better documentation

* Replace convert-*-hf-to-gguf.py files with convert-hf-to-gguf.py

.github/workflows/build.yml

@@ -288,6 +288,7 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
+      SDE_VERSION: 9.21.1-2023-04-24
 
     strategy:
       matrix:
@@ -383,11 +384,23 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
+        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
         run: |
           cd build
           ctest -C Release --verbose --timeout 900
 
+      - name: Test (Intel SDE)
+        id: cmake_test_sde
+        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz"
+          # for some weird reason windows tar doesn't like sde tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+          cd build
+          & $sde -future -- ctest -C Release --verbose --timeout 900
+
       - name: Determine tag name
         id: tag
         shell: bash

.gitignore

@@ -15,6 +15,7 @@
 .DS_Store
 .build/
 .cache/
+.ccls-cache/
 .direnv/
 .envrc
 .swiftpm
@@ -45,7 +46,7 @@ models-mnt
 /infill
 /libllama.so
 /llama-bench
-/llava
+/llava-cli
 /main
 /metal
 /perplexity
@@ -64,7 +65,7 @@ models-mnt
 /parallel
 /train-text-from-scratch
 /vdot
-build-info.h
+/common/build-info.cpp
 arm_neon.h
 compile_commands.json
 CMakeSettings.json

CMakeLists.txt

@@ -10,7 +10,7 @@ endif()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     # configure project version
@@ -94,46 +94,12 @@ option(LLAMA_CLBLAST                         "llama: use CLBlast"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
-option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
 
-#
-# Build info header
-#
-
-# Generate initial build-info.h
-include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
-
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-
-    # Is git submodule
-    if(NOT IS_DIRECTORY "${GIT_DIR}")
-        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
-        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
-    endif()
-
-    # Add a custom target for build-info.h
-    add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
-
-    # Add a custom command to rebuild build-info.h when .git/index changes
-    add_custom_command(
-        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
-        COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        DEPENDS "${GIT_DIR}/index"
-        VERBATIM
-    )
-else()
-    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-endif()
-
 #
 # Compile flags
 #
@@ -278,13 +244,8 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_K_QUANTS)
-    set(GGML_HEADERS_EXTRA k_quants.h)
-    set(GGML_SOURCES_EXTRA k_quants.c)
-    add_compile_definitions(GGML_USE_K_QUANTS)
-    if (LLAMA_QKK_64)
-        add_compile_definitions(GGML_QKK_64)
-    endif()
+if (LLAMA_QKK_64)
+    add_compile_definitions(GGML_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
@@ -549,6 +510,10 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
+        # instruction set detection for MSVC only
+        if (LLAMA_NATIVE)
+            include(cmake/FindSIMD.cmake)
+        endif ()
         if (LLAMA_AVX512)
             add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
@@ -673,6 +638,8 @@ add_library(ggml OBJECT
             ggml-alloc.h
             ggml-backend.c
             ggml-backend.h
+            ggml-quants.c
+            ggml-quants.h
             ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}

Makefile

@@ -1,7 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf llama-bench llava baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
 	speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
@@ -342,13 +342,9 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifndef LLAMA_NO_K_QUANTS
-	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
-	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
 	MK_CPPFLAGS += -DGGML_QKK_64
 endif
-endif
 
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
@@ -365,7 +361,7 @@ ifdef LLAMA_MPI
 	MK_CPPFLAGS += -DGGML_USE_MPI
 	MK_CFLAGS   += -Wno-cast-qual
 	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS     += ggml-mpi.o
+	OBJS        += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
@@ -382,7 +378,7 @@ endif # LLAMA_BLIS
 ifdef LLAMA_CUBLAS
 	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
-	OBJS      += ggml-cuda.o
+	OBJS         += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
@@ -497,11 +493,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_K_QUANTS
-k_quants.o: k_quants.c k_quants.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_NO_K_QUANTS
-
 # combine build flags with cmdline overrides
 override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
 override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
@@ -542,13 +533,16 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
 
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
-COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -569,46 +563,46 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.h ggml.o llama.o common.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -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 examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+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 examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
@@ -620,28 +614,31 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
+
+llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifdef LLAMA_METAL
@@ -654,7 +651,7 @@ swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif
 
-build-info.h: $(wildcard .git/index) scripts/build-info.sh
+common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
@@ -662,13 +659,16 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi
 
+build-info.o: common/build-info.cpp
+	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
+
 #
 # Tests
 #
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
@@ -682,40 +682,40 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h

Package.swift

@@ -42,13 +42,12 @@ let package = Package(
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "k_quants.c",
+                "ggml-quants.c",
             ] + additionalSources,
             resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_K_QUANTS"),
                 .define("GGML_USE_ACCELERATE")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14

README.md

@@ -2,7 +2,6 @@
 
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
-[![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
@@ -11,8 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- LLaVA support: https://github.com/ggerganov/llama.cpp/pull/3436
-- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
+- ⚠️ **Upcoming change that might break functionality. Help with testing is needed:** https://github.com/ggerganov/llama.cpp/pull/3912
 
 ----
 

build.zig

@@ -10,7 +10,6 @@ const Maker = struct {
     builder: *std.build.Builder,
     target: CrossTarget,
     optimize: Mode,
-    config_header: *ConfigHeader,
     enable_lto: bool,
 
     include_dirs: ArrayList([]const u8),
@@ -41,26 +40,24 @@ const Maker = struct {
         const commit_hash = try std.ChildProcess.exec(
             .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
         );
-        const config_header = builder.addConfigHeader(
-            .{ .style = .blank, .include_path = "build-info.h" },
-            .{
-                .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
-                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
-                .BUILD_TARGET = try target.allocDescription(builder.allocator),
-            },
-        );
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
+            \\int LLAMA_BUILD_NUMBER = {};
+            \\char const *LLAMA_COMMIT = "{s}";
+            \\char const *LLAMA_COMPILER = "Zig {s}";
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
+            \\
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
         var m = Maker{
             .builder = builder,
             .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
-            .config_header = config_header,
             .enable_lto = false,
             .include_dirs = ArrayList([]const u8).init(builder.allocator),
             .cflags = ArrayList([]const u8).init(builder.allocator),
             .cxxflags = ArrayList([]const u8).init(builder.allocator),
             .objs = ArrayList(*Compile).init(builder.allocator),
         };
+
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
@@ -72,7 +69,7 @@ const Maker = struct {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
         if (o.target.getAbi() != .msvc)
             o.defineCMacro("_GNU_SOURCE", null);
-        o.addConfigHeader(m.config_header);
+
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
@@ -85,7 +82,6 @@ const Maker = struct {
                 o.linkLibCpp();
             }
         }
-        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -105,7 +101,6 @@ const Maker = struct {
             // linkLibCpp already add (libc++ + libunwind + libc)
             e.linkLibCpp();
         }
-        e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
         return e;
@@ -116,16 +111,12 @@ pub fn build(b: *std.build.Builder) !void {
     var make = try Maker.init(b);
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
-    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
-        try make.addFlag("-DGGML_USE_K_QUANTS");
-        const k_quants = make.obj("k_quants", "k_quants.c");
-        try make.objs.append(k_quants);
-    }
-
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
+    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const llama = make.obj("llama", "llama.cpp");
+    const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
@@ -133,14 +124,14 @@ pub fn build(b: *std.build.Builder) !void {
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser, clip });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

cmake/FindSIMD.cmake

@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(LLAMA_AVX OFF)
+else()
+    set(LLAMA_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(LLAMA_AVX2 OFF)
+else()
+    set(LLAMA_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(LLAMA_AVX512 OFF)
+else()
+    set(LLAMA_AVX512 ON)
+endif()

common/CMakeLists.txt

@@ -1,8 +1,47 @@
 # common
 
+
+# Build info header
+#
+
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+
+    # Is git submodule
+    if(NOT IS_DIRECTORY "${GIT_DIR}")
+        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
+        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
+        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+    endif()
+
+    set(GIT_INDEX "${GIT_DIR}/index")
+else()
+    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
+endif()
+
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
+set(TARGET build_info)
+add_library(${TARGET} OBJECT build-info.cpp)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+
 set(TARGET common)
 
-add_library(${TARGET} OBJECT
+add_library(${TARGET} STATIC
+    base64.hpp
     common.h
     common.cpp
     sampling.h
@@ -21,4 +60,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
+target_link_libraries(${TARGET} PRIVATE llama build_info)

common/base64.hpp

@@ -0,0 +1,392 @@
+/*
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or
+distribute this software, either in source code form or as a compiled
+binary, for any purpose, commercial or non-commercial, and by any
+means.
+
+In jurisdictions that recognize copyright laws, the author or authors
+of this software dedicate any and all copyright interest in the
+software to the public domain. We make this dedication for the benefit
+of the public at large and to the detriment of our heirs and
+successors. We intend this dedication to be an overt act of
+relinquishment in perpetuity of all present and future rights to this
+software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org>
+*/
+
+#ifndef PUBLIC_DOMAIN_BASE64_HPP_
+#define PUBLIC_DOMAIN_BASE64_HPP_
+
+#include <cstdint>
+#include <iterator>
+#include <stdexcept>
+#include <string>
+
+class base64_error : public std::runtime_error
+{
+public:
+    using std::runtime_error::runtime_error;
+};
+
+class base64
+{
+public:
+    enum class alphabet
+    {
+        /** the alphabet is detected automatically */
+        auto_,
+        /** the standard base64 alphabet is used */
+        standard,
+        /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/
+        url_filename_safe
+    };
+
+    enum class decoding_behavior
+    {
+        /** if the input is not padded, the remaining bits are ignored */
+        moderate,
+        /** if a padding character is encounter decoding is finished */
+        loose
+    };
+
+    /**
+     Encodes all the elements from `in_begin` to `in_end` to `out`.
+
+     @warning The source and destination cannot overlap. The destination must be able to hold at least
+     `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than
+     8 bits
+     @tparam Output_iterator the destination; the elements written to it are from the type `char`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @returns the iterator to the next element past the last element copied
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet = alphabet::standard)
+    {
+        constexpr auto pad = '=';
+        const char* alpha  = alphabet == alphabet::url_filename_safe
+                                ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
+                                : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+        while (in_begin != in_end) {
+            std::uint8_t i0 = 0, i1 = 0, i2 = 0;
+
+            // first character
+            i0 = static_cast<std::uint8_t>(*in_begin);
+            ++in_begin;
+
+            *out = alpha[i0 >> 2 & 0x3f];
+            ++out;
+
+            // part of first character and second
+            if (in_begin != in_end) {
+                i1 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)];
+                ++out;
+            } else {
+                *out = alpha[(i0 & 0x3) << 4];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // part of second character and third
+            if (in_begin != in_end) {
+                i2 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)];
+                ++out;
+            } else {
+                *out = alpha[(i1 & 0xf) << 2];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // rest of third
+            *out = alpha[i2 & 0x3f];
+            ++out;
+        }
+
+        return out;
+    }
+    /**
+     Encodes a string.
+
+     @param str the string that should be encoded
+     @param alphabet which alphabet should be used
+     @returns the encoded base64 string
+     @throws see base64::encode()
+    */
+    static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(str.length()) + 1);
+
+        encode(str.begin(), str.end(), std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Encodes a char array.
+
+     @param buffer the char array
+     @param size the size of the array
+     @param alphabet which alphabet should be used
+     @returns the encoded string
+    */
+    static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(size) + 1);
+
+        encode(buffer, buffer + size, std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`,
+     in other words: inplace decoding is possible.
+
+     @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`,
+     otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `char`
+     @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the iterator to the next element past the last element copied
+     @throws base64_error depending on the set behavior
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet          = alphabet::auto_,
+                                  decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        //constexpr auto pad = '=';
+        std::uint8_t last  = 0;
+        auto bits          = 0;
+
+        while (in_begin != in_end) {
+            auto c = *in_begin;
+            ++in_begin;
+
+            if (c == '=') {
+                break;
+            }
+
+            auto part = _base64_value(alphabet, c);
+
+            // enough bits for one byte
+            if (bits + 6 >= 8) {
+                *out = (last << (8 - bits)) | (part >> (bits - 2));
+                ++out;
+
+                bits -= 2;
+            } else {
+                bits += 6;
+            }
+
+            last = part;
+        }
+
+        // check padding
+        if (behavior != decoding_behavior::loose) {
+            while (in_begin != in_end) {
+                auto c = *in_begin;
+                ++in_begin;
+
+                if (c != '=') {
+                    throw base64_error("invalid base64 character.");
+                }
+            }
+        }
+
+        return out;
+    }
+    /**
+     Decodes a string.
+
+     @param str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(str.length()));
+
+        decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string.
+
+     @param buffer the base64 encoded buffer
+     @param size the size of the buffer
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(size));
+
+        decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string inplace.
+
+     @param[in,out] str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @throws base64::decode_inplace()
+    */
+    static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_,
+                               decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin());
+    }
+    /**
+     Decodes a char array inplace.
+
+     @param[in,out] str the string array
+     @param size the length of the array
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the pointer to the next element past the last element decoded
+     @throws base64::decode_inplace()
+    */
+    static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_,
+                                decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        return decode(str, str + size, str, alphabet, behavior);
+    }
+    /**
+     Returns the required decoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{4} \rceil \cdot 3
+     $$
+
+     @param size the size of the encoded input
+     @returns the size of the resulting decoded buffer; this the absolute maximum
+    */
+    static std::size_t max_decode_size(std::size_t size) noexcept
+    {
+        return (size / 4 + (size % 4 ? 1 : 0)) * 3;
+    }
+    /**
+     Returns the required encoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{3} \rceil \cdot 4
+     $$
+
+     @param size the size of the decoded input
+     @returns the size of the resulting encoded buffer
+    */
+    static std::size_t required_encode_size(std::size_t size) noexcept
+    {
+        return (size / 3 + (size % 3 ? 1 : 0)) * 4;
+    }
+
+private:
+    static std::uint8_t _base64_value(alphabet& alphabet, char c)
+    {
+        if (c >= 'A' && c <= 'Z') {
+            return c - 'A';
+        } else if (c >= 'a' && c <= 'z') {
+            return c - 'a' + 26;
+        } else if (c >= '0' && c <= '9') {
+            return c - '0' + 52;
+        }
+
+        // comes down to alphabet
+        if (alphabet == alphabet::standard) {
+            if (c == '+') {
+                return 62;
+            } else if (c == '/') {
+                return 63;
+            }
+        } else if (alphabet == alphabet::url_filename_safe) {
+            if (c == '-') {
+                return 62;
+            } else if (c == '_') {
+                return 63;
+            }
+        } // auto detect
+        else {
+            if (c == '+') {
+                alphabet = alphabet::standard;
+
+                return 62;
+            } else if (c == '/') {
+                alphabet = alphabet::standard;
+
+                return 63;
+            } else if (c == '-') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 62;
+            } else if (c == '_') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 63;
+            }
+        }
+
+        throw base64_error("invalid base64 character.");
+    }
+};
+
+#endif // !PUBLIC_DOMAIN_BASE64_HPP_

common/build-info.cpp.in

@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/common.cpp

@@ -1,5 +1,4 @@
 #include "common.h"
-#include "build-info.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -91,6 +90,19 @@ void process_escapes(std::string& input) {
                 case '\'': input[output_idx++] = '\''; break;
                 case '\"': input[output_idx++] = '\"'; break;
                 case '\\': input[output_idx++] = '\\'; break;
+                case 'x':
+                    // Handle \x12, etc
+                    if (input_idx + 2 < input_len) {
+                        const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
+                        char *err_p = nullptr;
+                        const long val = std::strtol(x, &err_p, 16);
+                        if (err_p == x + 2) {
+                            input_idx += 2;
+                            input[output_idx++] = char(val);
+                            break;
+                        }
+                    }
+                    // fall through
                 default:   input[output_idx++] = '\\';
                            input[output_idx++] = input[input_idx]; break;
             }
@@ -103,9 +115,24 @@ void process_escapes(std::string& input) {
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    bool result = true;
+    try {
+        if (!gpt_params_parse_ex(argc, argv, params)) {
+            gpt_print_usage(argc, argv, gpt_params());
+            exit(0);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        gpt_print_usage(argc, argv, gpt_params());
+        exit(1);
+    }
+    return result;
+}
+
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
-    gpt_params default_params;
     const std::string arg_prefix = "--";
     llama_sampling_params & sparams = params.sparams;
 
@@ -204,12 +231,52 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scaling") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
+            else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rope_freq_scale = 1.0f/std::stof(argv[i]);
+        } else if (arg == "--yarn-orig-ctx") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_orig_ctx = std::stoi(argv[i]);
+        } else if (arg == "--yarn-ext-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_ext_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-attn-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_attn_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-fast") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_fast = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-slow") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_slow = std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -218,6 +285,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             sparams.top_p = std::stof(argv[i]);
+        } else if (arg == "--min-p") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.min_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -343,6 +416,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_sequences = std::stoi(argv[i]);
+        } else if (arg == "--p-accept" || arg == "-pa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.p_accept = std::stof(argv[i]);
+        } else if (arg == "--p-split" || arg == "-ps") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.p_split = std::stof(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -548,11 +633,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
         } else if (arg == "-h" || arg == "--help") {
-            gpt_print_usage(argc, argv, default_params);
-#ifndef LOG_DISABLE_LOGS
-            log_print_usage();
-#endif // LOG_DISABLE_LOGS
-            exit(0);
+            return false;
+
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
         } else if (arg == "--in-prefix-bos") {
@@ -611,22 +693,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         // End of Parse args for logging parameters
 #endif // LOG_DISABLE_LOGS
         } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            gpt_print_usage(argc, argv, default_params);
-            exit(1);
+            throw std::invalid_argument("error: unknown argument: " + arg);
         }
     }
     if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        gpt_print_usage(argc, argv, default_params);
-        exit(1);
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
              params.instruct)) {
-        fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
-        gpt_print_usage(argc, argv, default_params);
-        exit(1);
+
+        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
     if (params.escape) {
@@ -645,6 +722,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     const llama_sampling_params & sparams = params.sparams;
 
