv) {
 
             const int nargs = 1;
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[1], ggml_repeat(ctx0, x[0], x[1]))));
 
                 check_gradient("repeat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
             }
+        }
 
+        // repeat back
+        {
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            ne2[0] = ne[0] * ne2[0];
+            ne2[1] = ne[1] * ne2[1];
+            ne2[2] = 1;
+            ne2[3] = 1;
+
+            const int nargs = 1;
+            for (int ndims = 1; ndims <= 2; ++ndims) {
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[0], ggml_repeat_back(ctx0, x[1], x[0]))));
+
+                check_gradient("repeat back", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
+            }
         }
 
         // abs (finite differences do not work)
@@ -566,7 +668,7 @@ int main(int argc, const char ** argv) {
 
         //    for (int ndims = 1; ndims <= 2; ++ndims) {
         //        for (int i = 0; i < nargs; ++i) {
-        //            x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+        //            x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
         //            ggml_set_param(ctx0, x[i]);
         //        }
 
@@ -576,17 +678,82 @@ int main(int argc, const char ** argv) {
         //    }
         //}
 
+        // sgn
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_sgn(ctx0, x[0]));
+
+                check_gradient("sgn", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // neg
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_neg(ctx0, x[0]));
+
+                check_gradient("neg", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // step
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_step(ctx0, x[0]));
+
+                check_gradient("step", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // tanh, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_tanh(ctx0, x[0]));
+
+                check_gradient("tanh", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // mul_mat
         {
             const int nargs = 2;
 
             for (int ndims = 2; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 {
                     int64_t ne2[4];
                     get_random_dims(ne2, 4);
                     ne2[0] = ne[0];
-                    x[1] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                    x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 }
 
                 ggml_set_param(ctx0, x[0]);
@@ -602,13 +769,63 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // elu, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_elu(ctx0, x[0]));
+
+                check_gradient("elu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // relu
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_relu(ctx0, x[0]));
+
+                check_gradient("relu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            }
+        }
+
+        // gelu, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_gelu(ctx0, x[0]));
+
+                check_gradient("gelu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // silu
         {
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -629,11 +846,11 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_rms_norm(ctx0, x[0]));
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_rms_norm(ctx0, x[0], 1e-6f));
 
                 check_gradient("rms_norm", ctx0, x, f, ndims, nargs, 1e-4f, 1.0f, INFINITY);
             }
@@ -647,8 +864,8 @@ int main(int argc, const char ** argv) {
             ne2[0] = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
                 ggml_set_param(ctx0, x[1]);
@@ -659,20 +876,37 @@ int main(int argc, const char ** argv) {
             }
         }
 
-        // cpy
+        // cpy f32
         {
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
                 // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
 
-                check_gradient("cpy", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("cpy f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            }
+        }
+
+        // cpy f16
+        {
+            const int nargs = 2;
+
+            for (int ndims = 1; ndims <= 2; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+                // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
+
+                check_gradient("cpy f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
             }
         }
 
@@ -689,8 +923,8 @@ int main(int argc, const char ** argv) {
                 for (int i = 0; i < ndims; ++i) {
                     ne2[0] *= ne[i];
                 }
-                x[0] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
 
@@ -712,8 +946,8 @@ int main(int argc, const char ** argv) {
                 for (int i = 0; i < ndims; ++i) {
                     ne2[0] *= ne[i];
                 }
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
 
@@ -729,7 +963,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 1);
@@ -737,7 +971,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 1);
                 }
 
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
@@ -758,7 +992,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 2; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 2);
@@ -766,7 +1000,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 2);
                 }
 
-                x[1] = get_random_tensor(ctx0, 2, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -790,7 +1024,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 3; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 3);
@@ -798,7 +1032,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 3);
                 }
 
-                x[1] = get_random_tensor(ctx0, 3, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 3, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -824,7 +1058,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 4; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 4);
@@ -832,7 +1066,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 4);
                 }
 
-                x[1] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -858,7 +1092,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 1);
@@ -866,7 +1100,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 1);
                 }
 
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
@@ -887,7 +1121,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 2; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 2);
@@ -895,7 +1129,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 2);
                 }
 