+    printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
@@ -679,6 +757,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
+    printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
     printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
     printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
     printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
@@ -701,9 +780,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --cfg-negative-prompt-file FNAME\n");
     printf("                        negative prompt file to use for guidance. (default: empty)\n");
     printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
-    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-scaling {none,linear,yarn}\n");
+    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
+    printf("  --rope-scale N        RoPE context scaling factor, expands context by a factor of N\n");
     printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
+    printf("  --yarn-orig-ctx N     YaRN: original context size of model (default: 0 = model training context size)\n");
+    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
+    printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
+    printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
+    printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
     printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
@@ -717,6 +803,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
     printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -pa N, --p-accept N   speculative decoding accept probability (default: %.1f)\n", (double)params.p_accept);
+    printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
     printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
     printf("  --image IMAGE_FILE    path to an image file. use with multimodal models\n");
@@ -755,6 +843,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
     printf("\n");
+#ifndef LOG_DISABLE_LOGS
+    log_print_usage();
+#endif // LOG_DISABLE_LOGS
 }
 
 std::string get_system_info(const gpt_params & params) {
@@ -808,17 +899,23 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
-    cparams.n_ctx           = params.n_ctx;
-    cparams.n_batch         = params.n_batch;
-    cparams.n_threads       = params.n_threads;
-    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-    cparams.mul_mat_q       = params.mul_mat_q;
-    cparams.seed            = params.seed;
-    cparams.f16_kv          = params.memory_f16;
-    cparams.logits_all      = params.logits_all;
-    cparams.embedding       = params.embedding;
-    cparams.rope_freq_base  = params.rope_freq_base;
-    cparams.rope_freq_scale = params.rope_freq_scale;
+    cparams.n_ctx             = params.n_ctx;
+    cparams.n_batch           = params.n_batch;
+    cparams.n_threads         = params.n_threads;
+    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q         = params.mul_mat_q;
+    cparams.seed              = params.seed;
+    cparams.f16_kv            = params.memory_f16;
+    cparams.logits_all        = params.logits_all;
+    cparams.embedding         = params.embedding;
+    cparams.rope_scaling_type = params.rope_scaling_type;
+    cparams.rope_freq_base    = params.rope_freq_base;
+    cparams.rope_freq_scale   = params.rope_freq_scale;
+    cparams.yarn_ext_factor   = params.yarn_ext_factor;
+    cparams.yarn_attn_factor  = params.yarn_attn_factor;
+    cparams.yarn_beta_fast    = params.yarn_beta_fast;
+    cparams.yarn_beta_slow    = params.yarn_beta_slow;
+    cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
 
     return cparams;
 }
@@ -889,7 +986,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_kv_cache_clear(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -1128,8 +1225,8 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
     const llama_sampling_params & sparams = params.sparams;
 
-    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
-    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
     fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
     fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
     fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
@@ -1275,6 +1372,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }

common/common.h

@@ -9,6 +9,7 @@
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
+#include <cmath>
 #include <string>
 #include <vector>
 #include <random>
@@ -25,35 +26,51 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                             \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+#define print_build_info() do {                                                                     \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
+// build info
+extern int LLAMA_BUILD_NUMBER;
+extern char const *LLAMA_COMMIT;
+extern char const *LLAMA_COMPILER;
+extern char const *LLAMA_BUILD_TARGET;
+
 //
 // CLI argument parsing
 //
 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_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
-    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_draft                         = 16;   // number of tokens to draft during speculative decoding
-    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
-    int32_t n_sequences                     = 1;    // number of sequences to decode
-    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
-    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_beams                         = 0;    // if non-zero then use beam search of given width.
-    float   rope_freq_base                  = 0.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
+    int32_t n_threads_batch                 = -1;    // number of threads to use for batch processing (-1 = use n_threads)
+    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_draft                         = 16;    // number of tokens to draft during speculative decoding
+    int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel                      = 1;     // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;     // number of sequences to decode
+    float   p_accept                        = 0.5f;  // speculative decoding accept probability
+    float   p_split                         = 0.1f;  // speculative decoding split probability
+    int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    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_beams                         = 0;     // if non-zero then use beam search of given width.
+    float   rope_freq_base                  = 0.0f;  // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
+    float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
+    float   yarn_attn_factor                = 1.0f;  // YaRN magnitude scaling factor
+    float   yarn_beta_fast                  = 32.0f; // YaRN low correction dim
+    float   yarn_beta_slow                  = 1.0f;  // YaRN high correction dim
+    int32_t yarn_orig_ctx                   = 0;     // YaRN original context length
+    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
+                                                                              //       pinging @cebtenzzre
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -77,7 +94,7 @@ struct gpt_params {
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                     //                                       (which is more convenient to use for plotting)
                                     //
-    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    bool   hellaswag       = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
@@ -110,6 +127,8 @@ struct gpt_params {
     std::string image = ""; // path to an image file
 };
 
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
+
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);

common/log.h

@@ -97,38 +97,56 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
-// NOTE: currently disabled as it produces too many log files
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
 // Utility to obtain "pid" like unique process id and use it when creating log files.
-//inline std::string log_get_pid()
-//{
-//    static std::string pid;
-//    if (pid.empty())
-//    {
-//        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
-//        //  it's not the same as "pid" but is unique enough to solve multiple instances
-//        //  trying to write to the same log.
-//        std::stringstream ss;
-//        ss << std::this_thread::get_id();
-//        pid = ss.str();
-//    }
-//
-//    return pid;
-//}
+inline std::string log_get_pid()
+{
+   static std::string pid;
+   if (pid.empty())
+   {
+       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+       //  it's not the same as "pid" but is unique enough to solve multiple instances
+       //  trying to write to the same log.
+       std::stringstream ss;
+       ss << std::this_thread::get_id();
+       pid = ss.str();
+   }
+
+   return pid;
+}
 
 // Utility function for generating log file names with unique id based on thread id.
 //  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
 //  where the number is a runtime id of the current thread.
 
-#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
 
 // INTERNAL, DO NOT USE
-inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
 {
+    static bool _multilog = false;
+
+    if (multilog != LogTriStateSame)
+    {
+        _multilog = multilog == LogTriStateTrue;
+    }
+
     std::stringstream buf;
 
     buf << log_file_basename;
-    //buf << ".";
-    //buf << log_get_pid();
+    if (_multilog)
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
     buf << ".";
     buf << log_file_extension;
 
@@ -213,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     #define LOG_TEE_FLF_VAL ,""
 #endif
 
-// Utility for synchronizing log configuration state
-//  since std::optional was introduced only in c++17
-enum LogTriState
-{
-    LogTriStateSame,
-    LogTriStateFalse,
-    LogTriStateTrue
-};
-
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
@@ -315,16 +324,23 @@ enum LogTriState
 #endif
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
 {
-    static bool _initialized{false};
-    static bool _disabled{(filename.empty() && target == nullptr)};
+    static bool _initialized = false;
+    static bool _append = false;
+    static bool _disabled = filename.empty() && target == nullptr;
     static std::string log_current_filename{filename};
     static FILE *log_current_target{target};
     static FILE *logfile = nullptr;
 
     if (change)
     {
+        if (append != LogTriStateSame)
+        {
+            _append = append == LogTriStateTrue;
+            return logfile;
+        }
+
         if (disable == LogTriStateTrue)
         {
             // Disable primary target
@@ -377,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
             }
         }
 
-        logfile = fopen(filename.c_str(), "w");
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
     }
 
     if (!logfile)
@@ -398,9 +414,9 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
 }
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
 {
-    return log_handler1_impl(change, disable, filename, target);
+    return log_handler1_impl(change, append, disable, filename, target);
 }
 
 // Disables logs entirely at runtime.
@@ -411,7 +427,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTri
 // INTERNAL, DO NOT USE
 inline FILE *log_disable_impl()
 {
-    return log_handler1_impl(true, LogTriStateTrue);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
 }
 
 // Enables logs at runtime.
@@ -420,19 +436,31 @@ inline FILE *log_disable_impl()
 // INTERNAL, DO NOT USE
 inline FILE *log_enable_impl()
 {
-    return log_handler1_impl(true, LogTriStateFalse);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
 }
 
 // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
 #define log_set_target(target) log_set_target_impl(target)
 
 // INTERNAL, DO NOT USE
-inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
-inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
 
 // INTERNAL, DO NOT USE
 inline FILE *log_handler() { return log_handler1_impl(); }
 
+// Enable or disable creating separate log files for each run.
+//  can ONLY be invoked BEFORE first log use.
+#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
+// Enable or disable append mode for log file.
+//  can ONLY be invoked BEFORE first log use.
+#define log_append(enable) log_append_impl(enable)
+// INTERNAL, DO NOT USE
+inline FILE *log_append_impl(bool enable)
+{
+    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
+}
+
 inline void log_test()
 {
     log_disable();
@@ -494,6 +522,18 @@ inline bool log_param_single_parse(const std::string & param)
         return true;
     }
 
+    if (param == "--log-new")
+    {
+        log_multilog(true);
+        return true;
+    }
+
+    if (param == "--log-append")
+    {
+        log_append(true);
+        return true;
+    }
+
     return false;
 }
 
@@ -523,7 +563,9 @@ inline void log_print_usage()
     printf("  --log-disable         Disable trace logs\n");
     printf("  --log-enable          Enable trace logs\n");
     printf("  --log-file            Specify a log filename (without extension)\n");
-    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+    printf("  --log-new             Create a separate new log file on start. "
+                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
+    printf("  --log-append          Don't truncate the old log file.\n");
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)

common/sampling.cpp

@@ -39,6 +39,7 @@ void llama_sampling_free(struct llama_sampling_context * ctx) {
 void llama_sampling_reset(llama_sampling_context * ctx) {
     if (ctx->grammar != NULL) {
         llama_grammar_free(ctx->grammar);
+        ctx->grammar = NULL;
     }
 
     if (!ctx->parsed_grammar.rules.empty()) {
@@ -89,10 +90,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
-            params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp,
+            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
             params.mirostat, params.mirostat_eta, params.mirostat_tau);
 
     return std::string(result);
@@ -110,6 +111,7 @@ llama_token llama_sampling_sample(
     const float   temp            = params.temp;
     const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
     const float   top_p           = params.top_p;
+    const float   min_p           = params.min_p;
     const float   tfs_z           = params.tfs_z;
     const float   typical_p       = params.typical_p;
     const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
@@ -190,6 +192,7 @@ llama_token llama_sampling_sample(
             llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
             llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
             llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
+            llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep);
             llama_sample_temp     (ctx_main, &cur_p, temp);
 
             id = llama_sample_token(ctx_main, &cur_p);

common/sampling.h

@@ -14,6 +14,7 @@ typedef struct llama_sampling_params {
     int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
     int32_t top_k             = 40;    // <= 0 to use vocab size
     float   top_p             = 0.95f; // 1.0 = disabled
+    float   min_p             = 0.05f; // 0.0 = disabled
     float   tfs_z             = 1.00f; // 1.0 = disabled
     float   typical_p         = 1.00f; // 1.0 = disabled
     float   temp              = 0.80f; // 1.0 = disabled

common/train.cpp

@@ -1045,6 +1045,7 @@ struct train_params_common get_default_train_params_common() {
     params.n_batch    =    8;
     params.n_gradient_accumulation = 1;
     params.n_epochs   = -1;
+    params.n_gpu_layers = 0;
 
     params.custom_n_ctx = false;
 
@@ -1080,6 +1081,7 @@ struct train_params_common get_default_train_params_common() {
     params.adam_beta2          = 0.999f;
     params.adam_gclip          = 1.0f;
     params.adam_eps_f          = 0.0f;
+
     return params;
 }
 

common/train.h

@@ -44,6 +44,7 @@ struct train_params_common {
     int n_batch;
     int n_gradient_accumulation;
     int n_epochs;
+    int n_gpu_layers;
 
     bool custom_n_ctx;
 

convert-baichuan-hf-to-gguf.py

@@ -163,7 +163,8 @@ gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
 if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
     if "type" in hparams["rope_scaling"]:
         if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+            gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
 
 
 # TOKENIZATION

convert-bloom-hf-to-gguf.py

@@ -1,247 +0,0 @@
-#!/usr/bin/env python3
-# HF bloom --> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import re
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-# Supported Models:
-#   https://huggingface.co/bigscience/bloom-1b7
-#   https://huggingface.co/bigscience/bloom-3b
-#   https://huggingface.co/bigscience/bloom-7b1
-#   https://huggingface.co/Langboat/bloom-1b4-zh
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a Bloom model to a GGML compatible file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    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 (*.bin)")
-    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "BloomForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.BLOOM
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("Bloom")
-n_embed = hparams.get("hidden_size", hparams.get("n_embed"))
-n_head = hparams.get("n_head", hparams.get("num_attention_heads"))
-gguf_writer.add_context_length(hparams.get("seq_length", n_embed))
-gguf_writer.add_embedding_length(n_embed)
-gguf_writer.add_feed_forward_length(4 * n_embed)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(n_head)
-gguf_writer.add_head_count_kv(n_head)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-added_vocab = tokenizer.get_added_vocab()
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    if i not in reverse_vocab:
-        tokens.append(f"[PAD{i}]")
-        toktypes.append(gguf.TokenType.USER_DEFINED)
-    elif reverse_vocab[i] in added_vocab:
-        tokens.append(reverse_vocab[i])
-        if tokenizer.added_tokens_decoder[i].special:
-            toktypes.append(gguf.TokenType.CONTROL)
-        else:
-            toktypes.append(gguf.TokenType.USER_DEFINED)
-    else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
-
-# params for qkv transform
-n_head_kv = hparams.get("n_head_kv", n_head)
-head_dim = n_embed // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    has_lm_head = True
-    if "lm_head.weight" not in model_part.keys() and "output.weight" not in model_part.keys():
-        has_lm_head = False
-
-    for original_name in model_part.keys():
-        data = model_part[original_name]
-        name = re.sub(r'transformer\.', '', original_name)
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
-            # Map bloom-style qkv_linear to gpt-style qkv_linear
-            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
-            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
-            qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
-            data = np.concatenate(
-                (qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
-                 qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
-                 qkv_weights[:, 2, :, :].reshape((-1, n_embed))),
-                axis=0
-            )
-            print("re-format attention.linear_qkv.weight")
-        elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
-            qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
-            data = np.concatenate(
-                (qkv_bias[:, 0, :].reshape((n_embed,)),
-                 qkv_bias[:, 1, :].reshape((n_embed,)),
-                 qkv_bias[:, 2, :].reshape((n_embed,))),
-                axis=0
-            )
-            print("re-format attention.linear_qkv.bias")
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-        if not has_lm_head and name == "word_embeddings.weight":
-            gguf_writer.add_tensor("output.weight", data)
-            print(name, "=>", "output.weight" + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))  # noqa
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-falcon-hf-to-gguf.py

@@ -1,253 +0,0 @@
-#!/usr/bin/env python3
-# HF falcon--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import contextlib
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path, prefix: str) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith(prefix):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] not in ("RWForCausalLM", "FalconForCausalLM"):
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model, "model-00")
-if num_parts:
-    is_safetensors = True
-    from safetensors import safe_open
-else:
-    is_safetensors = False
-    num_parts = count_model_parts(dir_model, "pytorch_model-")
-
-ARCH=gguf.MODEL_ARCH.FALCON
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams.get("num_hidden_layers")
-if block_count is None:
-    block_count = hparams["n_layer"]  # old name
-
-n_head = hparams.get("num_attention_heads")
-if n_head is None:
-    n_head = hparams["n_head"]  # old name
-
-n_head_kv = hparams.get("num_kv_heads")
-if n_head_kv is None:
-    n_head_kv = hparams.get("n_head_kv", 1)  # old name
-
-gguf_writer.add_name("Falcon")
-gguf_writer.add_context_length(2048) # not in config.json
-gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(n_head)
-gguf_writer.add_head_count_kv(n_head_kv)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i])
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-head_dim = hparams["hidden_size"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-elif is_safetensors:
-    part_names = (
-        f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
-    )
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    if is_safetensors:
-        ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
-    else:
-        ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
-
-    with ctx as model_part:
-        for name in model_part.keys():
-            data = model_part.get_tensor(name) if is_safetensors else model_part[name]
-
-            old_dtype = data.dtype
-
-            # convert any unsupported data types to float32
-            if data.dtype != torch.float16 and data.dtype != torch.float32:
-                data = data.to(torch.float32)
-
-            # QKV tensor transform
-            # The original query_key_value tensor contains n_head_kv "kv groups",
-            # each consisting of n_head/n_head_kv query weights followed by one key
-            # and one value weight (shared by all query heads in the kv group).
-            # This layout makes it a big pain to work with in GGML.
-            # So we rearrange them here,, so that we have n_head query weights
-            # followed by n_head_kv key weights followed by n_head_kv value weights,
-            # in contiguous fashion.
-            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
-
-            if "query_key_value" in name:
-                qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-                q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
-                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                data = torch.cat((q,k,v)).reshape_as(data)
-
-            data = data.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-            if new_name is None:
-                print("Can not map tensor '" + name + "'")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-            gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-gptneox-hf-to-gguf.py

@@ -1,221 +0,0 @@
-#!/usr/bin/env python3
-# HF gptneox--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTNeoXForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.GPTNEOX
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-
-gguf_writer.add_name(dir_model.name)
-gguf_writer.add_context_length(hparams["max_position_embeddings"])
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"])))
-gguf_writer.add_head_count(hparams["num_attention_heads"])
-gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-added_vocab = tokenizer.get_added_vocab()
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    if i not in reverse_vocab:
-        tokens.append(f"[PAD{i}]")
-        toktypes.append(gguf.TokenType.USER_DEFINED)
-    elif reverse_vocab[i] in added_vocab:
-        tokens.append(reverse_vocab[i])
-        if tokenizer.added_tokens_decoder[i].special:
-            toktypes.append(gguf.TokenType.CONTROL)
-        else:
-            toktypes.append(gguf.TokenType.USER_DEFINED)
-    else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name.endswith(".attention.masked_bias") or name.endswith(".attention.bias") or name.endswith(".attention.rotary_emb.inv_freq"):
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-hf-to-gguf.py

@@ -0,0 +1,890 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import json
+import os
+import re
+import sys
+from enum import IntEnum
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
+
+import numpy as np
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+
+###### MODEL DEFINITIONS ######
+
+class SentencePieceTokenTypes(IntEnum):
+    NORMAL = 1
+    UNKNOWN = 2
+    CONTROL = 3
+    USER_DEFINED = 4
+    UNUSED = 5
+    BYTE = 6
+
+
+class Model:
+    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
+        self.dir_model = dir_model
+        self.ftype = ftype
+        self.fname_out = fname_out
+        self.is_big_endian = is_big_endian
+        self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.is_safetensors = self._is_model_safetensors()
+        self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
+        self.part_names = self._get_part_names()
+        self.hparams = Model.load_hparams(self.dir_model)
+        self.model_arch = self._get_model_architecture()
+        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for part_name in self.part_names:
+            print(f"gguf: loading model part '{part_name}'")
+            ctx: ContextManager[Any]
+            if self.is_safetensors:
+                from safetensors import safe_open
+                ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
+            else:
+                ctx = contextlib.nullcontext(torch.load(self.dir_model / part_name, map_location="cpu"))
+
+            with ctx as model_part:
+                for name in model_part.keys():
+                    data = model_part.get_tensor(name) if self.is_safetensors else model_part[name]
+                    yield name, data
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_block_count(self.hparams.get(
+            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
+        ))
+        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
+            self.gguf_writer.add_context_length(n_ctx)
+        if (n_embd := self.hparams.get("hidden_size")) is not None:
+            self.gguf_writer.add_embedding_length(n_embd)
+        if (n_ff := self.hparams.get("intermediate_size")) is not None:
+            self.gguf_writer.add_feed_forward_length(n_ff)
+        if (n_head := self.hparams.get("num_attention_head")) is not None:
+            self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+    def write(self):
+        self.write_tensors()
+        self.gguf_writer.write_header_to_file()
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.write_tensors_to_file()
+        self.gguf_writer.close()
+
+    def write_vocab(self):
+        self.gguf_writer.write_header_to_file()
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.close()
+
+    @staticmethod
+    def count_model_parts(dir_model: Path, prefix: str) -> int:
+        num_parts = 0
+        for filename in os.listdir(dir_model):
+            if filename.endswith(prefix):
+                num_parts += 1
+
+        return num_parts
+
+    @staticmethod
+    def load_hparams(dir_model):
+        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def from_model_architecture(model_architecture):
+        if model_architecture == "StableLMEpochForCausalLM":
+            return StableLMModel
+        if model_architecture == "GPTNeoXForCausalLM":
+            return GPTNeoXModel
+        if model_architecture == "BloomForCausalLM":
+            return BloomModel
+        if model_architecture == "MPTForCausalLM":
+            return MPTModel
+        if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return BaichuanModel
+        if model_architecture in ("FalconForCausalLM", "RWForCausalLM"):
+            return FalconModel
+        if model_architecture == "GPTBigCodeForCausalLM":
+            return StarCoderModel
+        if model_architecture == "GPTRefactForCausalLM":
+            return RefactModel
+        if model_architecture == "PersimmonForCausalLM":
+            return PersimmonModel
+        return Model
+
+    def _is_model_safetensors(self) -> bool:
+        return Model.count_model_parts(self.dir_model, ".safetensors") > 0
+
+    def _get_part_names(self):
+        if self.is_safetensors:
+            if self.num_parts == 1:  # there's only one .safetensors file
+                return ("model.safetensors",)
+            return (f"model-{n:05}-of-{self.num_parts:05}.safetensors" for n in range(1, self.num_parts + 1))
+
+        if self.num_parts == 1:  # there's only one .bin file
+            return ("pytorch_model.bin",)
+        return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
+
+    def _get_model_architecture(self) -> gguf.MODEL_ARCH:
+        arch = self.hparams["architectures"][0]
+        if arch == "GPTNeoXForCausalLM":
+            return gguf.MODEL_ARCH.GPTNEOX
+        if arch == "BloomForCausalLM":
+            return gguf.MODEL_ARCH.BLOOM
+        if arch == "MPTForCausalLM":
+            return gguf.MODEL_ARCH.MPT
+        if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return gguf.MODEL_ARCH.BAICHUAN
+        if arch == "FalconForCausalLM":
+            return gguf.MODEL_ARCH.FALCON
+        if arch == "GPTBigCodeForCausalLM":
+            return gguf.MODEL_ARCH.STARCODER
+        if arch == "GPTRefactForCausalLM":
+            return gguf.MODEL_ARCH.REFACT
+        if arch == "PersimmonForCausalLM":
+            return gguf.MODEL_ARCH.PERSIMMON
+
+        raise NotImplementedError(f'Architecture "{arch}" not supported!')
+
+    def _set_vocab_gpt2(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        tokenizer = AutoTokenizer.from_pretrained(dir_model)
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        assert max(tokenizer.vocab.values()) < vocab_size
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                pad_token = f"[PAD{i}]".encode('utf-8')
+                tokens.append(bytearray(pad_token))
+                toktypes.append(gguf.TokenType.USER_DEFINED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                if tokenizer.added_tokens_decoder[i].special:
+                    toktypes.append(gguf.TokenType.CONTROL)
+                else:
+                    toktypes.append(gguf.TokenType.USER_DEFINED)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _set_vocab_sentencepiece(self):
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        tokens: list[bytes] = []
+        scores: list[float] = []
+        toktypes: list[int] = []
+
+        if not tokenizer_path.is_file():
+            print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
+            sys.exit(1)
+
+        tokenizer = SentencePieceProcessor(str(tokenizer_path))
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        for token_id in range(vocab_size):
+            piece = tokenizer.id_to_piece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.get_score(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.is_unknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.is_control(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.is_unused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.is_byte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+
+                for key in added_tokens_json:
+                    tokens.append(key.encode("utf-8"))
+                    scores.append(-1000.0)
+                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+
+class StableLMModel(Model):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_dimension_count(
+            int(self.hparams["rope_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
+        )
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+
+
+class GPTNeoXModel(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(
+            int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
+        )
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+
+
+class BloomModel(Model):
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name("Bloom")
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
+        self.gguf_writer.add_embedding_length(n_embed)
+        self.gguf_writer.add_feed_forward_length(4 * n_embed)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        block_count = self.hparams["n_layer"]
+        tensors = dict(self.get_tensors())
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        has_lm_head = True
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        for name, data_torch in tensors.items():
+            if "lm_head.weight" not in tensors.keys() and "output.weight" not in tensors.keys():
+                has_lm_head = False
+
+            name = re.sub(r'transformer\.', '', name)
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+                # Map bloom-style qkv_linear to gpt-style qkv_linear
+                # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+                # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+                qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
+                data = np.concatenate(
+                    (
+                        qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.weight")
+            elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+                qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
+                data = np.concatenate(
+                    (
+                        qkv_bias[:, 0, :].reshape((n_embed,)),
+                        qkv_bias[:, 1, :].reshape((n_embed,)),
+                        qkv_bias[:, 2, :].reshape((n_embed,)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.bias")
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+            if not has_lm_head and name == "word_embeddings.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+                print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+
+class MPTModel(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layers"]
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"])
+        self.gguf_writer.add_head_count(self.hparams["n_heads"])
+        if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"):
+            self.gguf_writer.add_head_count_kv(kv_n_heads)
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+        if self.hparams["attn_config"]["clip_qkv"] is not None:
+            self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"])
+        self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers"))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+            # note: MPT output is tied to (same as) wte in original model;
+            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+            if new_name == "token_embd.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+
+
+class BaichuanModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        for i in range(block_count):
+            if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None:
+                print(f"Unpacking and permuting layer {i}")
+                model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \
+                    self._reverse_hf_permute_part(w, 0, head_count, head_count)
+                model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \
+                    self._reverse_hf_permute_part(w, 1, head_count, head_count_kv)
+                model_kv[f"model.layers.{i}.self_attn.v_proj.weight"] = \
+                    self._reverse_hf_part(w, 2)
+                del model_kv[f"model.layers.{i}.self_attn.W_pack.weight"]
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+    def _reverse_hf_permute_part(
+        self, weights: Tensor, n_part: int, n_head: int, n_head_kv: int | None = None,
+    ) -> Tensor:
+        r = weights.shape[0] // 3
+        return self._reverse_hf_permute(weights[r * n_part:r * n_part + r, ...], n_head, n_head_kv)
+
+    def _reverse_hf_part(self, weights: Tensor, n_part: int) -> Tensor:
+        r = weights.shape[0] // 3
+        return weights[r * n_part:r * n_part + r, ...]
+
+
+class FalconModel(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams.get("num_hidden_layers")
+        if block_count is None:
+            block_count = self.hparams["n_layer"]  # old name
+
+        n_head = self.hparams.get("num_attention_heads")
+        if n_head is None:
+            n_head = self.hparams["n_head"]  # old name
+
+        n_head_kv = self.hparams.get("num_kv_heads")
+        if n_head_kv is None:
+            n_head_kv = self.hparams.get("n_head_kv", 1)  # old name
+
+        self.gguf_writer.add_name("Falcon")
+        self.gguf_writer.add_context_length(2048)  # not in config.json
+        self.gguf_writer.add_tensor_data_layout("jploski")  # qkv tensor transform
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        block_count = self.hparams.get("num_hidden_layers")
+        if block_count is None:
+            block_count = self.hparams["n_layer"]  # old name
+
+        n_head = self.hparams.get("num_attention_heads")
+        if n_head is None:
+            n_head = self.hparams["n_head"]  # old name
+
+        n_head_kv = self.hparams.get("num_kv_heads")
+        if n_head_kv is None:
+            n_head_kv = self.hparams.get("n_head_kv", 1)  # old name
+
+        head_dim = self.hparams["hidden_size"] // n_head
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            # QKV tensor transform
+            # The original query_key_value tensor contains n_head_kv "kv groups",
+            # each consisting of n_head/n_head_kv query weights followed by one key
+            # and one value weight (shared by all query heads in the kv group).
+            # This layout makes it a big pain to work with in GGML.
+            # So we rearrange them here,, so that we have n_head query weights
+            # followed by n_head_kv key weights followed by n_head_kv value weights,
+            # in contiguous fashion.
+            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+
+            if "query_key_value" in name:
+                qkv = data_torch.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+                q = qkv[:, :-2].reshape(n_head * head_dim, head_dim * n_head)
+                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+class StarCoderModel(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
+        self.gguf_writer.add_name("StarCoder")
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(1)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+
+class RefactModel(Model):
+    def set_gguf_parameters(self):
+        hidden_dim = self.hparams["n_embd"]
+        inner_dim = 4 * hidden_dim
+        hidden_dim = int(2 * inner_dim / 3)
+        multiple_of = 256
+        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        block_count = self.hparams["n_layer"]
+
+        self.gguf_writer.add_name("Refact")
+        # refact uses Alibi. So this is from config.json which might be used by training.
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+
+        self.gguf_writer.add_feed_forward_length(ff_dim)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(1)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        hidden_dim = self.hparams["n_embd"]
+        inner_dim = 4 * hidden_dim
+        hidden_dim = int(2 * inner_dim / 3)
+        multiple_of = 256
+        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+        n_head = self.hparams["n_head"]
+        n_head_kv = 1
+        head_dim = self.hparams["n_embd"] // n_head
+        block_count = self.hparams["n_layer"]
+
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        tensors = dict(self.get_tensors())
+        for i in range(block_count):
+            if (w := tensors.get(f"transformer.h.{i}.attn.kv.weight")) is not None:
+                tensors[f"model.layers.{i}.self_attn.k_proj.weight"] = w[:n_head_kv * head_dim]
+                tensors[f"model.layers.{i}.self_attn.v_proj.weight"] = w[n_head_kv * head_dim:]
+                del tensors[f"transformer.h.{i}.attn.kv.weight"]
+            if (w := tensors.get(f"transformer.h.{i}.attn.q.weight")) is not None:
+                tensors[f"model.layers.{i}.self_attn.q_proj.weight"] = w
+                del tensors[f"transformer.h.{i}.attn.q.weight"]
+            if (w := tensors.get(f"transformer.h.{i}.mlp.gate_up_proj.weight")) is not None:
+                tensors[f"model.layers.{i}.mlp.gate_proj.weight"] = w[:ff_dim]
+                tensors[f"model.layers.{i}.mlp.up_proj.weight"] = w[ff_dim:]
+                del tensors[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+
+        for name, data_torch in tensors.items():
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+class PersimmonModel(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = head_count
+        hidden_size = self.hparams["hidden_size"]
+
+        self.gguf_writer.add_name('persimmon-8b-chat')
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+        # self.gguf_writer.add_bos_token_id(71013)
+        # self.gguf_writer.add_eos_token_id(71013)
+
+    def write_tensors(self):
+        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            if name.endswith(".self_attention.rotary_emb.inv_freq"):
+                continue
+            old_dtype = data_torch.dtype
+            # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
+            data = data_torch.to(torch.float32).squeeze().numpy()
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+            n_dims = len(data.shape)
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+###### CONVERSION LOGIC ######
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a huggingface model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16"], default="f16",
+        help="output format - use f32 for float32, f16 for float16",
+    )
+    parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file",
+    )
+
+    return parser.parse_args()
+
+
+args = parse_args()
+
+dir_model = args.model
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file=sys.stderr)
+    sys.exit(1)
+
+ftype_map = {
+    "f32": gguf.GGMLQuantizationType.F32,
+    "f16": gguf.GGMLQuantizationType.F16,
+}
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
+
+print(f"Loading model: {dir_model.name}")
+
+hparams = Model.load_hparams(dir_model)
+
+model_class = Model.from_model_architecture(hparams["architectures"][0])
+model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+
+print("Set model parameters")
+model_instance.set_gguf_parameters()
+
+print("Set model tokenizer")
+model_instance.set_vocab()
+
+if args.vocab_only:
+    print(f"Exporting model vocab to '{fname_out}'")
+    model_instance.write_vocab()
+else:
+    print(f"Exporting model to '{fname_out}'")
+    model_instance.write()
+
+print(f"Model successfully exported to '{fname_out}'")

convert-mpt-hf-to-gguf.py

@@ -1,227 +0,0 @@
-#!/usr/bin/env python3
-# HF mpt--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "MPTForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.MPT
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layers"]
-
-gguf_writer.add_name(dir_model.name)
-gguf_writer.add_context_length(hparams["max_seq_len"])
-gguf_writer.add_embedding_length(hparams["d_model"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
-gguf_writer.add_head_count(hparams["n_heads"])
-if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
-    gguf_writer.add_head_count_kv(kv_n_heads)
-gguf_writer.add_layer_norm_eps(1e-05)
-if hparams["attn_config"]["clip_qkv"] is not None:
-    gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
-gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but
-# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to
-# accomodate some "reserved" tokens; this is causing problems down the line in
-# llama.cpp, so we pad the vocab with dummy tokens:
-
-vocab_size = hparams["vocab_size"]
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-added_vocab = tokenizer.get_added_vocab()
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    if i not in reverse_vocab:
-        tokens.append(f"[PAD{i}]")
-        toktypes.append(gguf.TokenType.USER_DEFINED)
-    elif reverse_vocab[i] in added_vocab:
-        tokens.append(reverse_vocab[i])
-        if tokenizer.added_tokens_decoder[i].special:
-            toktypes.append(gguf.TokenType.CONTROL)
-        else:
-            toktypes.append(gguf.TokenType.USER_DEFINED)
-    else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Cannot map tensor '" + name + "'")
-            continue # for the sake of compatibility with some old published models, don't quit
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-        # note: MPT output is tied to (same as) wte in original model;
-        # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
-        if new_name == "token_embd.weight":
-            gguf_writer.add_tensor("output.weight", data)
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-refact-hf-to-gguf.py

@@ -1,272 +0,0 @@
-#!/usr/bin/env python3
-# HF refact--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import sys
-from pathlib import Path
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if "NO_LOCAL_GGUF" not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
-import gguf
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Convert a Refact model to a GGML compatible file"
-    )
-    parser.add_argument(
-        "--vocab-only",
-        action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype",
-        type=int,
-        choices=[0, 1],
-        default=1,
-        nargs="?",
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f"Error: {args.model} is not a directory", file=sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf"
-
-print("gguf: loading model " + dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTRefactForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH = gguf.MODEL_ARCH.REFACT
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-# Get refact feed forward dimension
-hidden_dim = hparams["n_embd"]
-inner_dim = 4 * hidden_dim
-hidden_dim = int(2 * inner_dim / 3)
-multiple_of = 256
-ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("Refact")
-# refact uses Alibi. So this is from config.json which might be used by training.
-gguf_writer.add_context_length(hparams["n_positions"])
-gguf_writer.add_embedding_length(hparams["n_embd"])
-
-gguf_writer.add_feed_forward_length(ff_dim)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-gguf_writer.add_head_count_kv(1)
-gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-added_vocab = tokenizer.get_added_vocab()
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    if i not in reverse_vocab:
-        tokens.append(f"[PAD{i}]")
-        toktypes.append(gguf.TokenType.USER_DEFINED)
-    elif reverse_vocab[i] in added_vocab:
-        tokens.append(reverse_vocab[i])
-        if tokenizer.added_tokens_decoder[i].special:
-            toktypes.append(gguf.TokenType.CONTROL)
-        else:
-            toktypes.append(gguf.TokenType.USER_DEFINED)
-    else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
-
-# params for qkv transform
-n_head = hparams["n_head"]
-n_head_kv = 1
-
-head_dim = hparams["n_embd"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    for i in range(block_count):
-        if f"transformer.h.{i}.attn.kv.weight" in model_part:
-            data = model_part[f"transformer.h.{i}.attn.kv.weight"]
-            model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[
-                : n_head_kv * head_dim
-            ]
-            model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[
-                n_head_kv * head_dim :
-            ]
-            del model_part[f"transformer.h.{i}.attn.kv.weight"]
-        if f"transformer.h.{i}.attn.q.weight" in model_part:
-            model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[
-                f"transformer.h.{i}.attn.q.weight"
-            ]
-            del model_part[f"transformer.h.{i}.attn.q.weight"]
-        if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part:
-            data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
-            model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim]
-            model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:]
-            del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if (
-            ftype == 1
-            and data_dtype == np.float32
-            and name.endswith(".weight")
-            and n_dims == 2
-        ):
-            data = data.astype(np.float16)
-
-        print(
-            new_name
-            + ", n_dims = "
-            + str(n_dims)
-            + ", "
-            + str(old_dtype)
-            + " --> "
-            + str(data.dtype)
-        )
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-starcoder-hf-to-gguf.py

@@ -1,210 +0,0 @@
-#!/usr/bin/env python3
-# HF starcoder --> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    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 (*.bin)")
-    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.STARCODER
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("StarCoder")
-gguf_writer.add_context_length(hparams["n_positions"])
-gguf_writer.add_embedding_length(hparams["n_embd"])
-gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-gguf_writer.add_head_count_kv(1)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-added_vocab = tokenizer.get_added_vocab()
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    if i not in reverse_vocab:
-        tokens.append(f"[PAD{i}]")
-        toktypes.append(gguf.TokenType.USER_DEFINED)
-    elif reverse_vocab[i] in added_vocab:
-        tokens.append(reverse_vocab[i])
-        if tokenizer.added_tokens_decoder[i].special:
-            toktypes.append(gguf.TokenType.CONTROL)
-        else:
-            toktypes.append(gguf.TokenType.USER_DEFINED)
-    else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# params for qkv transform
-n_head    = hparams["n_head"]
-n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
-
-head_dim = hparams["n_embd"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert.py