-                x[1] = get_random_tensor(ctx0, 2, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -915,7 +1149,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -941,7 +1175,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 get_random_dims(ne2, 2);
                 while (ne2[0]*ne2[1] > ggml_nelements(x[0])) {
@@ -971,7 +1205,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 get_random_dims(ne2, 3);
                 while (ne2[0]*ne2[1]*ne2[2] > ggml_nelements(x[0])) {
@@ -1010,7 +1244,7 @@ int main(int argc, const char ** argv) {
                 for (int i=ndims; i<4; ++i) {
                     ne2[i] = 1;
                 }
-                x[0] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -1043,7 +1277,7 @@ int main(int argc, const char ** argv) {
                 for (int i=ndims; i<4; ++i) {
                     ne2[i] = 1;
                 }
-                x[0] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -1060,8 +1294,8 @@ int main(int argc, const char ** argv) {
             int64_t ne3[4] = {1+irand(ne[1]), 1, 1, 1};
             const int nargs = 1;
             const int ndims = 2;
-            x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
-            x[1] = get_random_tensor_int(ctx0, 1, ne3, 0, ne2[1]);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+            x[1] = get_random_tensor_i32(ctx0, 1, ne3, 0, ne2[1]);
 
             ggml_set_param(ctx0, x[0]);
 
@@ -1075,7 +1309,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             const int ndims = 2;
 
-            x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
             ggml_set_param(ctx0, x[0]);
 
             int n_past = irand(ne[0]);
@@ -1090,7 +1324,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             const int ndims = 2;
 
-            x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
             ggml_set_param(ctx0, x[0]);
 
             int n_past = irand(ne[0]);
@@ -1108,7 +1342,7 @@ int main(int argc, const char ** argv) {
             get_random_dims(ne2, 4);
 
             for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_soft_max(ctx0, x[0]));
@@ -1125,8 +1359,8 @@ int main(int argc, const char ** argv) {
             get_random_dims(ne2, 4);
 
             for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
@@ -1136,7 +1370,7 @@ int main(int argc, const char ** argv) {
             }
         }
 
-        // rope
+        // rope f32
         {
             const int nargs = 1;
 
@@ -1148,7 +1382,7 @@ int main(int argc, const char ** argv) {
             for (int ndims = 3; ndims <= 4; ++ndims) {
                 for (int mode = 0; mode < 4; ++mode) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
-                        x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                        x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
 
                         ggml_set_param(ctx0, x[0]);
 
@@ -1163,14 +1397,48 @@ int main(int argc, const char ** argv) {
 
                         struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
 
-                        GGML_PRINT_DEBUG("rope: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
+                        GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
+                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
                     }
                 }
             }
         }
 
-        // flash_attn
+        // rope f16
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+            ne2[0] += ne2[0] % 2;
+            int n_rot = ne2[0];
+
+            for (int ndims = 3; ndims <= 4; ++ndims) {
+                for (int mode = 0; mode < 4; ++mode) {
+                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
+                        x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);
+
+                        ggml_set_param(ctx0, x[0]);
+
+                        const bool skip_past = (mode & 1);
+                        if (skip_past) {
+                            // we have no past, so this would have to work on uninitialized memory.
+                            // we only test the gradients here;
+                            // skip_past should have no influence on gradient computation.
+                            // so when other modes work, we assume that this does as well.
+                            continue;
+                        }
+
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+
+                        GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
+                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
+                    }
+                }
+            }
+        }
+
+        // flash_attn f32
         {
             const int nargs = 3;
 
@@ -1196,16 +1464,57 @@ int main(int argc, const char ** argv) {
                         nek[3] = 1;
                         nev[3] = 1;
                     }
-                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
                     ggml_set_param(ctx0, x[0]);
                     ggml_set_param(ctx0, x[1]);
                     ggml_set_param(ctx0, x[2]);
 
                     struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                    check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
+
+        // flash_attn f16, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor_f16(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor_f16(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor_f16(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
                 }
             }
         }

tests/test-opt.c

@@ -125,9 +125,9 @@ int main(void) {
     };
     struct ggml_context * ctx = ggml_init(params);
 
-    int64_t ne1[4] = {4, 1024, 1, 1};
-    int64_t ne2[4] = {4, 2048, 1, 1};;
-    int64_t ne3[4] = {1024, 2048, 1, 1};
+    int64_t ne1[4] = {4, 128, 1, 1};
+    int64_t ne2[4] = {4, 256, 1, 1};;
+    int64_t ne3[4] = {128, 256, 1, 1};
 
     struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
     struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);

tests/test-sampling.cpp

@@ -200,4 +200,6 @@ int main(void) {
     test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f);
 
     printf("OK\n");
+
+    return 0;
 }