@@ -26,7 +26,7 @@ from pathlib import Path
 from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
 
 import numpy as np
-from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+from sentencepiece import SentencePieceProcessor
 
 import os
 if 'NO_LOCAL_GGUF' not in os.environ:
@@ -151,8 +151,11 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
+    rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
     f_rope_scale: float | None = None
+    n_orig_ctx: int | None = None
+    rope_finetuned: bool | None = None
 
     ftype: GGMLFileType | None = None
 
@@ -198,20 +201,20 @@ class Params:
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab          = config["vocab_size"]
-        n_embd           = config["hidden_size"]
-        n_layer          = config["num_hidden_layers"]
-        n_ff             = config["intermediate_size"]
-        n_head           = config["num_attention_heads"]
-        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
-        f_norm_eps       = config["rms_norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
-
+        rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
-        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
-            f_rope_scale = config["rope_scaling"].get("factor")
-        else:
-            f_rope_scale = None
+
+        if rope_scaling is not None and (typ := rope_scaling.get("type")):
+            rope_factor = rope_scaling.get("factor")
+            f_rope_scale = rope_factor
+            if typ == "linear":
+                rope_scaling_type = gguf.RopeScalingType.LINEAR
+            elif typ == "yarn":
+                rope_scaling_type = gguf.RopeScalingType.YARN
+                n_orig_ctx = rope_scaling['original_max_position_embeddings']
+                rope_finetuned = rope_scaling['finetuned']
+            else:
+                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
@@ -222,16 +225,19 @@ class Params:
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
         return Params(
-            n_vocab          = n_vocab,
-            n_embd           = n_embd,
-            n_layer          = n_layer,
-            n_ctx            = n_ctx,
-            n_ff             = n_ff,
-            n_head           = n_head,
-            n_head_kv        = n_head_kv,
-            f_norm_eps       = f_norm_eps,
-            f_rope_freq_base = f_rope_freq_base,
-            f_rope_scale     = f_rope_scale,
+            n_vocab           = config["vocab_size"],
+            n_embd            = config["hidden_size"],
+            n_layer           = config["num_hidden_layers"],
+            n_ctx             = n_ctx,
+            n_ff              = config["intermediate_size"],
+            n_head            = (n_head := config["num_attention_heads"]),
+            n_head_kv         = config.get("num_key_value_heads", n_head),
+            f_norm_eps        = config["rms_norm_eps"],
+            f_rope_freq_base  = config.get("rope_theta"),
+            rope_scaling_type = rope_scaling_type,
+            f_rope_scale      = f_rope_scale,
+            n_orig_ctx        = n_orig_ctx,
+            rope_finetuned    = rope_finetuned,
         )
 
     # LLaMA v2 70B params.json
@@ -240,17 +246,8 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
-        n_embd           = config["dim"]
-        n_layer          = config["n_layers"]
-        n_ff             = -1
-        n_head           = config["n_heads"]
-        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
-        f_norm_eps       = config["norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
-
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if f_rope_freq_base == 1000000:
+        if config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -260,22 +257,16 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
-        if n_vocab == -1:
-            n_vocab = model["tok_embeddings.weight"].shape[0]
-
-        if n_ff == -1:
-            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
-
         return Params(
-            n_vocab          = n_vocab,
-            n_embd           = n_embd,
-            n_layer          = n_layer,
+            n_vocab          = config.get("vocab_size", model["tok_embeddings.weight"].shape[0]),
+            n_embd           = config["dim"],
+            n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
-            n_ff             = n_ff,
-            n_head           = n_head,
-            n_head_kv        = n_head_kv,
-            f_norm_eps       = f_norm_eps,
-            f_rope_freq_base = f_rope_freq_base,
+            n_ff             = model["layers.0.feed_forward.w1.weight"].shape[0],
+            n_head           = (n_head := config["n_heads"]),
+            n_head_kv        = config.get("n_kv_heads", n_head),
+            f_norm_eps       = config["norm_eps"],
+            f_rope_freq_base = config.get("rope_theta"),
         )
 
     @staticmethod
@@ -337,7 +328,7 @@ class BpeVocab:
 
     def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.bpe_tokenizer
-        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
+        from transformers.models.gpt2 import tokenization_gpt2
         reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
 
         for i, _ in enumerate(tokenizer):
@@ -831,8 +822,16 @@ class OutputFile:
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
-        if params.f_rope_scale is not None:
-            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+        if params.rope_scaling_type:
+            assert params.f_rope_scale is not None
+            self.gguf.add_rope_scaling_type(params.rope_scaling_type)
+            self.gguf.add_rope_scaling_factor(params.f_rope_scale)
+
+        if params.n_orig_ctx is not None:
+            self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx)
+
+        if params.rope_finetuned is not None:
+            self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
 
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)

examples/batched-bench/batched-bench.cpp

@@ -185,7 +185,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_tokens_rm(ctx, -1, -1);
+                llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);

examples/benchmark/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/benchmark/benchmark-matmult.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 

examples/embedding/CMakeLists.txt

@@ -3,6 +3,3 @@ 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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/embedding/embedding.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 

examples/finetune/finetune.cpp

@@ -642,8 +642,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx,
-            rope_freq_base, rope_freq_scale);
+            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+            rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        );
     };
 
     set_name(tokens_input, "tokens_input");
@@ -652,7 +653,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
 
     auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-        if (ggml_is_quantized(a->type)) {
+        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);
@@ -1459,6 +1460,17 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_w3 = std::stoi(argv[i]);
             params->custom_n_rank_w3 = true;
+        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params->common.n_gpu_layers = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1545,6 +1557,7 @@ int main(int argc, char ** argv) {
     srand(params.common.seed);
 
     struct llama_model_params llama_mparams = llama_model_default_params();
+    llama_mparams.n_gpu_layers = params.common.n_gpu_layers;
     llama_mparams.vocab_only = false;
 
     printf("%s: model base = '%s'\n", __func__, params.fn_model_base);

examples/finetune/finetune.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+cd `dirname $0`
+cd ../..
+
+EXE="./finetune"
+
+if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi
+if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi
+
+# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses.
+MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing.
+
+while getopts "dg" opt; do
+  case $opt in
+    d)
+      DEBUGGER="gdb --args"
+      ;;
+    g)
+      EXE="./build/bin/Release/finetune"
+      GPUARG="--gpu-layers 25"
+      ;;
+  esac
+done
+
+$DEBUGGER $EXE \
+        --model-base $MODEL \
+        $GPUARG \
+        --checkpoint-in  chk-ol3b-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \
+        --lora-out lora-ol3b-shakespeare-ITERATION.bin \
+        --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \
+        --save-every 10 \
+        --threads 10 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing

examples/infill/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} infill.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)
-    add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/infill/infill.cpp

@@ -2,7 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"
 
 #include <cassert>
@@ -184,8 +183,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);

examples/llama-bench/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} llama-bench.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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/llama-bench/llama-bench.cpp

@@ -19,7 +19,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "build-info.h"
 #include "ggml-cuda.h"
 
 // utils
@@ -641,8 +640,8 @@ struct test {
     }
 };
 
-const std::string test::build_commit = BUILD_COMMIT;
-const int         test::build_number = BUILD_NUMBER;
+const std::string test::build_commit = LLAMA_COMMIT;
+const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::metal        = !!ggml_cpu_has_metal();
@@ -1037,7 +1036,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1048,7 +1047,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_tokens_rm(ctx, -1, -1);
+            llama_kv_cache_clear(ctx);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llava/CMakeLists.txt

@@ -1,20 +1,36 @@
-set(TARGET clip)
-add_library(${TARGET} clip.cpp clip.h)
-install(TARGETS ${TARGET} LIBRARY)
-target_link_libraries(${TARGET} PRIVATE common ggml ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if (NOT MSVC)
-    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
-    endif()
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
+add_library(llava OBJECT
+            llava.cpp
+            llava.h
+            clip.cpp
+            clip.h
+            )
+
+target_link_libraries(llava PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(llava PUBLIC .)
+target_include_directories(llava PUBLIC ../..)
+target_include_directories(llava PUBLIC ../../common)
+
+target_compile_features(llava PRIVATE cxx_std_11)
+
+add_library(llava_static STATIC $<TARGET_OBJECTS:llava>)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(llava PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(llava PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(llava_shared SHARED $<TARGET_OBJECTS:llava>)
+    target_link_libraries(llava_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS llava_shared LIBRARY)
 endif()
 
-set(TARGET llava)
-add_executable(${TARGET} llava.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if (NOT MSVC)
+    target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
+    endif()
 if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
+    add_dependencies(llava BUILD_INFO)
 endif()
+
+set(TARGET llava-cli)
+add_executable(llava-cli llava-cli.cpp)
+install(TARGETS llava-cli RUNTIME)
+target_link_libraries(llava-cli PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(llava PRIVATE cxx_std_11)

examples/llava/README.md

@@ -9,12 +9,12 @@ models are available.
 After API is confirmed, more models will be supported / uploaded.
 
 ## Usage
-Build with cmake or run `make llava` to build it.
+Build with cmake or run `make llava-cli` to build it.
 
-After building, run: `./llava` to see the usage. For example:
+After building, run: `./llava-cli` to see the usage. For example:
 
 ```sh
-./llava -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+./llava-cli -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
@@ -51,7 +51,6 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 
 ## TODO
 
-- [ ] Support server mode.
 - [ ] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.cpp

@@ -680,26 +680,44 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     return new_clip;
 }
 
-clip_image_u8 * make_clip_image_u8() { return new clip_image_u8(); }
-
+clip_image_u8 * make_clip_image_u8() {
+    auto img = new clip_image_u8();
+    return img;
+}
 clip_image_f32 * make_clip_image_f32() { return new clip_image_f32(); }
 
-bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
-    int nx, ny, nc;
-    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
-    if (!data) {
-        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname);
-        return false;
-    }
+void clip_image_u8_free(clip_image_u8 * img) { if (img->data) { delete[] img->data; } delete img; }
+void clip_image_f32_free(clip_image_f32 * img) { if (img->data) { delete[] img->data; } delete img; }
 
+static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
     img->nx = nx;
     img->ny = ny;
     img->size = nx * ny * 3;
     img->data = new uint8_t[img->size]();
     memcpy(img->data, data, img->size);
+}
 
+bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
+    if (!data) {
+        fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
     stbi_image_free(data);
+    return true;
+}
 
+bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    if (!data) {
+        fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
     return true;
 }
 
@@ -714,39 +732,40 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
 
-    clip_image_u8 temp; // we will keep the input image data here temporarily
+    clip_image_u8 * temp = make_clip_image_u8(); // we will keep the input image data here temporarily
     if (pad2square && img->nx != img->ny) {
         int longer_side = std::max(img->nx, img->ny);
-        temp.nx = longer_side;
-        temp.ny = longer_side;
-        temp.size = 3 * longer_side * longer_side;
-        temp.data = new uint8_t[temp.size]();
+        temp->nx = longer_side;
+        temp->ny = longer_side;
+        temp->size = 3 * longer_side * longer_side;
+        temp->data = new uint8_t[temp->size]();
         uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
 
         // fill with background color
-        for (size_t i = 0; i < temp.size; i++) {
-            temp.data[i] = bc[i % 3];
+        for (size_t i = 0; i < temp->size; i++) {
+            temp->data[i] = bc[i % 3];
         }
 
         // copy from the input image
         for (int y = 0; y < img->ny; y++) {
             for (int x = 0; x < img->nx; x++) {
                 const int i = 3 * (y * img->nx + x);
-                const int j = 3 * (y * temp.nx + x);
-                temp.data[j] = img->data[i];
-                temp.data[j+1] = img->data[i+1];
-                temp.data[j+2] = img->data[i+2];
+                const int j = 3 * (y * temp->nx + x);
+                temp->data[j] = img->data[i];
+                temp->data[j+1] = img->data[i+1];
+                temp->data[j+2] = img->data[i+2];
             }
         }
     } else {
-        temp.nx   = img->nx;
-        temp.ny   = img->ny;
-        temp.size = img->size;
-        temp.data = img->data;
+        temp->nx   = img->nx;
+        temp->ny   = img->ny;
+        temp->size = img->size;
+        temp->data = new uint8_t[temp->size]();
+        *temp->data = *img->data; // copy
     }
 
-    const int nx = temp.nx;
-    const int ny = temp.ny;
+    const int nx = temp->nx;
+    const int ny = temp->ny;
 
     const int nx2 = ctx->vision_model.hparams.image_size;
     const int ny2 = ctx->vision_model.hparams.image_size;
@@ -785,10 +804,10 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
                 const int j10 = 3 * (y1 * nx + x0) + c;
                 const int j11 = 3 * (y1 * nx + x1) + c;
 
-                const float v00 = temp.data[j00];
-                const float v01 = temp.data[j01];
-                const float v10 = temp.data[j10];
-                const float v11 = temp.data[j11];
+                const float v00 = temp->data[j00];
+                const float v01 = temp->data[j01];
+                const float v10 = temp->data[j10];
+                const float v11 = temp->data[j11];
 
                 const float v0 = v00 * (1.0f - dx) + v01 * dx;
                 const float v1 = v10 * (1.0f - dx) + v11 * dx;
@@ -803,6 +822,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
             }
         }
     }
+    clip_image_u8_free(temp);
 
     return true;
 }
@@ -1049,16 +1069,16 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     return true;
 }
 
-int clip_n_mmproj_embd(struct clip_ctx * ctx) {
+int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     return ctx->vision_model.mm_2_b->ne[0];
 }
 
-int clip_n_patches(struct clip_ctx * ctx) {
+int clip_n_patches(const struct clip_ctx * ctx) {
     auto & params = ctx->vision_model.hparams;
 
     return (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 }
 
-size_t clip_embd_nbytes(struct clip_ctx * ctx) {
+size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
     return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }

examples/llava/clip.h

@@ -1,7 +1,22 @@
 #ifndef CLIP_H
 #define CLIP_H
 
-#include "ggml.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
 
 struct clip_ctx;
 
@@ -20,19 +35,20 @@ struct clip_vision_hparams {
     float eps;
 };
 
-struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
+/** load mmproj model */
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
+/** free mmproj model */
+CLIP_API void clip_free(struct clip_ctx * ctx);
 
-void clip_free(struct clip_ctx * ctx);
-
-size_t clip_embd_nbytes(struct clip_ctx * ctx);
-int clip_n_patches(struct clip_ctx * ctx);
-int clip_n_mmproj_embd(struct clip_ctx * ctx);
+size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+int clip_n_patches(const struct clip_ctx * ctx);
+int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 // RGB uint8 image
 struct clip_image_u8 {
     int nx;
     int ny;
-    uint8_t * data;
+    uint8_t * data = NULL;
     size_t size;
 };
 
@@ -41,7 +57,7 @@ struct clip_image_u8 {
 struct clip_image_f32 {
     int nx;
     int ny;
-    float * data;
+    float * data = NULL;
     size_t size;
 };
 
@@ -57,7 +73,12 @@ struct clip_image_f32_batch {
 
 struct clip_image_u8 * make_clip_image_u8();
 struct clip_image_f32 * make_clip_image_f32();
-bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+CLIP_API void clip_image_u8_free(clip_image_u8 * img);
+CLIP_API void clip_image_f32_free(clip_image_f32 * img);
+CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
 bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square);
 bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec);
 

examples/llava/llava-cli.cpp

@@ -0,0 +1,313 @@
+#include "ggml.h"
+#include "common.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+
+#include "base64.hpp"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+    return true;
+}
+
+// TODO: use common/sampling.h
+static llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
+    auto & sparams = params.sparams;
+
+    // out of user input, sample next token
+    const float   temp      = sparams.temp;
+    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
+    const float   top_p     = sparams.top_p;
+    const float   tfs_z     = sparams.tfs_z;
+    const float   typical_p = sparams.typical_p;
+    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
+    // const float   repeat_penalty  = sparams.repeat_penalty;
+    // const float   alpha_presence  = sparams.presence_penalty;
+    // const float   alpha_frequency = sparams.frequency_penalty;
+    const int     mirostat     = sparams.mirostat;
+    const float   mirostat_tau = sparams.mirostat_tau;
+    const float   mirostat_eta = sparams.mirostat_eta;
+    // const bool    penalize_nl     = sparams.penalize_nl;
+
+    llama_token id = 0;
+    {
+        auto logits  = llama_get_logits(ctx_llama);
+        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
+
+        // Apply params.logit_bias map
+        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
+            logits[it->first] += it->second;
+        }
+
+        std::vector<llama_token_data> candidates;
+        candidates.reserve(n_vocab);
+        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        }
+
+        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+        if (temp <= 0) {
+              // Greedy sampling
+            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
+        } else {
+            if (mirostat == 1) {
+                static float mirostat_mu = 2.0f * mirostat_tau;
+                const  int mirostat_m    = 100;
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+            } else if (mirostat == 2) {
+                static float mirostat_mu = 2.0f * mirostat_tau;
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+            } else {
+                  // Temperature sampling
+                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
+                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
+                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
+                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token(ctx_llama, &candidates_p);
+            }
+        }
+    }
+
+    return id;
+}
+
+static const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
+    int id = sample_id(ctx_llama, params);
+    static std::string ret;
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        fprintf(stderr, "%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        fprintf(stderr, "%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            printf("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        if (!embed) {
+            fprintf(stderr, "%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
+        if (!embed) {
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) {
+    int n_past = 0;
+
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+
+    // llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
+    eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, true);
+    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
+    eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);
+
+    // generate the response
+
+    printf("\n");
+
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(ctx_llava->ctx_llama, *params, &n_past);
+        if (strcmp(tmp, "</s>") == 0) break;
+
+        printf("%s", tmp);
+        fflush(stdout);
+    }
+
+    printf("\n");
+}
+
+
+static struct llava_context * llava_init(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+
+    llama_backend_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        gpt_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    auto ctx_llava = llava_init(&params);
+    if (ctx_llava == NULL) {
+        fprintf(stderr, "%s: error: failed to init llava\n", __func__);
+        return 1;
+    }
+
+    auto image_embed = load_image(ctx_llava, &params);
+
+    // process the prompt
+    process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+    llama_print_timings(ctx_llava->ctx_llama);
+
+    llava_image_embed_free(image_embed);
+    llava_free(ctx_llava);
+    return 0;
+}

examples/llava/llava-utils.h

@@ -1,147 +0,0 @@
-#pragma once
-
-// this one and clip lib will be eventually merged to a single lib, let's keep it this way for now
-
-#include "common.h"
-#include "llama.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <vector>
-
-inline bool eval_image_embd(llama_context * ctx_llama, float * embd, int N, int n_batch, int * n_past) {
-    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
-
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = N - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
-        if (llama_decode(ctx_llama, batch)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-inline bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
-    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-    return true;
-}
-
-// TODO: use common/sampling.h
-inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
-    auto & sparams = params.sparams;
-
-    // out of user input, sample next token
-    const float   temp      = sparams.temp;
-    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
-    const float   top_p     = sparams.top_p;
-    const float   tfs_z     = sparams.tfs_z;
-    const float   typical_p = sparams.typical_p;
-    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
-    // const float   repeat_penalty  = sparams.repeat_penalty;
-    // const float   alpha_presence  = sparams.presence_penalty;
-    // const float   alpha_frequency = sparams.frequency_penalty;
-    const int     mirostat     = sparams.mirostat;
-    const float   mirostat_tau = sparams.mirostat_tau;
-    const float   mirostat_eta = sparams.mirostat_eta;
-    // const bool    penalize_nl     = sparams.penalize_nl;
-
-    llama_token id = 0;
-    {
-        auto logits  = llama_get_logits(ctx_llama);
-        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
-
-        // Apply params.logit_bias map
-        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
-            logits[it->first] += it->second;
-        }
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        // TODO: Apply penalties
-        // float nl_logit = logits[llama_token_nl(ctx)];
-        // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-        // llama_sample_repetition_penalty(ctx, &candidates_p,
-        //      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-        //      last_n_repeat, repeat_penalty);
-        // llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-        // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-        // last_n_repeat, alpha_frequency, alpha_presence);
-        // if (!penalize_nl) {
-        //     logits[llama_token_nl(ctx)] = nl_logit;
-        // }
-
-        if (temp <= 0) {
-              // Greedy sampling
-            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
-        } else {
-            if (mirostat == 1) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                const  int mirostat_m    = 100;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-            } else if (mirostat == 2) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-            } else {
-                  // Temperature sampling
-                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
-                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
-                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
-                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token(ctx_llama, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
-    int id = sample_id(ctx_llama, params);
-    static std::string ret;
-    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
-        ret = "</s>";
-    } else {
-        ret = llama_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}

examples/llava/llava.cpp

@@ -1,164 +1,156 @@
 #include "clip.h"
-#include "llava-utils.h"
 #include "common.h"
 #include "llama.h"
+#include "llava.h"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
 
-static void show_additional_info(int /*argc*/, char ** argv) {
-    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
+#include "base64.hpp"
 
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        show_additional_info(argc, argv);
-        return 1;
+static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
+    clip_image_f32 * img_res = make_clip_image_f32();
+    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+        fprintf(stderr, "%s: unable to preprocess image\n", __func__);
+        clip_image_f32_free(img_res);
+        return false;
     }
 
-    if (params.mmproj.empty() || params.image.empty()) {
-        gpt_print_usage(argc, argv, params);
-        show_additional_info(argc, argv);
-        return 1;
-    }
-
-    const char * clip_path = params.mmproj.c_str();
-    const char * img_path = params.image.c_str();
-
-    if (params.prompt.empty()) {
-        params.prompt = "describe the image in detail.";
-    }
-
-    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
-
-    // load and preprocess the image
-    clip_image_u8 img;
-    clip_image_f32 img_res;
-
-    if (!clip_image_load_from_file(img_path, &img)) {
-        fprintf(stderr, "%s: is %s really an image file?\n", __func__, img_path);
-
-        clip_free(ctx_clip);
-        return 1;
-    }
-
-    if (!clip_image_preprocess(ctx_clip, &img, &img_res, /*pad2square =*/ true)) {
-        fprintf(stderr, "%s: unable to preprocess %s\n", __func__, img_path);
-
-        clip_free(ctx_clip);
-        return 1;
-    }
-
-    int n_img_pos  = clip_n_patches(ctx_clip);
-    int n_img_embd = clip_n_mmproj_embd(ctx_clip);
-
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
-
-    if (!image_embd) {
-        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
-
-        return 1;
-    }
+    *n_img_pos = clip_n_patches(ctx_clip);
 
     const int64_t t_img_enc_start_us = ggml_time_us();
-    if (!clip_image_encode(ctx_clip, params.n_threads, &img_res, image_embd)) {
+    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+    clip_image_f32_free(img_res);
+    if (!encoded) {
         fprintf(stderr, "Unable to encode image\n");
 
-        return 1;
+        return false;
     }
+
     const int64_t t_img_enc_end_us = ggml_time_us();
+    float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
-    // we get the embeddings, free up the memory required for CLIP
-    clip_free(ctx_clip);
+    printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
 
-    llama_backend_init(params.numa);
-
-    llama_model_params model_params              = llama_model_default_params();
-                       model_params.n_gpu_layers = params.n_gpu_layers;
-                       model_params.main_gpu     = params.main_gpu;
-                       model_params.tensor_split = params.tensor_split;
-                       model_params.use_mmap     = params.use_mmap;
-                       model_params.use_mlock    = params.use_mlock;
-
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
-    if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
-        return 1;
-    }
-
-    llama_context_params ctx_params = llama_context_default_params();
-
-    ctx_params.n_ctx           = params.n_ctx < 2048 ? 2048 : params.n_ctx; // we need a longer context size to process image embeddings
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-    ctx_params.seed            = params.seed;
-
-    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
-        return 1;
-    }
-
-    // make sure that the correct mmproj was used, i.e., compare apples to apples
-    const int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
-
-    if (n_img_embd != n_llama_embd) {
-        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_img_embd, n_llama_embd);
-
-        llama_free(ctx_llama);
-        llama_free_model(model);
-        llama_backend_free();
-        free(image_embd);
-
-        return 1;
-    }
-
-    // process the prompt
-    // llava chat format is "<system_prompt>USER: <image_embeddings>\n<textual_prompt>\nASSISTANT:"
-
-    int n_past = 0;
-
-    const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-
-    eval_string(ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params.n_batch, &n_past, true);
-    eval_image_embd(ctx_llama, image_embd, n_img_pos, params.n_batch, &n_past);
-    eval_string(ctx_llama, (params.prompt + "\nASSISTANT:").c_str(), params.n_batch, &n_past, false);
-
-    // generate the response
-
-    printf("\n");
-    printf("prompt: '%s'\n", params.prompt.c_str());
-    printf("\n");
-
-    for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_llama, params, &n_past);
-        if (strcmp(tmp, "</s>") == 0) break;
-
-        printf("%s", tmp);
-        fflush(stdout);
-    }
-
-    printf("\n");
-
-    {
-        const float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
-
-        printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / n_img_pos);
-    }
-
-    llama_print_timings(ctx_llama);
-
-    llama_free(ctx_llama);
-    llama_free_model(model);
-    llama_backend_free();
-    free(image_embd);
-
-    return 0;
+    return true;
+}
+
+bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
+        // make sure that the correct mmproj was used, i.e., compare apples to apples
+    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+    auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
+    if (n_image_embd != n_llama_embd) {
+        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        return false;
+    }
+    return true;
+}
+
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
+    if (!image_embd) {
+        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
+        free(image_embd);
+        return false;
+    }
+
+    int n_img_pos;
+    if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
+        fprintf(stderr, "%s: cannot encode image, aborting\n", __func__);
+        free(image_embd);
+        return false;
+    }
+    *image_embd_out = image_embd;
+    *n_img_pos_out = n_img_pos;
+
+    return true;
+}
+
+bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+
+    for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
+        int n_eval = image_embed->n_image_pos - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        if (llama_decode(ctx_llama, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+    clip_image_u8 * img = make_clip_image_u8();
+    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
+        clip_image_u8_free(img);
+        fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);
+        return NULL;
+    }
+
+    float* image_embed = NULL;
+    int n_image_pos = 0;
+    bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
+    if (!image_embed_result) {
+        clip_image_u8_free(img);
+        fprintf(stderr, "%s: coulnd't embed the image\n", __func__);
+        return NULL;
+    }
+
+    clip_image_u8_free(img);
+    auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    result->embed = image_embed;
+    result->n_image_pos = n_image_pos;
+    return result;
+}
+
+static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
+    auto file = fopen(path, "rb");
+    if (file == NULL) {
+        fprintf(stderr, "%s: can't read file %s\n", __func__, path);
+        return false;
+    }
+
+    fseek(file, 0, SEEK_END);
+    auto fileSize = ftell(file);
+    fseek(file, 0, SEEK_SET);
+
+    auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
+    if (buffer == NULL) {
+        fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        perror("Memory allocation error");
+        fclose(file);
+        return false;
+    }
+    fread(buffer, 1, fileSize, file); // Read the file into the buffer
+    fclose(file); // Close the file
+
+    *bytesOut = buffer;
+    *sizeOut = fileSize;
+    return true;
+}
+
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+    unsigned char* image_bytes;
+    long image_bytes_length;
+    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
+    if (!loaded) {
+        fprintf(stderr, "%s: failed to load %s\n", __func__, image_path);
+        return NULL;
+    }
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    free(image_bytes);
+
+    return embed;
+}
+
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+    free(embed->embed);
+    free(embed);
 }

examples/llava/llava.h

@@ -0,0 +1,50 @@
+#ifndef LLAVA_H
+#define LLAVA_H
+
+#include "ggml.h"
+
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAVA_API __declspec(dllexport)
+#        else
+#            define LLAVA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAVA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAVA_API
+#endif
+
+struct clip_ctx;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct llava_image_embed {
+    float * embed;
+    int n_image_pos;
+};
+
+/** sanity check for clip <-> llava embed size match */
+LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
+
+/** build an image embed from image file bytes */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
+/** build an image embed from a path to an image filename */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
+/** free an embedding made with llava_image_embed_make_* */
+
+/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
+LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

examples/main/CMakeLists.txt

@@ -3,6 +3,3 @@ 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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/main/README.md

@@ -208,6 +208,14 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho
 
 Example usage: `--top-p 0.95`
 
+### Min P Sampling
+
+-   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05).
+
+The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.
+
+Example usage: `--min-p 0.05`
+
 ### Tail Free Sampling (TFS)
 
 -   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).

examples/main/main.cpp

@@ -2,7 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -153,8 +152,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -298,7 +297,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_tokens_rm(ctx, n_matching_session_tokens, -1);
+        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOGLN(

examples/parallel/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} parallel.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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/parallel/parallel.cpp

@@ -1,8 +1,6 @@
 // A basic application simulating a server with multiple clients.
 // The clients submite requests to the server and they are processed in parallel.
 
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 

examples/perplexity/CMakeLists.txt

@@ -3,6 +3,3 @@ 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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/perplexity/perplexity.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -210,7 +209,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -339,7 +338,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -573,7 +572,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {

examples/quantize-stats/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize-stats/quantize-stats.cpp

@@ -1,5 +1,4 @@
 #define LLAMA_API_INTERNAL
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"

examples/quantize/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/quantize/quantize.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -18,7 +17,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
     { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
-#ifdef GGML_USE_K_QUANTS
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
@@ -31,7 +29,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
     { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
-#endif
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
@@ -70,13 +67,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
@@ -103,6 +101,8 @@ int main(int argc, char ** argv) {
             params.quantize_output_tensor = false;
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
+        } else if (strcmp(argv[arg_idx], "--pure") == 0) {
+            params.pure = true;
         } else {
             usage(argv[0]);
         }

examples/save-load-state/CMakeLists.txt

@@ -3,6 +3,3 @@ 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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/save-load-state/save-load-state.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 

examples/server/CMakeLists.txt

@@ -6,11 +6,8 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/server/README.md

@@ -7,7 +7,7 @@ Command line options:
 -   `--threads N`, `-t N`: Set the number of threads to use during generation.
 -   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
 -   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
--   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
+-   `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 -   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
 -   `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
@@ -122,6 +122,8 @@ node index.js
 
     `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
 
+    `min_p`: The minimum probability for a token to be considered, relative to the probability of the most likely token (default: 0.05).
+
     `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
 
     `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded.

examples/server/public/index.html

@@ -219,6 +219,7 @@
       repeat_penalty: 1.18, // 1.0 = disabled
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.5, // 1.0 = disabled
+      min_p: 0.05, // 0 = disabled
       tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
@@ -744,6 +745,7 @@
             ${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 })}
+            ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })}
           </fieldset>
           <details>
             <summary>More options</summary>

examples/server/server.cpp

@@ -1,6 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"
 
 #include "../llava/clip.h"
@@ -149,6 +148,7 @@ struct task_server {
     task_type type;
     json data;
     bool infill_mode = false;
+    bool embedding_mode = false;
 };
 
 struct task_result {
@@ -371,6 +371,7 @@ struct llama_client_slot
     std::vector<completion_token_output> generated_token_probs;
 
     bool infill = false;
+    bool embedding = false;
     bool has_next_token = true;
     bool truncated = false;
     bool stopped_eos = false;
@@ -678,6 +679,7 @@ struct llama_server_context
         slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
         slot->sparams.top_k           = json_value(data, "top_k",             default_sparams.top_k);
         slot->sparams.top_p           = json_value(data, "top_p",             default_sparams.top_p);
+        slot->sparams.min_p           = json_value(data, "min_p",             default_sparams.min_p);
         slot->sparams.tfs_z           = json_value(data, "tfs_z",             default_sparams.tfs_z);
         slot->sparams.typical_p       = json_value(data, "typical_p",         default_sparams.typical_p);
         slot->sparams.temp            = json_value(data, "temperature",       default_sparams.temp);
@@ -857,7 +859,7 @@ struct llama_server_context
 
     void kv_cache_clear() {
         // clear the entire KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
         clean_kv_cache = false;
     }
 
@@ -1112,6 +1114,7 @@ struct llama_server_context
             {"temp",              slot.sparams.temp},
             {"top_k",             slot.sparams.top_k},
             {"top_p",             slot.sparams.top_p},
+            {"min_p",             slot.sparams.min_p},
             {"tfs_z",             slot.sparams.tfs_z},
             {"typical_p",         slot.sparams.typical_p},
             {"repeat_last_n",     slot.sparams.penalty_last_n},
@@ -1244,13 +1247,14 @@ struct llama_server_context
         queue_results.push_back(res);
     }
 
-    int request_completion(json data, bool infill)
+    int request_completion(json data, bool infill, bool embedding)
     {
         std::lock_guard<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
         task.data = data;
         task.infill_mode = infill;
+        task.embedding_mode = embedding;
         task.type = COMPLETION_TASK;
         queue_tasks.push_back(task);
         return task.id;
@@ -1376,7 +1380,7 @@ struct llama_server_context
                     {
                         LOG_TEE("slot unavailable\n");
                         // send error result
-                        send_error(task.id, "slot unavaliable");
+                        send_error(task.id, "slot unavailable");
                         return;
                     }
 
@@ -1388,6 +1392,7 @@ struct llama_server_context
                     slot->reset();
 
                     slot->infill = task.infill_mode;
+                    slot->embedding = task.embedding_mode;
                     slot->task_id = task.id;
 
                     if (!launch_slot_with_data(slot, task.data))
@@ -1695,7 +1700,7 @@ struct llama_server_context
                 }
 
                 // prompt evaluated for embedding
-                if (params.embedding)
+                if (slot.embedding)
                 {
                     send_embedding(slot);
                     slot.release();
@@ -1751,12 +1756,18 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("options:\n");
     printf("  -h, --help                show this help message and exit\n");
     printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N,  --threads N        number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.n_threads);
     printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
-    printf("  -c N,  --ctx-size N       size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  -c N, --ctx-size N        size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  --rope-scaling {none,linear,yarn}\n");
+    printf("                            RoPE frequency scaling method, defaults to linear unless specified by the model\n");
     printf("  --rope-freq-base N        RoPE base frequency (default: loaded from model)\n");
-    printf("  --rope-freq-scale N       RoPE frequency scaling factor (default: loaded from model)\n");
-    printf("  -b N,  --batch-size N     batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --rope-freq-scale N       RoPE frequency scaling factor, expands context by a factor of 1/N\n");
+    printf("  --yarn-ext-factor N       YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
+    printf("  --yarn-attn-factor N      YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
+    printf("  --yarn-beta-slow N        YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
+    printf("  --yarn-beta-fast N        YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
+    printf("  -b N, --batch-size N      batch size for prompt processing (default: %d)\n", params.n_batch);
     printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
     printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
@@ -1877,6 +1888,19 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_ctx = std::stoi(argv[i]);
         }
+        else if (arg == "--rope-scaling")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
+            else { invalid_param = true; break; }
+        }
         else if (arg == "--rope-freq-base")
         {
             if (++i >= argc)
@@ -1895,6 +1919,38 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.rope_freq_scale = std::stof(argv[i]);
         }
+        else if (arg == "--yarn-ext-factor")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_ext_factor = std::stof(argv[i]);
+        }
+        else if (arg == "--yarn-attn-factor")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_attn_factor = std::stof(argv[i]);
+        }
+        else if (arg == "--yarn-beta-fast")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_fast = std::stof(argv[i]);
+        }
+        else if (arg == "--yarn-beta-slow")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_slow = std::stof(argv[i]);
+        }
         else if (arg == "--memory-f32" || arg == "--memory_f32")
         {
             params.memory_f16 = false;
@@ -2209,8 +2265,8 @@ int main(int argc, char **argv)
 
     llama_backend_init(params.numa);
 
-    LOG_INFO("build info", {{"build", BUILD_NUMBER},
-                            {"commit", BUILD_COMMIT}});
+    LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER},
+                            {"commit", LLAMA_COMMIT}});
 
     LOG_INFO("system info", {
                                 {"n_threads", params.n_threads},
@@ -2274,7 +2330,7 @@ int main(int argc, char **argv)
     svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res)
             {
                 json data = json::parse(req.body);
-                const int task_id = llama.request_completion(data, false);
+                const int task_id = llama.request_completion(data, false, false);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
@@ -2329,7 +2385,7 @@ int main(int argc, char **argv)
     svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
             {
                 json data = json::parse(req.body);
-                const int task_id = llama.request_completion(data, true);
+                const int task_id = llama.request_completion(data, true, false);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
@@ -2433,7 +2489,7 @@ int main(int argc, char **argv)
                 {
                     prompt = "";
                 }
-                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false);
+                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true);
                 task_result result = llama.next_result(task_id);
                 return res.set_content(result.result_json.dump(), "application/json");
             });

examples/speculative/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} speculative.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)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/speculative/speculative.cpp

@@ -1,5 +1,3 @@
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 
@@ -39,9 +37,11 @@ int main(int argc, char ** argv) {
     // max number of parallel drafting sequences (i.e. tree branches)
     const int n_seq_dft = params.n_parallel;
 
-    // TODO: make this configurable
-    const float p_accept = 0.80f;
-    const float p_split  = 0.10f;
+    // probability threshold for accepting a token from the draft model
+    const float p_accept = params.p_accept;
+
+    // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
+    const float p_split  = params.p_split;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("speculative", "log"));

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

@@ -349,9 +349,9 @@ static struct ggml_tensor * llama_build_train_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx,
-            rope_freq_base, rope_freq_scale);
+        return ggml_rope_custom(
+            ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        );
     };
 
     set_name(tokens_input, "tokens_input");

flake.lock

@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1692799911,
-        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
+        "lastModified": 1694529238,
+        "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
+        "rev": "ff7b65b44d01cf9ba6a71320833626af21126384",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1698134075,
-        "narHash": "sha256-foCD+nuKzfh49bIoiCBur4+Fx1nozo+4C/6k8BYk4sg=",
+        "lastModified": 1698318101,
+        "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "8efd5d1e283604f75a808a20e6cde0ef313d07d4",
+        "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c",
         "type": "github"
       },
       "original": {

flake.nix

@@ -11,8 +11,7 @@
         meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
-        osSpecific = with pkgs; buildInputs ++
-        (
+        osSpecific = with pkgs; buildInputs ++ (
           if isAarch64 && isDarwin then
             with pkgs.darwin.apple_sdk_11_0.frameworks; [
               Accelerate
@@ -96,12 +95,15 @@
         };
         packages.rocm = pkgs.stdenv.mkDerivation {
           inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_HIPBLAS=1"
             "-DCMAKE_C_COMPILER=hipcc"
             "-DCMAKE_CXX_COMPILER=hipcc"
-            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+            # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
+            # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
+            # and select the line that matches the current nixpkgs version of rocBLAS.
+            "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
           ];
         };
         apps.llama-server = {

ggml-alloc.c

@@ -378,9 +378,13 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
-static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) {
+static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view, bool update_backend) {
     assert(view->view_src != NULL && view->view_src->data != NULL);
-    view->backend = view->view_src->backend;
+
+    if (update_backend) {
+        view->backend = view->view_src->backend;
+    }
+
     view->buffer  = view->view_src->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
@@ -394,7 +398,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            init_view(alloc, node);
+            init_view(alloc, node, true);
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -424,15 +428,14 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                                 AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
                                 node->view_src = view_src;
                                 view_src_hn->n_views += 1;
-                                init_view(alloc, node);
+                                init_view(alloc, node, false);
                                 return;
                             }
-                        }
-                        else {
+                        } else {
                             AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
                             node->view_src = parent;
                             p_hn->n_views += 1;
-                            init_view(alloc, node);
+                            init_view(alloc, node, false);
                             return;
                         }
                     }
@@ -463,7 +466,7 @@ size_t ggml_allocr_alloc_graph_n(
                 hash_get(ht, view_src)->n_views += 1;
                 if (node->buffer == NULL && node->data != NULL) {
                     // view of a pre-allocated tensor, didn't call init_view() yet
-                    init_view(alloc, node);
+                    init_view(alloc, node, true);
                 }
             }
 
@@ -474,7 +477,7 @@ size_t ggml_allocr_alloc_graph_n(
                 }
                 hash_get(ht, parent)->n_children += 1;
                 if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
-                    init_view(alloc, parent);
+                    init_view(alloc, parent, true);
                 }
             }
         }

ggml-cuda.cu

@@ -513,6 +513,15 @@ static __global__ void add_f16_f32_f16(const half * x, const float * y, half * d
     dst[i] = __hadd(x[i], __float2half(y[i]));
 }
 
+static __global__ void add_f16_f32_f32(const half * x, const float * y, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = __half2float(x[i]) + y[i];
+}
+
 static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -973,7 +982,7 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
 
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     if (row > nrows) return;
 
     const int num_blocks_per_row = ncols / QK_K;
@@ -1077,7 +1086,7 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 
 static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     if (row > nrows) return;
 
     const int num_blocks_per_row = ncols / QK_K;
@@ -1181,7 +1190,7 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
 
 static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     if (row > nrows) return;
     const int num_blocks_per_row = ncols / QK_K;
     const int ib0 = row*num_blocks_per_row;
@@ -1435,7 +1444,7 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
 
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     if (row > nrows) return;
 
     const int num_blocks_per_row = ncols / QK_K;
@@ -4245,7 +4254,7 @@ template <bool need_check> static __global__ void
 
 template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
 static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows) {
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
 
     if (row >= nrows) {
         return;
@@ -4285,7 +4294,7 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
 static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
     // qk = quantized weights per x block
     // qr = number of quantized weights per data value in x block
-    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
 
     if (row >= nrows) {
         return;
@@ -4484,11 +4493,41 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
-// rope == RoPE == rotary positional embedding
+static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
 
+struct rope_corr_dims {
+    float v[4];
+};
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static __device__ void rope_yarn(
+    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
+    float * cos_theta, float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+    }
+    *cos_theta = cosf(theta) * mscale;
+    *sin_theta = sinf(theta) * mscale;
+}
+
+// rope == RoPE == rotary positional embedding
 template<typename T, bool has_pos>
-static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale,
-                            const int p_delta_rows, const float theta_scale) {
+static __global__ void rope(
+    const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims
+) {
     const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (col >= ncols) {
@@ -4500,10 +4539,10 @@ static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t
     const int i2 = row/p_delta_rows;
 
     const int p = has_pos ? pos[i2] : 0;
-    const float p0 = p*freq_scale;
-    const float theta = p0*powf(theta_scale, col/2);
-    const float sin_theta = sinf(theta);
-    const float cos_theta = cosf(theta);
+    const float theta_base = p*powf(freq_base, -float(col)/ncols);
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
     const float x0 = x[i + 0];
     const float x1 = x[i + 1];
@@ -4513,8 +4552,10 @@ static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t
 }
 
 template<typename T, bool has_pos>
-static __global__ void rope_neox(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale,
-                                 const int p_delta_rows, const float theta_scale) {
+static __global__ void rope_neox(
+    const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims
+) {
     const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (col >= ncols) {
@@ -4525,11 +4566,14 @@ static __global__ void rope_neox(const T * x, T * dst, const int ncols, const in
     const int i = row*ncols + col/2;
     const int i2 = row/p_delta_rows;
 
+    // simplified from `(ib * ncols + col) * (-1 / ncols)`, where ib is assumed to be zero
+    const float cur_rot = -float(col)/ncols;
+
     const int p = has_pos ? pos[i2] : 0;
-    const float p0 = p*freq_scale;
-    const float theta = p0*powf(theta_scale, col/2);
-    const float sin_theta = sinf(theta);
-    const float cos_theta = cosf(theta);
+    const float theta_base = p*powf(freq_base, cur_rot);
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
     const float x0 = x[i + 0];
     const float x1 = x[i + ncols/2];
@@ -4538,8 +4582,10 @@ static __global__ void rope_neox(const T * x, T * dst, const int ncols, const in
     dst[i + ncols/2] = x0*sin_theta + x1*cos_theta;
 }
 
-static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const int32_t * pos, const float freq_scale,
-                                    const int p_delta_rows, const float theta_scale, const int n_ctx) {
+static __global__ void rope_glm_f32(
+    const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
+    int n_ctx
+) {
     const int col = blockDim.x*blockIdx.x + threadIdx.x;
     const int half_n_dims = ncols/4;
 
@@ -4551,7 +4597,7 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol
     const int i = row*ncols + col;
     const int i2 = row/p_delta_rows;
 
-    const float col_theta_scale = powf(theta_scale, col);
+    const float col_theta_scale = powf(freq_base, -2.0f*col/ncols);
      // FIXME: this is likely wrong
     const int p = pos != nullptr ? pos[i2] : 0;
 
@@ -4693,6 +4739,11 @@ static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, co
     add_f16_f32_f16<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
 }
 
+static void add_f16_f32_f32_cuda(const half * x, const float * y, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
+    add_f16_f32_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
+}
+
 static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
     const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
     mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
@@ -4816,7 +4867,8 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(1, block_num_y, 1);
+    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
+    const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -4825,7 +4877,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -4834,7 +4886,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -4843,7 +4895,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -4852,7 +4904,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -4862,7 +4914,7 @@ static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, f
     GGML_ASSERT(ncols % QK_K == 0);
     const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
     const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(32, ny, 1);
     dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -4871,7 +4923,7 @@ static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, f
     GGML_ASSERT(ncols % QK_K == 0);
     const int ny = 2 / K_QUANTS_PER_ITERATION;
     const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(32, ny, 1);
     dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -4880,7 +4932,7 @@ static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, f
     GGML_ASSERT(ncols % QK_K == 0);
     const int ny = 2 / K_QUANTS_PER_ITERATION;
     const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(32, ny, 1);
     dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -4895,7 +4947,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
     GGML_ASSERT(ncols % QK_K == 0);
     const int ny = 2 / K_QUANTS_PER_ITERATION;
     const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(32, ny, 1);
     dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -4903,7 +4955,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
 static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK4_0 == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4912,7 +4964,7 @@ static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK4_1 == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4921,7 +4973,7 @@ static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK5_0 == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4930,7 +4982,7 @@ static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK5_1 == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4939,7 +4991,7 @@ static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK8_0 == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4948,7 +5000,7 @@ static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4957,7 +5009,7 @@ static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4966,7 +5018,7 @@ static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4975,7 +5027,7 @@ static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -4984,7 +5036,7 @@ static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float *
 static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
@@ -5003,7 +5055,7 @@ static void convert_fp32_to_fp16_cuda(const void * vx, half * y, const int k, cu
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     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_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<1, 1, convert_f16>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
@@ -5570,40 +5622,54 @@ static void clamp_f32_cuda(const float * x, float * dst, const float min, const
 }
 
 template<typename T>
-static void rope_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
-                          const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
+static void rope_cuda(
+    const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+) {
     GGML_ASSERT(ncols % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nrows, num_blocks_x, 1);
     if (pos == nullptr) {
-        rope<T, false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+        rope<T, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
     } else {
-        rope<T, true><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+        rope<T, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
     }
 }
 
 template<typename T>
-static void rope_neox_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
-                          const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
+static void rope_neox_cuda(
+    const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+) {
     GGML_ASSERT(ncols % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nrows, num_blocks_x, 1);
     if (pos == nullptr) {
-        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
     } else {
-        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
     }
 }
 
-static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
-                              const int p_delta_rows, const float theta_scale, const int n_ctx, cudaStream_t stream) {
+static void rope_glm_f32_cuda(
+    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, int n_ctx, cudaStream_t stream
+) {
     GGML_ASSERT(ncols % 4 == 0);
     const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
     const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
     const dim3 block_nums(num_blocks_x, nrows, 1);
-    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale, n_ctx);
+    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx);
 }
 
 static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
@@ -5724,6 +5790,11 @@ static void ggml_cuda_pool_free(void * ptr, size_t size) {
     CUDA_CHECK(cudaFree(ptr));
 }
 
+static bool g_cublas_loaded = false;
+
+bool ggml_cublas_loaded(void) {
+    return g_cublas_loaded;
+}
 
 void ggml_init_cublas() {
     static bool initialized = false;
@@ -5737,7 +5808,12 @@ void ggml_init_cublas() {
         CUDA_CHECK(cudaDeviceSynchronize());
 #endif
 
-        CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
+        if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
+            initialized = true;
+            g_cublas_loaded = false;
+            return;
+        }
+
         GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
         int64_t total_vram = 0;
 #if defined(GGML_CUDA_FORCE_MMQ)
@@ -5785,6 +5861,7 @@ void ggml_init_cublas() {
         // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
 
         initialized = true;
+        g_cublas_loaded = true;
     }
 }
 
@@ -5996,7 +6073,10 @@ inline void ggml_cuda_op_add(
         add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
         add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        add_f16_f32_f32_cuda((const half *) src0_dd, src1_dd, dst_dd, ggml_nelements(src0), main_stream);
     } else {
+        fprintf(stderr, "src0->type: %d  dst->type: %d\n", src0->type, dst->type);
         GGML_ASSERT(false);
     }
 
@@ -6460,17 +6540,20 @@ inline void ggml_cuda_op_rope(
     const int64_t ne2 = dst->ne[2];
     const int64_t nrows = ggml_nrows(src0);
 
-    //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];
+    //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];
+    const int n_orig_ctx  = ((int32_t *) dst->op_params)[4];
+
     // RoPE alteration for extended context
-
-    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);
+    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
     const int32_t * pos = nullptr;
     if ((mode & 1) == 0) {
@@ -6482,24 +6565,39 @@ inline void ggml_cuda_op_rope(
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
+
     // compute
     if (is_glm) {
         GGML_ASSERT(false);
-        rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, n_ctx, main_stream);
+        rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream);
     } else if (is_neox) {
         GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet");
         if (src0->type == GGML_TYPE_F32) {
-            rope_neox_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+            rope_neox_cuda(
+                (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, main_stream
+            );
         } else if (src0->type == GGML_TYPE_F16) {
-            rope_neox_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+            rope_neox_cuda(
+                (const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, main_stream
+            );
         } else {
             GGML_ASSERT(false);
         }
     } else {
         if (src0->type == GGML_TYPE_F32) {
-            rope_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+            rope_cuda(
+                (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, main_stream
+            );
         } else if (src0->type == GGML_TYPE_F16) {
-            rope_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+            rope_cuda(
+                (const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, main_stream
+            );
         } else {
             GGML_ASSERT(false);
         }
@@ -6610,8 +6708,10 @@ inline void ggml_cuda_op_clamp(
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const float min = ((float *) dst->op_params)[0];
-    const float max = ((float *) dst->op_params)[1];
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
 
     clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
     CUDA_CHECK(cudaGetLastError());
@@ -6804,6 +6904,8 @@ static void ggml_cuda_op_mul_mat(
     int64_t  row_low[GGML_CUDA_MAX_DEVICES];
     int64_t row_high[GGML_CUDA_MAX_DEVICES];
 
+    int used_devices = 0;
+
     for (int64_t id = 0; id < g_device_count; ++id) {
         // by default, use all rows
         row_low[id]  = 0;
@@ -6831,6 +6933,8 @@ static void ggml_cuda_op_mul_mat(
             continue;
         }
 
+        used_devices++;
+
         const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
         const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
 
@@ -6869,12 +6973,12 @@ static void ggml_cuda_op_mul_mat(
 
     // if multiple devices are used they need to wait for the main device
     // here an event is recorded that signals that the main device has finished calculating the input data
-    if (split && g_device_count > 1) {
+    if (split && used_devices > 1) {
         CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0]));
     }
 
-    const int64_t src1_col_stride = split && g_device_count > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
+    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
     for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
         const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0;
         const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
@@ -6990,6 +7094,9 @@ static void ggml_cuda_op_mul_mat(
     }
 
     for (int64_t id = 0; id < g_device_count; ++id) {
+        if ((!split && id != g_main_device) || row_low[id] == row_high[id]) {
+            continue;
+        }
         CUDA_CHECK(ggml_cuda_set_device(id));
 
         // free buffers again when done
@@ -7014,6 +7121,9 @@ static void ggml_cuda_op_mul_mat(
 
         CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         for (int64_t id = 0; id < g_device_count; ++id) {
+            if (row_low[id] == row_high[id]) {
+                continue;
+            }
             for (int64_t is = 0; is < is_max; ++is) {
                 CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
             }
@@ -7059,6 +7169,8 @@ static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src
 }
 
 bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    if (!g_cublas_loaded) return false;
+
     const int64_t ne10 = src1->ne[0];
 
     const int64_t ne0 = dst->ne[0];
@@ -7135,6 +7247,30 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
 }
 
+__global__ void k_compute_batched_ptrs(
+        const half * src0_as_f16, const half * src1_as_f16, half * dst_f16,
+        const void ** ptrs_src, void ** ptrs_dst,
+        int ne12, int ne13,
+        int ne23,
+        int nb02, int nb03,
+        int nb12, int nb13,
+        int nb2, int nb3,
+        int r2, int r3) {
+    int i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    int i12 = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i13 >= ne13 || i12 >= ne12) {
+        return;
+    }
+
+    int i03 = i13 / r3;
+    int i02 = i12 / r2;
+
+    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02   + i03*nb03;
+    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)     dst_f16 + i12* nb2/2 + i13* nb3/2;
+}
+
 static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
@@ -7236,49 +7372,45 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     } else {
         // use cublasGemmBatchedEx
-        // TODO: https://github.com/ggerganov/llama.cpp/pull/3749#discussion_r1369997000
         const int ne23 = ne12*ne13;
 
-        // TODO: avoid this alloc
-        void ** ptrs = (void **) malloc(3*ne23*sizeof(void *));
+        const void ** ptrs_src = nullptr;
+              void ** ptrs_dst = nullptr;
 
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
+        size_t ptrs_src_s = 0;
+        size_t ptrs_dst_s = 0;
 
-                ptrs[0*ne23 + i12 + i13*ne12] = (char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3];
-                ptrs[1*ne23 + i12 + i13*ne12] = (char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2;
-                ptrs[2*ne23 + i12 + i13*ne12] = (char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2;
-            }
-        }
+        ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s);
+        ptrs_dst = (      void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s);
 
-        // allocate device memory for pointers
-        void ** ptrs_as = nullptr;
-        CUDA_CHECK(cudaMalloc(&ptrs_as, 3*ne23*sizeof(void *)));
-
-        // TODO: this does not work for some reason -- not sure why?
-        //size_t ptrs_s = 0;
-        //ptrs_as = (void **) ggml_cuda_pool_malloc(3*ne23*sizeof(void *), &ptrs_s);
-
-        // copy pointers to device
-        CUDA_CHECK(cudaMemcpy(ptrs_as, ptrs, 3*ne23*sizeof(void *), cudaMemcpyHostToDevice));
-
-        free(ptrs);
+        dim3 block_dims(ne13, ne12);
+        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+                src0_as_f16, src1_as_f16, dst_f16,
+                ptrs_src, ptrs_dst,
+                ne12, ne13,
+                ne23,
+                nb02, nb03,
+                nb12, nb13,
+                dst->nb[2], dst->nb[3],
+                r2, r3);
+        CUDA_CHECK(cudaGetLastError());
 
         CUBLAS_CHECK(
         cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                &alpha_f16, (const void **) (ptrs_as + 0*ne23), CUDA_R_16F, nb01/sizeof(half),
-                            (const void **) (ptrs_as + 1*ne23), CUDA_R_16F, nb11/sizeof(float),
-                &beta_f16,  (      void **) (ptrs_as + 2*ne23), CUDA_R_16F, ne01,
+                &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, nb01/sizeof(half),
+                            (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, nb11/sizeof(float),
+                &beta_f16,  (      void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
                 ne23,
                 CUBLAS_COMPUTE_16F,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
 
-        // free device memory for pointers
-        CUDA_CHECK(cudaFree(ptrs_as));
-        //ggml_cuda_pool_free(ptrs_as, ptrs_s);
+        if (ptrs_src_s != 0) {
+            ggml_cuda_pool_free(ptrs_src, ptrs_src_s);
+        }
+        if (ptrs_dst_s != 0) {
+            ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s);
+        }
     }
 #endif
 
@@ -7291,10 +7423,12 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool all_on_device =
-        (src0->backend == GGML_BACKEND_GPU) &&
+        (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
         (src1->backend == GGML_BACKEND_GPU) &&
         ( dst->backend == GGML_BACKEND_GPU);
 
+    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+
     int64_t min_compute_capability = INT_MAX;
     for (int64_t id = 0; id < g_device_count; ++id) {
         if (min_compute_capability > g_compute_capabilities[id] && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
@@ -7316,13 +7450,13 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+    if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         // KQ single-batch
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
+    } else if (!split && all_on_device && use_tensor_cores && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
     } else if (src0->type == GGML_TYPE_F32) {
@@ -7722,6 +7856,8 @@ void ggml_cuda_free_scratch() {
 }
 
 bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
+    if (!g_cublas_loaded) return false;
+
     ggml_cuda_func_t func;
     const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
         || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))

ggml-cuda.h

@@ -17,7 +17,12 @@ extern "C" {
 
 #define GGML_CUDA_MAX_DEVICES       16
 
+// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
 GGML_API void   ggml_init_cublas(void);
+
+// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
+GGML_API bool   ggml_cublas_loaded(void);
+
 GGML_API void * ggml_cuda_host_malloc(size_t size);
 GGML_API void   ggml_cuda_host_free(void * ptr);
 

ggml-impl.h

@@ -0,0 +1,237 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // __ARM_NEON
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+#endif
+
+    // TODO: backend v2 PR
+
+#ifdef __cplusplus
+}
+#endif

ggml-metal.m

@@ -1001,11 +1001,15 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            const int nth = MIN(32, ne00);
+                            int nth = 32; // SIMD width
 
                             if (ne00%4 == 0) {
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
                             } else {
+                                do {
+                                    nth *= 2;
+                                } while (nth <= ne00 && nth <= 1024);
+                                nth /= 2;
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max];
                             }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -1013,8 +1017,9 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:MAX(16, nth/32*sizeof(float)) atIndex:0];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DIAG_MASK_INF:
                         {
@@ -1343,7 +1348,7 @@ 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:MAX(16, nth*sizeof(float)) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -1395,14 +1400,19 @@ void ggml_metal_graph_compute(
 
                             const int nth = MIN(1024, ne00);
 
-                            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_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];
+                            // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
+                            const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
 
-                            float freq_base;
-                            float freq_scale;
-                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
+                            float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+                            memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+                            memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+                            memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+                            memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+                            memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+                            memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
                             switch (src0->type) {
                                 case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
@@ -1410,30 +1420,35 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false);
                             };
 
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:2];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
+                            [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
+                            [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
+                            [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                            [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                            [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                            [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                            [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                            [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;

ggml-metal.metal

@@ -184,36 +184,73 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY;
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
+    float lmax = tpitg < ne00 ? psrc0[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) {
         lmax = MAX(lmax, psrc0[i00]);
     }
-    const float max = simd_max(lmax);
+
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         const float exp_psrc0 = exp(psrc0[i00] - max);
         lsum += exp_psrc0;
         // Remember the result of exp here. exp is expensive, so we really do not
-        // whish to compute it twice.
+        // wish to compute it twice.
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = simd_sum(lsum);
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
 
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
+
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         pdst[i00] /= sum;
     }
 }
@@ -224,37 +261,73 @@ kernel void kernel_soft_max_4(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
-    float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY;
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    float4 lmax4 = tpitg < ne00/4 ? psrc4[tpitg] : -INFINITY;
+
+    for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) {
         lmax4 = fmax(lmax4, psrc4[i00]);
     }
-    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
 
-    const float max = simd_max(lmax);
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         const float4 exp_psrc4 = exp(psrc4[i00] - max);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
-    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
 
-    const float sum = simd_sum(lsum);
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
 
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
+
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         pdst4[i00] /= sum;
     }
 }
@@ -274,7 +347,7 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
-     }
+    }
 }
 
 kernel void kernel_diag_mask_inf_8(
@@ -988,6 +1061,45 @@ kernel void kernel_alibi_f32(
     }
 }
 
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
+    return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
+}
+
 typedef void (rope_t)(
         device const    void * src0,
         device const int32_t * src1,
@@ -1011,8 +1123,13 @@ typedef void (rope_t)(
         constant         int & n_past,
         constant         int & n_dims,
         constant         int & mode,
+        constant         int & n_orig_ctx,
         constant       float & freq_base,
         constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]);
@@ -1041,8 +1158,13 @@ kernel void kernel_rope(
         constant         int & n_past,
         constant         int & n_dims,
         constant         int & mode,
+        constant         int & n_orig_ctx,
         constant       float & freq_base,
         constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]) {
@@ -1052,19 +1174,22 @@ kernel void kernel_rope(
 
     const bool is_neox = mode & 2;
 
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
+
     device const int32_t * pos = src1;
 
     const int64_t p = pos[i2];
 
-    const float theta_0 = freq_scale * (float)p;
+    const float theta_0 = (float)p;
     const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
         for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
 
             const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
-            const float cos_theta = cos(theta);
-            const float sin_theta = sin(theta);
+            float cos_theta, sin_theta;
+            rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
             device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
             device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -1079,9 +1204,12 @@ kernel void kernel_rope(
         for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
             for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
 
-                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
-                const float cos_theta = cos(theta);
-                const float sin_theta = sin(theta);
+                // simplified from `(ib * n_dims + ic) * inv_ndims`
+                const float cur_rot = inv_ndims*ic - ib;
+
+                const float theta = theta_0 * pow(freq_base, cur_rot);
+                float cos_theta, sin_theta;
+                rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
                 const int64_t i0 = ib*n_dims + ic/2;
 

ggml-quants.h

@@ -1,11 +1,63 @@
 #pragma once
 
-#include "ggml.h"
+#include "ggml-impl.h"
+
+// GGML internal header
 
 #include <stdint.h>
-#include <assert.h>
 #include <stddef.h>
 
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
 // Super-block size
 #ifdef GGML_QKK_64
 #define QK_K 64
@@ -15,18 +67,6 @@
 #define K_SCALE_SIZE 12
 #endif
 
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-
-//
-// Super-block quantization structures
-//
-
 // 2-bit quantization
 // weight is represented as x = a * q + b
 // 16 blocks of 16 elements each
@@ -127,6 +167,13 @@ static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_
 
 
 // Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
 void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
 void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
 void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
@@ -134,6 +181,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
 void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
 
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
 void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
@@ -142,6 +196,13 @@ void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
 
 // Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
 void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
 void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
 void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
@@ -150,16 +211,14 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
 
 // Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
 void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-// Quantization with histogram collection
-size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-

ggml.h

@@ -219,7 +219,7 @@
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          64
-#define GGML_MAX_OP_PARAMS     32
+#define GGML_MAX_OP_PARAMS     64
 #define GGML_DEFAULT_N_THREADS 4
 
 #if UINTPTR_MAX == 0xFFFFFFFF
@@ -1326,8 +1326,13 @@ extern "C" {
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
+            int                   n_orig_ctx,
             float                 freq_base,
-            float                 freq_scale);
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
@@ -1337,8 +1342,17 @@ extern "C" {
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
+            int                   n_orig_ctx,
             float                 freq_base,
-            float                 freq_scale);
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    // compute correction dims for YaRN RoPE scaling
+    void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // xPos RoPE, in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
@@ -1358,8 +1372,13 @@ extern "C" {
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
+            int                   n_orig_ctx,
             float                 freq_base,
             float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow,
             float                 xpos_base,
             bool                  xpos_down);
 
@@ -1930,12 +1949,19 @@ extern "C" {
     // quantization
     //
 
+    // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
     GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
 
+    GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+
     GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
 
     //
@@ -2001,6 +2027,7 @@ extern "C" {
     GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
     GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
     GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
     GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);

gguf-py/gguf/gguf.py

@@ -7,7 +7,7 @@ import shutil
 import struct
 import sys
 import tempfile
-from enum import IntEnum, auto
+from enum import Enum, IntEnum, auto
 from io import BufferedWriter
 from pathlib import Path
 from typing import IO, Any, BinaryIO, Callable, Sequence
@@ -53,9 +53,12 @@ KEY_ATTENTION_LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
 KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
 
 # RoPE
-KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count"
-KEY_ROPE_FREQ_BASE       = "{arch}.rope.freq_base"
-KEY_ROPE_SCALE_LINEAR    = "{arch}.rope.scale_linear"
+KEY_ROPE_DIMENSION_COUNT         = "{arch}.rope.dimension_count"
+KEY_ROPE_FREQ_BASE               = "{arch}.rope.freq_base"
+KEY_ROPE_SCALING_TYPE            = "{arch}.rope.scaling.type"
+KEY_ROPE_SCALING_FACTOR          = "{arch}.rope.scaling.factor"
+KEY_ROPE_SCALING_ORIG_CTX_LEN    = "{arch}.rope.scaling.original_context_length"
+KEY_ROPE_SCALING_FINETUNED       = "{arch}.rope.scaling.finetuned"
 
 # tokenization
 KEY_TOKENIZER_MODEL      = "tokenizer.ggml.model"
@@ -390,6 +393,7 @@ class TensorNameMap:
             "layers.{bid}.attention_norm",                         # llama-pth
             "encoder.layer.{bid}.attention.output.LayerNorm",      # bert
             "language_model.encoder.layers.{bid}.input_layernorm", # persimmon
+            "model.layers.{bid}.ln1",                              # yi
         ),
 
         # Attention norm 2
@@ -461,6 +465,7 @@ class TensorNameMap:
             "layers.{bid}.ffn_norm",                                        # llama-pth
             "encoder.layer.{bid}.output.LayerNorm",                         # bert
             "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
+            "model.layers.{bid}.ln2",                                       # yi
         ),
 
         # Feed-forward up
@@ -577,6 +582,11 @@ class TokenType(IntEnum):
     UNUSED       = 5
     BYTE         = 6
 
+class RopeScalingType(Enum):
+    NONE   = 'none'
+    LINEAR = 'linear'
+    YARN   = 'yarn'
+
 #
 # implementation
 #
@@ -636,18 +646,17 @@ class GGUFValueType(IntEnum):
             sys.exit()
 
 
+class WriterState(Enum):
+    EMPTY   = auto()
+    HEADER  = auto()
+    KV_DATA = auto()
+    TI_DATA = auto()
+
+
 class GGUFWriter:
     fout: BufferedWriter
-    arch: str
-    offset_tensor = 0
-    data_alignment = GGUF_DEFAULT_ALIGNMENT
-    kv_data = b""
-    kv_data_count = 0
-    ti_data = b""
-    ti_data_count = 0
-    use_temp_file: bool
-    temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None
-    tensors: list[tuple[np.ndarray[Any, Any], int]]
+    temp_file: tempfile.SpooledTemporaryFile[bytes] | None
+    tensors: list[np.ndarray[Any, Any]]
 
     @property
     def pack_prefix(self):
@@ -673,27 +682,47 @@ class GGUFWriter:
             GGUFValueType.FLOAT64: f"{self.pack_prefix}d",
             GGUFValueType.BOOL:    "?" ,
         }
-        self.add_architecture()
+        self.offset_tensor = 0
+        self.data_alignment = GGUF_DEFAULT_ALIGNMENT
+        self.kv_data = b""
+        self.kv_data_count = 0
+        self.ti_data = b""
+        self.ti_data_count = 0
         self.use_temp_file = use_temp_file
+        self.temp_file = None
         self.tensors = []
         endianess_str = "Big Endian" if self.endianess == GGUFEndian.BIG else "Little Endian"
         print(f"This gguf file is for {endianess_str} only")
+        self.state = WriterState.EMPTY
+
+        self.add_architecture()
 
     def write_header_to_file(self):
+        if self.state is not WriterState.EMPTY:
+            raise ValueError(f'Expected output file to be empty, got {self.state}')
+
         self.fout.write(struct.pack("<I", GGUF_MAGIC))
         self.fout.write(struct.pack(f"{self.pack_prefix}I", GGUF_VERSION))
         self.fout.write(struct.pack(f"{self.pack_prefix}Q", self.ti_data_count))
         self.fout.write(struct.pack(f"{self.pack_prefix}Q", self.kv_data_count))
         self.flush()
-#        print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
+        self.state = WriterState.HEADER
 
     def write_kv_data_to_file(self):
+        if self.state is not WriterState.HEADER:
+            raise ValueError(f'Expected output file to contain the header, got {self.state}')
+
         self.fout.write(self.kv_data)
         self.flush()
+        self.state = WriterState.KV_DATA
 
     def write_ti_data_to_file(self):
+        if self.state is not WriterState.KV_DATA:
+            raise ValueError(f'Expected output file to contain KV data, got {self.state}')
+
         self.fout.write(self.ti_data)
         self.flush()
+        self.state = WriterState.TI_DATA
 
     def add_key(self, key: str):
         self.add_val(key, GGUFValueType.STRING, add_vtype=False)
@@ -786,6 +815,9 @@ class GGUFWriter:
         return ((x + n - 1) // n) * n
 
     def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None):
+        if self.state is not WriterState.EMPTY:
+            raise ValueError(f'Expected output file to be empty, got {self.state}')
+
         assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
 
         encoded_name = name.encode("utf8")
@@ -815,23 +847,22 @@ class GGUFWriter:
         shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape
         self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype)
 
-        pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
-
-        if  self.temp_file is None:
-            self.tensors.append((tensor, pad))
+        if self.temp_file is None:
+            self.tensors.append(tensor)
             return
 
         tensor.tofile(self.temp_file)
+        self.write_padding(self.temp_file, tensor.nbytes)
 
-        if pad != 0:
-            self.temp_file.write(bytes([0] * pad))
-
-    def write_padding(self, fp: BinaryIO, n: int, align: int | None = None):
+    def write_padding(self, fp: IO[bytes], n: int, align: int | None = None):
         pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
         if pad != 0:
             fp.write(bytes([0] * pad))
 
     def write_tensor_data(self, tensor: np.ndarray[Any, Any]):
+        if self.state is not WriterState.TI_DATA:
+            raise ValueError(f'Expected output file to contain tensor info, got {self.state}')
+
         if self.endianess==GGUFEndian.BIG:
             tensor.byteswap(inplace=True)
         self.write_padding(self.fout, self.fout.tell())
@@ -844,10 +875,13 @@ class GGUFWriter:
         self.write_padding(self.fout, self.fout.tell())
 
         if self.temp_file is None:
-            for (currtensor, currpad) in self.tensors:
-                currtensor.tofile(self.fout)
-                if currpad != 0:
-                    self.fout.write(bytes([0] * currpad))
+            while True:
+                try:
+                    tensor = self.tensors.pop(0)
+                except IndexError:
+                    break
+                tensor.tofile(self.fout)
+                self.write_padding(self.fout, tensor.nbytes)
             return
 
         self.temp_file.seek(0)
@@ -948,8 +982,17 @@ class GGUFWriter:
     def add_rope_freq_base(self, value: float):
         self.add_float32(KEY_ROPE_FREQ_BASE.format(arch=self.arch), value)
 
-    def add_rope_scale_linear(self, value: float):
-        self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value)
+    def add_rope_scaling_type(self, value: RopeScalingType):
+        self.add_string(KEY_ROPE_SCALING_TYPE.format(arch=self.arch), value.value)
+
+    def add_rope_scaling_factor(self, value: float):
+        self.add_float32(KEY_ROPE_SCALING_FACTOR.format(arch=self.arch), value)
+
+    def add_rope_scaling_orig_ctx_len(self, value: int):
+        self.add_uint32(KEY_ROPE_SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)
+
+    def add_rope_scaling_finetuned(self, value: bool):
+        self.add_bool(KEY_ROPE_SCALING_FINETUNED.format(arch=self.arch), value)
 
     def add_tokenizer_model(self, model: str):
         self.add_string(KEY_TOKENIZER_MODEL, model)
@@ -983,11 +1026,8 @@ class GGUFWriter:
 
 
 class SpecialVocab:
-    load_merges: bool = False
-    merges: list[str] = []
-    special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad')
-    special_token_ids: dict[str, int] = {}
-    n_vocab: int | None = None
+    merges: list[str]
+    special_token_ids: dict[str, int]
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
@@ -997,8 +1037,11 @@ class SpecialVocab:
         self.special_token_ids = {}
         self.n_vocab = n_vocab
         self.load_merges = load_merges
+        self.merges = []
         if special_token_types is not None:
             self.special_token_types = special_token_types
+        else:
+            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad')
         self._load(Path(path))
 
     def _load(self, path: Path) -> None:

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.4.5"
+version = "0.4.6"
 description = "Write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

llama.h

@@ -106,6 +106,14 @@ extern "C" {
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
+    enum llama_rope_scaling_type {
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
+    };
+
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -167,15 +175,21 @@ extern "C" {
     };
 
     struct llama_context_params {
-        uint32_t seed;            // RNG seed, -1 for random
-        uint32_t n_ctx;           // text context, 0 = from model
-        uint32_t n_batch;         // prompt processing maximum batch size
-        uint32_t n_threads;       // number of threads to use for generation
-        uint32_t n_threads_batch; // number of threads to use for batch processing
+        uint32_t seed;              // RNG seed, -1 for random
+        uint32_t n_ctx;             // text context, 0 = from model
+        uint32_t n_batch;           // prompt processing maximum batch size
+        uint32_t n_threads;         // number of threads to use for generation
+        uint32_t n_threads_batch;   // number of threads to use for batch processing
+        int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float rope_freq_base;  // RoPE base frequency, 0 = from model
-        float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
+        float    rope_freq_base;   // RoPE base frequency, 0 = from model
+        float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, NaN = from model
+        float    yarn_attn_factor; // YaRN magnitude scaling factor
+        float    yarn_beta_fast;   // YaRN low correction dim
+        float    yarn_beta_slow;   // YaRN high correction dim
+        uint32_t yarn_orig_ctx;    // YaRN original context size
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
@@ -191,6 +205,7 @@ extern "C" {
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
     } llama_model_quantize_params;
 
     // grammar types
@@ -286,6 +301,23 @@ extern "C" {
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
+    // Functions to access the model's GGUF metadata scalar values
+    // - The functions return the length of the string on success, or -1 on failure
+    // - The output string is always null-terminated and cleared on failure
+    // - GGUF array values are not supported by these functions
+
+    // Get metadata value as a string by key name
+    LLAMA_API int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size);
+
+    // Get the number of metadata key/value pairs
+    LLAMA_API int llama_model_meta_count(const struct llama_model * model);
+
+    // Get metadata key name by index
+    LLAMA_API int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size);
+
+    // Get metadata value as a string by index
+    LLAMA_API int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size);
+
     // Get a string describing the model type
     LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
 
@@ -333,17 +365,14 @@ extern "C" {
     LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
             "avoid using this, it will be removed in the future, instead - count the tokens in user code");
 
-    // Remove all tokens data of cells in [c0, c1)
-    // c0 < 0 : [0,  c1]
-    // c1 < 0 : [c0, inf)
-    LLAMA_API void llama_kv_cache_tokens_rm(
-            struct llama_context * ctx,
-                         int32_t   c0,
-                         int32_t   c1);
+    // Clear the KV cache
+    LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
@@ -600,6 +629,13 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
+    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    LLAMA_API void llama_sample_min_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
     LLAMA_API void llama_sample_tail_free(
             struct llama_context * ctx,

mypy.ini

@@ -3,3 +3,4 @@ strict = true
 allow_untyped_calls = true
 allow_untyped_defs = true
 allow_incomplete_defs = true
+disable_error_code = import-untyped

scripts/build-info.cmake

@@ -1,5 +1,5 @@
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in")
-set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
 set(BUILD_NUMBER 0)
 set(BUILD_COMMIT "unknown")
 set(BUILD_COMPILER "unknown")
@@ -24,15 +24,21 @@ if(Git_FOUND)
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         OUTPUT_VARIABLE HEAD
         OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
     )
+    if (RES EQUAL 0)
+        set(BUILD_COMMIT ${HEAD})
+    endif()
     execute_process(
         COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         OUTPUT_VARIABLE COUNT
         OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
     )
-    set(BUILD_COMMIT ${HEAD})
-    set(BUILD_NUMBER ${COUNT})
+    if (RES EQUAL 0)
+        set(BUILD_NUMBER ${COUNT})
+    endif()
 endif()
 
 if(MSVC)
@@ -53,22 +59,22 @@ else()
     set(BUILD_TARGET ${OUT})
 endif()
 
-# Only write the header if it's changed to prevent unnecessary recompilation
-if(EXISTS ${HEADER_FILE})
-    file(READ ${HEADER_FILE} CONTENTS)
-    string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS})
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
     set(OLD_COMMIT ${CMAKE_MATCH_1})
-    string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
     set(OLD_COMPILER ${CMAKE_MATCH_1})
-    string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
     set(OLD_TARGET ${CMAKE_MATCH_1})
     if (
         NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
         NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
         NOT OLD_TARGET   STREQUAL BUILD_TARGET
     )
-        configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
     endif()
 else()
-    configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
 endif()

scripts/build-info.h.in

@@ -1,9 +0,0 @@
-#ifndef BUILD_INFO_H
-#define BUILD_INFO_H
-
-#define BUILD_NUMBER @BUILD_NUMBER@
-#define BUILD_COMMIT "@BUILD_COMMIT@"
-#define BUILD_COMPILER "@BUILD_COMPILER@"
-#define BUILD_TARGET "@BUILD_TARGET@"
-
-#endif // BUILD_INFO_H

scripts/build-info.sh

@@ -24,12 +24,7 @@ if out=$($CC -dumpmachine); then
   build_target=$out
 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_COMPILER \"$build_compiler\""
-echo "#define BUILD_TARGET \"$build_target\""
-echo
-echo "#endif // BUILD_INFO_H"
+echo "int LLAMA_BUILD_NUMBER = ${build_number};"
+echo "char const *LLAMA_COMMIT = \"${build_commit}\";"
+echo "char const *LLAMA_COMPILER = \"${build_compiler}\";"
+echo "char const *LLAMA_BUILD_TARGET = \"${build_target}\";"

scripts/server-llm.sh

@@ -0,0 +1,391 @@
+#!/bin/bash
+#
+# Helper script for deploying llama.cpp server with a single Bash command
+#
+# - Works on Linux and macOS
+# - Supports: CPU, CUDA, Metal, OpenCL
+# - Can run all GGUF models from HuggingFace
+# - Can serve requests in parallel
+# - Always builds latest llama.cpp from GitHub
+#
+# Limitations
+#
+# - Chat templates are poorly supported (base models recommended)
+# - Might be unstable!
+#
+# Usage:
+#   ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose]
+#
+#   --port:       port number, default is 8888
+#   --repo:       path to a repo containing GGUF model files
+#   --wtype:      weights type (f16, q8_0, q4_0, q4_1), default is user-input
+#   --backend:    cpu, cuda, metal, opencl, depends on the OS
+#   --gpu-id:     gpu id, default is 0
+#   --n-parallel: number of parallel requests, default is 8
+#   --n-kv:       KV cache size, default is 4096
+#   --verbose:    verbose output
+#
+# Example:
+#
+#   bash -c "$(curl -s https://ggml.ai/server-llm.sh)"
+#
+
+set -e
+
+# required utils: curl, git, make
+if ! command -v curl &> /dev/null; then
+    printf "[-] curl not found\n"
+    exit 1
+fi
+if ! command -v git &> /dev/null; then
+    printf "[-] git not found\n"
+    exit 1
+fi
+if ! command -v make &> /dev/null; then
+    printf "[-] make not found\n"
+    exit 1
+fi
+
+# parse arguments
+port=8888
+repo=""
+wtype=""
+backend="cpu"
+
+# if macOS, use metal backend by default
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    backend="metal"
+elif command -v nvcc &> /dev/null; then
+    backend="cuda"
+fi
+
+gpu_id=0
+n_parallel=8
+n_kv=4096
+verbose=0
+
+function print_usage {
+    printf "Usage:\n"
+    printf "  ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose]\n\n"
+    printf "  --port:       port number, default is 8888\n"
+    printf "  --repo:       path to a repo containing GGUF model files\n"
+    printf "  --wtype:      weights type (f16, q8_0, q4_0, q4_1), default is user-input\n"
+    printf "  --backend:    cpu, cuda, metal, opencl, depends on the OS\n"
+    printf "  --gpu-id:     gpu id, default is 0\n"
+    printf "  --n-parallel: number of parallel requests, default is 8\n"
+    printf "  --n-kv:       KV cache size, default is 4096\n"
+    printf "  --verbose:    verbose output\n\n"
+    printf "Example:\n\n"
+    printf '  bash -c "$(curl -s https://ggml.ai/server-llm.sh)"\n\n'
+}
+
+while [[ $# -gt 0 ]]; do
+    key="$1"
+    case $key in
+        --port)
+            port="$2"
+            shift
+            shift
+            ;;
+        --repo)
+            repo="$2"
+            shift
+            shift
+            ;;
+        --wtype)
+            wtype="$2"
+            shift
+            shift
+            ;;
+        --backend)
+            backend="$2"
+            shift
+            shift
+            ;;
+        --gpu-id)
+            gpu_id="$2"
+            shift
+            shift
+            ;;
+        --n-parallel)
+            n_parallel="$2"
+            shift
+            shift
+            ;;
+        --n-kv)
+            n_kv="$2"
+            shift
+            shift
+            ;;
+        --verbose)
+            verbose=1
+            shift
+            ;;
+        --help)
+            print_usage
+            exit 0
+            ;;
+        *)
+            echo "Unknown argument: $key"
+            print_usage
+            exit 1
+            ;;
+    esac
+done
+
+# available weights types
+wtypes=("F16" "Q8_0" "Q4_0" "Q4_1" "Q5_0" "Q5_1" "Q6_K" "Q5_K_M" "Q5_K_S" "Q4_K_M" "Q4_K_S" "Q3_K_L" "Q3_K_M" "Q3_K_S" "Q2_K")
+
+wfiles=()
+for wt in "${wtypes[@]}"; do
+    wfiles+=("")
+done
+
+# sample repos
+repos=(
+    "https://huggingface.co/TheBloke/Llama-2-7B-GGUF"
+    "https://huggingface.co/TheBloke/Llama-2-13B-GGUF"
+    "https://huggingface.co/TheBloke/Llama-2-70B-GGUF"
+    "https://huggingface.co/TheBloke/CodeLlama-7B-GGUF"
+    "https://huggingface.co/TheBloke/CodeLlama-13B-GGUF"
+    "https://huggingface.co/TheBloke/CodeLlama-34B-GGUF"
+    "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF"
+    "https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF"
+    "https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-GGUF"
+    "https://huggingface.co/TheBloke/CausalLM-7B-GGUF"
+)
+
+printf "\n"
+printf "[I] This is a helper script for deploying llama.cpp's server on this machine.\n\n"
+printf "    Based on the options that follow, the script might download a model file\n"
+printf "    from the internet, which can be a few GBs in size. The script will also\n"
+printf "    build the latest llama.cpp source code from GitHub, which can be unstable.\n"
+printf "\n"
+printf "    Upon success, an HTTP server will be started and it will serve the selected\n"
+printf "    model using llama.cpp for demonstration purposes.\n"
+printf "\n"
+printf "    Please note:\n"
+printf "\n"
+printf "    - All new data will be stored in the current folder\n"
+printf "    - The server will be listening on all network interfaces\n"
+printf "    - The server will run with default settings which are not always optimal\n"
+printf "    - Do not judge the quality of a model based on the results from this script\n"
+printf "    - Do not use this script to benchmark llama.cpp\n"
+printf "    - Do not use this script in production\n"
+printf "    - This script is only for demonstration purposes\n"
+printf "\n"
+printf "    If you don't know what you are doing, please press Ctrl-C to abort now\n"
+printf "\n"
+printf "    Press Enter to continue ...\n\n"
+
+read
+
+if [[ -z "$repo" ]]; then
+    printf "[+] No repo provided from the command line\n"
+    printf "    Please select a number from the list below or enter an URL:\n\n"
+
+    is=0
+    for r in "${repos[@]}"; do
+        printf "    %2d) %s\n" $is "$r"
+        is=$((is+1))
+    done
+
+    # ask for repo until index of sample repo is provided or an URL
+    while [[ -z "$repo" ]]; do
+        printf "\n    Or choose one from: https://huggingface.co/models?sort=trending&search=gguf\n\n"
+        read -p "[+] Select repo: " repo
+
+        # check if the input is a number
+        if [[ "$repo" =~ ^[0-9]+$ ]]; then
+            if [[ "$repo" -ge 0 && "$repo" -lt ${#repos[@]} ]]; then
+                repo="${repos[$repo]}"
+            else
+                printf "[-] Invalid repo index: %s\n" "$repo"
+                repo=""
+            fi
+        elif [[ "$repo" =~ ^https?:// ]]; then
+            repo="$repo"
+        else
+            printf "[-] Invalid repo URL: %s\n" "$repo"
+            repo=""
+        fi
+    done
+fi
+
+# remove suffix
+repo=$(echo "$repo" | sed -E 's/\/tree\/main$//g')
+
+printf "[+] Checking for GGUF model files in %s\n" "$repo"
+
+# find GGUF files in the source
+# TODO: better logic
+model_tree="${repo%/}/tree/main"
+model_files=$(curl -s "$model_tree" | grep -i "\\.gguf</span>" | sed -E 's/.*<span class="truncate group-hover:underline">(.*)<\/span><\/a>/\1/g')
+
+# list all files in the provided git repo
+printf "[+] Model files:\n\n"
+for file in $model_files; do
+    # determine iw by grepping the filename with wtypes
+    iw=-1
+    is=0
+    for wt in "${wtypes[@]}"; do
+        # uppercase
+        ufile=$(echo "$file" | tr '[:lower:]' '[:upper:]')
+        if [[ "$ufile" =~ "$wt" ]]; then
+            iw=$is
+            break
+        fi
+        is=$((is+1))
+    done
+
+    if [[ $iw -eq -1 ]]; then
+        continue
+    fi
+
+    wfiles[$iw]="$file"
+
+    have=" "
+    if [[ -f "$file" ]]; then
+        have="*"
+    fi
+
+    printf "    %2d) %s %s\n" $iw "$have" "$file"
+done
+
+# ask for weights type until provided and available
+while [[ -z "$wtype" ]]; do
+    printf "\n"
+    read -p "[+] Select weight type: " wtype
+    wfile="${wfiles[$wtype]}"
+
+    if [[ -z "$wfile" ]]; then
+        printf "[-] Invalid weight type: %s\n" "$wtype"
+        wtype=""
+    fi
+done
+
+printf "[+] Selected weight type: %s (%s)\n" "$wtype" "$wfile"
+
+url="${repo%/}/resolve/main/$wfile"
+
+# check file if the model has been downloaded before
+chk="$wfile.chk"
+
+# check if we should download the file
+# - if $wfile does not exist
+# - if $wfile exists but $chk does not exist
+# - if $wfile exists and $chk exists but $wfile is newer than $chk
+# TODO: better logic using git lfs info
+
+do_download=0
+
+if [[ ! -f "$wfile" ]]; then
+    do_download=1
+elif [[ ! -f "$chk" ]]; then
+    do_download=1
+elif [[ "$wfile" -nt "$chk" ]]; then
+    do_download=1
+fi
+
+if [[ $do_download -eq 1 ]]; then
+    printf "[+] Downloading weights from %s\n" "$url"
+
+    # download the weights file
+    curl -o "$wfile" -# -L "$url"
+
+    # create a check file if successful
+    if [[ $? -eq 0 ]]; then
+        printf "[+] Creating check file %s\n" "$chk"
+        touch "$chk"
+    fi
+else
+    printf "[+] Using cached weights %s\n" "$wfile"
+fi
+
+# get latest llama.cpp and build
+
+printf "[+] Downloading latest llama.cpp\n"
+
+llama_cpp_dir="__llama_cpp_port_${port}__"
+
+if [[ -d "$llama_cpp_dir" && ! -f "$llama_cpp_dir/__ggml_script__" ]]; then
+    # if the dir exists and there isn't a file "__ggml_script__" in it, abort
+    printf "[-] Directory %s already exists\n" "$llama_cpp_dir"
+    printf "[-] Please remove it and try again\n"
+    exit 1
+elif [[ -d "$llama_cpp_dir" ]]; then
+    printf "[+] Directory %s already exists\n" "$llama_cpp_dir"
+    printf "[+] Using cached llama.cpp\n"
+
+    cd "$llama_cpp_dir"
+    git reset --hard
+    git fetch
+    git checkout origin/master
+
+    cd ..
+else
+    printf "[+] Cloning llama.cpp\n"
+
+    git clone https://github.com/ggerganov/llama.cpp "$llama_cpp_dir"
+fi
+
+# mark that that the directory is made by this script
+touch "$llama_cpp_dir/__ggml_script__"
+
+if [[ $verbose -eq 1 ]]; then
+    set -x
+fi
+
+# build
+cd "$llama_cpp_dir"
+
+make clean
+
+log="--silent"
+if [[ $verbose -eq 1 ]]; then
+    log=""
+fi
+
+if [[ "$backend" == "cuda" ]]; then
+    printf "[+] Building with CUDA backend\n"
+    LLAMA_CUBLAS=1 make -j server $log
+elif [[ "$backend" == "cpu" ]]; then
+    printf "[+] Building with CPU backend\n"
+    make -j server $log
+elif [[ "$backend" == "metal" ]]; then
+    printf "[+] Building with Metal backend\n"
+    make -j server $log
+elif [[ "$backend" == "opencl" ]]; then
+    printf "[+] Building with OpenCL backend\n"
+    LLAMA_CLBLAST=1 make -j server $log
+else
+    printf "[-] Unknown backend: %s\n" "$backend"
+    exit 1
+fi
+
+# run the server
+
+printf "[+] Running server\n"
+
+args=""
+if [[ "$backend" == "cuda" ]]; then
+    export CUDA_VISIBLE_DEVICES=$gpu_id
+    args="-ngl 999"
+elif [[ "$backend" == "cpu" ]]; then
+    args="-ngl 0"
+elif [[ "$backend" == "metal" ]]; then
+    args="-ngl 999"
+elif [[ "$backend" == "opencl" ]]; then
+    args="-ngl 999"
+else
+    printf "[-] Unknown backend: %s\n" "$backend"
+    exit 1
+fi
+
+if [[ $verbose -eq 1 ]]; then
+    args="$args --verbose"
+fi
+
+./server -m "../$wfile" --host 0.0.0.0 --port "$port" -c $n_kv -np "$n_parallel" $args
+
+exit 0

tests/test-double-float.cpp

@@ -4,7 +4,7 @@
 
 #undef NDEBUG
 #include <cassert>
-#if !defined(__riscv) && !defined(__s390__)
+#if !defined(__riscv) && !defined(__s390__) && !defined(__ARM_NEON)
 #include <immintrin.h>
 #endif
 #include <cmath>

tests/test-quantize-fns.cpp

@@ -129,6 +129,13 @@ int main(int argc, char * argv[]) {
         ggml_type type = (ggml_type) i;
         ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
 
+        // deprecated - skip
+        if (qfns.blck_size == 0) {
+            continue;
+        }
+
+        printf("Testing %s\n", ggml_type_name((ggml_type) i));
+
         if (qfns.from_float && qfns.to_float) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
             const float max_quantization_error =